To more thoroughly compare L1000 to RNA-seq, we then computed sample self-correlations (using Spearman rank correlation) for the 3,176 samples in the space of the 970 genes directly measured by both platforms. There are 8 L1000 landmark genes that were not included in the DSGEO-RNA-seq. Level 3 L1000 data were used, and the GTEx RNA-seq data were quantile normalized, log2 scaled 1+RPKM values. We then computed sample self-correlations for the 3,176 samples and the median sample self-correlation was 0.84, with a notably right-shifted distribution relative to non-self correlations (Figure 1E, lower panel left). We also measured sample Recall (Rsample, see STAR Methods), wherein a given L1000 profile is forced to compete with all other RNA-seq profiles in order to find its RNA-seq counterpart. This analysis yielded 3,103/3,176 samples (98%) with a Rsample > 0.99 (indicating 99th percentile) and all but 5 (99.84%) had a Rsample > 0.95 (Figure S1D).
