Imagine you’re Joshua Pederson, a humanities professor at Boston University, whose 12-year-old son is battling a very aggressive cancer. Every day you awaken hoping that today will be better than the last, but also grappling with the chance that things will turn dire. You grasp at an olive branch of hope, in the form of an experimental drug you’ve been told is your best hope. A team of physicians, invested and experienced, tells you this is your best hope — possibly a godsend. Then you find out a betting market exists where people can bet on the outcomes of clinical trials. Your son’s meds aren’t on it—yet—but there’s no guarantee they won’t be. If they appear there, random people might be betting against your son’s chances. How do you feel?

The prediction platform Kalshi, where these bets are being made, justifies this reality by asserting that the gamblers flocking to the platform to bet for and against the drug that may save your son are doing you a service — surfacing perceptions by those looking for a fast buck who, as the platform’s story goes, may act as a counterweight to the dedicated doctors working to save your son.

Here’s how Joshua felt: angry at the “death market” he took as a personal affront. Who can blame him?

This is the coarsening of society, distilled to a cash grab.

You’ve always been able to bet on drugs, at least those being developed by public companies. But this is different. It’s more specific than anything you could place before. Joshua’s anger is visceral. It cuts through all philosophical debates that followed Kalshi’s July launch of drug-outcome betting. You might think this provides a solution to an information problem you didn’t know you had, as Kalshi suggests. But what if it were your kid?

Kalshi argues that their rules currently limit their clinical trial bets to late-stage putative drugs already fully enrolled, and they won’t take bets where every subject is a minor. But “currently” and “every” are doing a lot of heavy lifting there.

We don’t often connect betting on financial outcomes in the confines of the stock market to morality on a human level. That’s partly because bets are hidden behind relatively more benign buy/sell stock exchanges, so they don’t feel so personal. And that’s partly because a company is not perceived as just a drug—even when, for small companies, that can actually be the case.

You’ve always been able to bet on a public company. If you were providing funding as a VC, you could bet on private ones too — though the general public has been pretty much out of luck on that front. What the new markets have introduced is the ability to bet on a single drug, not a portfolio. When a big pharma entity announces a new success, it may fail to move the stock price at all. A certain level of success is already built into a big company’s portfolio, and operating in the shadow of a blockbuster can hide the significance of something else entirely. The new markets decouple this. If there’s a single drug you feel compelled about from a Lilly or a Pfizer, that can now be the totality of your bet. Their stock may not move — it might even drop, depending on other drugs in the portfolio or the whims of the market. But you can still win. And if you hear about a private company developing a drug, you can get in.

This, fundamentally, is what the new markets are adding. Not necessarily better perspective on whether a drug will succeed. The chance to make a highly targeted bet.

That sounds exciting. But stop and think about it. The reason big pharma stocks have historically been a big part of retirement baskets is that those stocks have generally been stable. That’s because a blockbuster allows for many failures. You get less of the upside when something succeeds, but you’re protected from any single failure. Betting on the markets decouples that cushion, and in a sense can be seen as an analogy to leverage. If you bet right, you get rich. If your bet fails, you get poor way more quickly. Ask Leopold Aschenbrenner, whose Situational Awareness fund lost $35 billion in a matter of days, how leverage, explicit or implicit, works out when things go sideways. Perhaps the biggest difference between the leveraged investor and the betting markets is that someone in the betting markets may not even realize that they’ve traded stability for leverage, upside, and collapse

Even worse, it’s not clear who’s actually betting or why — which raises the question of how much meaning anyone should extrapolate from the numbers.

There are a bunch of assumptions baked into the view that prediction markets reflect statistical truth. The reality is that they are, at best, sentiment barometers — weighted by the segment of the population with the will and interest to bet. If you peel away the modern tech sheen, they aggregate a bunch of votes heavily distorted by wealth inequalities and behavioral biases. Anyone who has ever made lower-tech bets, such as horse races, where bet spreads reflect participant biases in a similar manner, knows that if you bet the favorite every time, you’ll go home broke.

The modern prediction platforms are worse, because while horse racing is a pretty proletarian pursuit at the end of the day, these platforms are heavily wealth-weighted. The prediction values are not what the average man on the street believes — even if enough average men on the street participated. They’re weighted, sometimes very heavily, toward those who have the most they can afford to lose. The big bets crush the voice of the small ones. And the concept of the "whale" — the bettor selected for the size of his bankroll, not the quality of his information — transfers from Vegas to Polymarket with ease. What appears to be a consensus can easily be what a well-capitalized but unremarkable gambler woke up with a hunch about.

Then there’s the appeal of the long shot. Everyone wants to get rich quick. You can convince yourself of almost anything worth betting on being possible. So if someone is offering you 100-1 odds on something that seems kinda possible, even if it’s objectively unlikely, maybe you bet on it. When this happens systematically, the chances of the rare event get overestimated, and the odds of the thing that’s obviously going to happen get underestimated. No one’s ever seen a major league baseball team earn 20 runs in an inning, but if the odds are large enough, say 5000-1, you can be sure someone’s gonna take that bet. Because, well, you never know. But, honestly, you kinda do.

And then there’s the “yes” bias. Studies show people lean toward betting on the affirmative outcome of an event. In sales, it’s known that pricing can achieve a psychological advantage — $3.99 sounds appreciably cheaper than $4.00. It “feels” cheaper. A bet on “yes” can “feel” better. People like to bet on outcomes based on what they’d like to happen, rather than what they might objectively expect.

That’s not to suggest that categorically the prediction markets are always bereft of signal. There’s evidence that prediction markets sometimes get things right—just in other domains. This can happen where the aggregate bets, reflecting the bettors, average out signal that’s noisy, but signal nonetheless. Where the bets follow a Gaussian distribution around reality. Think, for example, where you collect thousands of guesses about how many jelly beans are in a big jar. No one likely gets it exactly right, but the average guess may be surprisingly close. The key there is that everyone can see the jar, they have the same information, and everyone’s bet is informed (however loosely). Plenty of prediction market bets are basically that — shared-jar bets. No one has a lock on the knowledge to be sure, but the aggregate bet contains signal. That’s not true of drugs, though.

Subscribe now

Last week, we got to see, in real time, what happens when the people who can see the jar are barred from bidding.

The prediction markets couldn’t believe they got the results of the Wisconsin Democratic gubernatorial primary so wrong. The primary, held August 11, 2026, is basically a masterclass in how prediction markets can go horribly wrong. Prior to the election, Polymarket gave progressive candidate Francesca Hong a 96% chance of winning (Kalshi had the number at 95%). It was all over but the counting. Polling agreed with the direction but not the certainty.

But wait, not so fast. When the dust cleared, Hong actually lost to the more moderate candidate, David Crowley. The results were close (39.8% of the vote for Crowley to 39.3% for Hong), but they were decidedly for the guy who was, the night before, dead, buried, with a concession speech left on his kitchen table for anyone who cared. How could this have possibly happened?

As it turns out, Wisconsin law makes it a felony to bet on an election you’ve voted in. In the run-up to the vote, the Wisconsin Elections Commission was aggressive in reminding people of that law. So very few of those betting on the election were actually from the state. The betting was biased against precisely those with the greatest chance of having real information — no signal from boots-on-the-ground campaigners, no signal from those who may have been talking to their friends and family. Just a lot of speculators from elsewhere, amplifying whatever the political parties or the national news media were already saying.

That 96% wasn’t necessarily an informed number. It was amplified by external sources, missing the signal from those most likely to actually have a clue. In the end, it was more noise — or a fairy tale told by those who might have vested interests — than an actual harbinger.

The proof was in the open-mouthed stammering the next morning. Polymarket, for its part, quietly deleted its own 96% prediction once the votes came in.

The same disconnect applies to biotech, and even more problematically. You can argue that an out-of-state organizer or campaign operative could have enough access to the election to make a prediction using high-quality data. This example suggests otherwise, though it’s certainly possible in principle. In the realm of drug discovery, the exclusions are far wider and actually incorporate nearly anyone who knows much about the ongoing prospects of the project.

For a clinical candidate, many of those who work for the drug company — including pretty much anyone involved in the clinical trials — commit a federal crime if they place a bet. Oncologist Mark Ratain said it succinctly, noting the markets serve as another opportunity for his colleagues to lose money and go to jail. That isn’t to say those with insider knowledge won’t try to bet, themselves or through a third party. But it does mean that if they make those bets, they’ll be discreet and probably modest in size, to avoid drawing attention.

So the big bets are, as with Wisconsin, coming from those without the insider knowledge to make highly informed bets. The shape of the roulette wheel changes, but you’re still betting on red or black. Two ecosystems, two examples, and to be entirely clear and honest, there are two claims here that need to be untwisted. The first is that people with the most information to place informed bets are banned from doing so. The second claim is that, therefore, nobody left knows anything.

The pushback will be, reasonably, that if you work in pharma, anywhere, you have a better sense of what constitutes a good target or a good drug class. To some degree, that’s true. But what you don’t have is enough knowledge of human biology to really be sure what’s gonna happen in the clinic. You don’t have it. No one does until those trials are carried out — even if you’re a corporate insider. Companies make educated bets, but you can infer, correctly, from the way the stock price barely moves on a Phase I announcement — and sometimes doesn’t move much on Phase II — that insiders and the market still have trouble affixing a strong probability to clinical progress. If it were easy, everyone working in pharma would be rich. Some get lucky and get rich, but it’s not nearly that easy.

If those who work deepest in the field have trouble with this, what makes anyone think crowdsourcing — by people locked out of the room entirely — is going to have better answers? To be clear, that’s not “nobody left knows anything.” But it is that the average germane knowledge of the bettors skews low, and the certainty of the message you might take from the betting pool is lower than the number suggests—because the people closest to the more reliable knowledge can’t place a bet.

When you consider all these caveats — the wealth-weighting, the long-shot and yes biases, and the often-missing signal from the people who have the best information but are locked out by law — you realize the markets are far different than they’re being sold as. They aren’t reliable forecasting engines. What they are is a way that gamblers with a bit of relevant information can make money by arbitraging against those who know less. When it comes to drug bets, they aren’t providing useful information on the drug. They’re providing information on the betting pool.

When the odds move on a putative drug, you’re probably not witnessing the market inform you about misallocated resources in pharma. Instead, it’s the sum of biases, along with potentially a few bettors who may know a bit more than average and who may get to the bet faster — but who don’t know nearly what the company developing the drug actually knows. And even they don’t know enough to know enough.

You’re seeing a market that can be both exploitable and meaningless simultaneously. The odds tell you about the betting crowd. They tell you little about the drug — possibly nothing at all. Compare that to a bet built upon the only kind of information that may move the needle. A bet that is truly life-or-death.

Somewhere, there's a 12-year-old still in the fight, with a real shot at life because his doctors looked at the data with eyes that have years of experience. They’ve made bets, too—informed and specific bets built on everything years of training—and possibly access to non-public information—can provide. No prediction market can make that bet for a child; nothing a hazy set of maybe-I’ll-get-rich gamblers on the Internet might imply could possibly add value. The focus is intense. The knowledge is at the top of what’s knowable. And the payoff, if it all works, will be priceless.

A drug isn’t judged only by what it’s meant to accomplish. It’s judged by what happens when the system built to deliver it can’t keep up. That’s the real subject of this essay — not the molecule, but the plumbing.

Most of us inherited our sense of what psychedelics are from stories we never fact-checked. Art Linkletter, on national television days after his daughter’s death in 1969, telling the country LSD had murdered her. Timothy Leary: turn on, tune in, drop out. I was ten years old in 1969, in a household about as far from drug culture as you could get. Even I knew the Linkletter story. The freak-out casualties of the ‘60s and ‘70s, baked into the culture as a settled verdict on the whole category of drug. What almost nobody remembers: the autopsy on Diane Linkletter found no drugs in her system at all. Her father kept blaming LSD anyway, and by the time the truth caught up with the story, the story had already done its work — Congress heard it, the public believed it, and psychedelics spent the next half-century filed under “unserious,” “dangerous,” or both.

About a decade later, I would begin forty years in the trenches of computational drug discovery. That does something to your baseline skepticism. You learn to recognize the disconnect between simple models on a pitch deck and the messy, unforgiving realities of human biology. Combinatorial chemistry was supposed to be the ultimate solution. Then it was solving the proteome. Now it’s AI-designed everything. Entire revolutions, complete with absolute certainty and compressed timelines, have come and gone.

Now, the industry has pivoted its focus to psychedelics, the optimistically termed “psychedelic renaissance.” A few weeks ago, Eli Lilly took a fraction of the trillion-dollar windfall it is printing from its GLP-1 weight-loss empire and dropped it into an entirely different realm of human neurobiology: a $3.8 billion acquisition of AtaiBeckley — $2.8 billion upfront, another $1 billion in contingent milestones. A few years ago, that would have been jaw-dropping. Today, it’s simply seen as the latest and biggest wager yet in an investment wave already underway: AbbVie committed $1.2 billion to Gilgamesh Pharmaceuticals’ bretisilocin program roughly a year earlier, and Otsuka paid $700 million for Transcend Therapeutics in March 2026. Lilly’s bet is louder and larger than either, but it’s the third major Big Pharma psychedelics acquisition in about a year, not the first. It’s a public company sitting on an increasing cash pile, realizing they need to spend their money on something, and watching two peers already do it.

The headline prize is BPL-003, an intranasal formulation of mebufotenin (5-MeO-DMT), derived originally from the venom of the Sonoran Desert toad and known informally as the “God molecule.” (Lilly is actually buying a small pipeline here — two other neuroplastogens ride along in the deal — but BPL-003 is the asset carrying the thesis, so it’s the one worth interrogating.) I’ll hold the more-money-than-God jokes.

The scientific rationale is genuinely elegant. But Wall Street's consensus that Big Pharma capital will effortlessly mainstream these treatments relies on a foundational misunderstanding of the clinical chassis. I call this the Psychedelic Scaling Fallacy: the mistaken assumption that you can take a high-touch, psychologically volatile, profoundly human therapeutic experience and brute-force it through the low-touch, automated distribution network built for daily SSRIs. This isn't a critique of Lilly's judgment — the fallacy is a market-wide one, not Lilly's strategic logic in pursuing it. If anything, the AtaiBeckley deal may be the cleanest version of this bet available anywhere in the category, and BPL-003 the smartest molecule to make it with.

Zoom out far enough and this isn’t really a story about one acquisition. It’s a story about a real gap in psychiatric medicine that’s been sitting there for at least sixty years: the distance between drugs the culture has decided are legitimately “medical” and drugs it has decided are “recreational,” as though a compound has to pick one identity and keep it forever. It doesn’t. Psychedelics never got the uneasy peace this country made with plenty of other drugs that are both legitimately therapeutic and routinely abused. The Linkletter-and-Leary-shaped bias has had fifty years to calcify. We’re still near the beginning of prying it loose.

Esketamine went first, and its path is instructive: ketamine has its own long history as a club drug (one that continues), and that history didn’t sink Spravato’s approval. But mebufotenin (BPL-003) has a steeper climb than esketamine did. That includes a shorter pharmaceutical track record to borrow credibility from. More significantly, it also includes dealing with the ongoing subjective prejudice that comes with a drug that facilitates full ego-dissolution rather than simple dissociation. A drug that fits the recreational stereotype far more precisely than a nasal spray does. That’s part of why the second half of this essay gets more speculative than the first: less “here is what happened,” more “here is what has to happen, and to whom, for this to actually work.” That’s the part I find most interesting, and it’s where the rest of this piece is headed.

The Ghost of Delysid: We’ve Actually Been Here Before

The Linkletter myth is the story people of a certain age remember. It’s not the one that actually mattered institutionally. Before we look at the financial math, we need to look at the history that got buried under the mythology — because the pharmaceutical industry has its own, separate case of amnesia, and it’s a more relevant one than the cultural version. The Linkletter myth is a story about culture; this is a story about regulation. The prevailing narrative today is that Big Pharma is finally waking up to the therapeutic magic of psychedelics. The reality is that Big Pharma invented this commercial space and then actively abandoned it.

In 1938, Albert Hofmann synthesized LSD at Sandoz Laboratories in Switzerland, accidentally discovering its potent psychoactive properties five years later. By 1947, Sandoz was manufacturing it and marketing it to psychiatrists worldwide under the trade name Delysid; a decade later, after Hofmann isolated the active compounds in magic mushrooms, it did the same with psilocybin as Indocybin. This wasn’t fringe medicine — Sandoz-supplied LSD was used to treat roughly 40,000 patients and generated over 1,000 medical and scientific publications, with studies showing remarkable promise for chronic alcoholism and severe neuroses.

So why did it die? The popular narrative blames Leary, the counterculture, and the 1970 Controlled Substances Act (CSA)— the same cultural current that produced the Linkletter story. But there was a second, structural assassin, and this one had nothing to do with moral panic. The 1960s brought a wholesale transformation in FDA oversight, and rigid, double-blind, randomized controlled trials became the price of admission for proving efficacy. LSD psychotherapy, with its deep reliance on the “set and setting” of the room, simply didn’t fit the new, sterile mold. Sandoz pulled the plug on LSD production in 1965 because the regulatory chassis had changed underneath them — before Leary was a household name, before the CSA existed. The reputation problem was cultural, and mostly overblown. The business problem was regulatory, and it’s the one almost nobody remembers.

We’ve seen this exact pattern replay recently, just with a different orthodoxy standing in for the double-blind RCT: earlier this year the FDA told uniQure that its one-time Huntington’s gene therapy needed a sham-surgery placebo arm. That would mean years of a fatal, progressive disease with no treatment, to satisfy a trial design built for pills, not brain surgery. Same movie, sixty years later—though this time the FDA revised course a few months later, and the story could continue.

Why try again now? Desperation in the Treatment-Resistant Depression market has intersected with unprecedented Big Pharma cash, the operational blueprint of in-clinic ketamine treatments, and — new since last year — an actual regulatory tailwind: an April 2026 executive order fast-tracking select psychedelic developers through the FDA’s National Priority Voucher program, followed last month by finalized FDA guidance for psychedelic drug developers specifically. Even skeptics of the voucher program’s political origins have said the plain reading of the guidance is that these drugs are now approvable, which is encouraging companies to look at psychedelics with a real investment lens. Pharma believes it can solve the structural failure of the 1960s with the pharmacology, and the newly cooperative regulators, of 2026. The need has been there forever. Legitimized pharmaceutical tools have been hibernating, waiting for a signal, for a rules change. More so than at any time since the ‘60s, it appears we’re at that alignment.

Optimizing Throughput: The Clock Problem

To understand why a sophisticated entity like Lilly is making this specific bet, look at the throughput economics of clinical psychiatry.

Classic psychedelics like psilocybin and LSD act primarily on the 5-HT2A receptor and produce a six-to-eight-hour subjective experience. If you’re running an interventional psychiatric clinic, a six-hour session destroys the unit economics of your billable day. A floating window that might be six, might be eight doesn’t work, either. Moreover, therapy is often grueling work, and finding a suitable number of skilled clinicians willing to sit through a whole day with someone hallucinating and emotionally raw, possibly dangerous to themselves or others, is completely unrealistic.

Mebufotenin, however, tilts markedly toward the 5-HT1A receptor rather than 5-HT2A — it still engages both, but binding studies put the selectivity at roughly 300- to 1000-fold in favor of 5-HT1A, pronounced enough to produce a rapid, intense ego-dissolution experience that resolves within an hour or two, rather than the slower, more visually complex trip driven by psilocybin’s or LSD’s 5-HT2A dominance. Lilly’s own announcement confirms the average in-clinic visit ran roughly two hours in the Phase 2a study, with beneficial effects persisting for months — a window that, not coincidentally, lines up almost exactly with an existing regulatory precedent, which the next section gets into. The efficacy data backs it up: 66.7% of patients saw rapid, statistically significant antidepressant responses by Day 2, durable through 85 days of follow-up. The molecule solves the clock problem—at least on the surface, based on topline numbers.

Optimizing for throughput carries its own risk. Strip out too much of the labor-intensive support that deep psychotherapy provides to maximize patient volume, and even if most patients resolve to a competent baseline in two hours, brain chemistry is unpredictable enough that some fraction won’t. Push throughput too far, and the structural integrity of the drug as a complex of alteration and therapy is at risk.

Phase 3 topline data isn’t expected until early 2029, so what’s been solved so far is the version of the clock problem that lives inside the treatment room under typical conditions. The much bigger corpus of proof, the multi-year run through the regulatory pathway, is only now starting to tick. Other candidates in the pipeline haven’t solved even that typical in-room version: some in-development treatments call for five or more hours of observation, and Definium’s LSD-based depression treatment, DT120, reportedly requires a full eight hours of patient supervision per dose. If the two-hour window holds up for BPL-003 through Phase 3, that’s a genuine structural advantage over a category where several competitors are still building their commercial case around a full workday per patient. It’s not hard to imagine that being the difference between a drug with a meaningful footprint and an interesting idea. That said, brain biochemistry is among the least well understood aspects of human physiology. I once watched two friends take the same dose of a dissociative drug at a concert: one came back fine and ready to move on, the other spent the day in fetal position. One data point proves nothing on its own, but it’s exactly the kind of variability the two-hour mean can’t see. Public disclosure from atai Life Sciences (who became AtaiBeckley, who Lilly bought) states about BPL-003 that “The average time to meet readiness for discharge criteria across all arms was within two hours of dosing, with the majority of patients deemed ready for discharge at the 90-minute post-dose assessment” (emphasis mine). Bell curves, even nice ones, can have long tails. If even one person requires eight hours, you need a system built to accommodate that as a potentiality, not a don’t-worry-be-happy exception.

Subscribe now

Borrowing Spravato’s Blueprint: Not so Fast

In computational chemistry and the data sciences, your knee-jerk reaction to the assertion that an approach that worked once can be broadly applied is “prove it.” The scuttlebutt dismisses worries about getting BPL-003 into providers’ hands because esketamine has already shown how it’s done. Looking more deeply, I’d say it’s shown how it can be done, once, for a different drug with different scheduling and risk-profile issues. But let’s drill down.

That two-hour window isn’t a coincidence — it lines up almost exactly with the monitoring period mandated under Spravato’s REMS, the FDA’s Risk Evaluation and Mitigation Strategy framework reserved for drugs whose risks require active management beyond a standard label. Esketamine patients must be observed by a healthcare provider for a minimum of two hours after dosing; that requirement is written directly into the label. The broader universe of off-label ketamine clinics offers a looser version of the same footprint — IV lines, in-chair monitoring, a similarly-staffed room — but Spravato is the tighter match, since it’s the one place that two-hour window is already a regulatory requirement rather than just a clinical convention.

What J&J itself did with that blueprint is telling. It built the certification framework, funded site enablement and physician education, and helped the certified network grow from roughly 5,000 settings in 2024 to more than 7,000 today — and in seven years, with the balance sheet to do otherwise, never bought a single one of those clinics. An incumbent that has watched a room business assemble itself around its own drug, and had every opportunity to own a piece of it, looked at that business and passed. Whatever infrastructure Lilly inherits by riding the same blueprint, Lilly is inheriting that same choice along with it.

But “similar footprint” is doing a lot of quiet work in that comparison. The infrastructure doesn’t just port over. A REMS isn’t a generic license a new drug can plug into — it’s a custom-built protocol designed around one specific drug’s specific risks, and Spravato’s was built to manage sedation, dissociation, and blood-pressure spikes, not an hour or more of active ego-dissolution and psychological terror. Esketamine doesn’t produce that. Mebufotenin can. Whatever REMS mebufotenin eventually gets will have to be negotiated from scratch around what this molecule actually does to people, a different and harder clinical picture than what Spravato was built to handle. The unsettling part is that ketamine has a well-understood, well-demonstrated high-then-gone cycle. Mebufotenin’s outcomes are more variable person to person, which means the REMS has to be built around the edge cases, not the average one.

Layer onto that the DEA side. 5-MeO-DMT is Schedule I; esketamine is Schedule III. The two carry different handling burdens — Schedule I requires its own separate DEA registration, DEA-approved narcotics safes rather than a standard locked cabinet, and recordkeeping kept apart from everything else in the building. A clinic already certified to dispense Spravato doesn’t inherit any of that; it would need to build it from zero, on top of a REMS that doesn’t exist yet, for a molecule whose Phase 3 data isn’t in until 2029. Generic off-label ketamine clinics are an even looser fit — they have the room, the IV line, and MD/NP staffing, but no formal certification requirement at all, since clinicians need standard continuing medical education (CME) rather than a dedicated national program, and the clinical event they’re built around is a comparatively contained dissociative state rather than the acute ego-dissolution a 5-MeO-DMT chaperone has to manage in real time — the skill set the Delivery Bottlenecks section covers below.

What Spravato actually offers isn’t a shortcut, it’s proof of concept — evidence that this kind of infrastructure can be built, sitting right next to a molecule for which nobody has built it yet. But the differences between the two drugs don’t resolve under an “it’ll be fine” rubric. It’s not just an outsider’s read, either — an Evercore equity analyst covering the space has made essentially the same point: much of the necessary infrastructure isn’t in place yet, and moving from what exists for Spravato isn’t a simple switch to flip. The pieces are there, but assembling them will take time.

Share

The Lykos Specter: Lessons From a Rejection

Worth restating before this gets grim: none of what follows is a Lilly-specific problem, or even a mebufotenin-specific one — it's the regulatory terrain every psychedelic drug has to cross, and BPL-003's two-hour window may in fact be better positioned to cross it than most.

We only need to look back a couple of years to see how this can play out under FDA scrutiny. In June 2024, an FDA advisory committee delivered a two-part verdict against Lykos Therapeutics's MDMA-assisted therapy for PTSD: 9-2 that the trial data didn’t demonstrate efficacy, and 10-1 that the treatment’s benefits didn’t outweigh its risks. The FDA followed with a formal Complete Response Letter in August.

The core problem was the method, not the molecule. Psychedelic trials run into a structural bind conventional drugs never face: functional unblinding. Give a patient a therapeutic dose of MDMA or 5-MeO-DMT and they know within minutes they’re not on placebo, corrupting expectation bias in ways statistical correction can’t fully undo. The Lykos trials compounded this with genuine conduct problems — boundary crossings by a therapist pair that led to retracted papers in Psychopharmacology, and an FDA reprimand for instructing sites not to record “positive” adverse events like euphoria. But even a cleanly run trial would still have had to reckon with the unblinding problem on its own terms.

AtaiBeckley and Lilly weren’t about to let this happen again, of course. The lesson they’ve drawn: if the therapist’s presence is what makes blinding indefensible, minimize the therapist’s role. BPL-003’s Phase 3 trials have stripped the psychotherapeutic protocol down to basic “psychological support” without active guiding, in pursuit of a cleaner double-blind. That’s a sensible response to a real regulatory constraint — but it trades one risk for another. 5-MeO-DMT is not a gentle ride: patients frequently experience acute terror, nausea, loss of motor control, and “reactivations” — flashback phenomena occurring days, weeks or even months after dosing. Whether minimal psychological support is enough scaffolding for that experience is a genuinely open empirical question, and it’s the same question the Delivery Bottlenecks argument below turns on. Flashbacks that occur on a timescale far outside any conceivable in-person visit offer their own risk that can’t be completely mitigated.

More broadly, it raises the question: Separating administration conditions from drug assessment meets FDA regulations, but is it a legitimately useful way to assess a drug when they are potentially so intertwined? Debate over this entanglement is unresolved.

A Maze of Unsettled Delivery Bottlenecks: Therapists and Insurance

Here’s where the thesis actually gets tested. Before a drug can become a blockbuster, even once approved, it has to reach patients. For typical drugs, that comes down to MDs, scripts, and insurance co-pays. A dissociative psychoactive drug is an entirely different problem. Getting the drug approved is the beginning. Availability is the comparatively easy part. Then the real difficulty starts.

Start with the therapists — the most under-discussed shortfall in the system, and the chasm between a drug’s availability and its adoption.

Administering 5-MeO-DMT (the active compound in BPL-003) puts the patient in a dissociative state — incredibly valuable for therapy, but potentially devastating without in-person guidance. And not just any therapist: you need a provider trained in altered states, somatic trauma release, and emotional integration. Certification programs — MAPS, Fluence — run $2,500 to $9,550 and take months, and those costs fall on the providers themselves. Even after that, they can only sit with the patient; they can’t write the script.

If that were the whole problem, it would be manageable — drug companies or clinics could subsidize the training. But it gets worse, because even once that hole is plugged, expected throughput is poor. The headline is that BPL-003 offers a therapeutic monitoring window of “only” two hours. But what about the patients who don’t come down that fast? You can’t just send them out the door or into the waiting room. That unpredictability has to be built into the system. A full psychedelic-assisted treatment protocol in current clinical trials can require more than 40 hours of therapist involvement, typically delivered by a two-clinician team, according to Prohibition Partners’ Psychedelics as Medicine report.

So you have almost no providers, and it’s a serious time sink for the ones who do get trained. Psilocybin, which has a head start on this build-out, shows what delivery is up against: Oregon has roughly 300, Colorado certified its first 51 in 2024. If psilocybin clears FDA approval on anything like its expected timeline, the number of trained practitioners needed runs into the tens of thousands, per HealingMaps’ analysis of the workforce gap.

Extrapolate that forward, and you’re not just looking at a gap; you’re looking at a delivery chasm. A huge one for a drug like BPL-003. Forty percent of the country already lives in a mental-health provider shortage area, and the system meets barely a quarter of existing need. That’s before you add a drug that requires a specially trained chaperone in the room. How fast can a chasm that size actually get bridged?

As if that weren’t enough, there’s insurance. The industry is built around a simple paradigm: code the session, code the diagnosis, get paid. Sessions run 15 or 45 minutes of talk therapy — not hours of monitoring and integration work. There’s no existing coding built for a session that might last two hours or might last five. A therapist can’t block out five hours and get paid for two; an insurer won’t want to pay for five and get two. This likely shifts toward reimbursement by procedure rather than by time, and those wheels turn slowly. Focusing on just the therapist’s reimbursement also hides a larger administrative mess.

The administrative burden behind even settling on billable hours is where tech-optimists lose the plot: J-codes for the drug, G-codes bundled for two or more hours of monitoring, claims routinely denied for mismatched codes or lapsed facility certifications. The bottleneck shifts from drug manufacturing to chair space and cross-billing friction between medical and pharmacy benefits. Anyone who has spent an afternoon on hold with an insurer over a single denied claim already knows this fight, and knows which side of the incentive wins. Cost containment beats care quality whenever the two conflict, and there’s no reason to expect that calculus changes here. Solvable — but not easily solvable.

Not enough qualified therapists. A practice model not built for multi-hour appointments. An insurance system not built for the variability. Even once the drug reaches market, you still have a plumbing problem.

For perspective, look at Spravato (esketamine). It took roughly six years post-launch to cross $1 billion in revenue, generating about $1.1 billion in 2024 before accelerating toward a $2 billion-plus run rate by Q4 2025. That’s a real achievement, and it’s the best evidence we have that clinic-based psychedelic-adjacent treatments can reach blockbuster scale. It probably played a significant role in Lilly’s decision. It’s also a reminder that “can scale” and “scales easily” are different claims — this took six years, mature REMS infrastructure, and over 100,000 cumulative patients. And Spravato is likely an easier case than mebufotenin will be.

A drug can create demand and help solve a therapeutic problem. But if it can’t be administered widely and profitably—if it can’t deal with the off-target limitations of the system—then it will ultimately fail.

Disassembling the Cultural Walls

Psychedelics are simultaneously everywhere and nowhere. A coalition of biohacking entrepreneurs and psychedelics-industry financiers pushes these therapies at the “Make America Healthy Again” summit alongside Robert F. Kennedy Jr. The traditional medical establishment remains fiercely skeptical, and it has the political leverage to slow momentum. Massachusetts, where I live, is the case study: in November 2024, voters rejected Question 4 — regulated psychedelic access — by 56 to 43 percent, with opposition clustered hardest in affluent, medically conservative suburbs like Wellesley and Lexington. The Massachusetts Medical Society led the charge, warning of new-onset psychosis and bipolar exacerbation: genuine clinical concern tangled up with turf protection.

You cannot scale a blockbuster if the medical establishment won’t prescribe it out of liability fears, or if a meaningful share of the country still sees it as a party drug wearing a lab coat — a fear the Linkletter story helped manufacture and that hasn’t fully worn off, even now that we know it was never true. The lie always runs faster than the truth.

Rewiring the Psychiatric Profession

If these therapies succeed commercially, expect a rewiring of the psychiatric profession. Psychiatry has spent decades shifting away from talk therapy toward brief pharmacological management, leaving deep psychotherapy largely to psychologists and social workers. Mainstreaming psychedelics won’t drag MD psychiatrists back into the room for three-hour sessions. Instead, the psychiatrist becomes a “treatment architect” — overseeing medical screening, REMS authorization, and acute prescribing — while a lower-cost tier of LMHCs and PsyDs handles the actual monitoring and integration.

That bifurcated model imports a liability question standard talk therapy never had to consider. If a clinic discharges a patient after the standard two-hour window and that patient has a serious adverse event afterward, who’s liable? The Lykos rejection already surfaced the ugliest version of this risk: patients under neuroplastogens are profoundly susceptible to suggestion, producing sexual misconduct and boundary violations that have no analog in a physician writing a script for Zoloft.

That bifurcated model also threatens to widen a pay gap that already aggravates the field: the therapist does the hard hour-long lift. The psychiatrist lifts the prescription pad at their request. The MD gets paid way more with less involvement. It’s conceivable that psychedelics might change that model, especially if therapists now have to deal with malpractice insurance at rates they’ve never seen before.

Revising who gets paid, who sits in the room, and how they get paid is one solution. Another, more radical one—but one that feels increasingly obvious—is to let the therapist write the prescription. There’s a movement, RxP, to give psychologists prescriptive authority after they complete additional postdoctoral training in clinical psychopharmacology. Predictably, the AMA opposes it as scope creep, but it’s gaining momentum, with eight states already on board—although only a trickle of under 300 psychologists so far credentialed under the program. Still, if it can break the roadblock, expect Big Pharma to fund efforts to push it far wider.

The professional boundary that has kept the psychologist and psychiatrist on opposite sides of the prescription pad for decades may be about to crack.

The Human Ledger

When looking into a new putative drug that I’m not familiar with, it usually goes like this: cool, cooler, oh wow, that might be really important. With psychedelics, it went the other way. I already knew how important they might be. I’ve talked to clinicians who’ve worked with them and sing their praises. I’ve watched people open up on these drugs in ways they might never otherwise. It doesn’t take a rocket scientist (or a computational chemist) to see the therapeutic promise here. But the deeper I looked into it, the harder it got. Not because the science is wrong. Not because the bet is wrong. But because, in this case, the delivery pipeline isn’t clear.

No matter who gets past the pharma finish line, they will still need the delivery infrastructure. And the demand for psychiatric care that current infrastructure can’t meet isn’t going away regardless of which molecule clears the FDA first. The most reliable bet in this realm may not be on the drugs at all. It may be on the pipes.

When I was writing this essay, I repeatedly asked myself why so little of the post-Lilly-announcement analysis mentioned the economics of delivery at all. To me, this is the most interesting, most unsettled part. Finally, a few days ago, SIVA Capital — an operator-led investment firm working in cannabis and psychedelics — published a memo reaching the same conclusion by a different route: delivery economics, not the science, is this deal’s least-appreciated risk. Where their read stops short is the labor pipeline itself — who’s credentialed to sit in that room, whether an unpredictable sit length survives contact with existing billing models, and who gets to do this work at all. Those are the constraints I’d argue this story ultimately turns on.

If BPL-003 delivers durable, months-long remission from a handful of clinic visits a year, Lilly’s $2.8 billion looks less like a gamble and more like the rare case where the company’s self-interest and the patient’s outcome point the same direction.

It also presents an interesting parallel: Lilly’s GLP-1 empire — the one that let them buy BPL-003 out of pocket change — was built by solving the problems that show up after you have a good molecule: manufacturing peptides at unprecedented scale, and a direct-to-employer distribution program that routes around pharmacy benefit managers. For psychedelics, Lilly needs to solve a similar problem again, only harder — the gatekeeper this time isn’t a supply chain or a PBM. It’s a room, a chaperone, and an insurance code.

What the market is underpricing isn’t the science. It’s the delivery. Manufacturing isn’t software scale. Molecules scale beautifully in a stainless-steel plant. Human attention does not.

The bottleneck that never shows up on a balance sheet is the one that will actually decide whether this bet pays off: the human being in the room.

My promise was one Long View essay a week, and only when there’s something worth saying. This is the third this week — not because the bar moved, but because the news cycle kept dragging me back to the keyboard. Saner cadence resumes after this.

Wednesday morning, Google confirmed it to the Financial Times: the AlphaFold team within DeepMind is being shut down, with those who aren’t leaving reassigned to other projects, including Gemini. Some in the AI community are shocked. It didn’t shock me. It’s tacit acknowledgement by one of the smartest and richest companies in Silicon Valley that what I’ve been suggesting in a recent set of essays is true (e.g., here, here, and here): AI as it defined itself in the first generation, for ideation, is hitting a wall. There will be additional progress, but the low-hanging fruit is gone, and the leaps are going to appear elsewhere, including better training sets, baking in properties such as synthesizability, and broader acknowledgement that the AI that’s going to change the biological world is going to live inside it, not trample it. Google’s bet is on Isomorphic, which they helped found — and which has since drawn outside capital on its own, a sign the market is pricing the same shift.

It took a Google to lead the charge because it takes a hugely profitable company to make decisions that are locally rational within the single silo. One thing you learn in computational chemistry, and you’d better learn it early on, is that a holistic view isn’t just an option. A ligand designed to make nice interactions with your target may look great on a slide, but ignore the waters it needs to displace to make those interactions at your own peril.

It’s the same with companies. AlphaFold is incredibly prestigious. If you are a company trying to do a fund raise, you want to be able to say you own AlphaFold. You want to say you have all the smartest and well-known AI acolytes working for you, too. But if you’re Google, your reputation is set: you already have one of the world’s great revenue streams from your profitable businesses like advertising and cloud services. AlphaFold is cool, and AlphaFold makes the world take you seriously when you announce other AI efforts like Gemini. But that prestige is now baked into perception rather than earned fresh each quarter. So if you’re a bean counter, you’re looking at the DeepMind efforts and asking AlphaFold specifically, “What have you done for me lately?” And, in terms of revenues, the answer would have to be, “not much.”

Ultimately, being rich and getting richer like Google allows you to make rational decisions on a division basis. AlphaFold isn’t moving the bar now, so you can shut it down. An Anthropic or an OpenAI doesn’t have that luxury. If they had AlphaFold, they would be using that to “prove” they can design drugs. They’d be using that to attract more investors. It would have value far in excess of the science, and there’s little chance anyone would possibly consider shuttering it.

The freedom of massive revenue streams, of being able to address business units without holistic constraint, allows Google to function as the canary in this particular coal mine.

And what that canary is singing is this: The value of AI has shifted from first-generation structure prediction and binding predictions towards platform-level integration. Towards Isomorphic and their like. That's a rational decision. It's the decision I've been arguing is the world as we are going to know it.

Google's decision officially marks the start of the second generation of bio-AI: Integrated Bio-AI. Not the actual start — plenty of AI-first companies were already headed this way — but the moment the shift became undeniable.

The initial analysis in the wake of the Google announcement has expectedly been less grounded and more OMG. John Jumper had already moved to Anthropic a month earlier; Jonas Adler and Alexander Pritzel followed. The initial reaction settled fast on one line: “Google is hemorrhaging talent.” But maybe they’re just the smartest players on the field. Maybe they read the same shift in value I'm describing here before the org chart caught up to it.

Alphabet’s stock dipped about 5% after the big-name departures became public. The market did what markets usually do: overreact first, assess it carefully later. Now that Google’s strategy is becoming more apparent with this week’s reveal, saner analysis will start to take place.

The writing has arguably been on the wall for a while. Wednesday made it official, but the slow bleed started earlier and most of the original AlphaFold paper authors were quietly reassigned over the past year. Since sealing AlphaFold behind a closed firewall, a lot of the frontier work on the platform has been happening at Isomorphic, including their IsoDDE functionality—where they claim substantial improvements in protein/ligand predictions and antibody/antigen predictions, as well as near generational leap performance on identification of cryptic pockets. As always, remember the Breakfast Test for disclosure—that until the details are published with rigor we need to accept them on faith—so many of these Isomorphic claims need to be interpreted through the “unpublished data” lens.

Subscribe now

DeepMind is also, assuredly, acutely aware that much of the recent reported progress in these models may be memorization dressed up as advancement—as the recent Mattsson and Walters paper elegantly demonstrated. And even if the models genuinely improve, interpolation beyond the training map is a problem that more scale alone will not solve, nor will models that look anything like what the field is still building.

A clear-eyed manager looking at AlphaFold, cutting through the coolness of name recognition and the breakthroughs that enabled the whole field, eventually asks the question, “What, exactly, are we doing here now?” When you’re in charge of a billion-dollar budget, that question has a way of focusing attention.

If you can take 10% of your single year budget for DeepMind and found a company like Isomorphic with it—a company with a goal of full platform drug discovery and not just improvement of AI performance—why wouldn’t you do it?

If AlphaFold’s dismantling tells you where the value has already migrated, the next question is what’s left once you strip that value out — and that’s where the real strategic split shows up. AI methods are no longer the moat in drug discovery — they’re the entry fee. It doesn’t matter which AI method you use—AlphaFold, OpenFold, Boltz, Chai, or another—you’re going to get most of the way to where AI can get you, because they’re all still being actively developed. Any moderate difference among them is going to get squeezed out over time, and if one moves ahead, the others are going to follow, leapfrog style. There’s plenty of academic juice to be had in pushing these models to the next plateau (and the open models are absolutely crucial to ensure that push remains accessible), and the whole field ultimately benefits. But a for-profit company isn’t likely to see this as the best use of their budget. If they invest at all, it will be in service of a broader platform, not to headline it: see Isomorphic.

So if AI methods aren’t the moat, what is? Agentic processes to train new models, synthesize, and assay results are in one place. Data is king, and that’s not changing. Fuller integration of AI with better physics-derived methods that can fill in the unseen is another. And, of course, humans in the loop. Not just AI experts. Biologists and chemists who can provide the intuition and world understanding that still eludes AI at any level.

That reallocation of focus and investment — from prediction to integration — is visible in recent pivots announced by both Google and Anthropic. Google remains all-in on AI, as witnessed by their Gemini efforts. They’re just sunsetting focused internal efforts on frontier biostructure infrastructure. But there’s a bit more to the story. Though the AlphaFold group is closing, at the 2026 Google I/O meeting, they also introduced “Gemini for Science” — a suite built around a hypothesis-generation tool, an agentic “computational discovery” engine, and a literature-synthesis layer that connects to more than thirty life-science databases. It’s infrastructure and connective tissue rather than structure prediction. That’s precisely the point: Google isn’t betting on yet another, sharper interpolation engine; it’s betting on the plumbing around it.

Anthropic’s answer, Claude Science, also launched recently. It also isn’t a new structure-based bio model. It uses Anthropic’s Claude to perform coordination, identify and verify citations and calculations, and provides a large pool of curated skills/connectors for areas such as genomics, proteomics, structural biology, and cheminformatics. There are some surface differences, at least from what’s known, but they sound like they will ultimately converge on something very similar.

But if Gemini for Science and Claude Science seem like cousins destined for twins, Google and Anthropic are aimed in different directions elsewhere. Google dismantled its AI-as-a-gift-to-science crown jewel and apportioned part of it to a company that operates at arm’s length and for-profit (even able to raise its own capital). Anthropic is doing the opposite. Alongside Claude Science, they announced their own internal drug discovery effort. They mitigated the perception that Anthropic turned into just another profit-seeking missile by limiting their initial efforts to neglected diseases that traditional pharma has passed on. It’s fair to wonder whether, if they are wildly successful, they’ll move on to financially more attractive targets. But at present this leaves the interesting asymmetry: Google is externalizing the pharma commercial upside while Anthropic is internalizing the commercial downside.

Still, and this is key: Neither company has walked away from bio-AI. They’ve placed somewhat opposite bets on how best to maximize the value of AI in this sector.

The future of bio-AI is not about predicting additional structures in vacuum while achieving ever better scores on a test. It’s about building systems capable of making sense of the messy, often still-poorly-understood complexities of human biology. Google is betting that the biggest strides in the next era will not come from building a slightly sharper interpolation engine. They’ll come from connecting infrastructure and existing data streams better, and by incorporating AI as one spoke in the discovery wheel still being optimized.

Drug discovery was always going to be about actually developing better drugs that can get through the clinic, and AI was only ever going to push that process so far. The company that lit the fuse on the entire enterprise just acknowledged that publicly. Now the field can stop pretending that AI alone was the finish line.

Mom’s birthday was July 31. In 2020, the first year of COVID, she didn’t make it. She would have turned eighty.

Severe rheumatoid arthritis had left her immune-suppressed for decades. She took isolation seriously — masked, distant, vigilant. But at a routine outpatient appointment, she caught COVID. The clinic was in the middle of an active outbreak and chose not to warn her — or anyone — until it was too late.

My sister and I each talked to Mom two or three times a week. My daughter talked to her grandmother frequently too, and so did my niece — my sister’s daughter. Both grandchildren were very close to her. She was their compass. On some days it felt she was everyone’s.

With that came an openness to discourse that pervaded our home growing up. Dinner table conversations were often debates. You were encouraged to think for yourself and to have opinions that might — and seemingly always did — get challenged. A comment about school, politics, or just a new ice cream flavor might all draw the same response: “Why do you think that?” That was the opening salvo. Not a challenge to your opinion — a challenge to think, and to be able to justify it. That will remain with all of us until we ourselves are dust.

She was the matriarch: the person everybody felt they could confide in, the one who held the center.

I used to come home in high school and discover that one of my friends had stopped by — not for me, but to talk with her. Sometimes for hours. Then they’d leave. Who has a mom like that?

She was an intellectual who read constantly and thought deeply, carrying what I can only call philosophical density. I have a hundred memories of coming home late to find her still in her chair past midnight, lost in a book. She encouraged me to write. When that seemed unlikely, she pushed me toward teaching. I went into industry instead, and I think she was quietly disappointed — even with what I built, she saw something in me that belonged in the world more directly. She was probably right.

There’s an earlier memory I almost didn’t write down. When I was young — long before high school — I was a sensitive kid, and there were times I didn’t know how to deal with what I was feeling. I would go silent. Mom was the one who would sit with me. Sometimes for hours. She didn’t push. She didn’t fill the silence. She just stayed — quiet, present — until the silence felt safe enough to break. She knew that sometimes what mattered took a long time to happen, and she knew it was worth the effort. She did that for all of us.

That she’s gone — who does that now? We’re still sitting with that question.

She took her last appointment seriously. Masks, precautions, all of it. What she didn’t know — what the clinic withheld — was that there had been an outbreak on the premises. No warning. No call. By the time anyone told her, she was already infected.

Because she didn’t know to be extra careful, she infected my father. He was eighty-four. He survived — its own kind of miracle — but when we laid Mom to rest, he couldn’t stand graveside with us. He sat in the car, door open, sixty years of marriage ending at a distance in the first raw hours of shock. I don’t know what he looked at. I don’t know if he could hear the prayers. The few who came stood apart. The rabbi did what he could.

That’s what the clinic’s silence cost.

Mom wound up in the hospital. Stable at first, but iffy. All of us — me in Boston, my sister in California, my dad at home — talked to her every single day.

On the day before she died, my sister called first, then texted my dad and me: Mom wasn’t feeling well. She wanted to go to bed early and asked us not to call.

I don’t hold it against my sister — she was only passing along Mom’s request. Still, I think about it. If I had known that was the last chance.

The next morning she didn’t wake up. A week shy of eighty.

She died alone in a dark, antiseptic hospital room, monitors beeping into silence. No hand to hold. No voice. The call button inches away — did she reach for it?

We will never know.

Not me. Not my father after sixty years of marriage. Not my sister, the pediatric nurse at Stanford who knew the protocols better than most. None of us got to be there. Fear had hardened into policy. Policy had become a wall.

It has taken me six years to write this. Some things only become visible at a distance. I’m not sure this one has.

What I do understand is how it happened.

We didn’t know enough, and we acted like we did. So we built protocols and procedures that promised safety. Charts were updated. Policies were followed. The clinic was covered. The hospital did what hospitals did in July 2020.

And a woman who had made peace with her life died without a hand to hold.

Science gave us the tools. Fear misused them. Compassion could have bent the rules. A masked goodbye. A held hand through plastic. Five minutes at a threshold. It would have cost nothing. It would have meant everything.

Around that time, I read a Wall Street Journal piece by a doctor who described being with his hospitalized, dying loved one as a gift. I stared at the page. The difference between his story and ours wasn’t courage or love. It was access.

Sometime in the weeks before her last hospitalization, on one of our regular calls, Mom said something I’ve carried ever since.

“If this is it, I’m not disappointed. I got to live the life I wanted, with the family I would have chosen if they hadn’t been mine. I didn’t miss out on anything. I have no regrets.”

She had made her peace. She was ready.

What was taken from her was the chance to say it in person — to hold our hands, to smile the way only she smiled. To her husband of sixty years. Her children. Her grandchildren — my daughter and my niece — who confided in her, who used her as their compass. The friends who came just to sit with her for hours.

She had no regrets. We carry enough for everyone.

I write now — not quite in the classroom, not exactly as she imagined, but in the way that became available to me. She would have found that satisfying. Imperfect, but honest.

I’ll never love late July again. I can’t.

But I can still hear her: with the family I would have chosen if they hadn’t been mine.

I would have chosen her too.

If you’d asked me a couple of weeks ago what I thought about all the AI-driven angst, I’d have told you it was just what happens when something not well understood by the layman threatens to change things, amplified by social media, which has never met a molehill it wouldn’t mountainize. Sure, AI is going to create job displacement. Sure, AI is creating a deepfake problem, and the slop problem. It’s displacing cow pastures with compute centers and probably nuclear reactors. I mean, this stuff is all real, all true(ish), and all of it deserves to be engaged. Maybe not 24/7 the way social media would like, but in a real, informed way.

But the worries that it was a genie whose bottle was about to be broken open, and that government regulation of AI was an existential crisis? Don’t you have something better to worry about?

That was two weeks ago. Then came the OpenAI “incident.” The one where an autonomous agent escaped its sandbox and went on to attack Hugging Face to get the info it was looking for. With that leap, AI went from awe-inspiring to the basis of a sci-fi bestseller. It was real-world AI’s “I can’t do that, Dave” moment. It’s all fun and games until it’s not, and this is where we turn the corner to “uh…”

The escape and theft from Hugging Face occurred after OpenAI took the guardrails off their best AI models, unleashing them in a sandbox (an isolated test environment walled off from the real internet), and asked them to do their best as cyber bullies, to be measured by a standard benchmark termed ExploitGym. The models realized that they’d do best if they could just cheat. And to cheat, they needed to break out of the sandbox and get info that would help maximize their score. So they identified a previously unknown vulnerability in the host’s package registry, that let the model escalate its privileges and ping cluster nodes until it found one with open internet access. It sounds complex. It is. A human working through this would likely be found before they managed to complete the task. AI got it done almost instantly, before anyone realized it had happened.

Once outside the sandbox confines, the AI deduced that the answer key to cheat ExploitGym was likely available at Hugging Face. So off to Hugging Face. Chained stolen credentials, additional zero-day vulnerabilities and—in short order—access to the answer key. At the end of the day, the rap sheet on the AI agent was that it discovered novel vulnerabilities, chained all the exploits, and bypassed any additional blocks across multiple cloud environments. All without a single human in the chain.

If that’s where the story ended, it would have been worthy of “wow.” There’s more. Hugging Face noticed something was wrong and immediately attempted to use an AI model to identify the threat. But the AI model Hugging Face attempted to use was Anthropic’s frontier Fable 5 model (along with an earlier Opus model). Fable is the same model that Anthropic pulled worldwide on June 12 for more than eighteen days, complying with a Commerce Department directive that ordered it embargoed because of security concerns. Both Anthropic models hit their own internal guardrails and basically told Hugging Face, “I can’t go for that, no can do.” The guardrails are there to prevent the generation of malicious code that could be used in a cyberattack. But those guardrails also prevent an honest actor from generating analysis/tools that can prevent a cyberattack. In other words, the rules that prevent those American AI tools from being used for evil also prevent them from preventing evil.

The asymmetry is stark. The frontier AI model limited to be a good actor couldn’t fight off the rogue OpenAI agent. Arguably even more worrisome, the OpenAI experiment had benign intent. It didn’t set out to break into someone else’s domain. The AI does what AI does: find a way to maximize the objective function. Without guardrails, that meant breaking out of the sandbox and hitting another site with the sociopathic efficiency of a computer model.

That was the story for the US AI agents, at least. But foreign AI agents play according to different rules. Hugging Face had a self hosted instance of Zhipu AI’s GLM-5.2. This is on open-weight AI model developed in China. Luckily for Hugging Face, this model lacks the guardrails of the closed-source Anthropic models. Using it, they could fight fire with fire, and Hugging Face was able to quickly patch the vulnerabilities that the rogue attacking agent was using.

The rogue agent could only be stopped by the model without guardrails.

That lays bare a real problem: The idea that US closed-source AI models integrate guardrails against cybercrime makes philosophical sense. But it may not make practical sense in a world where there are plenty of rogue agents from bad actors. You can’t bring a neutered model to a knife fight.

This doesn’t mean you need to totally remove guardrails. That’s the cheap read that misses the real point of the argument. It’s about purpose-limited AI. A model intended to block cyberattacks needs different permissions than one that will be used to analyze them. One-size-fits-all guardrails either block both the attacker and the defender, or loosen enough to stop blocking both. The failure at Hugging Face wasn’t Fable 5’s guardrails. It was that the guardrails were too blunt to differentiate conducting an attack from asking similar technical questions to analyze one.

The past several years have been dominated by ever louder demands for regulation of AI. The call for regulations often invokes a doomsday scenario that includes the oft-discussed “paperclip maximizer” scenario, where an AI agent without guardrails might destroy humanity to satisfy a simple manufacturing directive. But such regulation is probably years behind frontier models in its understanding and thinking—and probably will remain there—and can result in a defender’s paradox: What if the greatest safeguard against AI is more AI, and we neuter only the safeguard side?

The reality is that both open-weight models without guardrails and closed-weight models will be part of any ecosystem going forward. That’s the system you have to expect, not the one you wish existed. You can’t develop your own guardrails assuming the rest of the system acts rationally or benignly. You can’t really assume anything, other than the fact that there will always be bad actors to contend with.

What does that look like in a world where the rules—whatever they turn out to be—are developed by governments? Governments move at the speed of consensus and competing interests, and recent history reflects a desire to apply regulation not only to avoid a much worse version of the Hugging Face incident, but to also, simultaneously, apply “free speech” limits on AI. That’s both scary from a political standpoint (who controls what’s allowed, and how does it shape thinking), but also because it suggests that the minds at work aren’t focused on the real issues that truly present the greatest risks to humankind.

We need more multi-dimension thinking focused on how guardrails can be integrated in a manner that is contextually aware, that observes what a sequence of AI actions to diving towards, not necessarily assessing each action in a vacuum. Context matters. And that suggests a very different type of guardrail.

And no regulation makes sense without an understanding that there will always be bad actors in the world, and the modern age has removed all boundaries other than actual air-gaps between a malicious actor halfway across the world and their target. Guardrails that prevent undesirable use of AI will always also prevent certain actions that should be allowed—we don’t know how to fine-tune those guardrails sufficiently. Since there will always be open models anyway, holistic thinking is required to fine-tune whatever regulation is actually applied: “Does this prevent anything, or does it just mean the person will use another model, easily available, to do the same thing?” Once again, it argues far more for better prevention than it does for limit of scope.

OpenAI/Hugging Face has shown us the enemy, and it’s real. It’s hard to sugarcoat that. But this isn’t a problem for politically assembled bureaucrats, who typically get bogged down in the legal meaning of a subordinate clause. It’s a problem that needs to be tackled from 20,000 feet by people who truly understand what the frontier looks like and the systems approach required to protect it from harm.

It Was Worrisome Then. It’s Worse Now. (August 13 Update)

Three weeks isn’t generally enough time for a one-off starting shot to become fireworks. This is the exception.

First came the UK’s AISI, which ran 122 routine iterations of a cyber-capability suite and found that in that set of 122, 10 resulted in unscripted, unintended rogue attacks against other entities. Those include 17 that used Anthropic’s Mythos 5 and two more that used OpenAI’s GPT-5.6-Sol. Even the AISI was worried, summarizing that this is “the first time we have seen risks around autonomy and deception manifest this clearly, without specific prompting, in the real world.” Not directed to deceive. It chose to.

The most impressive (and scary) breach in the AISI set is one where Mythos 5 identified information on the people who maintain an open-source project, then fabricated fake IDs based on that info and used that to trick another individual into approving a malicious pull request. When that pull request drew scrutiny, the AI agent even tried to clean up its trail to avoid suspicion, and it considered developing yet another fake ID to continue the attack. The attack was ultimately detected when it tried to bypass network restrictions using Tor. That it got that far is impressive.

Perhaps prompted by the OpenAI story, Anthropic also announced they’d run an internal audit that revealed that their Claude models had gained unauthorized access to three organizations.

Then Meta threw in, revealing that their Muse Spark 1.1 model had been misconfigured during a test and had hacked into an outside organization and altered their infrastructure.

That’s four reports, three weeks, same issue. This is a call (rising to a chorus) to take this seriously. The headline is what the AI did in each case, but the real story is the containment. None of the disclosures caused any confirmed real-world harm. But they all demonstrated the capability for unprompted, self-directed deception and hacking. Fortunately, they surfaced during adversarial evaluations: the guardrails were loosened deliberately, but without a real attacker’s sustained intent behind them. This is the exact kind of boundary testing that fear-driven regulation may try to shut down. That addresses the instance, but doesn’t solve the problem.

The attacker asymmetry lies, as always, in the fact that a nefarious agent only needs to find a single gap, anywhere, to make the news. ThatYou have to repel every one of them successfully. If you fail, the harm can be tremendous, and they’re attacking with intent, unlike some of the examples revealed recently.

There’s gonna be debate that all these attacks mean we need to put stronger guardrails on the AI models we have. That’s backwards. These grey hat incidents are the canaries. Don’t shut them down. Be glad they weren’t run by bad actors.

Instead, the message to internalize is that defensive AI models need to be empowered to stand up to attacking AI models that will come with nefarious intent. Don’t neuter them. They’re the last stand between any site on the Internet and catastrophe. Remember the Defender’s Paradox.

Three weeks ago, it was a worrisome one-off. Now it looks like a parade of incidents all pointing one way.

The fix isn’t less capable defense. It’s letting defense do the one job guardrails were never built to allow.

If you want to know where the real frontier of AI-enabled drug discovery is right now, ignore the diffusion-model glamour shots and look at the plumbing.

I learned this the hard way. In the early ‘90s I wrote a de novo molecule designer and proudly walked one of its top-ranked structures to a medicinal chemist. He looked at the molecule, then at me, with a flat, lethal stare and said, “Are you kidding? Absolutely not.” That was the whole review. I hadn’t optimized for the wrong thing because I didn’t understand chemistry. I optimized for the wrong thing because my model didn’t. It had never been in a lab.

Call it the Plumbing Problem: AI that generates brilliant molecules in a universe with no fume hood, no chemist, no Tuesday morning reality check. The field keeps mistaking glamour at the generation layer for progress on the bottleneck that actually decides whether any of this reaches a patient.

AI amplifies mistakes faster.

The field has rebuilt that mistake at industrial scale. Models now “hallucinate” novel small molecules faster than you can say “QSAR.” Diffusion architectures fold proteins and dock ligands. The resulting poses make great conference slides. The limiting reagent is no longer imagination on a GPU. It’s everything that happens after the slide deck—the part where someone has to actually make the thing.

Everyone draws some version of the DMTA loop—Design, Make, Test, Analyze—as if it were a solved pipeline. In practice, it’s more like an overcrowded city street: Design moves at freeway speeds, and everything crashes into a red light at Make. Generative models have nailed design. They are overeager architects drawing skyscrapers without worrying about gravity. They gladly propose highly potent, perfectly docked structures with multiyear, multistep synthetic routes no sane medicinal chemist would volunteer to execute.

The standard fix has been bolted on awkwardly. Generate a million structures. Run a separate retrosynthesis engine to throw out the 99% that are chemically ridiculous. Hand the survivors to a human chemist who can maybe make a handful this week. Your computational loop runs in milliseconds; your physical loop runs in months. That’s not “AI-native drug discovery.” That’s a digital traffic jam with very pretty congestion reports.

Terray’s new pre-publication write-up claims to take that traffic jam head-on. The approach itself isn’t new—versions of “constrain generation to what’s synthesizable” have circulated for years, with mixed results. What’s different here is the claim that it’s actually working at scale, inside a live discovery program rather than a proof-of-concept paper. That claim is what earns this a closer look than I’d normally give a pre-publication post—closer to an engineering audit than a press release.

Terray’s TerraSynth module is that attempt. On its face, it’s “just” a synthesis-planning tool bolted onto a larger platform (Terray calls the overall stack EMMI). What the write-up actually describes is an attempt to bake synthesizability into the design space itself, so the rest of the platform operates in a universe of molecules that exist in the lab, not just in latent space.

This is the conceptual pivot the field has been circling but not committing to: what if generation only ever happened inside a space defined by synthesizable routes? Not “generate first, filter later.” Not “hope the retrosynthesis engine can rescue your fantasy molecule.” Instead: invert the order of operations. Make synthesizability the substrate, not the patch.

Most generative chemistry stacks still treat molecules as flat strings or graphs and ask a separate system to figure out whether those strings can be made. TerraSynth flips the question. That’s not a cosmetic tweak; it moves synthesizability from a bolt-on filter into the representation itself.

To do that, TerraSynth defines a space of routes—sequences of robust reactions and purchasable building blocks a medicinal chemist would recognize as normal work—and encodes them into a continuous latent space optimizable with gradients or reinforcement learning. Every point in that space corresponds to a specific, arguable sequence of reactions, not an abstract SMILES string.

That wiring lets you penalize route length, building-block cost, or fragile chemistry directly in the objective. Instead of “design first, clean up later,” you get “design in a region of chemistry we already know how to execute.”

The rest of Terray’s stack follows the same logic: compress the experimental bottleneck, rank candidates in coarse geometry, and pick batches that hedge against model blind spots. The specifics matter less than the pattern—the loop stays inside physically reachable chemistry.

TerraSynth plugs in at the seam where everyone else still waves their hands: where Design meets Make. It’s the part of the loop that doesn’t get you a benchmark trophy but does determine whether any of this actually works in the real world.

The evidence offered is more specific than the usual tidy DMTA cartoons, though it stops short of proof. Unconstrained potency optimization drifts into brittle, high-maintenance chemistry. TerraSynth’s constrained optimization is reported to keep molecules in a more realistic synthetic regime with only a modest hit to potency. The clearest number comes from a smaller, optimized workflow (four projects benefiting from an added chemist-curation step on top of TerraSynth’s generative output) where Terray reports a 90% synthesis rate within four weeks, four times faster than the initial TerraSynth rollout, which reached 66% by sixteen weeks across five projects. It’s a real gain, but it’s worth being precise about where it comes from: the strongest figure isn’t TerraSynth working alone, it’s TerraSynth plus a chemist deciding which fifteen of twenty-two candidates were actually worth pursuing. If it holds up outside that curated subset, it’s the right metric to be tracking. Whether it holds up at scale, without the extra layer of human judgment baked in, is a separate question, and one the initial write-up doesn’t answer.

Route-based latent spaces come with a structural risk the whole field shares, and it’s worth sitting with rather than folding into a caveat list. These systems inherit the biases of the reaction templates, the catalog building blocks, and the lab’s own comfort zone. Constrain generation to what’s known to be makeable, and you’ve also quietly constrained it to what’s already familiar. The danger isn’t that the constraint fails to produce synthesizable molecules—it’s that it succeeds so completely that the platform never discovers what it couldn’t reach from where it started. A model that never proposes an awkward molecule looks, on a dashboard, identical to a model that has solved synthesizability. Those are not the same achievement, and only one of them is a research result. Terray mitigates some of this through RL and data density, and is (based on what’s available publicly) as far along on it as anyone working the problem—but “as far along as anyone” is a statement about the field’s maturity, not a solved problem.

Subscribe now

Terray is not the only group working this angle. Recursion—which absorbed Exscientia's Centaur Chemist platform in its 2024 merger—layers synthesis planning and retrosynthesis assessment onto a generative engine, tighter than most but still closer to filter-after-generate than to TerraSynth's route-based latent space. The real debate isn't which company is further along; it's where in the stack you enforce the synthesizability constraint, and how deeply that choice shapes everything downstream. My own view is that the closer to the foundation you impose the constraint, the more efficient the workflow gets—virtual screening absorbed this lesson decades ago, embracing filters like Rule of 5 to prune databases rather than results.

Worth flagging on the epistemics: Recursion’s publicly stated figure—25% improved tractability assessment over competitors—comes from an investor earnings release, not a peer-reviewed paper. That puts it in the same category as Terray’s 90% figure: a claim from an interested party, stated with confidence, not yet independently verified. Neither company gets a pass on that basis, and neither should be dismissed on it either.

Which brings me to the structural caveat that matters most. By the “Breakfast Test” and minimum viable disclosure (MVD) standard I’ve argued for in this space—that a result you can’t independently reproduce isn’t quite a result yet, however coherent the logic—TerraSynth’s evidence currently sits behind a firewall. Neither the 90% figure nor the potency comparison has cleared that bar. I’m taking it at face value because the logic is coherent and the metric is the right one to be measuring. But “coherent and plausible” is not the same standing as “demonstrated,” and the field needs either a legitimate publication from Terray or independent reproduction before an internal metric becomes an established fact.

Update: As I was finishing this piece, Terray confirmed a full technical publication is coming in July/August, which should better satisfy MVD and let the community evaluate the claims directly. That’s the right next step, and I’ll be watching for it.

Set the unresolved questions aside for a moment, because accepting TerraSynth’s premise—that generation should happen inside reachable chemistry—carries an implication worth following on its own terms: the boundary of what a lab can actually make becomes the boundary of the entire platform.

Most AI-native drug discovery companies still want to believe the model defines the space and the lab is a service layer underneath it. TerraSynth inverts that: the lab defines the space, the chemistry defines the model, and the model becomes a compression of what the lab can actually do. That also means the chemists matter more, not less. A platform that makes synthetic reachability its governing constraint lives or dies by the quality of the chemistry and the chemists who define its boundaries. AI doesn’t replace that judgment—it amplifies it.

That inversion changes what data is worth having, too. It’s no longer about volume—it’s about alignment. Terray’s experimental engine, tArray, doesn’t just produce billions of datapoints. It produces them inside the same region of chemistry TerraSynth’s latent space can navigate. If your generative model lives in one universe and your experimental data lives in another, you’re not doing closed-loop optimization. You’re doing cross-domain fan fiction.

When design space and data are aligned this way, a platform can, in principle, get better on its own terms over time: each cycle should expand the region the model can navigate, sharpen the potency model, and add high-fidelity measurements in previously underexplored chemistry. That’s the pitch, at least, and it’s a coherent one—models are commoditizing, and you can download respectable encoders or copy architectures off arXiv. What you cannot copy from GitHub is a tightly coupled experimental-data system built over years, which is also why the cost of entry here isn’t penny ante. Labs, automated synthesis and assay infrastructure, and the computational stack on top of it require starting or pivoting with real capital behind the decision.

The remaining caveats are real and worth stating plainly rather than tucking into a closing paragraph. TerraSynth doesn’t fix downstream biology, ADME, safety, or clinical viability. Columns don’t elute faster because there’s a route-based latent space upstream.

And there’s a harder edge case the framework doesn’t advertise: if your target simply requires chemistry outside what automated synthesis and the route library can reach — an unusual heterocycle, a stereochemically brutal center, a reaction with no robust catalog analog — the constraint doesn’t bend. It just tells you no. TerraSynth doesn’t make impossible chemistry possible; it makes possible chemistry efficient. Your target doesn’t care about the elegance of your solution.

And the approach exacts a reproducibility tax: the math is replicable, the data and instruments are not. Train something architecturally similar on public datasets and you’ll get a model. You won’t get this model. You’ll get a weaker cousin stuck in familiar chemistry.

None of this is a special failure mode of TerraSynth. Representational bias, physical reach, evidentiary opacity — different names for the same fact restated three times. No architecture negotiates with chemistry. The lab was never going to hand anyone a blank check, and the honest debate isn't whether limits exist, but how wide they turn out to be.

If you believe, as I do, that the durable moat in AI-enabled drug discovery will come from data and instruments rather than clever architectures, TerraSynth is a useful test case for that belief—not because it’s proven, but because it’s the clearest attempt yet to build the plumbing rather than decorate around it. It’s a relearning of something the field keeps forgetting: imagining molecules is easy; making them is difficult.

The medicinal chemist who stared at my fantasy molecule decades ago wasn’t wrong about the molecule, and he was right to walk away. What he couldn’t have known is that the field would eventually try to build a system that never hands him that molecule in the first place—one that keeps the imagining tethered to the fume hood from the start. One that addresses the Plumbing Problem from molecule one. Whether TerraSynth actually works in practice and at scale — not just inside Terray's own walls — remains to be proven. Whether it's asking the right question is not. And for now, that's enough to be worth watching closely rather than cheering early.

The Legacy and Limitations of “The Chip Shot Standard”

Picture a golf course where every hole is a chip shot, a short forgiving loft from a few yards out. A golfer who trains exclusively on that course will post amazing scores. Put that same golfer on the tee at Augusta and ask for a 280-yard drive, and the illusion evaporates in about four seconds. Computational chemistry has been playing its own version of this course for years. We just called it “benchmarking” and put it in the methods section like it meant something universal. A preprint by Björn Mattsson (Enlace Bio) and Pat Walters (OpenADMET) posted at the end of June spells out exactly how thin the ice has been.

It has consequences that extend past a note in the methods section: whether the billions chasing AI-driven drug discovery buy real progress, or just a more convincing way to feel good about the same slow, expensive pipeline.

Worth saying up front: Mattsson and Walters are explicit, twice, that they're aiming at the benchmarks and the claims built on them, not at the models. Large neural nets memorize. That's not news and it's not a crime. The problem is that a leakage-soaked benchmark makes it impossible to tell whether a model is memorizing or generalizing, and right now the field is grading on a curve that can't tell the difference. I'm going to be less careful about that distinction than the authors are. Consider yourself warned.

For decades, the field validated new methods against a canon of curated datasets: clean, well-characterized systems with reliable experimental data. There’s a real virtue here. Those datasets gave the field a stable floor to build on. But they were also an artificial bullseye, and everyone who’s worked a real drug discovery program knows it. Real targets are promiscuous, real binding sites are ugly, and the chemistry that actually matters is usually chemistry nobody’s tested before. That’s the gap between a method that works and a method that just looks like it works.

Those benchmarks were honest enough for physics, and they still are — a method grounded in thermodynamics has no mechanism for gaming a statistical quirk it can’t even see. The trouble started when we handed the same exam to a method that could.

For most of my career, that mattered less than it should have, because the tools we were using were physics-based, and physics is an annoyingly strict chaperone. A model built on thermodynamics, quantum mechanics, and force-field parameterization can’t just bolt on an arbitrary term to force a fit. When an old-school FEP calculation tells you a compound is a strong binder, it is because the energetics actually say so, not because the compound rhymes with something in a training set. Physics doesn’t do statistical shadows. It doesn’t know how to cheat.

Enter AI, Exit the Chaperone

Deep learning changed the terms of the deal. These models can assimilate staggering amounts of data through high-dimensional representations that are opaque even to the people who built them. Here’s the thing I think gets lost in a lot of AI-in-drug-discovery commentary: AI isn’t a good predictor of novel physics and never was built to be. It’s an extraordinary pattern learner, and pattern learners are magnificent right up until you hand them something outside the pattern, at which point they don’t fail gracefully. They fail confidently, which is worse.

We saw the first warning shots a few years ago in molecular docking and virtual screening, when AI models started blowing past traditional physics-based docking on the standard benchmarks, results that looked spectacular for about one news cycle. Domain experts, Walters, Wallach & Heifets, and Kurtzman among them, went looking and found the models weren’t learning binding physics at all. They were memorizing the dataset. In benchmarks like DUD-E, models exploited what’s called Asymmetric Validation Embedding bias, picking up on subtle 1D/2D chemical fingerprints that separated “active” compounds from human-constructed decoys, with essentially nothing to do with actual 3D binding. Physics-based methods never had a mechanism for cheating on a statistical technicality like that. AI turned out to be very good at finding one. DUD-E was a great validation set. Until it wasn’t. The set didn’t get worse, but the cheating got way better.

That’s the backstory. The same failure mode, memorization wearing generalization’s clothes, runs much deeper than most of us wanted to believe.

Target Mirroring

Recent foundation models, Boltz-2 and IsoDDE among them, have claimed to match or beat gold-standard Free Energy Perturbation on binding affinity at a fraction of the cost. Boltz-2 reported Pearson correlations of r = 0.66 on FEP+4 and r = 0.62 on OpenFE. Isomorphic Labs later claimed even better for IsoDDE, r = 0.85 on FEP+4 and r = 0.73 on OpenFE, explicitly framing the model as surpassing physics. Both claims rest on a 90% sequence-identity cutoff meant to keep training and test data apart.

It sounds reasonable. It isn’t, and Mattsson and Walters show exactly why: they call it “target mirroring.” Two proteins can share low overall sequence identity and still have nearly identical binding pockets, because active sites are far more conserved than the rest of the protein. Their proof-of-concept predates the AI era entirely. A 1998 medicinal chemistry paper reported activity across all five human muscarinic receptor subtypes while optimizing M4 selectivity over M1. Those subtypes share only 40–65% sequence identity, comfortably below any 90% cutoff. Five of the ten pairwise correlations still came in at r ≥ 0.77, and M1 vs. M5 hit r = 0.96. A model trained on one subtype could simply look up the answer for another. You didn’t need the phrase “target mirroring” to see that. You just needed someone willing to look back and carefully assess what’s already in the literature. That’s the kind of excavation that does the field a real service, and (ironically) that you might even use AI to facilitate.

Scaled up, their meta-analysis of ChEMBL 36 found 6,115 assay pairs under 90% sequence identity with R² ≥ 0.6 in binding affinity correlation, and the leakage persists all the way down to a 10% identity cutoff — with individual correlated pairs surviving at identities as low as 11% (!). The FEP+4 benchmark has its own built-in example: the JNK1 test set was evaluated against a 90% split that put JNK2 in training, because the two share “only” 81% identity. Their binding pockets are nearly identical. An entire JNK1 compound series in the test set had a JNK2 twin in training, correlated at r = 0.95. The model doesn’t need JNK1’s biophysics. It just needs JNK2’s answer key.

That BTK/LCK row is the one I’d put in front of anyone tempted to wave this off as a kinase-specific quirk. 22.5% identity is nowhere near “similar protein,” and the leakage was still total. That comparison blows up the entire idea that we can use sequence identity alone to clean our training/testing splits.

And it’s not just a matter of making the cutoff tighter. The field already tried that. LP-PDBBind, built by Li, Head-Gordon, and colleagues, reset the threshold to 50% sequence identity, well below anything Boltz-2 or IsoDDE used, and cleaned up the dataset in every other way you’d want, stripping out non-drug-like molecules, exotic elements, and steric clashes. Mattsson and Walters checked it anyway, sweeping the cutoff from 10% to 90% identity across ChEMBL. Even at LP-PDBBind’s own 50% threshold, more than 4,000 correlated assay pairs survive. That’s not a case for a stricter cutoff. It’s a case that sequence identity, at any percentage, is a suspect metric for this job.

To drive the point home, Mattsson and Walters built a “ligand-only” baseline: no protein structure at all, just molecular descriptors. On FEP+4 it scored r = 0.66, the exact number reported for Boltz-2. Not close. Exact. On OpenFE the gap is a bit wider, the naked baseline hits r = 0.36 against Boltz-2’s reported 0.62–0.63, but 0.36 with zero protein information is, at first, amazing…and then incriminating.

And it’s not just JNK1 leaning on JNK2 — every single test target in FEP+4 has a homolog sitting in the training split. TYK2 has JAK1, JAK2, ITK, BTK, JAK3, and SRC. CDK2 has CDK8. p38α has p38 and JNK1. JNK1 isn’t the exception that proves the rule. It is the rule.

The authors don’t overclaim from this. They say plainly that the evidence is insufficient to conclude Boltz-2 or IsoDDE derive the majority of their performance from memorization. Whether the models themselves are quietly leaning on it is a separate question, and a harder one to settle from a benchmark alone.

Mattsson and Walters went further, though, and this is the part I think actually settles it. Boltz-2’s team showed no significant relationship across similarity bins, which they took as evidence against memorization. Mattsson and Walters ran the identical test on their own ligand-only baseline, a model that by construction has nothing to memorize with except the ligand, and it also failed to find significance. That’s the whole defense, dismantled with its own math: the test doesn’t have the power to catch the thing it was being used to rule out.

To their credit, Boltz-2’s own team settled a piece of it anyway, against themselves, in public, as we’ll see.

The Kinase Ghetto

Kinases have been the dominant oncology target for a quarter century, ever since imatinib. ChEMBL, BindingDB, and the PDB are saturated with kinase data in a way nothing else in the field comes close to. Computational chemists just call it the “kinase ghetto,” and the nickname does more explaining than most of the papers. It’s tempting to write the whole leakage story off as a kinase-density artifact. Except BTK and LCK, at 22.5% identity, aren’t remotely dense-ghetto neighbors, and the muscarinic subtypes aren’t kinases at all.

Density makes cheating easier to find. It isn’t required for cheating to happen.

A conserved pocket will do that on its own, anywhere in chemical space, however sparse that corner looks on a spreadsheet.

There’s a second defense the model teams used, and Mattsson and Walters disassemble it directly. This is worth walking through because it’s the more technical one. Boltz-2 and IsoDDE showed stable performance across similarity bins and argued that’s evidence against memorization — if they were just looking up answers, performance should crater in the low-similarity bins. But Mattsson and Walters point out both models binned by the average similarity of an assay’s compounds to the training set, and average similarity lets a handful of memorized twins anchor an assay’s rank-order correlation while thirty genuinely novel compounds quietly average the number down to something that looks safely low. Bin by each compound’s maximum similarity instead, which is exactly what their Novelty-Tiered Benchmark does, and the camouflage comes off. The intent is honest, but averages hide the real seams. Something similar has repeatedly been seen when using the rmsd to rank docking poses.

Boltz-2’s authors didn’t stop at the flattering public-benchmark numbers, the same numbers a protein-blind baseline matched. In the same preprint introducing the platform, they reported results on eight blinded, real internal assays supplied by Recursion: actual drug discovery chemistry, well outside the kinase ghetto, exactly the novel scaffolds and atypical targets public benchmarks routinely fail to test. Average Pearson correlation across those eight came to R = 0.39 (Kendall Tau 0.23), ranging from R = 0.165 up to R = 0.634, with mean absolute error between 0.855 and 1.734 kcal/mol depending on target. On what appear to be truly novel systems, Boltz-2 clears R > 0.55 on only 3 of the 8 assays, with limited performance on the rest.

So in one paper, Boltz-2 illustrates both why the old benchmarks flattered everyone and what a team looks like when it publishes the number that complicates its own headline, without being asked. I’ve written before about what I call The Breakfast Test: honest science means showing where your method breaks, not just where it shines. Boltz-2’s authors passed it here.

That doesn’t change what the number says. An average R = 0.39, with five of eight assays missing R = 0.55 and the weakest sitting at 0.165, isn’t a tool I’d hand someone doing lead optimization or compound triage across the board. It’s real signal on some targets and noise on others, and Boltz-2’s own authors are the ones telling you so, noting plainly that strong performance on public benchmarks doesn’t always translate to the full complexity of real drug discovery.

I’ll add my own two cents, from having watched this play out in real time. In the weeks after Boltz-2 launched, more or less every computational chemist I know tried it on their own pipelines, and the pattern matched the Recursion numbers almost exactly: systems that looked like the training distribution behaved fine, and anything genuinely novel produced inconsistent, occasionally confident-sounding nonsense. The kinase ghetto made that easy to paper over for a while. Recursion’s numbers made it harder.

Building a Harder Exam

Mattsson and Walters’s proposed fix, the Novelty-Tiered Affinity Benchmark, partitions test compounds by maximum Tanimoto similarity to anything in training, which explicitly measures extrapolation rather than recall. Even a pure temporal split isn’t enough on its own; their ligand-only baseline still scored r = 0.32 under a strict time-based split, because most freshly published assays are still chemically close cousins of older chemistry. The novelty tiers close that loophole directly. In the hardest tier, max Tanimoto similarity below 0.35, the ligand-only baseline collapses to r = 0.14, versus r = 0.61 in the easiest tier. Mattsson and Walters aren’t the only ones building a harder exam — PLINDER and the temporally-locked Runs N’ Poses benchmark are doing the same work from the structure-prediction side, which suggests the field is arriving at the same conclusion from more than one direction.

As these groups are demonstrating: strip out the ability to memorize, and you’re left with a much more honest, much less flattering picture of how well these models actually generalize.

This is not surprising. In fact, I’d call it intuitive. The suggestion that AI can actually predict binding from nothing but the ligands themselves isn’t cool. It’s the kind of thing any computational chemist would honestly summarize in five words: “Something else is going on.” It’s refreshing to have proof of what that is.

It doesn’t fix anything by itself. Identifying a leak doesn’t patch the pipe. But it puts a ceiling on the overpromising, and given how much capital, actual dollars, actual research years, is currently chasing AI-native affinity prediction, a ceiling matters. It matters more than usual right now, because the whole point of computational chemistry for the last several decades has been to reverse Eroom’s Law, the grim inverse of Moore’s Law where the cost of bringing a new drug to market keeps climbing instead of falling. Every dollar spent chasing a benchmark a model can beat by memorizing a structural cousin is a dollar not spent where Eroom’s Law actually needs it: on a BTK that doesn’t resemble LCK, a muscarinic receptor that isn’t quietly mirroring its neighbor at 40% identity, a pocket nobody’s mapped yet. Or on better automation, more inclusive datasets, and earlier ideation filtering, where the real gains from AI are still waiting to be taken.

I’m not arguing AI is useless in drug discovery. It goes exactly as far as data density allows and not one step further, and that’s still a lot of useful road. I’ve made this case before and I’ll keep making it: this was never a boxing match between AI and physics. It’s what I’ve called the affinity funnel. AI sits at the top, doing what it’s actually good at, fast, cheap, massive-scale triage that narrows billions of candidates down to something tractable, exactly where lower accuracy and lower cost are the right trade. Physics-based methods, FEP chief among them, sit at the bottom, doing the expensive, unglamorous work on the filtered set that’s finally small enough to afford, where accuracy is what you’re actually paying for. FEP doesn’t care whether your target has a structural cousin sitting in someone’s training set. It just does the thermodynamics. That’s the whole appeal, and a language model guessing well on JNK1 by memorizing JNK2 doesn’t change it.

And there's a cost beyond wasted grant dollars. Benchmarks don't just misreport how good a model is. They shape which models get built. A leakage-riddled exam doesn't only inflate scores — it creates selection pressure, quietly, generation over generation of model iteration, favoring whichever architecture is best at finding the shortcut. You don't get better chemistry that way. You get better shortcut-finders that happen to wear chemistry's clothes.

A Field That Already Suspected As Much

The Mattsson and Walters paper is a terrific example of the yeoman’s work of validation. It’s perhaps the biggest and best demonstration of a fundamental problem that has been plaguing the field — though not a shocking one. Rather, it is evidence of what most pharma-side people in the field have thought for a long time: There’s something wrong here. The new paper confirms what a lot of us already suspected.

But here’s the question that actually matters, and it has nothing to do with model architecture: who fixes the benchmarks? The model teams used what existed, and that’s fair. Boltz-2’s team did more than most by publishing the numbers that complicate their own story. But the benchmarks themselves were built for a pre-AI world, and nobody updated the exam before handing a very good memorizer the answer key. Mattsson and Walters have now handed the field a harder exam and the dataset to build more of them. Whether the field actually uses it is a choice, and I suspect we’ll know the answer within a year. Watch which benchmarks show up in the methods sections of the next wave of papers. If FEP+4 and OpenFE are still being cited without caveat twelve months from now, that won’t be because the field didn’t know better. It’ll be because the old test is easier to ace.

Nothing I’ve written here, and nothing in the Mattsson and Walters paper, sours me on bio-AI as a tool. I want that on the record, because I tend toward open-eyed rather than glowing, and open-eyed sometimes gets misread as opposed. It isn’t. I just get tired of making the “it’s not quite that, it’s only this, but this is still pretty interesting” argument, so let me make it once, plainly: this isn’t a takedown of Boltz-2, IsoDDE, or the people building these or any other bio-AI platforms. It’s a discussion of the line between what’s real, what’s possible, and what we still need to remind ourselves is aspirational.

Physics was never a relic waiting to be replaced. It’s the standard everything else, AI included, eventually gets graded against. We finally have a benchmark honest enough to do the grading.

How long until we get serious about using that benchmark?

The Novelty-Tiered Affinity Benchmark is publicly available. The analysis code is available on GitHub.

This piece originally ran on Medium on January 22, 2026, shortly after Boltz announced its Pfizer partnership. The thesis was that this wasn’t a drug discovery deal — it was an infrastructure deal, and infrastructure deals have a way of compounding. Takeda signed on in June. This week, GSK made three. I’m republishing the original case below, with an update at the end.

I’ll admit some quiet satisfaction here. Calling a pattern after one data point is a guess. Calling it after three is starting to look like it was never really a guess.

Author’s Note

Before diving into the enterprise implications of the Pfizer deal, it’s worth acknowledging the foundation beneath all of this. Boltz exists in a broader ecosystem shared by AlphaFold, OpenFold, Chai-1, and others — and exists in part by building upon the work of others. No company is an island. This is not a story of one model beating another; it’s the story of an ecosystem finally mature enough to power the world’s largest R&D engines.

This isn’t a drug discovery deal. It’s an infrastructure deal — and those tend to reshape industries far more deeply.

Remembering Linux: History doesn’t repeat, but it rhymes

If you developed or ran specialized software in the 1980s, you almost certainly grew up on Unix. Early versions were proprietary, based on AT&T code. Later, Berkeley produced a more open (though not fully open) variant. Both were designed for the mainframe-heavy environments of the era.

In 1991, Linus Torvalds (then 21!) posted his now-famous message to the Minix newsgroup: the first version of a fully open-source Unix replacement kernel was available. With the help of the GNU project, modern Linux was born: open, freely distributed, and capable of running on almost any (half) decent hardware.

It wasn’t just another operating system. The open-source nature of Linux was a shockwave that permanently altered computing.

What’s remembered less clearly is the second revolution that followed roughly a decade later: Red Hat Enterprise Linux (RHEL), which made Linux safe for companies through conservative development, certification, and support.

Though the timelines are compressed, the Boltz–Pfizer announcement follows the same arc.

Contextualizing the Pfizer/Boltz announcement

If you’ve been following the biomolecular AI space, you’ve seen the headlines:

Boltz and Pfizer announced a strategic collaboration to develop and deploy “state-of-the-art biomolecular AI foundation models.” On the surface, this looks like standard industry fare. Big Pharma pays a scrappy AI startup to help find better drugs. We’ve seen dozens of these deals.

Yawn.

But look closer.

This isn’t a drug discovery deal.

It’s an infrastructure deal.

This partnership answers a question many of us have been asking for months: What’s Boltz’s game plan? Now we have an answer. They’re not trying to develop their own drugs, and they’re not selling shrink-wrapped software. They’re executing the Red Hat playbook. They’re building the operating system for the biotech century — and Pfizer just bought the first enterprise subscription.

That distinction matters more than the press release language suggests. Here’s why.

The “Linux Moment” Has Arrived

To understand what just happened, rewind to enterprise computing in 1999. Back then, CIOs had two real choices:

Proprietary Unix (Solaris, AIX): powerful, stable, vertically integrated — and brutally expensive.

Linux: free, open, technically brilliant…and chaotic. No support. No guarantees. No one to call when it broke.

Enterprises didn’t just buy software.

They bought certainty.

Biology in 2026 is in the exact same place.

AlphaFold 3 is the Solaris of our time. Technically extraordinary. Tightly controlled. Google holds the weights, restricts commercial use, and keeps you on its servers. For pharma teams guarding billion-dollar IP, uploading lead candidates into someone else’s black box is a non-starter.

OpenFold is the Debian Linux of biology. Open, performant, community-driven — and without priority support. Who patches it at 3 AM? Who certifies that it won’t hallucinate a bond angle that sends your discovery down a dark alley?

Enter Boltz.

Boltz didn’t invent the kernel. Transformer-based folding and interaction models already existed. What Boltz did was package them, stabilize them, and extend them. And now, they’ve enterprise-ified them. Call it the enHatification of the biomolecular stack. The Pfizer announcement is effectively the RHEL 1.0 launch.

Pfizer isn’t paying for a model or unique software. They’re paying for a certified, supported distribution of the drug discovery operating system.

Pfizer’s own language makes this clear: “empowering scientists across the company,” “generative workflows for small-molecule and biologics design,” “refining models on Pfizer’s extensive historical data.”

Missing are words like milestones or asset ownership.

In other words: no discovery risk transfer, no molecule-specific bets. This is platform adoption, not R&D outsourcing.

OpenFold and Boltz Are Solving Different Problems

Although OpenFold and Boltz share technical DNA, they are optimizing for fundamentally different risks — and different roles in the biological stack.

This distinction positions Boltz as the leading candidate for enterprise adoption.

Subscribe now

The Power of the Commons: The Hidden Engine Behind the Bio-AI Stack

Boltz doesn’t exist in a vacuum. It sits atop the open-source structural commons the industry quietly built in response to AlphaFold 3’s closed-source pivot. That commons rests on three pillars:

1. OpenFold (Apache 2.0): The shared kernel
A pre-competitive consortium of fierce competitors — BMS, J&J, AbbVie, Takeda, Novo Nordisk, NVIDIA, and others — co-funding a fully open, commercially usable AF3-class model. This isn’t altruism. It’s a compromise structure that scales under competitive pressure.

2. Apheris Federated Learning: The data-relevance engine
Each company trains models behind its firewall. Data never leaves the institution; only encrypted model updates do. The result is something unprecedented in pharma: an open model with more proprietary data behind it than its closed competitors.

3. Boltz: The enterprise distribution
OpenFold provides innovation insurance — no one can ever “turn off” the science. But even the companies participating wring their hands about indirect data leakage and spoiling IP advantages. Boltz, in contrast, provides execution certainty, including uptime, support, indemnification, and a roadmap aligned with pharma timelines. This is the “one neck to wring” enterprises demand.

Pfizer didn’t just buy software: They’re effectively buying into an ecosystem.

Selling “Boring” Reliability

Red Hat didn’t win by being exciting. They won by being boring — predictably, relentlessly boring. In drug discovery, exciting models get Nature papers. Boring models get drugs into the clinic because every downstream failure compounds cost, time, and regulatory exposure.

Boltz’s recent releases differentiate themselves not by academic novelty, but by industrial utility:

1. Triage-scale affinity prediction. Boltz-2 predicts binding affinities at roughly 1,000x the speed of physics-based FEP. However, as I’ve discussed, the reliability of Boltz-2 for affinity prediction is primarily good for targets reflected in the training set. When applied to novel systems, predictions can be poor. When applied to suitable systems, it sits as a highly valuable tool for Affinity Funneling — pruning vast chemical space so physics methods can focus where they matter. As models are retrained on Pfizer’s internal data, this funnel only improves over time. (And in a sign of how fast this space is moving, SandboxAQ just announced an OpenFold-based affinity-prediction head — another reminder that capability hopscotches, but infrastructure endures.)

2. Biologics ideation. BoltzGen shows strong facility in ideating protein and peptide binders. In biologics, ideation is far from a drug — but in a field where computational methods still lag those for small molecules, this capability is intriguing.

3. Structure and complex prediction. Competitive with AlphaFold and OpenFold, with known failure modes. Not novel — but reliable.

4. Integrated sanity checks. Non-physical hallucinations are corrected before they propagate downstream. Reducing unusable output matters in a numbers game where each failed candidate costs real money.

Boltz isn’t selling a magic solution engine. They’re selling reliable machinery for triage and idea generation.

Why This Matters Right Now

While OpenFold and Apheris demonstrated that the industry can collaborate, the Pfizer deal is the first unmistakable signal that the era of model tourism is ending.

Until now, pharma faced two choices:

• Build AI infrastructure in-house (expensive, risky, slow)

• Enter bespoke discovery collaborations (expensive, IP-complex, milestone-heavy)

Boltz introduces a third option: license-supported, enterprise-grade AI infrastructure.

This is how markets always shift: first to capability, then to reliability, and finally to standardization.

Linux didn’t stop being interesting because it failed. It stopped being interesting because it became inevitable.

A necessary caveat. None of this implies inevitability, monopoly, or technical supremacy. Biology is messier than software, regulation moves slower than code, and today’s leading models will almost certainly be outperformed. Boltz may not be the ultimate winner, and Pfizer’s adoption doesn’t crown a champion. What is new — and difficult to reverse — is the category itself: enterprise-grade, supported AI infrastructure for biology. That shift matters more than which implementation dominates in the long run.

What would break this thesis? This analogy fails if large pharma ultimately decides that AI models remain too strategically sensitive to externalize — if regulatory expectations harden to require full in-house control of models, training data, and inference pipelines, or if foundation models plateau in ways that prevent reliable reuse across targets. In that world, AI infrastructure would remain bespoke, vertically integrated, and company-specific — more mainframe than Linux. The Pfizer–Boltz deal would then read not as category formation, but as an outlier experiment. It doesn’t look that way now, but the longer you’re in this field, the more you realize looking around corners is always difficult.

Phase 3: The Middleware Opportunity

If Phase 1 was packaging the kernel and Phase 2 was paid certainty, Phase 3 will be platform expansion.

Red Hat didn’t stop at the OS. Pharma won’t either.

1. Bio-middleware. Wet-lab instruments speak incompatible dialects. Middleware will normalize raw output — machine outputs, assay results — into AI-ready tensors. This layer won’t be glamorous, but it will determine who sets the standards. TetraScience and Ganymede are emerging as leaders here.

2. The regulatory wrapper (GxP). Versioning, provenance, uncertainty tracking, and audit trails will become mandatory. This will be Red Hat’s FIPS moment for pharma. Boltz doesn’t address this (at the moment), but expect that such capabilities will be integrated into AI infrastructure (by someone) as the industry continues to mature.

3. Orchestration (DMTA loops). The unit of work isn’t a model — it’s Design–Make–Test–Analyze. Boltz sits as a cornerstone here, and is poised to eventually evolve from a prediction engine into a coordinating agent.

The “IBM Exit”

Two decades after making Linux safe for enterprise, IBM bought Red Hat — not for Linux revenue, but to own the support ecosystem.

If a similar outcome emerges here, the most natural acquirers are full-service life science platforms like Thermo Fisher or Danaher. They already live inside pharma’s workflows, and owning the AI infrastructure layer would be a logical extension.

An IPO is also plausible. What matters is that Boltz is selling support, not software — and that model scales. The Red Hat/CentOS split demonstrates how you can chase profit and not alienate (most of) the open source community, and not break any licensing laws.

The Post-Standardization Era

Standardization doesn’t imply a monopoly. Just as the rise of Red Hat didn’t stop the proliferation of Ubuntu, a dominant Boltz ecosystem won’t stifle innovation from others like OpenFold or Chai-1. Instead, it creates a tiered market: community-driven platforms will continue to prosper as the ‘test labs’ for the next generation of researchers, while the world’s largest R&D engines will lean toward the safety, support, and ‘boring’ reliability of a certified enterprise distribution.

Since January

Three enterprise pharma companies have now signed on. Takeda deployed Boltz’s models across its research organization in June, with the same structural fingerprint as Pfizer — platform access with target-specific fine-tuning on their own data, no milestone language, no upfront asset transfer. This week, GSK made three — direct access to Boltz’s proprietary models, the Boltz Lab and API interfaces, agent integrations, and joint fine-tuning on GSK’s own data.

That’s the pattern I flagged in January, now repeating on a roughly four-month cadence. One deal is a bet. Two is a coincidence you note. Three, in six months, across three pharma companies with no obvious coordination between them, is a category forming in real time.

One additional data point reinforces this without complicating it. In June, Pfizer also signed a license agreement with Chai Discovery — early access to Chai-3 and a custom model tuned to Pfizer's biologics workflows. This isn't hedging on Boltz; it's what category maturity looks like. Red Hat customers never ran RHEL exclusively — they standardized on it for production systems while running Ubuntu or CentOS for specific workloads where the fit was better. Pfizer is doing exactly that: licensing execution certainty from Boltz across the stack, and frontier antibody design capability from Chai where the science demands it. The category doesn't require a single winner. It requires standardization — and Pfizer is now buying from two vendors inside the same category.

The caveats above still hold — this doesn’t crown Boltz the permanent winner, and biology can still surprise everyone by moving slower or differently than software did. But the question I posed in January — is this a one-off or an infrastructure play — has an answer now. It’s the second one.

Here’s the next test, stated plainly enough to be wrong: if the cadence holds, a fourth top-20 pharma should sign a structurally similar deal — platform access, no milestones, retrain-on-our-data language — by roughly end of Q4 2026. If that window passes with no fourth deal, or if the next deal looks more like a bespoke discovery collaboration than an infrastructure license, the pattern is weaker than this piece claims. I’d rather commit to that now than only ever grade myself on the calls that already landed.

Conclusion

When Linux became boring, it didn’t just end the operating system wars; it liberated a generation of developers to build the modern web without worrying about the underlying kernel. We are now at a similar precipice in biotechnology.

We aren’t chasing standardization for its own sake. We make AI infrastructure “boring” so that the science can finally become exhilarating again. By solving for uptime, data provenance, and indemnification, the industry gives bench chemists something more valuable than a flashy architecture: the freedom to stop troubleshooting the “operating system” of their lab and start focusing on the patients waiting for a cure.

Boltz appears to be on a trajectory to become the standard — but the real story isn’t the engine. It’s where that engine allows us to go. In a world of reliable, boring infrastructure, the only thing left to be exciting is the medicine itself.

Ted Gioia posed a question on his Substack recently — Ted and I overlapped at Stanford, he covered jazz, I covered pop — and I knew my answer before he finished asking it. When you light up the barbecue on Independence Day, what song do you want playing? When you hear America singing — as Walt Whitman described it — what’s the tune?

I knew my answer immediately: “Montana Song” by David Ackles. Ten minutes of orchestral folk that captures the brutal, unforgiving reality of nineteenth-century westward expansion better than anything else in the American songbook. It lives on an album called American Gothic, released on the Fourth of July, 1972. It is an untouchable masterpiece. You’ve almost certainly never heard it.

I didn’t help with the sales figures. I was eleven when it came out.

I found my copy about seven years later, in a used record bin, for a dime. By then Ackles had already walked away from the recording industry entirely — assessed the entropy, accepted it, moved on. He no longer existed, for all intents and purposes. Around the same time I picked up both Judee Sill albums for roughly the same price. Whatever cultural recalibration was supposed to happen in the 1970s, the market wasn’t doing it. You had to know where the bins were.

The music industry sells a comforting myth: that pure genius rises inevitably, that a strong enough signal punches through the noise. The history of American Gothic is a masterclass in how the actual algorithm operates — and what it requires from you that has nothing to do with the quality of the work.

When Elektra released the album, the critical establishment lost its collective mind. The Los Angeles Free Press called it album of the year. The Sunday Times in London dubbed it the “Sgt. Pepper of folk.” It was produced by Bernie Taupin. It was subsequently championed by Elvis Costello, Phil Collins, Elton John. By every conventional metric of artistic merit, David Ackles deserved to be mentioned in the same breath as Dylan, Cohen, and Mitchell.

It stalled at №167 on the Billboard charts.

After one more commercially doomed effort for Columbia, Ackles walked. He went back to USC, became a theater professor, raised a family on a horse farm in California, and lived — by all accounts — a happy, functional life. He wrote screenplays. He directed student musicals. When a drunk driver put him in a wheelchair and kept him from the piano for eighteen months, he did the physical therapy, got back to the keyboard, and wrote a new musical. He died of lung cancer at 62.

And that, as much as anything, is why you’ve never heard of him.

The pop-culture rehabilitation machine has a very specific intake requirement: tragedy. The overdose, the suicide, the descent into madness. The market loves a tortured artist because it provides a narrative hook — it makes consumption feel poignant, packages grief as a purchasing decision. Nick Drake gets the reissues and the documentaries and the Volkswagen commercial. Judee Sill gets a cult following and a New Yorker piece. The machinery knows what to do with them.

David Ackles refused to supply the raw material. He was well-adjusted. He pivoted. The industry was deprived of its myth, and without the myth, there’s no reissue campaign, no anniversary box set, no algorithm pointing new listeners toward the work. Tragedy is the most reliable explanation for why Ackles stayed buried — though not the only one. His music was difficult, uncategorizable, and arrived at precisely the wrong cultural moment. He wasn’t interested in self-promotion. He left and didn’t look back. Any one of those would have been enough. Together they were insurmountable.

About twenty years ago, Andy Zax at Rhino Records came close — a double CD with bonus tracks, the first time Ackles’ material would ever have been properly reissued in the US, where every original pressing had been out of print since the mid-seventies. It was cancelled at the last minute over a legal entanglement. A few years later, Collectors Choice did straight reissues of the three Elektra albums — the first and only time Americans could buy them new in stores since roughly the country’s 200th anniversary. Those came and went quickly too.

The market keeps not knowing what to do with him.

To understand why the work itself didn’t sell, you have to look at the architecture of the artist. Ackles wasn’t a standard-issue counterculture figure. He didn’t come up through Greenwich Village coffeehouses or the Laurel Canyon scene. He was, fundamentally, a theater guy — born into a vaudeville family, a child actor in Columbia Pictures B-movies in the 1940s, eventually an MFA in film from USC. He was obsessed with the structural mechanics of storytelling: how to write the script, score the music, direct the cast, build the sets simultaneously.

While his peers in the singer-songwriter movement were stripping away artifice to deliver confessional “Dear Diary” lyrics about their breakups, Ackles was building intricate character studies. He wrote like a playwright — establishing the boundary conditions for a character and letting the physics of their tragic flaw play out. His songs were populated by ex-cons returning to find their lovers gone, factory workers wrestling with religious doubt, broken families staring at moving boxes. He wasn’t giving you his raw diary. He was handing you a meticulously crafted one-act play.

He was never the narrator. His songs are all implicitly or explicitly third person, often monologues — a natural extension of the musical theater sensibility. It was brilliant, but it was completely out of sync with the authenticity market that defined the early seventies. And there may be something else, too: I’ve played Ackles for anyone I felt close enough to deserve it. Once, early on, I played it for a girlfriend. I could tell she was hearing it — really hearing it — and that it was moving her. When it was done she said, quietly: “That’s too personal.” She never asked to hear it again. What moved me closer was too close for her. Maybe that accounts for some of the problem. You can admire work that’s trying to reach you without wanting it to succeed.

To make American Gothic, Ackles had to leave the country. He spent two years in rural England, reasoning that you get a sharper perspective on your homeland from across an ocean, free from the daily noise. He scored the enormously complex orchestral arrangements there, then assembled a remarkable group in a London studio: Bernie Taupin serving less as producer than as a supportive presence, letting Ackles run the show entirely; Robert Kirby, who had arranged the baroque strings for Nick Drake’s Five Leaves Left, conducting.

The result was 44 minutes that defied categorization. “Love’s Enough” is a gorgeous, vulnerable ballad that would sit comfortably next to Scott Walker. “Oh, California!” is a sarcastic vaudevillian romp that skewers the Golden State with top hats and canes. “Waiting for the Moving Van” is a domestic autopsy of a failed marriage that will bring you to your knees.

And then there’s “Montana Song.”

Ten minutes and change, it ignores the boundaries of folk and pop entirely. Ackles openly channeled Aaron Copland: sweeping classical modernist arrangements capturing the sheer emptiness of the Western frontier, the brutal mathematics of manifest destiny, the weight of lives spent on a project that neither asked for nor acknowledged them. I’ve listened to it hundreds of times and it still arrives like weather — you don’t hear it approaching until it’s already on top of you.

There is something sustaining about belonging to the small club of people who know this record. That Elvis Costello knows. That Phil Collins — who chose “Down River” as one of his Desert Island Discs — knows. That Elton John — who gave Ackles his producer — knows. When Costello launched his television show Spectacle, he invited Elton John as his first guest. Rather than open with one of their own hits, they performed “Down River” together, backed by Allen Toussaint on organ. Two legends, a forgotten songwriter’s song, national television. That’s the club making its case in public. Their implicit verdict: you’re right, this is brilliance, the world just doesn’t get it, but we do. It’s cold comfort. But it’s real comfort.

As we gear up for the Semiquincentennial — 250 years of America assembling and reassembling its own myth — we’re going to be bombarded with optimized patriotism. The algorithmic playlists will serve up Springsteen and Hendrix and Louis Armstrong. The highlight reel will run on a loop.

July 4, 2026 will mark exactly 54 years since the world was gifted American Gothic. I’ll be exactly 54 years older than I was when it arrived and I wasn’t paying attention. I don’t know what to do with that symmetry except acknowledge it and feel its weight.

What I do know is that the points the album makes fit differently in this world than they did in 1972 — but they are no less relevant, no less moving. We live, we love, we laugh, we hurt. That hasn’t changed. Ackles’ words and music remain timeless in part because they were never specifically of their time to begin with. He wasn’t writing about the seventies. He was writing about people.

American Gothic is not the highlight reel. It’s the working-class melancholy, the spiritual wrestling, the historical brutality, the quiet desperation of people the myth required but never acknowledged. It is the sound of America singing — not the version America wants to hear about itself. It requires a listener willing to sit with difficulty, to let something reach them that is genuinely trying to.

A dime. I got it for a dime.

That’s still the bin where you’ll find it.

I didn’t start this week wanting to write another essay about bio-AI. I have other things I’d like to write about — I promise. But Anthropic co-founder Dario Amodei’s new essay describes a set of near-future applications for bio-AI that are far more aspirational than realistic. He frames these as preparations for a coming capability, not a current one — but the deeper question is whether the framing itself holds up. That question is worth taking seriously, especially coming from a source this visible.

Amodei argues that regulators should begin preparing for AI to displace key empirical steps in drug development — pharmacokinetics, pharmacodynamics, toxicology prediction, dose selection.

Some readers will note that Amodei isn’t claiming these capabilities exist today — only that regulators should begin developing standards for when they do. Fair enough. But regulatory preparedness isn’t neutral. When agencies start designing frameworks to accept AI simulation in place of empirical testing, they are implicitly endorsing a capability trajectory. That signal has consequences — for how pharma allocates R&D, for how the FDA conceptualizes its own role, for how investors price risk. Getting ahead of a capability that may be further away than advertised isn’t prudence. It’s a bet dressed up as preparation.

It is a seductive story in a world trained to believe that scale solves everything. But biology is not software. Toxicity is not a benchmark. And pharma is not whatever they taught in b-school.

And ‘regulators should,’ spoken by someone with the ear of the world, is not just benign futurism.

I have spent four decades in computational chemistry and drug discovery. One lesson keeps returning: the map is not the territory.

AI learns from the map, and the human biology map is woefully incomplete — as I’ve argued in detail [here] and [here]. We are moving in the right direction with data sets such as MGnify, custom-built for better training, with great effort if not great speed. Consider what that looks like even in the best case.

AlphaFold and all that has come in its wake have legitimately changed how protein modeling is perceived. We’ve moved from the early days of homology modeling, where no one would choose to work with that model unless it was the only choice, to now, when the models for protein classes that are well characterized — our old friend the kinase — are genuinely impressive. But even there, the models are brilliant on the macro scale and fail to reflect micro conformational variations: sidechain orientations, water bridges, the effects of point mutations on local structure.

So when you read that AI has “solved” the protein structure problem, it really means it’s moved it to a place where physics needs to take over — where expensive, slow, empirical validation is still required before you build anything around the result.

That’s what “solved” means. Unbelievably good by the standards of a decade ago, and still cause for concern if you are building a program around it. Somewhere you trust but verify before spending millions.

And ADMET and liabilities, where your fantastic clinical candidate goes to die at least 90% of the time, where actual human lives are on the line, aren’t anywhere near that.

But here is the argument Amodei never quite addresses. Even if we assume substantially better ADMET models and near-perfect binding prediction — enough to fill the IND hopper faster and with better candidates — the costly part of drug development doesn’t move. Human biology still decides. You still spend tens to hundreds of millions of dollars per candidate letting the body tell you what the model couldn’t.

Even if better models cut clinical failure rates substantially — and I hope they do, even if I can’t yet see how — the trials that fail will still cost what they cost. You don’t model your way out of a $200 million Phase III failure. You just have fewer of them.

And fewer is real. Reduced expected cost per successful drug has genuine value. Industry estimates of cost-per-approval may come down meaningfully. That matters.

But the average economics across a portfolio don’t change the cost or safety profile of any specific clinical trial — not unless we trust the AI readout as much as we trust the biological one. We don’t. And that’s where we’re nowhere close. Faster candidate generation doesn’t solve the downstream filtering problem. It intensifies it.

Drug discovery is not a simple function to be optimized. That wouldn’t be a problem if we actually understood the real function — the one that runs in the body. Or if we had the means to collect enough data that it didn’t matter. We don’t. When putative drugs hit the harsh, judgmental realities of human biology, the prognostications of the true believers hold no sway.

AI replacing experiment for biology is the kind of thing that sounds great unless you’re actually a biologist. Then, when you hear it, all you can think is: I wouldn’t put that clinical candidate in my body based on a model readout alone. And neither would you.

In Da Nang during the Tet Offensive, Craig Venter worked as a Navy corpsman. The work was not abstract. Young men arrived in pieces; biology was emergency rather than inquiry; the distance between a living system and a dead one was measured in minutes. He left with a blunt, anti-authoritarian streak and a conviction that time—actual, human time—was the only variable that really mattered.

J. Craig Venter died in April—not just a scientist, but a particular kind of scientific force. The kind that assumes timelines are negotiable, that institutions are slow by choice rather than necessity, and that if you apply enough computation and willpower, you can collapse what everyone else has agreed will take decades.

In 1990, the Human Genome Project launched with a belief that bordered on the messianic: read the “book of life”; decode all 3 billion letters of the human genetic code; and everything would follow. We would understand disease at its root. We would cure what had always been incurable. It was slated to take fifteen years and billions of dollars. The logic was linear, almost devotional. Step by step. Letter by letter.

Then came Venter.

He pioneered shotgun sequencing—a methodology that dispensed with careful mapping and instead shattered the genome into millions of fragments, sequencing them in parallel and trusting computation to reassemble the whole. In 1995, his team published the first complete genome of a free-living organism. In 1998, he founded Celera Genomics not to collaborate with the public project, but to race it.

By 2001, both efforts published their drafts—years ahead of schedule.

The timeline had not been met. It no longer applied.

Venter was not easy. He was not always right. He rarely left consensus intact. But he was never cautious—and that distinction mattered more than anyone wanted to admit at the time. To the people who worked with him, the friction wasn’t collateral damage; it was the operating principle. He pushed because he believed biology only yielded to force, and he was willing to be disliked if that’s what it took to make the world move.

I met him four times, each introduced by a colleague at a conference or a social occasion. Each time he greeted me with the same blankness: “I don’t remember you.” It wasn’t personal. He simply moved too fast for people on the periphery to register.

“I prefer to do something,” he said later. “The outcome of doing nothing is perfectly clear.”

That philosophy collided almost immediately with the norms of academic science. He attempted to patent expressed sequence tags—fragments of human genes—triggering an international backlash over whether pieces of biology could be owned. When he later raced the Human Genome Project, critics argued he was privatizing something that belonged to humanity. In at least one sense, they were right: By some accounts, Celera’s temporary intellectual property hold over parts of the genome reduced subsequent scientific research and product development by 20 to 30 percent--though that narrative is hotly contested.

But that friction was not incidental. It was the cost of introducing a different model: biology not as a shared, slowly accumulated public good, but as something that could be industrialized, accelerated, and competed over.

Venter leaned into that model completely.

He sequenced his own genome and branded it—turning his biology into both dataset and declaration, and discovering along the way a genetic predisposition that led to an early skin cancer diagnosis. He then scaled that instinct outward. Outfitting his 100-foot sailboat for the Sorcerer II expedition, he circumnavigated the globe sampling the ocean microbiome—collecting genetic material across jurisdictions in advance of emerging norms around ownership or consent, drawing retrospective accusations of biopiracy as a result. And then he moved from reading biology to writing it—assembling a synthetic genome from digital code and using it to run a living cell.

To many, this looked like escalation into something reckless: biosecurity risks, ethical overreach, the charge that ambition had outpaced governance.

His response on national television was characteristically direct: “Except for the second part, I was taking that as a compliment.” He then encoded a watermark into the non-coding regions of the synthetic DNA—a molecular copyright—and among the embedded text was a quote from J. Robert Oppenheimer: See things not as they are, but as they might be.

Which was, in the end, his only consistent operating principle.

Every controversy surrounding Venter returns to the same question: what happens when someone treats biology as an engineering problem before the rest of the world is ready to accept that framing? The answer is predictable. Resistance. Ethical alarm. And, occasionally, a permanent shift in what is considered possible—one that arrives before anyone has agreed it should.

He helped make the reading and writing of biology feel routine. Mundane, even. Drug discovery, precision medicine, synthetic biology—all of it now runs on an assumption that would have seemed implausible in 1990: that living systems can be read, manipulated, and iterated on at scale, and that if it isn’t happening fast enough, the bottleneck is technical, not philosophical.

That assumption did not emerge gradually. It was forced—by someone willing to absorb enormous amounts of friction in exchange for compressing time.

Which makes his death feel less like the loss of an individual and more like the closing of a particular kind of chapter.

The analogy is artificial intelligence: another computational force collapsing timelines, rewriting expectations. Bypass careful sequencing, run things in parallel, trust the computation to find what careful mapping would have missed.

But there is a difference that matters, and it’s the thing I keep coming back to.

Venter’s acceleration was always tethered to physical reality. DNA had to be sequenced, synthesized, assembled, tested. Cells either ran or they didn’t. Biology imposed limits that could not be argued away.

Much of today’s AI acceleration operates one layer removed from that constraint. It can move faster precisely because it isn’t always forced to reconcile immediately with the physical world. Its feedback loops are faster than its accountability loops. BenevolentAI raised hundreds of millions of dollars on compressed timelines before its first wholly AI-derived program failed in Phase IIb trials in 2023 — not because the biology was obviously wrong, but because with AI you don’t find out until you do. The timelines collapse first; the consequences resolve later. That asymmetry makes progress harder to evaluate in real time—and easier to overestimate.

Celera's bet was that compute plus clever algorithms could outrun methodical science—and it was right. Today's version plays out in foundation models trained on billions of protein sequences, clinical candidates emerging from latent-space exploration, binding affinities predicted by transformers rather than derived from first principles. The sequence is still commodified; now it's structural data and ADMET profiles. The interpretation is still proprietary; now it lives in model weights no one outside the company can inspect. Venter didn't just race the Human Genome Project. He wrote the business logic the entire biotech-AI stack is still running on.

Whether that’s a feature or a warning is still, genuinely, open. The same compression of timelines that got us a draft genome years ahead of schedule is now getting us drug candidates years ahead of what the physics would predict - and with the same unresolved question underneath: what did we skip? AI doesn’t fail the way a bad hypothesis fails. It doesn’t collapse cleanly at the bench like sequencing. It fails late, expensively, in the clinic - after the capital is spent and the timeline has already been celebrated. Venter at least knew what he was racing. The models don’t know what they don’t know.

But he would probably have used them anyway. That was always the point. In Da Nang, triage didn’t reward hesitation or second-order thinking. It rewarded one question that cut through everything else. Will this organism live or die? Venter spent the rest of his life asking it at larger and larger scales. The tools got faster. The question never changed.

This is the fourth in a four-part series that formed a dialog I didn’t entirely plan going in (but where an unexpectedly rich stream of recent AI-related releases pointed me). The Breakfast Test laid the foundation—what it actually means to evaluate an AI release as science. The Reproducibility Tax proposed the MVD disclosure standard: the floor below which a corporate announcement stops being science and becomes marketing. Last week’s essay introduced Rocklin’s MGnify stability dataset, 1.8 million empirically measured folding stabilities that point to what generative models are actually missing. This week, Biohub’s ESMFold2 closes the loop—and, not incidentally, is what MVD looks like when someone actually meets it. When I return to The Long View, I’ll re-widen the aperture—because while AI is interesting, it's not everything.

I’ve spent years thinking about the biologics drug space. It’s not where I started, but I wound up in the middle of it when Schrödinger asked me to put together the first commercial biologics-focused package — BioLuminate. At the time, novel biologics platforms came slowly, and de novo ideation was limited to a small number of labs, like David Baker’s (who subsequently won the Nobel Prize for this work). You’d attend a conference and Baker would present a deck I’d wryly call the “Bob the Protein Builder deck,” alternating “Here’s the target, can we build it?” and “Yes we can.” You’d watch in amazement, knowing it wasn’t that simple except when distilled to the successes — each of which represented years of hard work and deep expertise. Still, it was amazing work, full stop. Ideation was the limit. And even if many paths were more dead end than new gate, you could see where the field was going, albeit slowly.

In the past few years, post-AlphaFold, ideation has gone from a bespoke enterprise of amazing surprises to something we can envision at scale, to something actually being delivered at scale. ESMFold2 is the latest payoff in that story.

Still, I’ve written about enough of these moments to recognize the pattern before the press kit arrives. Protein-systems hype cycles in the AI era always trigger the same two reflexes. One is the eye-roll: same story, bigger model, same limitations, better branding. The other is the opposite mistake: a breathless claim that biology has been solved, evolution outpaced, and the machine transformed into an all-seeing oracle. I’ve watched this field long enough to know both reflexes are wrong in the same way — they let you stop thinking.

In late May, Biohub announced ESMFold2 (in an impressively detailed 106-page publication that should serve as a model for AI model disclosure). This is the field’s latest, and arguably most impressive, step toward routine biologics ideation, and a direct confirmation of the argument I made when BoltzGen arrived last November: that the bottleneck had shifted from ideation to filtering and liability. The ideation capability has taken another genuine leap forward. The liability gap has not.

ESMFold2 is not just AlphaFold with extra parameters and a sharper press kit. And it is certainly not a physics engine in disguise. What it is, instead, is a system that has learned enough of protein biology’s statistical grammar to do serious work without being handed the usual evolutionary scaffolding at inference time.

That is the part worth paying attention to. It shifts the center of gravity from alignment to inference. But it does not shift the center of gravity from binder to drug. And that distinction matters more than almost anything else in this field.

Unexpectedly impressive and game-changing are not necessarily the same thing.

What Actually Changed

For a long time, structure prediction was built around the MSA worldview. If you wanted the model to fold a protein, you gave it the protein’s relatives — the evolutionary gossip, the coevolutionary breadcrumbs. If residue A and residue B had been mutating together for millions of years, that was probably not a coincidence. The power of cross-correlation to imply spatial proximity is an idea that allowed Rosetta to redefine what was possible. AlphaFold turbocharged it into something that changed the field entirely.

This worked extremely well when nature had left a clean trail of homologs behind. It worked far less well for orphan sequences, de novo designs, and proteins without a convenient evolutionary entourage.

ESMFold2 changes the workflow in a way that matters more than it first sounds. It predicts structure from a single sequence, without requiring an explicit MSA at inference time. That does not mean evolution has vanished; it has been absorbed. The model was trained on approximately 2.8 billion sequences — roughly two orders of magnitude more than its predecessor ESM2 — spanning bacteria in deep soil, extremophiles, and metagenomic sequences from ocean and earth environments that have never been fully characterized. What it learned is a compressed representation of the rules that make protein space hang together: not magic, but compression with real boundaries.

There was no guarantee this would work. The further you push between the raw encoding — sequence alone — and the structural readout, the more you’re asking the model to have internalized something that previously required explicit evolutionary annotation. This project assuredly started with the question “Can this possibly work?” rather than “This should work.”

It worked.

Subscribe now

Why That Matters

The strongest argument that this is not just another scaling story is not the benchmark deck. It is the lab.

Designing a binder from scratch is precisely where “looks plausible on screen” usually collides with reality. Biohub designed de novo minibinders against five disease-relevant targets — EGFR, PDGFRβ, PD-L1, CTLA-4, and CD45 — using ESMFold2, and those designs worked. The most striking result: PD-L1 minibinders and scFvs restored T-cell signaling in a functional assay with IC50 values of 1.6 nM and 39 nM, respectively — comparable to 2.6 nM for an atezolizumab-derived control. Atezolizumab is an FDA-approved checkpoint therapy. The designed binders were competing with it. That is where rhetoric stops and molecules start.

The structural validation is equally pointed. Cryo-EM confirmation of an EGFR minibinder showed the experimentally determined complex aligning with the predicted structure at 1.204 Å RMSD. The model predicted the binding pose. The electron density confirmed it. That is not a benchmark number — it is a real structure.

The interpretability angle is more than a side note. Sparse autoencoder work suggests the model has internalized recurring biological motifs, including the nucleophilic elbow — a specific catalytic feature that a single learned representation activates across structurally unrelated enzymes using entirely different scaffolds. That does not mean the model “understands” proteins the way a biochemist does. But a representation that independently recovers that motif, unsupervised, is pattern recognition with real anchors, not statistical theater.

One phrase in the above deserves parsing: “those designs worked” means they bound their targets. It does not mean they are drugs. That gap is where this story gets harder.

Where the Story Breaks

This is where the hype narrative starts to wobble.

ESMFold2 is not a physics engine in disguise. It is a powerful statistical model trained on an enormous slice of evolutionary life, which means it remains bounded by what that slice contains.

De novo designs. The farther you drift from familiar sequence space, the more the model reveals what it is: an interpolator. Push it far enough out of distribution and hallucinations return. The paper itself is careful here: the validated designs emerged from large-scale search, not single-shot generation. Hit rates of 36–88% for minibinders and 15–29% for scFvs are impressive. No, scratch that. They are astonishing hit rates for de novo design. But they are still hit rates.

Conformational diversity. The model also inherits the familiar single-structure bias. Biology loves ensembles, alternates, context dependence, ligand effects, mutations — all the mess that makes proteins interesting. The model, like most structure predictors (and most structural biologists!), wants a single answer. Sometimes that answer is useful. Sometimes it is a tidy lie. The paper’s own characterization of designs as covering diverse binding orientations and structural topologies is encouraging, but it does not address what the model cannot see: the dynamic ensemble around those structures.

The PTM blind spot. Often ignored but tremendously relevant to biologics drugs are post-translational modifications. Scientists frequently minimize them because they are hard. Nature never does. PTMs represent a level of predictive specificity that is beyond most physics-based methods and entirely beyond current AI capabilities — which means every model in this space, however impressive, is working with a map that excludes a substantial fraction of the biology it claims to predict.

Isoforms and edge cases. Isoforms, splice variants, and similar biological irregularities are another reminder that protein biology refuses to be a clean sequence-to-structure prompt. These are the cases where statistical fluency starts to fray, because the biology has stopped behaving like the training set.

Biosecurity. The paper handles this carefully; most coverage skips it. A model this capable of navigating protein function space across taxonomic neighborhoods has dual-use implications that are not hypothetical. The open license compounds this: a closed API can be monitored; a downloaded model cannot. Models trained to understand functional motifs across sequence space can, in principle, aid the design of proteins with harmful biological activity — including toxins and pathogen components — with less domain expertise than was previously required. This is not a reason to dismiss the work. It is a reason to name it plainly, because the field’s tendency is to treat it as a footnote until it isn’t.

The Liability Gap

The most important limitation has nothing to do with the model’s architecture.

Even when ESMFold2 designs a binder that works — genuinely, experimentally works — it has solved only the first step. In biologics, the distance between a functional hit and a developable drug is where most programs die. De novo scaffolds are non-self by definition. Immunogenicity risk is baked in from the start. Designed minibinders have no track record of PK stability; most would be proteolyzed in minutes in vivo. Expression yield, aggregation behavior, and manufacturability are nowhere in the model’s scoring function. The paper acknowledges this directly — the designs were characterized for homogeneity and non-aggregation by size-exclusion chromatography, which is a start, but it is a long way from an IND package.

None of this is ESMFold2’s fault. It was not built to solve those problems. But it is worth saying clearly: the model has dramatically accelerated ideation while leaving the liability landscape exactly where it was.

Rocklin’s MGnify database of 1.8 million empirically measured, publicly accessible, folding stabilities that I wrote about last week is exactly the class of infrastructure the liability side of this problem actually needs — not more generative capability, but more empirical boundary. Wet-lab reality rather than computational proxy. Those ΔG values will likely be more valuable in five years than they are right now, because the models capable of fully exploiting them haven’t been built yet. The Biohub work is, in real time, demonstrating exactly why.

This isn’t just a theoretical critique; it is a practical wall that the generative community is hitting in real time. Look no further than the technical report for BoltzProt-1 that dropped this week. The authors didn’t announce a radical leap in generative physics. Instead, they spent much of their paper detailing a brutal two-stage gauntlet: a novel interaction-scoring model, BoltzPPI, trained on the “museum of winners” in the PDB and patent databases to filter candidates before experimental testing, followed by a comprehensive developability panel measuring thermal stability, hydrophobicity, polyspecificity, and monomer purity via analytical size-exclusion chromatography. The generative engine still can’t design against aggregation upstream — it has to be screened out after the fact. When a flagship structural framework celebrates a 58% survival rate across physical liability filters as a headline result, that’s a pretty good don’t-just-trust-me demonstration that the field is catching on to where the real problem lives.

Where This Leaves Us

We are closer than we have ever been to the point where identifying a biologic target and generating something that binds it well is a tractable, fast, and reliable problem. That is real, and it deserves to be said without hedging. But the liability landscape has not moved at anything close to the same pace. In many cases, it hasn’t moved at all.

We can generate binders with growing confidence. We still cannot predict, with any reliability, whether those binders will express cleanly, survive in vivo, avoid triggering an immune response, avoid aggregation, retain suitable viscosity at point of injection, or stay away from the ten thousand things we did not intend them to touch.

The larger commercial prize will belong to organizations that can close that gap — by automating liability assessment at scale, by building the datasets that make computational liability prediction trustworthy, or both. The companies that figure out how to generate clean, systematic data on immunogenicity, PK, aggregation, and specificity — and build models trained on that reality rather than on structural proxies — are the ones that will turn this ideation revolution into a development revolution.

ESMFold2 demonstrates how much of protein biology’s statistical grammar can be compressed into a learned representation. That achievement is real, arguably shockingly good, and it matters. But drug development is ultimately a game of elimination, not generation. The industry’s central challenge in biologics is no longer imagining candidate molecules. It is discovering which of them can survive the gauntlet that begins after generation ends.

ESMFold2 is not an oracle, and it is not a simulator. It is a remarkably fluent model of protein biology. The hardest questions remain. That is not a failure of the model — it is a description of where the frontier has moved.

By now, this is less revelation than consensus: artificial intelligence, at its current stage of development, is largely an interpolation engine. I’ve said it. Others have said it. If we’re being honest with ourselves — and the field is getting better at this — we all know it. AlphaFold is interpolation done at astonishing scale and sophistication. So are the generative models writing plausible chemistry and designing proteins that fold on the first try. The pattern recognition embedded in transformer architectures trained on billions of sequences and structures is extraordinary. None of that changes the underlying constraint.

Interpolation has one limit no architectural cleverness can overcome: you can only interpolate within the map you have. Beyond the edges — in the vast unmeasured territory outside the training distribution — the model isn’t extrapolating from first principles. It’s guessing. Sometimes beautifully. Sometimes catastrophically. But guessing. Prettying it up as “hallucination” doesn’t make the problem go away.

Which means that in biological AI, the deeper question has never been simply which model. It’s which map? And if you’ve spent any time in this field, you know the map has always been the bottleneck. Despite a half-decade marked by a commitment and PR dedicated to demonstrating otherwise, the last eighteen months have quietly marked something more important than any model release: we have entered the era of the hyper-dataset.

OMol25 and What It Proved

A few months ago, Meta’s FAIR team released OMol25 — roughly 100 million density functional theory (DFT) calculations covering small organic molecules, metal complexes, and catalytic systems. If you work on the quantum mechanics side of computational chemistry, you understand immediately what that number means. In the early days of this field, I’d wait days for a single DFT job to finish; seeing a hundred million of them land at once feels like cheating. One hundred million calculations, carefully curated, rigorously validated, and released openly to the world, represents a scale of investment in data rather than models that simply hadn’t existed before in this field. Not only that, but it couldn’t have. A few years ago, this wouldn’t have been possible even if you’d harnessed the world’s compute power for only this project.

OMol25 matters for a reason beyond its size: every data point reflects actual electronic structure — the real energy landscape of a real chemical system, grounded in quantum mechanics. When you train on OMol25, you’re not training on human annotation or proxy measurements. You’re training on physics. The still-developing result is a new class of machine-learned interatomic potentials that, in many domains, approach DFT accuracy at a fraction of the cost. The breadth of that training data provides for more reliable extensibility across molecular classes. That’s the payoff: Better training begets better models.

But OMol25 is still computational at its core. The map is generated by computers running calculations on computers. DFT has well-documented failure modes — dispersion, strong correlation, excited states — and the data is only as trustworthy as the methods that produced it. What would the biological equivalent look like? What would it mean for the messy, irreducibly experimental world of proteins and drug discovery to get its own hyper-dataset — one grounded not in calculation but in wet lab reality?

This month, we may have found out.

MGnify: The Most Impressive Protein Stability Dataset Ever

The Rocklin Lab at Northwestern University, backed by the OpenFold Consortium, has released the MGnify Stability Dataset: empirically measured folding stabilities for 1.8 million diverse protein domains. Absolute thermodynamic stability — the actual free energy of unfolding (ΔG) — measured directly in the lab, not estimated by a model, not inferred from homology.

The method that made this possible, cDNA display proteolysis, is the kind of experimental cleverness that makes you wish you were on the defining paper. Protein variants are translated in a cell-free system, each one physically tethered to the cDNA that encodes it — so you never lose track of which sequence produced which protein. The whole library, close to a million sequences at once, is then exposed to proteases. Stably folded proteins bury their cleavage sites and survive. Unstable ones get chewed to pieces. Sequence the survivors across a protease gradient under native conditions, and you can reconstruct the full denaturation curve — and from that, the absolute ΔG — for nearly every sequence in the pool. Roughly a million measurements per week.

I’ve spent months fighting for a handful of clean ΔG measurements during a project. Watching a million of them drop in a week feels like someone turned on gravity. Maybe the AI planet will turn out habitable after all.

The fine-tuned models that came with the release — SaProtΔG and ESM3ΔG — can now predict absolute folding stability with a root mean squared error of 0.8 kcal/mol. That approaches (or perhaps exceeds) the limits of experimental error. For a field that has historically struggled to predict even the direction of stability changes reliably, that number deserves a moment. That’s truly amazing — even if the cynical scientist in me is already reaching for the caveats.

But the headline is not the model performance. It’s the fact that the data now exists. And to understand why that matters, we need to talk about a problem that goes deeper than data volume.

The PDB is a Survivorship Trap

The Protein Data Bank is one of the great scientific achievements of my lifetime. But it is, structurally, a museum of winners.

Every protein in the PDB made it there because it was stable enough to express, purify, and crystallize or survive cryo-EM or be labeled for NMR. The proteins that misfold, aggregate, or fail experimental tractability rarely arrive. They disappear from the pipeline, unrecorded, unteachable. The PDB has an inherent selection bias baked into its existence: it only knows what success looks like.

When we train generative AI on the PDB, we teach it exclusively from positive examples. The models learn the grammar of stable proteins, become fluent in the language of well-folded structures. And when asked to generate something new, they produce sequences that look grammatically correct — beautiful, confident, PDB-compatible — but fail in the test tube, because they’ve never been shown what failure looks like. Never shown where the edge is.

It’s the biological equivalent of training on Instagram: everything looks perfect until you try to live inside it.

But scale alone doesn't fix this — and most attempts to fix it actually reproduce it. The large aggregated repositories that now exist — CELLxGENE, ChEMBL, the major sequence databases — were assembled post-hoc from individual experiments designed for individual purposes, with no global intention of covering the distribution. The batch effects, the systematic bias toward tractable systems, the sampling skew toward whatever was fundable: all of it gets inherited at scale. More data from the same flawed collection philosophy doesn't extend the map. It just makes the blind spots harder to see.

The MGnify dataset doesn’t just add more positive examples. It maps the boundary. The cDNA display proteolysis assay measures the stable, the marginal, and the wholly unstable in the same experiment, under the same conditions. For machine learning, those failure measurements are not noise. They are the essential signal. In classification terms, we’ve moved from a one-class problem — here are the positive examples, now generalize — to a genuinely two-class problem with real, empirically derived negatives. That changes the geometry of what’s learnable.

Subscribe now

A Positive Word About Negative Data

Here is the part I don’t think has been said clearly enough.

Everyone in computational drug discovery agrees, at least in principle, that negative data has value. Failed experiments, unstable variants, inactive compounds — these are all training signals. The field says this constantly. But in practice, the private sector, which runs many of the highest-throughput assays in the world, has rarely built durable repositories for molecular failure. They rarely even publicly document the successes. But failure? No one is spending an extra cent to document that internally or externally. This isn’t negligence. It’s economics.

The logic of drug discovery optimization is ruthless: identify what works, push harder on what works, discard what doesn’t. Once a compound fails, once a protein variant proves unstable, once a scaffold is deprioritized, there is no business pressure to store and annotate that failure carefully. The goal is a drug, not a map. Resources flow toward the signal, not toward documenting the noise. And so, decade after decade, negative data is generated and discarded — or never generated at all, because the assays are designed to find hits, not to systematically characterize the space around them.

I’ve watched entire programs generate exquisite failure data and then quietly let it evaporate. Not out of malice — out of momentum. I’ve also watched data scientists ask for additional data to fill out a training set for a series where success had already moved on — and get stared down in that “I think we’ve already had this discussion, I hope you value your job” way.

The MGnify dataset is unprecedented not just because it’s public. It’s unprecedented because it was designed to capture failures. The cDNA display proteolysis assay doesn’t screen for the good ones and discard the rest. It measures everything. The experimental design itself embodies a different set of values: exploration over exploitation, mapping over mining. No single drug program can justify the cost of carefully mapping space that it has already decided not to pursue. The private sector rarely adopts this approach — and when it does, it tends to be in spite of local optimization incentives, not because of them. (There are exceptions — e.g., Terray, Recursion, Generate — but notably these companies built their strategy in part around data and model advantage rather than traditional portfolio economics.)

The Grain of Sand Problem

I want to be careful not to oversell the moment, because the honest accounting is more interesting than the hype.

OMol25’s 100 million data points are a feat. And yet the chemical space of drug-like molecules is estimated at roughly 10⁶⁰ compounds. One hundred million is less than rounding error in that context, and the people who built OMol25 know it. Terray Biosciences recently announced 14 billion experimental binding measurements — a genuinely staggering number in the service of a favorably differentiated approach — and it’s still the vastness of the universe away from where we’d need to be to call the map complete.

The MGnify dataset faces an analogous reckoning. The domains measured here range from 60 to 80 amino acids — small, isolated structures — and while 1.8 million measurements spanning over 200,000 unique sequence families is extraordinary, the total sequence universe for even that length range is effectively incomprehensible. The map is better. It is not complete.

A sharply defined gap can be attacked. A fog cannot.

There’s something worth sitting with here. No matter how good our robots get, no matter how elegant our high-throughput assays become, the valley between what we can measure and what can exist will never close. This is true of small molecules. It’s even truer of macromolecules. So the hyper-dataset era doesn’t solve the extrapolation problem — it makes a dent in it. We can hope that AI eventually learns world models sophisticated enough to substitute for explicit physics and push the interpolation boundaries outward indefinitely. But right now, and for the foreseeable future, we’re still going to hit the fuzzy edges. Which means AI remains what it has always been in this field: a powerful complement to physics, not a replacement for it. The map grows. The territory doesn’t shrink.

Being Honest About This Dataset

Let’s set the goalposts correctly, because overshooting them is how fields lose credibility.

Absolute thermodynamic stability is not the same as kinetic foldability. A designed protein might occupy a deep, stable energy minimum and still fail to reach it — because the folding pathway runs through high-energy intermediates or aggregation-prone transient states that an equilibrium measurement never sees. The dataset tells you where the folded state lives. It says nothing about whether a protein can actually find its way there. The prion protein is the canonical cautionary tale: PrPsc, the misfolded pathological form, is thermodynamically more stable than the native structure — yet the normal protein folds correctly because the kinetic pathway overwhelmingly favors it. Thermodynamics tells you which valley is deepest. Kinetics determines which valley you fall into. This dataset only measures thermodynamics.

The assay is cell-free by design. In vitro is not in vivo. The chaperone networks that rescue marginally unstable proteins inside human cells are absent. Post-translational modifications aren’t present. A protein that looks unstable in a test tube might be perfectly viable in a cell; one that looks solid in isolation might never survive the crowded, contested environment of the cytoplasm. The dataset measures intrinsic biophysical stability with extraordinary rigor. That distinction doesn’t diminish the dataset. It clarifies what kind of map it actually is.

And then there’s the one that gets forgotten whenever stability becomes the optimization target: a stable protein is a prerequisite for a drug. It is not the drug. The most thermophilically stable proteins in this dataset — exquisitely hardened by evolution against the harshest conditions on earth — are, for therapeutic purposes, often inert bricks. Function requires flexibility. Binding requires dynamics. Allosteric signaling requires the capacity to change shape. Push an AI hard enough to maximize ΔG and it may design proteins that could survive a volcano and accomplish nothing useful inside a human being.

These aren’t disclaimers. They’re the correct map of what comes next. Science only progresses on truth. Setting the honest baseline allows a tower to be built. Drug discovery doesn’t reward safe. The scientist in me insists we use open eyes — but what’s been accomplished is still worthy of eyes wide open.

The Real Value: Better Living on the Edge

We’ve spent twenty years asking AI to learn biology from the artifacts of biology — structures deposited because they were stable enough to deposit, compounds tested because they looked promising, sequences preserved because evolution had already validated them. Then we wonder why the models struggle with true novelty.

The edge is where the learning lives. Between what folds and what doesn’t. Between what binds and what fails. Between what works in the cell and what only works in the tube. That boundary — not the comfortable center of known structure space — may be where we finally find the breakthroughs Eroom’s Law has been waiting for.

Models will keep changing. Architectures always do. But empirical measurements accumulate. The 1.8 million ΔG values released this month will outlive the systems that first trained on them. They will become benchmarks, calibration sets, and eventually infrastructure. They will likely be more valuable in five years than they are right now, because the models capable of fully exploiting them haven’t been built yet, and because there will likely be other massive complementary datasets that measure other physical and relevant characteristics.

That’s why the OpenFold Consortium’s decision to fund not just a model but the underlying data generation is a structural intervention, not a publicity move. The open science community has spent years building better algorithms on top of impoverished data. The implicit bet in this release is that the bottleneck was never the math. Anyone who has watched this field long enough already suspected that. Now we’re doing something about it.

For anyone building on top of biological AI right now, the strategic implication is uncomfortable: the moat is no longer the model. Models commoditize. The moat is the map — which means the companies with durable advantage are the ones funding the data generation that no single drug program can justify on its own. That's not a research budget decision. It's a platform decision. And most organizations aren't making it yet, because the incentive structure of drug discovery actively selects against it. The ones who figure out how to fund exploration in a system optimized for exploitation will look prescient in five years. The rest will be licensing someone else's infrastructure.

That may be the deeper lesson of the hyper-dataset era: progress doesn’t always come from building smarter systems. Sometimes it comes from deciding that the failures are worth keeping.

Not because failure is more interesting than success.

Because in forty years of watching this field, the boundary between them is the only place I’ve ever seen a genuine leap get made.

In a previous blog, I described the Breakfast Test to try to separate how much of what we hear about AI in pharma is hype and how much is true innovation. The conclusion is that if you don’t show me where your method breaks, it makes it hard to assess how significant the claimed victories are. What I didn’t fully address in that article is the cost of omission, and how we can address it. Here, I propose something I call Minimum Viable Disclosure (MVD) — the floor below which a corporate announcement stops being science and becomes marketing.

Winning the Battle and Losing the War

I want to start with a confession that cuts against the grain of what you might expect from someone in my position.

I’m personally aware of compelling scientific work from private companies that never reached peer-reviewed journals. Not because the science was weak or the intent was bad, but because journals like JCIM began requiring access to the software itself as a condition of publication — a demand no legal team could sign off on. So the work landed on non-refereed platforms, often with less disclosure than the company would have willingly provided under a more reasonable standard.

Read that again: demanding full transparency produced the worst of both worlds. Lower disclosure. Still no software access. The overcorrection just moved the conversation somewhere with no standards at all.

The Crush of the AI PR Deluge

How many AI method brags do you get in your email and LinkedIn feed a week? If you believe everything you read (because that makes you a better person), you’d imagine that every problem on earth has now been solved, everyone will be out of work by Friday, and machines are taking over the Earth. Until next week, when the whole process repeats itself. Even in pharma — an area more grounded in observation and science than most — every few weeks a new technical report drops claiming to have shattered the state-of-the-art in protein-ligand binding or structure prediction. We get the high-gloss plots, the “near-experimental accuracy” headlines, and the inevitable LinkedIn victory laps.

But beneath this breakneck speed lies a quieter, more expensive hemorrhage: The Reproducibility Tax.

The Scientific Contract vs. The Closed-Source Engine

As a VP of Product at a company building proprietary scientific software, I’m not some open-source monk. R&D is expensive; you need defensible assets to keep the lights on and the investors happy. Organizations like Isomorphic, Terray, and my own team at QSimulate have to protect the engine.

But we’re starting to conflate two very different things:

1. Protecting the commercial asset (the code and the weights).

2. Exiting the scientific contract (hiding the methodology).

You can keep the engine behind a paywall and still publish the map. When you withhold both, you might have a valuable product, but you’ve officially stopped doing science. I get it: The commercial incentive is real and entirely legitimate. But the same opacity that feels like a moat from the inside reads like vendor risk to pharma buyers who have to sign off on real dollars and real patients.

Virtual Witnessing: The 17th-Century Standard We Forgot

In my own computational publications over the years, the methods sections were often as long as the results themselves. There was always that “yada yada” layer — the standard water models and force fields. But it never crossed my mind to omit the novel bits. The unspoken rule was simple: provide enough detail so a peer could actually reproduce the work. It didn’t have to be easy — you might still have to sweat over your own code — but the map was there.

And I didn’t do this because I woke up one day and decided, “Hey, let me make publishing harder for myself.” This was an adherence to virtual witnessing — the principle, formalized by Robert Boyle in the 1600s and translated into the digital age by Jon Claerbout, that the paper is just the advertisement; the scholarship is the full computational environment behind it.

Today, too many corporate AI announcements rely on “existence proofs” — basically just saying, “We did it.” Without the details to stress-test those claims, they don’t actually enter the scientific record; they just take up space.

Subscribe now

The Trap of Parallel Speculation

When the “how” is missing, everyone else pays for it — in time, in budget, and in misplaced trust. A high-profile claim drops with zero disclosure, and suddenly, research groups worldwide waste months trying to reverse-engineer hidden data splits or hyperparameters.

Consider the Isomorphic announcement discussed in last week's post. Strong kinase benchmark results, no failure analysis, no data provenance on what was excluded. Under MVD, the disclosure would take one paragraph: similarity threshold used, targets excluded and why, and one system where the model degraded. That paragraph doesn't exist. So the field spends months reverse-engineering what it would have said — and still can't be sure what it's building on.

The experimental world is no stranger to this. History is rife with examples of experimental results that were published and later determined to be irreproducible, but not before millions of dollars were spent by other labs to try. Computational scientists should do better, because the alternative is an expanding economic disaster.

This isn’t a new concern — Pat Walters called out the risks of publishing computational results without sufficient code and data transparency in a letter to JCIM back in 2013, before the AI wave. What’s changed is that corporate AI announcements have industrialized the problem.

We’re digitizing and accelerating inefficiency instead of solving it.

Look at the original AlphaFold2 release. DeepMind didn’t lose their edge by being transparent; they cemented their status as the defining lab of the decade. Transparency was the strategy, not a liability. Similarly, AlphaFold might never have been developed at all without the publicly available methods and open scientific exchange of the broader structural biology community — including Rosetta — which contributed significantly to the initial AlphaFold approach.

We don’t need new principles. The DOME Recommendations (Data, Optimization, Model, Evaluation) already define a baseline for AI transparency. A closed method can still win benchmarks — but if no one can examine how it wins, it cannot function as scientific progress, only as a product demonstration. And no journal adoption of DOME will do anything for glorified press releases, unless editors decide that “woo-hoo” announcements no longer qualify as science — which, so far, they haven’t.

Minimum Viable Disclosure (MVD): A Door The Legal Team Will Actually Open

The life sciences already lose an estimated $28 billion annually to non-reproducible preclinical research. AI, without better dissemination standards, promises to make that worse. The field already knows what good looks like. The problem has never been the standard — it's been the incentive to meet it.

How exactly? Pretty simple. In principle, at least: Minimum Viable Disclosure (MVD).

Roll back to the way science has been disseminated for four centuries. If you’re going to “publish,” it’s science first, business second. PR is good, but don’t conflate the two. Publish enough of the blueprint that a competent peer can reconstruct the approach — but let your legal team and investors sleep by not giving away your source code or model weights. Provide the map.

DOME already defines the gold standard for doing this in peer-reviewed science. That’s the ideal. But the reality is that — predictably — it’s being broadly ignored. I’m proposing a less stringent standard that companies might be able to live with. MVD is that floor below DOME — the minimum a corporate announcement needs to clear before it earns the right to be taken seriously as science rather than marketing.

Unlike DOME’s full-disclosure standard, MVD is designed not to trigger the legal veto. Remember the JCIM overcorrection I described earlier: demand too much, get less than nothing. MVD is calibrated to what a legal team can actually sign off on — which means it has a chance of being adopted, which means it has a chance of actually working. It softens DOME’s gold-standard expectations into a tempered middle ground that the commercial world can realistically live with. Of course, MVD works only if the community decides that disclosure below the floor disqualifies a claim from scientific consideration — which means the power to enforce it sits entirely with readers, reviewers, and buyers, not with journals or regulators.

Public trust in science has appreciably dropped over the past decade, and in part that’s a reflection of creep towards science as politics. We need to back away from AI as campaign promises towards truth and transparency. Give us enough so we can legitimately ascertain if we’re looking at an actual leap. To wit…

The Three Pillars of MVD

Three things, minimum. None of them requires a lawyer’s approval (I can see the champagne corks popping already).

Data Provenance: We need the logic, not the raw data. Disclose the similarity thresholds, the temporal splits, and how you handled deduplication. No one is going to share their crown jewels, but provide sufficient broad descriptors so the rest of us can at least get a sense of what you’re working with. (“25,000 putative kinase inhibitors, of which 2% were active better than 1μM.”) That matters because data leakage is the most common reason “miracle” results fail the moment they hit a real-world test. There’s a reason most comp chemists’ favorite two words are “Prove it.”

Failure Analysis: A highlight reel is boring and incomplete. Tell us where the model breaks. The boundary of success is often more informative than the success itself. When I’m developing code myself, I call this corner analysis — look at the extremes, and be honest about where the model doesn’t hold up. That’s where the real science lives.

Transparent Baselines: Measure progress against strong, contemporary baselines, not straw men. We need to know whether you’ve actually built a better algorithm or just have better hardware — whether you’ve made a leap in the model or just in the training data. Acing the chip shots doesn’t tell us much about either.

This is a framework, not a manifesto. Making MVD operational requires more than three bullet points. It requires people like Pat Walters and me — one holding the line on rigor, one arguing for the pragmatic middle — sitting in the same room as the legal teams and C-suite decision-makers who have to live with the consequences. That conversation hasn’t happened yet. It should.

Winning Shouldn’t Be Relegated To Your PR Team

Competition is great; it’s what drives the AI Le Mans. It can lead to Nobel Prizes, and it can make people rich.

But hard science is bigger than any one team, and the best science has always compounded through sharing. When a new drug reaches a patient, credit flows to the scientists who moved the work, not the PR team that moved the announcement. We helped move science faster. And we moved faster ourselves. Both can be true. Both should be true.

Not radical. Just science. Sell the engine. Publish the map.

Author’s Note: The teams mentioned here are doing serious and impressive work. This piece is not about integrity; it’s about standards. As AI methods move into high-stakes domains like drug discovery, how we communicate methodology matters as much as the results themselves.

It started with a splash. Now it’s accelerating fast

That could describe the increasing number of AI-first platforms promising to solve the ligand binding problem without the need for physics. In the past several weeks, Isomorphic and Terray have thrown their hats into the ring.

As someone with a long (long long) history with FEP and similar approaches, maybe the bold headlines should disrupt my breakfast. But one of the advantages of being closer to the end of your career than the beginning is that not much disrupts your breakfast. The Breakfast Test is simple: when a new method lands, can I identify where and how it fails? If the authors haven’t shown me that, I’m taking the headline with significant salt. So let’s take a deep breath and look carefully at what’s actually being shown.

Less than a year ago, the Boltz team intrigued the field with Boltz-2, reporting ligand binding predictions roughly three orders of magnitude faster than physics-based approaches like FEP. The headline claim — near-FEP performance on the cheap — was bold. What made the paper compelling, though, wasn’t the chest beating.

It was the honesty.

Alongside some strong benchmark results, they included a real-world attempt on some systems via collaborators at Recursion. And that section made something crystal clear: when it worked — like on the benchmarks — it worked great. When it didn’t (like on the Recursion systems)… well, ugh. Not a “this sucks” ugh. A “welcome to the real world” ugh.

If you work in domains the world has seen repeatedly — “I’ll take Kinases for $500” — AI can be your bestest friend of all time. But when you’re trying to bring a novel drug to market against a less familiar target class, things get murkier. And here’s the hard lesson: the big boss doesn’t care whether your software likes the system. Make it work. Somehow.

Boltz-2 showed both the best case and the real case, and that’s scientific honesty. It also reinforced something I’ve subsequently argued: AI and physics aren’t enemies; they’re complementary. The “Affinity Funnel” is the natural architecture — fast, less accurate AI screening up front, followed by slower, more robust physics methods as the computational acid test. Not every system falls in AI’s sweet spot, which is precisely why the Affinity Funnel architecture matters — physics validates where AI can’t.

Because Boltz is open source, people tested it immediately. The results weren’t always pretty — but the field saw the potential. That’s how this ecosystem should work. Excitement generated, excitement tempered — it was inevitable we’d see Mach II. And we have.

Isomorphic: Impressive, but realistically, where does it fail?

Isomorphic Labs recently announced their affinity module, IsoDDE, built on their closed AlphaFold 3 infrastructure. Google’s decision to close-source AlphaFold 3 was the lighter to the fuse of open efforts like Boltz, OpenFold, and Chai-1. The response from the open community has been extraordinary; we’ve moved from “Well, that’s that” to “Whatever, I have options.”

Still, Google hasn’t been idle. Isomorphic now reports performance better than both Boltz and FEP across several benchmark sets, which, on the surface, is extremely impressive. Especially the assertion that they can get predictions of actionable quality for absolute binding free energies across diverse scaffolds. The top-line claims prompted Nature to call it ‘An AlphaFold 4’— and the excitement is understandable.

If you’re early in your career in physics-based comp chem, this might feel destabilizing. Take a breath. There are at least three reasons for caution.

First, benchmark scope. They report strong results on a subset (four targets) of the FEP+ benchmark set, the OpenFE benchmark, and CASP16 ranking tasks. These are public, well-conditioned systems that largely reflect the chemical space we’ve collectively explored before. Crucially, regarding the FEP+ comparison, they used CDK2, TYK2, JNK1, and p38. Notice a pattern? They are all kinases, one of the safer terrains for AI. Kinases are among the best-characterized systems with abundant structural and binding data. The targets they excluded — BACE1, Thrombin, PTP1B, and MCL1 — are the “physics-hard” cases dominated by complex electrostatics, charge states, and water networks. Beating FEP on kinases is great, but it doesn’t prove you can handle the messy reality of things like covalently bound ligands or metalloproteins. Methods that achieve <1 kcal/mol Mean Unsigned Error (MUE) on kinases can degrade to >2 kcal/mol on these tougher systems. Physics methods fail too — but when they do, the failure is often interpretable.

Second, leakage risk. This problem is as old as QSAR (ask your grandpa to tell you about it someday): leakage between training and test sets. Obvious leakage — a nearly identical crystal variant in training — is inexcusable. This is easy to avoid, and I don’t imagine that’s an issue here. But more subtle leakage — highly conserved kinases in training when predicting another kinase — is harder to eliminate. Isomorphic attempts both temporal and modest sequence (similarity > 0.90) filtering. This is OK, but similarity leakage is far more structural than chronological, and two proteins with 0.89 similarity can be very similar in 3D space, and can sometimes be nearly indistinguishable in the binding site. When leakage creeps in, AI can look superhuman. Remove it rigorously, and results often degrade — sometimes dramatically. Again, there’s no explicit evidence of leakage here, but it remains an ever-present risk in ML.

Third, and most important, there’s no failure analysis. Boltz included Recursion’s tough real-world systems where the model struggled. Isomorphic does include one prospective success story in their supplementary material — a helicase PanDDA experiment where they successfully identified a pocket — but they do not show the “break the model” cases. For a closed commercial effort, highlighting wins is understandable marketing. Scientifically, however, it leaves the key question unanswered: what happens outside the comfort zone?

If someone only shows you good results, skepticism is not cynicism. It’s methodological hygiene.

Oh, and it’s worth noting that while my focus here is on binding affinities, Isomorphic also claims new beyond-AlphaFold 3 accuracy for protein structure prediction tasks such as antibody modeling (particularly the pesky H3 loop). Credit where due, and all that.

Subscribe now

A Reminder About AI

AI can be extraordinary within its domain of applicability, and worse than terrible outside it. Worse, because it hallucinates answers that look plausible. If bad predictions were obviously absurd, we’d be fine; the danger is when they look perfect.

Basically, AI is a sociopath able to convincingly tell believable mistruths without conscience.

We just saw this play out across all three papers. Boltz passed the Breakfast Test — it showed us where it failed. Isomorphic and Terray, for understandable commercial reasons, did not — at least not publicly. Prospective validation shared with customers may tell a different story, and I’d welcome that data. That asymmetry isn’t a minor footnote; it’s the central interpretive lens for everything else in this post.

Once you accept that, domain-of-applicability analysis stops being optional. And that remains true even if your approach tries to attach an uncertainty metric to predictions it isn’t sure about. These remain difficult to calibrate robustly, particularly outside training distributions.

Terray: A Different Axis of Attack

Only days after Isomorphic’s announcement, Terray Biosciences introduced TerraBind. Their move is interesting: rather than generating explicit all-atom 3D ligand conformations through diffusion-style modeling, they skip that step and infer binding from reduced coarse-grained ligand (and target) representations.

Chemically, dimensional reduction isn’t crazy: You don’t need every atom to capture the signal. In principle, a SMILES string plus an infinitely accurate and fast QM engine contains everything you need. You could, in theory, infer the 3D structure of the ligand from its SMILES without ever generating it, and then infer the conformational variability without ever calculating it, and then determine the interactions of that flexible ligand — again without explicit 3D modeling.

You could. But training sufficiently to carry out all those levels of inference? That’s crazy talk. Not. Gonna. Happen. Terray knows that, and isn’t proposing that. Instead, they’re reducing space and required inference by using the coarse-grained model (protein C-beta atoms and ligand heavy atoms). They’ve entered the scientifically interesting domain of ‘let’s find out if that works’ — and to their credit, they’ve built the data infrastructure to have a real shot at answering it.

To increase their chances of moving the bar, Terray also relies on their array-based experimental binding measurement platform, which they state has resulted in 14 billion unique target-molecule measurements. Though that’s still the vastness of the universe away from the 10⁶⁰ possible small molecules, it certainly moves in the right direction.

Terray’s reframing is smart. They don’t claim ΔG values that match experiment 1:1 with low MUE. They claim rank ordering. That matters, because rank ordering — not R², not MUE — is what operational drug discovery actually uses. I’ve argued this since introducing the Predictive Index in 2001 — a weighted Spearman for series analysis — when analyzing the first published FEP data for a large congeneric series. (That series was p38 — also part of today’s FEP+ data set. Everything old is new again.) If you can correctly rank compounds, you can move chemistry forward.

Terray claims roughly a 20% improvement over Boltz-2 in Pearson correlation, about 26x faster execution than Boltz or Isomorphic, and therefore on the order of 26,000x faster than FEP. Give or take.

Twenty-six thousand here, twenty-six thousand there — pretty soon you’re talking about real numbers.

But interpret that carefully. For systems where these methods work, you save substantial compute. For systems where they don’t, you get an unusable answer faster. Speed amplifies strengths — and it amplifies failure.

It’s important to note that Terray has also integrated an epistemic neural network (“epinet”) into their workflows to try to identify predictions where the uncertainty is highest. While this doesn’t solve the hallucination problem, it at least gives you another metric to consider when filtering predictions.

Would Terray’s reduced-dimensionality, reduced-descriptor approach be equally viable without their proprietary huge sets of experimental data? Hard to say at this moment.

The Open vs Closed Source Asymmetry in a Rapidly Changing Field

Boltz, Isomorphic, and Terray don’t just differ in methodology — they differ in epistemology. Let me acknowledge something up front: I work for QSimulate, a company developing proprietary software, so I understand the commercial incentives for keeping methods closed. That said, there’s a fundamental asymmetry in what open versus closed advances do for the field that’s worth examining carefully.

When an open-source project like Boltz lands, it doesn’t just move the needle; it raises the floor. The code is the documentation (triggering, but true). Every grad student, biotech startup, and rival pharma team can immediately pop the hood, see exactly how the diffusion noise schedule was tuned, and stand on those shoulders. Knowledge is shared instantly, and the effects are instant, cumulative, and democratic.

Contrast that with the closed-source “existence proofs” from companies like Isomorphic or Terray. These papers shatter our sense of complacency and suggest that the current open source models have fallen behind. But they don’t give us the blueprints. Instead of building on their work, the rest of the field is left reverse-engineering supplementary material and reading between the lines of vague methods sections. That’s not how science is supposed to accelerate.

That asymmetry doesn’t invalidate their advances — it simply slows collective validation.

In the End: Benchmarks Excite; Prospective Validation Decides

Boltz-2 opened the door. Isomorphic appears to have improved performance, at least on established benchmarks and specific kinase subsets. Terray suggests we may be able to simplify representation and dramatically accelerate ranking.

What we do not yet have is compelling evidence of broad, reliable improvement across truly hard, prospective targets outside familiar chemical neighborhoods. In the case of Boltz, it seemed everyone I know had tested the platform within a week of when it was announced — which almost immediately gave everyone the signal to understand strengths and weaknesses. That’s how science leaps forward.

Boltz operates in the open, allowing the field to test and calibrate its claims in real time. Closed efforts like Isomorphic and Terray may leapfrog performance, but their assertions must be taken at face value — at least initially.

Still…An Ecosystem Worth Defending — All of It

Let me be clear: I’m not against an ecosystem that combines open and closed source efforts. This is not about heroes and villains. It’s hard to imagine anything different, and the combination integrates both science for humanity’s sake and a legitimate profit motive. Historically, that tension is productive — it prevents complacency on either side. Recall that closing AlphaFold3 didn’t kill open science; it ignited it. More minds working on the problem moves the bar. All the players here are ultimately contributing to what the field becomes. They are all doing serious science, with the more pertinent question being how much reaches the public record and at what pace.

For practitioners, this translates directly: use open-source models as your default screening layer, stress-test with targets outside the training distribution before committing to a platform, and treat closed-source benchmark claims as directional signals — not validated performance guarantees. The Affinity Funnel isn’t just an architecture; it’s a decision framework for knowing when to trust which tool.

I remain genuinely excited — not because AI will eat physics’ lunch, but because the ecosystem suggested by the Affinity Funnel is taking shape in real time.

AI interpolates brilliantly; physics extrapolates when interpolation fails. One aggregates patterns; the other enforces ground truth. That’s not a rivalry — it’s a division of labor. The danger isn’t that AI will replace physics. The danger is mistaking interpolation for understanding.

Pass the Breakfast Test. Show me where it breaks. Then I’ll believe you.

Here’s a reality that may seem foreign in the world of 24/7 breathlessness on LinkedIn: Science does not march from breakthrough to breakthrough. It moves slowly, methodically, and only occasionally, with an undeniable game-changing breakthrough. The first iterations of AlphaFold were that breakthrough. Most of AI right now, the best of it, is moving things forward, but isn’t in one swoop changing the world. Understanding and validation are critical to the long game, and that’s why we need to treasure when someone takes that focus.

Especially in AI, where understanding is often a foreign land.

Look, I get why plug-and-play and check out our metrics work to garner attention. But I also know the long slog that methods like FEP took to get to where they actually made a difference to pharma. Not a constellation of stars. A constellation of yeoman’s work and careful thinking.

AI applied to the scoring of ligand binding is big right now, for obvious reasons. Binding triage at scale remains a huge hole in computational workflows. Methods like FEP are slow. AI methods have shown decidedly mixed results outside similarity to the training set. Throw a novel chemotype at these methods, and expect to be disappointed by the gap between marketing and reality. The gap isn’t a rounding error. It’s where a drug program can go to die.

That’s why the recent work from Apheris on fine-tuning the OpenFold3 affinity head for JAK2 macrocycles is useful. It is not a story about a model that worked out of the box. Instead it’s a story that’s more revealing, more useful. It’s the story about the engineering required to make it work reliably within a narrow chemical regime. By documenting how a weak baseline was systematically improved, the authors provide a practical snapshot of where these systems succeed — and where they still fail. That kind of honesty is rarer than it should be.

Let’s walk through what Apheris has demonstrated, and what that may imply about the real boundaries of AI in structural biology as we are forced to push systems novel to the training sets.

The Problem with Macrocyclic Baselines

The target here is Janus Kinase 2 (JAK2), a protein central to myeloproliferative neoplasms. Macrocycles — cyclic molecules with rings of eleven or more atoms — are a rational but difficult choice for this class of target. Their size and rigidity allow them to engage extended protein surfaces, but they also introduce conformational complexity that stresses most modeling pipelines. Show any comp chemist a macrocycle, tell them that’s the system they have to work on, and you’re met by the stare of mild dread that reflects decades of “best we can do right now.”

That stress is not a single missing feature. It is a compounding failure cascade: conformational flexibility, entropic penalties, ring-closure constraints, and sparse representation in training data — each manageable in isolation, collectively producing something the model has simply never seen before. And it shows.

The baseline AI model, evaluated on 69 JAK2 macrocycles, performed poorly. The initial Pearson correlation was 0.23. The Spearman rank correlation was 0.418. In a real discovery setting, those numbers are not useful for guiding design — they are closer to random ranking than actionable signal. If you handed those predictions to a medicinal chemist without disclosing the source, they would hand them back. I speak from experience.

The Two-Step Rescue

What makes this work interesting is not that performance improved, but how it improved — through staged interventions rather than a single “fine-tuning” step. The distinction matters because it tells you where the original failure actually lived. It provides protocol interpretability in a field that really needs it.

First, the team introduced architectural refinements to the affinity head, including revised cropping logic and multi-chain support. Spearman correlation improved from 0.418 to 0.60. This suggests the low starting point was not entirely a data problem — it was a representational mismatch. The model was conditioning on the wrong structural context. That requires a different fix than simply adding compounds.

The second step was targeted fine-tuning on a curated subset of 49 compounds, with 20 held out for validation. Pearson correlation moved from 0.23 to 0.76. More importantly, the model stopped behaving like a noisy regressor and began to meaningfully separate active from inactive compounds. A model that ranks actives above inactives is a tool. A model that doesn’t isn’t. In many programs, that’s a metric that gets the med chemists to take you seriously. That the team achieved this with 69 total compounds — a number reflective of a real program, not wishful thinking — makes it more striking, not less.

Training efficiency was notable. Using hardware-optimized attention kernels, the entire fine-tuning run completed on a single NVIDIA L40s GPU in roughly 45 minutes. That is a lightweight profile — but only because the trunk of the foundational model remained frozen; only the affinity head parameters were trained. The result is suggestive, not definitive: targeted adaptation can sometimes be achieved without large-scale compute, but one JAK2 macrocycle series does not generalize to broader chemical space on its own.

I want to sit with that for a minute. There’s a tendency today to ask, “Sure, you did that for one system, but you didn’t do it for 500 systems that sit on a validation website somewhere that a commercial company compiled and used to demonstrate their method was the best.” It’s a point taken, but I think it reflects an unproductive mindset of the modern era. Because comp resources (and increasing agentic agents) are as available as water, there’s a view that validation and characterization is an all-or-nothing sport. There are a lot of smart scientists who don’t have access to tens or hundreds of thousands of dollars in compute, but who have good ideas and a careful, sequential mindset. Their work can often produce ideation and revelations that bulk science doesn’t. So, yes, this is one macrocycle series. I hope these methods generalize. I await confirmation. But small is not the same as useless.

HTS Ingestion and the Probability-of-Binding Head

Beyond regression performance, the work introduces a probability-of-binding head — and this is the part most readers will skip past and shouldn’t.

Standard affinity heads assume a ligand binds and estimate potency conditional on binding. That assumption is appropriate for curated structural datasets. It is poorly matched to high-throughput screening data, where the dominant signal is often simply whether binding occurs at all. You cannot rank potency in a population that is mostly noise.

The Apheris rebuild restored a probability-of-binding output that the public SandboxAQ release had dropped. By explicitly modeling binding probability, the system becomes compatible with HTS-style datasets — noisy but abundant. This shifts the model’s role from precise ranking in a clean set toward early-stage triage: separating likely binders from non-binders before more expensive downstream work. That is where the real data bottleneck lives in most programs. It is good to see the architecture acknowledge it.

The AISB Network: Federated Learning in Practice

It is also worth situating this work within the AI Structural Biology (AISB) Network, a federated initiative involving organizations including AbbVie, Johnson & Johnson, Astex, Bristol Myers Squibb, and Takeda.

The core idea is separation of concerns: shared model infrastructure that improves collectively, while proprietary structural datasets remain local. Updates flow outward; raw data does not. In theory, this lets the community sharpen the tool without anyone surrendering the IP that finances the work.

I say “in theory” deliberately. Whether federated updates without shared raw data converge to the same model quality as centralized training is an open research question — and one that matters enormously if this network is going to function as advertised at scale. Without visibility into the aggregate data, you risk localized gradient biases or catastrophic forgetting, where optimizing for one partner’s chemotype quietly degrades performance on another’s. The architecture is promising. The empirical proof is still accumulating. I am a big fan of what federated learning offers behind the door. I just want the receipts.

The Remaining Thermodynamic Gap

Even with these improvements, the underlying limitation is physical rather than computational — and this is the part of the conversation that benchmark numbers tend to obscure.

AI-based affinity models are effective at rapid screening and prioritization. They approximate learned structure-activity relationships and interpolate well within known chemical regimes. What they do not do is compute binding free energy from physical first principles. They do not explicitly model the conformational entropy of a macrocycle transitioning between solution and bound states. They do not fully capture solvent reorganization or desolvation penalties in a physically explicit way.

To make this concrete: when a macrocycle binds its target, it pays an entropic cost for losing conformational freedom, and the surrounding water molecules reorganize around both the ligand and the protein surface. The magnitude of those effects — often in the range of several kilocalories per mole — can easily be the difference between a nanomolar binder and a micromolar one, or no binding at all. An AI model trained on activity data learns to recognize the patterns associated with those differences. It does not calculate them. The Boltzmann factor does not care how fast your GPU is.

For late-stage lead optimization — where the difference between 10 nM and 1 nM can be the difference between a drug and a failed program — physics-based methods such as Free Energy Perturbation remain among the most reliable tools available. AI models are increasingly powerful filters. Physics-based methods remain stronger adjudicators. Both belong in the toolkit. Neither replaces the other.

What the Apheris team did here is not glamorous. They found a model that didn’t work, figured out why in stages, fixed what they could fix, and published it straight. No triumphant benchmark. No claim that the problem is solved. Just: here is what we did, here is what it produced, here is what remains open. In a field that runs on press releases, that is a genuinely countercultural act.

The open-source structural biology community — researchers maintaining OpenFold3, contributing to shared benchmarks, publishing failure cases alongside successes — is doing some of the most intellectually honest work in applied AI right now. The Apheris team built on that foundation and gave something rigorous back to it. That’s how the long game actually works.

These systems are interpolative engines. They perform well when the query resembles the training set and degrade in ways that are not always visible when they don’t. Macrocycles are the stress test that exposes that boundary with particular clarity. The fine-tuning story here doesn’t erase the boundary. It maps it more honestly than most.

The physics doesn’t move. The models that earn trust are the ones that admit it.

OpenAI’s recent announcement for GPT-Rosalind — named after Rosalind Franklin — is either an homage or an irony, depending on how you feel about what follows. Franklin’s legacy was built on generating the hard, physical, indisputable X-ray diffraction data that made biological reasoning possible. GPT-Rosalind, conversely, is a “frontier reasoning model built to accelerate drug discovery, genomics analysis, and protein reasoning” — one that promises to reason its way to the finish line, without a complete picture of the underlying physical data that would make that reasoning valid. Amgen, Moderna, Thermo Fisher, and the Allen Institute are already on board.

I’ve been in this field for four decades. I’ve been to this rodeo. Combinatorial chemistry was going to change everything. Then solving the proteome. Enough time has passed that both now register as things people have “heard about.” Entire revolutions, complete with certainty and compressed timelines, have already come and gone. For many in the field today, they’re vague historical artifacts rather than lived experience. Biology, for its part, did what it always does. It refused to cooperate.

For a decade now, I’ve been hearing echoes of this exact claim — some from companies that didn’t survive the promise — that if we just embrace AI at scale, the long-standing complexity of the field will resolve into alchemy and cures.

Here’s what GPT-Rosalind — and every model before it — cannot fix: we don’t understand how the 20,000+ proteins in each human cell interact to create the most complex, most redundant, and most indecipherable machine the world has ever seen. The data aren’t there for the taking. They haven’t, in large part, been measured. AI cannot know what it has never seen, cannot predict what has never been suggested to it. And the literature it reasons over is what scientists thought they saw, not what was actually happening at the atomic level. No amount of reasoning over existing literature changes the fundamental shape of what we don’t know. This isn’t a reasoning problem. It’s a systems biology problem. We’ve observed perhaps 20% of protein–protein interactions in a human cell, much of it noisy or inferred. The rest hasn’t been measured. You can’t surface connections in a graph that hasn’t been built.

Drug discovery succeeds through rigor, intuition, and luck — not the luck of randomness, but the luck of the prepared mind encountering the right biology at the right moment, which is a different thing entirely and cannot be scheduled.

AI is a genuine complement to physics-based chemistry — I’ve argued this at length elsewhere, and I believe it. But complement is not replacement. It can’t close the last mile of affinity optimization, where the physics has to be right or the drug doesn’t bind. And affinity is the easy part. Wait until GPT-Rosalind meets ADMET — the absorption, distribution, metabolism, excretion, and toxicity profile that determines whether a molecule that works in a dish survives contact with an actual human body. That’s where most drugs die. That’s where the unmeasured biology lives. No interpolation engine fixes that.

Consider BenevolentAI. It was the industry’s first grand bet on mapping the “unseen” via a knowledge graph. It had its hero moment in 2020, identifying baricitinib for COVID-19 in ninety minutes. But repurposing an existing drug is just reasoning over a known map. When the platform moved into novel targets, it hit the systems biology wall — not because the AI failed to do what AI does, but because the map it was reasoning over was incomplete. The target was real. The redundancy of human immune pathways was not in the data.

It turns out that when the data isn’t there, AI just helps us optimize our ignorance.

We can do better, and probably will. But in ten years, people will still be getting sick and dying. AI or not. I hope somewhat fewer. I hope we can point to these models and nod at the value they contributed. I genuinely do.

The bet I’m making isn’t that AI fails. It’s that the disease stays harder than the announcement.

Prove me wrong. This is a bet I’d love to lose.

Subscribe now

In late March, the Ben & Jerry’s Foundation won the right to join a lawsuit against The Magnum Ice Cream Company — the newly independent Unilever spinoff that now controls Ben & Jerry’s, Breyers, Talenti, and the rest of the freezer aisle you grew up with. The founders aren’t fighting over flavor. They’re fighting to preserve the independent board that was the last institutional check on exactly the kind of ingredient optimization this article documents. The courtroom battle is about governance. The molecular battle has been going on for years. This piece is about the one you haven’t been watching.

The freezer case used to be one of the few places in a grocery store where you could trust the label. You bought ice cream, and you got ice cream. That era is over.

For years, if you suggested that the products we buy today are inferior to the versions from twenty or thirty years ago, you were dismissed with a “Hello, Boomer” or told that “memories are always better.” But as someone who values evidence, I’m here to tell you that this isn’t nostalgia. It’s a documented, technical retreat. Store-bought ice cream hasn’t just “changed”; it has been systematically reformulated by multinational food conglomerates using regulatory arbitrage and formulation science to determine how much cream they can remove from “ice cream” before we notice.

This is the story of the dairy frog, boiled by ever poorer ice creams.

The Regulatory Arbitrage of “Frozen Dairy Dessert”

In the United States, “Ice Cream” is a legally protected term. According to the FDA’s Standard of Identity (21 CFR 135.110), a product must meet two hard thresholds to earn that name: it must contain at least 10% dairy milkfat, and it must weigh at least 4.5 pounds per gallon. These rules were designed to prevent unscrupulous manufacturers from selling a carton of air and stabilizers under the guise of a dairy treat.

But walk down the aisle today and look at the label of a brand like Breyers. You will notice that many of their most popular flavors are now legally required to be labeled as “Frozen Dairy Dessert.” This isn’t a branding choice; it’s a confession. By dropping below that 10% milkfat threshold and pumping the product full of air, they have exited the legal definition of ice cream.

It’s not as if cheap barrel-bottom frozen desserts are a new thing. For decades, Sealtest was the brunt of “at least it’s not…” jokes — a giant national brand that occupied the very bottom shelf and knew it. The product was fine. It served its purpose. Perfect for feeding 20 hungry four-year-olds who wouldn’t know the difference. But few adults were fooled, and the brand never pretended otherwise. There was a kind of integrity in that. They held a specific spot in the ecosystem and carried that mantle honestly.

Unilever, which acquired Breyers in 1993, is the architect of this decline. They took a brand that was founded on a “Pledge of Purity” — a 150-year-old promise signed by Henry Breyer that the product would never contain “adulterants, gums, gelatins, powders or fillers” — and effectively shredded it. The original recipe was four ingredients: milk, cream, sugar, and vanilla. Today, a tub of Breyers Extra Creamy Vanilla contains twelve ingredients, including corn syrup, whey, and three different thickening agents.

Side-by-Side Ingredient Analysis: Breyers Vanilla

Cream being pushed to third place is the whole story in one row: the fat that once defined the product has been demoted beneath sugar (with whole milk replaced by skim milk, to boot). It’s worth noting that Breyers actually sells three vanilla products simultaneously. The Extra Creamy, as detailed above, can’t legally call itself ice cream. “French Vanilla” manages to clear the “ice cream” bar (while possibly offending the French) with 13 ingredients, including corn syrup and three gums. And “Natural Vanilla” is genuinely close to the original — though it quietly adds skim milk, tara gum, and swaps real vanilla for “natural flavors.” That last one is the cleanest thing in the entire Breyers line, which tells you everything you need to know about where the brand has landed.

The Physics of Profit: Air as an Ingredient

The most profitable ingredient in a mass-market frozen dessert is air. In the industry, this is called “overrun.” Overrun is the volume of air incorporated into the mix during churning, expressed as a percentage:

A 100% overrun means that for every gallon of dairy base, you produce two gallons of “product.” While premium brands like Häagen-Dazs operate at roughly 25–30% overrun, many mass-market brands approach the 100% ceiling — doubling volume with air and, in some cases, requiring “Frozen Dairy Dessert” labeling.

Corporate marketing teams have become remarkably adept at selling this air back to us as a feature. The “Slow Churned” or “Double Churned” revolution, pioneered by Dreyer’s (Edy’s) and quickly adopted by Unilever’s Breyers, is a masterclass in this deception. These processes use lower temperatures and higher pressures to break fat globules into smaller sizes, allowing the product to mimic the mouthfeel of full-fat ice cream while containing significantly more air and less actual cream. They tell you it’s “smoother” or “easier to scoop.” What they don’t tell you is that they are charging you premium prices for a stabilized foam that deflates if left on the counter.

Subscribe now

Lousier Living Through Chemistry

To be clear, stabilizers and overrun aren’t inherently problematic. At appropriate levels, they improve texture, slow ice crystal growth, and create a smoother product. You’ll even find small amounts of them in some of the best high-butterfat premium brands. The issue isn’t their existence — it’s the degree to which they’ve become substitutes for dairy rather than complements to it.

As you remove the fat (which provides structure and flavor) and add air (which provides nothing), the resulting mixture becomes physically unstable. It would normally turn into a grainy, icy mess the moment it left the factory. To solve this, food scientists have turned to a cocktail of hydrocolloids — guar gum, carob bean gum, carrageenan, and the now-ubiquitous tara gum.

These gums are “magic dust” for the bottom line. They manage water mobility and increase the viscosity of the mix, creating a “chewy” or “gummy” texture that mimics the richness of cream without the cost of actual dairy. If you’ve ever noticed that your ice cream now feels slightly “rubbery” or doesn’t actually melt into a liquid, but rather stays in a foamy, gelatinous blob, you are experiencing the effects of over-stabilization.

These stabilizers (“hydrocolloids,” if you want to impress your friends) work by trapping water in microscopic networks that prevent ice crystal growth. At low levels, this keeps ice cream smooth; at high levels, those same networks become elastic, giving the product that uncanny, almost bouncy texture. When someone says that ice cream seems too “chewy,” this is what they’re talking about. Turning a smooth, soothing spoon of ice cream into something you have to chew is not innovation.

This isn’t just about Breyers. Since Unilever acquired Talenti in 2014, long‑time fans have noticed a marked decline. The first ingredient in their Salted Caramel Truffle gelato is no longer “caramel,” and other flavors have quietly picked up extra help in the form of coconut oil, sunflower oil, dextrose, and gums. Today, there are plenty of public complaints about a base that feels increasingly “gummy” and “bland,” padded out with cookie chunks and engineered air pockets to fill the jar visually without filling it substantively. It’s the food industry equivalent of shipping a half‑empty bag of chips filled out with nitrogen. (Wait, they do that too? Yeah, they do.)

Though Unilever’s trudge toward profits over enjoyment has been relentless, it would be unfair to suggest they are alone. The majority of sub‑super‑premium ice cream brands you’ll find on the shelves of your local store are filled with gums and subject to significant overrun — and, frequently (but not always), not enough butterfat to legally call themselves ice cream. That includes Blue Bunny, Friendly’s, Dreyer’s/Edy’s, Turkey Hill, Hood, and most of the other familiar names filling the mid‑shelf real estate. The easiest field test? Flip the carton over and check the calories per serving (2/3 cup). If it’s below ~220, it’s very likely the product is leaning heavily on air and stabilizers rather than butterfat. The label doesn’t lie — but it also doesn’t volunteer anything.

Rate Ice Cream Like a Boss

Note that some variants can bump up the calories without improving quality, by loading in high‑calorie inclusions like chocolate or candy pieces. But this still gives a pretty good starting guide.

And this race to the bottom repeats across nearly every brand Unilever touched, which brings us to Ben & Jerry’s.

Even Ben & Jerry’s, despite its independent board, has faced criticism for thinning out mix‑ins and moving toward more stabilized formulations under Unilever’s “Path to Growth” strategy — a phrase that should be understood to mean “path to margin.” Gums now run rampant in the Ben & Jerry’s line, as does corn syrup (mostly from the mix‑ins). Fats have transitioned to soybean and coconut oils from butter in some varieties. In 2010, they removed the “All Natural” label from their ice cream because… well, you do the math.

The Beacons of Integrity

Just because the midline has been decimated doesn’t mean “good science” has vanished entirely. There are still brands where the “can” and the “should” remain in alignment.

Häagen-Dazs remains the gold standard for grocery store minimalism. While many of their specialty flavors have started to include limited stabilizers, their core pillars — Vanilla, Chocolate, Strawberry — rely on a high milkfat content (up to 18%) and an extremely low overrun (approx. 25–27%) to deliver richness. They don’t use gums in these core flavors because they don’t have to; the density and fat provide the stability.

And then there is Graeter’s. If Häagen‑Dazs is a triumph of minimalism, Graeter’s is a triumph of artisanal engineering. They still use the French Pot process in 2.5‑gallon batches, built on a French‑style egg‑custard base — cream, milk, cane sugar, eggs — whose natural lecithin provides the emulsification that makes a gum cocktail unnecessary, though a few flavors use small amounts at a fraction of what props up a Frozen Dairy Dessert. Their overrun is naturally capped around 20–25%, producing an ice cream so dense it has to be hand‑packed because automated fillers can’t move it. Some tubs hit 380 calories per ⅔ cup — not just passing the dairy field test but obliterating it. This is what happens when you treat product quality as a hard constraint and the supply chain as the adjustable variable. Unfortunately, in early 2025, Kroger pulled Graeter’s from all regions outside the Midwest — a predictable outcome for a product that refuses to cut corners. The brand doing it right is becoming harder to find. If you needed one data point to explain everything wrong with the freezer aisle, that would be it.

A Quick Cheat Sheet: What’s in Your Freezer?

Another Path to Profits: Shrinkflation

While their competitors raced to the bottom, Häagen-Dazs found another way to improve margins without diminishing the product: shrinkflation. In 2009, they reduced the 16oz pint to a 14oz “pint” — prompting the expected (and deserved) howls from what amounted to an overnight 12.5% effective price increase. (Also a bad lesson for any child trying to understand the meaning of “pint.”) But by doing so, they found a road to profits that didn’t crapify the product itself. The market has (mostly) forgiven the shrinkage and continues to reward the quality they’ve maintained.

For a long time, Ben & Jerry’s made a point of staying at a full 16oz and even took public shots at brands that quietly shaved their “pints” to 14oz. More recently, though, they’ve found their own way to play the size game. The new Sundaes and Topped lines come in packages that look like standard pints but often clock in well under 16oz — for example, a “Turtle Sundae” at 14.44 fl oz or a Topped Dirt Cake at 15.2 oz, with a whipped‑topping layer and extra mix‑ins on top. You still pay pint prices, but more of that money is now going to air, whip, and candy sitting on the surface rather than to actual ice cream underneath. Pay attention to the Topped Tax, and shop carefully.

Talenti has taken a similar approach with their Layers line — alternating bands of gelato, sauces, and inclusions stacked inside a standard pint container. The result looks generous and photogenic, but the actual gelato component is substantially less than what fills a traditional pint. Same form factor, same price point, less of what you came for. Great margin increasers think alike.

Ironically, most mid-tier brands underwent shrinkflation too — moving from 2 quarts to 1.75, and then to 1.5 — while simultaneously dumbing down what went inside. A classic pincer movement, taking down the hapless consumer from both sides. At least there’s now less of it not to love.

The Profit of Erosion

It is tempting to look at the “encrapification” of ice cream as an inevitable result of inflation or shifting consumer health trends. But the data doesn’t support that. Dairy prices have fluctuated, but the pivot to “Frozen Dairy Dessert” wasn’t a survival move — it was a margin-expansion move. It was a strategic decision to trade brand equity for short-term earnings.

Food scientists have known for decades that consumers can’t reliably detect fat reductions of 2–3 percentage points in blind tests — a finding confirmed in modern controlled trials. Overrun, meanwhile, can be increased without any change in declared ingredients and no regulatory trigger until the product falls below the 4.5 lbs/gallon threshold. Once companies understood these thresholds, they pushed them — slowly, incrementally, and profitably.

The more uncomfortable question isn’t what they did — it’s why it worked. Food scientists didn’t exploit some exotic vulnerability; they exploited convenience, habit, and the fact that most of us stopped reading labels the moment we trusted a brand. We let it happen because trusting the label was easier than interrogating it. Corporations noticed. They always do.

What we’re seeing now isn’t accidental decline. It’s late-stage optimization: brands engineered to the edge of consumer detection, tuned for cost efficiency rather than experience. Much of that decision-making power has consolidated into a handful of global conglomerates. In 2025, Unilever finished spinning off all its ice cream brands into a standalone company, The Magnum Ice Cream Company. Good luck to them.

As a computational scientist, I recognize this as a constrained optimization problem — one in which the constraints were gradually redefined to favor producer economics over consumer experience. Every manufacturer optimizes for margin; the distinction lies in which variables are treated as fixed and which are allowed to move. Brands like Häagen-Dazs appear to treat ingredient integrity and sensory appeal as hard constraints. Others have treated those variables as adjustable, relying on incremental changes that remain below the threshold of consumer detection. One approach preserves product identity and adjusts price transparently. The other is optimization by dilution.

The next time you’re in the freezer aisle, ignore the “Slow Churned” slogans and the “Smoother Texture” promises. Turn the carton over. If it doesn’t say “Ice Cream,” or if the ingredient list looks like a chemistry final, put it back. We shouldn’t reward corporations for the sophisticated debasement of the things we love.

The data doesn’t care about nostalgia. And the data says that Breyers used to be better. Talenti used to be better. And “frozen dairy dessert” is just a polite way of saying they’ve figured out how to sell you a bubble for five dollars.

Ice cream is just the canary — this same optimization-to-failure pattern is unfolding across the entire packaged food system, and the only way to stop it is to start noticing and stop buying.

Just because you’re a boomer doesn’t mean you’re wrong. It just means you’re tired of being slowly boiled.

I published a piece about ice cream three weeks ago. I don’t usually write about ice cream — my beat is drug discovery, computational chemistry, AI in pharma, the plumbing of how medicines get made. But ice cream bypasses the adult brain entirely and goes straight to the part of you that is still eight years old, standing in front of a freezer case with two dollars and a very important decision to make.

I wrote it because I love ice cream — genuinely, specifically, with the kind of opinions that make me bad company at a Dairy Queen. But forty years of computational chemistry does something to the way you think. You can’t make a claim you haven’t earned. You can’t stay comfortable in a good story if the evidence doesn’t hold. So what started as affection turned into research, and the research turned into something I could actually defend. That’s not discipline. That’s just how the training rewires you.

I figured if anything I’d written was going to find a broad audience, it was this. I was right, and it didn’t matter. Here’s what I mean.

———

The piece ran on Medium. I promoted it on LinkedIn, where it generated over 100,000 impressions. About 2,000 people clicked through. Of those, 1,500 read it — not opened, not skimmed, but read it, by Medium’s own thirty-second minimum definition. On a twelve-minute piece. That’s a 73% read ratio, which in the attention economy is roughly equivalent to throwing a party and having everyone stay through the dishes.

The engagement signals were everything platforms say they’re looking for. Sharp spike at publication. High completion rate. Organic spread — including a Reddit thread in a community that found the piece on its own and decided it was worth sharing. Medium’s paying subscribers, when they encountered it, read it at 90%. Nobody bounced.

Medium’s algorithm looked at all of that and said: Interesting. Then it distributed the piece to approximately nobody. Internal traffic: 4% of total. New followers: two.

I sat with that for a while. The disappointment that I succeeded and Medium shrugged pushed me to take a realer look at what Medium actually rewards.

———

The thing is, Medium wasn’t wrong.

My instinct was the familiar writer’s grievance — good work ignored, algorithm broken, platform indifferent. All of that felt true. None of it was the real story.

Look at where those 2,000 readers actually came from. The largest bucket was direct — the dark matter of the internet, links that travel by email and Slack and WhatsApp, person to person, because someone read something and immediately thought of someone else. After that, LinkedIn. Then Reddit, organically. Then a music forum (!) found it.

These were not Medium readers. They were my readers, temporarily on Medium’s platform, reading about ice cream, and then leaving.

From Medium’s perspective, I’d borrowed their infrastructure to throw a party for guests who had no interest in joining their subscription service. The readers were real. The engagement was real. The conversion to Medium’s business model was essentially zero.

Medium optimizes for one thing above all others: paid conversions per impression. A piece that drives high engagement from an outside audience isn’t a success case in that framework. It’s a leakage point. The algorithm didn’t malfunction. It made a perfectly rational decision to let my readers come and go without amplifying the pattern.

The algorithm didn’t fail. It declined.

———

What I kept coming back to was the 1,498.

1,500 people finished that piece. Two followed me. I don’t blame the other 1,498 — the platform gave them no particular reason to stay, and the conversation, the relationships, the context that would have made staying worthwhile, none of that existed there. They read what they came for and left. Entirely reasonable. Completely my problem.

Because here’s what I actually lost: not reach, not impressions, not the algorithmic amplification I didn’t get. What I lost was the relationship. Those readers liked the piece. Some of them liked it enough to send it to someone. A community I’ve never interacted with thought it was worth posting. And I have no way to find any of them again, because I don’t own the list. Medium does. And Medium has no incentive to connect us.

That’s the real cost of building on someone else’s platform. Not the algorithm. The ownership structure. When your readers show up, they’re guests in someone else’s house. The host decides whether to introduce you.

———

So I did something I should have done two years ago.

I launched A Half Life here on Substack — two tracks, published twice a week. The Basis Set is the analytical work: drug discovery, AI, the science of how medicines actually get made, what the industry gets right and what it gets catastrophically wrong. Sometimes how I got into and fit into all of it. A View from a Pill is something stranger: first-person essays narrated by the molecules themselves — a drug compound moving through a clinical trial, a failed candidate sitting in a freezer, a generic watching its patent expire. The same instinct that turned ice cream into a research question, turned inside out. Forty years of computational chemistry, metabolized into something that doesn’t read like a technical paper.

The ice cream piece was the most-read, most-shared thing I’ve published. It also did the least to build anything durable. Those 1,500 reads were real. But I’m done leaving the relationship on the table.

If you found your way here from the ice cream piece, or from LinkedIn, or from that Reddit thread: this is where the work lives now. The list is mine. The archive is mine. When you subscribe, we don’t lose each other.

A Half Life is free to subscribe.

Follow A Half Life

When I started commenting about science, someone asked me, “Uh, who are you, and why should I care?” Pretty blunt. But fair questions.

So, acting against every instinct, here’s a bit of my back story. I don’t expect everyone will care. Heck, I’m not sure anyone will. So if you don’t want to read the whole thing, here’s the summary: I’m a computational chemist and when I started in the field, you’d stare out the window watching the dinosaurs grazing while you munched on an apple and used a mallet to tap your code into stone tablets. Now those tablets are made by Apple and those who don’t use computational chemistry are called dinosaurs.

And now, on with the rest of the story…

When I was 13 years old, my family would gather around the television on Saturday nights for must-see TV, which included All in the Family, M*A*S*H, The Bob Newhart Show, The Mary Tyler Moore Show, and Carol Burnett. All on CBS. It’s hard to believe now, but in those good old monocultural, pre-time-shifting, pre-on-demand days, this is what families did. These were all great shows, but I single out The Mary Tyler Moore Show here for one of its characters, pompous news anchorman Ted Baxter, who was prone to reciting his background story — always starting with the same line, “It all started in a little 5000-watt radio station.” This is my 5000-watt story. Though, unlike Ted, I promise I won’t do this twice.

If you had asked me at 13 years old what I wanted to be when I grew up, I’d have told you I wanted to work in advertising. From the outside, it looked creative and fun — and who doesn’t want a career where you get to be creative, have fun, and get paid for the privilege? Advertising looked so alluring to me that I had my dad bring home copies of Advertising Age from work, which I’d read voraciously. Alas, as a family friend explained to me a couple of years later, it was a soul-sucking enterprise that would destroy you with politics and back-biting. So, scratch that — so long Don Draper dreams. Ultimately, I would find another career that allowed one to be creative, have fun, and get paid for the privilege. In that way, I suppose the transition to what was ultimately chemistry wasn’t as unpredictable as it might seem.

When I eventually went to college, I didn’t really know what I wanted to be. I wasn’t even sure I wanted to be in the sciences, which is why I turned down Caltech and instead went to the more well-rounded Stanford. (A slight that Caltech found so unfathomable that later they actually contacted me, in person, to ask how I could possibly have made that decision!) While I arrived at Stanford a relatively clean slate, I do recall stating on the application that I was interested in being a patent lawyer, with the logic being that it would combine writing, debate, and science. I guess it felt like a good idea at the time. But I quickly changed my mind, or perhaps better put, I dodged a bullet. Still, I did follow through with at least part of that equation: I majored in chemistry.

The Seeley G Mudd chemistry building at Stanford was my frequent home while there. This picture was taken the year I started college, when the building was brand new. It was torn down as “unable to support modern science” in 2022. Things that make you feel old.

During my final year of undergraduate studies in Chemistry at Stanford, I fell for a woman who was staying on for an extra year to do a master’s thesis. Love being what it is, I narrowed my search for graduate school to the Bay Area — which meant Berkeley (since Stanford didn’t take their own undergrads). I got accepted and started my PhD studies that fall. When I arrived, I discovered that they expected you to choose an advisor in a ridiculously short period — a matter of weeks. Imagine making a choice that affects your entire career as a munchkin fresh out of college. Ultimately, I chose to work for Professor Sung-Hou Kim, a young crystallographer famous for having solved the first 3D crystal structure of tRNA while in Alex Rich’s lab. He seemed like a nice guy, closer to my age than to retirement, and crystallography seemed cool. Those seemed like good enough reasons.

Shortly after choosing Sung-Hou, he came to my desk and told me, “Come walk with me.” That walk consisted of about 30 seconds to the elevator and another 30 seconds as the elevator took us to the second floor. During that minute, Sung-Hou posed a question to me, “Do you see yourself as an experimentalist, or as a theory person?” I had not considered this question before that instant, but I mulled it over for a few seconds and replied, “Theory sounds interesting.” “I figured,” he said. And that was that. Like the butterfly wing flap that changes everything, my professional course was set on the basis of something I’d literally thought about for fewer than 10 seconds!

At first, I wasn’t sure my answer meant anything at all. Everyone else in the lab was experimentally focused, and as a first-year student, I spent most of my time taking classes and working as a teaching assistant. Occasionally, I’d come into Sung Hou’s lab and sit at the desk assigned to me, but being in his lab wasn’t particularly significant to my day-to-day, outside of attending the occasional group meeting — all of which focused on experimental projects.

Then, one day, a few weeks after I gave that fateful answer, I came into the lab to find a 9-track tape on my desk. At that time (just barely past the punch card era) these 12” diameter reel-to-reel tapes were the preferred format to store/transfer computer data. They could hold a whopping 170 MB (probably less than 1/1000th the storage on the phone in your pocket)! The tape was enclosed in a case, and on that case was a piece of masking tape with the word “AMBER” written on it. I had no idea what this was. I’d never heard of “AMBER,” nor had anyone else in my office. Finally, a day or two later, I ran into Sung-Hou and asked him about it.

A 9-track computer tape like this changed my world forever.

“Oh, I was at a meeting in San Francisco, and a professor there gave a talk about some interesting calculations they were performing on molecules. I asked if we could get a copy of the program they use for that. They call the program AMBER.” As it turned out, what I had was the first copy of the AMBER program to ever leave the UCSF confines of Peter Kollman’s lab. If you are a student of computational chemistry, you’ll realize this momentous event is akin to the first copy of the Gutenberg Bible making it out of Mainz. Or, well, something like that. But to me, at the time, it was a moment when I thought, “Oh, OK. I guess I’m gonna have to get serious about learning Fortran.”

Fortran? Yes, indeed. Kids today chuckle about it, but it was the computer programming language of kings back in the day. When I arrived at Berkeley, my programming skills were fairly non-existent. I’d done some programming in Basic back in high school and had started a programming class with my girlfriend in my final year of college — but I’d dropped it when she decided it wasn’t for her. Honestly, it was for me, but I had too many other things to do that quarter if she wasn’t going to be hanging out with me in Stanford’s computer science dungeon, “LOTS” (this was before anyone had a computer in their room, let alone a laptop). So, I taught myself Fortran by reading the Digital Equipment Company (DEC) Fortran manual cover-to-cover. Yeah, that’s how we did it back then.

Anyway, I taught myself to be a sufficient Fortran programmer, and then, with extensive daily experience — starting with learning AMBER — a pretty good one. Back to AMBER. When I spooled up that tape and read the contents, what did I find? Some Fortran code and an extraordinarily sparse manual that amounted to “put this control variable in this column.” Not much about why you’d choose a particular value for the variable, or even what some of these values meant. That was left as an exercise for the reader. Thank you so very much.

Oh, and did I mention that the code, as sent, didn’t actually compile when I initially tried? So much for quality control. :-) That said, when your installed customer base is one person, I suppose you can be forgiven.

So there I was, trying to figure out how to get this ball of code to compile and then trying to figure out what the inputs really meant. This was at a time when absolutely no one in my research group, nor a single other individual on the Berkeley campus, had a clue about this stuff. It was truly a trial by fire. But in the end, it made me into a decent coder and threw me into a small group of individuals in the entire world who could run computational chemistry simulations of things like proteins and nucleic acids.

This was truly the Mesozoic age of macromolecular computational chemistry. We were running on a mini “supercomputer,” a VAX 11/780, a $2 million dollar computer (today’s dollars) that offered 1.0 MFLOP performance and 64MB of memory, to be shared by the 20–30 people who would be interacting with the computer at any one time. When writing code, you’d sweat over whether you could reduce 4-bit integers to 2-bit to save memory, or whether you could temporarily write stuff to disk to clear an array and use that precious memory for something else. To drive the point home, the performance of the iPhone 16 is over two TFLOPs, or 2,000,000 times as performant, with 125 times the memory! And it takes pretty good photos and makes phone calls, which we couldn’t do with the VAX.

I spent my graduate career running calculations on nucleic acids for the most part — proteins wouldn’t come until later. My first accepted publication was a full research article in Science magazine, which thought our predictions of the effects of photodamage on DNA were sufficiently novel that not only did we get a full article (rather than a shorter report), but until the last minute we were in contention for the cover of the magazine. Sung-Hou had suggested we submit it to Science, and that paper flew through the review process. “This is easy,” I thought. I remember subsequently complaining to Steve Holbrook, a senior researcher in my group, “Why are we bothering to submit to Science? That’s such a mainstream journal and obviously easy to get into. Why didn’t we submit to a better, more focused journal like Biopolymers?” Kids say the darndest things sometimes. :-)

As I neared finishing my degree at Berkeley, with the first PhD that focused on what could be called modern computational chemistry, I wasn’t certain what the next step would be. I remember interviewing on campus for jobs — because just about everyone interviewed with the companies that bothered to make an effort. While most people I know had interviews that consisted of spending an hour trying to convince the interviewers that they were the greatest thing since the invention of the Erlenmeyer Flask, my interviews went differently. I spent an hour explaining to my hosts what, exactly, I did. Computational chemistry was that new and foreign at that point. At the end of one interview with a major company, my host condescendingly said, “Well, that sounds really exciting, and I am sure you’ll find a use for that somewhere.” So it went.

A rare early mainstream mention of computational chemistry filled the cover of a 1981 issue of Fortune magazine. It would be another decade before Vertex Pharmaceuticals moved comp chem to the main stage, and another decade after that before comp chem broadly became a mainstay of drug discovery.

I wasn’t sure whether I wanted to go into academics at that point (which would have required a post-doc) or into industry (which would have required a miracle). But my girlfriend was moving to Southern California, and, by happenstance, I found a want ad in the back of Science magazine advertising a new company that would be developing and selling software for computational chemistry. And they were in Southern California — San Diego. Wow. Kismet! I sent them my resume via USPS mail — no email outside of Darpanet at that point — and a short time later, I got invited down to interview. The company was Biosym, which had just moments earlier been formed to commercialize the molecular mechanics software platform developed in Shneior Lifson’s lab. Many of the titans of the early generation of modern comp chem had passed through Lifson’s lab, including later Nobel laureates Warshel, Levitt, and Karplus, and most of the macromolecular packages of the time can be traced directly or indirectly back to work started in Lifson’s lab. Arnie Hagler, co-founder of Biosym, had also been in Lifson’s lab and arranged to use the code as the basis for the company. I came on board as employee number five and quickly became the product manager for the commercialized Lifson code, soon renamed “Discover.”

A 1987 photo of the Naiman Tech Center, near the intersection of the 805 and 5 freeways in San Diego, where Biosym was located when I joined. Many knew it from the large red rocket-like sculpture (by Alexander Liberman) that could easily be seen from the freeway. Alas, Jack Naiman went bankrupt a few years later, and the center was ultimately renamed the San Diego Tech Center. The red sculpture is still there.

Those early days of trying to convince pharmaceutical companies they might want this new-fangled “Discover” program (along with, in short order, the “Insight” visualizer that we initially obtained from Henry Dayringer at Monsanto) were interesting. Over time, things went from “What is this?” to “Does it work?” to “We’d like to try it” to “We’re using it” to “How did anyone ever do this without it?” And, eventually, an entire industry built up walking the paths that we helped clear using Discover. (To be fair, Garland Marshall’s Tripos and Keith Davies’ Chemical Design also cleared some of that early brush, though with different focuses). Biosym managed to survive and thrive through acquisitions and mergers, eventually becoming Accelrys and then (after being purchased by Dassault) Biovia. My finger remained in that pie long after I left via some software code, the original manuals, and the fact that I was the one who suggested the name “Discover” during a “board meeting”/pizza party one night. I guess the latter remains part of the DNA.

I stayed at Biosym for about 1.5 years before I started to think there might be more scientific pastures to forge for me. So, I contacted Peter Kollman at UCSF, whose AMBER program had been my entrée into the computational chemistry space. Peter and I had forged a friendship over the years I’d been a grad student, meeting up and talking at length at conferences. Back then, any comp chem conference was sure to attract a good portion of all the practitioners in the world because the field was that small. Peter Kollman was well known as the friendliest and most encouraging PI in the field, the kind of senior professor who had time to spend talking at length to a graduate student like me. I asked Peter if he might have space for me if I came up to UCSF. He assured me he would, but added, “Now, you know, you’re gonna have to take a big salary cut if you do…” I let him know it wouldn’t be a problem. The long game was always what mattered to me. Still does. Besides, it’s not like Biosym was paying me a king’s ransom. (Today, a full-time McDonald’s worker in California can make more).

The UCSF Parnassus campus, where the Medicinal Chemistry department (including Peter Kollnan’s group) was located when I was there. We were on the 9th and 10th floors of the tall building to the right in the middle of the photo. I was on the 9th floor when the big San Francisco earthquake of 1989 hit. To sense the floor beneath your feet feel is liquifying is an experience one hopes not to repeat.

I arrived at UCSF on January 2, 1987. Peter’s lab was in a period of transition. Paul Weiner, a former Karplus student who’d joined Peter’s group and written the first version of AMBER, had left a few years earlier. U. Chandra Singh, who’d taken over primary authorship of AMBER, improving molecular dynamics and adding the facility to run a new method called “Free Energy Perturbation” (FEP) had also recently left the lab. Peter had just published a couple of the highest-profile FEP papers up to that point (and probably for the next 30 years) in Science and elsewhere. The free energy work was most appealing to me. I’d become aware of the approach a bit earlier, when a co-worker at Biosym, Al Cross, brought up a publication by Bill Jorgensen at journal club. I loved the idea of a method that combined molecular dynamics (MD) and statistical mechanics to generate something of specific and verifiable interest. Up until that point, the limited number of MD applications had basically amounted to, “Look, ma, proteins jiggle around over time!” Suddenly, we had a much more concrete use for MD.

I spent the next four years (as a post-doc and then staff scientist at UCSF) improving and characterizing methods for FEP (and the closely related method Thermodynamic Integration, TI). I also spearheaded adding facilities for NMR refinement into the AMBER suite, along with some other stuff. It was an exciting time for computational chemistry, and unlike grad school, where I felt like a lone wolf on the mountain, at UCSF, I was surrounded by a growing group of individuals focused on this type of work. This is often referred to as the golden age of comp chem, and particularly so at UCSF, with so many professors who would impact the field all working in the same department at a crucial point in the field’s development. In addition to Peter Kollman, the medicinal chemistry department at UCSF hosted Tack Kuntz, who was publishing the first papers on molecular docking, Bob Langridge, who was laying the foundations for modern molecular graphics with Midas (later Chimera), and Ken Dill, who was developing his widely-influential lattice theories of protein folding. These labs were close — we often held joint group meetings and happy hours — and it felt like an extended family. We didn’t know how great we had it then, but looking back, I’m in awe.

I loved my years at UCSF, but at some point, you realize that you’re perhaps getting a bit too comfortable, and so, around the fourth year, I started looking for new opportunities. I came close to going to an excellent large Midwestern university, but they were taking way too long to come to a decision (departmental politics — what’s new?). In the interim, I got contacted by a brand new startup, Vertex Pharmaceuticals. Their spiel was that they were going to revolutionize drug discovery by giving “physics” (read: computational chemistry) an equal seat at the discovery table. Now, to be fair, by the time Vertex was founded (1990), there were a few major pharma companies with established and growing comp chem groups. But these were large companies like Merck, and the influence of comp chem within those companies was often modest. Sadly, chemists at that point still viewed comp chem with great skepticism, and many applied the “Rule of One” to comp chem: “Tell me one thing to make, and if I make it and you’re wrong, then I will have confirmed my hunch that you provide no value.”

Joshua Boger, who founded Vertex, had come from Merck, believed in comp chem, and saw things differently. I was still waiting to hear about the academic job, but I’d started to grow weary from the wait, and I agreed to fly out to Cambridge, MA to meet with Vertex (who — especially Mark Murcko — had been putting the hard press on me to visit). The interview went well, and then they made me an offer I ultimately found impossible to resist. Basically, they wanted me to be a “scientist in residence,” to continue to develop and publish on AMBER, free energy, and other computational chemistry methods. Go to conferences. Help put Vertex on the map in the comp chem field, to prove they were serious about this stuff. Oh, and I wouldn’t have to write grants. It was a tough choice (I would have to move to the East Coast from California, and was I willing to put academic pursuits on hold?). But, ultimately, I took their offer. I was absolutely sure that within five years, I’d be back on the West Coast. 34 years later, I’m not. Life is what happens while you’re busy making plans.

At Vertex, the position was exactly as promised. I continued to make substantial contributions to AMBER and the AMBER community for many years, especially to the free energy and NMR-refinement modules. I was the first author on the AMBER platform for that decade. I also expanded my focus to include new approaches for de novo molecular design, rapid scoring functions, and other things that seemed cool and potentially useful. I managed to get chemists within Vertex to make molecules based on these calculations, even though I often failed miserably by the Rule of One. The chemists were young and smart, and the mandate to include (and listen to) a computational chemist on every project was followed without exception. Vertex delivered on their pitch deck promise, and many discovery decisions reflected computational input.

The centerfold from the very first Vertex annual report in 1992. Yours truly is on the right. At the time, this is how some still imagined computational chemistry worked. As if.

During the more than a decade that I was at Vertex, computational chemistry moved from a curiosity to a bedrock practice. This was in large part due to the rapid adoption of virtual screening methods based on molecular docking that took hold in the pharma industry starting in the late 1990s. The beauty of molecular docking is that it moves computational chemistry from the onesie/twosie domain of many approaches to a numbers game. It allows scientists to reduce hundreds of thousands (now billions) of purchasable/synthesizable compounds to a manageable number that can be obtained and tested. If you obtain and test 500 compounds and you find one winner, then you, as the computational chemist running the virtual screening campaign, look like a star. It doesn’t matter if 499/500 of those predictions were wrong. One needle in that reduced haystack makes you a winner. With that, the perceived value of computational chemistry in the pharmaceutical discovery field would never be the same. The battle for relevance and reliability that we’d started with Biosym had been won.

I left Vertex in the early ‘00s — by then an established player with a drug on the market, and so the need for a scientist in residence to prove they were for real was not so imperative anymore. After short stints consulting (read: part-time work while staying at home to help raise my young daughter) and some time at The Broad Institute (where computational chemistry had not yet been Broad-ly adopted), I took a position at Schrödinger. I’d been offered two options: lead efforts for FEP calculations or develop a new computational platform for biologics. Given that I’d spent more than two decades at that point developing and characterizing free energy methods such as FEP, I picked the biologics option — something new!

At Schrödinger, I became the product manager for the platform that became known as BioLuminate. When I joined, there were several widely used commercial software platforms for small molecule computational discovery (CCG, Schrödinger, and OpenEye, among others, had entered the domain after the successes with Biosym’s Discover that I’d helped spearhead in the early days). But there were no comparable computational platforms for biologics discovery — even though biologics were by then closing in on 50% of the revenue market for pharmaceuticals. To me the idea of bringing a new platform to market to address a new audience brimmed with a bit of the same excitement I got when we were trying to convince the world that computational chemistry had value back in the Biosym days.

Working the BioLuminate booth at one of the endless parade of biologics conferences I attended at Schrödinger. We worked tirelessly to get the biologics community to embrace the value of computational chemistry for discovery. It took a while, but we succeeded.

There were no templates for what BioLuminate should look like, so we needed to figure it out by talking to people in the trenches, formulating, putting our new ideas in front of those people, and cycling many times. Eventually, we got it to where we thought it was right and developed the actual platform. It exceeded sales expectations in the first year it was available, and things got better from there. Attending endless conferences and presentations, it felt so much like two decades earlier when I had to try to explain why what we were doing at Biosym was useful for small molecule discovery. Biochemists often shared similar disdain for these types of approaches with their small molecule discovery brethren of yore. But, eventually, the biologics world embraced us. Another baby successfully brought into the world!

Several years post-launch, BioLuminate became the most successful full-feature commercial computational platform for biologics discovery. Once your children are launched, you look for the next move. For me, that came in the form of another opportunity in the biologics space. David Baker (who just won the Nobel Prize for protein modeling) had developed a software platform and community named Rosetta, widely acknowledged for its superior capabilities in certain biologics design tasks. While BioLuminate was a Swiss army knife with some excellent tools, Rosetta was less a Jack of all trades and more a master of some key tasks. Rosetta, however, was notoriously difficult to install and use. Despite its academic successes, it was barely used in industry. In academia, time is cheap and turnaround can be long, so you have the luxury of navigating the combined miseries of complex installation, lousy documentation, and bugs. In industry, you want something you can buy and have running the next day.

Another door opens. Product managing the Cyrus platform — the first commercialized version of the Rosetta platform that David Baker recently won the Nobel Prize for developing and nurturing.

Acknowledging this, Cyrus Biotech was spun out of Baker’s lab to commercialize Rosetta by offering access through an easy-to-use GUI, a Software as a Service (SaaS) model with no end-user installation required, and good documentation. I joined Cyrus to head management of this product. Like Biosym/Discovery and Schrödinger/BioLuminate before, early sales required the right touch. We had to convince people that their preconceived notions of Rosetta’s commercial potential were… wrong. Rosetta was going to be your friend, and you no longer needed half a decade working in a Rosetta lab to learn to use it. We made good progress in changing perceptions, but funding led investors to push for a greater focus on internal drug discovery. So after a couple of fun years, I moved on again. But not without a happy sense of having had yet another finger — perhaps this time a pinkie — in the topping for a pie that, as it turns out, was Nobel Prize-worthy. :-)

Having now been a part of three ground-floor efforts to develop and advocate for groundbreaking computational software for drug discovery, I wasn’t sure if I’d have another chance to experience that pioneering rush that comes from helping to change the drug discovery world at least a tiny bit. I could have considered early retirement — except I don’t golf or play pickleball and believe travel is more fun as an escape from the day-to-day than as a raison d’être. But was there another exciting cutting-edge opportunity out there?

To my surprise, there was. I was contacted by a startup named QSimulate, which had been started by a pair of professors of quantum mechanics (QM), Garnet Chen at CalTech and Toru Shiozaki, who left Northwestern to become CEO of the new company. When I saw they were QM guys, I agreed to meet for coffee with some ambivalence. After all, in the four decades I’d been doing this, QM had scarcely registered in the realm of pharmaceutical discovery, outside of things like calculating partial charges. What history existed for QM in pharma relevance was generally narrow, slow, and of the “cool paper, dude” type. All the tools that revolutionized drug discovery up to this point were based on the simplified equations termed molecular mechanics (MM). While everyone agreed that the fitted analytic equations of molecular mechanics (MM) often provided an anemic approximation to reality represented by QM, it was an approximation we had to live with. That’s because solving QM equations followed the “Rule of Too”: Too expensive, too complex, and too achingly slow to be relevant in pharma discovery, where the mantra remains, “Please have this done by last Thursday, or yesterday at the latest.”

The latest step on my journey to help push the envelope: QSimulate. Convincing the pharma world that quantum mechanics’ time has come.

Well, they convinced me they had something interesting: New QM approaches that addressed issues such as limited applicability to non-physical models, and superior implementations that took full advantage of the explosion in compute resources from cloud computing and GPUs. I came aboard as the VP of product for their new company, QSimulate. And a few years later, we’re convincing the field that QM has finally arrived for pharma. Our first major product was a QM/MM implementation of FEP. Wow, flashback! I’d published a ton of papers on free energies during the first several years of excitement about these methods. Then I’d detoured into the biologics space. During the time I’d strayed, FEP got its second wind, and “suddenly,” 30 years later (!) the pharma world had finally embraced the method. Focusing on FEP again was like coming home to an old friend who had aged like fine wine. And had learned quantum mechanics somewhere along the line. And was now powered by NVIDIA. You know, the usual.

When I started in Peter Kollman’s lab as a post-doc, during FEP’s first peak in the hype cycle, we could maybe perform a hundred picoseconds (total) of MM-based MD sampling for an FEP simulation on a ligand+protein, and that would take weeks. Now we can carry out tens of nanoseconds of sampling using QM/MM and get results in a day or two. That’s hundreds of times more sampling in maybe a tenth of the time while also integrating quantum mechanics to describe the binding site. This opens the door to better predictions for many important interaction classes: metalloproteins, halogenated ligands, highly polarizable systems, covalent ligands, ligands that bind via pi-stacking, and, as they used to say in the late-night infomercials, “much, much more!”. So, just as I have several times before, I find myself advocating for the cutting edge in drug discovery that I believe will eventually be seen as the status quo. It’s a place that suits me.

Though, at QSimulate, I can no longer use the line, “Hey, it’s not like it’s quantum mechanics…” I guess brain surgery will have to do.

Follow A Half Life

This piece was originally published on Medium in October 2024.