Deep learning tools for protein structure prediction often generate physically and chemically implausible folds, especially for variant sequences and proteins with ionizable residues. Research published this week in the Proceedings of the National Academy of Sciences found that AI tools frequently overlook the underlying rules of protein folding, and notably, every tool tested rated its own accuracy higher than the results warranted. The work challenges the assumption that better AI automatically means better science. "You cannot blindly believe everything the model predicts," said researcher George Makhatadze.
The tools that were supposed to democratize structural biology are instead teaching scientists an old lesson: trust but verify.