Molecular descriptors selection and machine learning approaches in protein-ligand binding affinity with applications to molecular docking

In this paper, we propose algorithms for biomolecular docking sites selection problem by various machine learning approaches with selective features reduction. The proposed method can reduce the number of various amino acid features before constructing machine learning prediction models. Given frame boxes with features, the proposed method analyzes the important features by correlation coefficients to LE values. The algorithm ranks these possible candidate locations on the receptor before launching AutoDock. Given a small molecular, namely ligand, it is a time-consuming task to compute the molecular docking against a large, relatively stationary molecule, or receptor. Our methods divide the surface area of receptor to several subspaces and evaluate these subspaces before choosing the promising subspaces to speed up the molecular docking simulation. The method is implemented upon the widely employed automated molecular docking simulation software package, AutoDock. The paper examines three different machine learning prediction models including the support vector machines (LIBSVM), deep neural networks (H2O), and the logistic regression model (AOD). The proposed affinity estimation algorithm, incorporated with a ligand-specific SVM prediction model, achieves about 4 folds faster comparing with original Autodock searching the whole surface of the receptor with similar binding energy score (LE, lowest engery) measurement. Furthermore, the proposed method can be easily parallelized in the implementation. Hadoop MapReduce frameworks are used in our experiments to parallelize the underlying massive computation works corresponding to ligand-receptor pairs examined under the experiment.