Deep learning model of post-translational modification regulating liquid-liquid phase separation.

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Genomics

Background:

• Liquid-liquid phase separation (LLPS) is vital for forming membraneless organelles, impacting biological functions and disease pathogenesis.
• Protein post-translational modifications (PTMs) significantly regulate LLPS, but dedicated resources and predictive tools are scarce.

Purpose of the Study:

• To establish a comprehensive database (PTMPhaSe) and a predictive model (PhosLLPS) for PTMs regulating LLPS.
• To enhance understanding of PTMs' role in LLPS dynamics and facilitate drug discovery for LLPS-related disorders.

Main Methods:

• Construction of the PTMPhaSe database with manually curated experimental evidence on PTMs and LLPS.
• Development of a graph neural network-based deep learning model (PhosLLPS) for predicting functional phosphorylation sites involved in LLPS.
• Application of PhosLLPS for large-scale prediction across the human proteome.

Main Results:

• The PTMPhaSe database provides a curated repository of PTM regulation in LLPS.
• PhosLLPS demonstrated superior performance (AUC=0.9116) in identifying functional phosphorylation sites compared to baseline models and existing methods.
• PhosLLPS predictions were generated for the entire human proteome, offering insights into regulatory phosphorylation sites.

Conclusions:

• The developed database and predictor bridge the gap between PTM regulation and LLPS.
• These resources are valuable for advancing the study of LLPS molecular mechanisms and supporting the development of therapeutics for diseases linked to aberrant LLPS.