UC San Diego

Cancer, a complex and heterogeneous disease, continues to pose a significant challenge to global mortality. As such, there is a need for increasingly potent and precise therapies, developed rapidly and cost-effectively. Here, we describe a systematic method to uncover tumor-specific genetic co-dependencies and design small-molecule therapies that target those vulnerabilities using deep generative chemistry.

First, CRISPR-Cas9 perturbations are leveraged to profile a network of functional co-dependencies among cyclin-dependent kinases (CDK) genes and associated factors. In tandem, single-cell RNA sequencing is used to interrogate the transcriptomes of perturbed cells and investigate the potential mechanisms by which the observed synthetic lethalities arise. This study revealed a network of 12 synergistic interactions and 43 synthetic lethalities. The synthetic lethalities serve as direct target pairs for polypharmacology drug design.Second, a machine-learning model called CHEMIST is developed to generate de novo polypharmacology compounds that target synthetic lethalities in human cancer. Leveraging recent advances in deep generative chemistry and reinforcement learning, CHEMIST is designed to tune molecular structures to specific and potentially independent goals. CHEMIST is then utilized to design polypharmacology compounds against ten pairs of protein targets observed as synthetic-lethal in at least one human cancer context. Of these compound designs targeting the synthetic-lethal combination of MEK1 (mitogen-activated protein kinase kinase 1) and mTOR (mammalian target of rapamycin), 32 compounds were synthesized and validated in a human adenocarcinoma cell-line model. Together, these works present a framework for quickly designing small-molecule therapies that target tumor-specific vulnerabilities.