SSPSPredictor: A Sequence- and Structure-Based Deep Learning Model for Predicting Phase-Separating Proteins.

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Genetics

Background:

• Liquid-liquid phase separation (LLPS) drives the formation of membraneless organelles (MLOs), crucial for cellular organization.
• Phase-separating proteins (PSPs) are key drivers of LLPS, but experimental identification is challenging.
• Existing computational tools for PSP identification have limitations.

Purpose of the Study:

• To develop SSPSPredictor, a novel multimodal model for identifying PSPs, predicting LLPS propensity, and pinpointing LLPS-driving regions.
• To enhance the computational prediction of PSPs for both folded and intrinsically disordered proteins.
• To investigate the link between LLPS, intrinsically disordered proteins (IDPs), and pathogenic variants.

Main Methods:

• Developed SSPSPredictor, a multimodal model fusing protein language model (ESM-2) sequence data and graph neural network (GVP) structural insights.
• Employed a novel approach for PSP identification and LLPS propensity prediction.
• Analyzed the human proteome and pathogenic variants using SSPSPredictor.

Main Results:

• SSPSPredictor demonstrates balanced performance in PSP identification, LLPS propensity prediction, and identification of LLPS-driving regions.
• The model shows interpretability in identifying sequence regions critical for LLPS.
• Intrinsically disordered proteins (IDPs) are significantly more involved in LLPS than folded proteins.
• Pathogenic variants, particularly in disordered regions, show increased LLPS propensity, linking LLPS to disease at the amino acid level.

Conclusions:

• SSPSPredictor offers an effective and interpretable tool for studying LLPS and PSPs.
• LLPS plays a significant role in the function of IDPs.
• Aberrant LLPS driven by mutations in disordered regions may contribute to disease pathogenesis.