Combinatorial Cell State Engineering

Abstract Genome-wide screens in mammalian cells have emerged as a powerful tool for determining the relationship of individual genes to a chosen biological phenotype. However, biological systems often rely on the concerted action of multiple genes at once to elicit phenotypes. Nowhere is this more evident than in cellular differentiation, where cell state transitions often involve the modulation of 5-7 master regulatory factors. Consistent with this observation, successful efforts to reprogram cells, from Yamanaka on, have generally found that simultaneous expression of 3-5 transcription factors are needed to elicit cell state or type changes (similar to an “AND-gate- like” genetic circuit), and others have improved the efficiency or accuracy of these transitions by further perturbing other factors such as epigenetic remodelers. Given these observations, we posit that the ability to carry out highly combinatorial forward genetic screens for cell state phenotypes would produce a “sea change” in our ability to engineer cells with highly specific properties, transforming the quality of cells available for research and cell therapy applications. To this end, we propose an iterative platform that leverages a large multiplicity of perturbation (MOP) per cell, intelligent structuring of engineered perturbation libraries, and machine learning approaches to both identify combinations of perturbations most likely to elicit specific cellular phenotypes, and to engineer maximally informative new perturbation libraries. We have piloted this platform on a simple “toy model” wherein the simultaneous expression of 6 different proteins (across a total universe of 30 different potential factors) are required to elicit a phenotype. By overloading cells with ~14 perturbations per cell, structuring a library of ~80 perturbation combinations, then identifying further observations that would provide maximal information about the causative perturbation combination, we were able to confidently uncover this six- input “AND-gate” underlying state logic. While this initial ability to “solve” highly polygenic phenotypes is exciting, challenges to extending our platform to primary human cells include identification and minimization of dominant negative perturbations, identification of optimal MOP for each biological question, perfection of methods for high MOP of primary cells, exploration and optimization of the direction and mechanism of gene expression perturbation, and the engineering or selection of state changes sufficiently durable for therapeutic utility. We plan to initially apply this platform to the trans-differentiation of naive T cells into regulatory T cells and the generation of inexhaustible T-cells for cell therapies, with an eye toward establishing collaborations to deploy this platform to develop diverse cell types with regenerative or therapeutic value. In short, we posit that complex, therapeutically relevant phenotypes demand a polygenic design language that reflects the combinatorial vocabulary and grammar of human biology. We anticipate that our cell engineering platform will provide the first native implementation of this language.