Mount Sinai Reveals Hidden PKMYT1 Pocket AI Models Missed

Most stories about AI in drug discovery these days come with confetti. A new model claims to predict protein folds; a startup claims it has shrunk a multi-year discovery cycle to a long weekend; a press release insists the future of medicine is now a software problem. The newest paper from the Icahn School of Medicine at Mount Sinai, published in the Journal of the American Chemical Society on June 2, 2026, lands in a different register. It is a story about what AlphaFold got right, what AlphaFold missed, and what the chemists had to walk into the laboratory and find with their own hands. The result is not a takedown of AI tools. It is something rarer and more useful: a careful map of the boundary between confident prediction and stubborn biological reality, drawn around a cancer protein that nobody quite expected to bend in the way that it did.

What Happened

Researchers at the Icahn School of Medicine at Mount Sinai reported that they had identified a previously unrecognized druggable site on PKMYT1, a kinase that plays a central role in how cells coordinate their division. The work, led by co-senior authors Avner Schlessinger and Michael Lazarus and detailed in a paper from first author N. B. Herrington and colleagues, appeared in the June 2 online edition of the Journal of the American Chemical Society. The team described the discovery as a small but striking demonstration that protein surfaces commonly assumed to be static are anything but. PKMYT1, the paper argues, can shift into shapes that none of today's most widely used structural prediction tools see clearly enough to act on.

The approach the team used was, in outline, the standard new playbook for computational drug discovery. They began with AlphaFold2 to generate predicted structures of PKMYT1. They followed with virtual screening, sifting through libraries of small molecules to identify candidates that might interact with the protein. Where their work diverged from a purely in silico effort is in what came next. The team did not stop at simulated docking scores. They went on to characterize the candidate compounds with X-ray crystallography, ran biochemical assays to measure binding behavior, and tested how the molecules performed in cells. Each layer of evidence was meant to either confirm or correct the picture the algorithms had painted.

The corrective came almost immediately. When the researchers worked through the experimental data, they found that one of the small molecules they had screened did not bind PKMYT1 in the place anyone expected. Instead of latching onto the canonical ATP-binding site, the compound settled into a cleft on the protein that did not appear in any of the available AlphaFold predictions. The paper calls it an allosteric pocket, meaning the compound influences PKMYT1 from a position outside the active site, locking the protein into a unique, inactive conformation around its ATP-binding region. The researchers later went back and asked newer tools, including AlphaFold3 and the recently released Boltz-2, to predict this binding mode; even with the experimental answer in hand as a benchmark, the algorithms struggled to reproduce it.

The team did not soft-pedal the implication of that result. Schlessinger, who directs the AI Small Molecule Drug Discovery Center at Mount Sinai, framed the discovery as an honest accounting of where modern AI sits in 2026: very good at predicting structures that look like other structures it has been trained on, less reliable when nature decides to do something it has not seen before. The paper does not call for retiring AlphaFold; it calls for surrounding the predictions with experimental work that can catch the cases the model is most likely to misread, the cases that are also the most interesting medicinally.

Why It Matters

To appreciate why this particular paper has rippled through the drug discovery community, it helps to understand the long-standing problem with kinase inhibitors. PKMYT1 belongs to a family of about 500 human kinases, and most of them share a strikingly similar architecture at the active site where ATP, the cell's energy currency, slots in. A drug that targets the ATP-binding site of one kinase is therefore liable to also bind the ATP-binding site of several others, sometimes dozens. That cross-reactivity is the root of side effects that have dogged kinase inhibitors since the field's earliest successes. Decades of medicinal chemistry have gone into shaving away that promiscuity one functional group at a time.

An allosteric pocket sidesteps that whole problem at a stroke. Because allosteric sites are not under the same conservation pressure as ATP-binding sites, they tend to diverge across the kinase family. A molecule that binds an allosteric pocket on PKMYT1 has, in principle, a fair chance of leaving other kinases alone. That selectivity is the holy grail of targeted oncology: a drug that hits the protein driving disease without nudging the dozens of related proteins that keep the rest of the body running. The PKMYT1 finding does not deliver a finished therapy, but it does describe a real, structurally validated site on a clinically interesting protein and it identifies starting-point molecules that bind there. For medicinal chemists, that is the difference between gazing at a wall and being handed a foothold.

The relevance to cancer is direct. PKMYT1 has emerged in recent years as a promising target in tumors that carry CCNE1 amplification, a genetic alteration found in subsets of ovarian, endometrial, gastric, and other cancers. Several biotech companies have advanced ATP-competitive PKMYT1 inhibitors into clinical trials, with mixed early signals on tolerability. A selective allosteric handle would change the calculus. It would offer a chemically distinct chassis to build around, one whose toxicity profile is not predetermined by the ATP site's family resemblance. It would also give patients whose tumors become resistant to the first generation of PKMYT1 inhibitors a meaningful second option, because resistance mutations in one binding pocket rarely transfer cleanly to another.

There is a broader point to be drawn from the paper, one that goes beyond a single protein. If PKMYT1 hides a cryptic pocket that AlphaFold cannot see, how many other kinases do the same? The answer is almost certainly that quite a few do, and that the medicinal chemistry community has been overlooking them precisely because they do not appear in the most-used computational maps. A long line of papers over the last decade has chased cryptic and allosteric sites in proteins like PI3K, BRAF, MEK, and DDR1, often with significant clinical payoffs. The Mount Sinai work suggests that the modern, AI-first version of this hunt should treat cases of AI uncertainty as a feature rather than a nuisance: the places the algorithms struggle hardest may be exactly the places worth looking.

Reaction

The reception from the structural biology and drug discovery community has been, in a word, energized. Researchers who have spent years arguing that experimental validation cannot be quietly replaced by computational shortcuts have used the paper as a rallying point. On X and in lab Slacks across academic and biotech sites, the conversation has settled on a phrase from the paper's lead authors: AI is very accurate when predicting known protein shapes, but it can miss completely unexpected binding pockets that only experiments can uncover. That formulation is doing a lot of useful work. It does not dismiss AlphaFold; it locates AlphaFold within the actual workflow of how proteins reveal their secrets.

Among industry scientists, the more interesting reaction has been about chemical sensitivity. Lazarus, the study's co-senior author, emphasized that a very small chemical modification to the lead molecule caused it to abandon the hidden pocket and slide instead into the conventional ATP-binding site. That kind of switch is a familiar headache for medicinal chemists, who routinely watch promising compounds redistribute their binding modes after seemingly innocuous tweaks. The PKMYT1 paper offers a rare structurally resolved case of the phenomenon, complete with the X-ray data to show what is happening at the atomic level. For computational groups working on docking and pose prediction, the implication is uncomfortable: their tools are sometimes presenting confident answers about which pocket a compound prefers when the compound itself, in reality, is balanced on a knife's edge.

Some commentators have framed the result as a warning to the most aggressive forms of AI-first drug discovery now being pitched to investors. The arguments that have come out of Isomorphic Labs and other AI-native shops sometimes lean on the implicit premise that future models will reach a point where wet-lab validation becomes a formality. The Mount Sinai paper is not anti-AI, but it is a counterweight to that premise. It is a reminder that the universe of protein conformations available to a flexible kinase is enormous, that the conformations represented in current training data sample only a portion of that universe, and that the parts of the distribution the model has not seen are precisely the parts where unexpected druggable sites tend to hide. The paper offers an empirical case study, not a slogan.

Reactions inside Mount Sinai have leaned toward the practical. The team has already begun the work of optimizing the compound series that binds the hidden pocket, and it has started a parallel effort to screen other kinases for evidence of similar overlooked sites. Schlessinger has pointed to a second research thread that he considers equally important: feeding the experimental findings back into the development of new prediction models, so that the next generation of AI structure tools can begin to anticipate at least some of the cryptic conformations that today's tools miss. Read in that way, the paper is not a celebration of a limitation; it is an investment in closing the gap.

What's Next

The immediate next steps for the Mount Sinai group fall into three roughly parallel tracks. The first is straightforward medicinal chemistry: building a tighter, more potent set of compounds around the molecule that introduced the team to the hidden pocket. The compounds described in the paper are starting points, not optimized drug candidates. They serve as evidence that the pocket can be occupied and that occupying it produces a measurable biochemical effect on PKMYT1. The work of squeezing more potency, more selectivity, and better pharmacokinetic behavior out of that template is exactly the sort of project that has occupied generations of medicinal chemists, and it now has a fresh structural foundation to build on.

The second track is the kinase-wide hunt for analogous pockets in other proteins. The team has indicated that it intends to use the same general approach in other kinases that are implicated in cancer and other diseases, paying particular attention to cases where current AI tools generate uncertain or conflicting predictions. That triage strategy is itself something of an innovation. Rather than treating the limitations of AI tools as a problem to be hidden or worked around, the group is proposing to use those limitations as a signal: when AlphaFold's confidence drops, that may be a clue that an interesting conformation is lurking in the noise. The work of screening that hypothesis at scale will require the kind of collaboration between computational and experimental teams that has become the norm at large academic medical centers but is still sometimes patchy elsewhere.

The third track is feedback into model development. The Mount Sinai team and several outside groups have signaled interest in using the PKMYT1 result, along with similar cases yet to be reported, as training and benchmark data for a new generation of structure prediction models that explicitly account for protein dynamics. Existing models like AlphaFold2 and AlphaFold3 are powerful in part because they treat structure as a function of sequence; adding the dimension of conformational flexibility introduces serious technical challenges, including the need for richer training data and new architectures that can represent the distribution of states a protein samples in solution. Several labs have already published prototypes that move in this direction, and a result like the PKMYT1 paper gives them a concrete failure case to test against.

Beyond the lab, the next milestone for the broader cancer field is whether any of this translates into a clinical asset. That story will play out on a longer timescale than the rest of the program. Even with optimal lead optimization, a compound discovered today is unlikely to reach a Phase 1 trial in less than two to three years, and clinical readouts on tolerability and efficacy will take significantly longer. Investors should not expect a near-term inflection on the basis of a single JACS paper. What the paper does provide is the kind of upstream finding that, over time, tends to seed multiple downstream programs at several companies at once. CCNE1-amplified ovarian and endometrial cancers, in particular, remain among the most poorly served patient populations in solid tumor oncology, and any technique that widens the search space for selective PKMYT1 modulators will be received with interest.

Closing Thoughts

It is worth pausing on what the PKMYT1 story tells us about the maturation of AI in the sciences. The first wave of headlines about AlphaFold cast it as the end of structural biology, much as ImageNet was once cast as the end of computer vision and large language models have been cast as the end of writing. That framing has always made for clean copy and bad strategy. The closer one looks at the labs that have actually integrated these tools into their daily work, the more the picture comes to resemble what one finds in a competent design office: a tool that produces a confident first draft, surrounded by a workflow that has internalized exactly where the draft is most likely to be wrong. The Mount Sinai paper is a particularly clear example of that working pattern in action.

The deeper lesson is about humility, not about humanity versus machine. Proteins are, in their physical reality, restless and statistical objects. They breathe in solution, sampling a distribution of conformations on timescales that range from picoseconds to seconds. Any static snapshot, whether it comes from a crystal structure or a neural network, is a kind of average over that distribution. Most of the time the average is a good guide. Occasionally it is not, and in those occasional cases the divergence between the average and the underlying reality is where new biology, and new drugs, tend to live. AlphaFold has taught researchers to think of protein structure as an inference problem; the PKMYT1 paper reminds them that inference is always conditional on the data it was trained against.

That reminder has practical consequences for how laboratories should allocate scarce resources in 2026 and beyond. The temptation, especially at well-funded shops, is to scale up the computational pipeline until experimental work feels like a bottleneck. The more interesting move, suggested by results like this one, is to scale the experimental work in carefully chosen places: the cases where the model is least sure, the targets where the conservative answer feels too tidy, the families where prior medicinal chemistry has hit a wall that the algorithms cannot quite explain. That posture is not a rejection of AI. It is an acknowledgment that the value of a powerful first draft is highest when the editorial process around it is sharpest.

If the next decade of AI-driven biology unfolds along the lines the PKMYT1 paper suggests, the headline will not be that AlphaFold solved drug discovery, nor that it failed to. It will be that a generation of scientists learned to use a powerful inferential tool with a kind of disciplined skepticism, the way good empiricists have always used their best instruments. They will hold onto the speed and breadth that AI provides. They will also keep a crystallographer, a biochemist, and a few cellular assays in the loop, watching for the moments when the protein decides to do something the model never saw coming. It is, in its way, the most optimistic version of the AI-in-science story: a future in which the machines do their part, and the people, finally, do theirs.

한글 요약

미국 마운트 사이나이 아이칸 의과대학 연구팀이 2026년 6월 2일 Journal of the American Chemical Society 온라인판에 발표한 논문에서 암 표적 단백질인 PKMYT1에서 지금까지 알려지지 않았던 새로운 약물 결합 부위를 찾아냈다고 보고했다. 연구팀은 AlphaFold2로 단백질 구조를 예측한 뒤 가상 스크리닝을 거쳐 후보 분자를 골랐고, 이후 X선 결정학과 생화학 실험으로 검증을 진행했다. 그 결과 일부 분자가 ATP 결합 부위가 아닌, 기존 AI 도구들이 전혀 예측하지 못했던 알로스테릭 부위에 결합한다는 사실이 드러났다. AlphaFold3와 Boltz-2 같은 최신 모델도 이 결합 양상을 안정적으로 재현하지 못했다.

이번 발견이 중요한 이유는 키나아제 계열 약물의 오랜 난제, 즉 ATP 결합 부위가 키나아제들끼리 너무 비슷해 약물 선택성이 떨어진다는 한계를 우회할 수 있는 가능성을 열기 때문이다. 알로스테릭 부위는 키나아제마다 더 다양하기 때문에, 해당 부위를 노리는 분자는 다른 키나아제를 건드릴 가능성이 상대적으로 낮다. PKMYT1는 CCNE1 증폭을 가진 난소암, 자궁내막암 등 일부 고난도 암 환자군에서 유망한 표적으로 떠올라 있으며, 새로운 결합 부위는 향후 보다 정밀한 항암제 설계의 출발점이 될 수 있다. 단, 임상 적용까지는 수년의 추가 개발이 필요하다.

저자들은 이번 결과를 AI 무용론으로 받아들이지 말라고 분명히 선을 그었다. AlphaFold류 모델은 이미 알려진 단백질 구조를 빠르고 정확하게 예측하는 데 강력하지만, 학습 데이터에 잘 등장하지 않는 동적·숨겨진 구조에서는 한계를 보인다는 사실이 다시 확인된 것이다. 연구팀은 다른 키나아제에도 비슷한 숨은 부위가 있는지 탐색하는 동시에, 이번 발견을 차세대 AI 구조 예측 모델 학습에 환류시키겠다고 밝혔다. 결국 이 논문이 보여 주는 것은 'AI가 신약 개발을 끝낸다'는 단순한 이야기가 아니라, AI의 강점과 실험과학의 강점을 어떻게 조합해야 하는지에 대한 2026년 현재의 가장 실용적인 답이다.