Computational and machine learning approaches for functional characterization of chemical compounds using chemical-genetic interactions

Abstract

Chemical-genetic interactions provide an invaluable source of information about the cellular functions perturbed by compounds. The chemical-genetic interaction profile of a compound against genome-wide mutant collections captures the effects of that compound on the fitness of each mutant in the collection. Since such profiles cover the entire genome, they can be used as a systematic, unbiased functional proxy for the effects of compounds on cells. Therefore, exploring the chemical-genetic interaction profiles of compounds can elucidate the modes of actions of compounds and their genetic targets. Using large libraries of chemical-genetic interaction profiles in Saccharomyces cerevisiae, we improved the functional prediction of compounds from chemical structures. We benchmarked a set of molecular fingerprints and similarity coefficients to find the pair with superior prediction power for connecting the compound structural space to functional space. We also developed a machine learning approach for predicting compound functional similarities from molecular fingerprints, which demonstrated substantial improvement over existing similarity measurements based on chemical structure. Moreover, using large libraries of multimodal chemical-genetic interaction profiles in S. cerevisiae, we developed two scalable computational pipelines for target identification of compounds. In one pipeline, we integrated chemical-genetic interactions from several modes to assign a unified target score to each compound–gene pair and prioritize the most promising candidates for further experimental studies. In another pipeline, we modeled the gene target of a compound as the rumor source spreading inhibitory signals across a genetic network. Our validation results confirmed novel targets for several studied compounds.