ConfuseNN: Interpreting convolutional neural network inferences in population genomics with data shuffling

Area of Science:

• Population Genetics
• Computational Biology
• Genomics

Background:

• Convolutional neural networks (CNNs) are increasingly used for population genomic inference.
• However, the specific genomic features driving CNN performance are not well understood.

Purpose of the Study:

• To introduce ConfuseNN, a novel method for interpreting CNN performance in population genetics.
• To systematically evaluate the contribution of different population genetic features to CNN accuracy.

Main Methods:

• ConfuseNN systematically shuffles input haplotype matrices to disrupt specific population genetic features.
• The method sequentially removes signals like linkage disequilibrium and allele frequency to assess their impact on CNN performance.
• Applied ConfuseNN to existing CNNs for demographic history and selection inference.

Main Results:

• Confirmed the importance of specific genomic features for CNN performance in population genetic inference.
• Identified limitations in current CNN architectures and the design of simulated training/testing data.
• Demonstrated ConfuseNN's utility in evaluating CNNs across various population genetics tasks.

Conclusions:

• ConfuseNN offers a biologically motivated framework for interpreting CNNs in population genetics.
• This approach helps bridge the gap between deep learning methods and traditional population genetic theory.
• Facilitates a deeper understanding of how CNNs learn from genomic data.