DeepMVP: deep learning models trained on high-quality data accurately predict PTM sites and variant-induced...

Area of Science:

• Molecular Biology
• Bioinformatics
• Computational Biology

Background:

• Post-translational modifications (PTMs) are crucial for protein function, and their dysregulation by missense variants contributes to disease.
• Predicting PTM sites using deep learning is vital for identifying PTM-altering variants, but lacks sufficient high-quality training data.

Purpose of the Study:

• To introduce PTMAtlas, a comprehensive PTM site compendium, and DeepMVP, a deep learning framework for PTM site prediction.
• To evaluate DeepMVP's performance against existing tools and its utility in identifying PTM-altering missense variants.

Main Methods:

• Systematic reprocessing of 241 public mass-spectrometry datasets to curate PTMAtlas, containing 397,524 PTM sites.
• Development of DeepMVP, a deep learning model trained on PTMAtlas for predicting six types of PTMs: phosphorylation, acetylation, methylation, sumoylation, ubiquitination, and N-glycosylation.
• Validation of DeepMVP's predictions using literature-curated variants and cancer proteogenomic datasets.

Main Results:

• PTMAtlas provides a large, high-quality dataset for PTM research.
• DeepMVP significantly outperforms existing PTM prediction tools across all six PTM types.
• DeepMVP accurately predicts PTM-altering missense variants, showing strong concordance with experimental data.

Conclusions:

• PTMAtlas and DeepMVP offer a powerful, integrated platform for advancing PTM research.
• This framework enables scalable assessment of coding variant functional consequences via PTM disruption.
• The developed tools facilitate a deeper understanding of PTMs in health and disease.