In drug discovery, predicting how a small molecule (ligand) binds to a protein is key. But when computational models generate many possible binding positions (poses), telling the correct one from the others is surprisingly hard. Even expert chemists with 3D tools and scoring methods performed no better than random guessing on this task.
The researchers built a called Pyrite, containing 1,500 pairs of poses from diverse protein-ligand complexes. They then created two kinds of AI agents: one that uses language directly to reason about which pose is better, and one that writes code to compute structural features for a simple classifier.
The coding-based agent achieved 74% accuracy, the best result and higher than expert chemists. This shows that guiding a to discover informative features is more effective than asking it to reason directly, offering a promising way to combine computational chemistry tools for better pose ranking.