Detecting Positive Selection by Modeling Structure Within Images of Genetic Variation

A major challenge in population genomics is accurately identifying and characterizing natural selection from genomic data. The wide availability of dense whole-genome datasets has enabled researchers to analyze and localize genetic variation within populations. Powerful supervised machine learning methods allow researchers to extract spatial information about genetic variation across the genome and identify traces of natural selection. While convolutional neural networks capture correlations among neighboring features, design choices such as heavy-pooling or limited receptive fields can lead to loss of fine-grained spatial resolution. Extensions like dilated convolutions or attention mechanisms mitigate this issue of loss of spatial resolution but at the cost of increased architectural complexity and parameter count when capturing correlations at different scales. In contrast, trend filtering directly models the autocovariation of neighboring features, ensuring that spatial relationships remain intact without any architectural extensions. When integrated into a classical machine learning model, such as a support vector machine, trend filtering offers a natural framework to create powerful predictive models while retaining the spatial integrity of the input. Here, we introduce SKINET, which employs a novel trend filter kernel within a support vector machine framework and apply it to the task of detecting and characterizing regions affected by positive natural selection. Specifically, SKINET not only distinguishes regions under positive natural selection from neutrally evolving regions but also functions in a regression framework to estimate associated adaptive parameters. Moreover, applying SKINET to empirical human genome variation identifies adaptive candidate genes consistent with previous findings while also uncovering novel adaptation targets, such as FAM177A1, that are linked to cancer.