An xQTL map integrates the genetic architecture of the human brain's transcriptome and epigenome

Abstract

We report a multi-omic resource generated by applying quantitative trait locus (xQTL) analyses to RNA sequence, DNA methylation and histone acetylation data from the dorsolateral prefrontal cortex of 411 older adults who have all three data types. We identify SNPs significantly associated with gene expression, DNA methylation and histone modification levels. Many of these SNPs influence multiple molecular features, and we demonstrate that SNP effects on RNA expression are fully mediated by epigenetic features in 9% of these loci. Further, we illustrate the utility of our new resource, xQTL Serve, by using it to prioritize the cell type(s) most affected by an xQTL. We also reanalyze published genome wide association studies using an xQTL-weighted analysis approach and identify 18 new schizophrenia and 2 new bipolar susceptibility variants, which is more than double the number of loci that can be discovered with a larger blood-based expression eQTL resource.

Figure 1. Overview of xQTL analysis

(A) Graphical summary of our data and analyses. We first associate genetic variation with each data type separately to establish our xQTL reference. We then use these xQTLs to assess whether a given SNP influences more than one data type, whether epigenomic features mediate the effects of SNPs on gene expression, and whether our xQTLs can be leveraged to discover new susceptibility loci. (B) −log₁₀ p-value of Spearman’s correlation between SNPs and DNA methylation (mQTL), histone acetylation (haQTL), and gene expression (eQTL) vs. the SNPs’ physical positions in the genome. Each dot represents the strongest association within a cis window for each SNP. (C) Zoomed in Manhattan plot of chromosome 18 to illustrate p-value distribution of xQTLs at a higher resolution.

Figure 2. Cross-tissue replication analysis

(A) Scatter plot of −log₁₀ p-values of associations between the lead brain eQTL SNPs and their associated genes in brain and blood. The dashed red lines denote the significance threshold (α_FWER=0.05 with Bonferroni correction). (B) −log₁₀ p-value distribution of eQTLs that appear to be brain-specific (light and dark pink dots, the latter are specific to NLRP1). (C) Distribution of p-values from the DGN study restricted to brain eQTLs. Estimated replication rate (π₁ statistics) between blood and brain eQTLs is 0.83. (D) eQTL p-values at NLRP1 locus. Each dot represents one SNP tested in either brain (ROSMAP) or blood (DGN). The x-axis corresponds to the distance between each assessed cis SNP and NLRP1’s TSS, and the y-axis corresponds to −log₁₀ p-values for association between SNPs and NLRP1 expression. The LD between the lead SNP in blood and brain is r² < 0.1.

Figure 3. Genomic enrichment of xQTLs and their overlap

(A) Log odds ratio of xQTL SNP enrichment in 15 different chromatin states as defined by the Roadmap Epigenomics project via applying ChromHMM to DLPFC samples from two cognitively non-impaired ROSMAP subjects. The error bars reflect standard deviation. (B) Log odds ratio of xQTL SNP enrichment in exons and introns. The error bars reflect standard deviation. (C) Distribution of distance between each lead mQTL SNP and its nearest TSS. (D) π₁ statistics for assessing xQTL sharing across the three molecular features. Each cell (i,j) corresponds to the proportion of xQTLs of trait j that share the same xQTL SNPs identified in trait i.

Figure 4. Epigenetic mediation of eQTLs

(A) Sharing of SNPs between eQTLs, mQTLs, and haQTLs. 2,305,942 SNPs tested for all molecular phenotypes are considered. (B) Three models relating SNPs (s), epigenetic features (methylation/histone acetylation, m/h) and gene expression (g): i) independent model (IM) where effects of SNPs on epigenetic features and transcripts are unrelated, ii) epigenetic mediation model (EM) where epigenetic features mediate the effects of SNPs on gene expression, and iii) transcription mediation model (TM) where the effects of SNPs on epigenetics is mediated through its effect on gene expression. The causal inference test was used for assessing mediation. (C) Proportion of shared xQTL SNPs that are consistent with each model. (D) Expression level of IL1RL1 vs. number of minor alleles present for rs13015714, which is a shared xQTL SNP that impacts IL1RL1 expression and nearby DNA methylation and histone acetylation levels. The red line corresponds to the mean. The yellow region corresponds to the 95% confidence interval of the mean. The edges of the blue region correspond to ±1 standard deviation. The SNP effect disappears after regressing out the effect of the mQTL probes and haQTL peaks associated with rs13015714 from IL1RL1 expression. (E) Association between IL1RL1 expression and the levels of its associated methylation probes and acetylation peaks. Colors indicate the genotype of rs13015714: minor allele homozygotes (yellow), heterozygotes (green), major allele homozygotes (blue).

Figure 5. Application of the xQTL Resource for translational studies

(A) Enrichment of xQTL SNPs in published GWAS datasets based on the LDSR model. Enrichments are with respect to two sets of background SNPs: 1) all genome-wide SNPs and 2) SNPs falling in generic functional sites previously defined by LDSR. The error bars reflect standard deviation. (B) −log₁₀ p-value of interaction test in quantifying cell-specificity. 46 genes that survived FDR correction at a q-threshold of 0.2 shown. (C) Level of CPVL expression vs. a marker of microglial proportion (CD68 gene). CPVL expression is found to increase with increasing proportion of microglia, particularly among major allele homozygotes (pink dots). (D–E) Zoomed in Manhattan plot around the PCNX (D) and CPEB1 (E) loci, showing the results of the published standard GWAS (bottom panel) and our weighted GWAS (top panel). Each dot is one SNP. The dotted green line is the standard genome-wide significance threshold (p < 5x10⁻⁸).