Molecular Fingerprints Are a Simple Yet Effective Solution to the Drug–Drug Interaction Problem

Yanan Long, Horace Pan, Chao Zhang, Truong-Son Hy, Risi Kondor

Abstract:

Contemporary machine learning studies tackle the drug–drug interaction forecast problem by featurizing drugs using graph neural networks
(GNNs). This automated featurization allows to avoid laborious handcrafting chemical features. We demonstrate here that a simple neural networks using Morgan fingerprints of the drugs outperformed these more complicated GNN models while spending only a small fraction of the time in training. Furthermore, to improve training, we curated and made available a novel dataset with negative drug–drug interaction examples derived from a very large electronic health records dataset.