Protein-small Molecule Binding Site Prediction Based On Molecular Docking And Local Features

Protein-small molecule binding site prediction is a critical scientific question.It has an important significance on many issues,including structure-based drug design,elucidation of protein function,and molecular docking.After many years of development,there are various methods for predicting binding sites,including geometric feature-based prediction methods,energy-based prediction methods,template-based prediction methods,knowledge-based prediction methods,conservation-based prediction methods and so on.However,the situation of protein-small molecule binding sites is very complicated.Most existing software currently describe and predict the protein binding site from one or a few angles,which leads to low accuracy.In this paper we propose a new algorithm DFSite,which is an algorithm that combines molecular docking and local features to predict protein-small molecule binding sites.In this algorithm,we describe the characteristics of the binding site from the amino acid level,the atomic level and the binding site's own level,and use the linear regression algorithm to train a new set of scoring function.We combine this new scoring function with MDock molecular docking software to predict binding sites.We compared our method to the existing proteinsmall molecule binding site prediction method and found that our method surpasses most existing methods,including the widely used independent algorithm Fpocket,automatic binding site prediction algorithm DoGSite,effective detection algorithm LIGSITE for protein and small molecule binding sites,effectively integrates multiple binding site prediction tools MetaPocket 2.0,and based on 3D convolution Neural network to predict the binding site algorithm DeepSite.The DFSite algorithm is a fast algorithm that does not depend on any a priori binding site information.It increases the accuracy of the algorithm by combining molecular docking with a new scoring function.The test results show that our method is one of the two best ones among the tested binding site prediction algorithms.