Artists can only make educated guesses about Neanderthals' appearance. But DNA sequences and machine learning algorithms could help researchers determine ancient hominin phenotypes with much greater clarity. Image credit: Christian Jegou/Science Source.
the last few decades, researchers have sequenced a handful of genomes from the bones of Neanderthals and Denisovans, who are among modern humans’ closest relatives. Researchers now want to know what this DNA reveals about how extinct hominins lived, what they looked like, and even their risk for certain diseases.
But deciphering the phenotypes of long-gone hominins is no easy feat. “Even in living humans, it’s hard to go from genetic information to a prediction of a trait or phenotype,” says John “Tony” Capra, an evolutionary the University of California, San Francisco. In a recent article in Annual Reviews, Capra and coauthors proposed that rather than jumping directly from genomic data to predictions about complex phenotypes, researchers should instead take a stepwise approach by, for example, using variation in genetic sequences to predict variation in protein architectures. Researchers could then predict how those proteins might change cells and so on—to ultimately hypothesize about variation in the traits those proteins affect. “We want to focus on the building blocks,” says lead author Colin Brand, a postdoctoral evolutionary genomicist at UCSF.
Most traits are controlled by hundreds or even thousands of genes. Any single genetic change therefore typically has a minor effect on the overall phenotype, such as, say, risk for certain diseases or behaviors, Capra says. Predicting how variation in one gene will scale up to affect complex phenotypes is therefore “really, really challenging,” he explains.
So he and Brand asked what they could confidently predict from genomic data. They knew that machine learning algorithms developed in the last several years have become adept at predicting molecular traits, such as gene expression and protein structure from DNA sequence variation in modern humans. The authors propose applying these same algorithms to ancient DNA—as old as 120,000 years in the case of a Neanderthal from Siberia.
All of these algorithms work roughly the same way: They detect patterns in DNA sequence variation, and then correlate those patterns to molecular traits, such as the level of a specific gene’s expression in a given tissue or the likelihood that a protein has a certain folded conformation. For example, Capra says, imagine a gene in which an adenosine variant at one nucleotide position was associated with higher expression of the gene. The machine learning models would detect this. If past studies—say, in mice—had linked high expression of the gene to risk for heart disease, then researchers might hypothesize that people with the adenosine variant at that nucleotide position, might be at particularly elevated cardiac risk. “And what is really powerful about the machine learning approaches,” Capra says, “is that they can detect how many different genetic variants with small effects combine to have larger effects.”
Human geneticist Rajiv McCoy at Johns Hopkins University in Baltimore, MD, who was not involved in this latest work, notes that some labs are already using machine learning algorithms to make stepwise inroads, predicting phenotypes from DNA. Even so, he lauds Brand and Capra’s comprehensive review, noting that it’s the most direct argument he’s heard for studying molecular phenotypes, such as gene expression in ancient hominins, because they have simpler genetic bases than complex phenotypes such as disease risk, anatomy, or behavior. The approach is at the cutting edge of ancient hominid genetics, he says.
Brand expects that the approach could be useful across many groups besides hominins and other primates. One promising avenue is predicting the phenotype of ancient mammals, Brand says. Machine learning algorithms could similarly detect genetic variation and compare it to the outcome of variation in living mammals—even with relatively few genome samples from extinct species. “With ancient samples, all that’s left is little bits of DNA that have to be strung together,” Capra says. “That’s why these algorithms we’re applying are so essential. They let us go from just DNA information to something more...
Enjoyed this article? Stay informed by joining our newsletter!
You must be logged in to post a comment.