Solvent Accessible Surface Area-Based Hot-Spot Detection Methods for Protein–Protein and Protein–Nucleic Acid Interfaces

Abstract

Due to the importance of hot-spots (HS) detection and the efficiency of computational methodologies, several HS detecting approaches have been developed. The current paper presents new models to predict HS for protein-protein and protein-nucleic acid interactions with better statistics compared with the ones currently reported in literature. These models are based on solvent accessible surface area (SASA) and genetic conservation features subjected to simple Bayes networks (protein-protein systems) and a more complex multi-objective genetic algorithm-support vector machine algorithms (protein-nucleic acid systems). The best models for these interactions have been implemented in two free Web tools.

A Machine Learning Approach for Hot-Spot Detection at Protein-Protein Interfaces

“…For both sets, there is a natural expected tendency for a higher percentage of large hydrophobic or charged residues at the interfaces, in particular Tyr. Although different patterns could influence the training of a robust classifier, we have previously successfully constructed models that were bias-free for all different amino acids [14]. We randomly split this dataset (see for details Supplementary Information Table S1) into a training set consisting of 70% of data (382 mutations) and an independent test set (163 mutations, 30%).…”

“…It has become widely used in a variety of areas due to its reduced application time and high performance. Over the past few years, a few algorithms have been applied for the specific problem in this study: the detection of hot-spots at protein-protein interfaces [13,14,15,22,23,24,25,26,27,28,29,30,31,32,33,34,35]. …”

“…From a structural point of view, we compiled 12 previously-used different SASA descriptors for all interfacial residues [14,15]: (i) comp SASA i , the solvent accessible surface area of residue i in the complex form; (ii) mon SASA i , the residue SASA in the monomer form; (iii) ΔSASA i , the SASA difference upon complexation (Equation (1)); (iv) relSASA i , the ratio between ΔSASA for each residue and the mon SASA i value for the same residue (Equation (2)). A further four features ( comp/res SASA i , mon/res SASA i , Δ/res SASA i and rel/res SASA i ), defined by Equations (3)–(6), were determined applying amino acid standardization by dividing the previous features by the average protein res SASA r values as determined by Miller and colleagues [42,43], with r being the respective residue type.…”

“…However, due to the complexity and typical large size of protein-protein complexes, these methods are still computationally expensive. Recently, machine learning approaches trained on various features of experimentally-determined HS residues have been developed in order to predict HS in new protein complexes [6,12,13,14]. …”

“…In that study, we explored additional Machine Learning (ML) techniques, which led us to develop a more accurate and time-efficient HS detection methodology. This resulted in new HS predictor models for both protein-protein and protein-nucleic acid interactions, and we implemented the best performing models into two web tools [14]. …”