Modularity, Limited Sparsity, and Extensive Pleiotropy in a Genotype–Phenotype–Fitness Map of Drug Resistance

Understanding which phenotypic effects of mutation impact fitness is a central question in evolutionary biology. Yet genotype–phenotype relationships are not bipartite; a single mutation can influence hundreds of molecular-level phenotypes. This complexity poses a major challenge for predicting evolutionary outcomes, such as the relative fitness advantage of one mutant over another. Despite this, genotype–fitness relationships can sometimes be approximated by low-dimensional models that predict the fitness of diverse mutants in novel environments. But are these successes idiosyncratic to a few datasets, or do they reveal general features of genotype–phenotype maps? Here, we highlight the generality of this result and its limits in the context of drug resistance. We apply Singular Value Decomposition (SVD) to a dataset of 800 yeast mutants evolved in various concentrations and combinations of antifungals. This analysis identifies a small number of phenotypic dimensions that predict mutant fitness with up to 70% accuracy in unseen drug combinations. Predictive power plateaus rapidly with the addition of new training environments or model components, indicating that low dimensionality is not an artifact of sampling, but instead represents the funneling of mutational effects through a small number of major phenotypic axes. Our results support longstanding ideas that modularity and canalization constrain phenotypic variation, making evolutionary outcomes more predictable. Beyond modularity, our dataset also sheds light on pleiotropy. Unlike other systems where mutations affect only a subset of dimensions, we find that most mutations perturb most inferred phenotypic axes. This lack of sparsity suggests that drug environments may provoke unusually pleiotropic responses, perhaps because cells rarely encounter these stresses in nature. Natural environments, such as nutrient limitation, may instead reveal a sparser wiring where mutations act through more specific phenotypic routes. Finally, our results highlight that low-dimensional linear models do not capture everything: some drug combinations remain difficult to predict. In these environments, subsets of mutants often show nonmonotonic fitness responses to increasing drug concentrations, and their prediction accuracy does not improve simply by adding more training data. These limitations highlight the need for nonlinear approaches and models that integrate high-dimensional phenotypic measurements to fully uncover how mutations shape cellular processes and fitness.