Extended Data Figure 2: Transcriptome-wide DGE analysis.

a, We applied a classification method robust to overfitting (elastic net model⁴⁷) by training on the RNA-seq data from samples previously analysed in ref. 8 (Extended Data Fig. 1h, similar to the comparison in Extended Data Fig. 1i) and classifying ASD versus control status in independent samples. Results are shown as a comparison of classification scores (left) and area under the receiver operator characteristic curve (AUROC, right). Approximately 85% of ASD samples are classified successfully around a false positive rate of 20%. b, Summary table describing the subset of representative, covariate matched samples used for qRT–PCR validations. Supplementary Table 2 contains the underlying values. c, Fold changes from RNA-seq compared against fold changes from qRT–PCR (see Supplementary Table 2 for data). d, GO term enrichment analysis of genes that are upregulated or downregulated in individuals with ASD. e, Enrichment analysis of cell-type specific gene sets (defined as genes with fivefold higher expression in the cell type than in other cell types) with genes that are decreased or increased in ASD. f, g, Independent replication analysis of ASD versus control DGE fold changes between previously evaluated and new ASD samples from cerebellum by microarray and RNA-seq using samples from ref. 8 (similar to Fig. 1a and Extended Data Fig. 1i). The RNA-seq data show a replication signal between previously evaluated and new samples from this study. h, Comparison of fold changes that were significant at FDR < 0.05 in the ASD versus control DGE analysis from cortex compared with fold changes observed in cerebellum, revealing strong concordance but a lower average fold change in the cerebellum. i, Sample summary and quality control (QC) statistics for ref. 4. Compare to Extended Data Fig. 1b and see Supplementary Information for additional discussion. Compared to this study, samples from ref. 4 were prepared by poly(A) selection RNA-seq, exhibit lower RNA integrity number (RIN, median 4.8 versus 7.3), have lower median sequencing depth (11 million versus 40 million), exhibit greater 5′-3′ bias, and have generally greater variability across all QC metrics. j, Comparison of fold-changes for the top significant genes from ref. 4 (P < 0.01 as provided in their Supplementary Information) with the fold changes for the same genes in this study. Co-expression network analysis demonstrated that the moderate agreement is largely driven by concordance in upregulation of microglial genes in both studies (Extended Data Fig. 8e). k, Average linkage hierarchical clustering of lncRNAs in the DGE set. l, Boxplots of expression values of DGE lncRNAs across multiple tissue types from the Illumina Body Map (expression data from ref. 12). Lines above the plot indicate pairwise significance with a one-sided Wilcoxon rank-sum test between brain and the other tissues. m, Similar to l, except for embryonic stem cells and stem-cell-derived cell types. n, RT–PCR validation of the two lncRNAs shown in Fig. 1c, d; P values computed by two-sided Wilcoxon rank-sum test.