Protein-chemical interaction prediction via kernelized sparse learning SVM

Given the difficulty of experimental determination of drug-protein interactions, there is a significant motivation to develop effective in silico prediction methods that can provide both new predictions for experimental verification and supporting evidence for experimental results. Most recently, classification methods such as support vector machines (SVMs) have been applied to drug-target prediction. Unfortunately, these methods generally rely on measures of the maximum "local similarity" between two protein sequences, which could mask important drug-protein interaction information since drugs are much smaller molecules than proteins and drug-target binding regions must comprise only small local regions of the proteins. We therefore develop a novel sparse learning method that considers sets of short peptides. Our method integrates feature selection, multi-instance learning, and Gaussian kernelization into an L(1) norm support vector machine classifier. Experimental results show that it not only outperformed the previous methods but also pointed to an optimal subset of potential binding regions. Supplementary materials are available at "www.cs.ualberta.ca/~ys3/drug_target".