Fig. 4: GMM-based classification predicts that SRC but not BLK is actively expressed in ECs.
a, Illustrative kernal density estimates (KDEs) of log2-transformed TPM values for protein-coding genes in the bulk RNA-seq data from the indicated datasets. Expectation maximization was used to estimate the parameters of the low and high Gaussian distributions (predicted leaky versus active transcription), represented by black and gold fit curves, respectively. The log2(TPM) and P(active) values for PECAM1, YES1, SRC and BLK in each dataset are indicated together with the P(leaky) values for BLK. The illustrative HUVEC dataset is also shown without the estimated Gaussian distributions to its transcript distribution (top left). b, Stacked bar charts depicting the number of datasets per EC subtype in which the indicated genes were classified by the GMM method as active, leaky or undetermined, resolved by EC subtype; lung, brain and retina EC data were obtained from mouse datasets. The percentage of datasets in which each gene was classified as actively expressed is reported for each EC subtype below each bar (number of eligible, bimodally distributed datasets: PECAM1 n = 198, YES1 n = 198, SRC n = 194, BLK n = 64, KLF1 n = 30). Note that only datasets with a transcript level of >0 TPM are classified; therefore, datasets not shown have a transcript level of <0 TPM.
