Expression of carcinoma ecotypes in the tumor microenvironment predicts response to neoadjuvant therapy in early-stage breast cancer

Abstract

Immunotherapy is associated with modest pathologic complete response (pCR) rates in early-stage breast cancer, and a subset of patients still achieves a pCR after neoadjuvant chemotherapy (NAC) only. Identifying biomarkers in the complex tumor ecosystem which define the subsets of patients who achieve pCR benefit on immunotherapy versus not is of a critical need. Transcriptomic data for patients enrolled in the two neoadjuvant immunotherapy arms of ‘pembrolizumab’ (n = 69) and ‘durvalumab’ (n = 71), and the chemotherapy arm ‘control’ (n = 210) of the I-SPY2 breast cancer clinical trial were included. Using a machine learning algorithm for tumor ecosystem-based classification, we deconvoluted transcriptomic data into the established 10 multicellular organization systems known as ‘Ecotypes’. We found that the most pro-inflammatory carcinoma ecotype (CE)9 predicts pCR in the ‘pembrolizumab’ arm (OR = 2.07; 95% CI 1.35–3.8; adjusted P value = 0.01), in the ‘control’ arm (OR = 1.89; 95% CI 1.36–2.63; adjusted P value = 0.002), and in the ‘durvalumab’ arm (OR = 1.66; 95%CI 1.18–2.34; adjusted P value = 0.03). In contrast, the basal-enriched ecotype CE2 was the most significant predictor of pCR in the ‘durvalumab’ arm, which included the addition of olaparib (OR = 3.22; 95%CI 2.25–4.60; adjusted P value < 0.0001), but not in NAC (OR = 1.18; 95%CI 0.81–1.72; adjusted P value = 0.94). Our findings suggest that CE9 could identify early-stage breast cancer patients who achieve a pCR after neoadjuvant therapy and may have a good prognosis. Whether CE9 patients could still be considered for immunotherapy or be candidates for de-escalation strategies in the neoadjuvant setting requires further investigation in future studies with link to survival outcomes. In contrast, CE2 tumors would benefit from the combination of immunotherapy with olaparib. Integrating tumor ecosystem-based patient classification could guide effective clinical management in early-stage breast cancer.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-28810-z.

Introduction

Immunotherapy is revolutionizing the neoadjuvant treatment paradigm of early-stage breast cancer^1,2. In 2021, the PD-1 inhibitor pembrolizumab in combination with chemotherapy was approved for all high-risk early stage triple negative breast cancer (TNBC) patients based on results from the Keynote-522 phase III clinical trial showing improved pathologic complete response (pCR) and event free survival (EFS) rates when compared to chemotherapy alone^3,4. Most recently, emerging clinical trials including the CheckMate-7FL⁵ and the KEYNOTE-756⁶ along with earlier results from the I-SPY2 trial arms integrating pembrolizumab and durvalumab with neoadjuvant chemotherapy^7,8 have further supported the activity of immunotherapy in improving pCR among high-risk early-stage estrogen receptor positive (ER+) breast cancers. However, these trials are bringing into question several practical considerations in the clinical community regarding the integration of immunotherapy with chemotherapy¹ or with other approved adjuvant strategies such as CDK4/6 inhibitors in ER + breast cancer^2,9. Indeed, in the final analysis of the Keynote-522 TNBC clinical trial, pCR benefit was modest demonstrating 63% versus 56% in the pembrolizumab versus control arm respectively^3,4. Furthermore, a subset of patients still achieved a pCR on chemotherapy only and immunotherapy is associated with immune-related toxicity^1,10. As EFS data is currently still maturing for most neoadjuvant breast cancer immunotherapy trials^1,11,12, it is of a critical need to identify biomarkers for the subsets of early-stage breast cancer patients who achieve a pCR benefit from immunotherapy versus not to identify responders, to switch to alternative treatment strategies among non-responders, or even to consider de-escalation of immune checkpoint blockade among patients who still benefit from chemotherapy only.

While to date several biomarkers, such as PD-L1 immunohistochemical (IHC) assessment and tumor mutational burden (TMB), have been utilized in clinical practice to guide the selection of breast cancer patients to immunotherapy in early-stage and advanced settings, pCR rates and survival benefits are still not fully explained based on the evaluation of these biomarkers^2,4,13–15. Indeed, recent attempts to better match breast cancer patients to immunotherapy have demonstrated the superior clinical benefit of incorporating omics-based subtype evaluation over PD-L1 assessment in correlative analyses of two major breast cancer immunotherapy trials^16,17. However, these biological subtypes as currently derived from bulk omics analyses do not account for the complexity of the tumor microenvironment (TME) which has been shown to influence a multitude of biological pathways that play a fundamental role in modifying therapeutic response¹⁸.

Several spatial profiling studies have been conducted over the past five years in breast cancer demonstrating that the expression of biomarkers in different compartments within the TME could have distinct predictive value and clinical implications^19–28. However, most of these studies have only used limited panels of analytes to decipher the spatiotemporal heterogeneity of breast cancer at the cellular level in a way that relates to tumor ecosystems. As such, a detailed transcriptional taxonomy that integrates comprehensive omics-based analysis to interrogate cell types, their states and collectively their multicellular organization in the architecture of the TME would allow building a better classification that guides effective clinical management.

Recently, spatially-resolved atlases in several cancers including breast have become available defining different multicellular organization systems known as ecotypes^29,30 which appear to be conserved across different epithelial tumor types and have unique clinical associations²⁹. While these ecotypes were originally deconvoluted using single cell RNA-sequencing (scRNA-Seq) data, the advent of artificial intelligence has recently facilitated the development of a machine learning framework known as EcoTyper to identify 10 clinically-distinct ecotypes and their cell states from bulk gene expression data²⁹. This framework has revealed two spatially-defined pro-inflammatory ecotypes known as carcinoma ecotype (CE)9 and CE10 to display the best survival outcomes, while CE1 to display the worst survival²⁹. As currently limited data exist on the association of ecotypes with clinical outcomes in a formal analysis of clinical trials along with recent calls in breast cancer highlighting the importance of assessing their predictive capacity in relation to different therapies particularly immunotherapy³¹, we conducted this hypothesis-testing translational study. Using available RNA and clinical outcome data for patients enrolled in the two neoadjuvant immunotherapy arms (pembrolizumab and durvalumab) and the control arm (chemotherapy) of the I-SPY2 trial^7,8, we tested the association of ecotypes and their cell states with pCR versus non-pCR. Per the Reporting Recommendations for Tumor Marker Prognostic Studies (REMARK) criteria in clinical trials³², we hypothesized that (1) breast tumors with an increased abundance for the pro-inflammatory ecotypes CE9 and CE10 would be associated with pCR on immunotherapy, while the fibroblast-enriched ecotype CE1 will be associated with non-pCR on immunotherapy; (2) The abundance of the most pro-inflammatory ecotype CE9 will be the most significant predictor of pCR in both ‘immunotherapy’ and ‘control’ arms.

Methods

Study populations

I-SPY2 clinical trial neoadjuvant pembrolizumab and control arms

Data from patients enrolled in the I-SPY2 phase II randomized clinical trial (NCT01042379) arms of ‘pembrolizumab’ (n = 69) and ‘control’ (n = 210) were obtained from Wolf et al.⁷ for the current translational study. The original I-SPY2 trial randomized stage II/III breast cancer to multiple investigational arms, which consisted of a standard neoadjuvant chemotherapy (NAC) as the ‘control’ arm plus an investigational agent as previously described⁷. Patients in the ‘control’ arm received standard NAC consisting of intravenous paclitaxel 80 mg/m² weekly for 12 weeks, followed by intravenous doxorubicin 60 mg/m² for four 21-day cycles with a combination of intravenous cyclophosphamide 600 mg/m² every 2–3 weeks. Eligible patients randomly assigned to the ‘pembrolizumab’ arm had to have a histologically-confirmed human epidermal growth factor receptor 2 (Her2) negative status and received the same standard NAC described above plus intravenous pembrolizumab 200 mg every 3 weeks for four 21-day cycles concurrently with paclitaxel. Further details regarding treatment protocols and design of I-SPY2 have been published⁷.

I-SPY2 clinical trial neoadjuvant durvalumab arm

Data from patients enrolled in the I-SPY2 phase II clinical trial (NCT01042379) arm of ‘durvalumab’ (n = 71) were obtained from Pusztai et al.⁸ for the current translational study. Only stage II/III histologically-confirmed Her2-negative patients were eligible for randomization and received durvalumab in combination with the poly-ADP ribose polymerase (PARP) inhibitor of olaparib and the standard of care NAC as described above. Participants in the olaparib/durvalumab arm received 3 cycles of intravenous durvalumab 1500 mg every 4 weeks and 100 mg oral olaparib twice a day from weeks 1 through 11, concurrent with paclitaxel. Further details regarding treatment protocols and design of I-SPY2 have been published⁸.

Study design and prespecified hypotheses

The current study was designed as prospective-retrospective per the REMARK criteria that set guidelines for reporting recommendations for tumor marker prognostic studies³². Hypotheses were tested based on a prespecified analysis plan before obtaining the RNA data and execution of any clinical analysis. The primary hypotheses were that (i) breast tumors with increased abundance for the pro-inflammatory ecotypes CE9 and CE10 would predict pCR on immunotherapy, while the fibroblast-enriched ecotype CE1 will be associated with non-pCR on immunotherapy; (ii) The abundance of the most pro-inflammatory ecotype CE9 will be the most significant predictor of pCR in both immunotherapy arms (‘pembrolizumab’ and ‘durvalumab’) and the ‘control’ arm since it has previously shown an association with good survival outcomes across different treatments²⁹. Exploratory analysis in the current study investigated the predictive capacity of other ecotypes and cell states for pCR versus non-pCR in the ‘pembrolizumab’ and ‘control’ arms. Significant observations from the exploratory analysis on the ‘pembrolizumab’ arm were considered as discovery and further tested for validation using the ‘durvalumab’ cohort to define predictive biomarkers of pCR versus non-pCR.

Investigation of ecotypes and cell states using the machine learning ecotyper algorithm

Bulk RNA data, quantified through probe gene arrays, were obtained from Wolf et al.⁷ and Pusztai et al.⁸, deconvoluted and analyzed using the machine learning algorithm of EcoTyper as described by Luca et al.²⁹. Normalized data on the gene level was used for the gene expression analysis as originally described in the supplementary files of Wolf et al.⁷ and Pusztai et al.⁸ Probeset level data per array were mean-collapsed to the gene level and this data was used in the analysis. Firstly, the cell type specific transcriptional programs of “cell states” were quantified for their relative abundance for the following cell subsets including B cells, CD4 T cells, CD8 T cells, dendritic cells, endothelial cells, epithelial cells, fibroblast cells, mast cells, macrophages, natural killer (NK) cells, neutrophils, and plasma cells as previously described³¹. Next, the co-association patterns between cell states that form multicellular communities were assessed and their abundance calculated to define the different 10 known carcinoma ecotypes CE1-10. The abundance of each of the 10 known ecotypes was quantified in each I-SPY2 trial sample included in the current translational study. A full description of the methods used for RNA profiling in the original I-SPY2 trial can be found in Wolf et al.⁷ and Pusztai et al.⁸

Statistical analysis

The prespecified hypotheses and statistical analysis plan were finalized prior to testing any correlation between molecular and clinical data. The prespecified statistical hypotheses tested whether an increased continuous expression of ecotypes CE9 and CE10 will be significantly associated with pCR on immunotherapy, while the increased continuous expression of ecotype CE1 will be associated with non-pCR. In addition, an increased continuous expression of ecotype CE9 will be the most significant predictor of pCR in both ‘immunotherapy’ arms and the ‘control’ arm. Estimates of association between ecotype abundance and pCR were calculated using odds ratio (OR), 95% confidence interval (CI), and log-rank P values per each treatment arm assessed (i.e. ‘pembrolizumab’, ‘control’, and ‘durvalumab’) derived from a univariate logistic regression model. Results for the prespecified hypotheses tested were further adjusted for multiplicity using the Benjamini–Hochberg (BH) method. Results were depicted using forest plots. Multivariate analysis was adjusted for IHC subtype (TNBC vs. HR + /Her2−), pathological tumor size (T1 or T2) versus (T3 or T4), and nodal status negative versus positive. Exploratory analyses evaluating the association of cell states continuous expression and pCR were also derived from a univariate logistic regression model with estimation of OR and 95% CI for the association of each cell state feature with pCR. Predictor values for cell states were further multiplied by 10 so that more interpretable ORs could be derived and visualized on forest plots as previously recommended by Wang and Ali et al.²⁵. Exploratory analyses performed were considered hypothesis-generating and thus were not adjusted for multiplicity testing. In contrast, results of ecotypes and cell states found to be significant on the ‘pembrolizumab’ cohort and then validated using the ‘durvalumab’ cohort were adjusted for multiplicity using the BH method. Significant findings on both immunotherapy arms were displayed using a volcano plot showing differences between features associated with pCR versus non-pCR/residual disease based on the log2 OR transformed data. All tests were 2-sided, at a significance level of 0.05, and performed using the R statistical software.

Data availability

This study involved the collection and analysis of data from publicly-available datasets. Fully normalized and batch corrected RNA data on the gene level for cases in the ‘pembrolizumab’ and ‘control’ arms of the I-SPY2 were obtained from the "ISPY2ResID_AgilentGeneExp_990_FrshFrzn_meanCol_geneLevel_n988.txt" Supplementary files of Wolf et al.⁷ and downloaded from Gene Expression Omnibus (GEO) SubSeries GSE194040 (mRNA) as part of the SuperSeries GSE196096—unique identifier: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE194040. Clinical data, hormonal status, and IHC subtype information were retrieved from the Supplementary Table S2 of Wolf et al.⁷ and the SuperSeries GSE196096—unique identifier: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE196096). Normalized RNA data and clinical information for the ‘durvalumab’ arm cases were available from Pusztai et al.⁸ using “GSE173839_ISPY2_AgilntGeneExp_durvaPlusCtr_FFPE_MEANcOL_GENElevel_n105.txt” and downloaded from GEO under the accession number of GSE173839—unique identifier: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE173839. No custom code was generated to collect data in this study. Any queries for R codes used in the analysis of this study can be directed to the corresponding author. Information regarding trial registration numbers, CONSORT and trial protocols are found in the original publications of the I-SPY2 trial^7,8.

Results

Characteristics of the translational study cohorts

Deconvoluted RNA data and clinical response information were available from the 350 total patients enrolled in the I-SPY2 trial who all received their assigned treatments of ‘pembrolizumab’(n = 69), ‘control’ (n = 210), and durvalumab (n = 71). Details regarding the study workflow are displayed in Fig. 1. Among the total 350 cases analyzed, 98 (28%) had pCR (31 on ‘pembrolizumab’, 38 on ‘control’, and 29 on ‘durvalumab’ arms).

Fig. 1.

Schematic representation of the translational study workflow and design. The analysis of the translational I-SPY2 clinical trial cohorts followed a prespecified plan per the REMARK criteria for predictive biomarker evaluation. Figure was partially created by biorender.

To test whether the distribution of the different 10 carcinoma ecotypes would differ according to breast cancer subtypes, 280 cases (80%) had available information for both carcinoma ecotype assignment and IHC subtype. The distributions of IHC subtypes and carcinoma ecotypes for these cases within each of the translational I-SPY2 clinical trial arms are displayed in Table 1. The hormone receptor (HR)+ /Her2− subtype spanned the 10 different known ecotypes, comprising 33 CE8 cases (22%), 33 CE1 cases (22%), and others. In contrast, TNBC consisted mostly of CE2 (30%) which is known to be a basal-enriched ecotype³¹. The pro-inflammatory ecotypes of CE9 and CE10 comprised 16% and 12% of the TNBC cases respectively (Table 2). Of note, CE2 which is known to be a basal-enriched ecotype consisted primarily of TNBC (63%), while 35% of the CE2 cases were HR +/Her2− (Table 2).

Table 1.

Distribution of IHC subtypes and carcinoma ecotypes within the translational I-SPY2 clinical trial cohorts of ‘pembrolizumab’, ‘control’ and ‘durvalumab’ arms.

	Pembrolizumab cohort (n = 53)	Control cohort (n = 169)	Durvalumab cohort (n = 58)
IHC subtype
HR+/Her2−	30 (56.6%)	77 (45.56%)	40 (68.97%)
HR +/Her2+	0 (0%)	15 (8.88%)	0 (0%)
TNBC	23 (43.4%)	71 (42.01%)	18 (31.03%)
HR-/Her2+	0 (0%)	6 (3.55%)	0 (0%)
Carcinoma ecotype
CE1	13 (24.53%)	28 (16.57%)	15 (25.86%)
CE2	11 (20.75%)	31 (18.34%)	12 (20.69%)
CE3	1 (1.89%)	2 (1.18%)	0 (0%)
CE4	0 (0%)	0 (0%)	0 (0%)
CE5	5 (9.43%)	13 (7.70%)	4 (6.90%)
CE6	2 (3.77%)	23 (13.61%)	5 (8.63%)
CE7	1 (1.89%)	2 (1.18%)	1 (1.72%)
CE8	7 (13.21%)	32 (18.94%)	10 (17.24%)
CE9	7 (13.21%)	20 (11.83%)	6 (10.34%)
CE10	6 (11.32%)	18 (10.65%)	5 (8.62%)

IHC immunohistochemistry, Her2 human epidermal growth factor receptor 2, HR hormonal receptor, TNBC triple negative breast cancer.

Table 2.

The distribution of the different 10 carcinoma ecotypes across the immunohistochemical subtypes.

Carcinoma ecotype	IHC subtype	Frequency
CE1	HR−/Her2+	1
	HR+/Her2−	33
	HR+ /Her2+	8
	TNBC	14
CE2	HR−/Her2+	1
	HR+/Her2−	19
	HR+/Her2+	0
	TNBC	34
CE3	HR−/Her2+	0
	HR+/Her2−	2
	HR+/Her2+	0
	TNBC	1
CE5	HR−/Her2+	1
	HR+/Her2−	15
	HR+/Her2+	0
	TNBC	6
CE6	HR−/Her2+	0
	HR+/Her2−	16
	HR+/Her2+	1
	TNBC	13
CE7	HR−/Her2+	0
	HR+/Her2−	4
	HR+/Her2+	0
	TNBC	0
CE8	HR−/Her2+	0
	HR+/Her2−	33
	HR+/Her2+	3
	TNBC	13
CE9	HR−/Her2+	2
	HR+/Her2−	13
	HR+/Her2+	0
	TNBC	18
CE10	HR−/Her2+	1
	HR+/Her2−	12
	HR +/Her2+	3
	TNBC	13

IHC immunohistochemistry, Her2 human epidermal growth factor receptor 2, HR hormonal receptor, TNBC triple negative breast cancer.

The most pro-inflammatory carcinoma ecotype CE9 predicts pCR after neoadjuvant therapy in early-stage breast cancer: hypothesis-testing

In a prespecified analysis of the pro-inflammatory carcinoma ecotypes CE9 and CE10, the continuous expression of CE9 and CE10 was significantly associated with pCR on the pembrolizumab arm (OR for CE9 on pembrolizumab = 2.07; 95%CI 1.35–3.8; P = 0.001) and (OR for CE10 on pembrolizumab = 1.58; 95%CI 1.03–2.42; P = 0.04) (2. 3). In contrast, the abundance of CE1 was found to be significantly associated with non-pCR on the pembrolizumab arm (OR = 0.62; 95%CI 0.38–1.01; P = 0.05) (Fig. 2). Importantly, CE9 was the most significant predictor for pCR when adjusted for multiplicity testing in the ‘pembrolizumab’ arm (BH adjusted P value = 0.01) (Fig. 2), in the ‘control’ arm (OR = 1.89; 95%CI 1.36–2.63; P = 0.0002, BH adjusted P value = 0.002) (Fig. 2) and in the ‘durvalumab’ arm (OR = 1.66; 95%CI, 1.18–2.34; P = 0.004, BH adjusted P value = 0.03) (Table 3), confirming the prespecified hypothesis that was part of the original planned analysis. The association between CE9 and pCR remained significant in a multivariate analysis adjusted for IHC subtype, tumor size and lymph node status in both the ‘pembrolizumab’ and ‘control’ cohorts (Supplementary Table 1). To test the differential benefit from therapy, we performed interaction tests between the ‘pembrolizumab’ and ‘control’ cohorts and found that results were not significant (Supplementary Table 2). When testing whether the association of ecotypes with pCR would differ based on IHC subtype, results showed that CE9 was significantly associated with higher pCR on both ‘pembrolizumab’ and ‘control’ arms within HR+ /Her2− and TNBC subgroups (Supplementary Table 3).

Fig. 2.

Forest plots for pathologic complete response (pCR) in the translational I-SPY2 clinical trial cohorts of ‘pembrolizumab’ arm (a) and ‘control’ arm (b) according to their ecotype abundance. Estimates of association between ecotype abundance and pCR were calculated using odds ratio, 95% confidence interval, and log-rank P values per each treatment arm and derived from a univariate logistic regression model. Results were adjusted for multiplicity using the Benjamini–Hochberg (BH) method.

Table 3.

Correlation of carcinoma ecotypes and cell states abundance with response to neoadjuvant therapy in the ‘durvalumab’ cohort which included the addition of olaparib in the I-SPY2 trial.

Ecotypes and cell states abundance	Odds ratio	95% confidence interval	P value	BH adjusted P value
Carcinoma ecotypes
CE1 Abundance	0.44	0.25–0.79	0.006	0.048
CE2 Abundance	3.22	2.25–4.60	9.39e-09	9.39e-08
CE6 Abundance	0.63	0.43–0.91	0.01	0.07
CE8 Abundance	0.62	0.43–0.91	0.02	0.12
CE9 Abundance	1.66	1.18–2.34	0.004	0.03
B cells states
B cells S03 (Normal-enriched)	0.67	0.47–0.96	0.03	0.15
CD4 T cells states
CD4 T cells S01 (Exhausted/effector memory/Treg)	1.33	1.08–1.63	0.007	0.04
CD4 T cells S04 (Resting, normal-enriched)	0.70	0.58–0.85	0.0005	0.003
CD8 T cells states
CD8 T cells S03 (Exhausted/effector memory)	2.03	1.48–2.80	0.000035	0.0001
Dendritic cells states
Dendritic cells S03 (Mature immunogenic)	1.46	1.18–1.81	0.0008	0.006
Dendritic cells S05 (Mature, normal-enriched)	0.79	0.62–0.98	0.036	0.18
Dendritic cells S07 (Migratory activated)	1.36	1.14–1.61	0.0009	0.006
Dendritic cells S08 (Unknown)	0.66	0.51–0.87	0.003	0.02
Endothelial cells states
Endothelial cells S01 (Normal-enriched)	0.64	0.42–0.95	0.03	0.12
Epithelial cells states
Epithelial cells S01 (Basal-like)	3.48	1.93–6.27	0.00007	0.0003
Epithelial cells S03 (Pro-angiogenic)	0.47	0.28–0.79	0.005	0.01
Epithelial cells S04 (Pro-inflammatory)	4.21	2.25–7.90	0.00002	0.0001
Epithelial cells S05 (Unknown)	0.48	0.28–0.79	0.005	0.01
Epithelial cells S06 (Metabolic)	0.32	0.18–0.54	0.00007	0.0003
Fibroblast cells states
Fibroblast cells S02 (CAF2, Normal-enriched)	0.69	0.53–0.90	0.007	0.03
Fibroblast cells S03 (CAF1, tumor-associated)	0.53	0.36–0.76	0.001	0.005
Fibroblast cells S04 (Migratory-like)	0.75	0.60–0.95	0.02	0.06
Fibroblast cells S05 (Unknown)	0.61	0.47–0.80	0.0005	0.003
Fibroblast cells S08 (Pro-migratory-like)	3.11	2.14–4.52	0.00000006	0.0000004
Mast cells states
Mast cells S03 (Unknown)	0.62	0.38–0.98	0.04	0.26
Mast cells S05 (Unknown)	0.28	0.16–0.49	0.00003	0.0002
Mast cells S06 (Activated)	1.53	1.16–2.02	0.003	0.02
Monocytes/Macrophages cell states
Macrophages S02 (Classical M0, normal-enriched)	0.71	0.53–0.96	0.02	0.12
Macrophages S03 (Classical M1)	1.55	1.08–2.23	0.02	0.12
Macrophages S05 (M2-like, normal-enriched)	0.77	0.61–0.97	0.03	0.15
Macrophages S06 (M2 foam cell-like)	0.54	0.35–0.81	0.003	0.02
Macrophages S07 (M2-like proliferative)	3.02	2.04–4.47	0.0000004	0.000004
NK cells states
NK S01 (Classical)	1.44	0.998–1.08	0.05	0.15
NK S02 (Normal-enriched)	1.29	1.01–1.64	0.05	0.15
NK S03 (Unknown)	0.63	0.41–0.97	0.04	0.15
Plasma cells states
Plasma cells S02 (Unknown)	1.36	0.999–1.85	0.05	0.30

Odds ratios > 1 indicate good response while odds ratios < 1 indicate poor response.

Results were scaled by 10 to make it more visible for odds ratios and confidence intervals. Only biomarkers with significant uncorrected P value are displayed. The full list including other carcinoma ecotypes and cell states and their correlation with response to neoadjuvant therapy including durvalumab is displayed in Supplementary Table 4.

To further characterize CE9, we tested whether this ecotype would display a higher expression of tumor infiltrating lymphocytes (TILs) and PD-L1 signatures. Following previously-published methods on I-SPY2 by Campbell et al.³³, we found that CE9 significantly displayed a higher PD-L1 signature expression when compared to other ecotypes (p value < 0.0001) (Supplementary Fig. 1). However, results were not significant when testing the correlation of CE9 with TILs signature (p value = 0.17) (Supplementary Fig. 1).

Association of other ecotypes and cell states with response to neoadjuvant therapy: hypothesis-generating

Analysis of the predictive capacity of the rest of ecotypes and cell states that constitute them was performed as exploratory to identify biomarkers that could be related to pCR versus non-pCR in early-stage breast cancer and later confirmed on validation sets, particularly immunotherapy.

Within the ‘pembrolizumab’ cohort, we identified CE2 to significantly predict pCR benefit (OR for CE2 = 1.94; 95%CI 1.13–3.32; P = 0.02) (Fig. 2). However, this result was not significant when corrected for multiplicity testing including other ecotypes. In contrast, CE8 was found to be significantly associated with residual disease and this result was significant for multiplicity testing (OR = 2.07; 95%CI 1.35–3.8; P < 0.001, BH adjusted P value = 0.009) (Fig. 2). Exploratory analysis of cell states related to ecotypes revealed that those known to colocalize in CE9³¹ including activated B cells, effector memory CD4 and CD8 T cells, mature immunogenic dendritic cells, pro-inflammatory epithelial cells, NK cells, classical M1 macrophages, and a status of plasma cells known as S02 significantly predicted pCR on pembrolizumab (Fig. 3; Supplementary Table 4A). In addition, other cell states that are characteristic of CE10 (classic naïve B cells, naïve CD4 T cells, myeloid dendritic cells, monocytes, and classical plasma cells) and of CE2 (basal-like epithelial cells) were predictive of pCR on pembrolizumab (Fig. 3; Supplementary Table 4A). In contrast, cell states related to pro-angiogenic epithelial cells (characteristic of CE1), metabolic epithelial cells (characteristic of CE8), cancer associated fibroblasts (characteristic of CE1 and CE8), non-activated mast cells (characteristic of CE1 and CE8) and M2 macrophages predicted residual disease on pembrolizumab (Fig. 3; Supplementary Table 4A).

Fig. 3.

Forest plots for cell states significantly associated with pathologic complete response (pCR) in the translational I-SPY2 clinical trial cohort of ‘pembrolizumab’ arm. Estimates of association between cell state abundance and pCR were calculated using odds ratio, 95% confidence interval, and log-rank P values and derived from a univariate logistic regression model. Predictor values for cell states were further multiplied by 10 for more interpretable ORs and visualization on forest plots. The results of the predictive capacity for other cell states are found in Supplementary Table 2A. Figure was partially created by biorender.

Within the ‘control’ cohort, CE3 was found to be significantly associated with pCR (OR = 1.31; 95%CI 1.08–1.58; P = 0.006, BH adjusted P value = 0.048) (Fig. 4). In contrast, CE5, CE6, and CE8 were associated significantly with residual disease (Fig. 4). When testing the predictive capacity of cell states related to ecotypes, we found that those colocalizing in CE9 including activated B cells, effector memory CD4 and CD8 T cells, mature immunogenic dendritic cells, pro-inflammatory epithelial cells, NK cells, and classical M1 macrophages significantly predicted pCR benefit on the ‘control’ arm (Fig. 4; Supplementary Table 4A). Other cell states that are characteristic of CE10 (classic naïve B cells, myeloid dendritic cells, monocytes, and classical plasma cells) and of CE3 (late-stage differentiated effector CD8 T cells and activated mast cells) were predictive for pCR (Fig. 4; Supplementary Table 4A). However, cell states related to normal-enriched B cells (characteristic of CE5), resting/normal-enriched CD4 T cells (characteristic of CE6), normal-enriched dendritic cells (characteristic of CE6), endothelial cells (characteristic of CE1), non-activated mast cells (characteristic of CE8), and activated neutrophils (characteristic of CE1 and CE6), predicted residual disease on the ‘control’ arm (Fig. 4; Supplementary Table 4A).

Fig. 4.

Forest plots for cell states significantly associated with pathologic complete response in the translational I-SPY2 clinical trial cohort of ‘control’ arm. Estimates of association between cell state abundance and pCR were calculated using odds ratio, 95% confidence interval, and log-rank P values and derived from a univariate logistic regression model. Predictor values for cell states were further multiplied by 10 for more interpretable ORs and visualization on forest plots. The results of the predictive capacity for other cell states are found in Supplementary Table 2A. Figure was partially created by biorender.

Of note, receiver operating characteristics (ROCs) for response status over the different CE abundance scores demonstrated the high discriminatory ability of CE9 within both the ‘pembrolizumab’ and ‘control ‘cohorts (Supplementary Figs. 2, 3). Area under the ROC curves to discriminate responders was 0.77 on ‘pembrolizumab’ and 0.67 on the ‘control’ arm.

Validation of predictive ecotype and cell state biomarkers on neoadjuvant immunotherapy cohort

Per our prespecified analysis plan, we next aimed to validate candidate predictive ecotype and cell state biomarkers for pCR on immunotherapy from our exploratory analysis on the ‘pembrolizumab’ cohort using the I-SPY2 ‘durvalumab’ cohort.

In our analysis corrected for multiple testing for ecotype biomarkers, CE2 was found to significantly predict pCR benefit on immunotherapy (OR = 3.22; 95%CI 2.25–4.60; P < 0.00001, BH adjusted P value < 0.00001) (31; Fig. 5; Supplementary Table 4B). These results were further observed to be significant when restricting the analysis within the subgroup of HR + /Her2− and TNBC. In contrast, CE1 was associated with a residual disease (OR for CE1 = 0.44; 95%CI 0.25–0.79; P = 0.006, BH adjusted P value = 0.048) (Table 3; Fig. 5; Supplementary Table 4B). While CE8 showed significant results to predict residual disease after durvalumab, this result was not significant when adjusted for multiplicity testing (Table 3; Fig. 5; Supplementary Table 4B). ROCs for response status demonstrated the strong discriminatory ability of CE2 within the ‘durvalumab’ cohort to identify responders to neoadjuvant therapy with area under the ROC curve of 0.87 (Supplementary Fig. 4).

Fig. 5.

Volcano plot showing ecotype and cell state biomarkers significantly associated with pathologic complete response versus residual disease on immunotherapy. Results were derived from a univariate logistic regression model for significant biomarkers that were validated on the ‘durvalumab’ arm and thus were found to be significant on both the translational I-SPY2 clinical trial cohorts of ‘pembrolizumab’ and ‘durvalumab’. The x-axis is log2 odds ratio and the y-axis is − log10(P value). Negative and positive ORs indicate biomarkers associated with residual disease and pathologic complete response, respectively with the horizontal line indicating uncorrected P < 0.05. All biomarkers above this line and shown on the figure have a nominal P value < 0.05. Results were adjusted for multiplicity using the Benjamini–Hochberg (BH) method and those significant (i.e. with BH adjusted P value < 0.05) are denoted in red, while those not significant (with BH adjusted P value ≥ 0.05) are denoted in gray.

Our hypothesis-testing analysis for cell states validated the association of those characteristic of CE9 including effector memory CD4 and CD8 T cells, mature immunogenic dendritic cells, pro-inflammatory epithelial cells, and the basal-like epithelial cells characteristic of CE2 with pCR on immunotherapy (all with BH adjusted P value < 0.05) (Table 3; Fig. 5; Supplementary Table 4A). Among these cell states, the ones that were characteristic for CE9 were also observed to be associated with pCR on the ‘control’ arm (Fig. 4; Supplementary Table 4A). In contrast, the cell states of pro-angiogenic epithelial cells (characteristic of CE1), metabolic epithelial cells (characteristic of CE8), cancer associated fibroblasts (characteristic of CE1 and CE8), and non-activated mast cells (characteristic of CE1 and CE8) were confirmed to be associated with a residual disease on immunotherapy (all with BH adjusted P value < 0.05) (Table 3; Fig. 5; Supplementary Table 4A). Notably, among the cancer associated fibroblasts, fibroblast cells S03 (CAF1, tumor-associated) and fibroblast cells S05 (unknown) were the ones significantly predictive of non-pCR in both the ‘pembrolizumab’ and ‘durvalumab’ cohorts.

Discussion

The current study presents the results of a hypothesis-testing correlative analysis using data from patients enrolled in the I-SPY2 early-stage breast cancer clinical trial to assess the association of ecotypes and their cell states with pCR versus non-pCR after neoadjuvant therapy.

We demonstrate that breast tumors with increased abundance for CE9 identifies patients with early-stage breast cancer who likely to achieve a pCR after neoadjuvant therapy. In contrast, we show that CE2 tumors would identify patients that benefit the most from the combination of immunotherapy with olaparib.

The observation that CE9 is the most significant predictor in both immunotherapy and chemotherapy arms when adjusted for multiplicity testing is consistent with original findings in the EcoTyper framework reporting CE9 to outperform other ecotypes for predicting superior outcomes across different therapies²⁹. While the predictive capacity of CE9 for immunotherapy response was originally demonstrated using data from advanced melanoma and urothelial carcinoma patients treated with immune checkpoint inhibitors, its predictive capacity for early-stage breast cancer setting has not been previously investigated. Thus, our findings using early-stage breast cancer cohorts treated with contemporary guidelines including immunotherapy^1,2 could extend the clinical utility of CE9 to predict clinical response beyond conventional neoadjuvant chemotherapy protocols and can be even applied to the advanced breast cancer setting.

While the pro-inflammatory CE10 was found to significantly predict pCR on pembrolizumab, results were not significant when corrected for multiple testing and findings were not further confirmed on the durvalumab cohort as compared to CE9. These findings might be explained by the spatial distinction between CE9 and CE10 in breast cancer where CE9 is characterized with IFN-gamma signaling, T cells expressing activation and exhaustion genes that are more localized to the tumor core²⁹. In contrast, CE10 has features of T cells expressing naïve and central memory genes, localized more at the edge of tumor, and could be even expressed in adjacent normal tissue²⁹. The spatial organization of immune biomarkers to the tumor core in CE9 could suggest stronger chronic antigen stimulation and a link to an activated effector and exhausted T cell states³⁴ when compared to CE10 that could be related to a transitional effector memory T cell in the periphery²⁹. Indeed, an analysis of cell type states revealed that the co-occurrence of effector memory and exhausted CD4 and CD8 T cells significantly predicted pCR on both ‘pembrolizumab’ and ‘durvalumab’ cohorts and even chemotherapy only in the ‘control’ arm. Furthermore, other cell states that are characteristic of CE9- including mature immunogenic dendritic cells, and pro-inflammatory epithelial cells- were found to be independently predictive for pCR on ‘immunotherapy’ and ‘control’ arms suggesting that the simultaneous activity of these different cell states within CE9 elicits immune response signalling pathways in the TME that drive therapeutic response.

The finding that basal-like epithelial cells and their ecotype of CE2 were predictive of pCR on durvalumab could be explained by the fact that the ‘durvalumab’ arm contained a treatment with the PARP inhibitor olaparib for which basal tumors are also known to display sensitivity³⁵ and immune modulation induced by PARP inhibitors can enhance the activity of immune checkpoint blockade³⁶. Impaired nucleotide and base excision repair caused by PARP inhibition can further increase mutation and neoantigen loads leading to an increased sensitivity to immune checkpoint blockade³⁷. In addition, DNA fragments generated by impaired DNA repair can activate the interferon signaling pathway in the TME³⁸. Thus, it could be possible to propose that basal-like epithelial cells characteristic of CE2 along with other multiple immune cells that co-occur in this ecotype collectively define the multicellular organization system that benefits from a combination of immunotherapy with PARP inhibitors.

Our results demonstrating that cell states co-occurring in the fibroblast-enriched ecotype of CE1 including pro-angiogenic epithelial cells, CAFs, and non-activated mast cells predict residual disease on immunotherapy support their known participation in pathways that confer immune resistance^39–41. Several CAFs states revealed by scRNA-seq have been proposed to predict immunotherapy resistance in breast cancer as they possess functional properties that regulate extracellular matrix promoting an immunosuppressive TME^20,42. Of note, the fibroblast S03 cell state found to be predictive of residual disease on immunotherapy is characterized by several extracellular matrix components associated with collagen organization and angiogenesis processes²⁹, reflecting an immune-evasive, highly proliferative, and invasive TME that blocks immune cells from infiltrating the tumor and generating an effective anti-tumor response to immunotherapy⁴³. Furthermore, the fibroblast S05 cell state is characterized by several members that are associated with epithelial mesenchymal transition and an impaired anti-tumor immunity such as PPP1R32 (SOX4)⁴⁴. In addition, the findings on non-activated mast cells could suggest that these are ‘resting mast cells’ which have been linked to non-response to immunotherapy⁴¹. Mast cells have also been shown to play a role in inducing the activity of pro-angiogenic epithelial cells, essential for neovascularization during tumor progression, through mechanisms including the secretion of proangiogenic factors (e.g., VEGF, bFGF, TGFβ, TNF-α, IL8), and the release of proteases (e.g., tryptase and chymase) and heparin-binding growth factors^45–47. Altogether, these findings underscore the value of stromal-immune interactions for the design of better strategies that could overcome immunotherapy resistance in breast cancer.

The capacity of CE8 to predict residual disease after neoadjuvant therapy in both immunotherapy and ‘control’ arms could be partially explained by the existence of cell states such as CAFs, and non-activated mast cells that are also characteristic of CE1. However, CE8 was unique by the occurrence of metabolic epithelial cells which could reprogram the metabolic profile of TME inducing several pathways related to immune evasion and tumor growth such as expansion of myeloid-derived suppressor cells and inhibition of cytotoxic T cells and NK cells^48–50. Furthermore, CE8 had characteristics related to oncogenic signalling processes of PI3K/AKT and MYC that are known to foster the creation of an immunosuppressive stroma that induces T cell exclusion and dampens the capacity of antigen presentation^50,51. Altogether, these tumors may benefit from treatments targeting the immune stromal barrier along with immune augmentation strategies⁵².

Strengths of our study include (1) adherence to REMARK guidelines³² following a prespecified hypotheses and analyses. In addition, hypothesis-generating results were validated using a validation set containing immunotherapy. (2) The use of an in-depth analysis of the whole transcriptome with a recent machine learning tool to profile the spatial complex architecture of TME in a biomarker-agnostic manner. While most recent studies providing insights into the cellular communities of breast cancer have focused on a limited number of analytes^{21,22,25–27}, our comprehensive analysis revealing distinct ecotypes and cell states distilled from thousands of genes provide a more reliable way for the development of diagnostic and therapeutic strategies in breast cancer. Indeed, with the recent change in treatment paradigm of the neoadjuvant setting of breast cancer^3,4, we provide a pioneering attempt for a comprehensive assessment of biomarkers in the TME to predict immunotherapy response. (3) The I-SPY2 clinical trial represents the most comprehensive dataset resource reported to date with access to enough-powered number of cases that can be used for predictive biomarkers of neoadjuvant immunotherapy response in breast cancer⁷. The other attempt to spatially profile the TME in this setting of neoadjuvant immunotherapy using breast cancer clinical trial materials has assessed the in-situ expression of 43 biomarkers using imaging mass cytometry²⁵ on samples from the NeoTRIPaPDL1 trial¹². Interestingly, this study found that GZMB T cells which are also known to be the most characteristic for CE9¹² were more abundant in responsive tumors, supporting our findings on the predictive capacity of CE9. As immunotherapy agents are currently being used in the early-stage and metastatic settings of breast cancer^1,2, our findings could be further extended to inform prospective study designs for early-stage and advanced breast cancer patients, specifically classified based on their ecotypes.

Our study has some limitations. First, it included only pCR data and did not include information regarding survival endpoints as this data has yet to be matured for the I-SPY2 trial. Of note, the original EFS findings³ and the very recent positive results supporting the benefit of neoadjuvant immunotherapy on overall survival in the KEYNOTE-522 clinical trial⁵³ demonstrated a significant improvement even in the subpopulation of TNBC patients who did not achieve a pCR on pembrolizumab. Thus, whether the subset of CE9 patients could still be considered for immunotherapy in the neoadjuvant setting or be candidates for de-escalation of immunotherapy requires further investigation in future studies with enough power for correlation with survival outcomes. Furthermore, as our findings have demonstrated the association of CE9 with pCR in both ‘pembrolizumab’ and ‘control’ arm, CE9 could serve as a potential prognostic biomarker for neoadjuvant therapy in breast cancer and even investigated in currently-available de-escalation strategies in the clinic. Indeed, regimen-specific predictive biomarkers are still lacking in the context of neoadjuvant breast cancer therapy^1,2, and it would be valuable to apply our findings to cohorts that particularly explore regimens with a reduced intensity of chemotherapy in combination with immunotherapy such as those given in the NeoPACT trial⁵⁴. Second, none of the ecotypes we assessed selectively identified patients who would benefit from pembrolizumab with neoadjuvant chemotherapy but not from neoadjuvant chemotherapy alone. In fact, the associations between CE9 and CE10 and pCR from pembrolizumab may largely be driven by the pCR association from chemotherapy requiring further investigation in more powered studies. However, recent correlative studies of the I-SPY2 trial have also reported several immune response biomarkers to be positively associated with pCR in early-stage breast cancer in ‘pembrolizumab’, ‘durvalumab’ and ‘control’ arms^8,33. Thus, our findings on the pro-inflammatory ecotype CE9 and the immune signatures reported in these studies could be reflecting a similar “immune hot” biology that correlates with pCR. Third, while our results infer the organization of each ecotype in the TME based on bulk transcriptomic data, they do not provide a direct information regarding the spatial transcriptomic context that defines in situ morphologic characteristics and phenotypes. Thus, integrating our results with spatial multiomics methods which quantify biomarkers in tumor, stromal and immune cells compartments could leverage the quality and clinical utility of our findings to better define the relevant ecotypes that delineate therapeutic response. Of note, recent correlative analyses of the I-SPY2 trial described specific biomarkers in the spatial organization of the TME with gene expression signatures to selectively predict pCR on neoadjuvant immunotherapy and not chemotherapy in the I-SPY2^33,55. These biomarkers particularly showed that specific subpopulations of T cells, the PD-L1 gene of CD274 and the colocalization or close spatial proximity of PD-1 + T cells with PD-L1 + cells (immune and tumor cells), are significantly associated with response to neoadjuvant therapy in the ‘pembrolizumab’ arm, but not the ‘control’ arm of the I-SPY2 trial. The significant correlation between CE9 and PD-L1 signature we found could indicate that this ecotype, characterized with an immune hot TME, achieves a higher response to neoadjuvant therapy when compared to other ecotypes and that the spatial organization of T cells and PD-L1 could serve as more accurate predictive biomarkers to identify patients who selectively benefit from immunotherapy rather than the TILs assessment on H&E or PD-L1 by IHC. Indeed, the investigation of PD1/PDL1 or TILs biomarkers was not included in the design of our study since several recent correlative studies on the I-SPY2 and other neoadjuvant therapy trials have consistently highlighted that TIL counts and PD-L1 expression do not identify patients who selectively benefit from the inclusion of an immune checkpoint inhibitor in the neoadjuvant chemotherapy regimen^8,55–57. Altogether, these findings highlight the significance of the multiomic profiling of the TME and integrating our spatially-defined ecotypes with recently-reported spatial biomarkers as a better way to characterize the relevant multicellular ecosystem phenotype that selectively benefits from neoadjuvant therapy. Fourth, the high number of cell types tested, which could similarly correlate with other covariates, and the small number of cases included in each cohort when breaking them down according to their response status decreased the power to observe significant findings when correcting the results of cell types for multiple testing. As such, larger hypothesis-testing studies are needed to validate our observations with cell types. Fifth, while our results on CE9 in both the ‘pembrolizumab’ and ‘control’ cohorts remained significant, associations between other ecotypes and pCR were not significant in multivariate analysis indicating that clinicopathological variables could confound the univariate analysis results. Thus, these findings should be interpreted with caution and further investigated in future studies. Lastly, the immune-enriched ecotypes CE9 and CE10 constituted 28% of the TNBC cases and our interaction test for CE9 was not significant. Thus, investigating our findings using larger cohorts is essential to prove their applicability in clinical practice. As our findings have a direct clinical value in the context of ongoing approvals of neoadjuvant immunotherapy in early-stage breast cancer, they still require a validation in additional similar clinical trial series. Of note, the significant results we report in HR + breast cancers in particular, support the potential investigation of ecotypes as a biomarker for pCR response in emerging clinical trials such as CheckMate-7FL⁵ and the KEYNOTE-756⁶.

Conclusions

In this study, we present data from a prespecified correlative analysis of the phase II I-SPY2 breast cancer clinical trial reporting that by using machine learning on transcriptomic data to define ecotypes, CE9 and its cell states identify patients with early-stage breast cancer most likely to achieve pCR after neoadjuvant therapy and may have a good prognosis. Whether patients classified as CE9 could still be considered for immunotherapy or be candidates for de-escalation strategies in the neoadjuvant setting requires further investigation in future studies with link to survival outcomes. In contrast, CE2 identifies patients who could benefit the most from the combination of immunotherapy with PARP inhibitors. These findings represent a promising way to allow a better characterization of the heterogeneity of breast cancer which could be integrated in a spatially-resolved context to build an ecosystem-based patient classification.

Author contributions

K.A contributed to conceptualization, methodology, data analysis, data interpretation, writing original draft, review and editing. G.B contributed to review and editing. P.M contributed to data interpretation, review and editing. All authors read and approved the final manuscript.

Data availability

This study involved the collection and analysis of data from publicly-available datasets. Fully normalized and batch corrected RNA data on the gene level for cases in the ‘pembrolizumab’ and ‘control’ arms of the I-SPY2 were obtained from the "ISPY2ResID_AgilentGeneExp_990_FrshFrzn_meanCol_geneLevel_n988.txt" Supplementary files of Wolf et al.⁷ and downloaded from Gene Expression Omnibus (GEO) SubSeries GSE194040 (mRNA) as part of the SuperSeries GSE196096—unique identifier: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE194040. Clinical data, hormonal status, and IHC subtype information were retrieved from the Supplementary Table S2 of Wolf et al.⁷ and the SuperSeries GSE196096—unique identifier: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE196096). Normalized RNA data and clinical information for the ‘durvalumab’ arm cases were available from Pusztai et al.⁸ using “GSE173839_ISPY2_AgilntGeneExp_durvaPlusCtr_FFPE_MEANcOL_GENElevel_n105.txt” and downloaded from GEO under the accession number of GSE173839—unique identifier: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE173839. No custom code was generated to collect data in this study. Any queries for R codes used in the analysis of this study can be directed to the corresponding author.

Declarations

Competing interests

The authors declare no competing interests.