As we uniformly processed the transcriptomic and epigenomic data across the PsychENCODE, ENCODE, GTEx, and Roadmap datasets, we could compare the brain with other organs in a consistent fashion and also compare transcriptome variation with that of the epigenome (Fig. 3, C to F). Several approaches, including principal components anaylsis (PCA), t-distributed stochastic neighbor embedding (t-SNE), and reference component analysis (RCA), were tested to determine the best method for comparison. We found that, although popular and interpretable, PCA deemphasizes local structure and is overly influenced by outliers; by contrast, t-SNE preserves local relationships but “shatters” global structure. RCA is a compromise (21): It captures local structure while maintaining meaningful distances globally. We used RCA to project gene expression from PsychENCODE samples against a reference panel of gene expression for different tissues derived from GTEx and then reduced the dimensionality of the projections with PCA. RCA thus allowed us to represent high-dimensional expression data in a simple two-coordinate diagram.
