PROJECT SUMMARY/ABSTRACT Specific combinations of transcription factors (TFs) exhibit emergent properties when functioning together, enabling the generation of diverse cell types and behaviors. However, identifying which combinations regulate a behavior of interest requires overcoming a combinatorial explosion, as among the ~1,600 TFs in the human genome there are ~1.3 million possible pairs alone. This scaling challenge has forced past efforts at systematically mapping such genetic interactions (GIs) to rely on simple, parallelizable measures of phenotype such as growth rate. Each GI is then characterized only by a single number, obscuring the mechanistic or molecular basis for any particular interaction: put simply, there are many ways for cells to appear equally “unfit.” Finally, many human cell types are quiescent or post-mitotic, so that the growth-based measures of interaction that have been highly successful in model organisms such as yeast do not apply. Here we address these challenges by introducing a new, massively parallel method for studying GIs in human cells that combines rich phenotyping of single cells with an analytical framework for predicting which combinations are most informative to measure. We leverage the recently developed Perturb-seq screening technology, which allows pooled profiling of CRISPR-mediated genetic perturbations with single-cell RNA sequencing as the phenotypic readout. This approach allows us to overexpress many programmed combinations of TFs using CRISPR activation (CRISPRa) and obtain a direct readout of their transcriptional consequences. The resulting rich phenotypes yield insight into the biological origins of GIs, and can for example identify combinations of TFs that promote differentiation to diverse cell states. They also provide a critical “handle” to apply modern machine learning methods. Using techniques from the field of compressed sensing, we propose a predictive approach for searching combinatorial spaces of GIs that would be too large to profile exhaustively by any experimental technology. Since the transcriptome is a direct readout of TF function and TFs interact via specific mechanisms such as cooperative binding at target promoters, these large-scale experiments can also be used to study deeper questions on how GIs emerge mechanistically, and how neomorphic (i.e. entirely new or unpredictable) phenotypes are generated. Our research provides the first scalable method for simultaneously finding and characterizing GIs in any system, a technique for rapidly mapping the “levers” controlling cell fate in diverse models of development and disease, and a model for how machine learning can be used to design the large combinatorial genetics experiments made possible by Cas9.