Graph Neural Networks Improve Drug-Drug Interaction Type Prediction

A new study on arXiv (2605.27861) systematically compares three Graph Neural Network architectures for predicting drug-drug interaction types. Using a benchmark dataset of 38,337 positive pairs across 86 interaction types, the researchers found that a dual Message Passing Neural Network with four-head cross-attention (CrossAtt) improves multi-class F1-macro by 45% over a simple concatenation model, while binary detection AUC improves only 1.3%. This confirms that atom-level inter-molecular communication specifically enables mechanism-type classification.