Report 003 · AI in the Lab
Several state-of-the-art structure predictors nailed the shape of a cancer protein and missed a hidden druggable pocket. The lab found it anyway. The gap between those two facts is the whole story.
The headline going around was some version of "AI unveils hidden drug pocket in cancer protein." It's a good headline. It's also backwards. The AI didn't unveil the pocket. The AI is the reason we can say, precisely, that it missed one.
Here's the actual result, published this June in the Journal of the American Chemical Society by a team at the Icahn School of Medicine at Mount Sinai. Their target was PKMYT1, a kinase: an enzyme that helps control how cells divide, and a protein cancer drug developers care about. They started the way a modern lab does: they used AlphaFold2 to predict the protein's structure and ran computational screening for places a drug might bind. Then they did the slow part (X-ray crystallography, biochemical assays, and cell studies) to see what was really there.
What was really there was a pocket nobody's model saw coming: a previously unknown binding site that only opens up when the protein is pushed into an unusual, inactive shape. And this is the part worth sitting with: it wasn't just AlphaFold2 that missed it. The team reports that AlphaFold3 and a newer predictor, Boltz-2, plus unbiased physics simulations, all failed to predict this pocket, even as they reproduced the protein's known, ordinary shape well. The lead researcher, Dr. Avner Schlessinger, put it exactly right:
"AI was very accurate when predicting known protein shapes, but it missed a completely unexpected binding pocket that we could only uncover experimentally."
Credit where it's due, and only where it's due: the discovery came out of the combined workflow. The AI-seeded screening pointed the way, and then experimental structural biology, led by crystallography, resolved the new conformation. But the thing that made the pocket real, that turned "the model didn't rule it out" into "here it is, at this resolution, in this shape," was a physical experiment. Not a prediction.
This isn't an "AI is dumb" story, and I want to be careful not to let it become one. It's a story about what these tools are actually built to do. A structure predictor like AlphaFold is extraordinary at one task: given a sequence, guess the folded shape the protein most likely settles into. That's a single, representative snapshot: the protein at rest, being itself.
A cryptic pocket like this one is a different animal. It doesn't exist in the resting shape. It appears only when something (here, an inhibitor) perturbs the protein into a strained, non-obvious conformation it doesn't normally visit. Predicting the average fold and predicting the rare, drug-induced states a protein can be forced into are not the same problem, and the second one is much harder. The models did the job they were trained for and stopped exactly at the edge of it. The lab's instruments don't care what a protein "usually" does; they measure whatever is in front of them, including the shape no one expected.
I read this as an outside observer: someone who consults on where AI genuinely helps a lab and where it's expensive theater, and who, in a past research life, measured molecular structure the slow, physical way (different molecules, a different instrument, none of it anywhere near this study, but the same lesson). A measurement has a stubborn advantage a model doesn't: it can be surprised. A predictor trained on known structures is, almost by definition, biased toward the known. An X-ray doesn't have a prior.
It doesn't mean AlphaFold is overrated. The same paper shows the models were highly accurate on the protein's established structure, and the AI-guided screening is part of why the team was looking in a productive place at all. Skipping those tools would have made the work slower, not better. And it's one protein: a vivid, well-documented example of a limitation, not a universal law about every target. Someone will build the model that catches the next cryptic pocket; that's how this goes.
What it does mean is that "AI discovered X" deserves a reflex question every time you read it: what did the model actually do, and what did the humans and the instruments do? In this case the model predicted the easy part and a crystallographer found the hard part. Both mattered. Only one of them made the headline, and it was the wrong one.
U.S. Army combat veteran (Counter-IED / Electronic Warfare), peer-reviewed researcher in microwave spectroscopy, and founder & CEO of Shroombiosis. Consults on laboratory operations, AI, and supplement formulation.