INTERPRETING CONVOLUTIONAL NEURAL NETWORKS IN POPULATION GENETICS.

Area of Science:

• Population genetics
• Computational biology
• Machine learning

Background:

• Machine learning, particularly Convolutional Neural Networks (CNNs), offers powerful alternatives to traditional methods in population genetics.
• CNNs excel at tasks like inferring natural selection and recombination rates, but their interpretability remains a challenge.
• Understanding what CNNs learn is crucial for trusting their predictions in evolutionary studies.

Purpose of the Study:

• To investigate the interpretability of CNNs used in population genetics.
• To determine what evolutionary parameters and summary statistics CNNs learn implicitly.
• To clarify the relationship between CNN architecture, learned features, and traditional population genetic statistics.

Main Methods:

• Analyzed CNNs from pg-GAN and selective sweep detection models.
• Computed correlations between learned network features and traditional summary statistics.
• Assessed predictability of summary statistics from learned features using SHAP values and dimensionality reduction.
• Built interpretable models using decision trees and random forests.

Main Results:

• Some CNN architectures implicitly compute pairwise heterozygosity.
• The site frequency spectrum is less represented in learned CNN features compared to other statistics.
• CNNs effectively approximate long-range linkage disequilibrium, potentially more efficiently than traditional methods.
• Learned features show strong correlations with specific evolutionary parameters.

Conclusions:

• This work enhances the interpretability of deep learning in population genetics.
• CNNs can learn and approximate complex evolutionary statistics, offering computational advantages.
• Clarifies the link between CNN architecture, learned features, and established population genetic inference methods.