DrugTar Improves Druggability Prediction by Integrating Large Language Models and Gene Ontologies

Abstract

Target discovery is crucial in drug development, especially for complex chronic diseases. Recent advances in high-throughput technologies and the explosion of biomedical data have highlighted the potential of computational druggability prediction methods. However, most current methods rely on sequence-based features with machine learning, which often face challenges related to hand-crafted features, reproducibility, and accessibility. Moreover, the potential of raw sequence and protein structure has not been fully investigated. Here, we leveraged both protein sequence and structure using deep learning techniques, revealing that protein sequence, especially pre- trained embeddings, is more informative than protein structure. Next, we developed DrugTar, a highl7lperformance deep learning algorithm integrating sequence embeddings from the ESM-2 pre-trained protein language model with protein ontologies to predict druggability. DrugTar achieved areas under the curve and precision-recall curve values above 0.90, outperforming state-of-the-art methods. In conclusion, DrugTar streamlines target discovery as a bottleneck in developing novel therapeutics.