Fig. 4

Cell type DNA methylation patterns associated with schizophrenia. a DMPs associated with schizophrenia. Quantile-quantile plots of genome-wide P values for differential methylation between schizophrenia and control based on NeuN⁺ (left) and OLIG2⁺ (right) WGBS data. b Concordance between WGBS data and microarray-based data. Y-axis shows the ratio of sites showing the concordant direction in schizophrenia vs. control in our study at each P value bin compared with the Jaffe et al. study [7] (X-axis). Concordance was tested using a binomial test (stars indicate P < 0.05). Boxplots correspond to the directional concordance in 100 sets of association results after case-control label permutations. NeuN⁺ (left) and OLIG2⁺ (right). c szDMPs show altered cell type differences. Barplot shows the percentage of sites with larger cell type differences in controls than in schizophrenia and vice versa at different CpG classes. Absolute OLIG2⁺ vs. NeuN⁺ methylation differences are larger in controls than cases in szDMPs compared to cell type DMPs and non-DMP or background CpGs. szDMPs were detected as differentially methylated between cases and controls at FDR < 0.2 in NeuN⁺ (14 sites) and OLIG2⁺ samples (83 sites). Top 1000 szDMPs were selected as the top 1000 loci according to best P values in each cell type (N = 2000). Cell type DMPs were detected by comparing OLIG2⁺ vs. NeuN⁺ methylomes at Bonferroni P < 0.05. Background CpGs were sampled from CpGs showing non-significant P values for both case-control and OLIG2⁺ vs. NeuN⁺ comparisons. Stars represent P values for binomial tests with all comparisons showing P < 10⁻⁷. d Top 1000 szDMPs are enriched for SZ GWAS signals. szDMPs identified in our methylation study in both cell types consistently co-localize with genetic variants with moderate to large effect sizes for schizophrenia risk than expected. The table shows the empirical P values of szDMPs at each odds ratio (OR) percentile of different traits from genome-wide association studies (GWAS). The actual ORs corresponding to the schizophrenia percentiles are indicated at the top. Specifically, for each szDMP, we identified all SNPs reported by the GWAS study within a 1-kb window and counted the number of SNPs at different quantiles of odds ratio (OR). We used quantiles of OR so that we can compare the different diseases and traits among them. We repeated this step using the same number of random non-szDMPs 100 times. To obtain empirical P values, we calculated the number of times non-szDMP sets showed more SNPs in each OR quantile than szDMPs. SNPs with moderate-to-high OR in schizophrenia GWAS consistently showed low empirical P values for both cell type DMPs, implying that SNPs with large effect sizes in GWAS studies are closer to szDMPs than expected. Interestingly, this pattern was not observed for other traits, implying the co-localization is exclusive to the disease