CLASPP: A unified model for predicting post-translational modifications.

Area of Science:

• Computational Biology
• Proteomics
• Bioinformatics

Background:

• Post-Translational Modifications (PTMs) are crucial for protein function, but predicting their occurrence is challenging.
• Existing prediction models often focus on single PTM types or use fragmented ensemble methods.
• Data imbalance across different PTM types hinders the development of unified prediction models.

Purpose of the Study:

• To develop a unified deep learning model, CLASPP, for simultaneous prediction of multiple PTM types.
• To address data imbalance challenges in PTM prediction using novel learning and data curation strategies.
• To improve the accuracy and versatility of PTM prediction across diverse organisms.

Main Methods:

• Developed CLASPP, a contrastive learning-based, attention-driven stratified PTM predictor.
• Employed unsupervised clustering-based undersampling and hierarchical data organization for imbalance handling.
• Utilized a multi-stage training strategy with a curated dataset and a pre-trained protein language model.

Main Results:

• CLASPP successfully predicts multiple PTM types simultaneously from primary protein sequences.
• The model demonstrates superior performance compared to existing tools for 12 major PTM types.
• CLASPP shows robust prediction capabilities across different species and provides explainability through kinase substrate specificity profiles.

Conclusions:

• CLASPP offers a unified approach to PTM prediction, effectively addressing data imbalance issues.
• The developed model and curated dataset provide a valuable resource for the research community.
• CLASPP advances PTM prediction accuracy and broadens its applicability in functional proteomics.