LM-TLIPs: Integrating Large Model and Transfer Learning Technology for Precise Identification of Phosphorylation Sites in SARS-CoV-2

In recent years, the rapid spread of SARS-CoV-2 has triggered a global health crisis and socio-economic challenges. As a crucial post-translational modification, phosphorylation plays a vital role in the regulation of cellular functions. Given its close relationship with SARS-CoV-2 infection, accurately identifying virus-induce phosphorylation sites is essential for understanding the molecular mechanisms of viral infection and its impact on host cells. Although the development of various computational tools for predicting phosphorylation sites, these tools have several shortcomings, such as insufficient data and limited model generalization ability, which limit their effectiveness in practical applications. To overcome these limitations, this study proposes a novel method for predicting SARS-CoV-2 phosphorylation sites, LM-TLIPs, based on the latest technology. This method uses the most advanced large model technology ESM-2 to extract information from S/T sites and Y sites; by fine-tuning the large model and introducing transfer learning technology, it addresses the challenge of accurately predicting Y sites due to insufficient data in this study. Independent testing on S/T sites(Acc:0.8309, Sn:0.8443, Sp:0.8174, MCC:0.6620, AUC:0.8993) and Y sites(Acc:0.9048, Sn:0.9524, Sp:0.8571, MCC:0.8132, AUC:0.9388) has validated that LM-TLIPs outperforms existing optimal prediction tools, demonstrating its superior ability in identifying phosphorylation sites. Furthermore, we conducted an exhaustive interpretability analysis based on attention weight heatmaps and feature importance ranking to enhance the transparency and confidence of prediction results.