CRISPR-based genome editing technologies have enabled massively-parallel genomic screens, such as DepMap – a Broad Consortium effort to catalog gene knockouts in cancer cell lines. These projects find that the growth effects of a mutation depend heavily on the background genotype of a cell. Evolutionary theory has studied the effects of background genotype on mutations for generations and has uncovered general patterns across the tree of life These patterns found in evolving populations have culminated in a ‘Geometric Model’ of adaptation that has successfully predicted the effects of novel combinations of mutations in yeast and E. coli. This model could in principle be applied to DepMap and other massively-parallel genomic screens to learn genotype to phenotype to fitness mappings and potentially predict the evolution of a population. Fitting this model to large-scale real data, however, is challenging because the model infers a latent (hidden) space of phenotypes with mathematical symmetries which confuse regression methods. Here, we present a methodology for fitting a Geometric Model of adaptation to large-scale genomic screens that eliminates rotational, translational, and permutation symmetries in the inferred phenotype space and successfully reconstructs genotype to phenotype to fitness mappings of Liver cancer cell line knockout data. Thus, making comprehensive quantitative models of genotype to phenotype to fitness mappings possible in a multitude of diseases, which in turn will allow us to infer phenotypic complexity and predict treatment response.
