TripleBind: a generalizable deep learning framework for protein-nucleic acid and protein-ligand binding sites prediction based on pre-trained protein language models.

Protein-nucleic acid and protein-ligand interactions play important roles in regulating cellular processes and form a fundamental foundation for drug discovery and design. Although existing experimental assays can yield highly accurate measurements, they are often labor-intensive, costly, and insufficient to meet the rapidly growing demand for large-scale protein sequence annotation.Consequently, developing an efficient and reliable computational framework capable of accurately identifying protein-nucleic acid and protein-ligand binding residues has become crucial. In this study, we present TripleBind, a new deep learning architecture that predicts protein-nucleic acid and protein-ligand binding sites using sequence information alone. TripleBind integrates three Transformer-based pre-trained protein language models with a specially designed Multi-Branch Convolutional Network (MBCN) module. Experimental results on protein-nucleic acid benchmark datasets demonstrate that our method achieves MCC scores of 0.392 and 0.512 on independent protein-DNA test sets, and an MCC of 0.46 on an independent protein-RNA test set, shows improved performance compared with existing sequence-based methods. In addition, TripleBind can serve as a general predictor of protein-ligand interactions and exhibits robust performance across different tasks. Finally, we provide interpretability analyses to elucidate the behavior of our model and summarize its overall contributions.