2025/06/17 by Sandler, George, York, Ryan
#FOS: Biological sciences
paper · doi:10.57844/arcadia-25nt-guw3
Deep learning (DL) methods are becoming increasingly common in biological research. While powerful in some contexts, it's often unclear what biological patterns DL models end up learning and how much of an advantage they provide over simpler alternatives. Such questions can be probed most efficiently in highly distilled, simulated datasets, providing insight into the underlying behavior of DL models. Here we tackle this task in the context of epistatic interactions in genotype-to-phenotype mapping. We test the ability of a multilayer-perception (MLP) to beat conventional linear regression in three in silico experiments meant to probe the behaviour of DL across familiar quantitative genetics parameters space, namely numbers of QTLs, relative genetic variance components, and genetic correlations/pleiotropy among phenotypes. Our results help give us intuition about when and where applying DL is most likely to result in success in more complex, real-world biological datasets.