Protein binding affinity prediction using support vector regression and interfecial features

In understanding biology at the molecular level, analysis of protein interactions and protein binding affinity is a challenge. It is an important problem in computational and structural biology. Experimental measurement of binding affinity in the wet-lab is expensive and time consuming. Therefore, machine learning approaches are widely used to predict protein interactions and binding affinities by learning from specific properties of existing complexes. In this work, we propose an innovative computational model to predict binding affinities and interaction based on sequence, structural and interface features of the interacting proteins that are robust to binding associated conformational changes. We modeled the prediction of binding affinity as classification and regression problem with least-squared and support vector regression models using structure and sequence features of proteins. Specifically, we have used the number and composition of interacting residues at protein complexes interface as features and sequence features. We evaluated the performance of our prediction models using Affinity Benchmark Dataset version 2.0 which contains a diverse set of both bound and unbound protein complex structures with known binding affinities. We evaluated our regression performance results with root mean square error (RMSE) as well as Spearman and Pearson's correlation coefficients using a leave-one-out cross-validation protocol. We evaluate classification results with AUC-ROC and AUC-PR Our results show that Support Vector Regression performs significantly better than other models with a Spearman Correlation coefficient of 0.58, Pearson Correlation score of 0.55 and RMSE of 2.41 using 3-mer and sequence feature. It is interesting to note that simple features based on 3-mer features and the properties of the interface of a protein complex are predictive of its binding affinity. These features, together with support vector regression achieve higher accuracy than existing sequence based methods.