the basal subtype and TNBC; basal tumors comprised a small minority of clusters C and E in 7 VEGF and Sema Expression Define TNBC comprising 65% of the tumors in that cluster. This corresponds to the PCA results that demonstrated a strong association of PC2t with the MSL subtype. Most of the other subtypes were evenly distributed across the clusters, with the exception of the basal-like 1 subtype, which comprised 66% of cluster J. Patterns of expression for other genes when sorted in the order of the TNBC clusters were less apparent in Validation Using TCGA Data Using two TCGA data sets consisting of 537 tumors quantified using a different microarray platform and 750 tumors quantified using an RNA-Seq platform, we showed that the same patterns of gene expression that distinguish TNBCs from other tumors could be found in other patients using different technologies. PCA scores 
between the two platforms used in the TCGA datasets had strong correlations. Patterns of gene expression associated with the 2,656-tumor dataset were found in both TCGA datasets, including the low PC4a gene expression signature of high VEGFA and low SEMA3B, SEMA3C, SEMA3F, and PLXNB1. Some additional genes were altered consistently in the two validation data sets as well: SEMA5B, SEMA7A, and PLXNA1. All three of these had similar expression patterns to that of VEGFA. Survival Analysis of Clusters We performed Kaplan Meier survival analysis on the tumors to determine the impact of the PCA-derived clusters on patient prognosis. Triple negative status and increasing stage of the tumor were both correlated with poorer prognoses as expected. Multivariate survival analysis using a Cox proportional hazards model showed 9682837 that tumor stage, lymph node status, PC3a score, and PC4a score were all independent prognostic factors. Interestingly, triple negative status, which was clearly correlated with poor survival, was not significant in the multivariate model. The likely reason for this is that TN status is also highly correlated with PC4a scores. There were more non-TNBC samples with low PC4a scores than TNBC samples with high PC4a scores, possibly resulting in a stronger survival effect from PC4a score than from TN status. We also examined the interaction of ESR1 expression with PC4a; both low ESR1 expression and low PC4a scores were significantly associated with poor prognoses. Interestingly, PC4a score was significantly associated with survival in a subgroup consisting of tumors with high ESR1 expression. Although ER+ tumors are already associated with effective therapies, combination of existing therapies with angiogenesis inhibition may provide additional Chebulinic acid benefits for these low PC4a, ER+ tumors. Survival analysis by cluster showed that tumor clusters F and G had significantly better outcomes than the rest of the clusters. These were the only clusters that had both low PC3a scores and high PC4a scores, reinforcing the prognostic value of these two 7751958 principal components. The five TNBC clusters found here did not have significantly different prognoses; instead the survival curves of the clusters shared the same poor prognosis characteristic of TNBCs. Despite the lack of variability, the differences in patterns of VEGF and semaphorin gene expression may indicate different growth factor dependencies. For example, VEGFC-targeting therapies may be more effective in cluster M, while the rest may benefit more from VEGFA-targeting therapies. Within each PAM50 subtype, ESR1 was not s