Fig. 6. Repeated observation of LMS expression among primary human breast cancer. Hierarchical clustering of (A) the NKI-295 cohort (on Agilent oligonucleotide microarrays) and (B) the MSK-99 and EMC-344 breast cancer cohorts (both on the Affymetrix HG-U133A platform), using the 18-gene LMS. Dendrograms for the tumor samples are shown above the heatmaps, whereas the individual genes of the LMS are labeled and represented in rows. Yellow indicates high expression, and blue represents low expression. Below the heatmap is a row indicating the results of a nearest centroid classifier for LMS expression trained on a 78-tumor subset of the NKI-295 cohort, as previously described (15). Tumors labeled in black are negative for the LMS, whereas those that are in yellow express the LMS in a manner resembling lung metastatic breast cancer cell lines, and are thereby LMS positive. In the MSK-99 and EMC-344 cohorts, 21 and 76 LMS positive tumors are contributed by each cohort, respectively. Tumors labeled in gray have less than 3 years of clinical follow-up, were not classified, and were excluded from subsequent analyses. There are 52 LMS positive tumors in the NKI-295 cohort.

Fig. 7. Expression of the LMS predicts for increased risk of distant failure selectively in the lung. (A) Metastasis-free survival for the indicated distant sites measured as site(s) of first distant failure and (B) overall metastasis-free survival (Left) and overall survival (Right) in the NKI-295/EMC-344 cohort. P values were calculated by using the log-rank test. Patients with LMS⁺ primary tumors are shown in red, and LMS^- tumors are in blue.

Fig. 8. LMS enrichment during growth of parental MDA-MB-231 cells as mammary tumors. Relative levels of expression of LMS genes in parental MDA-MB-231 cells, mammary tumors generated by these cells, two different cell lines derived from these tumors, and the in vivo-selected LMS⁺ cell lines LM2-4175 (15). mRNA levels for thirteen of the genes in the LMS (11 up-regulated genes and two down-regulated genes) were determined by qRT-PCR. The values (n = 3 ± SD) are tabulated and graphically plotted. A heat-map conversion of these data are shown in Fig. 2C.

Fig. 9. Clinical and pathological factors that influence metastasis among LMS⁺ primary tumors. Factors that influence the risk of metastasis for patients from the NKI-295/EMC-344 cohort were determined by a random survival forest analysis as described in Fig. 4. The results for tumor size are shown in Fig. 4A. Results for other covariates are shown here either as a box-and-whisker plot or a scatter plot. For the scatter plots, events are indicated in red along with a lowess regression line through these points shown in magenta. Nonoverlapping notches on box-and-whisker plots are considered significant.

Fig. 10. Primary breast cancers with the wound response gene expression signature are associated with cell proliferation-related gene expression events. (A) Wound response signature positive and negative breast cancers from the NKI-295 cohort were used to determine genes that are differentially expressed between the two groups based on SAM. These 1,787 genes were then analyzed for enrichment of biological themes provided by Gene Ontology (GO) terms. The GO term with the highest node in the hierarchy that met a stringent P < 0.001 by a resampling probability analysis was GO: 0008283 representing cell proliferation. Of a possible 958 genes in the population that matched this GO term, 241 were differentially expressed. Shown is a hierarchical clustering using these 241 genes. (B) The difference in tumor size between wound signature positive and negative patients from the NKI-295 cohort is shown in the box-and-whisker plot. The P value was calculated by the Wilcoxon rank-sum test. For the analyses shown here, the wound response signature designation from the initial unsupervised classification were used (17). Similar results were obtained by using the optimal cut-off obtained from clinical training of the wound response signature (17).

Fig. 11. The influence of tumor size and poor prognosis gene expression signatures on the risk for lung metastasis. A random survival forest analysis was performed by using tumor size and the indicated gene expression signatures for the NKI-295 cohort as described in Fig. 4B. The expected frequencies for lung metastasis for each covariate are shown here either as a box-and-whisker plot or a scatter plot. For the scatter plots, events are indicated in red along with a lowess regression line through these points shown in magenta. Nonoverlapping notches on box-and-whisker plots are considered significant. The molecular subtypes luminal A (LuA), luminal B (LuB), normal-like (Nm), ERBB2/Her2 (Erb), and basal-like (Bas) are indicated.