NK Cell Infiltrates and HLA Class I Expression in Primary HER2⁺ Breast Cancer Predict and Uncouple Pathological Response and Disease-free Survival

This article is featured in Highlights of This Issue, p. 1439

HER2 overexpression and/or HER2 gene amplification occur in approximately 15% to 20% of breast tumors and are associated with aggressive disease (1). Combination of chemotherapy and anti-HER2 mAbs, trastuzumab as standalone or trastuzumab and pertuzumab, is the prevailing neoadjuvant approach for patients with primary HER2-positive breast cancer (2, 3). Achievement of pathologic complete response (pCR) to neoadjuvant treatment has been associated with improved disease-free and overall survival (OS); nonetheless, 35% to 50% of patients do not achieve pCR and/or eventually relapse. On the other hand, a subgroup of patients presents excellent clinical outcomes to anti-HER2 mAbs in the absence of concomitant chemotherapy (2, 3). This heterogeneity in clinical efficacy highlights: (i) the importance of further understanding the mechanisms of action underlying anti-HER2 mAb–based treatment efficacy in different patients, and (ii) the need for biomarkers aiding in patient stratification for tailoring treatment escalation or deescalation.

Approved anti-HER2 mAbs are immunoglobulins of the G1 subclass (IgG1) that specifically block HER2-mediated oncogenic signaling and can trigger antitumor immunity by engaging Fcγ receptors (FcγR) expressed by immune cells (4, 5). In experimental models, tumor antigen–specific mAbs rapidly kill tumor targets via FcγR-mediated cytotoxicity (ADCC; refs. 6–8). This short-term process also induces a vaccine effect, linking ADCC with the development of long-term antitumor adaptive immunity (9). Natural killer (NK) cells are innate lymphocytes capable of recognizing antibody-coated targets through the activating Fcγ receptor CD16A (FcγRIIIA). CD16-mediated NK-cell activation triggers the release of cytotoxic mediators, proinflammatory cytokines, and immune cell–recruiting chemokines (10–12). The contribution of NK cells, together with CD8 T cells, to anti-HER2 mAb activity (13) and their role as commanders of antitumor adaptive responses has been recently established in in vivo models (14, 15). In patients with breast cancer, the presence of tumor-infiltrating lymphocytes (TIL) has been associated with response to neoadjuvant treatments including anti-HER2 mAbs (16, 17) and increased tumor infiltration by NK cells after trastuzumab–docetaxel and the antibody–drug conjugate T-DM1 therapies has been reported (18–20). Despite these evidences, clinically relevant predictive biomarkers are lacking.

Expression of HLA class I molecules is an additional parameter determining the susceptibility of cancer cells to NK and CD8 T-cell recognition. Indeed, tumor–antigen (neoantigen) presentation by HLA class I is required for triggering tumor-specific CD8 T-cell cytotoxicity, a pathway commonly hijacked in different tumor types, including breast cancer (21–23). HLA-I downregulation facilitates NK-cell–mediated tumor cell recognition, whereas high HLA-I expression may repress NK-cell activation against tumors through interaction with inhibitory receptors of the KIR, CD94/NKG2, and LILRB1 families (10–12). Previous studies have reported high HLA-I expression as well as total HLA class I loss as good prognosis indicators in breast carcinomas (23–25).

We hypothesized that tumor-infiltrating NK cells and HLA class I expression in breast carcinomas could influence the efficacy of anti-HER2 mAb–based neoadjuvant treatment.

This study includes pretreatment biopsy samples from two cohorts of patients with primary HER2-positive breast cancer. The discovery cohort included patients prospectively recruited along 2014–2016 at Hospital del Mar (Barcelona, Spain; n = 42). A retrospective cohort including patients recruited along 2008–2013 at Hospital del Mar (Barcelona, Spain) and Hospital Fundación Jiménez Díaz (Madrid, Spain) was used as validation cohort (n = 71; cohort diagram in Supplementary Fig. S1). HER2-positive subtype classification was defined following 2013 ASCO/CAP guidelines (26). All patients received a neoadjuvant combination treatment of standard chemotherapy and anti-HER2 mAbs. The primary efficacy endpoint was pCR defined as ypT0ypN0 based on histopathologic analysis of the resection specimen (27); disease-free survival (DFS) as the time from surgery until any breast cancer relapse or death by any cause and distant metastasis–free survival (DMFS) as the time from surgery to distant metastasis occurrence were secondary efficacy endpoints.

The study was conducted following Declaration of Helsinki guidelines. All patients gave written informed consent for the analysis of tumor biopsies for biomarker assessment. This study was approved by the Hospital del Mar Ethics Committee (2013/5307) and is reported according to the REMARK guidelines.

Stromal tumor–infiltrating lymphocytes were quantified on hematoxylin and eosin sections of pretreatment tumor biopsies following the guidelines of the International TIL Working Group (17). NK cells were identified as CD56⁺CD3⁻ cells in tumor stroma–immune infiltrates by double IHC staining. Briefly, heat antigen retrieval was done in pH 9 citrate–based buffered solution and endogenous peroxidase was quenched. Mouse monoclonal anti-CD56 (clone 123-3, Dako-Agilent) and rabbit monoclonal anti-CD3 (2GV6, Ventana-Roche) antibodies were used, followed by incubation with a polymer coupled with peroxidase (UltraView, Ventana-Roche). Sections were visualized with 3,3′-diaminobenzidine and alkaline phosphatase, and counterstained with hematoxylin. All incubations were performed at Ultra Platform (Ventana-Roche). CD56 and CD3 were evaluated by a computerized measurement using a DM2000 Leica microscope equipped with the Nuance FX Multispectral Imaging System (PerkinElmer). A computer-aided analysis yielded quantitative data of CD56 and CD3 on pseudofluorescence images. CD56-positive CD3-negative cells were counted in 50 adjacent ×400 microscopic fields considering tumor stroma and expressed in absolute numbers. TILs and NK-cell scoring was performed by an expert pathologist (F. Rojo, Institut Hospital del Mar d'Investigacions Mèdiques, Barcelona, Spain.) blinded to clinical data.

HLA-I expression was evaluated in baseline biopsy sections using the mouse mAb HC10, kindly provided by Prof. Dr. Hidde Ploegh (Whitehead Institute for Biomedical Research, Cambridge, MA). HC10 mostly reacts with HLA-B and HLA-C heavy chains and some HLA-A (HLA-A10, -A28, -A29, -A30, -A31, -A32, and –A33; ref. 28). Staining of stromal cells served as internal positive control for HLA-I antibody reactivity. A semiquantitative histoscore was calculated for HLA-I by estimation of the percentage of cells positively stained with low, medium, or high staining intensity. The final score was determined after applying a weighting factor to each estimate. The following formula was used: histoscore = (low%)X1 + (medium%)X2 + (high%)X3; the results ranged from 0 to 300. HLA-I staining H score was separately quantified in tumor and stromal cells as well as in normal epithelium, when available. Four tumor groups were defined on the basis of HLA-I H score quartiles in transformed cells: (i) HLA-I negative (HLA-I^neg) group: tumors with HLA-I H scores ≤3 (n = 28); (ii) HLA-I low (HLA-I^low) group: HLA-I H scores >3 and ≤ 60 (n = 25); (iii) HLA-I normal (HLA-I ^norm) group: H scores >60 and ≤180 (n = 32), and (iv) HLA-I high (HLA-I^high) group: H score >180 (n = 20).

RNAseq and clinical data of the HER2-positive breast cancer cohort (n = 190) generated by the Cancer Genome Atlas (TCGA, http://cancergenome.nih.gov/abouttcga) were downloaded from the Firebrowse latest available release (2016/01; http://firebrowse.org). Relative infiltration levels of 16 immune populations were computed by gene set variation analysis (GSVA) through the gsva R package using nonoverlapping gene sets specifically overexpressed in each cell type as described previously (29). The GSVA produced normalized enrichment scores ranging from −1 to 1, which represent the abundance of the immune cell population in the sample relative to the remaining analyzed cohort. Hierarchical clustering of resulting immune infiltration patterns was done using the Ward linkage method (through Scipy Python library). Correlations between the relative abundance of cell populations across tumors were computed through a linear regression, using the Scipy Python library.

Bivariate analyses by Mann–Whitney U test were used to compare continuous variables in patients achieving pCR or not. pCR OR for TILs and tumor-infiltrating NK (TI-NK) cell numbers calculated by binary logistic regression are reported for 10% increments or individual units, respectively. Multivariate analysis was conducted with binary logistic regression in which patient baseline clinical and tumor characteristics: age (as a continuous variable), tumor size (T1 and T2 vs. T3 and T4), lymph node stage (LN⁺ vs. LN = 0), estrogen receptor status (ER⁺ vs. ER⁻), tumor grading (G1 and G2 vs. G3), and Ki67 status (Ki67<20% vs. Ki67 ≥20%) were added in the model separately. Thresholds for stromal TILs and TI-NK cells that best discriminated pCR were determined using receiver operating characteristic (ROC) curve analysis in the discovery cohort. Optimal cut-off points for TIL score and TI-NK cell numbers were determined by seeking the maximum Youden index (J = sensitivity+specificity-1). Predictive effects over pCR of TILs and TI-NK cells as categorical variables were calculated by Fisher exact test. Cox proportional hazards regression was used to estimate the HRs of TILs and TI-NK biomarkers in the analysis of DFS. Kaplan–Meier curves for DFS and DMFS were used to compare time to event in patients categorized according to pCR, TILs, or TI-NK cells analyzed by the log-rank test. All P values were two sided; P values lower than 0.05 were considered significant. Statistical analysis was performed using GraphPad Prism version 6 (GraphPad software) and STATA version 15 (STATA Corp.).

We hypothesized that TI-NK cell numbers could serve as predictive biomarkers of response to anti-HER2 mAb–based therapy. For that purpose, TILs and TI-NK cells were analyzed in diagnostic core biopsies from two patient cohorts with HER2-positive breast cancer presented as discovery and validation datasets (cohort characteristics in Table 1). NK cells, identified as CD56⁺CD3⁻ lymphocytes by double IHC, were mainly located in tumor stroma areas immersed within T-cell (CD3⁺) infiltrates (Fig. 1A). Eventually, intratumor NK cells in close proximity to transformed epithelia could also be detected, yet enumeration of TI-NK cells was restricted to stroma-associated NK cells.

In the discovery cohort, TIL and TI-NK cells were detected in 100% and 65% of core biopsies. Median TIL score and TI-NK cell numbers were 20% (IQR 10–40) and 1.5 cells/50x HPF (IQR 0–15), respectively. In the validation cohort, TIL and TI-NK cells were detected in 93% and 72% of core biopsies with a median TIL score of 20% (IQR 10–30) and 2 TI-NK cells/50xHPF (IQR 0–12). Higher TI-NK cell densities were detected in ER-negative as compared with ER-positive tumors in the discovery dataset, whereas TIL scores were comparable. TI-NK cell numbers were not consistently related to any other classic pathologic factor (Supplementary Table SI).

pCR rate was 48% and 42% in the discovery and validation cohort, respectively. Global pCR rate upon combining both cohorts was 43%. ER was the only traditional clinicopathologic factor associated with pCR in both cohorts. Baseline TIL score positively associated with pCR in the discovery, in the validation, and in the global analysis combining both patient cohorts (Fig. 1B). Baseline TI-NK cell numbers showed a remarkable association with pCR in all analyzed datasets (Fig. 1C). TIL score and TI-NK cell numbers remained significantly associated with pCR when adjusted by ER status in multivariate analysis [TILs discovery OR, 1.05 (1–1.09), P < 0.015; TILs validation OR, 1.04 (1.01–1.08), P = 0.006; TILs total OR, 1.04 (1.0–1.07), P < 0.0001; TI-NK discovery OR, 6.9 (1–40), P = 0.03; TI-NK validation OR, 1.45 (1.2–1.7), P < 0.0001; TI-NK total OR, 1.49 (1.2–1.79), P < 0.0001]. Remarkably, the association between TIL score and pCR vanished upon adjusting by TI-NK cell numbers in multivariate logistic regression [Discovery: TILs OR, 0.9 (0.76–1.08), P = 0.3; Validation: TILs OR, 1.0 (0.95–1.05), P = 0.8; Total: TILs OR, 0.9 (0.93–1.03), P = 0.4], indicating that pCR was selectively associated to NK-cell–enriched TILs.

ROC curves and cutoff values providing best discrimination between pCR and non-pCR were calculated for TI-NK cells and TILs in the discovery cohort (Supplementary Fig. S2) and used for patient stratification in subsequent analysis (Fig. 1D). ORs for pCR were higher in the group of patients with ≥50% TILs and remarkably high in tumors with ≥3 TI-NK cells/50xHPF upon adjusting by ER status in both patient cohorts (Fig. 1D). Tumors with ≥3 NK cells/50xHPF showed pCR rates of 100% and 77% in the discovery and validation cohorts, respectively. Remarkably, ≥3 TI-NK cells/50xHPF cutoff performed as a biomarker with great sensitivity, specificity, positive, and negative predictive value, superior to TILs cutoff, in both patient cohorts (Supplementary Fig. S2).

We also assessed the relationship between baseline TILs and TI-NK cells with DFS (Fig. 2). Median clinical follow-up was of 31 and 49 months in the discovery and validation cohort, respectively (Table I). Despite the limited follow-up, TI-NK cell presence (≥1/50xHPF) associated to improved DFS in the discovery, in the validation, as well as in the combined analysis of both patient cohorts (Fig. 2C, F, and I). TI-NK cell association with DFS was comparable with that of pCR in both patient datasets (Fig. 2A, D, and G). On the contrary, TIL score was not significantly associated to DFS in any of the datasets (Fig. 2B, E, and H). In an exploratory analysis of the association between TI-NK, TILs, and pCR with DMFS (Supplementary Fig. S3), TI-NK maintained its statistical significance (P = 0.03) but not pCR (P = 0.06), perhaps due to the reduced number of events in DMFS as compared with DFS. Overall, TI-NK cells predicted pCR and DFS better than TILs in patients with primary HER2-positive breast cancer treated with neoadjuvant anti-HER2 mAbs and chemotherapy.

To get an insight into the immune context associated with TI-NK cells, we estimated by GSVA the relative abundance of 16 immune cell populations [cytotoxic NK cells (NK^dim), B cells, eosinophils, macrophages, mast cells, CD56^bright NK cells (NK^bright), neutrophils, T-helper cells (Th), central memory T cells (Tcm), effector memory T cells (Tem), follicular helper T cells (Tfh), activated dendritic cells (aDC), immature dendritic cells (iDC), activated CD8 T cell (CD8), gamma delta T cells (Tgd), and regulatory T cells (Treg)] across HER2-positive breast tumors in the public TCGA dataset (n = 190). Figure 3A shows the distribution of the relative enrichment of these cell populations across 190 tumors. In agreement with IHC data, the relative enrichment in NK^dim gene signature varied in distinct tumors and tended to be higher in ER-negative as compared with ER-positive tumors (Mann–Whitney U test, P = 0.052; Fig. 3B), yet no significant association was found with other clinical parameters such as tumor size or lymph node status (data not shown). NK^dim gene set enrichment showed a positive correlation with gene signatures identifying activated dendritic and CD8 T cells, hallmarks of productive antitumor immunity, as well as B and regulatory T lymphocytes (Fig. 3A). Of note, tumor stratification according to the top quartile NK^dim gene set enrichment identified patients showing good prognosis (P = 0.008; Fig. 3C), whereas regulatory T cell (P = 0.07) or activated CD8 T-cell (P = 0.34) gene expression signatures showed no significant correlation with OS in the same dataset. Hence, NK^dim cell–associated gene expression signature may also perform as surrogate biomarker of effective antitumor immunity in primary HER2 breast carcinomas.

Because expression of HLA-I molecules regulates the susceptibility of cancer cells to NK and CD8 T-cell recognition, we next evaluated whether HLA-I tumor expression could further modulate the efficacy of anti-HER2 mAbs. A total of 107 biopsies from both patient cohorts were evaluated for HLA-I expression by IHC using the HC10 antibody and combined in the analysis of HLA-I association with clinical outcomes. The H score corresponding to tumor, stroma cells, and normal epithelia was separately annotated. The majority of transformed cells within a biopsy presented a homogeneous HLA-I labeling, though occasionally, tumor areas with distinct intensities were observed, likely reflecting tumor clone variants for HLA-I expression. HLA-I expression was variable in different tumors, ranging from complete absence to high intensity; independently of HLA-I levels in stroma or normal epithelial cells (Fig. 4A and B). As continuous variable, tumor HLA-I H score did not associate with pCR or with any tested clinicopathologic factor (Fig. 4C; Supplementary Table SII). Of note, HLA-I expression was higher in stroma from tumors achieving pCR in comparison with those not achieving it (Fig. 4C), correlating with higher TIL content (r = 0.27; P = 0.004). For the analysis of extreme HLA-I tumor phenotypes, patients were stratified in four quartiles as defined in methods. pCR rate in HLA-I^neg tumors was 57% and progressively decreased with increasing HLA-I tumor expression, being 40% in HLA-I^low and 26% in HLA-I^norm tumors, suggesting a relationship between pCR and susceptibility to NK-cell recognition. However, tumors in the HLA-I^high group showed the highest pCR rate (67%), perhaps reflecting tumor contexts in which tumor-specific CD8 T cells contributed to the efficacy of anti-HER2 mAb–based treatment (Fig. 4D). Interestingly, the group of patients with HLA-I^high tumors remained disease-free along the follow-up, regardless of prior pCR achievement, whereas patients with HLA-I^neg tumors displayed a trend for delayed relapses as compared with patients with HLA-I^low and HLA-I^normal tumors, despite showing high-pCR rates (Fig. 4E).

We next analyzed whether TILs and TI-NK cell numbers performed equally well as pCR predictors in tumors expressing different HLA-I levels. Average TIL scores and TI-NK cell numbers were comparable among tumors of all HLA-I expression groups and TI-NK cell numbers were significantly associated with pCR in all HLA-I subgroups (Supplementary Fig. S4). For subsequent analysis, the ≥3 Ti-NK cell/50xHPF cutoff was applied to the overall patient population.

Stratification of patients by tumor HLA-I quartiles generated three contexts with distinct clinical behavior (Fig. 5): (i) patients with HLA-I^high tumors: association between TI-NK cells and pCR was complete in this group of tumors with all 13 cases with baseline ≥3 TI-NK cells developing pCR posttreatment (Fig. 5A). Of note, all 20 patients remained relapse-free along the study, regardless of baseline TI-NK cell density and of pCR achievement (Fig. 5B); (ii) patients with HLA-I^{interm (low+norm)} tumors: TI-NK cells predicted pCR in 83% (15/18) of cases (Fig. 5C) and performed better than pCR in the prediction of DFS (Fig. 5D and E); tumors with HLA-I low and normal expression showed similar clinical behavior (Supplementary Fig. S6); and (iii) patients with HLA-I^neg tumors: TI-NK cells predictive performance on pCR and DFS was lower in this group of patients as compared with other HLA-I groups (Fig. 5F–H).

Despite many efforts, clinically useful predictive biomarkers of response to anti-HER2 mAb–based treatments in breast cancer remain elusive. Data of the herein reported study revealed a highly significant and independent association between tumor-infiltrating NK-cell numbers in the diagnostic core biopsy and the probability of achieving complete pathologic response to anti-HER2 mAb–based neoadjuvant treatment. Moreover, stratification of patients according to HLA-I tumor expression levels enhanced the prediction of long-term clinical outcomes, identifying patients with either decreased or increased relapse risk independently of pCR. From the clinical standpoint, our study identifies two independent and complementary biomarkers, which might support clinical decision making. Mechanistically, our data support the combined action of NK and CD8 T cells as major contributors to anti-HER2 mAb–based treatment efficacy in primary breast cancer and points to HLA-I expression as a tumor-intrinsic factor further influencing response to treatment.

A relationship between baseline TILs and anti-HER2 mAb efficacy has been described in primary HER2-positive breast cancer. However, the magnitude of the association and its clinical usefulness remains controversial. Some studies, including a recent analysis of 6 patient cohorts from Gepar trials (16), have reported a positive association between TILs as continuous variable and likelihood of pCR achievement (30, 31) or DFS (17, 32) upon anti-HER2 mAb–based treatment, whereas in others, the association with pCR was more evident in lymphocyte predominant breast tumors, defined by variable TIL cutoffs ranging from 30% to 60% in different studies (33–35). Hence, a consensus TILs cutoff associated to pCR or DFS prediction across different patient cohorts with HER2-positive breast cancer is still lacking. In our study, baseline TIL score was associated to higher pCR rates as continuous variable. Nonetheless, the optimal cutoff value determined in the discovery cohort performed with low sensitivity in the prediction of pCR and was not associated to DFS in any studied cohorts. It is recognized that TILs comprise diverse immune cell populations that might differentially influence tumor development and treatment outcome. Our study discloses TI-NK cells as an excellent biomarker labeling TIL contexts with antitumor potential in response to anti-HER2 mAb–based treatment. In fact, TI-NK cells not only associated to pCR but also underlie the association between TILs and pCR in both analyzed patient cohorts, as indicated by multivariate analysis. In addition, the quantitative approach for TI-NK cell determination allowed the identification of a precise cut-off value in the discovery cohort with a very significant predictive potential for pCR in the validation cohort, as well as in the global analysis including both patient cohorts. TI-NK cell numbers also associated with DFS in both patient cohorts. However, the follow-up of the patients was limited, precluding definite conclusions regarding the durable clinical outcome, particularly in the ER-positive subgroup. Nonetheless, the association between OS and NK-cell–rich tumors could also be detected by the analysis of NK-cell gene expression signature in HER2 tumors from the TCGA.

Beyond the biological difference between TILs and NK cells that can justify their different prognostic and predictive value, the fact that prognostic and predictive differences between TILs and NK cells could be also due to the precision on their semiquantitative versus quantitative enumeration cannot be ruled out. However, TI-NK cell numbers ranged from 0 to 30 cells in 50 high-power fields (which include roughly a total of 3 × 10³ cells) precluding their semiquantitative assessment.

In a complementary perspective, analysis of baseline HLA class I expression on tumor cells identified patients with distinct long-term outcomes, in some cases discordant for pCR (patients in tumor HLA-I high and negative quartiles). Stratification of intermediate HLA-I–expressing tumors enhanced TI-NK cell performance for pCR and DFS prediction. However, due to the very small numbers of patients in each subgroup, results should be considered hypotheses-generating and no clinical conclusions should be drawn from that.

From a mechanistic perspective, our data support the complementary action of NK and CD8 T-cell responses along tumor control by anti-HER2 mAb–based treatment. Baseline TI-NK cell presence may indicate tumor permissiveness to NK-cell infiltration/proliferation/survival. Indeed, an increase in NK-cell infiltration in tumors treated with anti-HER2 mAbs has been reported previously (18, 20, 36). Enhanced NK-cell infiltration could directly contribute to anti-HER2 mAb–triggered ADCC as well as facilitate the development of tumor-specific T-cell immunity, as recently described in preclinical models (9, 14, 15). In this regard, the positive correlation between gene expression sets identifying NK cells, activated dendritic cells, and CD8 T cells would support a role of TI-NK cells as orchestrators of effective antitumor immunity in HER2⁺ breast tumors, as recently proposed in patients with melanoma (14). In our study, the association between pCR and TI-NK cells in HLA-I^neg tumors indicates the dispensable action of CD8 T cells along tumor control in some neoadjuvant contexts. In contrast, the elevated pCR rates together with the absence of relapses in patients with high HLA-I–expressing tumors supported the decisive contribution of CD8 T-cell memory both for treatment-dependent tumor control as well as for metastasis surveillance in HER2+ breast cancer along anti-HER2 mAb therapies. Hence, patients with HLA-I negative or low tumors at diagnosis may benefit from treatments enhancing NK cell rather than T-cell antitumor immunity, whereas in patients with high HLA-I–expressing tumors substitution of chemotherapy treatment by PD1/PD-L1 checkpoint blockade, enhancing tumor T-cell immunity, could be a reasonable option.

Limitations of this study include a modest cohort size, the heterogeneity in the anti-HER2 mAb regimen as well as the lack of standardized methodology for TI-NK and HLA-I assessment. In contrast, its hypothesis-driven prospective–retrospective design and the strength of the associations reported enhance its intrinsic value. For HLA-I staining by IHC, the use of HC10 mAb, provided a general yet partial view biased toward HLA-B and HLA-C expression assessment, including β2 microglobulin–free HLA-I heavy chains (37). Despite the fact that further studies addressing the molecular mechanisms underlying the specific HLA class I alterations in distinct tumors (21, 22, 38) would be of interest, the analysis implemented here might facilitate a meaningful stratification of patients with distinct clinical behaviors for treatment adaptation. The predictive value of both biomarkers deserve validation for their potential clinical utility in HER2-positive breast cancer, likely extending to general biomarkers for patients' stratification and selection for immunotherapy.

A. Muntasell reports receiving speakers bureau honoraria from Roche. S. Servitja reports receiving speakers bureau honoraria from and is a consultant/advisory board member for Roche. M. Martínez-Garcia is a consultant/advisory board member for Roche. A. Lluch is a consultant/advisory board member for Novartis, Pfizer, Roche, Eisai, and Celgene. J. Albanell reports receiving commercial research grants and speakers bureau honoraria from and is a consultant/advisory board member for Roche. No potential conflicts of interest were disclosed by the other authors.

Conception and design: A. Muntasell, F. Rojo, S. Servitja, M. López-Botet, J. Albanell

Development of methodology: A. Muntasell, F. Rojo, S. Menéndez, J. Albanell

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): F. Rojo, S. Servitja, M. Martínez-Garcia, S. Menéndez, I. Vazquez, A. Lluch, J. Albanell

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): A. Muntasell, F. Rojo, S. Servitja, C. Rubio-Perez, M. Cabo, D. Tamborero, M. Costa-García, A. Gonzalez-Perez, M. López-Botet, J. Albanell

Writing, review, and/or revision of the manuscript: A. Muntasell, F. Rojo, S. Servitja, M. Cabo, D. Tamborero, M. Martínez-Garcia, A. Lluch, A. Gonzalez-Perez, A. Rovira, M. López-Botet, J. Albanell

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): F. Rojo, S. Servitja, S. Menéndez, J. Albanell

Study supervision: A. Muntasell, S. Servitja, A. Gonzalez-Perez

Others (Interpretation and discussion of results): A. Rovira

The authors thank the technical help of Andrea Vera, Gemma Heredia, Sara Santana, and Laura Soria. We acknowledge Xavier Duran [Consulting Service on Methodology for Biomedical Research (AMIB) at IMIM] for assistance in statistical analysis. We also thank the patients for their generous participation in the study. The authors are supported by coordinated research projects from Asociación Española contra el Cáncer (GCB15152947MELE) and Proyecto Integrado de Excelencia ISCIII/FEDER (PIE 2015/00008); M. López-Botet and A. Muntasell are supported by Worldwide Cancer Research (15-1146) and by Plan Estatal I+D Retos (SAF2013-49063-C2-1-R; SAF2016-80363-C2-1-R), Spanish Ministry of Economy and Competitiveness (MINECO, FEDER) and by Generalitat de Catalunya (2017 SGR 888); J. Albanell, A. Rovira, and F. Rojo are supported by ISCiii/FEDER (PI15/00146 and CIBERONC), by Generalitat de Catalunya (2017 SGR 507) and by Fundació Privada Cellex (Barcelona). A. Gonzalez-Perez is supported by a Ramón y Cajal contract (RYC-2013-14554). The results shown here are in part based on data generated by the TCGA Research Network.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.