AlphaFold 3

Physics limits

• Too little good data about how drugs bindOnly about 230,000 molecular structures have been solved by experiment, and far fewer have reliable binding measurements. Models trained on that are weakest on exactly the novel chemistry drugmakers want.
• It predicts one snapshot, not motionProteins flex, open and close. The model gives a single likely shape, and struggles with floppy regions and with how structures change when a drug binds.
• Confident shapes can still be wrongTrained to produce plausible structures, it can place atoms convincingly where no real structure exists, especially for new kinds of molecules unlike its training data.