Integrative Multiomic Profiling of Triple-Negative Breast Cancer for Identifying Suitable Therapies

Triple-negative breast cancer (TNBC) is a highly heterogeneous disease characterized by the absence of estrogen (ER), progesterone (PR), and human epidermal growth factor 2 (HER2) receptors (1, 2) and an increased likelihood of distant metastasis and death compared with other subtypes of breast cancer (3). Treatment of patients with TNBC has been complicated by the heterogeneity of the disease and the absence of well-defined molecular targets agreeable to therapeutic intervention (4). The standard treatment for TNBC is centered on cytotoxic chemotherapies, antibody-drug conjugates, and immune checkpoint blockade in a subset of cases. Unfortunately, there are currently no effective therapeutic options for patients with advanced TNBC that have progressed after several lines of cytotoxic chemotherapy. Thus, although TNBC constitutes approximately 20% of all breast cancers, it is the cause of almost half of all breast cancer deaths (5), and novel treatment approaches are required.

Several therapies that are either used or under development for patients with TNBC are targeted to tumors that harbor specific biomarkers. Examples include (i) anti-programmed death-1/programmed death ligand-1 (PD1/PDL1) antibodies for tumors expressing the PDL1 protein, (ii) PARP inhibitors for tumors in patients with germline mutations in BRCA1/2 and possibly other hereditary breast cancer genes, (iii) androgen receptor (AR) antagonists for tumors expressing AR protein or a gene expression signature consistent with androgen dependence, (iv) cyclin-dependent kinase 4 and 6 (CDK4/6) inhibitors for tumors expressing the retinoblastoma (Rb) protein and/or a “luminal” gene expression profile, and (v) PI3K and AKT inhibitors for tumors harboring genetic alterations associated with hyperactive PI3K-AKT-mTOR signaling. In addition, new antibody-drug conjugates such as trastuzumab deruxtecan might be more effective in a subset of TNBC that display cell surface expression of HER2. Having parallel biomarker-driven therapies, the lack of an established treatment algorithm or head-to-head comparison of any targeted therapies can be a source of confusion, as any individual tumor is likely to test positive for more than one, but not all, of the biomarkers. Adding to this complexity is that TNBC are variably classified in terms of their mutational states or gene expression profiles (i.e., TNBC subtypes defined by transcriptomic profiling).

To help address this complexity, we have performed multiomic profiling on a cohort of 95 TNBC (both primary and metastatic tumors). By analyzing DNA, RNA, protein, and cellular biomarkers in these tumors, we determine how various biological parameters are correlated, whether they are associated with superior or inferior clinical outcomes, and how they might be utilized to inform decisions for drug development. Moreover, by including complete de-identified data for each TNBC case analyzed (including gene expression profiles, outcome data, and information about all treatments received for metastatic disease), we provide a rich resource for other investigators to generate or validate novel hypotheses relevant to TNBC.

We recently conducted an Institutional Review Board–approved, single-arm phase 2 trial of the CDK4/6 inhibitor abemaciclib as a treatment for advanced TNBC that express the Rb protein as measured by IHC (NCT03130439). As part of prescreening for this trial, patients with advanced TNBC provided consent for archival tumor tissue (either primary breast cancers or metastatic biopsies) to be tested for Rb expression by IHC. In addition, patients provided consent for broader molecular profiling of their tumor tissue, including all analyses performed for this study. The current analysis provides data for 95 TNBC samples from patients who were originally prescreened for a clinical trial (Supplementary Fig. S1). The study was approved by the Dana-Farber/Harvard Cancer Center Institutional Review Board and was conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from each subject. Please see Supplementary Table S1 for data on the representativeness of our study population.

Formalin-fixed paraffin-embedded samples were acquired from 95 TNBC. IHC for AR (clone AR441, Dako, catalog No. M3562–RRID: AB_2060174; used at 1:200), Rb (clone G3-245–RRID: AB_395259, BD Bio, catalog No. 554136; used at 1:75), and PDL1 (clone 405.9A11, Cell Signaling Technology, catalog No. 29122–RRID: AB_2798970; used at 1:100) was performed on all tumors. Antigen retrieval for AR was performed using citrate buffer (pH 9.0) in a decloaking chamber at 98°C (Biocare Medical). All antibodies were incubated for 30 minutes at room temperature. Visualization was achieved using the 4 plus Detection System (BioCare) with 3,3′- diaminobenzidine (DAB; DAKO) as the chromogen.

Two breast pathologists (S. Schnitt and E. T. Richardson) quantified stromal tumor-infiltrating lymphocytes (sTIL) on hematoxylin and eosin–stained TNBC sections (various sites, listed in Table 1) according to guidelines from the International TILs Working Group (6). sTIL were defined as low (sTIL-lo, ≤10%) or high (sTIL-hi, >10%). The pathologists also quantified the staining intensity (0, 1+, 2+, and 3+) and the percentage of cell staining for AR, Rb, and PDL1. Rb positivity by IHC was defined as ≥50% of tumor cells stained (7). AR positivity was defined as ≥1% of tumor cells stained (8). PDL1 staining was scored separately for both tumor cells and immune cell (IC), and in each case, positivity was defined as ≥1% of the respective cell type stained. For Rb and AR stains, an H-score [defined as % cells stained (cs) at score 1+ plus (%cs at score 2+ *2) plus (% cs at score 3+ *3); range, 0–300] was also calculated (9).

Hematoxylin and eosin–stained sections were used to identify tumor-containing regions in each specimen (marked by breast pathologists S. Schnitt and E. Richardson). Marked samples were then used to guide the microdissection of tissue from adjacent, serial 5-µm tissue sections. RNA was isolated using All-Prep extraction kit (Qiagen) and quantified using RNA detection (Life Technologies). RNA was analyzed using the Breast Cancer 360 (BC360) panel (NanoString Technologies, catalog No. XT-CSPS-HBC360-2-12). The BC360 panel measures expression of 758 genes, some of which are used in algorithms to score 48 biological signatures (e.g., breast cancer subtyping, immune signaling, tumor regulation, differentiation, and mutational response).

For panel standard (NanoString Technologies, catalog No. PSTD-HBC360-2-12), a DNA oligo blend containing all BC360 probe target sequences was used and run on each cartridge within the experiment for normalization of non-PAM50 genes, whereas for reference samples, an RNA oligo blend containing PAM50 probe target sequences for PAM50 genes was used. The normalization was performed in two steps. In the first step, the process was adjusted based on whether the genes were part of the PAM50 signature, the Tumor Inflammation Signature (TIS), or neither. Details of this adjustment are provided below. Before normalization, any zero counts in the raw data were converted to ones.

Genes in this signature were normalized using a ratio of the expression value to the geometric mean of the housekeeper genes used only for the TIS.

Genes in this signature were normalized using a ratio of the expression value to the geometric mean of the housekeeper genes used only for the PAM50 signature.

Genes were normalized using a ratio of the expression value to the geometric mean of all housekeeping genes on the panel.

Genes not in the PAM50 signature were additionally normalized using a ratio of the housekeeper-normalized data and the average of two panel standards run in the batch cartridges as the observed data.

Genes in the PAM50 signature were additionally normalized using a ratio of the housekeeper-normalized data, and a reference sample run on the same codeset lot from a NanoString archive was used.

The housekeeper-normalized and panel standard-normalized data were log2 transformed. A constant of eight was added to TIS so that it was on the same scale as investigational use–only TIS, making scores comparable across research use–only and investigational use–only assays. Other non-TIS were also adjusted with constants to express values in a similar range.

Differential expression was fit on a per-gene or per-signature basis using a linear model for analyses without a blocking factor. The statistical model was based on the usage of the expression value or signature score as the dependent variable and fits a grouping variable as a fixed effect to test for differences in the levels of that grouping variable. P values were adjusted within each analysis, gene, or signature, and on the grouping variable level difference t test using the Benjamini and Yekutieli FDR adjustment to account for correlations among the tests. All models were fitted using the limma package in R.

For each patient, we calculated overall survival (OS), defined as the time from diagnosis of metastatic disease to the time of death. Kaplan-Meier curves were created using the survFit function from the R package survival and then plotted using the survMiner package in R. For all Kaplan‐Meier plots, a log-rank test was used for generating P values.

Analyses herein (i.e., IHC, sTIL, PAM50 subtypes, TNBC subtypes, and genotypes) were performed as follows: (i) The Fisher exact test and χ² were used for between-group comparisons among the categorical variables of interest, and (ii) the Mann‐Whitney test was used for testing the continuous outcomes between the independent study groups. All tests were two-tailed, the statistically significant level was set at 0.05, and P values were not adjusted for multiple comparisons. All analyses were performed with PRISM software.

OncoPanel is an in-house (Dana-Farber Cancer Institute) next-generation sequencing platform for surveying coding regions of 447 cancer-related genes (10). OncoPanel testing was performed at the Center for Advanced Molecular Diagnostics, a Clinical Laboratory Improvement Amendments–certified laboratory in the Department of Pathology at Brigham and Women’s Hospital. Of note, tumor samples used for OncoPanel analyses were not always the same as those used for other assays described in this study, and thus in some cases may reflect sequencing results from tissue procured at a different time or from a different organ site. DNA from formalin-fixed, paraffin-embedded patients’ samples was isolated using standard extraction methods (Qiagen) and quantified using PicoGreen dsDNA detection (Life Technologies). DNA sequencing results were available from 64 out of 95 (67%) cases (Table 1). We focused our analyses on commonly altered and/or clinically relevant genes including members of the PI3K pathway (aggregation of PIK3CA, PTEN, AKT1, AKT2, and AKT3), BRCA1, and BRCA2, TP53, and RB1. We performed classification based on the following criteria: (i) RB1 alteration was defined as mutation and/or deletion (regardless of being mono- or bi-allelic), (ii) BRCA1/2 alterations were defined as detection of pathogenic variants in tumor tissue but this was not confirmed with germline analysis, and (iii) PI3K pathway alterations were defined as activating mutations in PIK3CA or AKT1-3 and/or loss of/mutation in PTEN.

Supplementary Tables S2 through S6 contain data used to generate results in this study, including gene counts for gene expression profiling. Raw data files for exome sequencing are not available for public dissemination. For any data requests, please contact the corresponding author: [email protected].

Our cohort comprised 95 cases of TNBC, and the characteristics of these cases are provided in Table 1. The complete data for all the parameters analyzed in this study are contained in Supplementary Table S2.

CDK4/6 kinases exert their effect in large part through phosphorylation of the Rb tumor suppressor protein, and lack of functional Rb is associated with resistance to these agents (11, 12). Given this, trials of CDK4/6 inhibitors for TNBC often restrict eligibility to tumors most likely to have retained Rb function. At present, there is no standardized assay to determine Rb functional status in a tumor. Clinical studies often rely on RB1 gene deletion or mutation as evidence of nonfunctional Rb (13), but preclinical studies typically measure RB1 mRNA levels and/or the presence of Rb protein (11, 14). Thus, our goal was to evaluate associations between RB1 DNA mutation status, mRNA levels, and Rb protein levels in our cohort of TNBC and to determine their respective correlations with inferred markers of Rb function. We observed a significant positive correlation between RB1 mRNA and Rb protein expression regardless of whether Rb protein expression was classified as a categorical (positive vs. negative) or continuous (H-score) variable (Fig. 1A and B). In contrast, we did not observe a significant correlation between genetic alterations in RB1 (mutation and/or deletion) with either RB1 mRNA or Rb protein levels (Fig. 1C and D).

Previous studies have shown a reciprocal association between levels of Rb and p16 proteins in solid tumors, and high levels of p16 (encoded by CDKN2A) predict the presence of nonfunctional Rb (15, 16). When assessed by IHC, Rb-positive tumors showed significantly lower CDKN2A expression (Fig. 1E; Supplementary Table S3). In contrast, genetic alterations in RB1 were not associated with changes in CDKN2A expression (Fig. 1F). These observations suggest that assessment of RB1 mutation/deletion may not be sufficient to determine Rb function in TNBC and that assessment of Rb by IHC should be considered.

We next explored associations between Rb expression (protein and/or mRNA) and other parameters in TNBC. Transcriptomic profiling has uncovered the presence of several TNBC subtypes including the luminal androgen receptor (LAR) subtype (17, 18), and preclinical studies have suggested that LAR cell lines are sensitive to CDK4/6 inhibition in vitro (19). In our cohort, Rb IHC positivity was not associated with the TNBC subtype or the PAM50 subtype (as determined by the NanoString BC360 assay); however, we did observe a positive (nonsignificant) correlation between the expression of Rb and AR at the protein level (Supplementary Fig. S2A–S2C).

For DNA alterations, mutations in both PI3K pathway members and TP53 were evenly distributed between Rb IHC-positive and -negative tumors (Supplementary Fig. S2D). BRCA mutations were significantly more frequent in Rb IHC-positive tumors, but the number of BRCA-mutant tumors was low, and the significance of this association is unclear (Supplementary Fig. S2D).

Preclinical studies have identified several links between Rb and tumor cell immunogenicity. On the one hand, Rb hypophosphorylation by CDK4/6 inhibitors can induce a senescent-like state in luminal breast cancers, eliciting an interferon-driven gene expression profile and enhancing tumor cell immunogenicity (20). On the other hand, hyperphosphorylated Rb has also been shown to enhance tumor cell immunogenicity through inhibition of both CD274 (encoding for PDL1) transcription and NF-κB activity (21). We observed a significant increase in CD274 transcript levels in tumors with high versus low RB1 mRNA levels (Supplementary Fig. S2E). However, Rb status by IHC was not associated with sTIL infiltration or PDL1 expression on tumor cells or IC (Supplementary Fig. S2F). Finally, neither Rb IHC status nor RB1 alteration status was significantly associated with OS measured from the time of diagnosis of metastatic TNBC (Supplementary Fig. S2G and S2H).

As expected, the top upregulated gene in AR IHC-positive tumors was AR itself, and in keeping with this finding, we observed positive correlations between AR protein and mRNA levels (Fig. 2A–C). At the transcriptomic level, AR-positive cancers were more likely to be luminal or HER2-enriched by PAM50 classification and luminal AR on TNBC subtyping and showed significantly increased transcript levels for FOXA1, SPDEF, and GATA3, genes encoding for transcription factors associated with a luminal phenotype (Fig. 2A, D, and E; Supplementary Table S4). Consistent with this, several AR target genes (e.g., TFF3 and TSPAN1) were upregulated in AR-positive tumors, and basal cytokeratins (e.g., KRT5) were downregulated. We also observed significant downregulation of gene expression signatures related to genomic instability (i.e., “BRCAness” and “BC p53”) in AR-positive tumors (Fig. 2A; Supplementary Table S4).

Notably, when clustering all tumors in our cohort by gene expression patterns, the LAR tumors fell into two separate groups: LAR-CL1 and LAR-CL2 (Fig. 2F). Although both LAR-CL1 and LAR-CL2 tumors showed significantly higher levels of AR expression compared with non-LAR tumors, at a global level their transcriptomes were distinct, with LAR-CL1 tumors clustering with mesenchymal-type TNBC (and often classified as “HER2-enriched” by PAM50), and LAR-CL2 tumors clustering with basal-type TNBC (both basal-like immune-activated and basal-like immune-suppressed; Fig. 2F). Notably, AR, FOXA1, SPDEF, and GATA3 mRNA levels were significantly higher in LAR-CL1 than LAR-CL2 (Fig. 2G; Supplementary Fig. S3A–S3C). Of note, AR IHC-positive tumors as a group showed increased expression of ERBB2 (Fig. 2H), and on closer inspection, this seemed to be driven exclusively by the LAR-CL1 tumors (Fig. 2F and I).

Furthermore, although no significant difference was observed in RB1 or CD274 mRNA levels between AR IHC-positive and AR IHC-negative cases (Supplementary Fig. S3D and S3E), we did observe significantly higher expression of RB1 and CD274 (PDL1) in LAR-CL1 tumors compared with LAR-CL2 tumors (Supplementary Fig. S3F). Differential gene expression analyses between the two LAR clusters also showed an upregulation of signatures/genes related to adaptive immunity and T-cell presence/response (i.e., PDL2, PD1, and cytotoxicity/CD8⁺ T cells) in LAR-CL1 (Supplementary Table S5).

When comparing the immune profiles of AR-positive and AR-negative cases, we found no major differences in sTIL, PDL1 positivity rates by IHC (Supplementary Fig. S3G), or gene expression except for significantly higher TIGIT expression in AR-positive cases (Fig. 2J). At the genomic level, we detected the expected enrichment of alterations in PI3K pathway members in AR-positive tumors but no differences in the rates of alterations in BRCA genes, TP53, or RB1 (Supplementary Fig. S3H). Finally, OS from the time of diagnosis with metastatic disease did not differ significantly between patients with AR IHC-positive and AR IHC-negative tumors (Supplementary Fig. S3I).

Anti-PD1/PDL1 therapies (pembrolizumab and atezolizumab) have been shown to have efficacy in the treatment of certain early- and advanced-stage TNBC. In the case of advanced disease, such therapies are selectively effective in “PDL1-positive” cancers (22–24). A variety of PDL1 antibody–based IHC assays are used to define PDL1 positivity, and each uses a different antibody clone and measures PDL1 protein expression on different cell types (e.g., IC alone vs. all cells).

We assessed PDL1 protein expression by IHC using a nonclinical assay (405.9A11 antibody) and classified tumors as PDL1 positive based on PDL1 expression in tumor cell (TC) or IC (PDL1-positive TC or PDL1-positive IC). In each case, positivity was defined as ≥1% of the relevant cells staining positive for PDL1. There was a nonsignificant (P = 0.056) positive correlation between the percentage of PDL1-positive TC and IC, although several tumors showed a distinctly higher fraction of one cell type or the other expressing PDL1 protein (Fig. 3A). Both PDL1-positive TC and PDL1-positive IC tumors showed increased CD274 expression, and sTIL infiltration compared with their PDL1-negative counterparts (Supplementary Fig. S4A–S4F). To further compare tumors classified as PDL1 positive by TC or IC staining, we compared their gene expression profiles as measured by the NanoString BC360 assay (Fig. 3B). We observed very high concordance in gene expression patterns in the two groups when looking at all of the 758 genes included in the NanoString BC360 panel.

In keeping with the notion that PDL1 positivity by IHC connotes a “warm” tumor microenvironment, PDL1-positive tumors demonstrated marked upregulation of numerous immune gene signatures including the NanoString TIS, signatures pertaining to the presence of cytotoxic cells and cytotoxicity, signatures relating to antigen presentation, and signatures of macrophage infiltration (Supplementary Fig. S4G and S4H; Supplementary Table S6). Notably, PDL1 status by IHC was not correlated with (i) PAM50 subtype, (ii) TNBC subtype, or (iii) genomic alterations in the PI3K pathway, BRCA genes, or TP53 or RB1 (Fig. 3C–E). Finally, although the OS from the time of diagnosis with advanced TNBC was not significantly different between PDL1-positive and PDL1-negative cases (as defined by staining on TC), we observed a trend toward longer OS in patients with PDL1-positive tumors (Supplementary Fig. S4I).

Our cohort was heterogeneous, consisting of a mix of primary and metastatic lesions. In order to determine if there were differences between the primary and metastatic tumor samples, we compared RB1, AR, PDL1 biomarkers, TNBC and PAM50 subtypes, and mutational profiles. We did not observe any significant differences between the primary and metastatic tumors (Supplementary Fig. S5A–S5E). Although this does not exclude biological differences between the groups, this result does suggest that this entire cohort could be treated as one group.

Currently, there is a focus on “HER2-low” breast cancer, defined as cancers that are 1+ or 2+ for HER2 on IHC but without amplification of HER2 by ISH (25). This interest is driven by the evolving use of HER2-targeted antibody-drug conjugates (26, 27). We compared the phenotypic profiles of HER2-low and HER2-zero (0 score on HER2 IHC) TNBC. Overall, we did not find any significant differences between HER2-low and HER2-zero cases across several parameters. These include ERRB2 expression (Supplementary Fig. S6A), Rb or AR IHC positivity (Supplementary Fig. S6B), and PDL1 IHC positivity (Supplementary Fig. S6C). Additionally, there were no differences in sTIL percentage (Supplementary Fig. S6D), PAM50 and TNBC subtypes (Supplementary Fig. S6E and S6F), and the frequency of genomic alterations in the PI3K pathway members, BRCA genes, TP53, or RB1 (Supplementary Fig. S6G). Overall survival (OS) also showed no significant differences (Supplementary Fig. S6H).

We next determined whether genomic alterations in key genes/pathways assessed in our cohort were associated with other phenotypic parameters beyond those described above. Key findings included a positive association between genetic alterations in PI3K pathway members (PIK3CA, PTEN, AKT1, AKT2, and AKT3) and (i) classification of tumors as HER2-enriched or luminal (PAM50) or LAR or mesenchymal (TNBC subtype) and (ii) expression of AR and RB1 mRNA (Fig. 4A–D). In addition, TP53 mutations were more common in basal tumors (Fig. 4B). There were no differences in tumor subtype distribution between BRCA1/2 WT or mutant tumors or between RB1 WT and mutant tumors. PDL1 IHC positivity rates did not differ with alteration status in any of the genes assessed (Fig. 4E and F). Full gene expression data for every tumor and matched genomic information are provided in Supplementary Table S2.

We next evaluated associations with OS (defined as the time from diagnosis of advanced TNBC to death) in 94 patients from our TNBC cohort. At the time of analysis, 85% of the patients had died and 15% remained alive. Notably, OS times ranged from a few months to more than 9 years, with a median OS in the entire cohort of 517 days (Fig. 5A).

One notable finding from this analysis was a prolongation of OS in patients whose tumors, although TNBC by histology, classified as HER2-enriched on PAM50 subtyping (Fig. 5B). These HER2-enriched tumors (compared with basal tumors) showed higher ERBB2 gene expression, increased expression of AR, enrichment for the LAR subtype (Supplementary Fig. S7A–S7C), and a trend toward more PI3K pathway gene alterations, but no significant differences in the expression of RB1 or CD274 or in rates of PDL1 positivity or sTIL infiltration (Supplementary Fig. S7D–S7G). We also observed insignificant differences in OS between TNBC subtypes (Supplementary Fig. S7H), but there was a trend with basal-like immune-suppressed and mesenchymal tumors associated with a numerically shorter OS.

To identify other molecular differences in TNBC associated with OS, we classified patients in our cohort into good- and poor-outcome groups and performed differential gene expression/gene signature analysis between these groups. We observed similar trends whether classifying patients as (i) above versus below median OS restricting the analysis to patients who were deceased only (Fig. 5C), (ii) above versus below median OS for all patients (Supplementary Fig. S7I); or (iii) good- (>4-year OS) versus poor-prognosis groups (<1-year OS) (Supplementary Fig. S7J). In each case, there was a trend toward increased expression of immune genes (e.g., IFNG, IDO1, and CD274), improved outcome, and increased expression of signatures suggesting genomic instability (HRD and BRCAness) and worse outcome. We also observed a consistent increase in CDK4 gene expression in poor-outcome groups, the significance of which is unclear. Importantly, we acknowledge that outcomes are influenced not only by tumor characteristics but also by therapy, and as such, we have included information about lines of treatment received for metastatic disease for each patient (treatment and duration for each line) in Supplementary Table S2.

In the last decade, significant progress has been made in understanding TNBC biology. Preclinical and clinical research have identified a small number of predictive biomarkers that guide therapeutic decision-making in patients with TNBC, including PDL1 protein expression to guide use of anti-PD1 and anti-PDL1 antibodies in the metastatic setting, BRCA1/2 sequencing to identify patients suitable for PARP inhibitor therapy, and HER2 1+ or 2+ staining by IHC to identify patients with metastatic breast cancer who are candidates for trastuzumab deruxtecan. Despite these options, however, chemotherapy remains a mainstay of treatment.

The complexity of TNBC raises the need for an integrated approach for assessing biomarkers to assign appropriate targeted therapies to patients with TNBC and to understand overlap in biomarker positivity in considering sequencing strategies in patients with metastatic breast cancer. Thus, our study was devised to characterize the cross-association of TNBC biomarkers and to uncover novel features of TNBC biology. Moreover, by providing individual case-level data—for TILs, protein expression (AR/Rb/PDL1), the transcriptome (expression of 758 genes by NanoString), genomic alteration status (for PI3K pathway genes, BRCA genes, RB1, and TP53), and OS—we provide a rich data resource for investigators wishing to explore new aspects of TNBC biology. One strength of our study in this regard is that all our analyses (except for DNA sequencing in some cases) were performed on the same tumor specimen, using serial sections for pathology and RNA extraction. This increases the likelihood that associations reported are reflective of intrinsic TNBC biology and not artifacts created by comparing different features from one patient in tissue specimens separated by space or time.

One notable finding from our study relates to the assessment of Rb functionality in TNBC. First, we found a significant correlation between RB1 mRNA and protein levels; however, these did not correlate with RB1 mutation status. Moreover, both mRNA and protein levels for Rb were significantly correlated with a marker of Rb functionality (reduced CDKN2A expression), whereas RB1 genomic alterations were not. This is consistent with early preclinical studies of CDK4/6 inhibition in breast cancer (14) and might be attributed to the facts that (i) RB1 mutations present on only one allele might not be associated with significant Rb dysfunction, and (ii) Rb function could be compromised through nongenetic mechanisms such as epigenetic silencing. In clinical studies, inferences about Rb function are typically made using DNA sequencing alone, and our results suggest that Rb function might be better assessed by Rb IHC and/or transcriptomic profiling. This in turn is important given that CDK4/6 inhibitors, which are in development as a possible therapy for TNBC, are only likely to exert direct antiproliferative effects in TC if Rb is functional, making an accurate assessment of Rb function critical (28).

Our comparisons between AR-positive and AR-negative tumors revealed several expected findings, including increased expression of AR, AR target genes, and luminal transcription factors and keratins, as well as enrichment for mutations in PI3K pathway genes in AR-positive tumors. Importantly, we found no association between AR status and immune markers such as TIL or PDL1 positivity. In clinical trials, AR positivity has been used as a biomarker to identify TNBC that might be responsive to AR antagonist therapy. These trials have consistently yielded low clinical benefit rates, and attention has recently shifted to identifying transcriptomic signatures that might better predict androgen dependence in these tumors. In our study, we found that the majority of AR-positive tumors were classified as LAR by transcriptomic profiling. Interestingly, these tumors fell into two distinct subgroups. Although both LAR subgroups showed higher AR expression than non-LAR tumors, their transcriptomes were otherwise distinct: tumors in the first group (LAR-CL1) clustered together with mesenchymal TNBC, showed higher levels of AR and ERBB2 expression, and were often classified as “HER2-enriched” by PAM50; tumors in the second group (LAR-CL2) clustered more closely to basal TNBC and were more commonly denoted as “luminal B” by PAM50. Critically, our sample size, especially in the case of LAR-cluster 2, is small and thus further investigation is warranted to determine if there is significant heterogeneity within LAR tumors which might have therapeutic implications.

The expression of PDL1 has been recognized as an important biomarker for predicting response to immunotherapy in advanced TNBC (22–24). In the case of pembrolizumab, a companion IHC diagnostic assay using the 22C3 anti-PDL1 antibody is used to identify PDL1-positive tumors, with positivity defined as a combined positive score (the number of PDL1-staining TC, lymphocytes, and macrophages, divided by the total number of viable TC, multiplied by 100) of ≥10. Although our study utilized a different antibody than those approved for companion diagnostics, it shows that tumors identified as PDL1 positive by protein expression on IC show very similar phenotypic profiles to those identified as defined by PDL1 expression on TC. We also observed a longer OS in patients with PDL1-positive tumors but are unable to determine from our cohort whether PDL1 is acting as a general prognostic biomarker or rather a predictive marker for response to immunotherapy, as we do not have treatment information available. We focused our DNA sequencing analysis on genes commonly mutated in TNBC and/or relevant to the selection of contemporary therapies (i.e., PI3K pathway genes, BRCA genes, TP53, and RB1; refs. 2, 29–32). For the most part, our results have further confirmed previously reported findings, such as an enrichment of PI3K pathway genes in luminal tumors (29). Notwithstanding this observation, perhaps more striking is the lack of correlation between alterations in these genes and PAM50 subtype, TNBC subtype, TILs, PDL1 status, and OS.

An important aspect of our study is the availability of OS data and treatment information (including details of lines of therapy received for metastatic disease, with duration of treatment for each line) for most patients. As expected, we consistently observed upregulation of immune genes/signatures among patients with longer OS and upregulation of signatures associated with genomic instability in patients with poorer outcomes. In addition, we found a significantly longer OS among the small number of patients whose tumors were classified as “HER2-enriched” by PAM50. This is, to our knowledge, a new observation and might reflect an inherent favorable biology in these tumors.

It is also noteworthy that our analyses comparing tumors classified as HER2-low and HER2-zero did not reveal any significant differences regarding tumor immune profile, PAM50 or TNBC subtype, genomic features, or OS. This is consistent with a growing body of evidence that HER2-low breast cancer is not a distinct biological entity per se (26, 27).

Our study has several limitations: (i) the sample size is small—with 95 patients total, certain subgroups are small, which limits the power of intergroup comparison; (ii) TNBC analyzed here were from patients undergoing prescreening for a clinical trial (requiring a certain performance status and level of organ function) in a large academic cancer center and may thus not be representative of a TNBC population seen in community practice. A significant proportion of our cohort shows OS of greater than 36 months, with six patients having OS of greater than 60 months; (iii) DNA sequencing was not done for all tumors and, in some cases, was done on a different tumor specimen to all other analyses; and (iv) the PDL1 antibody used, although well validated, was not used in clinical practice.

In summary, the high heterogeneity of TNBC poses a major hurdle to current treatments, thus the urgency to understand it better. Our study offers a cohort of 95 TNBC tumors that we assessed using multimarker-based profiling as well as survival information for the whole cohort. We show that RB1 mRNA and Rb protein expression are better correlated with downstream Rb function (i.e., CDKN2A expression) than RB1 mutational status, suggesting that RB1 mutation may not be the best guide to Rb function. Furthermore, LAR tumors cluster into two major groups, with both having a high AR expression but with transcriptomic profiles otherwise resembling basal or mesenchymal tumors. We also observed similar transcriptomic profiles between the c immune and TC. As for patients’ survival, we observed that the two major PAM50 subtypes had significantly different OS as well as a trend for higher immune-related gene expression in patients with better OS. This study is an excellent resource, as it provides a significant amount of usable data to both scientific and clinical communities. This study’s findings provide new insights into the utility of different markers and suggest alternative ways to address TNBC heterogeneity.

S.E. Church reports other support from NanoString Technologies outside the submitted work. K. North reports employment with NanoString Technologies Inc when completing this work. E.T. Richardson reports personal fees and nonfinancial support from Merck & Co, Inc and grants from AstraZeneca outside the submitted work. V. Attaya reports personal fees from Olema Oncology outside the submitted work. N.U. Lin reports grants from Genentech, Zion Pharmaceuticals, and Merck; grants and personal fees from Pfizer/SeaGen, AstraZeneca, and Olema Pharmaceuticals; and personal fees from Stemline/Menarini, Artera Inc, Daiichi Sankyo, Blueprint Medicines, and Janssen outside the submitted work. E.A. Mittendorf reports other support from AstraZeneca, BioNTech, Merck, Moderna, Merck Sharp & Dohme, Gilead, and American Society of Clinical Oncology; nonfinancial support from BMS and Roche/Genentech; and grants from Roche/Genentech and Susan G. Komen outside the submitted work. S.M. Tolaney reports grants from Eli Lilly during the conduct of the study as well as grants and personal fees from Genentech/Roche, Merck, Pfizer, Novartis, Bristol Myers Squibb, Eisai, AstraZeneca, Gilead, Seattle Genetics, and Jazz Pharmaceuticals; grants from Exelixis, NanoString Technologies, and OncoPep; and personal fees from Eli Lilly, Sanofi, CytomX Therapeutics, Daiichi Sankyo, OncXerna, Zymeworks, Zentalis, Blueprint Medicines, Reveal Genomics, ARC Therapeutics, Infinity Therapeutics, Sumitovant Biopharma, Umoja Biopharma, Artios Pharma, Menarini/Stemline, Aadi Bio, Bayer, Incyte Corp, Natera, Tango Therapeutics, Systimmune, eFFECTOR, Hengrui USA, Cullinan Oncology, Circle Pharma, Arvinas, BioNTech, and Johnson & Johnson outside the submitted work. S. Goel reports grants from Eli Lilly during the conduct of the study as well as grants from Gi Therapeutics and Incyclix Bio and personal fees from Pfizer, Novartis, Regor Pharmaceuticals, and Beigene outside the submitted work. No disclosures were reported by the other authors.

B. Jovanovic: Formal analysis, visualization, methodology, writing–original draft, project administration, writing–review and editing. S.E. Church: Software, formal analysis, methodology, writing–review and editing. K.M. Gorman: Data curation, software, formal analysis, validation, writing–review and editing. K. North: Software, formal analysis, methodology, writing–review and editing. E.T. Richardson: Investigation, writing–review and editing. M. DiLullo: Data curation, validation, investigation, writing–review and editing. V. Attaya: Data curation, validation, investigation, writing–review and editing. J. Kasparian: Data curation, validation, investigation, writing–review and editing. A. Mohammed-Abreu: Data curation, validation, investigation, writing–review and editing. G. Kirkner: Data curation, validation, writing–review and editing. M.E. Hughes: Resources, data curation, validation, writing–review and editing. N.U. Lin: Resources, data curation, validation, writing–review and editing. E.A. Mittendorf: Conceptualization, resources, writing–review and editing. S.J. Schnitt: Investigation, writing–review and editing. S.M. Tolaney: Conceptualization, supervision, funding acquisition, visualization, writing–original draft, project administration, writing–review and editing. S. Goel: Conceptualization, formal analysis, supervision, funding acquisition, visualization, methodology, writing–original draft, project administration, writing–review and editing.

This work was funded by Eli Lilly and the Mehlman Family Fund. S. Goel is supported by a Snow Fellowship (SF2020-47) from the Snow Medical Research Foundation. E.A. Mittendorf is supported by the Rob and Karen Hale Distinguished Chair in Surgical Oncology. We would like to acknowledge the National Comprehensive Cancer Network Oncology Research Program (in collaboration with Pfizer Independent Grants for Learning & Change), Pan-Mass Challenge Fund for Breast Oncology, Fashion Footwear Association of New York, and the Cross Family Fund, for help with data collection, patient screening, and collection of tissue.