Systematic screening identifies ABCG2 as critical factor underlying synergy of kinase inhibitors with transcriptional CDK inhibitors

All TNBC cell lines and ER + breast cancer cells were provided by Prof. J. Martens (Erasmus MC, the Netherlands). Cell lines were cultured in RPMI-1640 medium containing L-glutamine and 25 mM HEPES (Gibco, Fisher Scientific) supplemented with 10% foetal bovine serum, 25 U/mL penicillin, and 25 µg/mL streptomycin (Fisher Scientific) in a humidified incubator with 5% CO₂ at 37 °C.

For combination screening of kinase inhibitors with CDK inhibitors, a 378-kinase inhibitor library (L1200, Selleckchem) was used at 1 µM. For screening of ABCG2 activity using pheophorbide A, an updated batch of this library with 760 kinase inhibitors (363 overlapping inhibitors with previous library) was screened at 1 µM.

PHA-767491 (S2742), lapatinib (S2111), erlotinib (S7786) and flavopiridol (S1230) were from Selleckchem. CDKI-73 was provided by Prof. S. Wang (University of South-Australia) [22]. BAY-143572 (HY-12871B), THZ1 (HY-80013A), THZ531 (HY-103618), SR4835 (HY-130250), nilotinib (HY-10159), rabusertib (LY2603618, HY-14720) were purchased from MedChemExpress. Momelotinib (CYT387, S2219) and ralimetinib (LY2228820, S1494) were used from the Selleckchem L1200 kinase inhibitor library. Cisplatin was from the Leiden University Medical Center pharmacy (Leiden, The Netherlands).

Mouse RNA-polymerase II (RPB1 CTD) antibody (#2629) and rabbit MEK1/2 (#8727), phospho-MEK1/2 (Ser217/221, #9154), Akt (#9272), phospho-AKT (Ser473, #9271), ERK1/2 (#4695), phospho-ERK1/2 (Thr202/Tyr204, #9101), MCL-1 (#5453), phospho-RNA polymerase II (Ser2, #13499), phospho-RNA polymerase II (Ser5, #13,523), total PARP (#9542) and cleaved PARP (Asp214, #9541) were from Cell Signalling Technology. The mouse antibody against α-tubulin (#T-9026) was purchased from Sigma-Aldrich. Horseradish peroxidase (HRP)-linked anti-rabbit (#111-035-003) and anti-mouse (#115-035-003) and Alexa-647-linked anti-mouse (#115-605-146) secondary goat antibodies were from Jackson Immunoresearch.

Cell proliferation was evaluated after 4 days of treatment using sulforhodamine B (SRB) colorimetric assay [23]. TNBC cell lines were plated in 96-well plates and cells were treated with inhibitors the next day. After 4 days of treatment, cells were fixed by addition of 50% trichloroacetic acid (TCA, 30 µL in 100 µL medium, Sigma-Aldrich). Cells were stained with 0.4% SRB (Sigma-Aldrich) in 1% acetic acid (VWR) and washed 4X with 1% acetic acid to remove unbound SRB. Bound SRB was solubilized in 10 mM unbuffered Tris (Fischer Scientific) and SRB absorbance (540 nm) was measured after 2 h on a Tecan Infinite M1000 microplate reader. Dose–response curves and IC50 values were generated in GraphPad Prism (version 9). The percentage point of synergy was calculated as the difference between combined responses and the summed or additive response of each inhibitor as single treatment. Bliss and Loewe synergy scores were calculated using the SynergyFinderPlus web application [24]. Synergy scores for individual concentrations higher than 10 are generally considered synergistic. The mean of these individual synergy scores across the entire dose range was used for comparison of extent of synergy between different cell lines or conditions. Here, we considered combination treatments with a mean Bliss/Loewe synergy score > 20 as strongly synergistic.

Proteins were extracted, and western blotting was performed, using SDS-PAGE, as described previously [25]. Briefly, proteins were extracted using RIPA buffer and cellular proteins were denatured in soluble protein buffer containing 2-mercaptoethanol (final concentration 1.7%). Proteins (20 µg/lane) were loaded into 7.5% polyacrylamide gels for SDS-PAGE. Proteins were transferred to polyvinylidene fluoride (PVDF) membranes (Merck) and blocked with 5% bovine serum albumin (BSA; Sigma-Aldrich) in Tris-buffered saline with 0.05% Tween-20 (TBS-T_0.05). Membranes were subsequently incubated with primary and secondary antibodies in 1% BSA in TBS-T_0.05. Prior to imaging, HRP-conjugated secondary antibodies were incubated with ECL Prime Western Blotting Detection Reagent (GE Healthcare Life Sciences) for 3 min. Chemiluminescence or fluorescence was detected on an Amersham Imager 600 (GE Healthcare Life Sciences). Uncropped images used of western blots are available in Additional file Figures (Additional File 1: Fig. S12).

FUCCI plasmid pLL3.7 m-Clover-Geminin(1-110)-IRES-mKO2-Cdt(30-120) was a gift from Michael Lin (Addgene plasmid #83841) [26] and was used for lentiviral transduction of Hs578T cells. Hs578T cells expressing the construct were selected using FACS cell sorting.

For both AnnexinV/PI cell death and FUCCI cell cycle analysis, cells were seeded in µclear 96-wells black-bottom imaging plates (#655090, Corning). The next day, cells were stained with Hoechst-33342 (100 ng/mL) and then treated with inhibitors. For AnnexinV/PI assays, Annexin V-Alexa633 (0.05%) and PI (100 nM) were added to the cells simultaneously with the inhibitors. Live cell imaging of FUCCI cell cycle indicators and AnnexinV/PI signals was performed at indicated timepoints using a 10X objective on a ImageXpress Micro XLS imager (Molecular Devices). Images were analysed in Cell Profiler, and percentage of positive cells from total number of cells were reported. For FUCCI cell cycle analysis, we defined Cdt1-positive cells as cells in G1 phase, Cdt1- and Geminin-positive cells as cells in the G1/S transition, Geminin-positive cells as cells in S, G2 or M-phase and negative cells as cells in transition from M to G1 phase [26].

siRNAs were from the human siGENOME libraries (Dharmacon). siRNA knockdown was performed by reverse transfection with 25 nM SMARTpool siRNAs using INTERFERin transfection reagent (Polyplus, 409-50). Medium was refreshed after 18 h. 48 h after transfection, cells were treated with the inhibitors for 4 days. Day 0 plates for SRB assays were fixated on day of treatment. Controls were cells treated with transfection reagent only (mock) or a mixture of 720 siRNAs targeting different kinase genes, thereby not significantly affecting any gene expression (non-specific kinase pool, siKP). Knockdown data were normalized to siKP and day 0, unless stated otherwise.

Hs578T, SKBR7 and MDA-MB-231 with inducible Cas9 and gRNA’s were generated by lentiviral transduction with the Edit-R inducible lentiviral Cas9 plasmid (Dharmacon) and sgRNAs. Non-targeting control and sgRNAs against CDK12 and ABCG2 were from the human Sanger Arrayed Whole Genome Lentiviral CRISPR Library (Sigma–Aldrich) and were provided by Prof. R. Hoeben (Leiden University Medical Centre, The Netherlands). Cells were selected using puromycin (1 µg/ml). Knockouts were generated by induction of Cas9 expression with doxycycline (1 µg/mL) for 3 days. Knockout efficiency was evaluated after 4 days using TIDE analysis, as described previously [27] and/or Western blotting. For proliferation assays, cells were re-seeded 3 days after doxycycline induction and were incubated the next day with the inhibitors for an additional 4 days.

Genomic DNA was isolated using the Gentra^® Puregene^® kit (#158767, Qiagen) following the manufacture’s protocol specified for cultured cells and dissolved in the hydration solution overnight while shaking at room temperature. DNA yield ranged between 276 and 384 µg per sample. The genomic DNA was divided into multiple reactions per sample (50 µg each, using all material) and fragmented at 37 °C overnight, using 50 U Ndel enzyme (R0111L), 50U Pstl-HF enzyme (R3140L), and 50 µL 10X cutSmart^® buffer (B7204S) from New England Biolabs, supplemented to 500 µL with nuclease-free water (AM9932, Thermo Fisher). The reactions were heated to 100 °C for 10 min., and following addition of 500 µL 2 M NaCl, reheated to 100 °C for 5 min. and then immediately snap-frozen in liquid nitrogen. Per tube and prior to thawing, 1 µL of each 10 µM 5′ biotinylated capture oligo (TGCTTACCGTAACTTGAAAGTATTTCGATTTCTTGGCTTTATATATCTTG and TGCAGCCAGGTGGAAGTAATTCAAGGCACGCAAGGGCCATAACCCGTAAA) was added on top of the frozen solution, which was then immediately transferred to a thermoshaker for overnight hybridization at 60 °C. Next, to capture hybridized DNA, 20 µL Streptavidin T1 Dynabeads (#65602, ThermoFisher) were washed three times with 500 µL wash buffer (1 M NaCl, 10 mM Tris–HCl, pH 8), added to each tube, and incubated under rotation at room temperature for 2 h. The beads were washed twice with wash buffer and twice with 10 mM Tris–HCl (pH 8). Non-hybridized biotinylated oligos were digested in 50 µL reactions composed of 44 µL 10 mM Tris–HCl (pH 8), 5 µL 10X Exonuclease buffer, and 1 µL Exonuclease I (M0293L, New England Biolabs), at 37 °C for 1 h. Beads were washed 3 times with 10 mM Tris–HCl (pH 8) and resuspended in 20 µL 10 mM Tris–HCl (pH 8). Two rounds of PCR were performed to amplify the gRNA sequences. In the first PCR, distinct forward primers that each encode a unique barcode sequence and facilitate deconvolution of sequence reads of pooled samples (ACACTCTTTCCCTACACGACGCTCTTCCGATCTNNNNNNGGCTTTATATATCTTGTGGAAAGGACG with NNNNNN representing barcode sequences ACATCG, TGGTCA, CACTGT, ATTGGC, GATCTG, TCAAGT, and TACAAG) were used in combination with a common reverse primer (GTGACTGGAGTTCAGACGTGTGCTCTTCCGATCTTCTACTATTCTTTCCCCTGCACTGT). PCR mixture: 1 μL 10 μM forward primer, 1 μL 10 μM reverse primer, 1 μL 10 mM dNTPs (R0193, ThermoFisher), 0.5 μL Phusion polymerase and 10 μL 5X HF buffer (M0530L, New England Biolabs), supplemented with nuclease-free water to a total volume of 50 μL. PCR cycling conditions: 3 min. @ 98 °C, 20 times (30 s. @ 98 °C, 30 s. @ 60 °C, 30 s. @ 72 °C), and 5 min @ 72 °C. Per sample, products of individual reactions were pooled and 2 μL of each pool was used as template in the second PCR with conditions similar to the first, but having 15 instead of 20 cycles, to add the p5 and p7 adapter sequences (primers: AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACACGACGCTCTTCCGATCT and CAAGCAGAAGACGGCATACGAGATATCACGGTGACTGGAGTTCAGACGTGTGCTCTTCCGATCT). The PCR products were purified using the Bioline ISOLATE II PCR and Gel kit (BIO-52060, GC biotech) following the manufacture’s protocol, subjected to quality control assays (DNA concentration and purity, fragment size by gel electrophoreses), and subsequently pooled by combining 150 ng of each sample. The pool was sequenced on a Illumina HiSeq 2500, and the reads were mapped to the unique barcodes used for each sample and the Brunello library. Differential analysis on the single sgRNA level was performed using DESeq2 [28]. Subsequently, the MAGeCK ranking aggregation method was used to prioritize sgRNAs and calculate p values for each gene, which were corrected for multiple testing using the Benjamini–Hochberg method [29].

TNBC cell line mRNA microarray (GSE41313) and RNA sequencing data were previously established [30, 31]. Data were log-transformed (log₂), and rank-ordered genes (signal2Noise) were evaluated using GSEA software. Cell lines were grouped as either THZ531-sensitive (IC50 < 0.1 µM) and non-synergistic (either Bliss/Loewe synergy score lapatinib or nilotinib < 10) or THZ531-resistant (IC50 > 0.1 µM) and synergistic (both Bliss/Loewe synergy score lapatinib and nilotinib > 20) and used for phenotype analysis in GSEA. Cell lines that did not meet these conditions (MDA-MB-231, HCC1143, HCC1806) were excluded from GSEA analysis.

Cells were incubated with inhibitors and pheophorbide A (0.25 µM) for 23 h. 1 h prior to imaging, cells were stained with Hoechst-33342 (100 ng/mL) by directly adding this in the medium. Cells were imaged on a Nikon Eclipse TiE2000 confocal microscope using 20 × Plan Apo objective 24 h after treatment. Images were analysed in Cell Profiler. Normalized pheophorbide A intensity was calculated by dividing the total pheophorbide A intensity by the number of nuclei, and this value was normalized to DMSO. For kinase inhibitor screening a normalized total pheophorbide A intensity of > 1.2 in both replicates was considered as pheophorbide A accumulation. Kinase inhibitors that had a normalized intensity of > 1.2 in one of the two independent experiments, but < 1.2 in the other, were classified as “uncertain”. Normalized pheophorbide A accumulation was correlated to extent of synergy for kinase inhibitors with THZ531 only for kinase inhibitors that did not strongly affect proliferation as single treatment (> 40% proliferation compared to DMSO).

The dataset used for the training of the classification model of ABCG2 inhibition is based on previously published data [32]. This original data set was used to generate a prediction model based on a logistic regression algorithm and is freely available for the scientific community as a web service (https://​livertox.​univie.​ac.​at) [33]. The dataset used in the current study was further enriched by including compounds obtained from the open-source databases ChEMBL and PubChem. The retrieved compounds were standardized using a modified Atkinson standardization protocol (available at github.com/flatkinson/standardiser). In more detail, bonds to alkali metals and alkaline earth metals were removed. In case of a mixture, the fragments were split, and each fragment was standardized separately. Non-organic compounds were eliminated, and functional groups were neutralized. Duplicates were removed from the dataset including stereoisomers. Further, if duplicate compounds between different sources disagree in the classification label, they were excluded. Compounds with an activity value below 10 µM were labelled as an inhibitor and above this threshold as non-inhibitor. The acquired data set contained in total 1442 compounds, which was divided into 776 inhibitors and 666 non-inhibitors. The performance of the model was validated by a tenfold cross-validation as well as an external test set. This test set was obtained by excluding 10% of the original dataset using the sample method implemented in pandas with a random state of 0. Finally, 1298 compounds (694 inhibitors and 604 non-inhibitors) were used for the training and 144 compounds (82 inhibitors and 62 non-inhibitors) for the validation step.

For the chemical representation of the compounds, molecular descriptors as implemented in RDKit (version 2019.03.3) were used. For the training of the prediction model, a random forest classifier was chosen based on the scikit-learn [34] (version 0.21.2) libraries. The model showed an accuracy of 0.85, which corresponds to the rate of correct predictions. Eighty-five percentage of actual positives were correctly identified as well as 85% of actual negatives, which is indicated by the sensitivity and specificity. Further, the Matthews correlation coefficient (MCC) was determined to estimate the quality of the models. MCC differentiates between random and well performing predictions. While a value of 0 indicates that a prediction is just random, a value of 1 indicates a perfect prediction and − 1 a wrong prediction. The score for the classification model of ABCG2 activity is 0.70. For the external test set, the sensitivity was comparable with the tenfold cross-validation with a value of 0.84 and slightly lower for the specificity with a value of 0.76.

Kinase inhibitors that had a ABCG2 inhibitory prediction score ≥ 0.7 were classified as “active”, while kinase inhibitors with a score of 0.4–0.6 were classified as "weak or uncertain” and kinase inhibitors with a score ≤ 0.3 were classified as “inactive”.

RNA was isolated from Hs578T cells 8 h after treatment with the inhibitors using the RNeasy Plus Mini Kit (Qiagen). RNA quality and quantity was measured on an Agilent-4200 Bioanalyzer. Unstranded PolyA-selected libraries were prepared using the DNBseq platform, and 50 M 100-bp paired-end reads were sequenced on a DNBSEQ-G400 sequencer by BGI Europe. Filtered reads were aligned to Grch38 using Hisat2 [35]. For differential gene expression analysis, aligned reads were counted using featureCounts [36]. Normalization and differential gene expression analysis were performed using DESeq2 [28]. Pathway enrichment of ranked gene lists was evaluated in STRING [37].

For quantification of intronic polyadenylation (IPA), the PolyAsite atlas (version 2.0), a database of genomic locations of alternative polyadenylation sites previously determined by 3′ end sequencing, was used [38]. To evaluate IPA events, we used custom scripts to quantify differential exon usage after IPA sites and expression of IPA sites (scripts available upon request). Briefly, intronic reads spanning − 10 and + 10 basepairs of intronic polyadenylation sites (exonic reads excluded) were counted from the aligned reads of the RNA sequencing data using HTSeq-count and custom GTF files describing these regions and exons [39]. Log2 fold changes and adjusted p values of exons and IPA sites were quantified using DESeq2. Upregulation of IPA sites was considered significant for log2 fold changes > 1 and adjusted p values < 0.001. Stringent p values and minimum counts of 10 reads for a minimum of 3 samples were used to exclude falsely increased IPA usage due to noise in intronic regions.

Exon usage was quantified using DEXSeq-count [40]. Counts of exons before (upstream of), and after (downstream of) IPA sites of an individual gene were summed, creating overall exon counts from the part of the gene before and after the IPA site. These were normalized and quantified using DESeq2. Gene parts downstream and upstream of the IPA site were considered to be significantly differentially used when the difference in log2 fold change downstream and upstream of the IPA site was larger than 0.75 and the part downstream of the IPA site was significantly altered with adjusted p value < 0.01. IPA sites significantly upregulated and resulting in differentially expressed gene parts were considered significant IPA events.

Statistical tests were performed in GraphPad Prism and are further specified in corresponding figure legends. For all figures, significance was expressed as follows: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001.

We have previously demonstrated that the HER2/EGFR inhibitor lapatinib synergizes with the CDC7/CDK9 inhibitor PHA-767491 in TNBC cells [10]. To investigate whether this is also the case for other transcriptional CDK inhibitors, we treated the TNBC cell lines Hs578T, SKBR7 and MDA-MB-231 with a combination of lapatinib and CDC7/CDK9 inhibitor PHA-767491, CDK9 inhibitor CDKI-73, CDK9 inhibitor BAY-1143572, CDK7 inhibitor THZ1, CDK12/13 inhibitor THZ531 or CDK12/13 inhibitor SR4835. In addition to PHA-767491, lapatinib synergized most strongly with CDKI-73, THZ1 and THZ531 in Hs578T, SKBR7 and MDA-MB-231 cells (Fig. 1A and B, Additional file 1: Figure S1A–D). However, the combination of lapatinib with SR-4835, BAY-1143572, or flavopiridol did not induce a similarly strong effect in all of these cell lines.

Given the relatively high synergy scores for THZ1 and THZ531 and the fact that several studies demonstrated synergistic activity with other compounds [11, 13, 18‐20], we aimed to systematically assess the kinase inhibitor landscape that could sensitize TNBC to these CDK inhibitors. We therefore screened a combination of THZ531 or THZ1 with 378 kinase inhibitors in Hs578T and MDA-MB-231 TNBC cells. We found that a strikingly large number of kinase inhibitors synergized with THZ531 in Hs578T (153 kinase inhibitors with > 25 percentage point of synergy) and MDA-MB-231 cells (60 kinase inhibitors with > 25 percentage point of synergy) (Fig. 1C and D, Additional file 1: Fig. S2A; Additional file 2: Table S1). A smaller number of kinase inhibitors synergized with THZ1 in Hs578T cells (20 kinase inhibitors with > 15 percentage point of synergy), all of which also synergized with THZ531 (Fig. 1D, Additional file 1: Fig. S2B; Additional file 2: Table S1), suggesting a similar underlying mechanism. In MDA-MB-231 cells, THZ1 also synergized with many kinase inhibitors (58 kinase inhibitors with > 25 percentage point of synergy), although single treatment of THZ1 at this concentration already affected their proliferation slightly (Additional file 1: Fig. S2B; Additional file 2: Table S1). Although the kinase inhibitors that synergized with THZ531 have a wide range of targets across multiple pathways, inhibitors targeting (receptor) tyrosine kinases induced the most synergy with THZ531 (Fig. 1C, Additional file 1: Fig. S2C, p < 0.0001).

For further validation and exploration of the synergistic drug interactions, we selected several compounds, currently in clinical trials or FDA approved, that displayed more than 25 percentage point of synergy in both Hs578T and MDA-MB-231 cells and that inhibit different targets, including EGFR tyrosine kinase inhibitors lapatinib and erlotinib, Bcr-Abl tyrosine kinase inhibitor nilotinib, Chk1/2 serine/threonine inhibitor rabusertib, JAK1/2 tyrosine kinase inhibitor momelotinib and p38 MAPK serine/threonine kinase inhibitor ralimetinib. These compounds indeed induced synergy with THZ531 across a wide range of concentrations in TNBC cell lines Hs578T (Fig. 1E), SKBR7 (Additional file 1: Fig. S2D) and MDA-MB-231 (Additional file 1: Fig. S2E). Altogether, these data demonstrate that THZ531 strongly synergizes with a large variety of kinase inhibitors.

Next, we further investigated the mechanism behind the growth inhibition upon the combination treatments with THZ531. Therefore, we assessed the cell cycle progression using Hs578T cells expressing the fluorescent cell cycle indicators (FUCCI) [26]. Whilst the single treatments did not affect cell cycle progression, the combination treatments with THZ531 (0.1 µM) and lapatinib, nilotinib, rabusertib, momelotinib or ralimetinib synergistically halted Hs578T cells in the G1/S checkpoint, already within 12 h of combination treatment, which became more prominent after 24 h of treatment (Fig. 1F and G, and Additional file 1: Fig. S3A). In addition, using Annexin V and propidium iodide (PI) staining, we studied whether this combination treatment only arrests cells in the cell cycle, or whether it eventually also kills TNBC cells. Indeed, at 48 and 72 h, but not yet at 24 h, these combination treatments synergistically induced apoptotic cell death in Hs578T and SKBR7 cells (Fig. 1I, Additional file 1: Fig. S3B and C). Of note, the cell death and cell cycle arrest dynamics induced by the combination treatments were similar to treatment of THZ531 at tenfold higher dosage (1 µM). Together these data demonstrate that the combination of THZ531 and a diverse set of kinase inhibitors halts Hs578T cells in G1/S checkpoint and subsequently causes cell death.

To further examine the relevance of these synergistic interactions, we investigated whether they also apply to other TNBC cell lines. Indeed, the combination of THZ531 and lapatinib or nilotinib also synergistically inhibited cell proliferation in BT20, SUM229PE, SUM149PT, HCC1937 and SUM159PT TNBC cell lines (Bliss and Loewe synergy scores for lapatinib and nilotinib combinations > 20) (Fig. 2A, Additional file 1: Figure S4). Lapatinib or nilotinib also sensitized HCC1143 and HCC1806 cells to THZ531, albeit to a lesser extent (Bliss or Loewe synergy score of lapatinib or nilotinib between 10 and 20), while in MDA-MB-468, MDA-MB-436, BT549, HCC38, MDA-MB-453 and HCC70 the synergy was limited (Bliss or Loewe score of lapatinib or nilotinib < 10). Of note, cell lines that were sensitized to THZ531 by lapatinib or nilotinib were relatively resistant (IC50 ranging from 0.2 to 4.6 µM) to THZ531 treatment alone, while cell lines that did not demonstrate strong synergy were relatively sensitive (IC50 ranging from 0.032 to 0.1 µM) to THZ531 (Fig. 2B). This observation was not specific for TNBC cells, as it also applied to ER + cell lines MCF7 and T47D (Additional file 1: Fig. S4).

As THZ531 sensitivity was associated with the extend of synergy with tyrosine kinase inhibitors, we further investigated the mechanisms of resistance to THZ531 in TNBC cells. Using a whole genome pooled CRISPR/Cas9 knockout screen, we identified 11 genes that support the proliferation of Hs578T cells in the presence of THZ531 (Fig. 2C). Knockout of ABCG2, TTC14, CDK10, FKBP8, CDK13, CYP2A13 and ARHGEF1 only affected proliferation of Hs578T cells in the presence of THZ531, with ABCG2 being the strongest regulator of sensitivity (Fig. 2D, Additional file 1: Fig. S5A).

We also investigated which transcriptomic features distinguish the cell lines resistant to THZ531, and demonstrating synergy in combination with lapatinib or nilotinib, from the cell lines sensitive to THZ531 and not demonstrating this synergy. For this purpose, we utilized previously established microarray and RNA sequencing-based expression data [30, 31]. Both analyses pointed out a high expression of ABCG2 in the cell lines demonstrating the resistant and synergistic phenotype (Fig. 2E, Additional file 1: Fig. S5B). Importantly, previous studies have shown that multi-drug ABC transporters ABCB1 or ABCG2 can cause acquired resistance to THZ1 upon long-term drug selection by lowering intracellular THZ1 concentrations [41, 42] and that this can cause cross-resistance to THZ531 [41]. Notably, ABCG2 was the only gene associated with the resistant and synergistic phenotype that overlapped with the significant hits in the CRISPR-Cas9 knockout screen. Moreover, the genes identified by the CRISPR-Cas9 knockout screen and phenotypic associations do not overlap with any of the intended targets of the synergistic kinase inhibitors, suggesting that these inhibitors may sensitize the cells to THZ531 by directly interacting with one of the identified proteins, most likely ABCG2 (Additional file 1: Fig. S5C; Additional file 3: Table S2). Consistent with this hypothesis, multiple tyrosine kinase inhibitors, including lapatinib and nilotinib, have been previously identified to also inhibit ABCG2 [43]. To further rule out target-related mechanisms of synergy, we investigated whether lapatinib or nilotinib sensitized cells to depletion of CDK12 and CDK13. CDK12 knockout in Hs578T or MDA-MB-231 cells or knockdown of CDK12 and/or CDK13 in Hs578T cells did not increase sensitivity to lapatinib or nilotinib (Additional file 1: Fig. S5D–F).

We next investigated whether the role of ABCG2 in resistance to THZ531 in Hs578T cells extends to the other CDK inhibitors that synergized with lapatinib, and whether this also holds true for TNBC cell lines other than Hs578T. Indeed, doxycycline-induced knockout of ABCG2 sensitized Hs578T cells to THZ531, THZ1, CDKI-73 and PHA-767491, but barely to SR4835 and not to flavopiridol, thus following similar trends as for combined inhibition with lapatinib (Fig. 3A, Additional file 1: Fig. S6A and B). ABCG2 knockout also sensitized SKBR7 cells most strongly to THZ531, THZ1 and CDKI-73 (Additional file 1: Fig. S6A and C). In MDA-MB-231 cells, ABCG2 knockout mostly sensitized to THZ531 and THZ1, but to a lesser extent than in Hs578T, which also is in line with the increased sensitivity to the CDK inhibitors alone and limited sensitization by lapatinib in this cell line, as described above (Additional file 1: Fig. S6D). Moreover, in Hs578T, SKBR7 and MDA-MB-231 cells with ABCG2 knockout, the synergistic interaction of lapatinib with THZ531 was strongly reduced (Additional file 1: Fig. S6E). Similar to genetic knockout, ABCG2 knockdown also sensitized Hs578T and SKBR7 cells to THZ531, THZ1, CDKI-73 and, to a lesser extent, PHA-767491 (Fig. 3B, Additional file 1: Fig. S7A). Likewise, ABCG2 inhibitor KO143 sensitized Hs578T and SKBR7 cells to THZ531, THZ1, CDKI-73 and PHA-767491 with equal potency as lapatinib (Fig. 3C, Additional file 1: Fig. S7B). Like the combination treatment of various tyrosine kinase inhibitors with THZ531, combination of KO143 with THZ531 arrested cells in the G1/S checkpoint firstly and subsequently induced cell death in Hs578T (Fig. 3D).

Since previous studies have indicated that resistance to THZ1 and THZ531 can also be associated with ABCB1 overexpression [41, 42], we used siRNA knockdown to analyse the importance of other ABC-transporter family members for sensitivity to these compounds. In line with the results of our CRISPR/Cas9 screen, only knockdown of ABCG2 increased THZ531 sensitivity in Hs578T and SKBR7 cells (Fig. 3E, Additional file 1: Fig. S7D). In addition, unlike KO143, ABCB1 inhibitor verapamil did not sensitize Hs578T or SKBR7 cells to THZ531 (Fig. 3F, Additional file 1: Fig. S7E).

Although previous studies have also reported that some tyrosine kinase inhibitors, including lapatinib and nilotinib, can inhibit ABCG2 activity, this has not yet been demonstrated for the entire set of kinase inhibitors that synergized with THZ531 [43]. To investigate whether the selected kinase inhibitors in fact inhibit ABCG2 at the synergistic concentrations, we evaluated ABCG2 activity by measuring intracellular accumulation of the fluorescent ABCG2 substrate pheophorbide A [44, 45]. Indeed KO143, lapatinib, nilotinib, momelotinib, ralimetinib and rabusertib caused an increase in intracellular pheophorbide A intensity in Hs578T and SKBR7 cells, suggesting inhibition of ABCG2 activity (Fig. 4A and B). MEK inhibitor selumetinib, that did not synergize with THZ531, and THZ531 itself, did not affect pheophorbide A accumulation.

Next, we examined whether this is a general mechanism of synergy between THZ531 and the identified kinase inhibitors. We therefore screened the kinase inhibitor library for effects on pheophorbide A accumulation. In line with a general mechanism of synergy, we identified 103 kinase inhibitors (from a total of 303 evaluated with less than 40% effect on proliferation at used concentration) that caused pheophorbide A accumulation (fold change > 1.2). These kinase inhibitors also showed a higher extent of synergy in the previously described kinase inhibitor combination screen with THZ531, compared to the 112 kinase inhibitors that did not affect pheophorbide A accumulation (Fig. 4C, Additional file 4: Table S3). Although the extent of pheophorbide A accumulation could not fully describe the variation in extent of synergy with THZ531, there was a significant correlation, showing that a higher pheophorbide A accumulation is associated with increased potentiation of THZ531 treatment (Spearman r = 0.45, P < 0.0001, Fig. 4D). To explore the relationship between ABCG2 activity and sensitivity to THZ531 further, we used computational modelling to predict the probability of the inhibitory activity of the kinase inhibitors for ABCG2 based on several structural scaffolds described previously [32, 33] (see Methods). This model assesses the interaction profile of small molecules towards ABCG2 using a machine learning approach. It could be found that kinase inhibitors predicted to be inhibitors of ABCG2 synergized more strongly with THZ531 than those predicted to be inactive (Fig. 4E, Additional file 4: Table S3). Indeed, compounds with predicted ABCG2 inhibitory activity also had a higher extent of pheophorbide A accumulation (Additional File 1: Fig. S8). To conclude, these data indicate that a striking number of (mainly tyrosine) kinase inhibitors synergize with THZ531 and other transcriptional CDK inhibitors by inhibiting ABCG2 transport in TNBC cells.

To further substantiate the role of ABCG2 in the synergistic interactions between THZ531 and other kinase inhibitors, as opposed to other mechanisms such as adaptive responses or targeting of specific pathways, we investigated how the activity and expression of the targets of the synergistic treatments are affected upon treatment. Corresponding to decreased THZ531 efflux by ABCG2 inhibition, the combination treatments with THZ531 (0.1 µM) and lapatinib, nilotinib or rabusertib synergistically reduced phosphorylation of RNA polymerase II (Ser2) and to a lesser extent total expression of RNA polymerase II, in a similar fashion as THZ531 at a higher dose (1 µM), in Hs578T and SKBR7 cells (Fig. 5A, Additional file 1: Fig. S9A and B). Moreover, the short-lived pro-survival protein MCL-1, which is often downregulated by transcriptional CDK inhibitors as a direct result of transcriptional inhibition [46], was also downregulated upon treatments with combinations of THZ531 with lapatinib or nilotinib, or a high dose of THZ531.

In accordance with ABCG2 inhibition as the main mechanism of synergy, and not adaptive signalling or simultaneous inhibition of a specific pathway, we did not observe clear changes within MAPK or PI3K pathways, the key downstream pathways of EGFR signalling, with the individual or combination treatments after 24 h in Hs578T or SKBR7 cells (Fig. 5A and Additional file 1: Fig. S9A). Although after 48 h the combination treatment of lapatinib and THZ531 slightly increased levels of p-MEK and p-ERK, this may be related to induction of apoptosis, rather than stimulation of these growth factor pathways (Additional file 1: Fig. S9B). Indeed, the combination of lapatinib and THZ531 caused PARP cleavage at 48 h in both Hs578T and SKBR7 cells, indicative of caspase-dependent apoptosis. Collectively, these data show that the tyrosine kinase inhibitors lapatinib and nilotinib mainly sensitize the cells towards the inhibitory effect of THZ531 on RNA polymerase II activity, which subsequently leads to apoptotic cell death.

To explore whether additional mechanisms contribute to the effect of ABCG2 inhibition on THZ531 response, we studied the transcriptional changes caused by THZ531 (0.1 or 1.0 μM), lapatinib (3.16 μM), nilotinib (1.0 μM) and combination treatments. RNA sequencing on samples collected 8 h after treatment revealed that mainly combination treatments of 0.1 μM THZ531 with lapatinib or nilotinib and treatment with 1.0 μM THZ531 alone affected transcription, predominantly resulting in downregulation of gene expression (2953, 2883 and 2723 genes, respectively, log2 fold change (FC) < − 1, adjusted p < 0.05; Fig. 5B; Additional file 5: Table S4). Corresponding to the increased sensitivity to THZ531 caused by lapatinib or nilotinib mediated ABCG2 inhibition, the differences in gene expression caused by the treatment with a high dosage of THZ531 alone were nearly identical to the effects by the combination treatments (Fig. 5C and D). This further supports that nilotinib and lapatinib are sensitizing the tumour cells to THZ531 and further rules out that the synergy is occurring because THZ531 is suppressing adaptive transcriptional responses. Only a limited number of genes got slightly upregulated upon treatment with lapatinib alone (56 genes, log2 FC > 0.5, adjusted p < 0.05) or nilotinib alone (72 genes, log2 FC > 0.5, adjusted p < 0.05) (Additional file 1: Fig. S10A and B).

GO biological processes, that were enriched among the genes downregulated by combination treatment with THZ531 and lapatinib, were mostly related to regulation of RNA expression, cell cycle progression and DNA damage checkpoints (Fig. 5E). Biological processes, which were enriched among the upregulated genes, included mainly processes involved in translation and protein localization. Previous studies have highlighted that CDK12/13 inhibitors mainly disrupt the expression of DNA damage genes [5‐8]. Although the effects in our study are not limited to these genes, we do also observe that DNA damage response genes are mostly downregulated upon combination treatment with THZ531 (0.1 µM) or high dosage of THZ531 (1 µM) (Fig. 5F, Additional file 1: Fig. S10C). Moreover, compared to other genes, these treatments more specifically reduced the expression of many transcription factors, including mostly zinc finger (ZNF) transcription factors (Additional file 1: Fig. S10C and D). Altogether these data show that CDK12/13 inhibition with THZ531 strongly disrupts gene expression and further supports that combination with lapatinib or nilotinib sensitizes cells to THZ531 to inhibit CDK12/13-regulated gene expression.

Previous studies have demonstrated that a key mechanism behind transcription regulation by CDK12/13 inhibitors is an increase in premature cleavage and polyadenylation in introns, called “intronic polyadenylation” (IPA) [6, 8]. Indeed, our polyA-enriched RNA sequencing data also showed gene track patterns of upregulated intronic polyadenylation with concomitant downregulation of transcription of downstream exons after treatment with THZ531, for example in the SOS1 and JAK2 genes (Fig. 6A, Additional file 1: Fig. S11A). The combination treatment of lapatinib and nilotinib with THZ531 (0.1 µM) or a high dose of THZ531 (1 µM) more strongly reduced the expression of exons after (downstream of) potential IPA sites compared to the expression of exons before (upstream of) these sites, suggesting a premature end of the transcript around these IPA sites (Fig. 6B, Additional file 1: Fig. S11B). Moreover, these treatments increased the expression of many of these potential IPA sites themselves (Fig. 6C). Taken together, while the individual treatments with lapatinib, nilotinib and low concentration of THZ531 (0.1 µM) did not affect intronic polyadenylation, lapatinib and nilotinib sensitized the cells to CDK12/13-related defects in transcriptional processing by THZ531 (Fig. 6D–F and Additional file 6: Table S5).

Previous studies have demonstrated that IPA upon CDK12/13 inhibition mostly occurs in DNA damage response genes due to their relatively long length [6, 8]. Indeed, the high dose of THZ531 and combination treatments cause IPA of several genes, which are enriched for genes involved in DNA damage response pathways, such as ERCC6 (Fig. 6G, Additional file 1: Fig. S12A and B). However, our data demonstrate that also genes from other pathways are affected, including genes involved in processes such as regulation of transcription (e.g. SMARCA4), cell cycle (e.g. E2F3) and protein tyrosine kinase signalling (e.g. EGFR) (Fig. 6G and Additional file 1: Fig. S12A, B). Interestingly, several of these genes are associated with pathways targeted by lapatinib or nilotinib, including EGFR, JAK2, BRAF, and SOS1. Although IPA of these genes is likely not the driving force behind the synergy, this disruption of gene processing could further hamper the growth of TNBC cells with the addition of these treatments. Overall, these data further support that tyrosine kinase inhibitors sensitize cells to THZ531 and thereby CDK12/13 related transcriptional defects by inhibiting ABCG2-mediated drug efflux (Fig. 6H).