LEP-AD: language embedding of proteins and attention to drugs predicts drug-target interactions

Introduction Predicting drug–target interactions remains a significant challenge in drug development and lead optimization. Recent advances have leveraged machine learning algorithms to model drug–target interactions from molecular and sequence data. Materials and Methods In this work, we use Evolutionary Scale Modeling (ESM-3) to construct a transformer-based protein language representation for drug–target interaction prediction. We introduce LEP-AD (Language Embedding of Proteins and Attention to Drugs), a modular architecture that combines pretrained protein language models with graph-based molecular encoders to predict binding affinity values. Results We systematically benchmark LEP-AD alongside a range of established deep learning methods across multiple datasets—Davis, KIBA, DTC, Metz, ToxCast, and STITCH. To assess predictive validity, we compare model-derived rankings of drug–target interactions with experimental results reported in the literature. In addition, we perform new experimental assays to evaluate the binding of three ATP-competitive Src kinase inhibitors—Dasatinib, UM-164, and Saracatinib—where experimentally measured IC₅₀ and pKᵢ values are consistent with the predicted rankings. Conclusion In summary, our benchmark highlights the strengths and limitations of current drug–target interaction models across diverse datasets and evaluation settings. The results emphasize the impact of pretrained protein and molecular representations on predictive performance and illustrate the persistent challenges of generalization, while the modular LEP-AD framework provides a flexible reference point for comparative evaluation. Scientific Contribution This study presents LEP-AD, a modular deep learning framework for drug–target interaction prediction that integrates pretrained protein language representations with graph-based molecular encoders. Beyond introducing the architecture, we provide a systematic benchmark under similarity-aware evaluation settings and experimental validation, highlighting the impact of pretrained protein embeddings on predictive behavior across diverse datasets.