Activation of Host Wound Responses in Breast Cancer Microenvironment

The concept of cancers as wounds that do not heal arose >20 years ago based on the observation of gross cellular and molecular similarities between wounds and cancer tissue (1, 2). Wound response involves clotting and coagulation, tissue remodeling, cellular migration and proliferation, and angiogenesis. Many of these processes have well-recognized roles in promoting cancer. Thus, the hypothesis that tumors induce wound responses in cancer microenvironments has been the subject of recent reviews (2) and has led to suggestions that targeting the microenvironment could lead to new cancer therapies (3, 4) or chemoprevention opportunities (5). Other targets and drugs in the microenvironment have not been successful (reviewed in ref. 6], highlighting the need to better characterize of the microenvironment response to breast cancer.

The gene expression alterations present in the breast cancer microenvironment have been examined in a few previous studies. In a careful study by Allinen et al. (7), pure cell populations were isolated from two reduction mammoplasty, two ductal carcinoma in situ, and 10 invasive breast cancer patients. Serial analysis of gene expression of these purified cell populations documented widespread molecular changes in all cell types of breast cancer stroma (7). These molecular changes may reflect paracrine interactions among the diverse cell types that are present in the mammary gland during carcinogenesis: progression to carcinoma is associated with increases in myofibroblasts and immune cells, as well as increased vascularization (8,13). However, recent studies comparing stroma from reduction mammoplasty patients to stroma adjacent to breast cancers failed to identify any gene expression changes associated with disease status (14). Technical differences between these two studies may account for their distinct conclusions. Further research is needed to explain differences between these two studies. Because both studies relied on relatively small numbers of patients, little is known about interindividual variation in the microenvironment response. However, previous studies on breast cancers suggest that interindividual variation is an important feature of cancer-associated wound responses. Based on the idea that fibroblasts in vivo are exposed to serum only during wounding, Chang et al. (15) used cultured fibroblasts from many different anatomic sites to derive a wound response signature, that is highly prognostic in breast cancer patients (15). Whether that signature is altered in the microenvironment adjacent to breast cancer has not been evaluated previously.

We hypothesized that, with larger sample sizes, a reproducible gene expression signature of activated wound response would be detectable in vivo in the microenvironment of breast cancer. We used whole genome microarrays to investigate gene expression in 107 patients, including 60 reduction mammoplasty patients and 47 histologically normal samples from cancer patients (referred to as cancer-adjacent normal samples). Our results show that an in vivo wound response signature is identifiable in the histologically normal tissue adjacent to breast cancer and that this signature has strong predictive value in distinguishing normal tissue of cancer patients from that of disease-free individuals. This signature is also prognostic, confirming the importance of wound response in breast cancer progression.

Patients were women undergoing elective reduction mammoplasty at Baystate Medical Center in Springfield, MA, or mastectomy at University of North Carolina Hospitals in Chapel Hill, NC. All patients enrolled voluntarily under Institutional Review Boardapproved protocols. Reduction mammoplasty tissue was excluded if pathologic assessment of patient-matched, paraffin-embedded tissues suggested any abnormal malignant or premalignant findings. For histologically normal tissue adjacent to breast cancer (cancer-adjacent normal), a pathologist from the Tissue Procurement Facility at University of North Carolina at Chapel Hill confirmed histologically normal tissues. Only patients receiving no neoadjuvant therapy before surgery were included. All tissues were handled by snap freezing immediately after surgery, and RNA was isolated using a single protocol as described by Hu et al. (16).

All microarrays were done at the University of North Carolina at Chapel Hill. Two-color Agilent human arrays were used. Cy3-labeled reference was produced from total RNA from Stratagene Universal Human Reference (spiked with 1:1,000 with MCF-7 RNA and 1:1,000 with ME16C RNA to increase expression of breast cancer genes) following amplification with Agilent low RNA input amplification kit. The identical protocol was applied to total RNA from reduction mammoplasty or cancer-adjacent normal tissue, and all patient samples were labeled with Cy5. Some samples from reduction mammoplasty and cancer-adjacent normal groups were done on 4 44 k Agilent whole genome arrays, and some of each type were done on 244 k Agilent custom arrays that included all of the probes from the 4 44 k array. Only probes that were present on both formats were used in the final analysis and intraclass r's for technical replicates assayed on the same platform (range, 0.73-0.97; median, 0.89; n = 8) did not show meaningful differences from those conducted on different platforms (range, 0.708-0.938; median, 0.90; n = 12). Principal component analysis suggested no substantial difference between the two platforms. Duplicate microarrays corresponding to the same patient sample were combined by averaging. All data are publicly available through the Gene Expression Omnibus (GSE16113).

Data were lowess normalized, and only genes with signal intensity >10 in both channels were included in analyses. Because the reduction mammoplasty patients tended to be younger than the cancer-adjacent normal patients, we selectively adjusted for confounding by age using the following algorithm. First, we estimated the effect of tissue type (reduction mammoplasty versus cancer-adjacent normal) in a linear regression model that included tissue type as a main effect and age as a covariate (full model). We compared the main effect estimate from this full model to that of a reduced model that included only the main effect (no age parameter).We concluded that age was a confounder if inclusion of age in the model induced a change in the main effect parameter >10 and that change in effect was in the top 80th percentile. To ensure that the change-in-effect estimate was stable, we did 100 bootstrap samplings and confirmed that genes were similarly affected in the original analysis and in the bootstrap analysis. That is, in most cases, those genes that were unadjusted in the original data remained unadjusted in the bootstrap. We used the parameter estimates from the original analysis to select genes for age adjustment, and as a result, we adjusted for age in 19 of genes. This is concordant with previous literature from lymphoblastoid cell lines suggesting that at least 10 of genes were significantly associated with donor age (17). Identical normalization parameters and age-adjustment parameters and algorithms applied to the training set were used to adjust for age in the test set.

The age-adjusted training data set was analyzed using Distance Weighted Discrimination (DWD) to identify genes that distinguished reduction mammoplasty and cancer-adjacent normal samples (18). The top 200 DWD loadings were selected for the predictive signature based upon the observation that this was the smallest gene set that provided maximum predictive accuracy in the training set (by 10-fold cross-validation). Average predictive accuracy in 10-fold cross-validation was 95.4 (95 confidence interval, 95.1-96.3) in the training data set, with 91.9 sensitivity (95 confidence interval, 91.2-94.1) and 97.98 specificity (95 confidence interval, 97.92-97.92), and the same set of genes was used to assign class membership in the test set.

To evaluate the Chang et al. (15) serum response as a predictor of cancer-adjacent normal versus reduction mammoplasty status, all patients in the training set were classified according to the Pearson correlation of its expression levels of the core serum response genes to the serum activated fibroblast centroid. Patients with correlation >0.15 were classified as activated, and patients with correlation 0.15 were classified as quiescent, as described in the original paper.

To evaluate the prognostic value of our signature, we used published gene expression data for breast tumors from 295 patients, derived by researchers from Netherlands Cancer Institute and Rosetta Inpharmatics-Merck using Agilent oligonucleotide microarrays (15, 19, 20). To evaluate the prognostic value of our in vivo wound response signature, we mapped our 200 probes, representing 155 genes, to 112 genes on the Netherlands Cancer Institute data set. We then used principal component analysis to define three groups of equal size. These groups were then evaluated in univariate Kaplan-Meier analyses using WinSTAT. We compared our prognostic signature to the wound response signature published for these tumors by Chang et al. (15). For that comparison, we used the group definitions from the original paper (two groups, activated and quiescent). In addition, we compared our signature to the Stroma-Derived Prognostic Predictor of Finak et al. (21) and the 66-gene desmoids-type fibromatosis signature (22), both of which were used to define three groups of equal size by principal component analysis.

A series of cancer-adjacent normal (n = 7) and reduction mammoplasty (n = 8) samples that had been analyzed by microarray were also analyzed by immunohistochemistry. Five-micrometer sections cut from formalin-fixed, paraffin-embedded tissue blocks were deparaffinized and rehydrated following standard protocol. Sections were incubated with antisera against COX-2/PTGS2 (Cayman Chemical) and CYR61 (H-78, Santa Cruz), following manufacturers' instructions. Antigen-antibody complexes were visualized using horseradish peroxidaselabeled polymer kit following standard protocol (DAKO). Sections were counterstained in hematoxylin, dehydrated through graded alcohols, cleared in xylene, and mounted in Permount.

Using data from 48 reduction mammoplasty samples and 34 histologically normal samples adjacent to cancer, we did Distance Weighted Discrimination to select the top 200 probes that were differentially expressed between reduction mammoplasty and cancer-adjacent normal patients. In 10-fold cross-validation, these 200 probes, representing 155 genes, had 98 accuracy in predicting reduction mammoplasty versus cancer-adjacent normal status. Figure 1 shows a hierarchical cluster of these 82 patients using this gene list (complete gene list and values for all 200 probes shown in Supplementary Table S1).

The cancer-adjacent normal signature seems to be enriched for many genes previously shown to have an important role in wound healing and inflammation (Fig. 1; Supplementary Table S2). For example, extracellular matrix (ECM) deposition is an important process in repairing of wounds and ECM genes are differentially expressed between reduction mammoplasty and cancer-adjacent normal samples. Figure 1C shows a group of genes more highly expressed in reduction mammoplasty tissue than in cancer-adjacent normal tissue that includes genes in the Gene Ontology (GO) categories for ECM, such as EGFL6, BMPER, ELA2, and TFPI2. TFPI2 is the negative regulator of F3, and consistent with lower expression of its inhibitor, the transcript levels of F3 are higher in cancer-adjacent normal samples (Fig. 1D), along with a number of other genes involved in ECM formation, such as ADAMTS4, OGN, PDGFRL, FBLN1, LUM, and EFEMP1. Thus, clusters 1C and D show dysregulation of ECM in the microenvironment of cancer. Functional annotation clustering with DAVID⁸

Angiogenesis is an important process in wound healing. Consistent with a role for angiogenesis in the cancer microenvironment, Fig. 1E includes important immediate early angiogenesis genes such as PTGS2/COX-2, CYR61, and CTGF (23,26). Furthermore, well-documented binding partners for CYR61 and CTGF (including THBS1 and BMPER) are upregulated in cancer-adjacent normals. Thus, these pathways and the CCN family of genes may be important in regulating the host response to cancer. PTGS2 and CTGF are immediate early genes that are in the GO category response to wounding, which was enriched in our data (P = 1.4 10⁵), and the role of CYR61 in wound response is well established in recent literature (27). PTGS2 has been previously shown to be elevated in the microenvironment of breast cancer (28). Figure 1E therefore represents an important wound response angiogenic signature that is activated in the breast cancer microenvironment.

Genes from the GO cellular adhesion category were represented in Fig. 1F, including LAMA3 and LAMC2. The cluster in Fig. 1F also included a number of chemotaxis genes such as IL1B, CXCL1, and CXCL14, the latter of which has previously been shown to be upregulated in the histologically normal tissue of breast cancer patients (7). Figure 1G includes a wide range of genes involved in re-epithelialization after wounding, including keratins 6A, 6B, 6C and 14, 14p, and 16 (29,31). Re-epithelialization is one of the major processes necessary for successful wound repair (32). Keratins are overrepresented in our functional clusters (P = 5.2 10⁵), in a cluster that includes significant enrichment for GO categories epidermis development (P = 2.8 10⁴), ectoderm development (P = 2.8 10⁴), and cell communication (P = 5.5 10⁶). Figure 1G shows overrepresentation of cellular adhesion genes, including TP63 (33), DST, COL17A1, and DSC3. In sum, Fig. 1F and G show remodeling clusters that are dependent upon paracrine signaling and heterotypic interactions between cell types. Taken as a whole, Fig. 1 shows that the wound healing processes that have been well documented in the skin (32) are strongly induced in the microenvironment of normal breast tissue adjacent to breast cancers.

The original work hypothesizing that tumors resemble chronic wounds was referring to the tumor stroma rather than the adjacent normal tissue (1). Although we defined a wound response signature in adjacent normal tissue, we were interested to examine the relative expression of this signature in the tumors themselves. Supplementary Fig. S1 depicts the expression of the 200 probes across the 82 patients, including the paired tumor tissue for each of the cancer patients, with the groups of genes labeled according to the corresponding clusters from Fig. 1. This figure shows that, for some clusters (e.g., Fig. 1C) wherein cancer-adjacent normal has lower expression than reduction mammoplasty, the decreased expression is even more pronounced in the tumors. Likewise, many of the genes in Fig. 1F that were upregulated in cancer-adjacent normal are more strongly upregulated in tumors. However, there are groups of genes from Fig. 1D to F that are homogeneously expressed in cancer-adjacent normal tissue and heterogeneously expressed in tumors, and there are genes (e.g., Fig. 1G) that are upregulated in cancer-adjacent normal and expressed at lower levels in tumors. In the case of the cluster in Fig. 1G, many of these genes are immune response/chemotaxis genes, and therefore, downregulation in the immunosuppressive tumor microenvironment, but not the adjacent normal, may be expected. In sum, our gene list has a unique expression profile in each of the tissue types (cancer-adjacent normal versus reduction mammoplasty versus tumor), consistent with some of the known biological features of each tissue type, and our signature reliably segregates normal breast samples from patients with disease from those without disease (Fig. 1B; dendrogram in Supplementary Fig. S1).

To confirm the changes that we observed by microarray and to identify the cell types responsible for altered expression, we did immunohistochemistry for two markers from the Angiogenesis/IE cluster in Fig. 1E. Representative images are shown in Supplementary Fig. S2. COX-2 was more highly expressed in the luminal epithelium of cancer-adjacent normal patients than reduction mammoplasty patients (two-tailed t test with unequal variance; P = 0.0005), consistent with previous observations (28). Less than half of the reduction mammoplasty samples (5 of 13) were positive for COX-2 staining, with an average score of 0.7 (range, 1-2). However, 16 (89) of 18 of cancer-adjacent normal samples were positive for COX-2, with an average score of 2.1 (range, 1-3). CYR61 was moderately or strongly expressed in the luminal epithelium of cancer-adjacent normal and reduction mammoplasty patients, and we also observed CYR61 expression in fibroblasts, endothelial cells, plasma cells, and lymphocytes. Thus, the 2- to 4-fold change in CYR61 RNA levels in cancer-adjacent normal relative to reduction mammoplasty may reflect changes in expression levels of specific cellular populations or may reflect changes in the distribution of cell types present in the tissue.

Our in vivo wound response signature distinguished cancer-adjacent normal samples from reduction mammoplasty samples with 98 accuracy in the training set, 94 sensitivity, and 100 specificity. This accuracy and the striking expression differences between cancer-adjacent normal and reduction mammoplasty shown in Fig. 1 suggest that this signature can be used to predict the disease status of histologically normal samples in an independent test set. To our knowledge, there are no comparable, publicly available data set of grossly dissected human breast tissue, and therefore, we collected a separate test set of 25 additional arrays (12 reduction mammoplasty and 13 cancer-adjacent normal). In this independent test set, we correctly predicted cancer-adjacent normal versus reduction mammoplasty status with 96 accuracy (92 sensitivity and 100 specificity), with only 1 of 13 cancer-adjacent normal samples incorrectly classified as a reduction mammoplasty.

The population undergoing reduction mammoplasty surgery may be systematically different from general population. Whereas some women undergo unilateral reduction mammoplasty for cosmetic reasons, others undergo bilateral reduction mammoplasty to address ergonomic problems. This latter group may have more fatty tissue and larger breasts, and therefore, it is possible that some of these patients have biology that differs from normal breast biology in the general population. To evaluate generalizability of our study, we obtained a second test set comprised of 10 additional samples of normal breast tissue, none of which was from reduction mammoplasty patients: three normal breast samples from women with benign diagnoses (one benign cyst and two fibroadenoma) and seven normal breast tissue samples adjacent to ductal carcinoma in situ. In this independent test set, we applied the same normalization, data preprocessing algorithms, and parameters as were applied to the training set. Of the 200 probes that comprised our classifier in the training set, 149 were present with >80 good data in this second independent test set. All three of the normal tissues adjacent to benign lesions were classified as reduction mammoplastylike, and four of seven normal breast tissue samples adjacent to ductal carcinoma in situ samples were identified as cancer-adjacent normal. These results suggest that the normal breast signature that we observed in reduction mammoplasty is not peculiar to women who elect reduction surgery, and furthermore, documents that the cancer-adjacent normallike signature is also present adjacent to ductal carcinoma in situ in some patients, even before the disease becomes invasive.

Because Chang et al. (15) had previously published a serum treatment/wound response signature, we were interested in whether their in vitro signature, which had little overlap in terms of exact gene content but substantial overlap in terms of the processes represented, would also predict cancer-adjacent normal versus reduction mammoplasty status. We classified each of our samples as activated or quiescent using their signature and classification parameters, expecting that perhaps activated would be more common in the cancer-adjacent normal samples. However, 37 percent of reduction mammoplasty samples were classified as activated and 35 of cancer-adjacent normal samples were classified as activated, showing that the in vitro serum response signature does not offer predictive value in distinguishing reduction mammoplasty and cancer-adjacent normal samples (² test; 1 df; P = 0.8). This suggests that, although many of the processes involved in wound response may be common to different experimental and observational systems, the specific signatures of wound response may be tissue and/or cell-type dependent. Our wound response signature, being breast tissue derived, seems to offer advantages over the in vitro serum-derived signatures in identifying abnormal breast tissues.

Because the aforementioned signature derived from serum-treated fibroblasts, commonly referred to as the wound response signature, was highly prognostic for breast cancer patient outcomes, we hypothesized that our in vivo, 200-probe, breast tissuederived wound response signature might also have prognostic value. Thus we evaluated the prognostic value of our signature on a test set of 295 patients from the Netherlands Cancer Institute (15, 34). These 295 patients were divided into three groups based on the rank order of the first principal component of our in vivo wound response signature across their tumor expression profiles (gene expression shown in Fig. 2), and we conducted Kaplan-Meier analysis to evaluate the association between these groups and overall and relapse-free survival (Fig. 3).

In Fig. 2, the poor prognosis group (group I) had low expression of immediate early genes CTGF and CYR61 and low expression of GREM1 and GREM2. GREM1 is an important developmental signaling molecule that is part of a fibroblast growth factor/GREM1 inhibitory loop that halts tissue outgrowth during development and regeneration (35). These tumors also have low expression of angiogenesis inhibitor THBSP1 (36). In contrast, the good prognosis group III tumors had high expression of these same genes. The group I poor prognosis tumors had high expression of a number of chemotaxis and immunoregulatory genes, including IL1B, TLR2, CCL2, and indoleamine 2,3-dioxygenase, a targetable enzyme that modulates the response of T cells to tumors (37). These tumors have high expression of keratin 6 and 14, the former of which contains an IL-1responsive DNA element (38). Keratin 6 is expressed in the skin in hyperproliferative disorders and in response to stressful stimuli such as wounding (29). Keratin 14 is a basal keratinocyte marker that is differentially expressed in wound repair (30). We observed that basal-like tumors were more likely to have the group I wound response signature, whereas luminal tumors were more likely to be group III (Fig. 2). The correlations between subtype and some of the gene expression changes were consistent with those in previous studies. For example, PTGS2 has been shown to be highly expressed in basal-like tumors (39) previously, and we also observed this here.

Based on the three groups of equal size identified in Fig. 2, we then evaluated associations with survival. The three groups had different relapse-free (Fig. 3A) and overall survival (Fig. 3B), and the predictive value of these signatures compare favorably with other prognostic signatures that have used microenvironment biology to predict outcome. For example, our signatures show prognostic value similar to that provided by the Chang et al. (15) wound response signature in predicting relapse-free (Fig. 3C) and overall (Fig. 3D) survival.

To further evaluate concordance of our in vivo wound response with previous wound response or microenvironment-related signatures, we applied methods described in Fan et al. (34). Concordance between predictors was evaluated by classifying each patient and then conducting a ² test to identify associations between classifications made by each predictor. Table 1 shows that there was high concordance between those classified as activated by Chang et al. (15) and our group I and II tumors. Likewise, our predictor showed significant concordance with the three-class Stroma-Derived Prognostic Predictor (21) and desmoids-type fibromatosis signatures (22), both of which were more likely to classify group III tumors as good prognosis and group I tumors as poor prognosis.

Although our signature represents many of the processes observed to occur in healing skin (32), we sought to strengthen the claim that our signature reflects wound repair using results from previous gene expression studies on healing skin and wound response in breast. Using the same concordance evaluation methods (34) that we applied for Stroma-Derived Prognostic Predictor and desmoid-type fibromatosis signatures, we found a significant association between our signature and a previously published gene expression signature for the early response of skin to injury (40). This signature was measured only 30 or 60 minutes after skin injury and does not reflect chronic injury, and it is also derived from the skin, not the breast. Thus, we also sought comparisons with breast tissue injury. A breast injury signature has not been reported, but mammary gland involution has established features of wound repair (41). Therefore, we examined concordance with the mammary gland involution signature and found a strong and significant correlation. The results are presented in Table 1. Taken together, these findings show reproducible concordance between our signature and wound repair and microenvironment-based signatures.

It has been argued that solid tumors are composed of two distinct but interdependent compartments: the malignant parenchyma and its supporting microenvironment (42). In this article, we have studied how the normal microenvironment responds to the presence of the tumor. Our results show that a wound response signature is activated in the histologically normal tissue of cancer patients. Wound healing includes numerous overlapping processes such as blood clotting, inflammation, ECM alterations, angiogenesis, and tissue remodeling (reviewed in refs. 2, 32). Many of these processes are expressed differentially in normal tissue adjacent to breast cancer and in tumors but not in the normal tissue of women undergoing reduction mammoplasty.

Previous microarray studies have examined molecular similarities between wound repair and cancer using in vitro models. First, Iyer et al. (43) studied the response of fibroblasts to serum based on the idea that serum has a very specific meaning in vivo: cells encounter serum only in the event of local injury. In response, tissues mount a rapid concerted multicellular response to preserve tissue integrity. Serum promotes growth and survival of normal and cancer cells in culture, and in parallel, serum causes rapid cell proliferation in tissue-level wound repair responses. Given these parallels, Iyer et al. (43) were able to document a serum response of foreskin fibroblasts that included genes involved in clotting and coagulation during remodeling, including tissue factor 3, genes promoting migration and proliferation of fibroblasts (e.g., CTGF), and genes involved in angiogenesis (e.g. PTGS2 and IL1B). Many of these same genes were modified in our in vivo wound response signature. To extend the observations of Iyer et al. (43), Chang et al. (15) hypothesized that a canonical gene expression signature might be identified for serum response of fibroblasts from different anatomic sites. Using gene expression profiles of 50 samples from ten sites, Chang et al. (15) recapitulated a wound response signature originally reported by Iyer et al. (43) and documented that this core serum response was evident in a subset of breast tumors. Breast tumors that expressed the serum response signature had a poorer prognosis than tumors with a quiescent serum response phenotype. Although the in vitro wound response signatures of Iyer et al. (43) and Chang et al. (15) did not have predictive accuracy in distinguishing reduction mammoplasty and cancer-adjacent normal tissues in our data set, we did observe that the same biological processes are activated in association with breast cancer in vivo, outside the boundaries of the tumor itself.

The association between microenvironment characteristics, survival, and breast cancer subtype has been reported in previous articles. For example, in a tissue microarray study on 479 invasive breast carcinomas using 28 different markers, basal-like breast cancers were more likely to highly express genes (such as laminin) that are involved in ECM remodeling (44). In a separate study, the ECM composition of 28 primary breast carcinomas were examined using microarrays, and a gene expression profile of 278 ECM-related genes was used to divide the tumors into four main groups that correlated with outcome and basal-like subtype (45). The current results confirm that many ECM genes are enriched in basal-like cancers and that these genes have prognostic value but suggest that the enrichment may be part of a broader chronic response of the host as it attempts to repair the tissue injury caused by cancer.

Strengths of our investigation include the large number of samples relative to previous studies on histologically normal tissue, the application of methods to control for confounding by age, and the systematic comparison with previous microenvironment-derived signatures. In addition, we documented the generalizability of our findings to women with premalignant or benign breast disease and showed the validity of our predictor in two independent test sets. Our results show that the use of whole tissue, which includes epithelial cells, as well as stromal cells, does not hamper our ability to detect important microenvironment signatures. Many of the genes we detected as differentially expressed are reported to be expressed by fibroblasts (e.g., CTGF) and neutrophils (e.g. IL1B), for example, whereas others are epithelial contributors (keratins 6 and 14). This reiterates the findings of Perou et al. (46) showing that signatures from distinct cellular compartments can be identified in microarray data from mixed cell populations.

Our results suggest some important questions for future investigation related to the spatial and temporal parameters of the wound response to cancer. Although our tissue procurement protocols ensured that the histologically normal tissue was collected from regions outside of tumor margins, the distance to the margin was not measured precisely, so we are unable to document the geographic zones of gene expression alterations. Methylation changes have been reported as far as 4 cm from the site of the primary tumor (47); however, the geographic zones for gene expression alterations have not been characterized. Based on the notion that wound response is a local phenomenon, it might be expected that the wound response would not be present in contralateral breast or that it may dissipate as the distance from the primary tumor increases, but this has not been investigated. The geographic limits of the wound response to cancer may have translational implications. For example, most tumor recurrences occur at the site of previous resection even when margins are clear (48). If gene expression alterations with promoting characteristics have a defined range of action outside the primary tumor margins, future guidelines for surgical intervention may need to consider wider margins consistent with the geographic range of microenvironment alterations. It is also not known how early in development of breast cancer gene expression alterations are activated in adjacent normal tissue. Leukocyte infiltration occurs early in carcinogenesis (49), suggesting that aberrations in gene expression may already be present during ductal carcinoma in situ, but this has not been investigated. What is clear from our results is that wound response signatures are strongly activated in histologically normal tissue adjacent to invasive tumors.

A role for wound response in breast carcinogenesis is increasingly well documented. Differential expression of wound response signatures in different tumors suggests that individualized targeting of microenvironment may be a viable treatment strategy in breast cancer. For example, there is significant heterogeneity in ECM across tumors (44, 45), suggesting the differential efficacy of ECM-targeting drugs. The strong angiogenesis signature observed in group I tumors, which are enriched for basal-like breast cancers, supports the application of chemotherapy combinations that target VEGF and angiogenesis in triple-negative breast cancers (50). By better characterizing microenvironment heterogeneity, improved strategies for application of novel and existing therapies may be possible. The similarities between cancer and wound response are compelling, and continued investigation of host responses to cancer may lead to new biological insights and translational opportunities.

No potential conflicts of interest were disclosed.

We thank Dr. Melissa Johnson; Dr. Kristin Stueber; Christina Alves Baghat, PA; pathologist Dr. Giovanna Crisi at Baystate Medical Center; Dr. Lisa Carey; Emily Burrows; Amy Drobish; pathologists Chad Livasy and Leigh Thorne; and Dr. Keith Amos at the University of North Carolina Hospitals.