Predictive Biomarkers for Adjuvant Capecitabine Benefit in Early-Stage Triple-Negative Breast Cancer in the FinXX Clinical Trial

The introduction of anthracyclines and taxanes as adjuvant therapies has drastically improved outcome in patients with early-stage breast cancer (1–3). However, despite these effective adjuvant therapies, almost a third of patients with early-stage and locally advanced breast cancer still develop recurrence and metastasis (4). This observation highlights the need for additional therapies for patients with early-stage breast cancer.

Capecitabine is a nucleoside analogue commonly used in patients with metastatic breast cancer. It is a prodrug of 5′-deoxy-5-fluorouridine that is converted to 5-fluorouracil (5FU) by three sequential enzyme activities (5). First, capecitabine is metabolized to 5′-deoxy-5-fluorocytidine (5′-DFCR) by carboxylesterase. The second step involves cytidine deaminase, which converts 5′-DFCR to 5′-deoxy-5-fluorouridine (5′-DFUR). The final step of capecitabine conversion to 5FU requires thymidine phosphorylase, which has been shown to be highly expressed in breast cancer (6, 7). Several clinical trials have evaluated the benefit of capecitabine in the adjuvant setting. However, the results of these trials are conflicting and mostly showed no significant improvement in the outcome of unselected patients in the entire population of these trials (8–13).

The Finland Capecitabine trial (FinXX; refs. 11, 13) is a phase III trial, which randomized early-stage breast cancer patients to receive standard anthracycline and taxanes-based adjuvant chemotherapy with or without capecitabine. Overall, there was no significant improvement in recurrence-free survival (RFS) or overall survival (OS) in patients who received additional capecitabine (11, 13). However, in the subset analysis, patients with triple-negative breast cancer (TNBC) had 47% improvement in RFS with an addition of capecitabine (P = 0.02). More recently, the meta-analysis of eight clinical trials, which evaluated the benefit of capecitabine in combination with standard chemotherapy in either neoadjuvant or adjuvant setting, also demonstrated a significant improvement in disease-free survival (DFS) in patients with TNBC compared with non-TNBC with the hazard ratio (HR) of 0.72 versus 1.01, P-interaction = 0.02 (14).

Another approach that has been used to evaluate the benefit of capecitabine in patients with early-stage breast cancer is in the extended adjuvant therapy in patients with residual disease after neoadjuvant chemotherapy. This particular approach has been evaluated in the CREATE-X trial (12). This trial is a phase III trial which randomized patients with early-stage human epidermal growth factor receptor 2 (HER2)-negative breast cancer to receive additional 6 months of single-agent capecitabine versus standard postsurgical treatment without capecitabine. In this particular trial, there was a significant improvement in both DFS and OS among patients who received capecitabine. In the subset analysis, the benefit of capecitabine was more evident among patients with TNBC. Based on the result of the CREATE-X trial, capecitabine became the new standard-of-care option for patients with residual disease after neoadjuvant chemotherapy.

Nevertheless, additional capecitabine is associated with significant side effects, particularly grade 3 or 4 diarrhea and hand–foot syndrome. These undesirable side effects resulted in a significantly more frequent treatment discontinuation compared with standard chemotherapy without the addition of capecitabine (odds ratio = 3.80; ref. 14). Furthermore, TNBC is a heterogeneous disease with multiple distinct molecular subtypes and is driven by underlying biological processes that involve the tumor genome, gene-expression profiles, microenvironment, and immune response (15). Therefore, it is critical to identify potential biomarkers to determine which patients are more likely to benefit from this therapy. In this study, we evaluated the potential clinical utility of a 770-gene panel as well as an additional 30-custom gene panel to discover biomarkers that predict the benefit of capecitabine in TNBC patients in the FinXX trial. The test panel and methodology used is readily applicable to small amounts of variably-handled formalin-fixed paraffin-embedded (FFPE) standard clinical specimens and comprehensively evaluates 37 biological signatures, including PAM50-intrinsic subtype (16, 17), claudin-low subtype (18), triple-negative biology (19), tumor inflammation signature (TIS; ref. 20), DNA-repair deficiency (21–23), inhibitory immune signaling (24), and immune cell population abundance (25, 26). We also included a custom panel of 30 genes that have been associated with capecitabine metabolism and function (27–30).

FinXX (11, 13) is a multicenter, randomized phase III trial that was conducted in Finland and Sweden between January 2004 and May 2007. A total of 1,500 high-risk early-stage breast cancer patients with either lymph node involvement or node-negative breast cancer with tumor size ≥20 mm and progesterone receptor expression <10% were enrolled. Patients were randomly assigned to receive either 3 cycles of docetaxel followed by 3 cycles of cyclophosphamide, epirubicin, and fluorouracil (arm A: T + CEF) versus 3 cycles of docetaxel plus capecitabine followed by 3 cycles of cyclophosphamide, epirubicin, and capecitabine (arm B: TX + CEX). Full details on the FinXX treatment protocols have been published (11, 13). A subset of patients with TNBC who participated in the FinXX trial and had adequate archival samples was included in this analysis. TNBC was defined as estrogen receptor (ER) and progesterone receptor (PR) expression <10% with no HER2 overexpression. HER2 status was determined locally in accredited laboratories based on IHC staining or a positive in situ hybridization test (31, 32).

The FinXX trial (NCT00114816) was sponsored by the Finnish Breast Cancer Group and the study protocol was approved by the Ethics Committee of the participating medical institutions and the National Agency for Medicines, Finland (Finnish Breast Cancer Group Protocol No. 01-2003). All patients have signed a written informed consent to allow the use of their tumor tissue for clinical study–related research purposes. Consent for the use of previously collected patients’ specimens was obtained under waiver of informed consent policy with no revelation of identifiable patient information. Approval for the use of archival FFPE tissue specimens in the current translational study was granted by Institutional Review Boards at Helsinki University, Finland. The study was conducted in accordance with the recognized Declaration of Helsinki ethical guidelines.

Hematoxylin and eosin slides corresponding to FFPE tissue samples from 120 TNBC patients randomized in the FinXX trial were reviewed by a breast pathologist. Areas containing viable invasive tumor cells were circled and macrodissected from 10-μm FFPE tissue slices mounted on corresponding unstained slides.

RNA was extracted from these macrodissected slides using standard extraction kits, following protocols established in a previous registration study (33). Samples were analyzed on a NanoString nCounter system using the Breast Cancer 360 NanoString 770-gene code set, which comprised of 18 housekeeping genes and 752 target genes covering key pathways in tumor biology, microenvironment, and immune response (Supplementary Data S1; ref. 34).

The 770-gene panel was designed to specifically measure 37 biologically important gene and metagene signatures that capture 10 major areas of breast cancer tumor biology using an established signature algorithm trained and validated on data sets from The Cancer Genome Atlas (TCGA; ref. 34). These 37 biologically important gene and metagene signatures that can be derived through the commercially available 770-gene panel were specifically preselected as a planned analysis in our study.

We supplemented the standard 770-gene panel with probes for 30 additional genes reported in the literature to be related to capecitabine activation and fluorouracil absorption, distribution, metabolism, and/or elimination (27–30). These 30 additional genes consisted of five related to capecitabine activation prior to its conversion to 5FU, and 25 genes related to fluorouracil absorption, distribution, metabolism, and/or elimination that are common to the activation of both capecitabine and 5FU as chemotherapy drugs (Supplementary Table S1).

Following the manufacturer's protocol, 5 μL of RNA (250 ng total) at a final concentration of 50 ng/μL per sample was hybridized overnight with the NanoString code set using the high-sensitivity protocol. Samples were subsequently processed using the sample preparation robots. Multiple samples were loaded on to 12-well cartridges and analyzed overnight on the nCounter. Data from the NanoString output files were analyzed using the nSolver software package, and R statistical software was used for gene-expression data analysis.

Scores for the genes and metagenes included in the entire 800-gene panel were computed using normalized and log transformed gene-expression data (Supplementary Data S2). Normalization in the 770-gene panel was carried out in a multistep process using both housekeeping genes and “panel standard” to control for run-to-run variation. Quality control for gene-expression data was done largely following the “NanoString Gene Expression Data Analysis Guidelines” with a modification for the normalization part. First, the geomean of the housekeeping genes was subtracted from the log₂ transformed raw counts for each gene to normalize the data. Next, the average of the log₂ transformed raw counts was calculated for each gene across the six lanes of panel standard included in this study. Lastly, the average values across the panel standard lanes for each gene were subtracted from the housekeeper normalized data, and this data set was then used to calculate the signature scores.

There are 18 housekeeper genes in the panel. Ten were used in the development of the TIS, and eight are specific to PAM50. PAM50 subtype correlations were calculated by using the eight housekeeper genes specific to that assay for housekeeper normalization and the TIS score used the 10 housekeeper genes specific to the tumor inflammation signature assay for housekeeper normalization. All other scores were calculated using 18 housekeeper genes for normalization. Scores in the 30-gene custom panel were normalized to housekeeping genes only.

Housekeeper gene expression was used as a metric for data quality. After transforming the data as described above, samples with a mean housekeeper gene-expression score below 5 indicated low quality either due to low input or inefficient reaction and were excluded. To ensure a high data quality control for gene expression, we included samples with a mean housekeeper gene expression of at least 7.

PAM50-intrinsic subtype analysis was performed to identify the luminal A, luminal B, Her2-Enriched, and basal-like prototypical breast cancer subtypes as previously published (17). The analysis of TIS was performed as previously validated to predict response to pembrolizumab (20). This signature includes 18 genes that measure CD8⁺ T-cell abundance, NK-cell abundance, antigen presentation and processing, interferon gamma (IFNγ) signaling, cytotoxicity, and coinhibitory signaling (20, 35). Analysis of “immune cell populations abundance” signatures followed prespecified previously published algorithms based on TCGA data sets and validated on immunotherapy-independent data set (25). Signature scores for additional metagenes that cover other major areas of tumor biology in the 770-gene panel were calculated using prespecified established algorithms developed by NanoString Technologies (36).

Analysis of genes and metagenes included in the entire 800-gene panel was considered hypothesis generating, aiming to identify candidate biomarkers significantly associated with capecitabine benefit in TNBC.

The expression levels of 37 selected biologically important genes and metagenes present within the 770-gene panel were performed (Supplementary Data S3). Z scores for the different 37 signatures were calculated and when relevant they were mapped to TCGA quantile value (from 0 to 1) to generate a color-scaled heat map for clustering all the samples, and to plot gene and metagene expression for each individual sample.

Χ² tests were used to assess the associations between the categorical variables of clinicopathologic characteristics and each treatment arm in the FinXX trial. The assessment of continuous expression scores of selected genes and metagenes against categorical variables was performed using Wilcoxon rank sum tests for pair-wise comparisons. The correlation between continuous variables was tested using the Pearson correlation coefficient. Exploratory univariate and multivariate survival analyses were performed for continuous expression scores of genes and metagenes using Cox regression models and stratified log-rank tests. The associations between log HR and 95% confidence intervals (CI) with unit increase in each signature score were calculated in each treatment arm and results were demonstrated using forest plots. Multivariate analysis was adjusted for pathologic tumor size T1 versus T2 or T3, nodal status negative versus positive, grade 1 or 2 versus 3, and age at randomization >50 years versus ≤50 years. The association of biomarker expression with clinical outcomes within each treatment arm could report both prognostic and predictive effects. However, for this study, we focused on the predictive effect as our main analysis. To more clearly assess the predictive effects of gene and metagene expression with treatment arm, an interaction test of heterogeneity that tests the associations of biomarker expression with clinical outcomes between treatment arms was used.

Prognostic analyses of clinical outcomes in relationship to continuous increase in the scores of selected genes and metagenes were performed for the entire cohort (two arms combined).

Consistent with the primary endpoint of the original FinXX trial, RFS was used as the primary endpoint in the current correlative study. RFS was defined as the time interval between the dates of randomization and detection of invasive breast cancer recurrence (local or distant), or death if the patient died prior to documented recurrence (11, 13). Additional exploratory univariate and multivariate analyses were performed for categorical expression scores of selected genes and metagenes using the median score as a cutoff point. Kaplan–Meier curves and forest plots were used to display survival outcomes according to gene or metagene expression status. Multivariate analysis was adjusted for pathologic tumor size, nodal status, grade, and age at randomization as described for the analysis of the continuous expression scores. HR and 95% CI were derived using Cox regression models. When the number of censored cases exceeded 80%, the Cox regression Firth's penalized likelihood method was applied. All tests were performed two-sided at a significance level of 0.05. Analyses including multiple comparisons were adjusted for multiple testing using the Benjamini–Hochberg method to control for false discovery rate. Statistical analyses were conducted using R statistical software. Functional term enrichment analysis for genes found to be significantly associated with capecitabine benefit was performed using the DAVID bioinformatics Gene Ontology Tool (37) and selected for biological process annotation with a Fisher exact test and Benjamini–Hochberg adjusted P value. A false discovery rate of <0.05 was considered statistically significant.

Archival FFPE tumor tissue specimens from 120 of the 202 TNBC patients enrolled in the FinXX trial were available and had sufficient RNA isolated. Among these, 111 specimens passed the quality control assessment for gene-expression analysis. The CONSORT flow diagram demonstrating the cases included in this translational study is depicted in Supplementary Fig. S1. The current study cohort consisted of 57 patients treated without capecitabine in arm A: T + CEF and 54 patients treated with capecitabine in arm B: TX + CEX. The median follow-up of this cohort was 10.2 years (range, 2 months–12 years), 10-year RFS rate was 77%, and 10-year OS rate was 79%. There were no significant imbalances in clinical and pathologic characteristics observed between the two study populations (Supplementary Table S2). In addition, the current translational study cohort was representative of the TNBC subpopulation (n = 202) of the original FinXX trial (Supplementary Table S3).

Using the 770-gene RNA expression profile, we explored the expression levels of 37 selected biologically important individual genes and metagenes within the 111 TNBC cases. Overall, high expression levels of signatures associated with DNA damage repair including homologous recombination deficiency, BRCAness, p53, proliferation, and hypoxia were observed (Fig. 1). In contrast, expression levels of genes and metagenes associated with luminal A subtype, estrogen signaling, ESR1, PGR, FOXA1, AR, differentiation, and angiogenesis were lower in the majority of cases (Fig. 1). A wide variability of expression levels was observed for stromal, apoptosis, and claudin-low signatures. In addition, gene-expression plots based on the expression level of the biologically important selected genes and metagenes were produced for each individual sample (Supplementary Fig. S2).

Classification of the study cohort into different PAM50-intrinsic subtypes revealed that most of the cases (78%; n = 86) were assigned as basal-like subtype, 14% (n = 16) as Her2-enriched, 5% (n = 6) as luminal A, and 3% (n = 3) as luminal B (Supplementary Table S2).

Given the well-established evidence in the literature demonstrating that the basal-like subtype of breast cancer is the most distinctive gene-expression subtype within TNBC and most likely to display higher counts of immune-related genes (19), we performed a comparison between tumors classified as basal-like versus other subtypes based on the PAM50 classifier. Expression levels of immune-related genes and metagenes, including antigen-processing machinery (APM), TIS, CD8⁺ T cells, cytotoxic cells, IDO1, PDL1, PDL2, were significantly higher among cases classified as basal-like (Supplementary Fig. S3).

A classification of our cohort according to the four previously published TNBC molecular subgroups (19), including luminal/androgen receptor, mesenchymal, basal-like/immune suppressed, and basal-like/immune activated, revealed that 73% of the tumors were classified as basal-like/immune activated. However, the tumors classified as basal-like/immune suppressed constituted 15% of the cohort in our study (Supplementary Table S4).

Analysis according to TNBC clinical status showed that RFS among patients who received capecitabine in arm B: TX + CEX compared with arm A: T + CEF had a HR of 0.55 (95% CI, 0.23–1.31, P = 0.20; Supplementary Fig. S4).

An exploratory univariate analysis for RFS was conducted to assess the continuous increase in the scores of 37 selected biologically important individual genes and metagenes in the 770-gene panel according to treatment arm (Fig. 2). When comparing arm A (T + CEF) with arm B (TX + CEX), expression levels of PDL1, PDL2, IDO1, cytotoxic cells, mast cells, and endothelial signatures were significantly associated with improved survival on capecitabine compared with the no capecitabine arm by unadjusted P value (Fig. 2 and Supplementary Data S4). However, multiple testing corrections indicated none of these findings to be considered as statistically significant (Supplementary Data S4).

Higher expression levels of PDL1, PDL2, IDO1, and cytotoxic cells were significantly associated with improved outcome among patients who were treated with capecitabine in arm B: TX + CEX, whereas higher expression levels of mast cell and endothelial signatures were significantly associated with shorter RFS in patients who were treated in arm A: T + CEF (Fig. 2). These results were significant by unadjusted P value, but were not found to be statistically significant when adjusted for multiple testing corrections.

In a multivariate analysis, four genes and metagenes displayed a significantly improved RFS favoring addition of capecitabine, by interaction test. These genes and metagenes were the cytotoxic cells signature (HR, 0.38; 95% CI, 0.16–0.86, P-interaction = 0.01), endothelial signature (HR, 0.67; 95% CI, 0.20–2.22, P-interaction = 0.02), mast cell signature (HR, 0.78; 95% CI, 0.49–1.27, P-interaction = 0.04), and PDL2 (HR, 0.31; 95% CI, 0.12–0.81, P-interaction = 0.03; Table 1). The significantly higher magnitude of benefit from the addition of capecitabine in relation to T + CEF for these genes and metagenes is found in Supplementary Table S5. However, none of these results were found to be statistically significant when adjusted for multiple testing (Table 1). Results of the multivariate analysis and interaction tests for the remaining selected biologically important individual genes and metagenes in the 770-gene panel are displayed in Supplementary Table S6. Similarly, none of these results were found to be statistically significant when adjusted for multiple testing (Supplementary Table S6).

We further evaluated the predictive value of these genes and metagenes using categorical classification as “high” versus “low” based on the median cutoff. In the univariate analysis, high expression levels of PDL1, PDL2, cytotoxic cells, and endothelial signatures were associated with significantly improved 10-year RFS when patients were treated with added capecitabine in arm B: TX + CEX versus arm A: T + CEF (high PDL1: 88% vs. 58%; high PDL2: 85% vs. 58%; high cytotoxic cells: 80% vs. 63%; high endothelial: 87% vs. 52%; Fig. 3). However, these differences were not observed among patients classified as having low scores for these genes and metagenes (low PDL1: 73% vs. 63%; low PDL2: 71% vs. 64%; low cytotoxic cells: 67% vs. 69%; low endothelial: 75% vs. 58%; Fig. 3). These findings were also subsequently confirmed in a multivariate analysis with the interaction test of heterogeneity (Supplementary Fig. S5).

We then performed exploratory prognostic analyses of RFS in relationship to continuous increases in the scores of the biologically important selected genes and metagenes of the 770-gene panel in the entire cohort. Overall, when adjusted for tumor size, nodal status, tumor grade, and age at randomization, increases in the expression of IDO1 and PDL1 were found to be significantly associated with a longer RFS (Table 1). However, these findings were not statistically significant when adjusted for multiple testing.

To identify individual genes that could serve as potential predictive biomarkers for capecitabine benefit, we evaluated the association between the continuous increase in each individual gene and RFS within each treatment arm separately (Fig. 4; Supplementary Fig. S5). In arm A: T + CEF, there were 93 genes that were significantly associated with benefit or resistance by unadjusted P value, but were not found to be statistically significant when adjusted for multiplicity (Supplementary Data S5 and S6). In arm B: TX + CEX, 41 genes were associated with benefit or resistance (Supplementary Data S5 and S7). Among these 41 genes, 38 were significantly associated with capecitabine benefit by unadjusted P value, but not significant when adjusted for multiple testing (Supplementary Data S7 and S8). In addition, higher expression levels of these 38 genes were found among patients who did not experience a recurrence when compared with those who developed a recurrence on the capecitabine arm (Supplementary Fig. S6). In contrast, the remaining three genes that were associated with capecitabine resistance showed lower expression levels in patients who did not experience a recurrence (Supplementary Fig. S6).

To further evaluate the predictive capacity of these 38 candidate genes, a multivariate analysis and interaction tests of heterogeneity comparing arm A: T + CEF and arm B: TX + CEX treatment were performed. Most of these genes (28 of 38) showed a significant interaction test for capecitabine benefit (Table 2). However, these results were not found to be significant when adjusted for multiple testing.

Assessing the biological mechanisms of these 38 genes revealed that they were dominated by immune response and the STAT pathway (Supplementary Data S8), a finding that was further supported by gene functional term enrichment analysis showing that genes involved in the biological process of immune response were the most significantly associated with capecitabine benefit, at a false discovery rate <0.05 (Supplementary Data S9).

Among the five genes originally included in the customized 30-gene panel related to capecitabine activation prior to its conversion to 5FU, we identified CES1 and SLC28A1 to be among the 38 genes predictive for capecitabine benefit within the arm B: TX + CEX arm. However, none of the 25 genes in the customized 30-gene panel common to the downstream activation of both capecitabine and 5FU as a drug were found among these 38 genes. In addition, we identified TYMP—the gene encoding the thymidine phosphorylase enzyme—to be among the 38 genes that significantly predicted capecitabine benefit within the TX + CEX arm. Analysis of the correlation between TYMP expression and the 37 biologically important genes and metagenes in the 770-gene panel revealed a significantly large positive correlation (R > 0.50) with the immune-related genes and metagenes PDL1, PDL2, TIGIT, IDO1, cytotoxic cells, CD8 T cells, macrophages, APM, and major histocompatibility complex 2 (MHC2; Supplementary Fig. S7). However, a low positive but nominally significant correlation was observed between TYMP and basal and homologous recombination deficiency scores (R < 0.30; Supplementary Fig. S7). When assessing the expression of CES1 and SLC28A1 with immune-related genes and metagenes in the 770-gene panel, an overall low positive correlation was found (R < 0.30).

In the FinXX phase III clinical trial of women with early-stage breast cancer randomized to T + CEF versus TX + CEX, we present evidence that the sensitivity of TNBC to adjuvant capecitabine may be explained by interactions with the immune system.

Using a clinically applicable gene-expression panel methodology that captures the most biologically relevant information in breast cancer from standard clinical specimens, we found two immune subsets that characterize “immune-enriched” versus “immune-depleted.” These two groups primarily differed in gene expression related to antitumor immune activities, antigen processing, and immune inhibitory genes.

Our further analysis demonstrated that genes and metagenes related to antitumor immunity (cytotoxic T cells), inhibition of antitumor immune response (PDL2), mast cells, and endothelial cells were significantly associated with improved outcome among TNBC patients in the FinXX trial who received capecitabine. In our study, we have focused on assessing RFS rather than OS for capecitabine benefit, as RFS was defined as the primary endpoint in the phase III FinXX trial.

Overall, these data suggest that a subset of TNBC patients with preexisting inflamed tumors, displaying high cytotoxic cell counts as well as expression of checkpoint molecules, may be the subgroup most associated with better outcome with an addition of capecitabine. Although such an immune subset has been shown to associate with improved outcome with immune-checkpoint blockade agents (38, 39), less is known about the predictive utility of these immune signatures in the context of standard-of-care chemotherapies. 5FU-based cytotoxic drugs, such as capecitabine, have been reported to induce tumor cell death and antigen release, rendering these tumors more visible to the immune system—a process that promotes the activation of antigen-presenting cells, and subsequent presentation of tumor antigens to T cells (40, 41). As such, capecitabine, by virtue of generating higher amounts of fluorouracil preferentially in tumor tissues, could enhance immune responses that are more evident in immunogenic tumors such as TNBC.

Given the fact that capecitabine has to be metabolized into its active form by thymidine phosphorylase (an enzyme expressed preferentially in tumor tissue), one possible hypothesis is that the tumors with immune-enriched features may have increased levels of this converting enzyme resulting in a higher concentration of 5FU intratumorally. Interestingly, previous studies demonstrated that the expression of thymidine phosphorylase can be induced by several proinflammatory cytokines including IFNγ and TNFα, as well as under hypoxic conditions (6) that characterize aggressive tumors with unfavorable prognosis such as TNBC, sensitizing them to capecitabine (42). In addition, the administration of anthracyclines and taxane-based chemotherapy prior to capecitabine has been shown to upregulate the expression of thymidine phosphorylase, leading to an enhanced efficacy of the sequential administration of capecitabine (43, 44). As such, the administration of epirubicin and docetaxel in the FinXX trial could have also contributed to the upregulation of thymidine phosphorylase, leading to its high expression in a subset of breast cancer patients. Similar to previous publications (45), our study demonstrated that higher expression of TYMP was associated with improved outcome in patients who were treated with capecitabine. In addition, TYMP expression displayed a significantly large positive correlation with most of the immune-related genes and metagenes we assessed in our study. Such findings are consistent with recent reports supporting the evaluation of thymidine phosphorylase as a predictive biomarker for capecitabine benefit in early-stage TNBC (42), and in metastatic breast cancer clinical trials (46). Specifically, metastatic breast cancer patients with ER- and/or PR-positive disease have shown a significantly improved survival on capecitabine when compared with ER/PR-negative disease (47). However, predictive biomarkers that explain this finding are currently lacking, and evaluating TYMP expression might define the subset of metastatic ER-positive patients likely to respond to capecitabine.

Another hypothesis is that fluorouracil, as an active metabolite of 5FU-based drugs, has a specific immunomodulatory effect by depleting myeloid-derived suppressor cells (MDSC; ref. 48). MDSCs are immature myeloid cells that accumulate during tumor progression and suppress T-cell activation. MDSCs do so in a variety of ways, such as secretion of immunosuppressive enzymes and cytokines, including indoleamine 2,3-dioxygenase (IDO), arginase, and IL10 (49, 50). In addition, activated MDSCs have been shown to express high levels of PD-L1 (50) causing T-cell exhaustion that is further augmented by induction of regulatory T-cell expansion (51). The fluorouracil-mediated depletion of MDSCs is due to the low expression of thymidylate synthase, an enzyme targeted by fluorouracil, in MDSCs (48). MDSC depletion by capecitabine alleviates the inhibitory effects on T and NK cytotoxic cells in the tumor (52) leading to net enhanced immune activation, including high cytolytic activity and IFNγ production that could also activate components of the downstream STAT key pathway, resulting in a stronger antitumor immune response.

Our findings that tumors with high preexisting cytotoxic cells signature and checkpoint expression (PDL2) benefit the most from capecitabine may highlight tumors that underwent recognition by the immune system, but with responses attenuated partly by pathways of T-cell exhaustion. This subset may derive more benefit from capecitabine due to a potential relief of immune suppression after drug-induced MDSC depletion. Thus, inhibiting MDSCs using capecitabine might be one potential strategy to augment immunotherapy efficacy among this particular TNBC subset.

Our findings that mast cells significantly predict capecitabine benefit could be supported by recent evidence demonstrating that mast cells play a critical role in augmenting MDSC activity. Mast cells have been reported to be associated with the upregulation of different proinflammatory cytokines (e.g., TNFα and IL6) and enhance the immune-suppressive activity of MDSCs through the secretion of CCL2 and IL17 (53, 54). Thus, the subset of tumors displaying high counts of mast cells may have an enhanced MDSC activity that could gain the most benefit from MDSC depletion by capecitabine.

The identification of an endothelial signature as a significant predictor for capecitabine benefit in our study might be explained by preclinical findings reporting that 5FU-based drugs induce genes of the thrombospondin family that act as mediators of antiangiogenic effects (55). In addition, thymidine phosphorylase has been reported to enhance the migration of endothelial cells, promoting the formation of a proangiogenic tumor microenvironment (6).

With regard to other signatures and PAM50-intrinsic subtypes, our findings are consistent with previous analysis from the United States Oncology Trial 01062 (8), which demonstrated no significant improvement in DFS among patients with basal-like subtype with an addition of capecitabine (56). Although some preclinical studies have suggested that basal-like tumors might be particularly susceptible to nucleoside analogues due to their enrichment for DNA defects (57), in our exploratory analysis, continuous increase in basal scores and DNA damage repair signature scores were not found to be predictive for capecitabine benefit. In addition, these signatures displayed only a low positive correlation with TYMP expression.

Our finding that TYMP and two of the customized genes individually added to the code set were significant for capecitabine benefit is likely due to their involvement in the upstream activation of capecitabine prior to its conversion to 5FU. These customized genes were CES1—encoding an enzyme that converts capecitabine into 5′-deoxy-5-fluorocytidine—and SLC28A1 that encodes a protein involved in the transportation of the metabolite 5′-deoxy-5-fluorouridine in the liver (27).

The enzymes encoded by CES1 and SLC28A1 are involved in different steps than TYMP in the metabolic pathway of capecitabine. TYMP reflects the preferential activation of capecitabine in the tumor tissue, whereas CES1 and SLC28A1 are overexpressed in the liver. This fact could explain the observation that TYMP displayed a significantly large association with the expression of immune-related genes and metagenes, whereas such an association was low for CES1 and SLC28A1.

More interestingly, in the current study, we have identified 38 specific genes that were significantly associated with capecitabine benefit, most of which were enriched for immune response and capecitabine activation. These 38 candidate genes can help to identify the subset within TNBC patients most likely to benefit from capecitabine in the adjuvant setting. Such findings, when validated on independent data sets of larger clinical trials in a rigorous hypothesis-testing fashion, could aid the development of biomarkers and a signature that predict patient response to capecitabine.

Nevertheless, our study has several limitations. First, our study represents a discovery-based exploratory analysis, which is considered hypothesis generating rather than hypothesis testing. After quality-controlled process to verify sufficient RNA quality, there were only 111 samples with adequate RNA for the analysis. In addition, further breaking down according to different molecular subsets resulted in even smaller sample size, which limited statistical power to assess some biologically important signature scores and their interaction with treatment benefit. This limitation, along with the high number of genes (∼800) tested in our study and the small number of events identified in the entire cohort (n = 26), decreased the power to observe significant findings when correcting the results for multiple testing. The number of events was smaller than the number of biologically important signatures (n = 37) we assessed in our study and thus precluded the power to observe significant findings when corrected for multiplicity. Therefore, future studies using samples from larger clinical trials are critical to further validate our results. Additionally, some of the genes and metagenes that we have focused on assessing their predictive capacity based on continuous scores do not yet have existing predefined cutoff points that are analytically and clinically valid to allow their development in a patient-centered fashion.

Furthermore, although RNA expression allows a simultaneous, quantitative assessment of multiple genes, gene-expression analysis does not show the in situ morphologic context. Given the complexity of the immune landscape in the tumors, correlating our findings with assays that allow multiplex evaluation of cell type, the location of expressing cells within the tumor microenvironment and their coexpression pattern with other biomarkers could better define the phenotype and function of relevant subsets that relate to benefit from capecitabine.

In summary, our current study demonstrated that genes and metagene signatures related to antitumor immunity, modulation of immune response, as well as genes related to capecitabine activation such as CES1, SLC28A1, and TYMP, are associated with improved outcome among TNBC patients who received an addition of capecitabine. Moreover, we identified 38 genes that may serve as potential predictive biomarkers for capecitabine benefit in TNBC patients. Given that our study had a decreased power to observe significant findings when correcting the results for multiple testing, our findings require validation on larger clinical trials assessing capecitabine in the adjuvant setting. If successfully validated on independent series, these biomarkers may be used to identify TNBC patients who are more likely to benefit from an addition of capecitabine to standard adjuvant chemotherapy.

H.A. Brauer and A. Sullivan are employees/paid consultants for NanoString. T.O. Nielsen is an employee/paid consultant for and reports receiving speakers bureau honoraria from NanoString, and holds ownership interest (including patents) in Bioclassifier LLC. H. Joensuu is an employee/paid consultant for Orion Pharma and holds ownership interest (including patents) in Sartar Therapeutics, Neutron Therapeutics, and Orion Pharma. No potential conflicts of interest were disclosed by the other authors.

Conception and design: K. Asleh, T.O. Nielsen, E.A. Thompson, S. Chumsri

Development of methodology: K. Asleh, H.A. Brauer, A. Sullivan, S. Lauttia, H. Lindman, T.O. Nielsen, H. Joensuu

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): A. Sullivan, S. Lauttia, H. Lindman, H. Joensuu, E.A. Thompson, S. Chumsri

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): K. Asleh, H.A. Brauer, A. Sullivan, H. Joensuu, E.A. Thompson, S. Chumsri

Writing, review, and/or revision of the manuscript: K. Asleh, H.A. Brauer, A. Sullivan, T.O. Nielsen, H. Joensuu, E.A. Thompson, S. Chumsri

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): K. Asleh, A. Sullivan, S. Lauttia, S. Chumsri

Study supervision: H. Lindman, T.O. Nielsen, S. Chumsri

We thank Samantha Burugu for providing helpful comments. The research reported in this publication was supported in part by funds from the Breast Cancer Research Foundation (BCRF-17-161), Bankhead-Coley Research Program (6BC05), and the DONNA Foundation to Mayo Clinic. The research also supported in part by funds from the Canadian Cancer Society (grant #705463), Academy of Finland, Cancer Society of Finland, Sigrid Juselius Foundation, and Helsinki University Central Hospital, Finland. K. Asleh is supported by the Vanier Canada Graduate Scholarship–Canadian Institutes of Health Research. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.

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.