PSTP: accurate residue-level phase separation prediction using protein conformational and language model embeddings

Area of Science:

• Biochemistry
• Computational Biology
• Molecular Biology

Background:

• Phase separation (PS) is vital for cellular processes and disease pathogenesis.
• Existing predictive methods for PS often require extensive annotations or handcrafted features, limiting their applicability.
• There is a need for accurate, generalizable algorithms to predict PS from protein sequences.

Purpose of the Study:

• To develop a novel computational tool for predicting protein phase separation (PS) from amino acid sequences.
• To enable accurate PS predictions without relying on sequence annotations or handcrafted features.
• To investigate the link between pathogenic variants, intrinsically disordered regions, and PS propensity.

Main Methods:

• Developed Phase Separation's Transfer-learning Prediction (PSTP) by combining conformational and large language model embeddings.
• Utilized protein sequences as the sole input for PS prediction.
• Analyzed over 160,000 variants to correlate pathogenic mutations with residue-level PS propensities.

Main Results:

• PSTP achieves state-of-the-art PS prediction accuracy using only protein sequences.
• Residue-level predictions from PSTP correlate highly with experimentally validated PS regions.
• Identified a significant association between pathogenic variants and PS propensity in intrinsically disordered regions.

Conclusions:

• PSTP offers a powerful and versatile tool for predicting protein phase separation behavior.
• The method facilitates the analysis of mutation effects on PS, particularly in unconserved intrinsically disordered regions.
• PSTP's efficiency and accessibility (web server, Python package) support broad application in biological research and disease studies.