Systematic drug screening reveals specific vulnerabilities and co-resistance patterns in endocrine-resistant breast cancer

Human breast cancer cell lines MCF-7 (HTB-22, ATCC), T-47D (HTB-133, ATCC), ZR-75-1 (CRL-1500, ATCC) and BT-474 (HTB-20, ATCC) were obtained from the American Type Culture Collection. The cells were grown in DMEM with L-Glutamine (MCF-7 and BT-474, PAN Biotech, Aidenbach, Germany) or RPMI-1640 with L-Glutamine (ZR-75-1 and T-47D, PAN Biotech) supplemented with 10 % FCS (Gibco, Life Technologies, Carlsbad, CA) and 1 % penicillin/streptomycin (Gibco). Culture media for T-47D, MCF-7 and BT-474 additionally contained 0,1 % bovine insulin (Sigma. St. Louis, MO). The tamoxifen-resistant cell lines (MCF-7 Tam1, T-47D Tam1 & Tam2, ZR-75-1 Tam1 & Tam2, BT-474 Tam1 & Tam2) were derived from the parental cell lines by continuous exposure to 4-OH-tamoxifen (Sigma, 1 μM in ethanol) for 8–12 months. Culture media was replaced every 2–3 days. All cells were incubated at 37 °C with 5 % CO₂ and passaged when ca 80 % confluent. The approximate doubling times of the cells were as follows: parental MCF-7, T-47D, ZR-75-1 and BT-474: 1–3 days. Resistant MCF-7 Tam1, T-47D Tam1 and Tam2: 1–2 weeks, ZR-75-1 Tam1 and Tam2: > 1 week, BT-474 Tam1 and Tam2: 2 weeks. The cells were free of mycoplasma and verified for their authenticity (Technology Centre, Institute for Molecular Medicine Finland, Helsinki, Finland).

For viability measurements cells were seeded in 384-well culture plates with increasing tamoxifen concentrations (0–1,8 μM). After 120 h cell viability was evaluated by CellTiter-Glo Cell Viability Assay (Promega, Fitchburg, WI) with the PHERAstar plate reader (Agilent Technologies Santa Clara, CA). To measure the active DNA synthesis cell proliferation assays with Click-iT® EdU Alexa Fluor® 488 Flow Cytometry Assay Kit were performed according to manufacturer’s protocol (Life Technologies) with following minor modifications: Cells were plated on 10 cm plates and cultured with and without 1 μM tamoxifen until approximately 50 % confluent. Parental cells and their tamoxifen-resistant derivatives were then pulsed with 10 μM of EdU Alexa Fluor® 488 for 4 h (T-47D and MCF-7) or for 28 h (BT-474 and ZR-75-1). Cells were permeabilized with saponin-based permeabilization and wash reagent (MCF-7, T-47D, BT-474) or with 0,1 % TritonX-100-PBS (ZR-75-1) for 10 min. Additionally, DNA content staining was performed using FxCycle™ Far Red and the cell suspension treated with RibonucleaseA. Flow cytometry was carried out and results analyzed using Accuri C6 flow cytometer and associated software (BD Biosciences, Franklin Lakes, NJ). To measure estrogen-responsivity of the parental cells and their tamoxifen-resistant derivatives, the cells were grown on 6-well plates in hormone-deprived culture medium for 72 h (phenol red-free DMEM or RPMI, PAN Biotech), supplemented with 2,5 % dextran–charcoal-treated (Sigma-Aldrich) FCS and other additives (see above). 17β-estradiol (Sigma, 10⁻⁸ M in ethanol) was then added back to the cells for 4, 8 or 24 h and RNA isolated with Total RNA Purification kit (Norgen, Thorold, ON). 4 μg of total RNA were reverse transcribed with the High-Capacity cDNA Reverse transcription kit (Applied Biosystems, Thermo Scientific, Waltham, MA) as instructed. Quantitative-PCR was then performed on the LightCycler 480 system (Roche, Penzberg, Germany) using the DyNAmo colour flash SYBRGreen PCR kit (Thermo Scientific) with equal amounts of cDNA. The optimal internal reference gene was determined out of a pool of 16 different housekeeping genes for each parental-resistant cell line pair (18S for MCF-7 s, PPIA for T-47Ds and B2M for ZR-75-1 s and BT-474 s). Primer sequences can be found in Additional file 1. All experiments were done in triplicates. For Western Blotting cells were grown on 10 cm dishes, and lyzed in Laemmli buffer. Immunoblotting was done as previously described [25]. The used antibodies were as follows: ERα (Abcam, Cambridge, UK, ab16660), β-actin (Sigma-Aldrich, A1978), EGFR (Cell Signaling Technologies, CST4267), phospho-EGFR (CST3777), ERK1/2 (CST9107), phospho-ERK1/2 (CST4370). To test the effects the ERK inhibitor VX-11E and the MEK inhibitor selumintinib the cells were cultured either in their default culture media with no additional drug, or with increasing concentrations of VX-11E (50nM, 100nM and 250nM), or with 100nM VX-11E in combination with 1 μM selumetinib. The cells were then harvested and Western blotting with the above-mentioned antibodies performed.

Genomic DNA was isolated from the parental and tamoxifen-resistant cells using the DNeasy Blood & Tissue kit (Qiagen, Venlo, Netherlands and Hilden, Germany). Exome-capture was done on 3 μg DNA with the NimbleGen SeqCap EZ Human Exome v2.0 kit (Roche NimbleGen) and paired-end sequencing performed on Illumina HiSeq platform. Point mutations were detected as previously described [26], using the parental cell lines as controls. Briefly, point mutations specific to resistant samples were called with VarScan2 somatic [27], with the following parameters: strand-filter 1, min-coverage-normal 8, min-coverage-tumor 6, somatic-p-value 1, normal-purity 1, min-var-freq 0,05. Parental cell line was used as the normal control. Mutation calling was done within the exome capture target regions of the NimbleGen SeqCap EZ v2 capture kit and the flanking 500 bps. Mutations were annotated with SnpEff [28] using the Ensembl v66 annotation database. To filter out misclassified germline variants, common population variants included in dbSNP version 135 were removed. Remaining non-synonymous mutations were visually validated using the Integrated Genomics Viewer (Broad Institute). Mutations with p < 0,05 and resistant variant frequency >30 % were deemed high confidence. Known false positive point mutations [29] were excluded. Exome-sequencing data was also analyzed using the sequence alignment and variant calling pipeline VCP. As input in the CNV analysis we used alignments in BAM format, as well as identified variants in all samples [30] (and unpublished). All exome sequencing capture kit target regions less than 76 bp apart were merged with each other. An RPKM (reads per kilobase of target region length per million mapped reads) copy number value was calculated separately for every target region, followed by filtering out regions with sequencing coverage lower than 25x. Finally, relative log2 copy number ratios for sample (drug-resistant variant) divided by reference (parental cell line) were calculated and segmented using Circular Binary Segmentation [31]. Plots of copy number, segmentation and variant allele frequencies in capture target regions were visualized using R. Gene level copy number data for all human genes in Ensembl database v67 was calculated by assigning a gene the value of the CNV data segment that it overlapped. When a gene overlapped more than one segment, the gene was assigned a copy number value based on a modification of the extreme method option in GISTIC2 [32‐34] as follows: the gene was given the lowest segment log2 value in case any overlapped segment had log2 ratio < = −0,6 and the highest segment value if any overlapped segment had log2 ratio > = 0,5. If all segments the gene overlapped had log2 ratio > −0,6 and < 0,5; the gene was assigned the median log2 ratio of all overlapped segments. Thresholds for copy number changes were determined based on samples (not published here) with known copy number differences, such as male vs female comparisons on chromosome X, as well as trisomies observed in karyotyping of cells during routine diagnostics. Based on these, the limits were set at −0,4 (heterozygous deletion), −1,2 (homozygous deletion), +0,5 (gain) and +1,3 (amplification). Raw exome-sequencing data have been deposited in the NCBI Sequence Read Archive [SRP: SRP050366].

DSRT with the FIMM FO2Baq library containing 279 approved and investigational oncology drugs was done as previously described [26] with minor modifications. Briefly, drugs were dissolved and plated in five different concentrations covering a 10 000-fold concentration range into the wells of 384-well plates. Optimized amounts of cells were then seeded into the wells in their normal growth media, i.e. parental cells in normal media and tamoxifen-resistant cells in media supplemented with 1 μM 4-OH-tamoxifen. Thus, this set-up allows for measurement of permanent drug response changes corresponding to long-term tamoxifen treatment used in the clinical setting, and allows for any combinatorial drug responses to be observed. A further Cells were incubated at 37 °C for 72 h and viability measured by CellTiter-Glo Cell Viability Assay with the PHERAstar plate reader. Data were normalized to negative (DMSO only) as well as positive (100 μmol/l benzethonium chloride) controls. The logistic dose–response curves were estimated using the Marquardt-Levenberg algorithm and implemented in the in-house bioinformatic pipeline Breeze. The dose–response curves were then employed to quantitatively profile drug responses, i.e. the Drug Sensitivity Score (DSS), which is based on the estimated logistic model parameters, and the DSS difference, which quantifies the differential drug response between tamoxifen-resistant and parental cells, as previously described [26, 35]. We found that |dDSS| = 5 cutoff lies in the tail end of the distribution with 9.7 % of values above the cutoff. With a two-tailed distribution (signed DSS), this would correspond to ca. 5 % “hit rate”, which we deemed as appropriate for such a drug discovery approach. Clustering of the DSS response differences across resistant/parental cell line pairs was performed using unsupervised hierarchical complete-linkage clustering, using Spearman and Euclidean distance measures of the drug and sample profiles, respectively, and visualized as a heat map [35]. In order to identify drugs that significantly change their efficacy as the cells gain resistance to tamoxifen, we performed rank product analysis [36] at false discovery rate of 5 % (q < 0,05) by comparing drug response profiles in the parental cells against response profiles in the tamoxifen-resistant cells. The average DSS activity of a drug in all parental cell lines was plotted against the average DSS activity in the resistant clones. The drugs selected from the rank product analysis were considered as significant hits and displayed as colored dots. Luminal A or B subtype-specific drugs were identified (Additional file 2) based on the known subtypes of the parental cells [37].

To visualize drug sensitivity and resistance networks in each resistant/parental cell line pair, drugs with DSS difference >5 (sensitivity) or < −5 (co-resistance) were selected. For each drug, specific target molecules were defined using the KIBA (Kinase Inhibitor BioActivity) -score [38] as follows: First, drug target bioactivity (Ki, Kd and IC₅₀) values were extracted from the ChEMBL database (https://​www.​ebi.​ac.​uk/​chembl), and integrated KIBA-score was calculated for each drug-target pair. A low KIBA-score indicates a high binding affinity of the drug with the target. Amongst the set of targets for a drug, if the target with highest binding affinity had KIBA-score of <0,1, the specific target threshold was considered to be 50-fold the lowest KIBA-value; otherwise KIBA-score <3 units was considered as the specific target threshold. Second, to capture the connection between genomic changes and drug-target genes of the most effective drugs in each resistant cell model compared to its parental cell line, network analysis and visualization was done on KIBA-scored specific targets genes of the sensitizing or desensitizing drugs using Ingenuity Pathway Analysis application (Ingenuity® Systems, Qiagen). The IPA system is based on the Ingenuity Pathways Knowledge Base, which is a repository of curated biological interactions, and functional annotations between proteins, genes, complexes extracted from scientific articles. The network consists of thousands of nodes and the edges represent experimentally observed cause-effect relationships that are related to direct physical interactions or regulatory interactions events like expression, transcription, activation and molecular modifications. Network edges are also associated with a direction of the causal effect, i.e. either activating or inhibiting [39]. Sensitizing and desensitizing drugs were considered separately. We required each resistant/parental cell line pair to have at least three effector drugs with a DSS difference >5 to enable network building based on their target genes, discarding the sensitivity networks for MCF-7 Tam1 and ZR-75-1 Tam1 as each of them had one effector drug only. The target molecule networks representing the most effective sensitizing or desensitizing drugs were merged, and then adjusted to keep the edges that were human-specific and high confidence. The resulting networks were then extended to upstream neighboring molecules to reveal connections with genes containing copy number variations or somatic mutations within the target networks. In the IPA framework, upstream molecules are defined as the genes that have been shown to affect the gene expression in some direct/indirect way [39]. The IPA network generation algorithm considers all the immediate upstream molecules separated by one degree from the nodes in the current network, where priority is given to those genes having maximum number of overlaps with the existing network. Furthermore, we merged the IPA canonical pathway for Estrogen Receptor Signaling (www.​qiagen.​com) with the networks utilizing the IPA overlay tool.

To study the development of tamoxifen resistance across distinct molecular backgrounds, we established seven long-term tamoxifen-treated cell lines originating from four parental cell types; MCF-7, T-47D, ZR-75-1 and BT-474. We exposed these ER-positive, initially tamoxifen-responsive cells to continuous 1 μM 4-OH-tamoxifen treatment (hereafter tamoxifen) for 8 to 12 months (Fig. 1). A steady concentration of tamoxifen with clear inhibitory effect on ER-mediated transcription was chosen to mimic the exposure of patients to the drug in the clinic. In the course of the treatment, all cell cultures underwent cell death leaving a few founder cells that then recovered and repopulated the culture plates. Once stable, the cultures were molecularly characterized and subjected to pharmacogenomic profiling through drug sensitivity and resistance testing (DSRT) and exome-sequencing (Fig. 1). To confirm the resistant phenotypes, we assessed the viability of the cells upon increasing tamoxifen concentrations. All resistant cells showed increased tolerance towards tamoxifen compared to the parental cells, verifying the resistance development (Additional file 3A). For further characterization, we studied the cell cycle properties, functionality and levels of ER, and estrogen-responsivity of the long-term tamoxifen-treated cells. Resistant cell clones exhibited altered cell cycle properties, with a higher fraction of cells in the G0/G1 phase (Additional file 3B). All resistant cells had diminished ER target gene transcription, whereas the level of ER itself either increased or decreased depending on the clone (Additional file 4). Estradiol (E2) deprivation and subsequent addition of estradiol back to the cells revealed aberrant levels of ER target gene transcription, indicating that upon acquiring resistance to the ER antagonist tamoxifen, transcriptional regulation in response to the natural ligand of the receptor, estradiol, is disturbed.

To uncover drug response profiles associated with long-term tamoxifen treatment mimicking the prolonged exposure to the drug used in the clinic, we conducted drug sensitivity and resistance testing (DSRT) [26] with 279 FDA/EMA-approved and investigational oncology drugs ranging from conventional chemotherapeutics to a variety of targeted drugs (Additional files 5 and 6). We quantified the overall drug responses for each compound in each cell line using drug sensitivity score (DSS) [35]. Targeted drugs generally had higher efficacy than the standard chemotherapeutics, as seen for example with the HER2/EGFR inhibitors lapatinib, afatinib and neratinib that were efficacious only against the BT-474 parental cells that express these receptors (Additional file 2A and B). We then calculated the DSS difference between the tamoxifen-resistant and their parental cells to reveal the cell line-specific drug sensitivity and resistance patterns associated with the resistance to tamoxifen. Interestingly, all tamoxifen-resistant cell lines displayed distinct drug response profiles, with a few overlapping responses (Fig. 2a). Overall, the drug responses of the resistant clones derived from the same parentals resembled each other more than those of the other resistant clones, and a few shared sensitivities arose, such as towards tyrosine kinase (gefinitib, ibrutinib) and cdk inhibitors (alvocidib, SNS-032) in the BT-474 clones. However, in T-47D and ZR-75-1, only a few common drug sensitivities and co-resistance patterns evolved between the two subclones, indicating clonal evolution (Fig. 2b).

To define the mutational landscape of the tamoxifen-resistant cell clones and to identify potential genomic markers for drug sensitivity and resistance, we conducted exome-sequencing on each isogenic parental-resistant cell line pair. In order to identify resistance-specific point mutations and copy number variations, each tamoxifen-resistant cell line was compared to its parental cell line. Between 31,8 and 83,0 million uniquely mapping reads per sample were obtained with an average coverage of 43,7x (Additional file 7). We found genetic changes scattered along the genomes of the tamoxifen-resistant cells (Additional files 8, 9 and 10). We identified approximately 250 to 350 non-synonymous mutations per resistant cell line (Additional file 9), and for integration with copy number alterations, selected the high confidence ones. Interestingly, only a limited panel of mutated genes was shared between the two clones derived from the same parental cells including TIMM23 and RP11-368 J21.2.1 in ZR-75-1 Tam1 and Tam2, and TNS1, PTH2R and NHLRC2 in BT-474 Tam1 and Tam2. None of the point mutations were shared between T-47D Tam1 and Tam2 (Additional file 8). All resistant clones displayed marked copy number aberrations, which were primarily large deletions, with the exception of ZR-75-1 Tam1 that only harbored gains. Chromosomes X and seven were recurrently altered, with few larger deletions (such as a homo- and heterozygous deletion of Xq in MCF-7 Tam1 and heterozygous deletion of 7q in T-47D Tam1), as well as several smaller aberrations found in each resistant cell line (Additional files 8 and 10). Again, just a few genes displaying copy number alterations were shared between the two resistant clones derived from ZR-75-1 and BT-474, implying genetic clonal divergence.

To reveal the pharmacogenomic relationships between the drug response profiles and the genetic alterations, we integrated the drug response and genetic profiling data. Specifically, based on the DSS differences, we first selected the most pronounced drug response changes between each parental-resistant cell line pair. To correlate the genomic alterations with the drug responses we then constructed sensitivity and resistance networks among the target genes of the selected drugs, and mapped the point mutations and CNV changes onto these molecular networks as detailed in Materials and Methods, Construction of drug sensitivity and co-resistance networks. We also merged the obtained networks with the IPA canonical pathway for Estrogen Receptor Signaling. However, only in two out of fourteen networks more than two overlapping molecules were found, indicating that only a fraction of them are actually connected with the ER signaling pathway and hence, in the majority of cases, the sensitivity and resistance mechanisms arise independently of ER. Overall, the number of genetic alterations mapping onto the drug target networks varied between 11 (T-47D Tam2 sensitivity network) and 0 (some ZR-75-1 and BT-474 networks), and converged into a number of target genes. In all but two networks (BT-474 Tam2 and ZR-75-1 Tam2 sensitivity) all the observed CNV changes were heterozygous deletions (Figs. 3, 4, 5 and 6 and Additional files 11 and 12). Of the two tamoxifen-resistant clones generated from the parental T-47D, both exhibited mostly sensitized profiles. T-47D Tam1 and Tam2 both acquired sensitivity towards the SRC/ABL-family inhibitors; the Tam1 cells towards three of them: ponatinib, dasatinib and KX2-391; and the Tam2 cells against two of them: dasatinib and KX2-391 (Figs. 3 and 4). The Tam1 cells also became sensitive towards selumetinib, a MEK inhibitor, and the Tam2 cells towards VX-11E, an ERK1/2 inhibitor. In addition, sensitivity towards BCL-family inhibitors (navitoclax, obatoclax), and RAF inhibitors (regorafenib, RAF265) was observed in Tam1, but not in Tam2. Reflecting the acquired drug sensitivities, multiple target genes of these drugs or their upstream effectors converged into the same sensitivity networks, and notably, many of these target genes also harbored genetic changes.

Similar to the T-47D resistant clones, the tamoxifen-resistant BT-474 s displayed mainly acquisition of drug sensitivities. However, unlike any of the other tamoxifen-resistant clones, BT-474 Tam1 and Tam2 developed vulnerability to several EGFR/HER2-inhibitors (gefitinib, ibrutinib, neratinib, nintedanib; Fig. 5a, c and Fig. 6a, c). Notably, the parental BT-474 s also showed sensitivity towards HER2/EGFR inhibitors (Additional file 2), which was selectively enhanced upon developing resistance to tamoxifen. In addition, sensitivity towards cdk-inhibitors (SNS-032, alvocidib) emerged, and several of the drugs’ target genes or their upstream effectors also harbored CNV changes (Fig. 6d). Both tamoxifen-resistant BT-474 clones also developed sensitivity towards the ERK1/2-inhibitor VX-11E and the JAK2-inhibitor gandotinib. BT-474 Tam2 also displayed gain of JAK2, possibly explaining the developing sensitivity towards gandotinib (Fig. 6d).

In addition to emerging drug sensitivities, we also observed acquired co-resistances upon development of tamoxifen resistance. In contrast to the other resistant cells that developed both new sensitivities as well as co-resistances, the tamoxifen-resistant MCF-7 Tam1 displayed an overwhelmingly co-resistant drug response profile, including resistance towards many chemotherapeutics (camptothecin, vincristine, SNS-032), but also several mTOR- and two HDAC-inhibitors (dactolisib, AZD8055, sirolimus, panobinostat, belinostat) (Additional file 11). In the two T47-D tamoxifen clones, the resistance networks were markedly different, with T47-D Tam1 displaying resistance merely to four agents, two of which were chemotherapeutics, whereas T-47D Tam2 cells exhibited co-resistance to a large variety of drugs. These included some of the same drugs as for the Tam1 clone, and additionally Tam2 showed resistance to four mTOR- and two HDAC-inhibitors (Figs. 3 and 4). Both ZR-75-1 as well as the BT-474 resistant clones developed co-resistance to several chemotherapeutics (Figs. 5 and 6, Additional files 5, 6, 12), with the ZR-75-1 networks being nearly identical, reflecting the high similarity between their drug response patterns. Interestingly, the BT-474 Tam1 cells additionally showed resistance to rapamycin and everolimus, two mTOR inhibitors, as well as to AT 101, a BCL-family inhibitor (Fig. 5). Collectively, compared to the sensitivity profiles, the resistance networks demonstrated less variance between the different tamoxifen-resistant clones, with the majority of them developing co-resistance towards common chemotherapeutics (Table 1).

We next addressed the development of common sensitivities and co-resistances upon acquisition of tamoxifen resistance. Cross-comparison of the drug response profiles across all tamoxifen-resistant clones revealed that several of them developed sensitivity towards the ERK1/2-inhibitor VX-11E, the proteasome-inhibitor bortezomib, and the FAAH-inhibitor PF-3845. Common co-resistance towards the survivin-inhibitor YM155 and the chemotherapeutic agent paclitaxel also occurred (Fig. 7a). Furthermore, even with the limited sample size, these shared responses were statistically significant (rank product analysis, Fig. 7b). To further assess the EGFR/ERK signaling pathway in the resistant cells, we selected the MEK/ERK inhibitors for which a differential drug sensitivity score was obtained, i.e. VX-11E and selumetinib. The concentration range was selected based on the IC₅₀ of the drugs in the individual cell lines (Additional file 5). We then cultured the cells either in their default culture media, or with increasing concentrations of VX-11E, or with VX-11E and selumetinib concomitantly, and performed Western blotting with ERK1/2 and EGFR antibodies (Fig. 7c). The basal levels of these unphosphorylated as well as phosphorylated signaling proteins were lowered in the tamoxifen-resistant cells compared to the parentals (Fig. 7c), but upon increasing concentrations of VX-11E, slight increase in phosphorylated ERK1/2 was observed in the T-47Ds. However, upon additional inhibition of MEK with selumetinib, the increase in phosphorylated ERK1/2 was diminished.