Integrative genomics identifies a convergent molecular subtype that links epigenomic with transcriptomic differences in autism

Abstract

Autism spectrum disorder (ASD) is a phenotypically and genetically heterogeneous neurodevelopmental disorder. Despite this heterogeneity, previous studies have shown patterns of molecular convergence in post-mortem brain tissue from autistic subjects. Here, we integrate genome-wide measures of mRNA expression, miRNA expression, DNA methylation, and histone acetylation from ASD and control brains to identify a convergent molecular subtype of ASD with shared dysregulation across both the epigenome and transcriptome. Focusing on this convergent subtype, we substantially expand the repertoire of differentially expressed genes in ASD and identify a component of upregulated immune processes that are associated with hypomethylation. We utilize eQTL and chromosome conformation datasets to link differentially acetylated regions with their cognate genes and identify an enrichment of ASD genetic risk variants in hyperacetylated noncoding regulatory regions linked to neuronal genes. These findings help elucidate how diverse genetic risk factors converge onto specific molecular processes in ASD.

Fig. 1. SNF to identify ASD molecular subtypes.

a Overview of data integration and molecular subtyping to characterize the cascade of molecular changes in ASD. b Relationship between sample loadings on the first principal component of differential mRNA expression, miRNA expression, DNA methylation, and histone acetylation. c Identification of two sample clusters using SNF: SNF Group 1 and SNF Group 2. ASD samples in SNF Group 1 constitute the Disparate Subtype, whereas ASD samples in SNF Group 2 constitute the Convergent Subtype. d Number of samples classified into the two cluster groups using SNF and logistic regression (LR) classifiers. e Comparison of SNF clustering and logistic regression (LR) classification assignments when utilizing three out of four datasets (see Supplementary Fig. 5). The concordance of sample assignments to those when using the complete dataset are plotted.

Fig. 2. mRNA expression differences in ASD.

a Overlap in ASD downregulated genes identified in this study with Parikshak et al.. b Overlap in ASD upregulated genes identified in this study with Parikshak et al.. c Signed association of mRNA co-expression module eigengenes with diagnosis (Bonferroni-corrected p-value from a linear mixed effects model, see Supplementary Fig. 7e). Positive values indicate modules with an increased expression in ASD samples. Gray and black bars with labels signify ASD-associated modules identified in Parikshak et al., and those additionally identified in this study, respectively. Cell type enrichment for each module is shown in parenthesis: neuron (N), astrocyte (A), microglia (M), and no enrichment (−) (see Supplementary Fig. 7g). d Top 30 hub genes and 300 connections for co-expression module mRNA.M17. e Top gene ontology enrichments for co-expression module mRNA.M17. Ontology enrichments were calculated by g:Profiler with FDR corrected p-values. f Enrichment of ASD downregulated neuronal co-expression modules with neuronal cell-type markers identified from single-nuclei RNA sequencing. Enrichments were calculated using a logistic regression model and p-values were adjusted for multiple testing using FDR correction. Only those enrichments with odds ratio >1 and FDR corrected p-value < 0.05 are shown. g Top 30 hub genes and 300 connections for co-expression module mRNA.M15. h Top gene ontology enrichments for co-expression module mRNA.M15. Ontology enrichments were calculated by g:Profiler with FDR corrected p-values. i Enrichment of ASD upregulated glial co-expression modules with microglial activated genes and microglial cell-type markers. Enrichments were calculated using a logistic regression model and p-values, which are shown in parentheses, were adjusted for multiple testing using FDR correction. Only those enrichments with odds ratio >1 and FDR corrected p-value < 0.05 are shown.

Fig. 3. DNA methylation differences in ASD.

a Overlap in ASD hypermethylated gene promoters and gene bodies. b Overlap in ASD hypomethylated gene promoters and gene bodies. c Correlation between expression and methylation changes for genes that have differential promoter methylation and are differentially expressed. A linear model was used to correlate differential expression with differential methylation. P-value is from fit of linear model. d Correlation between expression and methylation changes for genes that have differential gene body methylation and are differentially expressed. A linear model was used to correlate differential expression with differential methylation. P-value is from fit of linear model. e Top 30 hub genes and 300 connections for promoter co-methylation module Prom.lightgreen. f Top gene ontology enrichments for promoter co-methylation module Prom.lightgreen. Ontology enrichments were calculated by g:Profiler with FDR corrected p-values. g Promoter co-methylation module Prom.lightgreen eigengene values for ASD and control samples. P-value is from fit of a linear mixed effects model (see Supplementary Fig. 9f). h Top 30 hub genes and 300 connections for gene body co-methylation module GB.darkgreen. i Top gene ontology enrichments for gene body co-methylation module GB.darkgreen. Ontology enrichments were calculated by g:Profiler with FDR corrected p-values. J Gene body co-methylation module GB.darkgreen eigengene values for ASD and control samples. P-value is from fit of a linear mixed effects model (see Supplementary Fig. 10f). For boxplots in g, j, the center of the box is the median value, the bounds of the box are the 75th and 25th percentile values, the whiskers extend out from the box to 1.5 times the interquartile range of the box, and outlier values are presented as individual points.

Fig. 4. Histone acetylation differences in ASD.

a Top gene ontology enrichments when linking ASD hyperacetylated regions to proximal genes using GREAT. P-values were adjusted for multiple testing by FDR correction. b Top gene ontology enrichments when linking ASD hypoacetylated regions to proximal genes using GREAT. P-values were adjusted for multiple testing by FDR correction. c Schema to link H3K27ac regions with their cognate genes. H3K27ac peaks within promoters were directly assigned to the proximal gene. Distal H3K27ac peaks were assigned to genes using eQTL and Hi-C datasets. d Correlation between expression and acetylation changes for genes that have a differentially acetylated region within their promoter and are differentially expressed. P-value is from a linear model used to correlate differential expression with differential acetylation. The four separate quadrants are marked. e Cell type enrichments for the four quadrants in d. Enrichments were calculated using a logistic regression model and p-values, which are shown in parentheses, were adjusted for multiple testing using FDR correction. Only those enrichments with odds ratio >1 and FDR corrected p-value < 0.05 are shown. f Enrichment of cognate genes linked to differentially acetylated regions within mRNA co-expression modules. Modules with a significant relationship to diagnosis are marked along the y axes (red: increased expression in ASD; blue: decreased expression in ASD). Enrichments were calculated using a logistic regression model and p-values, which are shown in parentheses, were adjusted for multiple testing using FDR correction. Only those enrichments with odds ratio >1 and FDR corrected p-value < 0.05 are shown. g Relationship between expression and acetylation changes for differentially acetylated peaks linked to gene co-expression modules. The functional annotation for each module is represented in the top left corner. The association of each module to ASD diagnosis is represented in the top right corner as well as whether acetylation changes are contributory or compensatory to changes in expression.

Fig. 5. ASD genetic risk variant enrichments.

a Partitioned heritability enrichments for ASD, Alzheimer’s, and IBD GWAS in differentially expressed, methylated, or acetylated regions of the genome. Uncorrected p-values < 0.05 are shown. P-values were also adjusted for multiple testing by FDR correction and those enrichments with adjusted p-value < 0.1 are marked with asterisks. b Partitioned heritability enrichments for ASD, Alzheimer’s, and IBD GWAS in differentially acetylated regions of the genome within, or distal to, gene promoters. Uncorrected p-values < 0.05 are shown. P-values were also adjusted for multiple testing by FDR correction and those enrichments with adjusted p-value < 0.1 are marked with asterisks. c Top 30 hub genes and 300 connections for co-expression module mRNA.M4. d Top gene ontology enrichments for co-expression module mRNA.M4. Ontology enrichments were calculated by g:Profiler with FDR corrected p-values. e, f Genomic region around DMTN e and STX1B f. ASD-associated hyperacetylated regions are shown along with eQTL and Hi-C linkages to the gene TSS.

Fig. 6. Schematic model of molecular dysregulation in ASD.

Integration across genetic, transcriptomic, and epigenomic levels in ASD. ASD risk variants primarily act with respect to neuronal genes, which are broadly down-regulated, likely indirectly, including via micro-RNA (e.g. yellow module). Other features, including another microRNA module (brown), and DNA hypomethylation are predicted to be compensatory or secondary and are associated with up-regulation of glial-immune genes. Histone acetylation patterns show a complex relationship, with some predicted to reflect an attempt to compensate for changes in gene expression (e.g. M4), while others are predicted to be likely driving the changes (microglia). Blue bar-headed and red arrows correspond to regulatory mechanisms predicted to decrease and increase gene expression, respectively.