Comprehensive Profiling of Poor-Risk Paired Primary and Recurrent Triple-Negative Breast Cancers Reveals Immune Phenotype Shifts

Abstract

Background:

Emerging data suggest immune checkpoint inhibitors have reduced efficacy in heavily pretreated TNBCs, but underlying mechanisms are poorly understood. To better understand the phenotypic evolution of TNBCs, we studied the genomic and transcriptomic profiles of paired tumors from TNBC patients.

Methods:

We collected paired primary and metastatic TNBC specimens from 43 patients and performed targeted exome sequencing and whole-transcriptome sequencing. From these efforts, we ascertained somatic mutation profiles, tumor mutational burden (TMB), TNBC molecular subtypes, and immune-related gene expression patterns. Stromal tumor-infiltrating lymphocytes (stromal TILs), recurrence-free survival, and overall survival were also analyzed.

Results:

We observed a typical TNBC mutational landscape with minimal shifts in copy number or TMB over time. However, there were notable TNBC molecular subtype shifts, including increases in the Lehmann/Pietenpol-defined basal-like 1 (BL1, 11.4-22.6%) and mesenchymal (M, 11.4-22.6%) phenotypes, and a decrease in the immunomodulatory phenotype (IM, 31.4-3.2%). The Burstein-defined basal-like immune-activated phenotype was also decreased (BLIA, 42.2-17.2%). Among down-regulated genes from metastases, we saw enrichment of immune-related KEGG pathways and GO terms, and decreased expression of immunomodulatory gene signatures (p<0.03) and percent stromal TILs (p=0.03). There was no clear association between stromal TILs and survival.

Conclusions:

We observed few mutational shifts, but largely consistent transcriptomic shifts in longitudinally paired TNBCs. Transcriptomic and immunohistochemical analyses revealed significantly reduced immune-activating gene expression signatures and TILs in recurrent TNBCs. These data may explain the observed lack of efficacy of immunotherapeutic agents in heavily pretreated TNBCs. Further studies are ongoing to better understand these initial observations.

INTRODUCTION

Triple-negative breast cancer (TNBC) represents approximately 15% of breast cancers and is characterized by the lack of expression of estrogen receptor (ER), progesterone receptor (PR) and human epidermal growth factor receptor-2 (HER2) (1). Despite recent FDA approval of atezolizumab for the first-line treatment of TNBC (2), the current standard-of-care treatment for TNBC is still cytotoxic chemotherapy, and the only FDA-approved targeted therapies are the poly-ADP ribose polymerase (PARP) inhibitors olaparib and talazoparib for BRCA1/2-mutant TNBCs (3,4). Regardless of early aggressive chemotherapy, the lack of effective targeted therapies in the advanced setting lends an overall poor prognosis for TNBC patients (5).

TNBC is clinically aggressive, with a high degree of chromosomal instability and extensive inter- and intra-tumoral heterogeneity (6,7). Differential gene expression profiling enables sub-classification of TNBCs into several molecular subtypes, the most commonly recognized of which are the Lehmann-Pietenpol (8) and Burstein (9) classification systems. In the former, TNBCs are molecularly grouped into six subtypes: basal-like 1 (BL1), basal-like 2 (BL2), immune-modulatory (IM), mesenchymal (M), mesenchymal stem-like (MSL), and luminal androgen receptor (LAR) (8). In a recent update, these subtypes have been revised and limited to four subtypes: BL1, BL2, M, and LAR (10). Similarly, Burstein et al. described four subtypes: luminal/androgen receptor (LAR), mesenchymal (MES), basal-like/immune suppressed (BLIS), and basal-like/immune-activated (BLIA) (9). In addition to tumor heterogeneity, acquired genetic subclones resulting from chemotherapy selection pressure may lead to treatment resistance (11-13).

With the development of large-scale molecular profiling technologies, understanding of the TNBC genomic landscape has improved, but currently available datasets from METABRIC [(14), Molecular Taxonomy of Breast Cancer International Consortium] and TCGA [(15), The Cancer Genome Atlas] are limited to primary, treatment-naïve breast cancers only. In-depth molecular analyses of metastatic TNBC compared to paired primary specimens are required to inform molecular changes as a result of chemotherapy selection pressure. Previous studies were limited by their sample size and the findings remain inconclusive (16).

Immune checkpoint inhibitors (ICIs) have been under rigorous investigation in multiple TNBC clinical studies (2,17-21). Recent results from IMpassion130 showed a promising response rate of 56% when atezolizumab was combined with nab-paclitaxel compared with 45.9% in TNBC patients treated first-line with nab-paclitaxel alone (2). In the neoadjuvant ISPY-2 study, adding pembrolizumab to a paclitaxel, adriamycin, and cyclophosphamide regimen increased the complete pathological response rate (pCR) from approximately 20% to 62% (n=21) (21). However, the efficacy of ICIs varies significantly depending on the line of therapy. For example, single agent checkpoint inhibition elicits a much lower response rate (5-6%) in the late-line setting compared to response rates of 19%-24% when administered as front-line treatment (17-20). These emerging clinical data suggest checkpoint inhibition is less effective in heavily pretreated TNBCs, but the underlying mechanisms are not well understood. To investigate the immuno-phenotypic evolution of TNBC, we studied the genomic and transcriptomic profiles of tumors from patients undergoing treatment for TNBC.

METHODS

Patient selection, ethical approval, and consent to participate:

Paired TNBC specimens were identified through a City of Hope IRB-approved retrospective protocol (IRB 07047) via the City of Hope (COH) Biorepository from patients with breast cancer treated at COH from 2002 to 2015. All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. Informed consent was obtained from all of the participants of this study. Eligible patients had the following features: TNBC with recurrence; at least one tumor biospecimen available from initial surgery or biopsy; at least one specimen available from relapsed disease biopsy; and clinical outcomes data. All collected samples were formalin-fixed paraffin-embedded (FFPE).

Histological assessments:

Histopathology of specimens was obtained from original pathology reports after being reviewed by two independent pathologists. ER, PR, and HER2 status were determined using standard ASCO/CAP guidelines. Immune cell subset profile changes were analyzed by TIL quantification of H&E stained slides, according to the International Immuno-Oncology Biomarker Working Group on Breast Cancer Guidelines (22).

Clinicopathological analysis:

Patient characteristics were obtained via chart review. Demographic data including age, gender, date of birth, date of diagnosis, date of relapse, and date of death (if applicable) were obtained. Tumor characteristics including tumor size, histology type, grade, lymph node involvement and treatment variables including chemotherapy were also obtained. Clinical outcomes including overall survival (OS), relapse-free survival (RFS), survival after relapse (SAR), and time between specimen collections (TBC) were also calculated. OS was defined as the date of surgery to date of death. RFS was defined as date of surgery to date of first relapse. SAR was defined as the date of relapse to the date of death. TBC was defined as the difference in time between the date of collection of the first specimen of a pair and the second specimen of a pair. Per physician decision, in the neoadjuvant or adjuvant setting [termed “adjuvant/neoadjuvant” herein], patients received either no chemotherapy, or one of the following regimens: Adriamycin-containing, platinum-containing, Adriamycin+platinum, or taxane-containing.

Targeted Next-Generation Sequencing (FoundationOne®):

Genomic alterations in FFPE specimens from primary and recurrent TNBCs were detected using the FoundationOne® targeted next-generation sequencing panel at a CLIA-certified, CAP-accredited reference laboratory (Foundation Medicine, Inc.). FoundationOne® identifies base substitutions, insertions and deletions (indels), amplifications with copy number ≥6, and rearrangements. More comprehensive details on the FoundationOne® platform version, sequencing, and mutation calling methodologies can be found in the Supplement.

Whole Transcriptome Sequencing (RNAseq):

Sequencing libraries were generated using the Illumina TruSeq RNA Access method, a hybridization-based protocol to enrich for coding RNAs from total RNA sequencing libraries. The method consists of two major steps: total RNA library preparation and coding RNA library enrichment. First strand cDNA synthesis is primed from total RNA using random primers, followed by the generation of second strand cDNA with dUTP (in place of dTTP) in the master mix. This facilitates the preservation of strand information, as amplification in subsequent steps will stall when it encounters Uracil in the nucleotide strand. Double stranded cDNA undergoes end-repair, A-tailing, and ligation of adapters that include index sequences. The resulting molecules are amplified via polymerase chain reaction (PCR), their yield and size distribution are determined using a BioAnalyzer, and their concentrations are normalized in preparation for the enrichment step. Libraries are enriched for the mRNA fraction by positive selection using a cocktail of biotinylated oligos corresponding to coding regions of the genome. Targeted library molecules are then captured through the hybridized biotinylated oligo probe using streptavidin-conjugated beads. After two rounds of hybridization/capture reactions, the enriched library molecules are PCR amplified, quantified, then normalized and pooled in preparation for sequencing. Please refer to the Supplement for comprehensive details on the methodologies for whole transcriptome data processing and data normalization.

Gene Expression and Pathway Analyses:

Differential gene expression was evaluated by applying an empirical Bayesian approach, using paired expression data from primary and metastatic samples for each patient. At the standard statistical threshold (false discovery rate (FDR)-adjusted p-value < 0.05), a total of 1,011 genes were significantly differentially expressed. Additionally, continuous Cox proportional hazards regression models were fit to the data to investigate the impact of individual gene expression on OS, RFS, SAR and TBC. In a Cox regression analysis, the effect of a predictor variable (e.g. the expression level of a gene) on the relative likelihood of an event (e.g. death or relapse) occurring at any given point in time can be expressed as the hazard ratio (HR). The HR value represents the predicted increase or decrease in this likelihood for a unit increase in the predictor. Hence, genes which are assigned HR values > 1 are associated with increased hazard rates (i.e. shorter time to event when the gene is more highly expressed), while those assigned HR values < 1 are associated with decreased hazard rates (i.e. longer time to event when the gene is more highly expressed). At a FDR-adjusted p-value < 0.05, few features had significant hazard ratios in the Cox regression analyses; therefore, a less stringent threshold of raw (unadjusted) p-value < 0.01 was used for comparisons.

Significant genes (raw p-value < 0.01) from each contrast were analyzed for significant enrichment of KEGG pathway membership and for GO terms across all three gene ontologies using a hypergeometric test. The resulting p-value was then corrected for tests over multiple KEGG pathways or GO terms using the method of Bejamini and Hochberg (1995) to yield an adjusted p-value. Enrichment (p < 0.01) was assessed for up- and down-regulated genes separately.

Gene Set Enrichment Analyses (GSEA):

For enrichment analyses, genes were ranked based on the p-value of differential expression. Specifically, log10(p-value) was used to rank genes, with negative / positive values used for down-regulated / up-regulated genes respectively. This ranked list was then used in Gene Set Enrichment Analysis (GSEA), as implemented by fgsea (23) to search for enrichment of KEGG pathway and GO term membership (across all three gene ontologies). The resulting p-values were corrected for tests over multiple KEGG pathways or GO terms using the method of Benjamini and Hochberg (1995) to yield an adjusted p-value.

Breast Cancer Molecular Subtyping:

The prediction of PAM50 subtype was based on a random-forest based classifier that was developed from the RNAseq data in TCGA (6) and 50 genes from the public PAM50 signature (24-26). Ninety-four randomly selected breast cancer samples from TCGA were used as an independent testing set, and the remaining samples were used to train the classifier. The Lehmann/Pietenpol classifier (8) was also trained by RNAseq data from TCGA with the subtype call from available microarray data. The subtype call from microarray was treated as the “true” label, and a nearest-centroid based classifier was developed. The classifier was then applied to the rest of TCGA samples without a microarray-based subtype call, and the performance was evaluated based on consensus clustering using all samples. The Burstein subtype classifier (9) was first evaluated using TNBC samples from TCGA with both available microarray and RNAseq data. Non-negative matrix factorization (NMF), a commonly used group of multivariate analysis algorithms used to cluster gene expression, was applied to both the microarray and RNAseq data to determine the gene and sample groups. Due to the high concordance of subtype calls by the 80 genes in the Burstein classifier across the microarray and RNAseq data (93%), we decided to directly use the 80 genes in the RNAseq data without further feature selection. The groups of 80 genes were then further evaluated in an independent training set of 215 TNBC samples from two historical clinical trials (NCT01375842 and NCT02162719), and a random-forest based classifier was trained with subtypes assigned by NMF. The classifier was implemented with an estimated “out-of-bag” error rate of 10%. The classifier was subsequently tested in an independent cohort of 51 TNBC samples from a third historical clinical trial (NCT02322814), and the NMF sample cluster result was ultimately considered as the label to evaluate performance.

Immune Gene Signature Scoring:

The immune gene signatures (gene lists) profiled in this study can be found in Supplementary Table S1. The composite score for the immune-activating genes (n = 7) from Denkert, et al., (27) was defined as the median log10(CPM) expression of the indicated genes. From Ayers, et al., (28) two gene expression signatures (lists) were analyzed: an 18-gene T-cell inflamed signature, and a 28-gene IFNgamma signature. As a result of RNAseq data processing and normalization procedures, the genes HLA-DRA and HLA-E were missing from the data and could not be analyzed in these two signatures. As such, scores for these two signatures were calculated based on the median log10(CPM) expression of the remaining genes in each signature. The score for the Th1 response activating gene signature (29-31) was calculated as the median log10(CPM) expression of the genes contained in the signature.

xCell Immune Cell Subtype Deconvolution Analysis:

The normalized RNAseq dataset was used to estimate cell subtypes using xCell (32) as implemented in the immunedeconv R package (33). Cell subtype proportions for each sample were summarised across the primary and metastatic samples and depicted as percentages in the heatmaps. Differences in proportions between the primary and metastatic groups were tested using a Wilcoxon signed rank test for each cell subtype, with p-values adjusted for multiple test inflation using Benjamini-Hochberg correction (Supplementary Table S2).

Survival Analysis:

To investigate the relationship between tumor-infiltrating lymphocytes (% stromal TILs) with survival (OS and RFS), Cox proportional hazard models were fit separately for primary and metastatic samples. Categorical classification of % stromal TILs (< 30% as “TIL-low” and ≥ 30% as “TIL-high”) were used as the predictor variables in the survival analyses. Hazard ratios (HR) were generated as in the above Methods section titled “Gene Expression and Pathway Analyses” and survival curves were generated using the Kaplan Meier method.

Other Statistical Analyses:

For all statistical analyses not listed in methods sections above, specific methods are delineated in the text and/or figure legends.

RESULTS

Summary of clinical characteristics:

Patient characteristics including histopathology, clinical staging, metastasis patterns, adjuvant/neoadjuvant chemotherapy, relapse-free survival (RFS), and overall survival (OS), are summarized in Table 1. Paired TNBC specimens from a total of 54 patients were analyzed, including 43 (79.6%) primary-metastatic (PM) pairs and 11 (20.4%) metastatic-metastatic (MM) pairs. Fifty-one (94.4%) patients had invasive ductal carcinoma (IDC), and 28 (51.9%) were stage II. Of the 43 paired PM samples, sites of metastases studied were: breast (n = 10, 23.3%); lymph node (n = 7, 16.3%), brain (n = 7, 16.3%), skin (n = 6, 13.9%), bone (n = 4, 9.3%), and other (n = 9, 20.9%). All but 7 patients received a chemotherapy-based treatment regimen. The most common regimen included Adriamycin with (n = 1, 1.9%) or without (n = 33, 61.1%) a platinum agent. Twelve (22.2%) patients received a taxane-based regimen. The majority (n = 37, 68.5%) of patients in the cohort recurred within 36 months (Table 1).

Table 1.

Summary of Patient Characteristics

Age (n = 54)	Years
Range	34-86
Median	51
Tumor Histology (n = 54)	N (%)
Invasive ductal carcinoma	51 (94.4)
Invasive lobular carcinoma	2 (3.7)
Metaplastic	1 (1.9)
Clinical Staging (n = 54)	N (%)
I	5 (9.3)
II	28 (51.9)
III	18 (33.3)
IV	3 (5.5)
Primary / Metastatic Pairs (n = 43)†	N (%)
Breast	10 (23.3)
Lymph node	7 (16.3)
Brain	7 (16.3)
Skin	6 (13.9)
Bone	4 (9.3)
Other sites*	9 (20.9)
Adjuvant/Neoadjuvant Chemotherapy (n = 54) ‡	N (%)
Adriamycin-containing	33 (61.1)
Platinum-containing	1 (1.9%)
Adriamycin + platinum	1 (1.9)
Taxane-containing	12 (22.2)
No chemotherapy	7 (12.9)
Relapse-Free Survival (n = 54)	N (%)
< 36 months	38 (70.4)
36 – 60 months	9 (16.7)
> 60 months	7 (12.9)
Overall Survival (n = 54)	N (%)
<36 months	18 (33.3)
36 – 60 months	14 (25.9)
> 60 months	13 (24.1)
Under active treatment	7 (13.0)
Lost to follow-up	2 (3.7)

^†11 paired samples were longitudinal metastatic pairs

^*Other sites included ovary (2), pleural effusion (2), liver (1), muscle mass (1), adrenal gland (1), pericardium (1), chest wall (1).

^‡Chemotherapy regimen listed was administered either as neoadjuvant or adjuvant treatment, per physician decision.

Genomic landscape of TNBC from primary to metastatic disease:

Thirty-four PM TNBC pairs (68 specimens) were successfully sequenced using the FoundationOne® targeted next-generation sequencing assay (Figure 1, Supplementary Figure S1). The most commonly observed genomic mutations were consistent with previously-reported TNBC genomics (15). Of 34 paired primary and metastatic specimens, respectively, TP53 was mutated in 29 (85.3%) and 30 (88.2%) tumors; MYC was mutated in 7 (20.6%) and 9 (26.5%); PIK3CA was mutated in 6 (17.6%) and 7 (20.6%); and PTEN alterations were observed in 6 (17.6%) and 6 (17.6%) (Figure 1A). Figure 1A only displays known oncogenic variants, per Foundation Medicine, Inc. curation (see Methods). As such, variants of unknown significance (VUS, Supplementary Figure S2A - S2C), inclusive of many BRCA1/2 variants, are not visible on the plot in Figure 1A. BRCA1 was mutated in 2 (5.9%) and 2 (5.9%); BRCA2 was mutated in 6 (17.6%) and 6 (17.6%). Of note, both the primary and metastatic specimens from two patients exhibited amplifications of CCND1 and FGFs 3, 4, and 19, which is typically observed in 30-40% of hormone receptor-positive breast cancers (34,35). The remaining observed mutations occurred sporadically throughout the cohort.

Figure 1. Genomic landscape of primary-metastatic TNBCs.

FoundationOne® targeted DNA sequencing was successfully performed on 34 primary-metastatic (PM) TNBC pairs. (A) The mutational landscape of these tumors was typical of TNBCs and similar between pairs. This figure displays data from patients who had a PM pair of specimens [i.e. no met-met (MM) specimens], and only known or likely oncogenic variants are shown (VUSs are not included). (B) Number and (C) percent of known or likely oncogenic variants unique to primary specimens, unique to metastatic specimens, or shared between pairs. (D) Of the 34 PM TNBC pairs, 21 pairs yielded tumor mutational burden (TMB) results. With the exception of one primary sample and one metastatic sample (from different pairs), TMB was low/intermediate (< 16 mutations per megabase) overall and relatively unchanged between TNBC pairs.

While the individual frequency of shared and unique mutations between PM TNBC pairs may have varied by patient (Figures 1B & 1C), overall, 50% or more of mutations were shared between PM TNBC pairs. Specifically, of mutations annotated to be of known/likely oncogenic significance, 50% (n = 123) were shared between pairs, 16.3% (n = 40) were unique to primary TNBC specimens, and 33.7% (n = 83) were unique to mTNBC specimens (Figures 1B & 1C). When VUSs were considered, 56.8% of mutations (n = 536) were shared, 12.1% (n = 114) were unique to primary specimens, and 31.1% (n = 294) were unique to metastatic specimens (Supplementary Figures S2B & S2C). Overall, the mutational differences appear largely sporadic and are not consistent with a given gene or pathway (Supplementary Figure S2D).

Similar to the mutational landscape, very few copy number changes were observed in the transition from primary to metastatic disease (Supplementary Figure S2E). Finally, tumor mutational burden (TMB) was less than 16 mutations per megabase (mut/Mb) (Figure 1D, TMB >16 in one primary tumor and one metastatic tumor from different pairs) and no significant changes in TMB were observed between primary and metastatic TNBC pairs.

Breast cancer molecular subtype shifts between primary and metastatic disease:

To understand the transcriptomic features of our TNBC cohort, we performed whole-transcriptome sequencing (RNAseq) on PM pairs from 35 patients (70 specimens) (Supplementary Figure S1). Intrinsic breast cancer subtype phenotyping by PAM50 analysis (38-40) confirmed that 88.9% of primary samples were basal-like and remained basal-like (87.5%) through the transition to recurrent disease (Supplementary Figure S3).

Previous microarray and whole-transcriptome breast cancer studies have led to the generation of TNBC molecular subtype classifiers. These include the 6-subtype Lehmann/Pietenpol classifier (BL1, basal-like 1; BL2, basal-like 2; IM, immunomodulatory; M, mesenchymal; MSL, mesenchymal stem-like; LAR, luminal androgen receptor) (8), and the 4-subtype Burstein classifier (LAR, luminal androgen receptor; MES, mesenchymal; BLIS, basal-like immune suppressed; BLIA, basal-like immune activated) (9). Similar to the PAM50 analysis above, we used both the Lehmann/Pietenpol and Burstein molecular subtype classification systems to characterize the tumors in our dataset and to observe any shifts in these molecular phenotypes from primary to metastatic disease (see Methods). From the Lehmann/Pietenpol system, we observed an increase in tumors defined as basal-like 1 (BL1, 11.4% to 22.6%), an increase in tumors defined as mesenchymal (M, 11.4% to 19.4%), and a drastic decrease in tumors defined as immunomodulatory (IM, 31.4% to 3.2%) from primary to metastatic tumors (Figure 2A). Similarly, using the Burstein classification system, we observed a downward shift in tumors characterized as basal-like immune activated (BLIA, 42.2% to 17.2%), with a concomitant increase in those characterized as basal-like immune suppressed (BLIS, 42.2% to 62.1%) (Figure 2B). Together, these data suggest that TNBCs become less immunologically rich over time.

Figure 2. Breast cancer molecular subtype shifts between primary and metastatic TNBC pairs.

(A) Classification of TNBC pairs into Lehmann/Pietenpol defined subtypes revealed an increase in the basal-like 1 phenotype (BL1, 11.4% to 22.6%), an increase in the mesenchymal phenotype (M, 11.4% to 19.4%), and a significant decrease in the immuno-modulatory phenotype (IM, 31.4% to 3.2%). BL1, basal-like 1; BL2, basal-like 2; IM, immunomodulatory; LAR, luminal androgen receptor; M, mesenchymal; MSL, mesenchymal stem like; UNS, unspecified. (B) Burstein-defined classifications revealed a decrease in the basal-like immune-activated phenotype (BLIA, 42.2% to 17.2%), a converse decrease in the basal-like immune-suppressed phenotype (BLIS, 42.2% to 62.1%), and an increase in the mesenchymal phenotype (MES, 0% to 6.9%). BLIA, basal-like immune-activated; BLIS, basal-like immune-suppressed; LAR, luminal androgen receptor; MES, mesenchymal. These analyses were performed on specimens from patients with primary-metastatic pairs and of those, specimens for which a breast cancer subtype score was able to be assigned. Please note percentages may not add exactly to 100% due to rounding.

Tumor-infiltrating lymphocytes and survival:

A number of prior TNBC studies have shown that stromal tumor-infiltrating lymphocyte (TIL) levels exhibit a prognostic association with outcome in patients receiving adjuvant or neoadjuvant chemotherapy (27,36-39). We observed a modest, yet statistically significant decrease in the percent of stromal TILs in primary versus recurrent specimens (p = 0.02, unpaired Mann-Whitney U test; Figure 3A). Of 37 patients for which stromal TILs could be scored for both these patients’ paired primary and metastatic specimens, 21 (56.8%) patients’ tumor pairs displayed a decrease in stromal TILs in their metastatic disease specimens; 8 (21.6%) patients’ tumor pairs displayed an increase in stromal TILs in metastatic disease; and 8 (21.6%) patients’ tumor pairs exhibited no change in stromal TILs (Figure 3B). Despite the PM pairs for which TILs either increased or did not change, the paired statistical analysis still revealed a statistically significant overall decrease in TILs in this population (p = 0.03, paired Student’s t-test, Figure 3B). No significant differences were observed when TILs were further sub-analyzed by line of treatment (data not shown). These data are consistent with our analysis of TNBC molecular subtypes, above.

Figure 3. Tumor infiltrating lymphocytes and survival.

(A) Histopathology-derived percent stromal TILs were significantly decreased in mTNBCs [p = 0.02, unpaired Mann-Whitney U (MWU) test]. (B) Plotting only paired PM specimens longitudinally, while TILs are overall significantly decreased (p = 0.03, paired Student’s t-test), some primary-to-metastatic pairs exhibited increases in stromal TILs in metastatic disease. (C) Higher stromal TILs (≥30%) were not associated with improved overall survival when measured in primary specimens (p = 0.2), or (D) with the time from relapse to death when measured in recurrent specimens (p = 0.74). See Methods section for statistical considerations.

Using a cut-off of < 30% stromal TILs for “TIL-low” and ≥ 30% stromal TILs for “TIL-high” in this cohort (40), we analyzed TIL levels in relation to overall survival (OS) and survival after relapse (SAR). For OS analysis, TILs scored in primary tumors were used; for SAR analysis, TILs scored in recurrent tumors were used. In this dataset, no significant difference in OS or SAR was observed at the defined 30% stromal TILs cut-point (Figure 3C, Figure 3D).

Finally, while difficult to make any statistical inferences due to subgrouping of an already limited sample size, we observed a correlative trend of elevated stromal TILs with the IM, BLIA, and Basal-like breast cancer molecular subtypes as defined by the Lehmann/Pietenpol, Burstein, and PAM50 classifications, respectively (Supplementary Figure S3B - S3D).

Differential gene expression analysis:

RNAseq was successfully performed on 35 PM pairs of specimens from the study cohort (Supplementary Figure S1) upon which we employed a series of downstream analyses to understand gene expression as it relates to disease state, survival, and adjuvant/neoadjuvant therapy. First, comparing gene expression between primary and metastatic disease, significantly up- and down-regulated differentially expressed genes were assessed for KEGG pathway (Figures 4A, 4B, and Supplementary Figures S4A, S4B) and GO term (Figures 4C, 4D, and Supplementary Figures S4C, S4D) enrichment. A clear enrichment of immune-related KEGG pathways was found among genes that were down-regulated in metastatic samples compared to primary samples (Figure 4B, Supplementary Figure S4B). These included cytokine-cytokine receptor interactions, Th1 and Th2 cell differentiation, Th17 cell differentiation, natural killer cell mediated cytotoxicity, T and B cell receptor signaling, and NFKB signaling. Interestingly, many of these pathways were also enriched among genes lending to hazard ratios (HR) < 1 in the Cox proportional hazards modeling for the OS, SAR, and TBC comparisons for primary (but not for metastatic) samples (see Methods, Supplementary Figure S4B). We observed similar findings in the comparison between primary and metastatic samples through a ranked Gene Set Enrichment Analysis (GSEA) (Supplementary Figures S4E - S4H).

Figure 4. KEGG pathway and GO term enrichment between primary and metastatic TNBC.

KEGG pathway enrichment analysis associated with genes with (A) up-regulated expression or (B) down-regulated expression in metastatic compared to primary TNBCs revealed a clear enrichment of immune-related KEGG pathways amongst genes that were significantly down-regulated in the metastatic samples. GO term enrichment analysis associated with genes with (C) up-regulated expression or (B) down-regulated expression in metastatic versus primary TNBCs revealed enrichment of a number of immune-related GO terms among down-regulated genes in metastatic specimens. Bar plots herein display the pathway/term enrichment scores with p < 0.01. If more than 50 pathways/terms were significantly enriched, only the top 50 pathways are displayed for visualization purposes.

Next, gene expression was analyzed separately in primary and metastatic samples using Cox proportional hazards regression modeling to understand whether expression of particular genes was associated with impacts to response variables, including overall survival (OS), relapse-free survival (RFS), survival after relapse (SAR), and time between specimen collections (TBC). A number of immune-related GO terms were enriched among down-regulated genes in metastatic compared with primary samples, and among genes with HR < 1 in the OS, SAR, and TBC comparisons (Figure 4D, Supplementary Figure S4D). These included T cell aggregation, B cell receptor signaling, immune response activation, and the adaptive immune response. Together, these data indicate that genes comprising these KEGG pathway and GO terms are expressed at lower levels in metastatic versus primary TNBC samples, and increased expression of these genes in primary samples is associated with longer survival times. Other KEGG pathway and GO term comparisons were less thematic when contextualized with our broader data. Among up-regulated genes, one KEGG pathway (spliceosome) was commonly enriched in the following comparisons: gene expression in metastatic versus primary specimens, gene expression in primary specimens versus RFS time, gene expression in metastatic specimens versus RFS time, and gene expression in primary specimens versus TBC (Figure 4A, Supplementary Figure S4A). Also among up-regulated genes, GO terms associated with cellular cross-talk including cell-cell adhesion, extracellular exosomes, and extracellular organelles were enriched in the following comparisons: gene expression in metastatic versus primary specimens, gene expression in primary specimens versus RFS time, and gene expression in primary specimens versus TBC (Figure 4C, Supplementary Figure S4C).

Finally, to compare gene expression in patients as it relates to adjuvant/neoadjuvant therapy received, we performed differential gene expression in primary and metastatic samples comparing between the following categories of adjuvant/neoadjuvant therapy: Adriamycin-containing (adriamycin cyclophosphamide, AC; cyclophosphamide adriamycin 5-FU, CAF; docetaxel adriamycin cyclophosphamide, TAC; or adriamycin cyclophosphamide paclitaxel, AC-T), platinum-containing (carboplatin paclitaxel), adriamycin and platinum-containing (adriamycin cyclophosphamide-carboplatin paclitaxel), taxane-containing (docetaxel cyclophosphamide, TC), or no adjuvant/neoadjuvant therapy. Gene expression in primary specimens from patients who received Adriamycin+platinum-containing regimens versus those who received Adriamycin-containing therapies (without a platinum-based agent), taxane-containing therapies, or no adjuvant/neoadjuvant therapy revealed the only notable significant associations with specific KEGG pathways or GO terms. Specifically, the KEGG pathway “microRNAs in cancer” was enriched amongst up-regulated genes as determined from primary specimens in the three adjuvant/neoadjuvant regimen comparisons above (Supplementary Figure S4A). The GO terms extracellular exosome, extracellular organelles, and vesicles were enriched amongst significantly down-regulated genes in these comparisons (Supplementary Figure S4D). No notable associations were observed in adjuvant/neoadjuvant therapy comparisons using gene expression from metastatic samples.

Immune phenotyping analysis of paired primary and mTNBCs:

Considering the findings from the differential gene expression analyses above revealed a strong association of significantly down-regulated genes in mTNBCs with immune-related KEGG pathways and GO terms, we further profiled the RNAseq data from our cohort of specimens against four published gene signatures (lists) associated with immunologically active phenotypes: a 14-gene Th1 response activating score, an 18-gene T cell inflamed score, a 28-gene IFNγ score, and a 7-gene immune-activating score (37-41). Refer to the Methods for details on scoring and Supplementary Table S1 for the genes that comprise each signature. Composite intensity scores from three of the four gene signatures were significantly downregulated in metastatic samples compared to primary samples [Figure 5A (heatmap) and Figure 5B - 5E (boxplots)]. Not surprisingly, we observed a concomitant trending decrease in the immune-related gene signature scores by line of therapy (Supplementary Figure S5A - S5D) and positive correlation of each of these scores with percent stromal TILs (Supplementary Figure S6A - S6D).

Figure 5. Immune phenotype profiling between TNBC pairs.

(A) Heatmap of composite expression scores of immuno-modulatory gene signatures between PM TNBC specimens. (B - E) Boxplots of the same, to illustrate decreases in immune-modulatory gene signature scores between PM TNBC samples. Statistical analyses between primary and metastatic specimens were performed using the Mann-Whitney U test (unpaired) and p-values < 0.05 were considered significant. (F) Cell subtype proportions were summarized across primary and metastatic samples and depicted as an average percentage in the heatmap. Differences in proportions between the primary and metastatic groups were tested using a Wilcoxon signed rank test for each cell subtype, with p-values adjusted for multiple test inflation using Benjamini-Hochberg correction. P-values are presented in Supplementary Table S2.

To further understand possible shifts in immune-related cell subtypes, we estimated cell subtypes on paired primary and metastatic specimens for which RNAseq was successful using the xCell immune cell deconvolution analysis (see Methods). The fraction of cell subtypes averaged across primary samples, and the fraction of cell subtypes averaged across metastatic samples, is summarized in Figure 5F. Consistent with the immune-related gene signature phenotyping, in the transition from primary to metastatic disease, we observed significant decreases in the proportions of B cell, CD4+ naive T cell, CD8+ T cell, and cancer-associated fibroblast subtypes. Conversely, we observed significant increases in the proportions of endothelial cell, macrophage and M1 macrophage subtypes. A heatmap of the fraction of cell subtype for individual samples can be found in Supplementary Figure S5E and p-values of the averaged fractions can be found in Supplementary Table S2.

DISCUSSION

Molecular evolution of TNBC through chemotherapy selection pressure and the impact of this evolution toward next therapeutic steps are well recognized but poorly understood. In this study, we described the molecular changes observed in the largest paired primary-mTNBC cohort reported to-date. We observed few mutational shifts and overall low TMB suggesting that tumors from patients treated with chemotherapies tend to maintain their genomic profiles over the course of therapy. We did, however, observe consistent transcriptomic shifts in longitudinally paired TNBCs. Transcriptomic and IHC analyses revealed significantly reduced expression of immune-activity associated gene expression signatures and of TILs in recurrent TNBCs. These data support the hypothesis that mTNBCs are less immunogenic compared to primary TNBCs and explains the lack of efficacy of ICIs in heavily pretreated patients in early clinical trials.

To better understand the molecular heterogeneity of TNBCs, several molecular classifiers have been developed (i.e. Lehmann/Pietenpol (8) and Burstein (9)). While helpful for our understanding of TNBC tumor biology, these molecular classifiers have little impact in guiding clinical practice. Previous studies comparing primary and metastatic breast cancers have provided variable results (41). Some studies indicated concordant overall expression patterns between primary breast cancers and their matched lymph nodes (42-45). Other studies identified discordant overall gene expression patterns between primary tumor and synchronous metastases (46). Weigelt, et al. studied 7 cases of primary breast cancer and asynchronous distant metastases and showed that a 70-gene prognostic signature (Mammaprint®) was generally maintained in the switch from primary to metastasis across most of the pairs (47). As we attempted herein, some studies have used targeted NGS to address genomic concordance between primary and metastatic disease. For example, and confirming our findings, Meric-Bernstam and colleagues reported that 86.6% of the somatic mutations and 62.3% of the copy number variations were concordant between primary tumors and recurrences (16). These studies highlight the limitation of tumor biopsies in capturing the heterogeneity of TNBC genomics and indicate the necessity of using alternative approaches for a more comprehensive understanding of the mutational spectrum of TNBCs (48). Furthermore, noncoding RNAs and epigenetic modifications may also play an important role in the metastatic process (42,49,50).

With recent FDA approval, ICI therapy is now a reality for patients with TNBC. As such, defining the appropriate population for such therapy is imperative. Atezolizumab increased PFS from 5.5 months to 7.2 months when added to a nab-paclitaxel as first-line treatment of patients with PD-L1 positive, mTNBC (2,19). In I-SPY2, more than 60% of early-stage TNBC patients treated with neoadjuvant ACT (Adriamycin, cyclophosphamide and paclitaxel) + pembrolizumab achieved a pCR, while patients who received chemotherapy alone only achieved a ~20% pCR rate (21). In heavily pretreated patients, however, response rates to single-agent ICIs decreases significantly. In cohort A of KEYNOTE-086 (PD-L1 agnostic cohort), 43.5% of patients with ≥ 3 lines of prior therapy achieved a response rate of 5.3% (18), suggesting that PD-L1 positivity is less of a response predictor in heavily pretreated patients. In contrast, in the PD-L1 positive cohort B of KEYNOTE-086, an overall response rate of 23% was observed in the first-line setting (17). Similarly, single agent atezolizumab elicited a response rate of 24% (5 of 21) in patients with first-line mTNBC but only 5% (6 of 94) in patients with second-line or greater disease (19). These clinical trial data demonstrate that early intervention with immune checkpoint inhibition is associated with enhanced efficacy. Our study confirmed that metastatic breast cancers are immunologically more inert than their corresponding primary tumor, which is consistent with a recently published study by Szekely et al (41). Our study provides a possible mechanistic explanation for clinical observations and further supports early incorporation of immune checkpoint blockade for TNBC patients in the adjuvant/neoadjuvant setting.

This study is limited by its sample size, retrospective nature, and data in a cohort of patients not treated with immunotherapy. The sample size also precludes the ability to perform a subset analysis of the association of TIL levels and immune signature scores with sites of metastasis. Furthermore, our study’s observed longitudinal genomic ‘similarities’ could be a limitation of targeted exome sequencing which is unable to capture the entirety of genomic changes and epigenomic variations.

TRANSLATIONAL RELEVANCE.

The molecular heterogeneity of triple-negative breast cancers (TNBCs) lends difficulty in identifying effective targeted therapies against TNBC. Atezolizumab was recently approved for the treatment of PD-L1-positive, unresectable locally advanced or metastatic TNBC in combination with protein-bound paclitaxel. Results from recent clinical trials, however, suggest immune checkpoint inhibitors have reduced efficacy in heavily pretreated, and presumably, more heterogeneous TNBCs. The work described herein suggests that in the transition from primary to metastatic disease, TNBCs treated with traditional chemotherapy regimens exhibit decreased immune activity over time. This is illustrated by decreased stromal tumor-infiltrating lymphocytes (TILs) and decreased expression of immune activity-related gene signatures in metastatic TNBCs. These phenotypes may partially explain the limited efficacy of immune checkpoint inhibitors in the metastatic setting and highlight the importance of their use instead in the early neoadjuvant setting.