BindProfX: Assessing Mutation-Induced Binding Affinity Change by Protein Interface Profiles with Pseudo-Counts

Abstract

Understanding how gene-level mutations affect the binding affinity of protein–protein interactions is a key issue of protein engineering. Due to the complexity of the problem, using physical force field to predict the mutation-induced binding free-energy change remains challenging. In this work, we present a renewed approach to calculate the impact of gene mutations on the binding affinity through the structure-based profiling of protein–protein interfaces, where the binding free-energy change (ΔΔG) is counted as the logarithm of relative probability of mutant amino acids over wild-type ones in the interface alignment matrix; three pseudo-counts are introduced to alleviate the limit of the current interface library. Compared with a previous profile score that was based on the log-odds likelihood calculation, the correlation between predicted and experimental ΔΔG of single-site mutations is increased in this approach from 0.33 to 0.68. The structure-based profile score is found complementary to the physical potentials, where a linear combination of the profile score with the FoldX potential could increase the ΔΔG correlation from 0.46 to 0.74. It is also shown that the profile score is robust for counting the coupling effect of multiple individual mutations. For the mutations involving more than two mutation sites where the correlation between FoldX and experimental data vanishes, the profile-based calculation retains a strong correlation with the experimental measurements.

iSEE: Interface Structure, Evolution and Energy-based machine learning predictor of binding affinity changes upon mutations

“…• mCSM [11] and BindProfX [12], two machine learning based approaches, with mCSM using distance-specific atom-contacts (calculated from the wild-type structures only) and pharmacophore changes of the mutation site as features of Gaussian processes to predict ∆∆G, and BindProfX extracting evolutionary interface profiles from structural homologs and using FoldX energy terms to predict ∆∆G through a random forest model. iSEE compares favorably with the eight other predictors considered here over the independent NM test set with a RMSE of 1.37 kcal mol 8 FoldX (0.72) and ZeMu (0.70).…”

“…For the NM dataset, the predicted ∆∆G values of pred1 [9], pred2 [9], CC/PBSA [8], BeAtMuSiC [10] and FoldX [6] were directly extracted from Li et al [9] and those of ZeMu from Dourado's paper [7]. Predictions of mCSM [11] and BindProfX [12] for the NM and MDM2-p53 test datasets were directly obtained from their respective webservers. The default parameters of BindProfX were used except the "Score to use" which was set to "interface profile and physics potential" (the authors reported it to work best for single point mutations [12]).…”

“…Predictions of mCSM [11] and BindProfX [12] for the NM and MDM2-p53 test datasets were directly obtained from their respective webservers. The default parameters of BindProfX were used except the "Score to use" which was set to "interface profile and physics potential" (the authors reported it to work best for single point mutations [12]). The MDM2-p53 predictions for BeAtMuSiC and FoldX were obtained using the webserver and a local version of FoldX4.0, respectively.…”

“…It is therefore not surprising that the publication of the SKEMPI database [15] (currently the largest mutation ∆∆G dataset for protein-protein complexes containing 3047 mutations in 85 complexes) quickly promoted the emergence of several machine learning-based ∆∆G predictors [11][12][13][14]. However, the SKEMPI dataset is still rather limited in size and one has to be careful not to use too many features to train a model to avoid overfitting problems.…”

“…Interface conversation has been used in several of the best performing ∆∆G predictors [12,14]. However, since the conservation measure they used is structure-based, relying on the availability of structural homologs [12,14], the application and prediction of these ∆∆G predictors are largely limited by the availability and the number of such homologs. By contrast, conservation from Position Specific Scoring Matrix (PSSM) is sequence-based and thus better applicable.…”