SPAR: a random forest-based predictor for self-interacting proteins with fine-grained domain information.

Area of Science:

• Bioinformatics
• Computational Biology
• Molecular Biology

Background:

• Protein self-interaction is vital for cellular regulation.
• Experimental methods for identifying self-interacting proteins are limited.
• Computational approaches are needed for accurate self-interacting protein prediction.

Purpose of the Study:

• To develop an improved computational approach for predicting self-interacting proteins (SIPs) from sequence information.
• To introduce a novel encoding scheme, critical residues substitution (CRS), incorporating domain-domain interaction data.
• To evaluate and optimize the SPAR (Self-interacting Protein Analysis serveR) model for enhanced prediction accuracy.

Main Methods:

• Developed the Critical Residues Substitution (CRS) encoding scheme, integrating fine-grained domain-domain interaction information.
• Employed the Random Forest algorithm to evaluate CRS performance against other common encoding schemes.
• Utilized the Minimum Redundancy Maximum Relevance (mRMR) feature selection method to identify key features for the SPAR model.
• Validated the SPAR model on independent human and yeast test datasets.

Main Results:

• The CRS encoding scheme achieved an average accuracy of 72.01% in tenfold cross-validation.
• The SPAR model, integrating CRS and selected features, reached 92.09% accuracy on a human-independent test set.
• SPAR demonstrated cross-species applicability, achieving 76.96% accuracy on an independent yeast test set.

Conclusions:

• The proposed SPAR model significantly improves self-interacting protein prediction accuracy using sequence-based features.
• The CRS encoding scheme effectively captures essential information for identifying protein self-interactions.
• SPAR provides a valuable, freely accessible bioinformatics tool for researchers studying protein self-interactions and cellular functions.