Support Vector Rank Regression And Its Application In Optimizing Docking Scores

Support vector machine (SVM) is a statistical learning strategy that has numerous applications in a wide range of areas. Typically, SVM is implemented as classification algorithms or regression algorithms, which are not ideal for problems where we aim to score examples in anticipated order. This work introduces a new category of SVM algorithm, which we call support vector rank regression (SVRR) that learns from the relative order of some training examples and scores testing examples with expected rank. The utility of SVRR is demonstrated with its application to optimizing molecular docking scores. Based on three well-established datasets for docking evaluation, we showed that scores produced by our SVRR model could identify predicted ligand-protein binding conformations that are closer to the crystal ones. Our SVRR scores were also generally more reliable in compound screening, i.e. finding compounds in a library that can bind a specific protein. In computational biology, there are many other scenarios SVRR may find its application. The characteristics and potential variants of SVRR are discussed.