Skip to main content
Premium Trial:

Request an Annual Quote

Autism Spectrum Disorder Risk Genes Predicted Using Machine-Learning Approach

NEW YORK (GenomeWeb) – Researchers at Princeton University have uncovered hundreds of candidate autism spectrum disorder risk genes, many of which are involved in similar biological pathways and are active during particular brain development stages.

Princeton's Olga Troyanskaya and her team developed a machine-learning approach that relies on a brain-specific functional interaction network to predict risk genes for autism spectrum disorder (ASD), which they described today in Nature Neuroscience. The researchers then linked many of these predicted candidate risk genes to certain functions.

"The genome-wide complement of autism candidate genes produced in this study can be used to systematically prioritize genes for resequencing, to guide the analysis of whole genome sequencing results, and to accelerate discovery of the full genetic spectrum underlying ASD, which is needed to refine genetic diagnosis and develop treatments," Troyanskaya and her colleagues wrote in their paper.

The researchers developed a gene interaction network model that included the predicted functional relationships in the brain for all pairs of genes in the 25,000-gene human genome. They further devised a machine learning-based method to take advantage of this network to uncover new candidate autism genes. They trained the classifier using a curated group of 594 genes previously linked to autism with varying levels of evidence supporting their association.

After cross-validating their method, Troyanskaya and her colleagues compiled a list of hundreds of genes associated with autism risk with varying degrees of confidence.

By drawing on an external exome sequencing cohort, the researchers found that the genes with loss-of-function mutations in autism probands were enriched among the top-ranked genes on their list. This enrichment, they noted, was not found in the unaffected siblings of the probands.

The top 10 percent of Troyanskaya and her colleagues' list of predicted autism spectrum disorder risk genes includes the targets of many known disease-associated regulators like FMRP, TOP1, and CHD8, as well as members of biological pathways previously implicated in autism like the Wnt-β-catenin and MAPK pathways.

By combining their gene predictions with a large-scale spatiotemporal gene expression dataset of the developing human brain, Troyanskaya and her colleagues homed in on brain regions and developmental stages likely involved in autism development. They reported seeing a pattern suggesting that autism-linked genetic changes influence fetal prefrontal, temporal, and cerebellar cortex development. At the same time, they observed broad activity of candidate genes throughout the brain, which they said could reflect the disease's heterogeneity and its effects on the visual, auditory, motor, and somatosensory cortices.

Meanwhile, the researchers also used a community-finding algorithm to identify clusters of autism-linked genes that share local network neighborhoods. This, they reported, grouped the autism network into functional modules that are affected by changes to the genes the researchers identified. This, they added, highlighted the roles of synaptic transmission and neuronal function in autism, as well as sensory perception and glucose metabolism.

Troyanskaya and her colleagues also noted that genes within eight most common copy-number variants linked to autism included many of the genes in the top decile of their list. For instance, they found that PPP4C, one of their top-ranked genes in the 16p11.2 locus, regulates Rho GTPase and filamentous actin levels. Haploinsufficiency at that spot, they suggested, could contribute to the autism phenotype. These CNVs, too, appear to converge on particular biological processes, the researchers reported.

"In addition to prioritizing hundreds of new ASD candidate genes, we used the underlying network to link individual ASD genes to perturbed cellular functions and higher-level phenotypes (such as brain development, sensory perception, and circadian rhythm), thus aiming to bridge the genotype–phenotype gap in autism," Troyanskaya and her colleagues wrote.

She and her team also developed a web interface to provide access to the autism gene predictions from this study and other analyses.