A prognostic gene signature for metastasis-free survival of triple negative breast cancer patients

Abstract

Although triple negative breast cancers (TNBC) are the most aggressive subtype of breast cancer, they currently lack targeted therapies. Because this classification still includes a heterogeneous collection of tumors, new tools to classify TNBCs are urgently required in order to improve our prognostic capability for high risk patients and predict response to therapy. We previously defined a gene expression signature, RKIP Pathway Metastasis Signature (RPMS), based upon a metastasis-suppressive signaling pathway initiated by Raf Kinase Inhibitory Protein (RKIP). We have now generated a new BACH1 Pathway Metastasis gene signature (BPMS) that utilizes targets of the metastasis regulator BACH1. Specifically, we substituted experimentally validated target genes to generate a new BACH1 metagene, developed an approach to optimize patient tumor stratification, and reduced the number of signature genes to 30. The BPMS significantly and selectively stratified metastasis-free survival in basal-like and, in particular, TNBC patients. In addition, the BPMS further stratified patients identified as having a good or poor prognosis by other signatures including the Mammaprint® and Oncotype® clinical tests. The BPMS is thus complementary to existing signatures and is a prognostic tool for high risk ER-HER2- patients. We also demonstrate the potential clinical applicability of the BPMS as a single sample predictor. Together, these results reveal the potential of this pathway-based BPMS gene signature to identify high risk TNBC patients that can respond effectively to targeted therapy, and highlight BPMS genes as novel drug targets for therapeutic development.

Figure 1. The optimized solutions yield larger cohort sizes and better p-values.

Distribution density plots for non-optimized (control) and optimized signatures verify that significantly better cohort sizes (A) and p-values (B) were generated using a cost function in conjunction with the Nelder-Mead optimization algorithm.

Figure 2. Optimization procedure for the BPMS.

After separating the overall training set (BrCa871) into a training set and a cross validation set, (A) a series of 24,800 potential solutions are produced by optimizing our cost function using the Nelder-Mead downhill simplex algorithm. These solutions were trained on survival data with no year-specific endpoint defined to maximize signal sensitivity (See Figure 4). Using these 24,800 potential solutions, (B) significance in both training and cross-validation sets was assessed. To control for over-fitting solutions, 556 solutions yielding significance in both sets were extracted and used to estimate the final BPMS signature.

Figure 3. The BPMS is a single patient predictor.

Using frozen RMA pre-processed data, the BPMS was trained to be applied on a patient-to-patient basis. The BrCa871 set was processed using fRMA, divided into the BrCa436-Train and BrCa435-CV sets and 7,500 potential solutions were optimized. Using a cross-validation strategy, a final set of BPMS parameters were trained for fRMA processed data. Shown is the application of these parameters to the fRMA processed BrCa341 data set.

Figure 4. The BPMS is prognostic for metastasis-free survival (MFS).

Patients from three breast cancer datasets, (A) BrCa871 (35 BPMS+ out of 871 patients), (B) BrCa443 (24 BPMS+ out of 443 patients) and (C) BrCa341 (6 BPMS+ out of 341 patients), were stratified for MFS using the BPMS. BrCa871 is shown with no year-specific clinical endpoint to reflect the training data. Red indicates patient tumors that express the BPMS signature while black indicates patient tumors that do not. Survival curves were generated by Kaplan–Meier analysis, and the indicated P-values were calculated by the log-rank test.

Figure 5. The BPMS is prognostic for metastasis-free survival of breast cancer patients with tumors of the basal subtype.

PAM50 was used to categorize breast tumors into (A) Basal (16 BPMS+ patients out of 120 Basal patients, χ² = 13.7), (B) luminal A (0 BPMS+ patients out of 110 luminal A patients), (C) luminal B (1 BPMS+ patient out of 97 luminal B patients, χ² = 0.5), (D) HER2 (4 BPMS+ patients out of 67 HER2 patients, χ² = 0) and (E) normal (3 BPMS+ patients out of 48 Normal patients, χ² = 0.8) subtypes as indicated. BrCa443 patients were stratified for MFS using the BPMS. Red indicates patient tumors that express the BPMS signature while black indicates patient tumors that do not. Survival curves were generated by Kaplan–Meier analysis, and the indicated P-values were calculated by the log-rank test.

Figure 6. The BPMS is prognostic for metastasis-free survival of TNBC patients.

The proliferation signature was used to categorize breast tumors into (A) ER-HER2- (15 BPMS+ patients out of 121 ER-HER- patients, χ² = 10.5), (B) TNBC (18 BPMS+ patients out of 118 TNBC patients, χ² = 9.4), (C) ER+HER2- (n = 1), and (D) HER2+ (8 BPMS+ patients out of 117 HER2 patients, χ² = 0). BrCa443 patients were stratified for MFS using the BPMS. Red indicates patient tumors that express the BPMS signature while black indicates patient tumors that do not. Survival curves were generated by Kaplan–Meier analysis, and the indicated P-values were calculated by the log-rank test.

Figure 7. The BPMS is prognostic for high risk patients among good prognosis patients.

Good prognosis categories examined were: (A) the 76-gene (20 BPMS+ patients out of 290 good-prognosis patients, χ² = 12.2), (B) 28-kinase metagene (8 BPMS+ patients out of 104 high immune response patients, χ² = 6.9), (C) GAB2 Scaffolding (23 BPMS+ patients out of 429 good prognosis patients, χ² = 9.7), and (D) glucocorticoid receptor signature (16 BPMS+ patients out of 121 GR-/ER- patients as defined by 50% cutoff, χ² = 10.5). Patients were stratified for MFS using the BPMS. Red indicates patient tumors that express the BPMS signature while black indicates patient tumors that do not. Survival curves were generated by Kaplan–Meier analysis, and the indicated P-values were calculated by the log-rank test.

Figure 8. The BPMS is prognostic for high risk patients among the clinically predicted poor outcome and high recurrence patients.

Clinically relevant gene signatures (A) Mammaprint® Poor (23 BPMS+ patients out of 226 Mammaprint Poor patients, χ² = 4.3) and (B) OncotypeDX® Recurrence High (16 BPMS+ patients out of 257 RS High patients, χ² = 6.7) were stratified for MFS using the BPMS. Red indicates patient tumors that express the BPMS signature while black indicates patient tumors that do not. Survival curves were generated by Kaplan–Meier analysis, and the indicated P-values were calculated by the log-rank test.