INTERPRETING GENERATIVE ADVERSARIAL NETWORKS TO INFER NATURAL SELECTION FROM GENETIC DATA

Area of Science:

• Population genetics
• Evolutionary biology
• Computational biology

Background:

• Detecting natural selection is crucial for understanding evolution in humans and other species.
• Current machine learning methods require computationally intensive simulations, often leading to mismatches between simulated and real data.
• Interpreting these models remains challenging, hindering understanding of selection drivers.

Approach:

• Developed a novel approach using a Generative Adversarial Network (GAN) for efficient detection of natural selection.
• The GAN comprises a generator for neutral data simulation and a discriminator (convolutional neural network) to distinguish real from simulated data.
• Fine-tuned the discriminator with minimal selection simulations to enhance its ability to identify regions under selection.

Key Points:

• The GAN approach requires significantly fewer selection simulations during training compared to existing methods.
• Demonstrated high power in detecting selection in simulations and identified known selected regions in human populations.
• Developed a method for interpreting the GAN's discriminator by analyzing hidden unit correlations with summary statistics.

Conclusions:

• This novel GAN-based method offers an efficient and powerful machine learning tool for detecting natural selection.
• The approach improves upon existing methods by reducing computational costs and enhancing interpretability.
• Provides a new avenue for advancing population genetic inference and understanding evolutionary processes.