Interpretable improving prediction performance of general protein language model by domain-adaptive pretraining on DNA-binding protein

Abstract

DNA-protein interactions exert the fundamental structure of many pivotal biological processes, such as DNA replication, transcription, and gene regulation. However, accurate and efficient computational methods for identifying these interactions are still lacking. In this study, we propose a novel method ESM-DBP through refining the DNA-binding protein (DBP) sequence repertory and domain-adaptive pretraining based the protein language model (PLM). Our method considers the lack of exploration of general PLM for DBP domain-specific knowledge, so we screened out 170,264 DBPs from the UniProtKB database to construct the model that more suitable for learning crucial characteristics of DBP. The evaluation of ESM-DBP is systematically performed in four different DBP-related downstream prediction tasks, i.e., DNA-binding protein, DNA-binding residue, transcription factor, and DNA-binding Cys2His2 zinc-finger predictions. Experimental results show that ESM-DBP provides a better feature representation of DBP compared to the original PLM, resulting in improved prediction performance and outperforming other state-of-the-art prediction methods. In addition, ESM-DBP incorporates the integrated gradient algorithm for interpretable analysis, which usually ignored in the previous methods. It reveals that ESM-DBP possesses high sensitivity to the key decisive DNA-binding domains. Moreover, we find that ESM-DBP can still perform well even for those DBPs with only a few similar homologous sequences, and this generalization performs better than the original PLM. The data and standalone program of ESM-DBP are freely accessible at https://github.com/pengsl-lab/ESM-DBP.