MPCutter: Predicting Protease-specific Substrate Cleavage Sites Using a Protein Language Model

Proteases can cleave peptide bonds of target substrate proteins. Their controlled proteolysis is vital for protein degradation, recycling, and physiological processes. Understanding the hydrolytic mechanisms of proteases is crucial, particularly for identifying their specific substrates and cleavage sites. Bioinformatics approaches can predict novel protease-substrate cleavage events with high accuracy using sequence and structural information. However, existing tools for cleavage site prediction face several limitations, including restricted accuracy due to limited data and cumbersome training processes that impede timely updates. To address these challenges, we developed MPCutter, which was created by fine-tuning a general-purpose protein sequence language model. This method combined the extensive knowledge of the general model with the targeted optimization of fine-tuning, providing a powerful tool for protease-substrate cleavage prediction. MPCutter offers optimized cleavage site prediction models with enhanced performance and broader coverage across proteases, encompassing four major protease families including 62 distinct proteases. Benchmarking experiments using independent test datasets demonstrated that MPCutter outperformed existing generic tools. In our case study and experiments, MPCutter precisely recognized the majority of cleavage sites and validated five caspase-3 cleavage sites crucial for cellular physiology. Notably, its application to the 10,260-protein human proteome and specific cancer pathways revealed potential new target substrates and provided insights into key biochemical behaviors of proteases. MPCutter is expected to serve as a powerful tool for high-throughput prediction of protease-specific substrates and to facilitate hypothesis-driven exploration of protease proteolytic events. The MPCutter code and associated data are freely available at https://github.com/2053798680wang/MPCutter.git.