Automatic inference of demographic parameters using generative adversarial networks

Abstract

Simulation is a key component of population genetics. It helps to train our intuition and is important for the development, testing and comparison of inference methods. Because population genetic models such as the ancestral recombination and selection graphs (Griffiths & Marjoram, 1997;Neuhauser & Krone, 1997) are computationally intractable for inference but relatively easy to simulate, simulations are also heavily used for parameter inference.Approximate Bayesian Computation (ABC; Beaumont et al., 2002) is a widely used example. Regardless of the application, the goal is to simulate data that is 'realistic' in the sense that it resembles real data from the population(s) of interest. Typically this is done by fixing some parameters that are fairly well-known, then choosing other parameters to match some property of the real data, usually based on summary statistics. However, this involves a potential loss of information in the reduction in summary statistics and then an implicit weighting on the relative importance of different summary statistics. Often, parameters that create simulations that match one type of summary statistic (e.g. the site frequency spectrum) do not match others (e.g. linkage disequilibrium patterns; Beichman et al., 2017). Here, we present a novel parameter learning approach using Generative Adversarial Networks (GANs). Our approach creates both realistic simulated data and a quantitative way of determining

Detecting adaptive introgression in human evolution using convolutional neural networks

“…Progress in this regard could involve the use of generative adversarial networks (GANs), which appears to be a fruitful way to address this. Indeed, recent work suggests that one can train a GAN to learn to generate realistic population genomic data for any population ( Wang et al, 2020 ).…”

Machine learning in molecular ecology

“…The GAN is trained until the generator produces realistic data, and the evolutionary parameters that produce these realistic data are then inferred to be the true evolutionary parameters. Wang et al (2021) demonstrate that this approach can accurately estimate parameters under a two-population isolation-with-migration model using simulation studies. They then apply pg-gan to estimate parameters across several human populations and demonstrate the power and flexibility of this approach.…”

“…While the previous three papers use deep learning algorithms trained on data sets generated under a wide range of parameters in a manner analogous to many ABC algorithms, Wang et al (2021) explore an approach that permits a heuristic exploration of parameter space. In their contribution, they introduce a method (pg-gan)…”

“…The ultimate goal of a GAN is to generate fake data that are indistinguishable from real data. By using a parameterized evolutionary model as a generator, Wang et al (2021) are able to adapt GANs to infer evolutionary parameters. The GAN is trained until the generator produces realistic data, and the evolutionary parameters that produce these realistic data are then inferred to be the true evolutionary parameters.…”

slimr: An R package for integrating data and tailor-made population genomic simulations over space and time

“…However, utilising flexible simulations requires exploration of large parameter space, which often generates large amounts of data that need sophisticated computational tools to unpack, interrogate and synthesize. Likewise, using simulations to model empirical data is an emerging field because it allows researchers to deal with complex situations where it is difficult to obtain a closed likelihood (Beaumont, Zhang, & Balding, 2002; Brehmer, Louppe, Pavez, & Cranmer, 2020; Cranmer, Brehmer, & Louppe, 2020; Marjoram, Molitor, Plagnol, & Tavare, 2003; Sisson, 2018; Torada et al, 2019; Wang et al, 2020). To facilitate more rapid and seamless interrogation and synthesis between empirical data and population genetics simulation, we present (https://rdinnager.github.io/slimr/).…”