MuFGPS: enhancing liquid-liquid phase separation protein prediction through multi-level features and ensemble...

Area of Science:

• Biochemistry
• Structural Biology
• Computational Biology

Background:

• Liquid-liquid phase separation (LLPS) drives the formation of membrane-less organelles and is implicated in cellular functions and diseases.
• Current computational methods for predicting LLPS proteins often lack explicit structural topology information, limiting accuracy due to sequence diversity and complex determinants.
• Accurate identification of LLPS-capable proteins is essential for biological and medical research.

Purpose of the Study:

• To develop a novel computational framework, MuFGPS (Multi-level Feature Graph-based Predictor for Phase-Separating proteins), for accurate prediction of LLPS-capable proteins.
• To integrate diverse biological data, including sequence features, secondary structures, and structural embeddings, for enhanced predictive power.
• To benchmark MuFGPS against existing methods and validate the contribution of its integrated features.

Main Methods:

• MuFGPS integrates sequence-derived physicochemical features, secondary structure information (from Define Secondary Structure of Proteins), and graph-based structural embeddings derived from AlphaFold residue contact maps.
• A multi-head Graph Attention Network is employed for processing structural information.
• Class imbalance is managed using Synthetic Minority Oversampling Technique (SMOTE), and classification is achieved via a stacking ensemble of Random Forest, XGBoost, and LightGBM.

Main Results:

• MuFGPS demonstrates superior performance compared to six representative methods across all evaluated metrics.
• Significant improvements were observed in F1-score and Matthews Correlation Coefficient (MCC).
• Ablation studies confirmed that both structural features and ensemble learning synergistically enhance prediction accuracy and robustness.

Conclusions:

• MuFGPS provides a scalable and highly accurate framework for predicting proteins involved in liquid-liquid phase separation.
• The integration of multi-level features, particularly structural information, is critical for improving LLPS protein prediction.
• This framework has the potential for broad application in proteome-wide LLPS protein identification and disease research.