WAE-DTI: Ensemble-based architecture for drug–target interaction prediction using descriptors and embeddings

Abstract

Drug Target Interaction (DTI) prediction is one of the main challenges in the pharmaceutical and drug discovery domains due to its high costs, time-consuming nature, and complexity of manual experiments required to evaluate interactions between large numbers of drugs and targets. In addition, a single drug can bind to multiple targets. In contrast, a single target can also bind to a number of drugs; this complicates the DTI task. Existing silico models often struggle with these challenges, particularly in managing diverse datasets. To address these issues, we introduce the Weighted Average Ensemble Drug–Target Interaction (WAE-DTI) model. Our approach integrates several descriptors and fingerprint representations to enhance prediction accuracy and generalization, namely atom pair fingerprint, Avalon, MACCS, MH, Morgan, RDKit, SEC, topological torsion, and LDP for drug representation, and ESM-2 for target representation. WAE-DTI employs a weighted average ensemble technique to handle diverse datasets effectively. The model demonstrates significant improvements over state-of-the-art methods, achieving an average mean squared error of 0.190 ±0.001 on the Davis dataset, 0.127 ±0.001 on Kiba, 0.143 ±0.001 on DTC, 0.284 ±0.004 on Metz, 0.308 ±0.001 on ToxCast, and 0.934 ±0.004 on STITCH. As for the classification task, WAE-DTI outperforms existing models with an AUPRC of 0.943 ±0.001 on BioSNAP, 0.474 ±0.011 on Davis, and 0.707 ±0.005 on BindingDB. Our code is publicly available at ¹.