Forecasting risk gene discovery in autism with machine learning and genome-scale data

Abstract

Genetics has been one of the most powerful windows into the biology of autism spectrum disorder (ASD). it is estimated that a thousand or more genes may confer risk for ASD when functionally perturbed, however, only around 100 genes currently have sufficient evidence to be considered true "autism risk genes". Massive genetic studies are currently underway producing data to implicate additional genes. This approach-although necessary-is costly and slow-moving, making identification of putative ASD risk genes with existing data vital. Here, we approach autism risk gene discovery as a machine learning problem, rather than a genetic association problem, by using genomescale data as predictors to identify new genes with similar properties to established autism risk genes. This ensemble method, forecASD, integrates brain gene expression, heterogeneous network data, and previous gene-level predictors of autism association into an ensemble classifier that yields a single score indexing evidence of each gene's involvement in the etiology of autism. We demonstrate that forecASD has substantially better performance than previous predictors of autism association in three independent trio-based sequencing studies. Studying forecASD prioritized genes, we show that forecASD is a robust indicator of a gene's involvement in ASD etiology, with diverse applications to gene discovery, differential expression analysis, eQTL prioritization, and pathway enrichment analysis. Autism Spectrum Disorder (ASD) is a heterogeneous group of developmental disorders caused by a range of genetic and environmental factors. The core diagnostic features of ASD-which manifest at a young age-are impairments in social communication and restrictive and repetitive behaviors and interests. Evidence for genetic causes of ASD is strong: monozygotic twins have near 90% concordance of ASD diagnosis 1. Population and twin studies confirmed these findings 2 and estimated ASD's narrow-sense heritability to be in the 50-95% range. While evidence of genetics's role in autism is abundant, our understanding of the specific genes underlying the disorder's etiology is still limited. It is estimated that a thousand or more genes may confer risk for ASD when functionally perturbed 3. This functional perturbation can come in a number of forms, including de novo mutation, expression quantitative trait loci (eQTL) disruption, and inheritance of damaging rare variants. Despite these projections, the existing list of high-confidence autism risk genes-depending on the source-stands closer to the 100-gene mark 3-6. This discrepancy is partly explained by the relatively limited number of genomic studies compared to the vast genetic heterogeneity underlying autism. To close the gap between anticipated and known autism risk genes, several network-biology approaches were applied in the past decade. These studies leveraged large, publicly-available datasets to add context and amplify the genetic signals observed in ASD sequencing studies. These network-biology studies predicted indi...

A Machine Learning Approach to Predicting Autism Risk Genes: Validation of Known Genes and Discovery of New Candidates

“…In the search for additional de novo mutations, sequencing studies continue to be an important approach, but the current sequencing cost is still very high, especially for large samples. As an alternative strategy, advanced analytical approaches, which leverage previously implicated genes and prior knowledge, have the potential to enhance risk gene discovery in an efficient and cost-effective manner ( Asif et al, 2018 ; Gök, 2018 ; Brueggeman et al, 2020 ).…”

Can machine learning aid in identifying disease genes? The case of autism spectrum disorder

“…forecASD analysis: We obtained code for the forecASD classifier from https://github.com/LeoBman/forecASD, and re-ran it locally (Brueggeman et al, 2020). A minor difference from the preprint is the GitHub code uses a different version of the randomForest R package (version 4.6-14 vs 4.6-12 in the preprint) (Brueggeman et al, 2020;Liaw & Wiener, 2002).…”

“…forecASD analysis: We obtained code for the forecASD classifier from https://github.com/LeoBman/forecASD, and re-ran it locally (Brueggeman et al, 2020). A minor difference from the preprint is the GitHub code uses a different version of the randomForest R package (version 4.6-14 vs 4.6-12 in the preprint) (Brueggeman et al, 2020;Liaw & Wiener, 2002). We refit the final ensemble model (03_ensemble_model.R) with different sets of the input features used in final ensemble model: the noClass (noC) model removed features from other classifiers listed above; the noClassPPI (noCP) model eliminated the other classifiers, and the STRING score; noClassPPIBS (noCPB) model eliminated the other classifiers, the STRING score, and the BrainSpan score; the PPIOnly (PPI) model only used the STRING score; and the BrainSpanOnly (BS) model only used the BrainSpan score.…”

“…forecASD (Brueggeman et al, 2020) is a stacked random forest ensemble classifier using BrainSpan (Sunkin et al, 2013) gene expression data, STRING (Szklarczyk et al, 2019) protein-protein interaction data, and genome-wide results from Princeton, DAWN, DAMAGES, Sanders and DeRubeis studies described above. They used 76 SFARI high-confidence genes and 1,000 randomly selected non-SFARI genes as positive and negative training examples, respectively.…”

A novel data-driven approach reveals gene networks and biological processes underlying autism

“…Several authors have investigated possible methods to produce network-based prioritization of ASD genes [15]. In particular, machine learning algorithms have been employed in order to delineate the ASD architecture [17] [18] [19]. Also cluster analysis has seen increasing applications in biomedicine to further the understanding of ASD [20].…”