ChemBERTaDDI: Transformer Driven Molecular Structures and Clinical Data for predicting drug-drug interactions

Abstract

The problem of polypharmacy arises when two or more drugs taken in combination cause adverse side effects, even when the use of the drugs individually causes no harm. Drug-drug interactions (DDIs) are a major cause of these reactions, contributing to increased morbidity and mortality. As the potential for harmful DDI grows combinatorially, the prediction of drug-drug interactions is increasingly critical for patient safety and effective healthcare management. In this paper, we present the ChemBERTaDDI frame-work a robust approach that uses transformer self-attention to extract latent molecular representations. By employing ChemBERTa-77M-MLM--a transformer-based language model pretrained on SMILES sequences--our approach generates enriched chemical embeddings that capture detailed molecular structural information. These embeddings are integrated with clinical mono side effect data and processed through a DNN predictor, enabling the learning of complex pairwise interaction patterns with minimal architectural overhead. Experiments performed using this combined data on a benchmark data set show superior performance compared with five state-of-the-art methods: Decagon, DeepDDI, MDF-SA-DDI, DPSP and NNPS. ChemBER-TaDDI outperforms the baseline architectures, as measured by F1 and AUROC, and generalizes to new introduced drug compounds.