Stratified linkage disequilibrium score regression

We used stratified linkage disequilibrium score regression (SLDSR), implemented in the LDSC [72] software, to evaluate the enrichment of common genetic variants from genome-wide association study (GWAS) signals to partition trait heritability within functional categories represented by our DMRs and VMRs. SLDSR estimates the proportion of genome-wide single nucleotide polymorphism (SNP)-based heritability that can be attributed to SNPs within a given genomic feature by a regression model that combines GWAS summary statistics with estimates of linkage disequilibrium from an ancestry-matched reference panel. Links to GWAS summary statistics are available in Additional file 16: Table S15. The GRCh38 “baseline-LD model v2.2” data files were downloaded from https://data.broadinstitute.org/alkesgroup/LDSCORE/ following instructions at https://github.com/bulik/ldsc/wiki.

We ran LDSC (v1.0.0; https://github.com/bulik/ldsc) to estimate the proportion of genome-wide SNP-based heritability of 30 traits (Additional file 16: Table S15) across the 97 “baseline” genomic features and our neuronal (NeuN+) DMRs and VMRs:

5-group CG-DMRs: CG-DMRs among 5 brain tissue groups (196 Mb) (Additional file 6: Table S5)

5-group CH-DMRs: union of CA-, CC-, and CT-DMRs (both strands) among 5 brain tissue groups (1010 Mb) (Additional file 12: Table S11)

Basal ganglia CG-DMRs: CG-DMRs among 3 basal ganglia tissues (24 Mb) (Additional file 8: Table S7)

Basal ganglia CH-DMRs: union of CA-, CC-, and CT-DMRs (both strands) among 3 basal ganglia tissues (284.4 Mb) (Additional file 13: Table S12)

Hippocampal CG-DMRs: CG-DMRs among the two hippocampal groups (24.4 Mb) (Additional file 10: Table S9)

Hippocampal CH-DMRs: union of CA-, CC-, and CT-DMRs (both strands) among the two hippocampal groups (596 Mb) (Additional file 13: Table S12)

VMRs identified in each brain tissue (Mb covered listed in Table ​Table1)1) (Additional file 14: Table S13)

We performed a standard SLDSR analysis, as suggested by the method authors, whereby each of the brain-specific features was added one at a time to a “full baseline model” that included the 97 “baseline” categories that capture a broad set of genomic annotations. We used SLDSR to estimate a “coefficient z-score” and an “enrichment score” for each feature-trait combination. A brief description of their interpretation is given below; we refer the interested reader to the Online Methods of [50] for the complete mathematical derivation. A coefficient z-score statistically larger than zero indicates that adding the feature to the model increased the explained heritability of the trait, beyond the heritability explained by other features in the model. The enrichment score is defined as (proportion of heritability explained by the feature)/(proportion of SNPs in the feature). The enrichment score is unadjusted for the other features in the model, but is more readily interpretable as an effect size. Particularly interesting are those feature-trait combinations with statistically significant z-score coefficients and large enrichment scores. z-score coefficient p-values within each trait were post hoc adjusted for multiple testing using Holm’s method [73] (Additional file 17: Table S16).