Machine learning in evolutionary studies comes of age

Viviane Callier; PNAS April 19, 2022 | 119 (17) e2205058119

Figure: A variety of new machine learning methods are shedding light on longstanding questions about the forces that shape genomes over time. Image credit: Shutterstock/alionaprof.

Whether flowers, mosquitoes, or humans, all organisms tend to interact and reproduce with others of their ilk who are close by. In principle, that means that their genetics can reveal not only their ancestry but also their geography. Taking advantage of this simple insight, evolutionary geneticists recently showed that it was possible to create models, based on scanning thousands of genomes from mosquitoes to elephants to humans (Fig. 1), that match individual genomes to spatial locations and therefore predict where a given individual animal was born (1, 2). The approach has practical import: It could be applied in the context of ecology and conservation by, for example, tracing the origins of elephant tusks and rare woods.

It’s one of a variety of new machine learning methods that are finding applications in the field, shedding light on longstanding questions about the forces that shape genomes, such as selection and genetic drift. These approaches have already demonstrated the potential to deal with the messiness of real data in ways that formal population genetic theory cannot. In the early 2000s, researchers who were “thinking a little bit ahead,” says evolutionary geneticist Andrew Kern at the University of Oregon in Eugene, “saw that there was going to be this collision happening, where we have this giant corpus of mathematical theory, and all of a sudden reality was going to intrude on our models in the form of all the genomic data that we’re about to collect.” What the field needed, he adds, were “more ways to bridge that gap between theory and biology.” Evolutionary biologists are starting to do exactly that.

See: https://www.pnas.org/doi/10.1073/pnas.2205058119