Unraveling the role of stromal disruption in aggressive breast cancer etiology and outcomes

Abstract

Background: Aggressive (typically high-grade) breast cancers (BCs) remain major contributors to BC-related mortality globally. The tissue changes underpinning their etiology and outcomes, however, remain poorly characterized.

Methods: Spatially resolved machine-learning algorithms were used to characterize "stromal disruption" as a morphological metric of reduced/altered extracellular matrix and increased immune, inflammatory, and/or wound response-related processes in normal, benign breast disease (BBD), and invasive hematoxylin and eosin (H&E)-stained breast tissues. Associations of stromal disruption with BC etiologic factors were assessed among 4023 healthy breast tissue donors, its impact on BC incidence was assessed among 974 BBD patients in a nested case-control study, while its prognostic associations were assessed in 4 BC patient cohorts (n = 4223).

Results: Epidemiologic risk factors for aggressive BC, including younger age, multiparity, Black race, obesity, and family history, demonstrated strong associations with increasing stromal disruption in H&E sections prior to tumor development. Substantial stromal disruption in BBD H&E was associated with ∼4-fold increased risk of aggressive (high-grade) BC and ∼3 years shorter latency from BBD to BC diagnosis, independently of BBD histology. Across BC cohorts, stromal disruption in H&E was associated with aggressive (mostly high-grade) tumor phenotypes and with markedly poor prognosis among ER-positive patients, irrespective of histology. The immunobiology of stromal disruption reflected heightened innate (CD68+), adaptive (CD3+CD4+, CD3+CD8+), immunoregulatory (CD3+CD4+FOXP3+), immune escape (PD1+PDL1+), endothelial (CD31+), and myofibroblast (α-SMA+) marker expression.

Conclusion: Our findings highlight the active stromal role in aggressive BC etiology and outcomes, opening possibilities for readily identifying high-risk women across the BC continuum that may benefit from stroma-centric preventative or therapeutic strategies.

Figure 1.

Machine-learning characterization of stromal architectural and cell composition features in normal, benign, and malignant breast tissues. Machine learning-based random forest algorithms (Halo, Indica Labs, Albuquerque, NM) were developed and optimized for the image analysis of normal, benign, and malignant breast tissues. A tissue classifier algorithm (A) was optimized to spatially characterize tissues into dense, mostly interlobular, connective tissue stroma (comprising pink regions on normal and benign breast disease images and green regions on invasive breast cancer images); loose, mostly intralobular, connective tissue stroma (comprising blue regions on images); epithelial/tumor (red regions on images), and adipose tissue (yellow) regions. In all tissue types, but particularly for malignant tissues, the stromal classifier algorithm was further trained to distinguish regions of stromal remodeling and/or denaturing, including regions of granulation tissue, desmoplasia, and/or necrosis from the “healthy” dense/collagenous stroma. Next (B), a cell classification script was optimized to detect nucleated cells (including immune and nonimmune cells) within the stromal compartments of normal, benign, and malignant breast tissues (B, green dots on analyzed images represent immune and nonimmune cells detected by machine-learning analysis of the stromal compartment). Counts of the total number of stromal cells were obtained by running the cell detection script restricted to the stromal compartment aided by the spatial separation that was performed by the tissue classifier algorithm described in (A). Cell classification was accomplished using several machine-learning parameters, including color deconvolution, nuclear detection weight, nuclear contrast threshold, nuclear segmentation aggressiveness, etc. Stromal disruption (C) was defined by combining data from tissue classification (ie, stroma vs epithelium/tumor classification), stromal phenotypic classification (ie, dense/collagenous vs nondense, eg, loose, remodeled, desmoplastic, or necrotic stroma), and stromal cell detection (immune and nonimmune stromal cells) algorithms. Categories (low [<25th percentile], intermediate [25th-75th percentile], and high [>75th percentile]) of dense stromal-to-epithelial ratio and stromal cellular density were combined to define stromal disruption phenotypes as follows: (1) none/no stromal disruption (high-dense stromal-to-epithelial ratio and low stromal cellular density); (2) minimal stromal disruption (high-dense stromal-to-epithelial ratio and intermediate or high stromal cellular density; intermediate-dense stromal-to-epithelial ratio and low or intermediate stromal cellular density; low-dense stromal-to-epithelial ratio and low stromal cellular density); (3) moderate stromal disruption (intermediate-dense stromal-to-epithelial ratio and high stromal cellular density, or vice versa); and (4) substantial stromal disruption (low-dense stromal-to-epithelial ratio and high stromal cellular density). Examples of H&E-stained imaging patches and corresponding machine learning-based maps showing probabilities for stromal architectural disruption in malignant breast biopsies are shown in (D). The blue color mask demonstrates regions of low probability of stromal disruption (ie, regions of dense connective tissue stroma or nonstromal tissue), the green color mask corresponds to regions of medium probability of stromal disruption, and the red color mask corresponds to regions of high probability of stromal disruption. As shown in the images (D), tissue regions of dense collagenous stroma had a low probability of stromal disruption, whereas regions of stromal desmoplasia, necrosis, or granulation tissue had a higher probability of stromal disruption.

Figure 2.

Associations of breast cancer etiologic factors with stromal disruption in normal and benign breast tissues. An optimized machine-learning tissue classifier algorithm was used for the analysis of standard histologic images of normal tissue biopsies donated by healthy women volunteers in the USA-based Komen Tissue Bank (A). Data on epithelial (red color), stromal (green color), and adipose (yellow color) tissue areas (mm²) were extracted from the whole slide images. A cell detection script was developed to identify and count all nucleated cells (green dots, B) within the stroma. In normal tissues, progressive stromal disruption was characterized by stromal depletion (C), reduction in the amount of dense collagenous, mostly interlobular, connective tissue stroma, concomitant increases in the amounts of loose, mostly intralobular, connective tissue stroma, and associated reduction in dense-to-loose stroma ratio (D). Progressive stromal disruption was also characterized by decreasing dense stromal-to-epithelial ratio (E), and by increasing stromal cellularity (F). G) Temporal associations of stromal disruption among women with a first and second breast tissue donation that were, on average, 3.5 years apart. The top figures are representative examples of analyzed whole slide images from a woman whose first tissue donation was in 2014 and her second donation was in 2016. A determination of substantial stromal disruption was made in both instances. The associations of breast cancer risk factors with stromal disruption in normal tissues of healthy donors are shown in (H). Multivariable ordinal logistic regression models were mutually adjusted for risk factors (including age, age at menarche, parity, family history of breast cancer (FHBC), body mass index (BMI), menopausal status, oophorectomy, race) and used to estimate odds ratios and corresponding 95% confidence intervals for 1-unit increase in stromal disruption. A separate model was used to estimate the associations of childbearing-related variables including breastfeeding, number of live births and age at first live birth (AFLB) with stromal disruption among parous women. The associations between menopausal hormone therapy (MHT) use and stromal disruption, overall and by type of MHT formulation (E = estrogen only; combined E+P = combined estrogen plus progesterone), were estimated among postmenopausal women. Associations of breast cancer risk factors with stromal disruption in benign breast disease (BBD) patients in the USA-based Kaiser BBD study are shown in (I). Benign breast disease models were additionally adjusted for BBD histological classification in addition to risk factors.

Figure 3.

Stromal disruption in benign breast disease biopsies and risk of subsequent invasive breast cancer. Representative diagnostic hematoxylin and eosin (H&E)-stained images from benign breast disease (BBD) patients with similar histological diagnosis of fibroadenoma but disparate patterns of stromal disruption are shown in (A). The morphological changes reflective of progressive stromal disruption in BBD are shown in (A) (from left to right). The corresponding probability maps (B) were generated by machine-learning algorithms that mapped regions on each slide with low (blue color), medium (yellowish green color), and high (red color) probability of stromal architectural changes. Sections from patients classified as having moderate and substantial stromal disruption had substantially larger regions with high (red) probability of stromal disruption in comparison to those from patients classified as having none or minimal stromal disruption, which had substantially larger regions with low (blue) probability of stromal disruption on BBD biopsies. In general (from left to right A and B), progressive stromal disruption in BBD was characterized by stromal depletion (C); progressive effacement of the dense/collagenized connective tissue stroma, its concomitant replacement by loose (or remodeled) stroma, and reduction in dense-to-loose (or remodeled) stroma ratio (D); decreasing (dense) stroma-to-epithelium ratio (E); and increasing stromal cellularity (F). In multivariable conditional logistic regression models adjusted for BBD histological classification (ie, nonproliferative, proliferative without atypia, atypical ductal/lobular hyperplasia) in addition to established breast cancer risk factors (family history of breast cancer, age at menarche, parity and age at first live birth, body mass index, bilateral oophorectomy, menopause and menopausal hormone therapy use, mammographic breast density, lobular involution, and tissue area) progressive stromal disruption (left to right; G) was strongly associated with increasing breast cancer risk (P<.001 for trend). In analysis stratified by BBD histological classification, increasing stromal disruption was more strongly associated with increasing risk of breast cancer among women with nonproliferative BBD (H) than those with proliferative BBD (I), with/without atypia (P=0.04 for heterogeneity). In analysis stratifying the cases according to levels of histologic grade of their subsequent invasive breast cancer (J), the magnitude of the association between progressive stromal disruption and breast cancer risk increased with increasing histologic grade (P=.002 for trend). In analysis comparing the associations of atypical hyperplasia and stromal disruption with risks of estrogen receptor-positive breast cancer stratified by levels of histologic grade (K), atypical hyperplasia (vs nonproliferative BBD) was more strongly associated with the risk of low-grade tumors, while stromal disruption (substantial vs minimal) was more strongly associated with the risk of high-grade tumors. The results were suggestive of an inverse and positive trend in risk with increasing grade for atypical hyperplasia and stromal disruption, respectively. Due to the small number of atypical ductal (n=20) and lobular (n=10) hyperplasia, the analyses could not be performed separately for these entities. *Further, due to the small number of atypical hyperplasia patients who developed grade 3 breast cancer (n=1) in this dataset, atypical hyperplasia was combined with proliferative disease for grade 3 case–control comparisons. In time-to-event analysis (L), increasing stromal disruption was associated with shorter latency between BBD diagnosis and breast cancer onset (log-rank P=.0003), irrespective of BBD histological classification.

Figure 4.

Stromal disruption in relation to invasive breast cancer biology and clinical outcomes. Representative diagnostic hematoxylin and eosin (H&E)-stained images from invasive breast cancer patients with different degrees of stromal disruption are shown in (A). The morphological changes reflective of progressive stromal disruption in invasive breast tissues are shown in (A) (from left to right). The corresponding probability maps (B) were generated by machine-learning algorithms that mapped regions on each slide with low (blue), medium (yellowish green), and high (red) probability of stromal architectural changes. Histological sections from patients classified as having moderate and substantial stromal disruption had substantially larger regions with high (red) probability of stromal disruption in comparison to those from patients classified as having minimal stromal disruption, which had substantially larger regions with low (blue) probability of stromal disruption (B). In general (from left to right in A [H&E] and B [image analysis]), progressive stromal disruption in invasive breast tissues was characterized by progressive effacement of the dense/collagenized connective tissue stroma, its concomitant replacement by remodeled and/or desmoplastic stroma, and the associated reduction in dense-to-remodeled stroma ratio (C). Progressive stromal disruption in invasive disease was further characterized by increasing necrosis (D), decreasing (dense) stroma-to-tumor ratio (E), and increasing tumor-associated stromal cellular density (F). The associations between stromal disruption on H&E images and tumor molecular characteristics (G) were assessed using ordinal logistic regression modeling with stromal disruption as the outcome and individual tumor characteristics adjusted for age and tissue area as the predictors. Analyses were performed among participants from 4 independent study populations internationally, including China (n =1807), Ghana (n=790), Poland (n=810), and the United States (n=816). The prognostic value of stromal disruption was assessed in study populations and among participants with follow-up and clinical outcomes data. Associations of stromal disruption with 10-year disease-free survival, DFS, among estrogen receptor-positive (ER+) patients from China (n=596) and 10-year overall survival, OS, among ER-positive patients from Poland (n=544) and the United States (n=480) are shown in (H). Associations between stromal disruption and 10-year OS among patients with ER-negative breast cancer from Poland (n=249) and the United States (n=153) are shown in (I).

Figure 5.

Immune and nonimmune infiltration landscape of stromal disruption in premalignant and preinvasive breast tissues. The immune, endothelial, and fibroblast infiltration landscape of stromal disruption in premalignant/preinvasive breast tissues was assessed by integrating computational pathology of hematoxylin and eosin (H&E)-stained images with multiplex immunofluorescence staining of 10 markers. Stromal disruption was characterized on H&E-stained images (A) by training algorithms to spatially characterize tissue morphology (B: epithelium [red] and stroma [green]) and to identify and count nucleated cells (shown as green dots), including immune, fibroblast, and endothelial cells, etc. within the stromal compartment (C). Regions of tissue characterized by increased stromal remodeling and/or stromal cellularity corresponded to regions of moderate (yellow) to high (red) probability of stromal disruption on machine learning (D). Immune infiltration was evaluated using an 8-plex immunofluorescence panel of seven immune markers, including CD3, CD4, CD8, CD68, PD1, PD-L1, and FOXP3, and pan-cytokeratin (CK) marker (E), while endothelial cells and fibroblasts were evaluated by using CD31 and alpha smooth muscle actine (αSMA) markers developed on a 3-plex (CD31, αSMA, and CK). The CK marker facilitated spatial resolution of the tissue into tumor (CK+) and stromal (CK-) compartments (F). Artificial intelligence algorithms were used to accomplish the cell segmentation task (G) for all markers in all tissues. Batch analysis was centralized and masked to patient demographic and clinicopathological characteristics. Data from the multiplex immunofluorescence panel were compared against categories of stromal disruption (minimal, moderate, and substantial) from the H&E image as previously defined (A–D). Overall, the density of the total immune cells in the stromal compartment increased with increasing stromal disruption (H). The distributions of the individual immune, endothelial, and fibroblast markers by categories of stromal disruption are shown in (I). The P-values in figures were obtained from Kruskal–Wallis test. (J) Expression of the different immune cell subsets and their co-localization phenotypes according to the degree of stromal disruption in premalignant/preinvasive breast tissues.

Figure 6.

Immune and nonimmune infiltration landscape of stromal disruption in invasive breast cancer. The immune, endothelial, and fibroblast infiltration landscape of stromal disruption in invasive breast cancer tissues was assessed by integrating stromal disruption data from image analysis of hematoxylin and eosin (H&E)-stained whole slide images (WSIs) with multiplex immunofluorescence staining of 10 markers on tissue microarrays (TMAs) from 2 independent study populations. As shown in (A), machine-learning algorithms were used to characterize stromal disruption on H&E-stained WSIs. The immune and nonimmune infiltration landscape of stromal disruption in invasive breast tissues was evaluated by linking slide-level stromal disruption data to TMA-level multiplex immunofluorescence data for CD3, CD4, CD8, CD68, PD1, PD-L1, FOXP3, CD31, and αSMA. Pan-cytokeratin was used to facilitate spatial analysis of immune cell populations within the stromal compartment. The distributions of the densities of the individual markers according to categories of stromal disruption are shown in (B). The data are presented for participants from study populations in the United States and Poland. P-values in figures were obtained using the Kruskal–Wallis test comparing the equality of medians across the stromal disruption categories. The immune co-localization landscape of stromal disruption was assessed using co-localization phenotypes shown using heatmaps (C) for US and Polish participants. In both study populations, substantial stromal disruption was characterized by higher immune and nonimmune cell infiltration (C).

Figure 7.

Conceptual model of invasive breast cancer risk and tumor heterogeneity incorporating premalignant (benign) epithelial and stromal tissue changes. We proposed a tissue-based model of invasive breast cancer risk and tumor heterogeneity that incorporates premalignant (benign) epithelial and stromal tissue changes. According to this model, tumor initiation in premalignant tissues depends on the fidelity of the initiating carcinogenic (etiological) signals toward epithelial or stromal tissue, or both. Carcinogenic signals acting primarily on the epithelium are more likely to initiate carcinogenesis via a “primarily epithelial” pathway, while those acting primarily on the stroma are more likely to do so via a “primarily stromal” pathway. Accordingly, tumor initiation in premalignant tissues can occur through at least 2 major tissue pathways, namely: primarily epithelial (A) and primarily stromal (B). A subset of tumors might acquire properties of both the epithelial and stromal pathways early during their development, hence considered here as arising via a “mixed” epithelial and stromal pathway (C). Blue color within images represents regions of low probability of stromal disruption; green color represents regions with medium probability of stromal disruption, while the red color depicts regions of high probability of stromal disruption. It is our observation that premalignant tissue features that are consistent with the epithelial, mixed, and stromal pathways were associated with elevated risks of biologically distinct tumor phenotypes, including low-grade, moderate-grade, and high-grade invasive breast carcinomas, respectively. Conceivably, the “Primarily Epithelial” pathway is initiated by exogenous and/or endogenous signals acting directly on the epithelial cells within the terminal duct lobular units in the setting of relatively intact stromal biology. Morphologically, it is characterized by progressive epithelial changes, including proliferative changes without atypia and atypical hyperplasia. Atypical hyperplasia with minimal or no stromal disruption represents the ‘worst’ premalignant histological entity in this pathway and our data revealed this entity to be most strongly associated with elevated risk of low-grade invasive breast cancer. The “Primarily Stromal” pathway might be initiated by exogenous and/or endogenous carcinogenic signals acting on the intralobular and/or interlobular stroma to disrupt or diminish its normal tumor suppressor function, thereby compromising stromal–epithelial crosstalk and the genomic integrity of epithelial cells. Morphologically, the primarily stromal pathway is characterized by disruptive changes in stromal architecture and/or cell composition in the absence of histological evidence for epithelial hyperplasia in premalignant tissues. The stromal tissue changes that characterize this pathway include progressive stromal depletion or denaturing and increased stromal cellularity (ie, heightened innate, adaptive, immunoregulatory, and immune escape pathways), both of which are reminiscent of chronic inflammation, and wound repair response. Substantial stromal disruption with no documented proliferative epithelial changes on biopsy represents the “worst” premalignant histological entity in this pathway and our data showed this entity to be most strongly associated with elevated risk of high-grade invasive breast cancer. In the third, that is, “Mixed” epithelial and stromal pathway, tumors may arise due to perturbations in both the epithelial and stromal pathways. Perhaps more likely, however, tumors arising from the epithelial pathway might acquire histological features that are consistent with stromal disruption during their evolutionary trajectory. Accordingly, this pathway is characterized by the co-occurrence of proliferative or atypical epithelial changes and moderate to substantial stromal disruption on premalignant biopsies. Atypical hyperplasia with moderate to substantial stromal disruption represents the “worst” premalignant histological entity in this pathway and our data revealed this entity to be most strongly associated with elevated risk of intermediate-grade invasive breast cancer. These observations suggest that the acquisition of disruptive stromal changes shifts the biological behavior of tumors that were otherwise predestined to be low grade to become more aggressive tumors. Conceivably, the primarily epithelial, mixed, and primarily stromal pathways may evolve through low-grade, intermediate-grade, or high-grade ductal carcinoma in situ (DCIS) preinvasive states prior to culminating as invasive breast cancers, respectively.