Counterfactual Debiased Co-Embedding Model for Enhanced Drug-Drug Interaction Prediction

Predicting drug-drug interactions (DDIs) is critical to drug discovery and development because adverse interactions pose serious health risks. Most of the existing studies utilize the properties of drugs or network topology information of DDIs to predict unknown interactions between drugs. However, DDI networks are usually sparse with insufficient interaction information, and these approaches lack in-depth integration of these two types of information to effectively exploit potential associations between DDI network nodes and properties. In this work, we present a novel co-embedding model, counterfactual debiased co-embedding (CDCE), for counterfactual-based analyses. The model mitigates the effects of sparse networks and information embedding loss through counterfactual debiasing without losing the original information. In addition, we fuse two attribute information, Anatomical Therapeutic Chemical (ATC) code and Simplified Molecular Input Line Entry System (SMILES), from different perspectives. The implicit information obtained from the ATC code is embedded into the DDI network and then fused with SMILES through the variational graph autoencoder model. We validated CDCE on the benchmark dataset BioSNAP, with experimental results showing that it outperforms state-of-the-art methods.