Using artificial intelligence, a Princeton University-led team has decoded the functional impact of such mutations in people with autism. The researchers believe this powerful method is generally applicable to discovering such genetic contributions to any disease.
The researchers analyzed the genomes of 1,790 families in which one child has autism spectrum disorder but other members do not. The method sorted among 120,000 mutations to find those that affect the behavior of genes in people with autism. Although the results do not reveal exact causes of cases of autism, they reveal thousands of possible contributors for researchers to study.
Much previous research has focused on identifying mutations in genes themselves. Genes are essentially instructions for making the many proteins that build and control the body. Mutations in genes result in mutated proteins whose functions are disrupted. Other types of mutations, however, disrupt how genes are regulated. Mutations in these areas affect not what genes make but when and how much they make.
Until now, it was not possible to look across the entire genome for snippets of DNA that regulate genes and to predict how mutations in this regulatory DNA are likely to contribute to complex disease, the researchers said. This study is the first proof that mutations in regulatory DNA can cause a complex disease.
In their new finding, the research team offers a method to make sense of this vast array of genomic data. The system uses an artificial intelligence technique called deep learning in which an algorithm performs successive layers of analysis to learn about patterns that would otherwise be impossible to discern. In this case, the algorithm teaches itself how to identify biologically relevant sections of DNA and predicts whether those snippets play a role in any of more than 2,000 protein interactions that are known to affect the regulation of genes. The system also predicts whether disrupting a single pair of DNA units would have a substantial effect on those protein interactions.
Prior to this computational achievement, the conventional way to glean such information would be painstaking laboratory experiments on each sequence and each possible mutation in that sequence. This number of possible functions and mutations is too big to contemplate — an experimental approach would require testing each mutation against more than 2,000 types of protein interactions and repeating those experiments over and over across tissues and cell types, amounting to hundreds of millions of experiments. Other research groups have sought to accelerate this discovery by applying machine learning to targeted sections of DNA but had not achieved the ability to look at each DNA unit and each possible mutation and the effects on each of more than 2,000 regulatory interactions across the whole genome.
The ability to predict the functional effect of each mutation was the key innovation in this new study. Previous studies had found it challenging to detect any difference in the number of regulatory mutations in people with autism compared to unaffected people. The new method, however, looked at mutations predicted to have a high functional impact, and found a significantly higher number of such mutations in affected people.