Interpreting Convolutional Neural Networks in Population Genetics.

Some CNN architectures in population genetics implicitly learn summary statistics like pairwise heterozygosity, with long-range linkage disequilibrium being efficiently approximated.

• Why it matters: Understanding what CNNs learn is crucial because they can outperform traditional methods but are often difficult to interpret, limiting their broader application and trustworthiness in population genetics.
• What they did: The study analyzed CNNs trained for data discrimination and detecting selective sweeps, using correlation analysis, SHAP values, dimensionality reduction, and interpretable models to examine learned features.
• The result: Findings suggest CNNs can implicitly compute certain summary statistics more efficiently than traditional methods, enhancing the interpretability of deep learning models in evolutionary parameter inference.