AI Drug Interaction Models Show Limitations in Understanding Chemical Details
-
Researchers examine inner workings of AI models predicting drug-protein interactions, finding they rely more on recalling data than learning specific chemical interactions.
-
Study analyzed 6 graph neural networks (GNNs) using new "EdgeSHAPer" method to uncover how they make predictions.
-
GNNs focused more on chemical similarities of previously seen molecules rather than learning protein-drug interactions to make predictions.
-
Findings indicate GNN predictions are overrated; simpler methods can give similar quality forecasts.
-
Developing methods like EdgeSHAPer to explain AI model predictions helps illuminate the "black box" of how they work.