Identification of Prognosis-Relevant Subgroups in Patients with Chemoresistant Triple Negative Breast Cancer

Abstract

Purpose

Triple-negative breast cancer (TNBC) patients with residual disease after neoadjuvant chemotherapy generally have worse outcome; however, some patients with residual tumor after neoadjuvant chemotherapy do not relapse. We hypothesize that there are subgroups of chemoresistant TNBC patients with different prognosis.

Experimental Design

Forty-nine chemoresistant cases from 111 TNBC patients treated with neoadjuvant chemotherapy (M.D. Anderson Cancer Center) constituted the discovery cohort, 25 chemoresistant samples from 47 neoadjuvant chemotherapy-treated TNBC (The Methodist Hospital) chosen for validation. Extended validation was performed in 269 operable TNBC predicted to be chemoresistant by expression pattern from published data sets.

Results

We established a 7-gene prognostic signature using dChip and gene set enrichment analyses. In the independent validation cohort, the classifier predicted correctly with positive predictive value of 75.0% and negative predictive value (i.e., relapse-free survival [RFS]) of 76.9% at 3 years. Those predicted to relapse had a hazard ratio (HR) of 4.67 (95%CI, 1.27–17.15) for relapse in 3 years. In extended validation, patients predicted not to relapse exhibited 3-year RFS of 78.9%, while the 3-year RFS was 48.5% for patients predicted to relapse, with HR of 2.61 (95%CI, 1.52–4.49). The TNBC subgroup predicted to have relatively favorable prognosis was characterized by high expression of “luminal-like” genes (androgen-receptor [AR] and GATA3); while the subgroup with worse prognosis was characterized by expression of cancer stem-cell markers.

Conclusion

We developed a clinically relevant signature for patients with chemoresistant TNBC. For these women, new therapeutic strategies like targeting AR-activation or cancer stem-cells may need to be developed.

Introduction

Triple-negative breast cancer (TNBC) is clinically defined by the lack of expression of estrogen receptor (ER), progesterone receptor (PgR) and the absence of amplification or over-expression of human epidermal growth factor receptor-2 (HER2), and accounts for 15%–20% of newly diagnosed breast cancer cases. In general, TNBC patients present with larger tumors, higher grade, increased number of involved nodes, and poorer survival compared to other subtypes.(1, 2) Increasing evidence indicates that TNBC is a highly heterogeneous disease(1) on a molecular(3) and genetic level.(4) Treatment of patients with TNBC has been challenging due to this heterogeneity, as well as the absence of well-defined molecular targets.

Despite having higher rates of pathologic complete response (pCR) to neoadjuvant chemotherapy, TNBC patients have a higher rate of distant recurrence and worse prognosis. Among TNBC patients receiving neoadjuvant chemotherapy, only those with pCR have improved survival. In contrast, more than 70% of TNBC patients have residual invasive disease after neoadjuvant chemotherapy and are at high risk of disease relapse, with significantly worse survival, particularly in the first three years.(5, 6) Paradoxically, not all TNBC patients with residual disease after neoadjuvant chemotherapy relapse. Identifying chemoresistant TNBC patients who relapse vs. those with relatively favorable prognosis would serve to distinguish clinically relevant subgroups for whom the targeting of different molecular pathways may be important. This study was designed to test our hypothesis that there are clinical prognosis-relevant subgroups within chemoresistant TNBC patients. Understanding the molecular pathways distinguishing prognostically significant subgroups will aid in the rationale design of future clinical trials.

Methods

Patients and samples from M. D. Anderson Cancer Center (MDACC)

To investigate the difference in genetic expression between chemoresistant TNBC patients who relapse vs. those without relapse, we chose patients treated with neoadjuvant chemotherapy (with residual cancer) and investigated survival outcomes, as our discovery and validation cohorts.

The samples of discovery cohort were from MDACC. Patients prospectively provided written informed consent to participate in an institutional review board-approved research protocol. As previously described, 313 HER2-negative samples from patients (45% of them were with operable stage I–II disease) treated with taxane and anthracycline-based neoadjuvant chemotherapy were obtained from Jun-2000 to Dec-2006.(7) Among them, 111 patients were identified to have TNBC, of whom 49 patients fulfilled the following criteria and were included in the discovery cohort: (1), having residual invasive disease either in the breast or in regional lymph nodes after neoadjuvant chemotherapy (i.e., non-pCR); (2), having grade II/III residual cancer burden (RCB);(8) (3), followed up for longer than 20 months. The information of cohorts are provided in Table 1. Chemoresistant tumors were defined in MDACC as “non-pCR and RCB-II/III after neoadjuvant chemotherapy”.

Table 1.

Pretreatment characteristics of the discovery and validation cohorts

	MDACC				TMH-BCM
	Original TNBC source (n=111)		Discovery cohort (n=49)		Original TNBC source (n=47)		Validation cohort (n=25)
	n	%	n	%	n	%	n	%
Follow-up time, months
Median (range)			25 (1–79)				36 (4–88)
Age, years
≤50	59	53.2	23	46.9	27	57.4	12	48.0
>50	52	46.8	26	53.1	20	42.6	13	52.0
Nodal status
Negative	26	23.4	9	18.4	10	21.3	8	32.0
Positive	85	76.6	40	81.6	37	78.7	17	68.0
Tumor size stage
T0–2	71	64.0	24	49.0	17	36.2	7	28.0
T3–4	40	36.0	25	51.0	30	63.8	18	72.0
Grade
I	0	0.0	1	0.0	3	6.4	3	12.0
II	13	11.7	9	18.4	7	14.9	3	12.0
III	98	88.3	40	81.6	37	78.7	19	76.0
pCR
No	69	65.7	49	100.0	35	74.5	25	100.0
Yes	36	34.3	0	0.0	12	25.5	0	0.0
Unknown	6		0		0		0
RCB
I	48	46.6	0	0.0	N.A.		N.A.
II	30	29.1	24	51.1	N.A.		N.A.
III	25	24.3	23	48.9	N.A.		N.A.
Unknown	8		2		N.A.		N.A.
Neoadjuvant and adjuvant CT
P×12→FAC×4→Surgery→no CT	111	100.0	49	100.0	0	0.0	0	0.0
T×4→Surgery→AC×4	0	0.0	0	0.0	24	51.1	11	44.0
AC×4→Surgery→T×4	0	0.0	0	0.0	23	48.9	14	56.0

Abbreviations: AC, doxorubicin and cyclophosphamide; TMH-BCM, The Methodist Hospital, Baylor College of Medicine; CT, chemotherapy; FAC, fluorouracil, doxorubicin and cyclophosphamide; MDACC, M.D. Anderson Cancer Center; N.A., not assessed; P, paclitaxel; pCR, pathologic complete response; RCB, residual cancer burden; T, docetaxel; TNBC, triple negative breast cancer

All gene expression microarrays were profiled in the Department of Pathology at the MDACC and the details of the methods for RNA purification and microarray hybridization have been reported previously.(9)

Patients and samples from The Methodist Hospital, Baylor College of Medicine (TMH-BCM)

The study protocol was approved by the institutional review board, and signed informed consent was obtained from all patients. From Jan-2002 to Dec-2006, 116 patients with locally advanced breast cancer presenting to the Breast Center at TMH-BCM were recruited into a taxane/anthracycline-based neoadjuvant chemotherapy trial. The inclusion criteria were described in the Methods in Appendix. Among them, 47 were identified as TNBC cases, and 25 were recognized as chemoresistant cases after neoadjuvant chemotherapy. The definition of chemoresistant was “pathological grade 3B-3D and grade 4 of modified Chevallier classification after neoadjuvant chemotherapy”.(10) The 25 patients from TMH-BCM constituted the validation cohort (Table 1). Patients without relapse or death events were followed up more than 48 months. The processes of RNA treatment and gene expression profiling have been described elsewhere.(11)

Assessment of pathological response and statuses of ER, PgR, and HER2 was described in the Methods in Appendix.

Identificationand validation of prognosis signature for chemoresistant TNBC

Because of word limit, the details of the finalization and validation of a 7-gene signature for chemoresistant TNBC are described in the Methods in Appendix.

Molecular classification of chemoresistant TNBC

In order to investigate the relationship between our 7-gene signature and the recently described TNBC subtype molecular classification,(3) we used 587 TNBC cases in that study. Gene expression profiles of individual case were read and subtyped by Pietenpol and colleagues (briefly described in the Methods in Appendix).(3)

Extended validation from published adjuvant TNBC microarray data

The chemoresistant, prognosis-relevant TNBC signature was then further validated in publically available datasets. A total of 579 adjuvant TNBC from 3,488 primary breast cancer gene expression profiles representing 28 individual datasets were identified.(12) We predicted the sensitivity to chemotherapy using a previously published signature.(13) The procedure of chemosensitivity prediction was described in the Methods in Appendix. Finally, 269 adjuvant cases predicted to be chemoresistant with at least 3 years follow-up and available survival outcome data were included. They could be grouped into 4 main sets according to patient sample size and patients’ characteristics (Table 2).

Table 2.

Characteristics of the extended validation cohort

	Extended validation cohort in adjuvant TNBC predicted to be chemo-insensitive and using no neoadjuvant chemotherapy
	Total (n=269)		Set I, Rotterdam (n=87)		Set II, Frankfurt (n=53)		Set III, New York/San Francisco/Stockholm/Uppsala (n=78)		Set IV, Mainz/TransBIG/Oxford/others* (n=51)
	n	%	n	%	n	%	n	%	n	%
Follow-up time, months
Median (range)	63 (0–120)		69 (0–120)		34 (6–120)		82 (2–120)		93 (6–120)
Relapse during follow-up
No	162	60.2	49	56.3	32	60.4	51	65.4	30	58.8
Yes	107	39.8	38	43.7	21	39.6	27	34.6	21	41.2
Age, years
≤50	118	47.8	45	52.9	23	43.4	25	42.4	25	50.0
>50	129	52.2	40	47.1	30	56.6	34	57.6	25	50.0
Unknown	22		2		0		19		1
Nodal status
Negative	183	79.6	69	100.0	36	69.2	29	50.0	49	96.1
Positive	47	20.4	0	0.0	16	30.8	29	50.0	2	3.9
Unknown	39		18		1		20		0
Tumor size stage
T1	78	31.6	30	35.3	14	26.4	19	32.2	15	30.0
T2–3	169	68.4	55	64.7	39	73.6	40	67.8	35	70.0
Unknown	22		2		0		19		1
Grade
I–II	78	33.1	24	28.2	16	30.8	24	46.2	14	29.8
III	158	66.9	61	71.8	36	69.2	28	53.8	33	70.2
Unknown	33		2		1		26		4
Adjuvant CT
No	179	73.1	83	98.8	6	15.4	39	54.9	51	100.0
Yes	66	26.9	1	1.2	33	84.6	32	45.1	0	0.0

Abbreviations: CT, chemotherapy; TNBC, triple negative breast cancer

^*others included 3 TNBC cases from London and Veridex studies.

We grouped the 269 eligible patients into 4 main sets according to patient sample size and patients’ characteristics: set I, Rotterdam (www.ncbi.nlm.nih.gov/gds, GSE2034, GSE5327, GSE12276) (n=87, almost all were node-negative and not treated with chemotherapy); set II, Frankfurt (GSE31519) (n=53, mixed node status); set III, New York (GSE2603)/San Francisco (available at www.ebi.ac.uk/arrayexpress with accession number E-TABM-158)/Stockholm (GSE1456)/Uppsala (GSE3494, GSE6232, GSE4922, GSE2990) (n=78, mixed nodes status in each subset); and set IV, Mainz (GSE11121)/TransBIG (GSE7390)/Oxford (GSE2990, GSE6532)/others (GSE12093, GSE9195, GSE6532) (n=51, almost all were node-negative and not treated with chemotherapy).

The normalization and rescaling of 269 samples to our discovery and validation cohorts were based on a median rank score based method(14) using ArrayMining online tools.(15) Predictions were generated by applying the exact SVM model that has been learned and validated from discovery and validation cohorts respectively.

Statistical analysis

In MDACC set, the study endpoint was distant relapse-free survival (DRFS), which was calculated from initial diagnostic biopsy of breast cancer to the occurrence of distant metastasis or non-breast cancer death. In TMH-BCM set, the study endpoint was relapse-free survival (RFS), calculated from initial diagnosis to the occurrence of local and regional recurrence, distant metastasis, or non-breast cancer death. Since distant metastasis is the major component of breast cancer early relapse events,(16–18) the DRFS and RFS are comparable in the first 3 years. As the relapse peak in TNBC patients occur within the first 3 years after surgery, the 3-year DRFS/RFS was calculated and evaluated. The log-rank test was used for comparison of differences between survival curves derived by the Kaplan-Meier method.

Predictive performance was assessed by the positive predictive value (PPV), defined as the cumulative relapse and death rate for patients predicted to relapse or death in 3 years; the negative predictive value (NPV), defined as the DRFS or RFS for patients predicted to be free of relapse or death within the first 3 years. The hazard or survival was calculated from the Kaplan-Meier estimators of the survival function based on cumulative events. Confidence intervals (CIs) for NPV and PPV were based on the Greenwood variance estimate. The independent prognostic value of signature was assessed in multivariate Cox regression analysis using the likelihood ratio test. The corresponding hazard ratio (HR) was calculated by the Cox model. Statistical analyses were performed in Stata 12.0 (StataCorp LP, TX). Two-sided P<0.05 was considered statistically significant.

Results

Establishment and performance of the prognostic signature in discovery cohort

To determine if chemo- prognostic predictors exist, we first examined the MDACC cohort. The study flow chart is shown in Figure 1. For the discovery cohort, 49 of 111 TNBC samples from breast cancer women treated with neoadjuvant chemotherapy were used. This cohort had a median follow-up of 25 months, with overall 3-year DRFS of 34.4% (95% CI, 20.1%–49.2%). We compared the relapsed cases (n=29) with non-relapsed cases (n=20) by dChip(19) and identified 246 genes significantly differentially expressed between the two groups, with at least a 2.14 (2^1.1)-fold difference for the ratio, with P<0.01. The gene set enrichment analysis was also employed to find the differentially expressed genes (see details in the Methods in Appendix). Afinal 7-gene signature with a minimal number and maximal prediction ability was determined.

Figure 1. Flow chart of decision algorithm used in the establishment and validation of prognostic signature in patients with chemo-insensitive triple negative breast cancer.

Abbreviations: NCT, neoadjuvant chemotherapy; pCR, pathologic complete response; RCB indicates residual cancer burden; TNBC, triple negative breast cancer

*sufficient follow-up is needed for the cases without relapse event.

The seven genes were AR (androgen receptor), ESR2 (estrogen receptor 2), GATA3 (GATA binding protein 3), GBX2 (gastrulation brain homeobox 2), KRT16 (keratin 16), MMP28 (matrix metallopeptidase 28), and WNT11 (wingless-type MMTV integration site family, member 11) (Table A1). The basal marker KRT16, stem cell maker WNT11, and epithelial-to-mesenchymal transition (EMT) marker MMP28, integrally defined a subset of TNBC with unfavorable prognosis.

In contrast, luminal hormone receptor AR and luminal marker GATA3 were relatively high expressed in TNBC tumors with favorable prognosis (Figure 2a). GATA3 is recognized as a marker of luminal ER-positive breast tumor and there is a strong relationship between co-expression of ERα and GATA3, as reported in the literature.(20, 21) Here, we observed that GATA3 was moderately-to-high expressed in approximately 30% of 49 TNBC. Similarly, among the 313 HER2-negative patients (all available cases from MDACC) (Figure 2b), some ER-negative tumors expressed GATA3 as high as that in ER-positive tumors. ER-negative tumors had a wide range of GATA3 expression, compared with ER-positive ones. We further plotted the GATA3 expression in 313 patients according to PAM50-predicted subtypes,(22) and the results reconfirmed a wide range of GATA3 expression in these non-luminal tumors (Figure 2c). No obvious association between GATA3 and ESR1 expression was observed in these basal-like cases (Figure 2d). Thus, GATA 3 expression is present in a subset of TNBC.

Figure 2. Expression of featured genes in the signature.

a, expression of two luminal-like genes (AR and GATA3) in 49 chemoresistant triple negative breast cancer (TNBC) in the MDACC cohort. Expressions of ESR1 and ERBB2 are shown as reference for “low-expression” and “non-overexpression”, respectively. b, the association between ESR1 and GATA3 in 313 HER2-negative breast cancer (all available cases from MDACC). c, box plot of expression of GATA3 in 313 patients according to PAM50-predicted subtypes. d, association between ESR1 and GATA3 in 313 patients according to PAM50-predicted subtypes. e, association between the 7-gene defined subgroups and Pietenpol’s TNBC subtypes classification in 587 TNBC in Pietenpol’s study. BL1 and BL2, basal-like 1 and 2; IM, immunomodulatory, M, mesenchymal, MSL, mesenchymal stem-like; LAR, luminal AR. The color scale is also shown: the red color represents expression level above mean expression of a gene across all samples, the white color represents mean expression, and the blue color represents expression lower than the mean.

In the discovery set, the 7-gene prognostic signature had PPV of 95.4% (95% CI, 81.7%–99.6%) and NPV (DRFS) of 100% (95% CI, 80%–100%) for the first 3 years after diagnosis (Table 3). Compared with other clinicopathological factors available, the 7-gene signature was the only factor that could effectively predict the outcome of TNBC patients with residual disease after neoadjuvant chemotherapy (log-rank P for 7-gene signature, <0.001; for age, 0.301; for tumor size, 0.114; for nodes status, 0.810; and for grade, 0.737).

Table 3.

Performance of signature for predicting prognosis of patients with chemoresistant triple negative breast cancer

	Discovery cohort	Validation cohort	Extended validation cohort
	MDACC	TMH-BCM	Overall	Set I, Rotterdam	Set II, Frankfurt	Set III, New York/San Francisco/Stockholm/Uppsala	Set IV, Mainz/TransBIG/Oxford/others
PPV (95% CI) at 3-year (CDRR or CRR)	95.4 (81.7 to 99.6)	75.0 (40.8 to 91.2)	51.5 (41.0 to 63.0)	66.7 (49.1 to 83.2)	56.3 (34.4 to 80.2)	34.8 (19.2 to 57.7)	41.7 (19.1 to 73.0)
NPV (95% CI) at 3-year (DRFS or RFS)	100 (80.0 to 100.0)	76.9 (52.5 to 94.4)	78.9 (72.4 to 84.1)	73.3 (60.2 to 82.7)	77.0 (59.0 to 87.9)	83.6 (70.9 to 91.1)	82.1 (66.0 to 91.0)
Univariate log rank P value at 3-year	<0.0001	0.009	<0.0001	<0.0001	0.006	0.063	0.066
Multivariate hazard ratio at 3-year*	/	4.67 (1.27 to 17.15)	2.61 (1.52 to 4.49)	4.00 (1.21 to 13.20)	3.55 (1.77 to 7.14)	2.40 (0.93 to 6.23)	2.46 (0.77 to 7.85)
Univariate log rank P value during follow-up	<0.0001	0.017	<0.0001	0.001	0.013	0.053	0.105
Multivariate hazard ratio during follow-up*	/	3.39 (1.14 to 10.10)	2.11 (1.33 to 3.34)	2.92 (1.51 to 5.65)	3.08 (1.10 to 8.63)	1.64 (0.66 to 4.09)	1.88 (0.75 to 4.75)

Abbreviations: CDRR, cumulative distant relapse rate; CI, confidence interval; CRR, cumulative relapse rate; DRFS, distant relapse-free survival; RFS, relapse-free survival; NPV, negative predictive value; PPV, positive predictive value

^*multivariate hazard ratio was adjusted for age, lymph nodes status, tumor size stage, grade, and neoadjuvant chemotherapy regiment in TMH-BCM cohort; for age, lymph nodes status, tumor size stage, grade, adjuvant chemotherapy, and subset group in the overall extended validation cohort; for age, tumor size stage, and grade in set I and IV; for age, lymph nodes status, tumor size stage, grade, and adjuvant chemotherapy in set II and III.

Association between chemoresistant prognosis-relevant subgroups and the Pietenpol’s molecular subtypes

As mentioned above, the two prognosis-relevant subgroups (early relapse vs. non-relapse) could be molecularly defined by the 7-gene signature. The subgroup expressing high luminal-like genes (AR, GATA3) was associated with good prognosis, while the subgroup expressing high cancer stem cell-like genes (WNT11, MMP28) was related to early metastasis. Recently, Pietenpol and colleagues(3) identified six TNBC subtypes including two basal-like (BL1/2), an immunomodulatory (IM), a mesenchymal (M), a mesenchymal stem-like (MSL), and a luminal AR (LAR) subtype. We examined the expression of our seven genes in their 587 TNBC samples (Figure 2e). The values for AR, GATA3, and KRT16 were higher than the rest of the genes. There was a clear absence of GATA3 in the M and MSL subtypes. AR and GATA3 were enriched in LAR subtype, while WNT11 and MMP28 were commonly expressed in M and MSL subtypes.

Performance of the prognostic signature in validation cohort

Independent validation was conducted in a second cohort from TMH-BCM which included 25 TNBC. This cohort was followed up for median 36 months, with 3-year RFS of 48.0% (95% CI, 27.8%–65.6%).

In this independent cohort, this 7-gene signature predicted correctly prognosis for 9 out of 12 patients predicted not to relapse in 3 years (NPV [RFS], 76.9%), and for 10 out of 13 patients predicted to relapse in 3 years (PPV, 75.0%) (Table 3, Kaplan-Meier plots in Figure 3a). Thus, the 3-year RFS estimate for the patients predicted to have good prognosis was 76.9%, compared to those predicted to relapse within 3 years was only 25.0%. Similarly, the likelihood ratio for relapse vs. absence of 3-year relapse was 4.67 (95% CI, 1.27–17.15), after adjustment for other clinicopathological factors (Table 3).

Figure 3. Kaplan-Meier estimates of relapse-free survival according to 7-gene chemoresistant prognostic signature in validation cohorts.

a, validation in the TMH-BCM cohort; b, extended validation in the whole cohort of operative triple negative breast cancer without neoadjuvant chemotherapy but predicted to be chemo-insensitive by a JAMA predictor. c, d, e, and f, four subsets within the whole cohort of operative triple negative breast cancer. Vertical ticks on the curves indicate censored observations.

Extended validation in operable TNBC patients treated with adjuvant chemotherapy

The 7-gene signature was useful in predicting the prognosis of TNBC patients with known resistance to neoadjuvant chemotherapy. The utility in the adjuvant TNBC was unclear. In order to validate the utility of this signature in adjuvant TNBC patients, chemoresistance to treatment was first determined. The previously established signature(7) which could discriminate between chemoresistant (RCB-II/III) and chemosensitive (pCR or RCB-I) in ER-negative and HER2-negative patients was utilized. As expected, the 7-gene signature could not accurately predict the prognosis in patients predicted to be chemosensitive (log rank P=0.172; data not shown). In contrast, the 7-gene signature discriminated well in women predicted to be chemoresistant, either in the overall cohort (Figure 3b) or in each subset (log rank P significant in set I and II; borderline in set III and IV; Figure 3c–3f).

Regarding the degree of accuracy, patients predicted not to relapse exhibited high 3-year RFS (NPV) of 78.9% (95% CI, 72.4%–84.1%), compared to only 48.5% (95% CI, 37.0%–59.0%) for those predicted to relapse (calculated by “1-PPV”) (Table 3). The results were concordant in each set, indicating the robustness of prediction. Moreover, the prediction of relapse by our signature was independent to clinicopathological factors such as nodal status, tumor size, age, etc (Figure A1). After adjustment, the 7-gene signature was independently and significantly associated with risk of relapse in 3 years among adjuvant TNBC predicted to be chemoresistant (HR=2.61, 95% CI, 1.52–4.49). Of note, the 7-gene signature had limits in predicting long-term relapse beyond 3 years and the predicted results were less reliable beyond this time frame. Relapse in TNBC after median follow-up of 3 years is rare, and the loss of prognostic accuracy of this signature most likely reflects small sample sizes.

Discussion

Although TNBC patients with residual disease after neoadjuvant chemotherapy have worse survival than those with luminal subtypes,(5) some of them do not relapse for a long time. In this study, we used gene expression data of TNBC patients with residual disease and different prognosis to molecularly define the clinically relevant subgroups, and developed a 7-gene prognostic signature for chemoresistant TNBCs. A favorable prognosis was observed in patients with TNBC tumors displaying high expression of “luminal-like” genes (AR, GATA3), while decreased survival was observed in patients with TNBC tumors expressing cancer stem cell-like (WNT11) or EMT-associated genes (MMP28). The signature not only predicted 3-year RFS, but also showed a clinically meaningful survival difference between patients predicted to relapse vs. no relapse. Furthermore, the signature is the only significant marker that can effectively predict prognosis of chemoresistant TNBC in a multivariate clinicopathological model (including age, tumor size, nodal status, grade, and adjuvant chemotherapy).

Although the majority of TNBCs classified as basal-like,(1, 3, 23) the clinically diagnosed TNBC is a heterogeneous collection of distinct phenotypes.(3) Our study, unlike previous reports, focuses on only chemoresistant TNBC and subdivides these cancers according to relapse outcomes. A simple combination of luminal-like genes and cancer stem cell-like genes defines the subgroup of TNBC with relatively favorable or unfavorable survival. Our discovery also challenge the value of non-pCR in TNBC and the universal applicability of the concept that non-pCR in TNBC equals to recurrence or poor survival.

AR and its ligand androgens may have some essential role in breast cancer.(24) AR expression was found in 20%–30% of the cases with TNBC.(25, 26) Most studies confirm a significantly positive correlation between AR expression and favorable survival in TNBC patients.(3, 26, 27) Our study suggests a relatively favorable prognosis in chemoresistant TNBC patients with higher expression of AR. Several novel and druggable pathways, including AR, are being studied in TNBC patients.(3) Another marker defining the favorable prognosis is GATA3. Previously, studies have shown that GATA3 expression is highly correlated with ERα (encoded by ESR1).(20, 21) We confirmed a high coincidence between GATA3 and ESR1 at mRNA level; however, when ESR1 is lowly expressed, the range of GATA3 expression is wide and ~30%–40% of TNBCs have moderately-to-high expression of GATA3. Low GATA3 expression is associated with aggressive phenotype, and in most studies, worse RFS.(28–31) Increasing evidence indicates that the role of GATA3 is not ER-dependent and that GATA3 is functional in TNBC cells.(29, 31, 32) Expression of GATA3 re-programs TNBCs to a less aggressive phenotype.(30)

The subgroup with unfavorable prognosis is characterized by the stem cell-like and EMT-associated genes. Inhibitors of WNT/β-catenin are of great interest for such a subtype and they currently are in preclinical development.(33)

Central to this study is whether chemotherapy is still needed for the TNBC subgroup with relatively favorable prognosis. With only a 78% 3-year RFS, chemotherapy as yet cannot be avoided. Generally, by consensus, low-clinical risk group is defined as patients with 10-year overall survival probabilities of at least 92% for ER-negative tumors.(34)Thus, even for TNBC patients predicted not to relapse, some alternate therapy to further decrease the risk of relapse is needed. However, the nature of these chemoresistant cases implies limited benefits from standard chemotherapy. Novel treatment strategies based on the biological features of chemoresistant TNBC need to be developed. According to our study, there are main two entities in chemoresistant TNBCs: one is AR-related luminal-like tumors and the other is stem cell-like tumors. For the former, an AR antagonist might be more effective than traditional chemotherapy;(3) for the latter, targeting proteins involved in cell-renewal or EMT may provide a more reasonable therapeutic strategy(3) since chemotherapy may not effectively eliminate tumor-initiating cells.(35)

Our observation is important since most currently available genomic prognostic signatures (e.g., 70-gene profile, Recurrence Score, Genomic Grading Index) assign poor prognostic risk status to all TNBC samples despite their variable outcomes. A few signatures have been developed to allow prognostic stratification of TNBC cancers with consideration of the chemosensitivity of the tumors.(12, 36) The implication of our study is that, ER/PgR/HER2 biomarkers have some limitation in defining a subtype with similar biological behavior and a TNBC patient could have received an inadequate or untargeted treatment and a more accurate evaluation of TNBC biology before planning neoadjuvant/adjuvant treatment is needed. Our signature, for the first time, considers chemosensitivity and excludes chemosensitive cases who achieve pCR and have excellent prognosis,(5) and focuses primarily on the chemoresistant tumors. Our 7-gene signature has the potential to assist treatment decision-making (e.g., guide to participate appropriate clinical trials) and predict clinical outcomes for chemoresistant TNBC. Of note, our signature should be utilized only in patients proven or predicted to be chemoresistant. There is a need for studies introducing molecularly targeted therapies in the adjuvant management of TNBC patients and the strategies to prospectively validate the signature as well as the novel therapeutic approach.

This study has several limitations. First, the sample size in the discovery cohort and in the homogeneous validation cohort is limited. Although our signature is successfully validated in the extended validation, further optimization is needed. Second, we used the normalized gene expression data as provided in public databases;(12) no attempts to renormalize the microarray data were made, although a robust rescaling procedure ensured that the gene expressions were similarly distributed across datasets.

In conclusion, we have developed a prognostic classifier specific to chemoresistant TNBC. It is derived from TNBC patients receiving neoadjuvant chemotherapy, and is further validated in patients with either locally advanced disease or operable tumors. This signature outperforms the classical clinicopathological features in predicting the prognosis of chemoresistant TNBC. More importantly, biologically relevant genes included in the signature might provide new potential therapeutic targets. Further validation in a large prospective series and additional research on new therapeutic strategy for chemoresistant TNBC is warranted.

Translational Relevance.

Although triple-negative breast cancer (TNBC) patients with residual disease after neoadjuvant chemotherapy have worse survival than those with luminal subtypes, some of them do not relapse for a long time. In this study, we used gene expression data of TNBC patients with residual disease and different prognosis to molecularly define the clinically relevant subgroups, and developed a 7-gene prognostic signature for chemoresistant TNBCs. A favorable prognosis was observed in patients with TNBC tumors displaying high expression of “luminal-like” genes (AR, GATA3), while decreased survival was observed in patients with TNBC tumors expressing cancer stem cell-like (WNT11) or EMT-associated genes (MMP28). The signature not only predicted 3-year RFS, but also showed a clinically meaningful survival difference between patients predicted to relapse vs. no relapse. This signature outperforms the classical clinicopathological features in predicting the prognosis of chemoresistant TNBC. More importantly, biologically relevant genes included in the signature might provide new potential therapeutic targets.