Second, we used an unsupervised analysis to identify the primary components of bulk expression variation. We decomposed the bulk gene expression matrix by using nonnegative matrix factorization (NMF) (B ≈ VH, where B, V, and H represent matrices) and determined whether the top components (NMF-TCs), capturing the majority of covariance (columns of V) (Fig. 2B), were consistently associated with the single-cell signatures (Fig. 2C) (21). A number of NMF-TCs were, in fact, highly correlated with cell types from matrix C for both TPM and UMI data—e.g., component NMF-17 is correlated with the Ex2 cell type (correlation coefficient r = 0.63) (Fig. 2C and fig. S9). This demonstrates that an unsupervised analysis derived solely from bulk data can roughly recapitulate the single-cell signatures, partially corroborating them.
