Deep-Learning Used to Identify Contribution of Noncoding Mutations to Autism

Leveraging artificial intelligence techniques, researchers have demonstrated that mutations in junk DNA can cause autism. The study, published yesterday in Nature Genetics, is the first to functionally link such mutations to the neurodevelopmental condition, according to a team led by researchers at the Flatiron Institute.

Machine learning was used to analyze the whole genomes of 1,790 individuals with autism and their unaffected parents and siblings. The analysis predicted the ramifications of genetic mutations in junk DNA. The number of autism cases linked to the noncoding mutations was comparable to the number of cases linked to protein-coding mutations that disable gene function.

Uncovering which noncoding mutations may cause autism is tricky. A single individual may have dozens of noncoding mutations, most of which will be unique to the individual. This makes the traditional approach of identifying common mutations among affected populations nonviable. Troyanskaya and her colleagues took a new approach. They trained a machine learning model to predict how a given sequence would affect gene expression.

"This is a shift in thinking about genetic studies that we're introducing with this analysis," says Chandra Theesfeld, a research scientist in Troyanskaya's lab. "In addition to scientists studying shared genetic mutations across large groups of individuals, here we're applying a set of smart, sophisticated tools that tell us what any specific mutation is going to do, even those that are rare or never observed before."

The researchers studied the genetic basis of autism by applying the machine learning model to genetic data in the Simons Simplex, which contains the whole genomes of nearly 2,000 'quartets' made up of a child with autism, an unaffected sibling, and their unaffected parents. The researchers used their model to predict the impact of non-inherited, noncoding mutations in each child with autism. They then compared those predictions with the effects of the same, unmutated strand in the child's unaffected sibling.

Troyanskaya says she and her colleagues will continue improving and expanding their method. Ultimately, she hopes the work will improve how genetic data are used for diagnosing and treating diseases and disorders. "Right now, 98 percent of the genome is usually being thrown away," she says. "Our work allows you to think about what we can do with the 98 percent."