Addressing model overcomplexity in drug-drug interaction prediction with molecular fingerprints

Accurately predicting interactions between drugs is critical for pharmaceutical research and clinical safety. The literature keeps moving toward increasingly complex architectures, yet gains on standard benchmarks are often small. We use a deliberately simple setup that keeps the classifier fixed and swaps only the molecular representation. We compare ECFP4 Morgan fingerprints (MFPs), pretrained graph convolutional networks (GCNs), and MoLFormer embeddings on common DrugBank DDI splits and on an FDA drug-drug affinity benchmark. This design lets us isolate the effect of representation and of split design with minimal confounders.Across leak proof DrugBank splits and the FDA task, MFPs with a shallow head match or surpass much more complex graph and knowledge graph systems while using far fewer parameters. On the Unseen DDI split, MFPs reach AUROC 99.4 and AUPR 98.4, slightly ahead of MoLFormer and the prior state of the art. When we impose a strict scaffold out-of-distribution split, a pretrained GCN leads with AUROC 73.99, which pinpoints when extra capacity helps. On standard splits the absolute gains from added complexity are small. We therefore argue that progress will come mainly from better curated datasets and from rigorous out-of-distribution evaluation, rather than from ever more elaborate models.