Kinase-substrate prediction using an autoregressive model.

Area of Science:

• Biochemistry
• Computational Biology
• Genomics

Background:

• Kinase-specific phosphorylation is crucial for cellular signaling and disease pathogenesis.
• Identifying kinase substrates remains a significant challenge, hindering a comprehensive understanding of cellular processes.
• Existing methods lack reliable approaches for predicting kinase-substrate relationships.

Purpose of the Study:

• To introduce an innovative autoregressive model for predicting kinase-substrate pairs at the protein level.
• To address the limitations of site-specific phosphorylation prediction by focusing on kinase-protein interactions.
• To provide a reliable computational tool for identifying kinase targets.

Main Methods:

• Utilized an autoregressive model integrating the ESM-2 protein large language model as an encoder.
• Redefined the problem as a protein-protein interaction prediction task with a binary classification approach.
• Implemented a hard negative strategy based on kinase embedding distances for robust model training.
• Performed top-k analysis to evaluate the prioritization of potential kinase candidates.

Main Results:

• The model successfully predicts kinase-substrate pairs at the protein level.
• Demonstrated capability for zero-shot prediction, identifying substrates for kinases with no prior known interactions.
• Showcased robust generalization to novel kinases and underrepresented protein groups.
• Effectively distinguished positive kinase-substrate interactions from negative ones through the hard negative strategy.

Conclusions:

• The developed autoregressive model offers a novel and effective solution for predicting kinase-substrate relationships.
• This approach significantly advances the field by enabling protein-level substrate identification and zero-shot prediction.
• The model's versatility and generalization capabilities highlight its potential utility in diverse biological research and disease studies.