Genetic screens in physiological contexts nominate metabolic pathways for deep learning-assisted drug discovery

Acquisition of host metabolites is a defining feature of parasitism, yet culture systems are typically replete with excess nutrients that may obscure metabolic relationships between parasites and hosts. We performed two genetic screens to identify metabolic restrictions that occur during infection. First, we conducted a genome-wide screen of Toxoplasma in mice, revealing dozens of metabolic pathways that are required for animal infection. We next assessed which of these pathways could be studied in culture by performing a targeted screen of parasites grown under physiological media conditions. Among the top hits was GTP Cyclohydrolase I (GCH), an enzyme responsible for the first step in the biosynthesis of two metabolites, biopterin and folate. In addition to being synthesized by parasites, these metabolites can be salvaged from the host. We demonstrate that parasites adaptively rely on salvage when host folate levels are artificially elevated by standard culture conditions. As the target of anti-parasitic drugs like pyrimethamine, folate metabolism has long been appreciated as critical for parasite replication. However, anti-folate drugs have traditionally targeted enzymes downstream of the convergence of folate biosynthesis and salvage. Our work demonstrates that targeting biosynthesis alone is sufficient to restrict parasite growth under physiological conditions, nominating new enzymes as putative anti-parasitic targets. As a proof-of-concept, we demonstrate that an inhibitor of mammalian GCH can be repurposed to inhibit parasite growth. We have additionally devised a screening platform to identify compounds that more potently inhibit GCH. Leveraging conservation in bacteria, we created E. coli strains in which bacterial GCH is replaced with parasite or human homologs, enabling rapid screening of large chemical libraries through simple measurements of bacterial growth. An initial screen of 2500 compounds revealed dozens of molecules that selectively inhibit parasite GCH. In ongoing work, we are combining the graph neural network Chemprop and large language models (LLM) to learn the molecular features that enable preferential binding of a compound to parasite GCH, while reducing inhibition of the human homolog. In total, this work demonstrates how examining parasite metabolism in a physiological context can reveal previously overlooked druggable targets, and suggests a new methodology for high-throughput drug screening via bacterial complementation.