Abstract
MOTIVATION: Protein-protein interactions are vital for protein function with the average protein having between three and ten interacting partners. Knowledge of precise protein-protein interfaces comes from crystal structures deposited in the Protein Data Bank (PDB), but only 50% of structures in the PDB are complexes. There is therefore a need to predict protein-protein interfaces in silico and various methods for this purpose. Here we explore the use of a predictor based on structural features and which exploits random forest machine learning, comparing its performance with a number of popular established methods. RESULTS: On an independent test set of obligate and transient complexes, our IntPred predictor performs well (MCC = 0.370, ACC = 0.811, SPEC = 0.916, SENS = 0.411) and compares favourably with other methods. Overall, IntPred ranks second of six methods tested with SPPIDER having slightly better overall performance (MCC = 0.410, ACC = 0.759, SPEC = 0.783, SENS = 0.676), but considerably worse specificity than IntPred. As with SPPIDER, using an independent test set of obligate complexes enhanced performance (MCC = 0.381) while performance is somewhat reduced on a dataset of transient complexes (MCC = 0.303). The trade-off between sensitivity and specificity compared with SPPIDER suggests that the choice of the appropriate tool is application-dependent. AVAILABILITY AND IMPLEMENTATION: IntPred is implemented in Perl and may be downloaded for local use or run via a web server at www.bioinf.org.uk/intpred/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.