Interpreting Supervised Machine Learning Inferences in Population Genomics Using Haplotype Matrix Permutations.

Area of Science:

• Population genomics
• Machine learning
• Bioinformatics

Background:

• Supervised machine learning, particularly convolutional neural networks (CNNs), are increasingly used for population genomics inference.
• A major limitation of these methods is their lack of interpretability, hindering biological insights and method development.

Purpose of the Study:

• To develop a systematic and interpretable framework for understanding the population genetics features driving machine learning model predictions.
• To assess the feature importance for existing CNNs used in population genomics.

Main Methods:

• Introduced a permutation-based approach to progressively disrupt population genetics features (linkage disequilibrium, haplotype structure, allele frequencies) in haplotype matrices.
• Measured performance degradation of CNNs after feature disruption to quantify feature importance.
• Applied the method to three published CNNs for positive selection and demographic history inference.

Main Results:

• The ImaGene CNN for positive selection heavily relies on haplotype structure and linkage disequilibrium.
• A demographic inference CNN primarily uses allele frequency information.
• The disc-pg-gan CNN achieved high accuracy with only allele counts, suggesting potential limitations in its learned features.

Conclusions:

• The developed permutation approach is a model-agnostic, biologically-motivated framework for interpreting haplotype matrix-based methods.
• Provides crucial insights into feature importance, guiding future method development and application in population genomics.
• Highlights variability in how different CNNs utilize population genetic information.