For three years, every major AI structure-prediction tool available to drug hunters has been trained to see something that, it turns out, none of them could actually see. Researchers at the Icahn School of Medicine at Mount Sinai in New York report that they discovered a previously unknown drug-binding pocket on PKMYT1, a cancer-related protein, using old-fashioned experimental biochemistry — after AlphaFold2, AlphaFold3, Boltz-2 and molecular dynamics simulations all failed to flag it.
The findings, published in the Journal of the American Chemical Society, land at a strange moment for the field. AI-guided drug discovery has raised roughly $8.9 billion in venture and pharma investment by some industry estimates, yet as of August 2026 not a single AI-discovered drug has received full FDA approval. Mount Sinai’s paper doesn’t argue AI is useless — it argues something more specific and, for the industry, more uncomfortable: the tools are excellent at confirming what is already known and strikingly blind to what is not.
A Kinase Everyone Thought They Understood
PKMYT1 is a kinase, a type of protein that acts as a molecular switch controlling how and when cells divide. It has drawn attention from cancer researchers because inhibiting it can force tumor cells with defective DNA-repair machinery into a lethal mitotic catastrophe, killing them while sparing healthy cells. Several companies, including Sierra Oncology and others, have already built clinical-stage PKMYT1 inhibitors on the strength of its known structure.
The Mount Sinai team, led by researchers in the department of pharmacological sciences, set out to map every possible small-molecule binding site on the protein using biophysical fragment screening — essentially soaking crystallized protein in a library of small chemical fragments and observing, atom by atom, where they stick. What they found was a pocket nobody had modeled: a druggable cavity that opens only under specific conformational conditions the AI systems had never been trained to anticipate.
Why the Machines Missed It
“AI was very accurate when predicting known protein shapes, but it missed a completely unexpected binding pocket that we could only uncover experimentally,” the study’s senior author said in comments accompanying the release. That is the crux of the finding: AlphaFold and its successors are trained on the Protein Data Bank, a library of structures that themselves came largely from crystallography and cryo-EM experiments. If a class of pocket has never been captured in that training data, the models have no way to conjure it out of first principles, no matter how sophisticated their attention mechanisms.
Boltz-2, an open-source diffusion model released in 2025 specifically marketed for its improved binding-affinity predictions, fared no better. Neither did classical molecular dynamics, which simulates atomic motion over time and is often treated as a physics-based check on AI predictions. All four approaches converged on the same known geometry and skipped past the pocket entirely.
The Industry’s Uncomfortable Math
The timing matters. Groups like Insilico Medicine, Recursion Pharmaceuticals and Isomorphic Labs have built entire business models around the promise that generative AI can shrink drug discovery timelines from years to months. Insilico’s own TNIK inhibitor, rentosertib, has become the industry’s most-cited proof point after posting encouraging Phase IIa lung-function results published in Nature Medicine. But skeptics, including several quoted in trade press coverage of the sector’s spending, note that hype has consistently outpaced regulatory outcomes, and that AI-native pipelines still lean heavily on wet-lab validation to catch what models miss.
Mount Sinai’s researchers don’t frame their work as a rebuke of AI drug discovery broadly — they frame it as a map of its blind spot. Their plan now is to develop more potent compounds targeting the newly discovered PKMYT1 pocket, and separately to investigate whether comparable hidden pockets exist on other cancer-related kinases that have already been “solved” by AI and considered done.
A Feedback Loop Back Into the Models
There’s a second, quieter ambition buried in the paper: using this newly characterized pocket as training data to make the next generation of structure-prediction tools better at anticipating conformational surprises rather than only reproducing known shapes. That is a slower, less glamorous version of AI-driven drug discovery than the “molecule-to-clinic-in-18-months” pitch decks that have circulated across biotech conferences this year, but researchers argue it is the version most likely to actually produce new medicines.
What Comes Next
Pharmaceutical companies running PKMYT1 programs will now have to decide whether to revisit compounds they may have deprioritized based on incomplete structural pictures. More broadly, the finding is likely to feed into a growing conversation — visible in recent FDA guidance discussions and in academic critiques of AI-native biotech valuations — about how much experimental verification should be mandatory before an AI-predicted structure or binding site is treated as reliable enough to build a clinical program on. For an industry racing to prove that algorithms can compress a decade of pharmacology into months, Mount Sinai’s quiet, unglamorous fragment-screening experiment is a reminder that the wet lab still catches things the models can’t yet imagine.
