Fig. 7

Principal components analysis of JHU-FFPE. SCPCs and adenocarcinomas were generally separated by principal components analysis. Mixed adenocarcinomas exhibited roughly intermediate behavior, although we questioned whether 56104_A contained an accidental admixture with its neighboring small cell component 56104_S. One SCPC (57915) and one pure adenocarcinoma (57634) clustered with opposite phenotypes, similar to meta-9 clustering. Among SCPCs with known AR or PSA-positivity, two clustered side by side in a relatively intermediate territory (57915, 56107) while the third was loosely in the vicinity (57912_S). Of all principal components, PC2 separated SCPCs from adenocarcinomas best (AUC 86%) and was highly correlated to the difference between ARS and CCP (r=0.93). By contrast in frozen tissue primary and xenograft NEPC datasets, respective PC1's separated SCPCs from adenocarcinomas best (AUC 100%) and was highly correlated to the difference between CCP and ARS (r=0.75-0.98) (Supp Figure 9). Thus in JHU-FFPE, PC1 represented a different source of greatest variability. Examination of its top coefficients by magnitude revealed that PC1 was highly anti-correlated to the average expression of various ribosomal subunits (r = -0.93) including RPL19, known to be an effective reference gene. Two SCPCs (56321_S, autopsy 54674) had the largest PC1 magnitudes and were archived 14-16 years (versus 0-6 years for other SCPCs), perhaps reflecting higher levels of RNA degradation; however the oldest adenocarcinoma (56321_A) did not exhibit this trend