Interpreting generative adversarial networks to infer natural selection from genetic data

Area of Science:

• Population Genetics
• Evolutionary Biology
• Machine Learning

Background:

• Detecting natural selection is crucial in population genetics, but current machine learning methods rely on computationally intensive simulations.
• Realistic selection simulations are slow and explore a vast parameter space, hindering efficient analysis.
• Interpreting neural network models for evolutionary processes remains challenging.

Purpose of the Study:

• To develop a computationally efficient machine learning approach for detecting natural selection and local evolutionary processes.
• To overcome the limitations of slow, simulation-heavy methods in population genetics.
• To improve the interpretability of machine learning models in evolutionary studies.

Main Methods:

• Utilized a generative adversarial network (GAN) framework trained on realistic neutral data.
• Employed a generator for simulating neutral demographic data and a discriminator (convolutional neural network) for distinguishing real from simulated data.
• Fine-tuned the discriminator with a small set of non-neutral simulations to specifically identify selection modes.

Main Results:

• The developed approach demonstrates high power in detecting various forms of selection within simulated data.
• Identified genomic regions under positive selection in three human populations, consistent with established population genetic methods.
• Successfully interpreted trained neural networks by clustering discriminator hidden units based on summary statistics.

Conclusions:

• This novel GAN-based method offers an efficient and interpretable alternative for detecting natural selection in population genetics.
• The approach effectively identifies targets of selection by leveraging deviations from neutral demographic models.
• Future work can build upon this interpretable framework for deeper insights into evolutionary processes.