IntPred: a structure-based predictor of protein-protein interaction sites

Area of Science:

• Computational biology and bioinformatics
• Structural biology
• Machine learning in bioinformatics

Background:

• Protein-protein interactions (PPIs) are crucial for protein function, with proteins interacting with 3-10 partners on average.
• Experimental structures of protein complexes are limited, with only 50% of Protein Data Bank entries representing complexes.
• Accurate prediction of protein-protein interfaces is essential due to the scarcity of experimental complex structures.

Purpose of the Study:

• To develop and evaluate a novel in silico predictor for protein-protein interfaces.
• To compare the performance of the new predictor against established methods.
• To assess the predictor's effectiveness on different types of protein complexes (obligate and transient).

Main Methods:

• Utilized a random forest machine learning approach.
• Employed structural features for interface prediction.
• Developed the IntPred predictor, implemented in Perl.
• Made IntPred available via a web server and for local download.

Main Results:

• IntPred demonstrated strong performance on an independent test set (MCC=0.370, ACC=0.811, SPEC=0.916, SENS=0.411).
• IntPred ranked second among six tested methods, closely following SPPIDER.
• SPPIDER showed higher overall performance but significantly lower specificity compared to IntPred.
• Performance varied with complex type: enhanced for obligate (MCC=0.381) and reduced for transient (MCC=0.303).

Conclusions:

• IntPred is a competitive tool for predicting protein-protein interfaces.
• The choice between IntPred and other methods like SPPIDER depends on specific application needs, particularly the sensitivity-specificity trade-off.
• The predictor's performance is influenced by the nature of the protein complexes being analyzed.