A Transcriptionally Definable Subgroup of Triple-Negative Breast and Ovarian Cancer Samples Shows Sensitivity to HSP90 Inhibition

Technical advances in DNA sequencing and molecular profiling have ushered in an age of “precision medicine,” where the genetics of a patient's tumor can be accurately interrogated to identify therapeutically targetable alterations. Tumor-targeted therapeutics are often selected on the basis of somatic DNA alterations in a drug target, such as anti-HER2 therapies for HER2/ERBB2-amplified breast cancer, and BRAF inhibitors for BRAF^V600-mutant lung cancers and melanomas. In such cases, drug development strategies have been straightforward. However, there are many cancer subtypes for which targeting individual somatic genetic alterations has not translated into therapeutic success, including triple-negative breast cancer [TNBC, which lacks expression of estrogen receptor α (ER), progesterone receptor (PR), and HER2] and ovarian cancer. TNBC comprises 10%–15% of breast cancer cases, and approximately one-third of the patients (∼50,000–75,000/year worldwide) treated for early-stage TNBC later experience recurrence (1, 2). Most of the approximately 239,000 new ovarian cancer cases/year worldwide ultimately recur, and ovarian cancer has the highest mortality rate among female cancers (3). The current standard treatments for both diseases in the primary and metastatic settings include anthracyclines, platinum agents, and microtubule-targeted chemotherapies (4, 5), highlighting a need to discover new tumor-targeted therapeutic opportunities in these cancer subtypes. Deficiencies in homologous recombination–mediated DNA repair, such as those caused by germline mutations in BRCA1 or BRCA2 in approximately 10% of TNBC and ovarian cancer, sensitize cancer cells to PARP inhibitors (6, 7), which are now approved for these cancers in patients with germline BRCA1/2 alterations. However, advanced TNBC and ovarian cancer typically develop resistance to all approved therapies.

Two obstacles to the development of effective tumor-targeted therapies for TNBC and ovarian cancer have been (i) heterogeneity within cancer subtypes, and (ii) intrinsic drug resistance. TNBC and ovarian cancer may contain as many as six and four molecular subtypes, respectively (8–10). Thus, development of a broadly effective “pan-TNBC therapy” or “pan-ovarian cancer therapy” is unlikely, and dissection of oncogenic pathways within subgroups of TNBC and ovarian cancer to identify therapeutic targets is warranted. Breast cancer and ovarian cancer have been shown to have similar (epi)genetic and transcriptional profiles (11, 12), which led us to hypothesize that analyzing these two cancer types as a single cohort may reveal novel molecularly identifiable mixed subgroups that are uniquely sensitive to certain drugs.

Robust multi-array average–normalized gene expression data for 1,074 cancer cell lines were downloaded from the Cancer Cell Line Encyclopedia (CCLE), and for 623 cancer cell lines from the Genomics of Drug Sensitivity in Cancer (GDSC) database (13, 14). Morpheus Software (Broad Institute) was used to collapse gene expression data to one probe set per gene using a maximum-mean collapsing strategy (15).

Level 4 normalized expression data from reverse-phase protein arrays (RPPA) for 452 (phospho)proteins across 651 cell lines were downloaded from the MD Anderson Cell Lines Project (MCLP), and were filtered manually using a complete-case-analysis approach (16).

Hierarchical clustering (Euclidean distance) of gene and (phospho)protein expression profiles from breast cancer and ovarian cancer cell lines was performed using package “gplots”, and heatmaps and dendrograms were generated with R software (17). We identified two mixed subgroups containing primarily TNBC and ovarian cancer cell lines, termed BR/OV-1 and -2 (Fig. 1A).

Breast cancer and ovarian cancer cell lines (Supplementary Table S1) were assigned to the BR/OV-1 or -2 subgroup based on CCLE gene expression data (Fig. 1A). Differentially expressed genes between BR/OV-1 versus -2 cell lines in the CCLE dataset were used to generate a BR/OV-1/2 gene expression classifier using two-sided t test and Bonferroni multiple-comparison adjusted P ≤ 0.05. The t statistic was calculated using the formula t = (⁠|\bar{x}$|_M - |\bar{x}$|_B)/s, where |\bar{x}$| and s are mean and SD, respectively. The classifier was applied to GDSC gene expression data, and clustering of cell lines as BR/OV-1 and -2 was validated.

Support vector machine (SVM) regression (SVR) was used to classify cell lines as BR/OV-1 or -2 in the GDSC gene expression datasets using genes from the BR/OV-1/2 classifier as features. One hundred iterations of Monte Carlo cross-validation were implemented to evaluate model performance: half of cell lines were randomly selected to train the classifier, which was then used to predict BR/OV-1 or -2 status in the remaining cell lines. After cross-validation, model accuracy was evaluated by calculating the AUC ROC. One hundred iterations of Monte Carlo cross-validation were then performed 10,000 times using cell lines randomly assigned to BR/OV-1 or -2 subgroups to generate a P value.

Breast cancer and ovarian cancer cell lines in GDSC were classified as BR/OV-1 or -2 based on CCLE gene expression data. Sensitivity of BR/OV-1 versus -2 cell lines [n = 24; ln(IC₅₀)] was compared in response to 99 drugs (GDSC v5) to generate P values by two-sided t test adjusted for multiple comparisons using the Benjamini–Hochberg FDR procedure. As a post hoc validation analysis, sensitivity of BR/OV-1 versus -2 cell lines (n = 36; AUC) was compared in response to 211 drugs (GDSC v7).

These studies were approved by the Dartmouth College Institutional Animal Care and Use Committee (IACUC) and performed on-site. Luciferase-expressing ovarian cancer cells (10⁶) were injected intraperitoneally into 4-week-old female NOD-scid/IL2Rγ^−/− (NSG) mice (obtained from Norris Cotton Cancer Center Mouse Modeling Shared Resource). Tumor burden was serially measured by bioluminescence imaging. When bioluminescent flux reached ≥10³ photons/second, mice were randomized to twice weekly treatment by intraperitoneal injection of AT13387 (70 mg/kg; provided by Astex Therapeutics, Ltd.) or vehicle (n = 6–8 mice/group). For molecular analyses, COV644 tumors were harvested 24 hours after drug treatment on day 36, and were snap-frozen for immunoblot analysis, or formalin-fixed and paraffin-embedded for histologic analyses. Because of near-complete tumor regression in most AT13387-treated mice bearing JHOC5 xenografts, half of the vehicle-treated mice at day 36 were treated with AT13387 for 24 hours prior to harvesting of tumors to provide specimens from vehicle- and AT13387-treated mice.

Expression levels of the 275 genes in the BR/OV-1/2 gene expression classifier were extracted from whole-transcriptome RNA sequencing data (complete dataset being reported elsewhere) obtained from 88 breast cancer patient-derived xenograft (PDX) models from 20 patients (18). Z-scores were generated for each gene across the 88 models and weighted on the basis of the gene expression fold-change from the classifier. The BR/OV-1/2 subgroup score for each PDX model represents the difference between the average weighted z-score for BR/OV-2 genes versus the average weighted z-score for BR/OV-1 genes.

These studies were approved by the Mayo Clinic IACUC and performed on-site. PDX tumor fragments were subcutaneously implanted into the flanks of 6- to 8-week-old female NOD-scid (NOD.CB17-Prkdc^scid/NCrCrl) mice (Charles River Laboratories). Tumor dimensions were measured twice weekly using calipers (volume = length × width²/2). When average tumor volume reached approximately 200 mm³, mice were randomized to treatment with AT13387 or vehicle (n = 9–13 mice/group). Tumors were harvested after 4–8 weeks of treatment.

For comparisons of cell growth, ln(IC₅₀), AUC, mitochondrial depolarization, and gene and protein expression levels, data were analyzed by two-tailed t test. For IHC and terminal deoxynucleotidyl transferase–mediated dUTP nick end labeling (TUNEL) assays, data were analyzed by ANOVA with Bonferroni multiple comparison-adjusted post hoc testing between groups. To estimate progression/regression of tumors, the following linear mixed model was employed: log₁₀(tumor volume_it) = a_i + b × t + e_it, where i represents the i^th mouse, t represents the time of tumor volume measurement, a_i represents the mouse-specific log tumor volume at baseline (t = 0), slope b represents rate of tumor volume growth (or reduction), and e_it represents deviation of measurements from the model over time (19). The variance of a_i is interpreted as mouse heterogeneity, and b*log_e(10)*100 estimates percent tumor volume increase per week. The computation was carried out in R using function “lme” from library “nlme.” Treatment groups were compared using Z-test for slopes with SE derived from lme.

Additional details are provided in Supplementary Materials and Methods.

An overview of our experimental approach is presented in Supplementary Fig. S1. We first performed hierarchical clustering of gene expression profiles from 63 breast cancer cell lines from the CCLE (13). Breast cancer cell lines mainly clustered by receptor status: receptor-positive lines (ER⁺, HER2⁺, and luminal androgen receptor–positive TNBC) clustered separately from receptor-negative TNBC lines (Supplementary Fig. S2A). Similar analysis of 47 ovarian cancer cell lines revealed only loose clustering by histologic subtype (e.g., clear cell and serous; Supplementary Fig. S2B). Integrated hierarchical clustering of both breast cancer and ovarian cancer cell lines as a single cohort revealed two novel mixed clusters containing predominantly ovarian cancer and receptor-negative TNBC; these subgroups were termed BR/OV-1 and -2 (Fig. 1A). The strength of the clusters was confirmed using the Partitioning Around Medoids algorithm (k = 4; Rand Index = 0.79; Supplementary Fig. S3). Clusters were further verified by multiscale bootstrap resampling (P = 0.10).

Drug sensitivity (IC₅₀) for 99 compounds was compared between subgroups of breast cancer and ovarian cancer cell lines in the GDSC database v5 (ref. 14; Fig. 1B; Supplementary Table S2). Because mutations in BRCA1 or BRCA2 are associated with sensitization to agents targeting DNA repair (e.g., PARP inhibitors; ref. 7), and BRCA1/2 mutations are more frequent among BR/OV-2 cell lines (Fig. 1A), BRCA1/2-mutant cell lines were excluded from these analyses to focus on cancer subgroups lacking known targetable alterations. We assessed sensitivity of 13 BR/OV-1 cell lines and 11 BR/OV-2 cell lines to the 99-compound panel. Among the top eight drugs with significantly different ln(IC₅₀) values between BR/OV-1 versus -2 cells, two Hsp90 inhibitors (Hsp90i; CCT018159 and 17-AAG) were more effective against BR/OV-2 cells (t test, P = 0.0007 and P = 0.01, respectively; FDR-adjusted P = 0.067 and P = 0.149, respectively; Fig. 1C and D; Supplementary Table S2). However, a third Hsp90i (AUY922) was not significantly more effective in either subgroup, possibly due to differences in cellular sensitivity to structurally distinct HSP90i across cell lines (20). These findings were confirmed using AUC sensitivity data for CCT018159 and 17-AAG in the updated v7 of GDSC with 211 drugs tested in 19 BR/OV-1 and 17 BR/OV-2 cell lines (P = 0.002 and P = 0.007, respectively; Supplementary Fig. S4). Given the correlation between CCT018159 and 17-AAG drug sensitivity profiles across BR/OV-1/2 cell lines (Fig. 1E), the >20-year history of clinical data demonstrating limited efficacy of Hsp90i in unselected populations of patients with cancer (21, 22), and the lack of a biomarker to identify tumors likely to respond to Hsp90i, we focused on the therapeutic potential of Hsp90i for the BR/OV-2 subgroup of breast cancer/ovarian cancer in functional validation studies.

To test the specificity of the Hsp90i sensitivity phenotype, the ln(IC₅₀) values of BR/OV-1 and -2 cell lines were compared with those of all other histologic cancer subtypes represented in GDSC. BR/OV-2 cells were more sensitive to CCT018159 than any other established cancer subtype (Fig. 1F), making the BR/OV-2 subgroup of breast cancer/ovarian cancer an attractive candidate for the development of Hsp90-directed therapeutics.

Genes differentially expressed (P_adj ≤ 0.05) between BR/OV-1 cell lines (10 breast cancer and 19 ovarian cancer) and BR/OV-2 cell lines (seven breast cancer and 22 ovarian cancer) were determined using transcriptomic data from CCLE, which yielded a 275-gene classifier (Fig. 2A; Supplementary Table S3). Gene Ontology (GO) classification of these genes indicated enrichment for growth/proliferation-related phenotypes in the BR/OV-2 subgroup, and enrichment for epithelial phenotypes (e.g., cell–cell adhesion) in the BR/OV-1 subgroup (Supplementary Table S4). Prior studies identified mesenchymal-like and epithelial/basal-like subtypes of TNBC and ovarian cancer based on gene expression profiles (8, 9, 23–25). Lehmann and colleagues assigned six of seven BR/OV-2 TNBC cell lines to mesenchymal or mesenchymal stem–like subtypes (8). However, a reanalysis using the TNBCtype tool (26) on the same breast cancer gene expression datasets, and using CCLE and GDSC gene expression datasets, did not indicate significant enrichment for mesenchymal-like TNBC cell lines in the BR/OV-2 subgroup (Supplementary Table S5). Similarly, we found only minimal overlap between BR/OV-2–upregulated genes and published mesenchymal gene signatures (Supplementary Table S6). Furthermore, a prior report classified 10 of 20 ovarian cancer cell lines of the BR/OV-2 subgroup as mesenchymal-like, while the remaining 10 cell lines were classified as epithelial or stem-like (9). Thus, the BR/OV-1 and -2 subgroups represent unique mixed groups within breast cancer and ovarian cancer.

The ability of the BR/OV-1/2 classifier (Supplementary Table S3) to create two mixed subgroups of breast cancer/ovarian cancer was tested using clustering of independent gene expression data from GDSC (14). On the basis of the BR/OV-1/2 assignments from Fig. 1A (from CCLE data), one GDSC cluster contained 12 BR/OV-1 cell lines, and the other cluster contained 11 BR/OV-2 cell lines and one BR/OV-1 line, demonstrating that the classifier distinguished BR/OV-1 from -2 cell lines with 96% (23/24) accuracy in an independent gene expression dataset (Fig. 2B). We further evaluated the robustness of this classifier in a supervised setting using SVR to predict BR/OV-1/2 status using expression of genes in the classifier as features. After Monte Carlo cross-validation, we found that the classifier could distinguish BR/OV-1 from -2 cell lines with high accuracy in the GDSC dataset (AUC = 0.87; P = 5.51 × 10⁻⁸⁹; Supplementary Fig. S5).

To confirm that these breast cancer/ovarian cancer subgroups also differed at the level of (phospho)protein expression, hierarchical clustering was performed for BR/OV-1 and -2 cell lines (n = 30) using RPPA data from the MCLP dataset (16). Clustering confirmed classification of cell lines into BR/OV-1 and -2 subgroups with 100% accuracy (Fig. 2C; Supplementary Fig. S6; Supplementary Table S7). Tandem Mass Tag (TMT)–based mass spectrometry was also performed to quantitatively profile the proteomes of three BR/OV-1 and three BR/OV-2 TNBC cell lines. Analysis revealed 588 proteins significantly differentially expressed between BR/OV-1 versus -2 cells (P ≤ 0.05; Supplementary Fig. S7A and S7B; Supplementary Table S8). Differentially expressed proteins were classified by GO, revealing functions classically associated with epithelial (e.g., epidermis and ectoderm development) or mesenchymal (e.g., TGFβ and bone morphogenetic protein signaling) phenotypes in BR/OV-1 or BR/OV-2 cells, respectively (Supplementary Fig. S7C and S7D; Supplementary Table S9).

Hsp90i sensitivity was tested in growth experiments using three TNBC and three ovarian cancer cell lines from each of the BR/OV-1 and -2 subgroups (total of 12 cell lines). BR/OV-2 cells showed overall increased sensitivity to CCT018159 compared with BR/OV-1 cells (⁠|${P_{{\rm{ln}}( {{\rm{I}}{{\rm{C}}_{{\rm{5}}0}}} )}}$| = 0.003; Fig. 3A; Supplementary Fig. S8). The same differential sensitivity was apparent when analyzing cell lines from a single tissue of origin (breast or ovary), indicating that sensitivity to CCT018159 is associated with BR/OV-1/2 status and not tissue of origin (Supplementary Fig. S9).

To confirm that differential sensitivity to Hsp90i was not compound specific, we similarly tested sensitivity of eight cell lines (two per BR/OV-1/2 subgroup and tissue of origin) to three other Hsp90i: 17-AAG, AT13387 (onalespib), and PU-H71. BR/OV-2 cells consistently showed increased sensitivity to these Hsp90i compared with BR/OV-1 cells (all |${P_{{\rm{ln}}( {{\rm{I}}{{\rm{C}}_{{\rm{5}}0}}} )}}$| < 0.01; Fig. 3B–D).

We then selected two TNBC and four ovarian cancer cell lines for which gene expression profiles were available (from CCLE), but drug sensitivity data were not (i.e., cell lines were not in GDSC v5). A BR/OV-1/2 subgroup score was generated for each cell line, classifying four of these cell lines (OVTOKO, JHOC5, MDA-MB-436, and ES2) as BR/OV-2, and two lines (COV644 and HDQ-P1) as BR/OV-1. Growth assays revealed that the cell lines classified as BR/OV-2 are more sensitive to Hsp90i than those classified as BR/OV-1, confirming the ability of the BR/OV-1/2 classifier to predict Hsp90i sensitivity in TNBC and ovarian cancer cell lines (⁠|${P_{{\rm{ln}}( {{\rm{I}}{{\rm{C}}_{{\rm{5}}0}}} )}}$| = 0.002; Fig. 3E; Supplementary Fig. S10).

To discover mechanisms underlying differential Hsp90i sensitivity between BR/OV-1 and -2 cells, four BR/OV-1 and four BR/OV-2 cell lines (balanced for breast and ovarian origin) were treated with a dose range of AT13387 (chosen for further studies for its clinical potential; ref. 27). Compensatory upregulation of Hsp70, inhibition of major signal transduction pathways including PI3K/AKT/mTOR, MEK/ERK, STAT1/3, and TGFβ, and downregulation of cell-cycle proteins, which are known effects of Hsp90i (28–32), occurred consistently in both BR/OV-1 and -2 cells in response to AT13387 (Supplementary Figs. S11 and S12). The most notable differences between BR/OV-1 and -2 cells in signaling response to AT13387 occurred at the level of apoptosis, where only BR/OV-2 cells demonstrated drug-induced increases in PARP cleavage, which coincided with upregulation of the proapoptotic BH3-only proteins Bim (particularly the potent Bim_L and Bim_S isoforms; ref. 33) and PUMA (Fig. 4A; Supplementary Fig. S12 and S13). Differential apoptotic sensitivity to Hsp90i was confirmed by Bim BH3 profiling, where AT13387 increased mitochondrial outer membrane depolarization in BR/OV-2 cells, but not BR/OV-1 cells (Fig. 4B).

RNAi-mediated knockdown of Bim, PUMA, and the combination was performed in two BR/OV-2 cell lines (CAL51 TNBC cells and JHOC5 ovarian cancer cells) to functionally test whether Bim and/or PUMA modulate Hsp90i-induced apoptosis. Interestingly, knockdown of either Bim or PUMA induced compensatory upregulation of the other (Fig. 4C), which may be because of increased availability of binding sites on antiapoptotic Bcl-2 family proteins to stabilize BH3-only proteins (34, 35). Dual Bim/PUMA knockdown effectively prevented Hsp90i-induced apoptosis in BR/OV-2 cells. These data collectively suggest that BH3-only proteins are differentially regulated between BR/OV-1 and -2 cells in response to Hsp90i, which may underlie differences in drug sensitivity between the BR/OV-1 and -2 subgroups of breast cancer and ovarian cancer.

COV644 and JHOC5 ovarian cancer cells were assigned to the BR/OV-1 and -2 subgroups, respectively, based on the BR/OV-1/2 classifier (Fig. 3E). Luciferase-labeled cells were injected intraperitoneally into mice to provide xenograft models of advanced ovarian cancer. Relative tumor burden was serially measured by bioluminescence imaging. When bioluminescent flux reached ≥10³ photons/sec, mice were randomized to treatment with AT13387 or vehicle. COV644 (BR/OV-1) tumors were completely resistant to Hsp90i, while JHOC5 (BR/OV-2) tumors were exquisitely sensitive (Fig. 5A and B; Supplementary Fig. S14), confirming the predicted responses of BR/OV-1 and -2 tumors to Hsp90i. AT13387 significantly decreased tumor cell proliferation (assessed by Ki67 IHC) and increased apoptosis (assessed by TUNEL) in JHOC5 tumors but not COV644 tumors (Fig. 5C and D; Supplementary Fig. S15). Because of small tumor volumes, tumor lysates from each treatment group were pooled for immunoblot analysis. AT13387 blocked Hsp90 activity (assessed by increased Hsp70) in both tumor models, while PARP cleavage and upregulation of Bim and PUMA occurred only in JHOC5 (BR/OV-2) tumors (Fig. 5E).

Expression profiles for the 275 genes in the BR/OV-1/2 classifier were then extracted from RNA-sequencing data from 88 breast cancer PDX models, which were generated from 20 patients enrolled in the BEAUTY study (18). A BR/OV-1/2 subgroup score was calculated for each PDX model (Supplementary Table S10), and the models with the highest score (most BR/OV-2) and the lowest score (most BR/OV-1) were selected for functional analysis (Fig. 6A). Mice bearing subcutaneous tumors were randomized to treatment with AT13387 or vehicle. Ex173957 (BR/OV-1) tumors were insensitive to AT13387, while the growth of Ex173852 (BR/OV-2) tumors was significantly suppressed by AT13387 (Fig. 6B and C; Supplementary Fig. S16). Toxicity (weight loss) was detected in seven of 10 mice bearing Ex173852 (BR/OV-2) tumors after 1 month of drug treatment (Supplementary Fig. S17), leading to a change in treatment schedule on day 36 from twice weekly to once weekly in this model that may have blunted treatment efficacy. Such toxicity was not observed in mice bearing Ex173957 (BR/OV-1) PDX tumors, likely due to shorter duration of study (Fig. 6B; Supplementary Fig. S17). In PDX models, AT13387 treatment significantly decreased tumor cell proliferation and increased apoptosis in Ex173852 (BR/OV-2) tumors but not Ex173957 (BR/OV-1) tumors (Fig. 6D and E; Supplementary Fig. S18). Immunoblot analysis of PDX lysates showed that AT13387 blocked Hsp90 activity (assessed by increased Hsp70) in both models, while PARP cleavage and upregulation of Bim and PUMA consistently occurred only in Ex173852 (BR/OV-2) tumors (Fig. 6F).

Homologous recombination deficiency (HRD), often involving mutations in BRCA1 or BRCA2, is known to increase sensitivity to DNA-damaging chemotherapy in breast cancer and ovarian cancer (36, 37). To rule out HRD-related alterations as a contributor to differences in drug sensitivity between BR/OV-1 and -2 tumors, DNA mutation and copy number data were evaluated for xenograft models. Analysis of common HRD-related genes revealed BRCA1 amplification in the JHOC5 (BR/OV-2) cell line, and PARP1 amplification in the BR/OV-1 PDX model, which are not known to be associated with phenotypes of drug sensitivity or resistance, respectively (Supplementary Fig. S19).

Finally, RNA sequencing data from TNBC (n = 136) and ovarian cancer (n = 303 serous cystadenocarcinoma) cases from The Cancer Genome Atlas (TCGA) were used to calculate a BR/OV-1/2 subgroup score for each primary tumor. A substantial proportion of TNBC and ovarian cancer tumors had relatively high scores, suggesting alignment with the BR/OV-2 subgroup (Fig. 6G); such tumors would be predicted to exhibit increased sensitivity to Hsp90i. On the basis of our finding that 29 of 41 TNBC/ovarian cancer cell lines with a score > 0 were in the BR/OV-2–like cluster in Fig. 1A, we predict that 71% (95% confidence interval, 55%–84%; exact binomial test P = 0.012) of primary human TNBC/ovarian cancer cases with a score > 0 are BR/OV-2 like; thus, approximately 49% and 35% of all TNBC and ovarian cancer cases, respectively, are predicted to be BR/OV-2 like.

We present an integrated strategy to uncover new therapeutic opportunities in cancer types with limited targetable genetic alterations, such as TNBC and ovarian cancer. To our knowledge, this is the first study to generate gene and protein expression classifiers, and integrate drug sensitivity analyses for molecularly identifiable subgroups that transcend cancer lineage. In so doing, we identified two mixed subgroups of ovarian cancer and breast cancer (consisting of mostly TNBC). We discovered that BR/OV-2 subgroup cell lines were sensitized to Hsp90i, which we validated with multiple cell lines and several Hsp90is. We developed a gene expression classifier that predicted sensitivity to Hsp90i in cultured cells, cell line–derived xenografts, and PDX models. Hsp90 thus represents a viable therapeutic opportunity in subgroups of TNBC and ovarian cancer. Furthermore, this cancer lineage–transcendent strategy may be applied more broadly to discover other novel cancer subgroups with unique therapeutic vulnerabilities.

Hsp90 is a family of ATP-dependent molecular chaperones consisting of multiple isoforms with high sequence identity that are expressed in different subcellular compartments. The two major isoforms, Hsp90α (stress-inducible, major isoform) and Hsp90β (constitutive, minor isoform), share a conserved amino-terminal ATP-binding domain targeted by most Hsp90 inhibitors (38). BR/OV-1 and -2 cells expressed similar levels of Hsp90α, while BR/OV-1 cells expressed more Hsp90β (Supplementary Fig. S13); however, an association between Hsp90 isoform levels and Hsp90i sensitivity in cancer cells has not been reported. Hsp90 chaperones are involved in the stabilization and activation of many client proteins essential for signal transduction and stress responses, including oncogenic proteins such as Src, Bcr/Abl, ErbB receptors, p53, and nuclear hormone receptors (39). As such, Hsp90 has been recognized as an important target for anti-cancer drug discovery. Most Hsp90 inhibitors have been developed to inhibit ATP binding, which decreases the affinity of Hsp90 for client proteins and results in their subsequent ubiquitination and degradation (40). Although Hsp90 is expressed in most mammalian cells, cancer cells are frequently more sensitive to Hsp90i than non-cancer cells, providing a therapeutic window and further highlighting the clinical appeal of Hsp90i (41). Several Hsp90 inhibitors have been tested clinically, but results have been disappointing with modest response rates in unselected patient populations (21, 22). Hsp90 inhibitors are being tested in ongoing clinical trials that include patients with breast cancer or ovarian cancer (e.g., NCT02474173, NCT01962948); our findings suggest that such trials evaluating broad patient populations will have mixed success due to differential Hsp90i sensitivity between cancers that are BR/OV-2-like vs. other, which could be tested through retrospective molecular analysis of tumors. Future Hsp90i clinical trials may benefit from prospectively screening for BR/OV-2-like TNBC or ovarian cancer to identify patients more likely to benefit.

In agreement with prior observations (29, 32), Hsp90i induced compensatory upregulation of Hsp70, inhibition of major signal transduction pathways (e.g., PI3K/AKT/mTOR, MEK/ERK, STAT, and TGFβ), and downregulation of cell-cycle proteins in both BR/OV-1 and -2 cells (Figs. 4A, 5E, and 6F; Supplementary Figs. S11–S13); thus, such effects do not define the mechanism of enhanced response to Hsp90i seen in BR/OV-2 cells. On the other hand, we found major differences in apoptosis, where BR/OV-2 cells and tumors showed Hsp90i-induced apoptosis, while BR/OV-1 cells and tumors did not (Figs. 4, 5D and E, and 6E and F; Supplementary Figs. S11–S13). Hsp90i-induced apoptosis coincided with upregulation of Bim and PUMA in BR/OV-2 cells but not BR/OV-1 cells, implicating differential response to Hsp90i at the level of proapoptotic BH3-only protein effectors. To confirm that these BH3-only proteins mediate response to Hsp90i, we showed that knockdown of Bim and PUMA abrogated apoptosis in BR/OV-2 cells (Fig. 4C). Bim and PUMA mediate apoptosis by (i) sequestering antiapoptotic Bcl-2 family proteins (e.g., Bcl-2 and Bcl-xL) away from Bax and Bak, which prevents Bax/Bak pore formation in the mitochondrial outer membrane (MOM) and cytochrome C release, and/or (ii) binding Bax/Bak directly in the MOM (42, 43). Mechanisms regulating Bim and PUMA expression have been partly elucidated. Bim is destabilized by ERK phosphorylation, and inhibition of MEK/ERK signaling leads to an abundance of Bim (44). PUMA is downregulated by PI3K/AKT/FOXO signaling, and PI3K/AKT inhibition leads to nuclear translocation of FOXO transcription factors to drive PUMA transcription (45, 46). Our data show differential modulation of Bim and PUMA in BR/OV-2 versus -1 cells despite consistent downregulation of MEK/ERK and PI3K/AKT upon Hsp90i, suggesting additional uncharacterized mechanisms regulating Bim and PUMA that warrant further investigation.

Our approach to identify novel multi-lineage cancer subgroups and subgroup-specific therapeutic vulnerabilities has broad implications for precision oncology in subtypes that do not have obvious targetable genetic alterations, such as TNBC and ovarian cancer without germline BRCA1/2 mutations. Our study is proof of the concept that transcriptional/protein classifier generation and drug sensitivity analyses in cell lines could provide the basis for future umbrella clinical trials where patients will get different drugs depending on tumor expression–based predictors of drug sensitivity. This approach may prove especially useful in trials involving drug targets without an obvious target patient population, as is the case for current Hsp90i trials. Furthermore, our initial results warrant additional post hoc analyses of human tumors from completed Hsp90i clinical trials, which may provide further insight into the clinical utility of a transcriptional/protein classifier to predict drug sensitivity.

M.P. Goetz is an employee/paid consultant for Lilly, Genomic Health, Biovica, Novartis, Sermonix, Contex Pharm, Pfizer, and Biotheranostics, and reports receiving commercial research grants from Pfizer and Lilly. T.W. Miller reports receiving commercial research grants from Takeda Pharmaceuticals. No potential conflicts of interest were disclosed by the other authors.

Conception and design: K. Shee, T.W. Miller

Development of methodology: K. Shee, J.D. Wells, M. Ung, C. Cheng, T.W. Miller

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): K. Shee, R.A. Hampsch, N.A. Traphagen, W. Yang, S.C. Liu, L. Wang, J. Yu, J.C. Boughey, A.N. Kettenbach, M.P. Goetz, T.W. Miller

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): K. Shee, J.D. Wells, M. Ung, K.R. Kalari, J.C. Boughey, E. Demidenko, A.N. Kettenbach, C. Cheng, M.P. Goetz, T.W. Miller

Writing, review, and/or revision of the manuscript: K. Shee, M. Ung, L. Wang, K.R. Kalari, M.P. Goetz, T.W. Miller

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): K. Shee, J.D. Wells, M.A. Zeldenrust, M.P. Goetz, T.W. Miller

Study supervision: K. Shee, M.P. Goetz, T.W. Miller

We thank Alan Eastman for sharing antibodies and suggestions, Jonathan Marotti for pathology interpretation, and Norris Cotton Cancer Center Shared Resources for assistance (Mouse Modeling, Genomics and Molecular Biology, and Pathology, Microscopy). This work was supported by grants from NIH (F30CA216966 to K. Shee; R01CA200994, R01CA211869, and a pilot grant via P20GM113132 to T.W. Miller; R35GM119455 to A.N. Kettenbach; Dartmouth College Norris Cotton Cancer Center Support Grant P30CA023108; Mayo Clinic Breast SPORE P50CA116201-9 to M.P. Goetz, L. Wang, K.R. Kalari; and Mayo Clinic Cancer Center Support Grant P50CA015083); American Cancer Society (RSG-13–292-01-TBE to T.W. Miller); Susan G. Komen (CCR1533084 to T.W. Miller); Mayo Clinic Center for Individualized Medicine; Nadia's Gift Foundation; John P. Guider; The Eveleigh Family; George M. Eisenberg Foundation for Charities; Prospect Creek Foundation; and Randy Shaver Cancer Research and Community Fund.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.