Pinpointing where and how the human genome is evolving can be like hunting for a needle in a haystack. Each person’s genome contains three billion building blocks called nucleotides, and researchers must compile data from thousands of people to discover patterns that signal how genes have been shaped by evolutionary pressures.

To find these patterns, a growing number of geneticists are turning to a form of machine learning called deep learning. Proponents of the approach say that deep-learning algorithms incorporate fewer explicit assumptions about what the genetic signatures of natural selection should look like than do conventional statistical methods.

“Machine learning is automating the ability to make evolutionary inferences,” says Andrew Kern, a population geneticist at the University of Oregon in Portland. “There is no question that it is moving things forward.”

One deep-learning tool called ‘DeepSweep’, developed by researchers at the Broad Institute of MIT and Harvard in Cambridge, Massachusetts, has flagged 20,000 single nucleotides for further study. Some or all of these simple mutations may have helped humans survive disease, drought or what Charles Darwin called the “conditions of life”, researchers reported last month at the annual meeting of the American Society of Human Genetics in San Diego, California.

Since the 1970s, geneticists have created mathematical models to describe the fingerprint of natural selection in DNA. If a mutation arises that renders a person better able to survive and produce offspring than their neighbours, the percentage of the population with that gene variant will grow over time.

One example is the mutation that gives many adults the ability to drink cow’s milk. It enables the body to produce lactase, an enzyme that digests the sugar in milk, into adulthood. By analysing human genomes with statistical methods, researchers discovered that the mutation spread rapidly through communities in Europe thousands of years ago — presumably because nutrients in cow’s milk helped people to produce healthy children ^{1 , 2} . Today, nearly 80% of people of European descent carry this variant.

Training camp

Yet geneticists have struggled to identify, and confirm, other specific snippets of the genome that spread throughout populations because they provided an adaptive edge. Deep learning excels at just this sort of task: discovering subtle patterns hidden in large amounts of data .

But there is a catch. Deep-learning algorithms often learn to classify information after being trained by exposure to real data; Facebook, for example, primes algorithms to recognize faces on the basis of pictures that people have already labelled. Because geneticists don’t yet know which parts of the genome are being shaped by natural selection, they must train their deep-learning algorithms on simulated data.

Generating that simulated data requires researchers to posit what the signature of natural selection looks like, says Sohini Ramachandran, a population geneticist at Brown University in Providence, Rhode Island. “We don’t have ground truth data, so the worry is that we may not be simulating properly.”

And because deep-learning algorithms operate as black boxes , it’s hard to know what criteria they use to identify patterns in data, says Philipp Messer, a population geneticist at Cornell University in Ithaca, New York. “If the simulation is wrong, it’s not clear what the response means,” he adds.

Researchers who use deep-learning algorithms do have some ways to peek into the black box. DeepSweep's creators trained their algorithm on signatures of natural selection that they inserted into simulated genomes. When they tested it on real human-genome data, the algorithm zeroed in on the lactase mutations that allow adults to drink milk. That bolstered the team's confidence in the tool, says Joseph Vitti, a computational geneticist at the Broad Institute.

The researchers then sifted through data from the 1000 Genomes Project — an international initiative that sequenced DNA from 2,504 people around the world — using a statistical method to identify regions that might be under evolutionary pressure. These amounted to about one-third of the three billion building blocks that comprise the human genome. Next, DeepSweep assessed each region. By the end of the analysis, it had delivered a list of 20,000 single mutations to explore.

In the coming months, Vitti and his colleagues will investigate what these mutations do by editing them in the DNA of living cells, and comparing what happens when they are there with when they are not.

On the hunt

Several other researchers are training deep-learning algorithms to search for signs of adaptation in genomes. A deep-learning model developed by Andrew Kern, a population geneticist at the University of Oregon in Eugene, suggests that at first, most mutations in humans are neither beneficial nor harmful ³ . Rather, they seem to drift along in populations, increasing natural genetic variability, and only become more frequent when a change in the environment gives people who possess the mutation an evolutionary edge.

In February, Ramachandran and her colleagues reported on a deep-learning algorithm they developed called SWIF(r) ⁴ . When they applied it to the genomes of 45 members of the Khomani San ethnic group from southern Africa, it flagged variations near genes associated with metabolism. The researchers speculate that the changes may have occurred thousands of years ago and that they helped members of the group to store fat during unreliable food cycles.

The effects of the mutations still need to be tested. But as with the variants identified by DeepSweep, the candidates singled out by SWIF(r) provide researchers with a place to start.

“These are incredibly powerful methods for looking for the signals of natural selection,” says Pardis Sabeti, a computational geneticist at the Broad Institute, and Vitti’s PhD supervisor. “Some people didn’t think you could pinpoint variants when I started. Some thought it was impossible.”