Prediction of protein-binding residues: dichotomy of sequence-based methods developed using structured complexes...

Area of Science:

• Computational biology
• Bioinformatics
• Structural biology

Background:

• Over 30 sequence-based predictors for protein-binding residues (PBRs) exist.
• These predictors are trained on either structure- or disorder-annotated datasets, creating a dichotomy.
• Structure-trained predictors often inaccurately predict PBRs for non-protein partners.

Purpose of the Study:

• To conduct a comparative study of existing PBR predictors.
• To analyze the cross-over prediction capabilities between structure- and disorder-annotated proteins.
• To evaluate cross-predictions with non-protein binding partners.

Main Methods:

• Comparative analysis of disorder- and structure-trained PBR predictors.
• Utilized a comprehensive benchmark set with diverse annotations (structure, disorder, protein, nucleic acid, small ligand).
• Developed a novel hybrid predictor, hybridPBRpred, by combining disoRDPbind and SCRIBER.

Main Results:

• SCRIBER, ANCHOR, and disoRDPbind showed accurate predictions within their respective training data types.
• Most predictors failed to cross-over between structure and disorder annotations, except for SCRIBER.
• Nearly all methods exhibited significant cross-predictions with non-protein partners, with exceptions for SCRIBER and disoRDPbind.
• hybridPBRpred demonstrated accurate cross-over predictions and low cross-prediction rates.

Conclusions:

• Existing PBR predictors have limitations in generalizability.
• The novel hybridPBRpred offers an accurate and versatile alternative for PBR prediction.
• hybridPBRpred successfully addresses the dichotomy and cross-prediction issues in PBR prediction.