For the past few years, brewer’s yeast (Saccharomyces cerevisiae) has been teaching a computer network how to look for new drugs. To be more precise, this powerful model organism is driving the development of an algorithm that researchers at Université de Montréal and McMaster University regard as the foundation for identifying novel antifungal compounds.

This work is rooted in a fresh approach to drug discovery, one that departs from the well-established model of searching for a “magic bullet” with precisely the right medicinal properties to neutralize the effects of harmful pathogens. Although such products have been found over the course of the last century, the quest for new ones represents a daunting trial-and-error exercise that even the largest of pharmaceutical firms now regard as being too expensive and time-consuming to be practical.

Instead, this quest is shifting to a new landscape: the intricate networks encoded by the genome. As we have unravelled  the complete genetic sequences of many different organisms — including ourselves — we have arrived at a network model for the way in which the complex array of proteins encoded by the genome governs physiology, from determining specific physical characteristics such as eye colour to guiding crucial processes such as growth and development. 

Whenever this network is disturbed, either from some internal mutation or interference by an external agent, organisms suffer. This is the case when pathogens interfere with our biochemical machinery but if the right combination of chemicals could be mustered to target these genetic interactions in the pathogens themselves, we could turn the tables on them. 

This principle became the foundation of ChemGRID, a publicly accessible portal of chemical-genetic interactions. This database was initiated about a decade ago by Université de Montréal systems biologist Mike Tyers to study the links between genotype and small molecule action. This approach entailed monitoring the genetic responses of brewer’s yeast — whose genome has been fully mapped — to an array of 4,915 compounds. This dataset of chemical-genetic interactions served as the input for machine learning algorithms to search for patterns of interactions that might reveal effective chemical combinations that inhibit yeast growth. 

The algorithm design and learning sequence was developed by McMaster postdoctoral fellows Michaela Spitzer and Jan Wildenhain, both data analysts specializing in bioinformatics who originally began their work with Tyers at the University of Edinburgh and now work in Gerard Wright’s group. In order to improve and assess the model’s predictive power, they narrowed the list of active compounds down to 1,221 candidate molecules that exhibited specific activity against particular genetic backgrounds. A subset of 128 precisely selected compounds was subsequently used to generate 8,128 combinations as a benchmark for algorithm performance. According to Wildenhain, at least 18 of these combinations proved to have potent activity against yeast pathogens that are harmful to human health. 

More generally, this study argues that such chemical-genetic datasets will prove a fertile new hunting ground for drug leads, but the sheer scale of these data matrices can only be effectively queried and modelled by machine intelligence rather than human intuition. And having tackled the field of fungal infection, the team would now like to apply this same machine learning strategy to chemical-genetic interactions found in human cells, which could shed entirely new light on potential cancer therapeutics. “You could come up with ways of targeting the specific mutations found in different cancer types,” Wildenhain says.