Phosformer: an explainable transformer model for protein kinase-specific phosphorylation predictions.

Area of Science:

• * Bioinformatics
• * Computational Biology
• * Molecular Biology

Background:

• * Protein kinases regulate cellular processes via phosphorylation, but identifying specific kinase-substrate relationships is challenging.
• * Current computational models for kinase-substrate prediction are limited to well-studied kinases and require manual feature engineering.
• * Existing methods struggle with dataset imbalances and lack a unified approach for whole kinome analysis.

Purpose of the Study:

• * To develop a unified computational framework for predicting kinase-specific phosphorylation sites across the entire human kinome.
• * To leverage universal protein language models for automated feature extraction from primary sequence data.
• * To overcome limitations of existing models, including reliance on curated features and restricted kinase coverage.

Main Methods:

• * Developed Phosformer, a deep learning model utilizing universal protein language models.
• * Trained Phosformer on kinase and substrate peptide sequences to predict phosphorylation probability.
• * Analyzed learned features to assess evolutionary, functional, and substrate specificity motif recognition.

Main Results:

• * Phosformer accurately predicts kinase-specific phosphorylation sites using only primary sequence information.
• * The model implicitly learns relevant evolutionary and functional features, eliminating the need for manual curation.
• * Phosformer demonstrates superior performance compared to state-of-the-art methods and distinguishes kinase families.

Conclusions:

• * Phosformer provides a generalized, unified, and interpretable framework for kinase-specific phosphosite prediction.
• * The model enables substrate predictions at the whole kinome level, enhancing investigative utility.
• * This approach represents a significant advancement in understanding kinase-substrate interactions and phosphorylation dynamics.