Background
The triple negative subtype of breast cancer (TNBC), devoid of the hormone estrogen/progesterone receptor (ER/PR+) expression and HER2 overexpression (HER2+), is a heterogeneous group of aggressive diseases that account for 15–20 % of all breast cancer cases. While targeted treatments exist for the receptor positive breast cancer subtypes, TNBC lacks such specific treatments. The current line of therapy is limited to surgery, radiation and chemotherapy [
1,
2]. TNBC patients have a worse prognosis than other breast cancer patients. In a breast cancer patient follow-up study, 93 % 5-year survival was seen in non-TNBC-patients, as compared to only 77 % of the TNBC-patients [
3]. Hence, there is an obvious need for better treatment options for TNBC.
Development of targeted therapeutics for TNBC diseases is challenging due to their heterogeneity. To address this challenge, several studies have assigned TNBC cases into multiple subtypes using transcriptomics approaches. For example, Kreike et al. [
1] studied 97 TNBC samples and 7700 genes and observed five groups (I-V) in a hierarchical analysis of gene expression data. In another study, TNBC was grouped into seven transcriptomics-based subtypes: basal-like 1 (BL1) and 2 (BL2), mesenchymal-like (M), mesenchymal stem cell-like (MSL), immunomodulatory (IM), luminal androgen receptor positive (LAR) type, and unclassified (UNC) [
4]. Recently, Burstein et al. [
5] performed a similar study and identified four subtypes; LAR, MES, BLIS and BLIA. By assessing expression of 13 biomarkers, TNBC was assigned to four subtypes by Elsawaf et al. [
6]. Based on intrinsic PAM50 subtyping, 80.6 % of TNBC were found basal-like, 14.6 % normal-like, 3.5 % luminal A, 1.1 % luminal B and 0.2 % HER2-enriched [
7,
8]. While there is overlap between the results in the different studies, such as identification of the LAR-type in two recent studies [
1,
5], it is evident that defining clear, distinct subgroups is challenging, highlighting the diversity of the disease.
Despite the diversity, a number of therapy-guiding biomarkers have been proposed and efforts for tailoring targeted therapies against TNBC are ongoing. Though
TP53,
BRCA1/2,
EGFR,
PIK3CA and
PTEN tend to be dominant mutations in TNBC, these markers have been elusive and inconsistently useful for guiding therapy [
9,
10]. An important finding is that Poly-ADP-ribose polymerase (PARP) inhibitors appear to be highly effective against the
BRCA1-mutant TNBC [
11]. The PARP inhibitor olaparib was recently approved for use in
BRCA-mutated ovarian cancers and several PARP inhibitors are currently in clinical trials against
BRCA-mutated TNBC. Furthermore, inhibitors of PI3K, mTOR, CDK, HDAC and androgen signaling are currently being explored in clinical trials as treatments of TNBC [
8]. Also drugs targeting different growth factor receptors such as EGFR, VEGFR and FGFR are explored in clinical trials [
7].
As an alternative strategy to tailor targeted therapies to breast cancers, chemosensitivity profiling of in vitro cell lines is applied increasingly. This functional profiling approach allows for identification of selective vulnerabilities in cell lines reflecting human diseases. Recently, Barretina et al. [
12] and Garnett et al. [
13] tested 25 TNBC lines against 24 anticancer agents and 10 TNBC lines against 130 compounds, respectively, as part of large comprehensive pharmacogenomics studies in hundreds of cell lines. Heiser et al. [
14] performed an analysis of 77 cancer drug compounds on 19 TNBC cell lines, and combined the drug-sensitivity data with gene expression and copy number interrogation. In a similar study by Daemen et al. [
15], 19 TNBC cell lines were screened against 90 compounds along with integration of multi-omic molecular profiling data to identify potential response-predictive markers. Another study by Lawrence et al. [
16] reported a combined proteomics, genomics, and drug sensitivity interrogation using 160 compounds and 16 TNBC cell lines and four tumor samples. Muellner et al. [
17] identified the broad-spectrum tyrosine kinase inhibitor midostaurin (PKC412) as a post-EMT-specific drug targeting spleen tyrosine kinase SYK in a subset of TNBC cells.
Together, these studies identified a number putative links between drug sensitivities, TNBC subtypes and genomic and proteomic markers, but also highlighted a striking functional heterogeneity among TNBC cell lines. Notably, each of these studies used cell viability readouts to monitor the drug responses. However, the consistency in Barretina et al. and Garnett et al. datasets [
18] was poor, possibly due to differences in experimental setup and especially in their viability readouts. These results highlight the importance of the well-defined functional readouts in chemo-sensitivity profiling studies.
By studying the response of 16 diverse TNBC cell lines to 301 compounds using our drug sensitivity and resistance (DSRT) approach [
19,
20], we investigated whether going deeper than the traditional cell viability measurement could provide us with improved drug response selectivity information of translational value. Our compound testing included a resazurin cytosolic reduction-based viability assay, a cellular ATP-based viability assay, and a cell membrane impermeable DNA-binding dye-based cytotoxicity assay [
21]. While the readouts of the two viability assays correlated well, we found that several classes of drugs caused an apparent drastic loss in viability and cell numbers but failed to induce cell death. However, among these generally cytostatic agents, sporadic cytotoxicity occurred in specific cell lines, suggesting that some cell lines, and presumably cancers with the same genotype and/or phenotype, may be strongly and selectively responsive to such chemotherapy. Furthermore, when testing drug combinations, the cytotoxicity readout revealed both antagonistic and synthetic lethal effects that were missed by the viability readouts.
Discussion
In this study, we systematically explored how comprehensive drug responses with different cell health readouts compared to current subgroupings and previously described biomarker information of triple-negative breast cancers. The results led us to several conclusions. First, the drug response clustering of the TNBC cell lines based on their differential drug vulnerabilities resulted in highly heterogeneous patterns of drug responses and distinctive grouping compared to gene expression derived grouping. Second, by studying cell death rather than cell viability, which has so far been the standard readout in other large-scale chemosensitivity profiling studies, we could separate static responses from the cytotoxic ones and identify several drug classes that exhibit broad viability readout effects but induce only limited or no cell killing responses. Third, the cytostatic responses seen by many drugs were reversible, and by studying the cells in real-time, we detected that some of the static responses were overcome even in the presence of the drugs. Fourth, by measuring the cytotoxic responses in drug combination studies, synergistic cytotoxic responses and even synthetic lethalities were detected that were not observed with the cell viability readout. In our combination studies, we explored the simultaneous exposure of compounds because we lacked the scientific evidence suggesting that one agent should be added before the other, and without that information, exploratory testing of many combinations in different addition orders became unmanageable in size and cost. However, as a proof of approach, we did perform combinatorial order of addition testing of dactolisib and trametinib in DU4475 and MDA-MB-231 to see if the antagonistic effects of this combination could be reverted by an appropriate sequential addition of the compound. In these experiments, the combination remained antagonistic in both cell lines regardless of whether simultaneous or sequential addition was applied (Additional file
11: Figure S8).
Similar to what others have shown [
15,
22] we detected a great heterogeneity in drug responses among the TNBC cell lines, re-emphasizing the heterogeneous nature of this breast cancer subtype. As has also been shown by others as well as by us in other cell systems [
19], the overall drug response profiles are not easily linked to genetic or transcriptional profiles, arguing that functional drug response profiling is currently the most powerful way to identify individualized vulnerabilities that can be used to target the disease. Alternatively, a more refined analysis of TNBC transcriptomics may be needed for effective linking to broad drug sensitivities.
In vitro anti-cancer chemosensitivity testing has traditionally been focused on growth inhibition measurements with the assumption that reducing or stopping cancer cell growth will translate into an anti-cancer activity of the agent in vivo. We hypothesized that by also following the drug-induced cytotoxicity, one can discover a deeper and different range of drug responses, which may also lead to more translationally-predictive results. Overall, our results with the cytotoxicity measurement strongly argue that high throughput chemosensitivity profiling of cancer cells need to go beyond the current standard viability measurement. There are various types of viability measurement reagents commonly employed in multiwell-based assays but most of them monitor the metabolic activity of the cells, measuring the amount of energy molecules like ATP, NADH, NADPH or the redox activity in the cells. Here, we show that using two different commonly used viability readouts, one measuring cellular ATP and one measuring reducing potential of the cells give highly correlated results. In other chemosensitivity profiling studies, it has often been implied that the loss in apparent cell viability should also strongly correlate to cytotoxicity. However, our data clearly show that this is often not the case. Only some of the compounds that appeared effective when assessed using cell viability readouts, were able to potently induce cell death. The easily adaptable multiplexed cell viability and cell death readout we applied allowed us to identify several drugs and drug classes that inhibited viability across the 19 breast cancer cell lines but failed to induce broad cell death responses. These included PI3K/mTOR inhibitors, CDK inhibitors, HSP90 inhibitors, anti-metabolites and antimitotic drugs. Importantly, we also showed that these cytostatic responses were fully reversible. Cells started growing as soon as inhibitory compounds were removed. Furthermore, in some cases (such as with rapamycin analogs and some CDK inhibitors), cell growth inhibition was bypassed over time even in the presence of the compounds, presumably an effect of cellular reprogramming in response to the drug as has been described in other model systems [
23,
24]. Among the compounds that caused a very preferential cytostatic effect, there was still heterogeneity in cytotoxic responses within the cell line panel. In most cases, there were subsets of the cell lines that exhibited strong cytotoxic responses, such as CAL-85-1, MDA-MB-231, CAL-51, Hs-578-T, to antimitotic taxanes and we hypothesize that these selective cytotoxic responding cell lines represent the TNBC subgroups that are more likely to respond to each specific type of therapy.
PI3K/AKT/mTOR signals have gained attention as potential therapeutic targets for several cancer types [
25‐
27]. Our results suggest that PI3K/mTOR inhibitors are able to induce viability inhibition in most of the cells lines, but fail to induce selective cell death. Given that mTOR inhibition is expected to slow down or halt cellular metabolism and thereby cell growth, our finding is not surprising
per se, but it emphasizes that using a cell viability/metabolic readout most likely is not relevant when assessing the effects of PI3K, AKT and mTOR inhibitors in vitro or ex vivo. Judging from the cell death readout, all the cell lines except CAL-148 were unresponsive or only had weak responses towards the inhibitors that target the PI3K/AKT/mTOR pathway. Furthermore, mTOR inhibitors, presumably through their antimetabolic activity, antagonize the activity of diverse classes of compounds, including many conventional antimitotic and cytotoxic drugs. Numerous clinical trials for PI3K and mTOR inhibitors along with conventional chemotherapy are ongoing, but our results argue that combining mTOR inhibitors with traditional chemotherapy such as doxorubicin, etoposide, gemcitabine should be considered with caution as combinations might turn out to be counterproductive. However, as our study was carried out using cell cultures, it may not fully reflect the responses in the considerably more complex biological settings when treating a cancer patient.
We found that mitotic and proteasome inhibitors had a heterogeneous cytotoxic effect on TNBC cell lines. This led us to try to find biomarkers that could be linked to the cytotoxic effects of the mitotic and proteasome inhibitors. We compared the basal protein and phosphoprotein levels in cell lines that were either sensitive or insensitive to the mitotic and proteasome inhibitors. Despite of the small overlap between our cell line collection and Daemen et al. [
15] study, we were able to identify some candidates that could potentially be further explored for predictive biomarkers.
The mitotic inhibitor-responsive cell lines expressed a higher level of the survival regulator PKCα [
28]; of FGFR1 that has been linked to TNBC cell growth [
25,
26]; of the cell cycle and apoptosis regulator c-Jun [
27]; and of caveolin-1, low level of which has been linked to poor clinical outcome in TNBC [
28]. The mitotic inhibitor-sensitive cell lines also expressed low levels of NOTCH3, which has been linked to induction of apoptosis in HER2-negative breast cancer cell lines [
29,
30]; of the small GTPase protein Rab25, which has been linked to aggressiveness of epithelial cancers [
31]; of Bcl-2 and Stat3 high expression of which have been linked to the development of chemoresistance [
32‐
34] and of the well-known driver of chemoresistance, HER2 [
35,
36].
Proteasome inhibitors have been found efficient against hematologic malignancies but less successful against solid tumors [
37]. We discovered that the proteasome inhibitor-sensitive cell lines exhibited high level of PAI1, a well established prognostic biomarker for the selection of chemotherapy [
38]; MKP-1 [
39]; AKT and p38. Low levels of NOTCH3; the cell cycle regulator Cyclin D1 that has been linked to chemoresistance in multiple cancers [
40,
41] and PTEN were observed in proteasome inhibitor-sensitive cell lines. Proteasome inhibition has been shown to activate phosphorylation of p38, MKP-1, and AKT that further activate resistance to proteasome inhibitors [
42,
43]. Similarly, suppression of PTEN has been linked to chemoresistance [
44]. Our findings suggest that some subgroups TNBC might be responsive to treatment with proteasome inhibitors.
In vitro/ex vivo drug sensitivity testing is a re-emerging area of research, thanks to improved possibilities to follow phenotypic drug responses and the possibility to link the responses to deep molecular profiling. However, there are limitations to such high throughput testing. Due to experimental logistics and scale it is still challenging to comprehensively address the complex pharmacology and metabolism of the compounds in vivo, extended time dependent effects of the drugs, as well as order of addition combination testing. Therefore, false negative results are always possible in these types of screening approaches. For example, some of the compounds in our collection represented prodrugs that are metabolized into active substances in the liver in vivo and are largely inactive in vitro. This group of compounds included most alkylating agents but these were still included in our collection because they are approved for human use. We also also attempted to use the active metabolites of several of these compounds in our screens but they were too unstable to be suitable for screening use. On the other hand, some other prodrugs, such as nucleoside analogs, are metabolized in the target cells and were therefore highly relevant to include in the in vitro testing. Furthermore, compounds with diverse mechanisms of action are likely to reach their cellular effects at different time points, a challenge when performing high throughput testing of broad arrays of agents where a single endpoint measurement becomes the most feasible assay readout. In our testing, we chose the 72 h endpoint for the experiments, as we found it sufficient to observe the activity of the majority of the compounds in our collection. Extending the incubation to up to 168 h did not significantly affect the overall results (Additional file
12: Figure S6), and hence the shorter time point was preferred because of assay logistics and robustness.
Here we studied cell lines, but novel technologies have emerged in recent years that may allow for systematically taking this approach on primary patient cells ex vivo, ultimately making it more directly translational. One example is the culture of 3-dimensional organoids closely recapitulating disease conditions [
45]. Organoid culture methods have already been established for human mammary tissue and primary breast cancer cells [
46,
47]. In an alternative approach, primary cells can be cultured on fibroblast feeder cells in the presence of a Rho kinase inhibitor resulting in immortalized, conditionally reprogrammed progenitor-like cells. This approach allows for 2- or 3-dimensional culture of patient-derived cells that maintain the heterogeneity of the initial tissue environment [
48,
49]. Finally, the prospect of generating these types of cultures from either circulating tumor cells or biopsies opens the possibility to explore ex vivo drug response testing without major surgical intervention [
50,
51], although the success rate and time to establish cultures for comprehensive testing are still bottlenecks.
Acknowledgements
We thank Laura Turunen, Swapnil Potdar and other members of the High Throughput Biomedicine Unit of the FIMM Technology Centre, (supported by University of Helsinki/Biocenter Finland research infrastructure funds) for technical assistance. We thank Astrid Murumägi, Disha Malani, Päivi Östling, Akira Hirasawa, Evgeny Kulesskiy and Sarang Talwelkar for their help with the drug sensitivity reference panel. Funding for the work was provided by the Jane & Aatos Erkko foundation (KW), the Sigrid Jusélius foundation (KW), the Academy of Finland (272577 and 277293 for KW and 272437, 269862, 279163 and 292611 for TA), Cancer Society of Finland (KW and TA), and the University of Helsinki Doctoral Program in Biomedicine (DPBM) for salary support to PG, AS and BY.
Competing interest
The authors declare that they have no competing interests.
Authors’ contributions
PG designed and performed experiments, prepared figures and data tables, and wrote the manuscript. LK planned the project and experiments and wrote the manuscript. AS performed the bioinformatic analysis and assisted in writing, SKJ performed CellTiterBlue based screening, BY calculated drug responses and assisted in writing. TA supervised the project and assisted in writing. KW conceived and supervised the project and wrote the manuscript. All authors read and approved the final manuscript.