Replication in Independent Datasets
Replication was performed in the HBCC microarray cohort (dbGaP: phs000979, see Web Resources) and in the ROSMAP22 RNA-seq cohort by fitting the stepwise regression models identified in the CMC data. For cases in which a marker was unavailable in the replication cohort, all models including that marker (i.e., for that eQTL and higher-order eQTL conditional on it, for a given gene) were omitted from replication.
Data from the HBCC cohort was QC’ed and normalized as described in Fromer et al.16 DLPFC tissue was profiled on the Illumina HumanHT-12_V4 BeadChip and normalized in an analogous manner to the CMC data. Genotypes were obtained using the HumanHap650Yv3 or Human1MDuov3 chips and imputed using the 1000 Genomes Phase 1 reference panel. Replication of the eQTL models was performed on 279 genetically inferred European-ancestry samples (76 control subjects, 72 SCZ-affected subjects, 43 BP-affected subjects, 88 MDD-affected subjects), adjusting for diagnosis and five ancestry components.
ROSMAP data were obtained from the AMP-AD Knowledge Portal (see Web Resources). Quantile normalized FPKM expression values were adjusted for age of death, RIN, PMI, and 31 hidden confounders from SVA, conditional on diagnosis. Only genes with FPKM > 0 in more than 50 samples were retained. QC’ed genotypes were also obtained from the AMP-AD Knowledge Portal and imputed to the Haplotype Reference Consortium (v.1.1)23 reference panel via the Michigan Imputation Server.24 Only markers with imputation quality score R2 ≥ 0.7 were considered in the replication analysis. GemTools was used to infer ancestry components as was done for the CMC data above. After QC, 494 samples were used for eQTL replication in a linear regression model that also adjusted for diagnosis (Alzheimer disease, mild cognitive impairment, no cognitive impairment, and other) and four ancestry components.