Cross-target view to feature selection: identification of molecular interaction features in ligand-target space

There is growing interest in computational chemogenomics, which aims to identify all possible ligands of all target families using in silico prediction models. In particular, kernel methods provide a means of integrating compounds and proteins in a principled manner and enable the exploration of ligand-target binding on a genomic scale. To better understand the link between ligands and targets, it is of fundamental interest to identify molecular interaction features that contribute to prediction of ligand-target binding. To this end, we describe a feature selection approach based on kernel dimensionality reduction (KDR) that works in a ligand-target space defined by kernels. We further propose an efficient algorithm to overcome a computational bottleneck and thereby provide a useful general approach to feature selection for chemogenomics. Our experiment on cytochrome P450 (CYP) enzymes has shown that the algorithm is capable of identifying predictive features, as well as prioritizing features that are indicative of ligand preference for a given target family. We further illustrate its applicability on the mutation data of HIV protease by identifying influential mutated positions within protease variants. These results suggest that our approach has the potential to uncover the molecular basis for ligand selectivity and off-target effects.