Dual perspectives with consistency training on predicting drug–drug interactions

Predicting drug–drug interactions (DDIs) is essential for pharmaceutical research, aiding in the prevention of adverse effects from drug combinations. Our study introduces a novel Dual Perspectives with Consistency Training (DP-CT) method that harnesses both graph structures and SMILES sequences for deep learning-based DDI prediction. Unlike previous methods that focus on a single representation, our DP-CT approach leverages the strengths of both, enhancing predictive accuracy. We employ a Transformer for SMILES encoding and a graph neural network for graph representation, integrating them with point-wise and pair-wise consistency regularizations to ensure alignment and robustness of the encoded drug features. Our method has been tested on the DrugBank and TWOSIDES datasets, demonstrating significant improvements in accuracy, particularly in inductive settings, where we observed an approximate 8-point increase over existing methods.