#37 · Genomics

Learned and sparse models infer nonlinear fitness landscapes and population dynamics

The three items share a concern with inferring structure from data in evolutionary and population models rather than specifying it a priori. A convolutional neural network trained in population genetics is interrogated with SHAP to determine which learned features track summary statistics such as heterozygosity and linkage disequilibrium, offering a route to interpreting otherwise opaque predictors. On the GFP random mutagenesis fitness landscape, Power Transforms are tested as a means of addressing nonlinearity; Box-Cox reduces local pairwise epistasis by 19.4%, suggesting that the measurement scale on which fitness is expressed alters the apparent epistatic structure. The published WSINDy extension learns structured population models with heterogeneous dynamics and boundary processes. What is new is the joint emphasis on interpretability and on measurement scale as determinants of inferred nonlinearity, though the items address these concerns to differing degrees.