Statistical quantitative trait loci analyses

For cis-meQTL analysis, we confirmed individual sample identity of the SNP and CpG methylation data sets based on the genotypes of 65 CpG/SNP probes on the HumanMethylation450 array compared with the SNP genotype calls assessed by the Illumina Human660W SNP array. The eQTL and meQTL analyses were performed for an additive linear regression model using the R package Matrix eQTL²⁸. Given that the majority of hippocampus biopsies showed neuropathological alterations with neuronal cell loss and gliosis, the inter-sample adjustment for the neuronal proportion as well as known and unidentified confounders is of critical relevance for QTL analyses. The neuronal and glial proportion of each hippocampus specimen was calculated using the CETS R package⁴². To design the model, we first build residuals of the expression and methylation matrix (lm, stats R package) over the known confounders: neuronal proportion, age, and gender. Moreover, the first three Eigenstrat⁶⁶ PCs from SNP genotype data were included to control for ethnicity differences. For the meQTL model, the six PCs from the 220 array control probes were added to control for technical bias. Subsequent principal component analyses (PCAs) on residuals were calculated using prcomb (stats R package). PCA for RNA expression resulted in a variance proportion of 17.1% and 11.9% for the first two principal components (PCs), respectively. PCA for DNA methylation revealed PCs #1–#3 as an important unknown confounder accounting for 13.5% of variance. These PCs have general impact on RNA expression and DNA methylation, respectively. Matrix eQTL²⁸ linear model eQTL analysis was performed using the covariates: gender, age at surgery, neuronal proportion (known confounder), three PCs for population stratification and two additional PCs as unknown confounder. The cis-meQTL analysis was conducted using the covariates: gender, age of surgery, neuronal proportion (known confounder), three PCs for population stratification, six PCs for the adjustment of technical bias and three additional PCs as unknown confounder. In addition, we extracted independent surrogate variables (ISVs) as an alternative method of correction for heterogeneity and latent variables in QTL analyses. We performed a surrogate variable analysis using ISVA⁴³, gender as the protected variable, fastICA for component extraction and an FDR threshold for feature selection of 5%. Each time 13 ISVs for the methylation and expression matrix, respectively, gender and the three Eigenstrat⁶⁶ PCs for correction of ethnic heterogeneity were used for the alternative ISV-adjusted QTL calculations using Matrix eQTL. Cis-QTL analyses were carried out for SNP/probe pairs spanning <500 kb. An FDR threshold of 1% was considered as significant. We did not perform trans-QTL analyses due to the relatively small sample size resulting in an insufficient power and the substantial impact of spurious trans-meQTL associations reflecting cross-hybridization of CpG probes to more than one genomic localization^65,67.