Binary Drug-Drug Interaction Prediction by Dual-View Graph Convolutional Networks with Semi-Hard Negative Sampling

Abstract

The issue of DDI prediction is a patient safety issue and the majority of benchmarks currently in use are ineffective to address it. Redundant entries and pair-wise cuts swell the figures with no one lying, because the model only memorizes, but does not generalize and conventional measures cannot detect it.. The TWOSIDES dataset is publicly available and is used in this pipeline: interaction records are aggregated into unique unordered drug pairs, and negatives are sampled with a Tanimoto cutoff of 0.4 with Morgan fingerprints to eliminate structural shortcuts, and the evaluation is run on both drug-level and pairlevel splits to decouple real generalization and data leakage. The model consists of a three layer GCN and a fingerprint projection head that is concatenated to form dual-view drug embeddings, which are then combined with interaction operators and then finally a compact focal loss MLP classifies them. AUPRC Splits in drug levels include: 0.6355, AUROC 0.6298, above the fingerprint-only baseline. Pair-level split: AUPRC 0.9760. The difference between them is 0.34 which is not a performance gain but rather an indication of the extent of split design distorting reported results. Light weight post hoc explanations are incorporated through the use of fingerprint ablation and gradient saliency on atomic features.