Abstract
Chemogenomics has emerged as an interdisciplinary field that aims to ultimately identify all possible ligands of all target families in a systematic manner. An ever-increasing need to explore the vast space of both ligands and targets has recently triggered the development of novel computational techniques for chemogenomics, which have the potential to play a crucial role in drug discovery. Among others, a kernel-based machine learning approach has attracted increasing attention. Here, we explore the applicability of several ligand-target kernels by extensively evaluating the prediction performance of ligand-target interactions on five target families, and reveal how different combinations of ligand kernels and protein kernels affect the performance and also how the performance varies between the target families.