Computational prediction of protein interfaces: A review of data driven methods

Abstract

Reliably pinpointing which specific amino acid residues form the interface(s) between a protein and its binding partner(s) is critical for understanding the structural and physicochemical determinants of protein recognition and binding affinity, and has wide applications in modeling and validating protein interactions predicted by high-throughput methods, in engineering proteins, and in prioritizing drug targets. Here, we review the basic concepts, principles and recent advances in computational approaches to the analysis and prediction of protein-protein interfaces. We point out caveats for objectively evaluating interface predictors, and discuss various applications of data-driven interface predictors for improving energy model-driven protein-protein docking. Finally, we stress the importance of exploiting binding partner information in reliably predicting interfaces and highlight recent advances in this emerging direction.

Keywords: Cross validation on instance level; Cross validation on protein level; Docking; Evaluation caveats; Machine learning; Partner-specific interface prediction; Protein–protein interaction.

Figure 1. A surface patch and a sequence window

A) A surface patch defined by a target residue (blue) and its spatial neighboring residues (magenta) that fall within a virtual sphere of diameter, d, centered on the target residue. B) A sequence window centered on a target residue (purple).

Figure 2. Partner-specific interface predictors outperform non-partner-specific predictors

(A) A comparison of the top 20 predicted interfacial residues for a complex of Acetylcholinesterase (blue ribbons) and Toxin F-VII Fasciculin-2 (red ribbons) (PDB ID: 1MAH) by the partner-specific method, PPIPP (Ahmad & Mizuguchi, 2011), and the corresponding non-partner-specific version. By including partner information, PPIPP is able to predict interfacial residues (green) clustering around the interaction location specific to the binding partner whereas those predicted by the non-partner-specific method (red) are scattered over the surface of query protein. Figure credit: (Ahmad & Mizuguchi, 2011)(B) Prediction performance comparisons over a set of 123 non-redundant protein-protein complexes in Docking Benchmark 3.0 (Hwang, Pierce, Mintseris, Janin, & Weng, 2008). We compared two partner-specific predictors, PAIRpred (Afsar Minhas et al., 2013) and PPiPP (a sequence-based predictor) (Ahmad & Mizuguchi, 2011), with two non-partner-specific machine learning predictors: PSIVER, a sequence-based predictor (Murakami & Mizuguchi, 2010) and SPPIDER, a structure-based predictor (Porollo & Meller, 2007). With partner information, PAIRpred and PPiPP outperform the two predictors that do not consider partner information when making predictions, improving Area Under Curves (AUCs) from 0.63 (PSIVER) and 0.58 (SPPIDER) to 0.73 (PPiPP) and 0.89 (PAIRpred). AUC values are extracted from (Afsar Minhas et al., 2013; Ahmad & Mizuguchi, 2011).

Figure 3. Protein-protein docking and its two major challenges

Figure 4. Machine learning toward improved 3D protein interaction prediction