Deep Learning Models for Predicting Drug–Drug Interactions and Clinical Safety Optimization

Abstract

Adverse drug-drug interactions (DDIs) are one of the most serious and expensive health issues facing the population as they lead to a significant morbidity rate, mortality, and costs. The conventional approaches to detection, such as controlled pharmacokinetic studies and spontaneous reporting systems, are inherently reactive, incomplete and cannot proactively evaluate the enormous combinatoric space of polypharmacy. The synergies of big biomedical data, i.e., chemical structures, biological targets, genomic profiles, electronic health records (EHRs), and scientific literature, with state-of-the-art deep learning (DL) architectures are an opportunity to implement a paradigm shift in predictive DDI safety. This article describes the entire data to bedside implementation pipeline. It outlines the mechanistic typology of DDIs (pharmacodynamic vs. pharmacodynamics), and the multi-modal data needed to predict them. The fundamental DL approaches are discussed, such as molecular representation learning through sequence-based models (Transformers), graph neural networks (GNNs), and geometric deep learning. Link prediction, multi-modal fusion, knowledge graph reasoning and natural language processing are noteworthy predictive paradigms discussed. More importantly, the discussion is not limited to model development to the aspects of the necessary pipeline to clinical translation: rigorous workflow optimization, explainable AI (XAI) to gain a mechanistic insight, and implementation in next-generation, risk-stratified clinical decision support systems (CDSS) to manage polypharmacy and drug development. The combination of these factors makes deep learning one of the major technologies of turning reactive pharmacovigilance into proactive and personalized, and pre-emptive DDI prevention.