A high-throughput drug combination screen of targeted small molecule inhibitors in cancer cell lines

While there is a high interest in drug combinations in cancer therapy, openly accessible datasets for drug combination responses are sparse. Here we present a dataset comprising 171 pairwise combinations of 19 individual drugs targeting signal transduction mechanisms across eight cancer cell lines, where the effect of each drug and drug combination is reported as cell viability assessed by metabolic activity. Drugs are chosen by their capacity to specifically interfere with well-known signal transduction mechanisms. Signalling processes targeted by the drugs include PI3K/AKT, NFkB, JAK/STAT, CTNNB1/TCF, and MAPK pathways. Drug combinations are classified as synergistic based on the Bliss independence synergy metrics. The data identifies combinations that synergistically reduce cancer cell viability and that can be of interest for further pre-clinical investigations.

In this paper, we report results from a high-throughput screen testing the effects of combining 19 small-molecule inhibitors on cancer cell viability. The drugs target signalling proteins involved in several well described signalling pathways, including PI3K/AKT, NFkB, JAK/STAT, CTNNB1/TCF, and MAPK pathways. Inhibitors were selected for high specificity and a minimum of known off-target activities, as judged from characterizations available from the MRC PPU lab (https://www.ppu.mrc.ac.uk/) 14 for 15 of 19 drugs. Eight cancer cell lines were treated with single inhibitors and 171 pairwise combinations at equimolar concentrations from 10-0.16 µM. Cell viability was determined after 48 hours from ATP content using CellTiter-Glo 2.0 (Promega). Several of the synergistic combinations involved the PI-103 inhibitor. Six of these combinations were investigated in a secondary screen. We here applied a matrix design allowing for synergy assessment outside the equimolar range and determined inhibitor effects from both viability (ATP content) and cell confluency (brightfield imaging). Figure 1 gives a schematic overview of the study design.
In the here presented dataset we rediscovered previously reported synergistic combinations and discovered several new synergies. The synergistic reduction of cell viability resulting from jointly targeting PI3K/mTOR and TAK1, previously reported by us for the AGS cell line 15 and later by others 16 , is observed in two of the tested cell lines. Furthermore, we rediscovered the synergistic effect of jointly targeting MEK and PI3K [17][18][19] . Recently, Sathe et al. have reported on increased effect of combined inhibition of PI3K/mTOR and PDPK1 in bladder carcinoma cells 20 . We identify that combined targeting of PI3K/mTOR and PDPK1 is potent across multiple cancer cell lines. The PI3 kinase is one of the canonical activators of PDPK1, and our dataset thus identifies this 'vertical synergy' intervention by targeting multiple points in one cascade as a potential chemical intervention for inhibiting PI3K-dependent cancer growth across cell line models representing different tissue types. Co-targeting of multiple vertical points in one cascade is also the rationale for the clinically verified 'vertical synergy' observed by jointly targeting MEK and BRAF [3][4][5][6] . Moreover, PI3K and PI3K/mTOR inhibitors tested in a clinical setting (e.g. pictilisib 21 and apitolisib 22 ) have suffered from toxic effects 23,24 . Thus, identification of compounds that increase the therapeutic index for PI3K inhibitors could enable new uses for PI3K inhibition in oncology.
The outcomes of our work contribute to expanding the available data sets on novel combinations of targeted drugs across a variety of cancer cell lines and can benefit the fields of cancer therapeutics, molecular signalling interventions and molecular cancer biology.
Primary screen -ray designed drug treatment. Cell lines were treated for 48 hours with 19 small-molecule inhibitors in single and pairwise applications, comprising a total of 171 drug combinations. All Fig. 1 Schematic representation of study design. Eight human cancer cell lines from different tissue origins were used in this study. Cells were incubated overnight prior to drug addition. In the primary screen, cells were screened against 19 small-molecule inhibitors in single and double application, with a total of 171 combinations. After 48 hours of drug exposure the assay was terminated, and cell viability was measured using CellTiter-Glo 2.0 (Promega). In the secondary screen, cells were screened against 7 single small-molecule inhibitors and 6 combinations for a duration of 48 hours. Drug effect was measured using automated brightfield imaging of confluency and CellTiter-Glo 2.0 (Promega).
www.nature.com/scientificdata www.nature.com/scientificdata/ drugs were pre-diluted in DMSO to a stock concentration of 20 mM. Drugs were applied in an equimolar ray design from 10 µM to 0.016 µM using the following procedure: First, 2.0 mM solutions of each inhibitor and inhibitor combination were prepared. These were then serially diluted in DMSO (four 5-fold serial dilutions). The solutions were further diluted in medium (8 µl inhibitor in DMSO in a total volume of 228 µl), of which 5 µl were added to each cell culture well. Total cell culture volume per well was 35 µl. Highest final concentrations were thus 10 µM in each well. All dilutions and liquid handling were performed with robotic equipment using sterile and pyrogen free pipette tips. The inhibitors are listed in Online-only Table 1. Final DMSO concentration in assay wells was 0.5%. Secondary screen -matrix designed drug treatment. In the secondary screen six combinations were screened in a matrix designed across all cell lines. As the PI inhibitor was identified as one of the top drugs involved in synergistic combinations, it was chosen as the anchor drug. The inhibitor was combined with BI, D1, JN, G2, PD, and 5Z. Cell lines were treated for 48 hours with the seven small-molecule inhibitors in single application and six drug combinations in all possible combinations of doses 10, 2, 0.4, 0.08 and 0.02 µM. Drugs, prepared as stocks of 20 mM in DMSO, were first diluted to 400x final concentration in DMSO. The solutions were further manually mixed 1:1 with either DMSO, for single drugs, or combination drug and applied to a 96 well plate. Further liquid handling was performed with robotic equipment. The solutions were diluted in medium (28.5x), of which 5 µl was added to each cell culture well leading to a total volume of 35 µl. We note that a new set of inhibitors was ordered for this screen.
High-throughput screen. Cells were seeded in 384-well plates (CORNING, 3712) at densities shown in Table 1 and allowed to attach overnight, before they were treated with small-molecule inhibitors and incubated for another 48 hours. The cell seeding densities were selected such that the cells were in a sub confluent stage at the time of drug addition. The variances in doubling time between the cell lines were taken into account and seeding densities were adjusted so that cell cultures in wells were either in late log phase or early plateau at the end of the incubation period. Cell viability was measured using the commercially available CellTiter-Glo 2.0 assay (Promega). In the secondary screen confluency was included as an additional readout and monitored by brightfield imaging using (SpectraMax i3x MiniMax 300 Imaging Cytometer, 2 views per well). A set of control wells with cells treated with 0.5% DMSO was included on all plates. The assay was performed with three biological replicates, if not stated otherwise.
For the first two biological replicates of the primary screen, two technical replicates are reported. Two of the four technical replicates assayed were excluded due to insufficient mixing of drug dilutions for these replicates. For the third biological replicate all four technical replicates assayed are included. For MDA-MB-468 and AGS cells, one or two biological replicates were performed, respectively. Some data points for the AGS cell line had to be excluded due to erroneous drug addition within the robotic script.
For the first biological replicate in the secondary screen, some technical replicates had to be excluded for the SW-620 cell line as too little drug volume was added. For the MDA-MB-468 cell line the first biological replicate was excluded due to low cell viability in the control wells. For the first, second and third biological replicate some technical replicates had to be excluded due to non-efficient cell attachment caused by an air bubble below the cell suspension. Additionally, some wells were excluded in the cell confluency data due to condensation of water on the bottom of the plate.
Data processing and representation. Viability data -Primary screen. For each 384-well plate, viability data of treated cells was normalised to the DMSO control cells measured in the same plate. Normalised viability data for each biological replicate where processed using "R-script-average-of-biological-replicates.R". The average viability score for each technical replicate as well as the average viability score across the biological replicates were calculated and exported as tab-separated values (tsv) files. Standard deviation across biological replicates was calculated using "R-script-calculate-mean-STDV.R". Dose response curves were generated using "R-script-sensitivity-of-drugs-across-cell-lines-plot.R" (Fig. 2, and in figshare deposit 25 : Drug_cell_response. pdf). Drug combination plots were generated from average viability scores with standard deviations using www.nature.com/scientificdata www.nature.com/scientificdata/ "R-script-plotting-drug-combination.R" (figshare deposit 25 : Combination_plots_primary_screen.pdf). Data points with missing values (14%) are represented as NA values.
Viability and confluency data -Secondary screen. For confluency data brightfield images were analysed using the SoftMax ® Pro Software field analysis setting. For this, the region of the image covering the well was selected and confluency was reported. We note that the absolute values for confluency for very flat and long cells (e.g. A498, SF-295) may be underestimated while absolute confluency for round cells (e.g. MDA-MB-498) may be overestimated. We further want to indicate that COLO 205 are growing sub-confluent. For completeness of the data set, confluency is reported for this cell line. For each 384-well plate, viability and confluency data of treated cells was normalised to the plate-internal DMSO control. Technical replicates and averages of biological replicates are reported as comma-separated files (csv). These were further processed using "Biological_Average_secondary_ screen.R" to calculate the biological average, standard deviation and coefficient of variation. Drug combination plots were generation from average viability scores with standard deviations using "Combination_plots_second-ary_screen.R" (figshare deposit 25 : Combination_plots_secondary_screen_viability.pdf). The same script was used to generate drug combination plots for confluency data (figshare deposit 25 : Combination_plots_second-ary_screen_confluency.pdf). In the primary screen statistical significance of the synergistic response was evaluated. For this the computed Bliss excesses for each drug combination-cell line was compared to a zero-centred normal distribution. T-tests were performed to test for statistically significant deviations from a zero-centred distribution. When the null hypothesis was rejected the alternative hypothesis was accepted, which states that the distribution is not zero-centred, i.e. either antagonistic or synergistic. A conservative Bonferroni correction was done to adjust p www.nature.com/scientificdata www.nature.com/scientificdata/ values for multiple hypotheses. Combinations with average Bliss excess values ≤ −0.08 and p-value ≤ 0.05 were classified as synergies. Bliss excess and significance classifications were performed using "p_values_Bliss.R". A synergy score heatmap and volcano plot visualising mean Bliss excess was generated using "R-script-heatmap.R" (Fig. 3a,b). For the secondary screen average Bliss excesses per dose, combination and cell line were computed using "Bliss_synergy_secondary_screen.R". A cumulative Bliss score was calculated using the same script by taking the mean over the computed Bliss excesses per dose. Combinations with average Bliss excess value < 0 were considered as synergies. The lower cut-off was motivated by the fact that five times more combination concentrations were tested in the secondary compared to the primary screen. Example plots of mean Bliss excess ± standard deviation for two combination tested in the secondary screen are illustrated in Fig. 3c. Table 1 lists inhibitor abbreviation, name and unique identifiers (InChiKey and PubChem CID) along with the provider and primary targets as given by provider. CHEBI IDs are provided when available. We provide extensive target profile information in the form of bioassay information extracted from PubChem (last accessed on 24-06-2019) and kinase profiling data from the MRC PPU International Centre for Kinase Profiling -Kinase Profiling Inhibitor Database (http://www.kinase-screen.mrc. ac.uk/kinase-inhibitors, last time accessed 24-06-2019) available in our figshare deposit 25 .

Drug target annotation. Online-only
For the inhibitors SF1670, BI605906 and 10058-F4 no bioassay information supporting targets given by the provider were found in PubChem or MRC PPU Inhibitor Database (proteins inhibited below 10% activity). For these inhibitors, we performed an exhaustive literature search to provide annotated information 25 .

Data records
All data is available at figshare 25 , https://doi.org/10.6084/m9.figshare.9810719. Below, the file names for the different data sets are listed.
Viability and confluency data. The raw luminescence data from the three biological replicates from the primary screen can be found in NTNU_HTS_Repl [1][2][3] Fig. 2, while all plots can be found in Drug_cell_response.pdf 25 . In Fig. 4, the combination effect of selected drug combinations is visualised. Visualised drug combinations are based on significance and strength of synergy (see section for Calculation of synergy and statistical significance of quantified synergy, Fig. 3). Figure 4b shows average viability of SF-295 cells and A498 cells treated with Doramapimod (BI) in combination with PI-103 (PI), and SF-295 cells treated with BI-D1870 (D1) in combination with PI from the primary and secondary screen. The complete file with all drug combination plots can be found in Combination_plots_primary_screen.pdf 25 for the primary screen and in Combination_plots_ secondary_screen_viability.pdf and Combination_plots_secondary_screen_confluency.pdf for the secondary screen for viability and confluency response, respectively 25 .
The raw luminescence data and relative viability data from the three biological replicates from the secondary screen can be found in NTNU_HTS_secondary_viability_REPL [1][2][3].csv 25 . Raw as well as relative confluency data can be found in NTNU_HTS_secondary_confluency_REPL [1][2][3].csv 25 . Mean viability and confluency across the three biological replicates are depicted in NTNU_HTS_secondary_viability_Biological_Average.tsv 25 and NTNU_HTS_secondary_confluency_Biological_Average.tsv 25 , respectively. Synergy data. The Bliss excess values for each combination tested in the primary screen across the eight cell lines are represented in bliss_significance.tsv 25 . Synergistic drug combinations are further presented in bliss_synergies.tsv 25 . A visual representation of Bliss excess is depicted in Fig. 3a. Cell lines were clustered according to Euclidean distances of quantified synergy strengths. 4% synergies were recorded according Bliss Independence model.
Bliss excess values for each combination per dose tested in the secondary across eight cell lines are represented in bliss_synergies_secondary_screen_viability.tsv 25 and bliss_synergies_secondary_screen_confluency.tsv 25 from viability and confluency data, respectively. A visual representation of Bliss excess ± standard deviation from the viability data is shown in bliss_synergies_secondary_screen_plots_viability.pdf 25 and in bliss_synergies_second-ary_screen_plots_confluency.pdf 25 from the confluency data. Cumulative bliss scores for the viability data can be found in Mean.Bliss.excess.matrix.tsv 25 .

Technical Validation
Cell line identity. The seven NCI-cell lines used were received from the NCI-DCTD Repository and cultivated no more than 25 passages after reception in the lab. All cell lines were validated by STR profiling performed by Cell Line Authentication Service at ATCC. Eight core STR loci (TH01, D5S818, D13S317, D7S820, D16S539, CSF1PO, vWA and TPOX) plus amelogenin were matched to database profiles obtained from ATCC and Deutsche Sammlung von Mikroorganismen und Zellkulturen GmbH (DSMZ). For two of the cell lines, SF-295 and UACC-62, no database profile was available for comparison from ATCC and DSMZ. However, both cell lines were found to match their STR profile listed on the ExPASy website (https://web.expasy.org/cellosaurus). All 7 cell lines were found to be identical or similar to their database profiles with percent matches at 95-100%. The AGS cell line was validated by STR profiling performed by the in-house Genomics Core Facility. Eight core STR loci (TH01, D5S818, D13S317, D7S820, D16S539, CSF1PO, vWA and TPOX) plus amelogenin had a match of 81% to the ATCC profile (CRL-1739).
www.nature.com/scientificdata www.nature.com/scientificdata/ Viability assay performance and variation. The performance of the viability assays was monitored using a sample set of small molecule positive controls. In the ray design nine concentrations of digitonin ranging from 100 µg/ml to 0.02 µg/ml was included in each biological replicate. In the matrix design a sample set with three concentrations of the positive controls digitonin, ranging from 26 µg/ml to 3 µg/ml, and staurosporine, ranging from 1.6 µM to 0.18 µM, were included in each plate. The standard deviation of the plate internal reference groups was also monitored and was less than 10% at maximum (on average below 5%) in the ray designed screen. Overall, we observed a mean coefficient of variation of 8% across all biological replicates in the primary screen and 7% in the secondary screen for both the viability and confluency data.
To assess reproducibility between biological replicates, we calculated Pearson correlation coefficients for all measured data points (normalised). We observe a high correlation between all comparisons with a correlation of R = 0.89 between Replicate 1 vs Replicate 2, R = 0.88 between Replicate 1 vs Replicate 3 and R = 0.87 between Replicate 2 vs Replicate 3 of the primary screen (See Fig. 5a-c). In the viability data of the secondary screen we    www.nature.com/scientificdata www.nature.com/scientificdata/ R = 0.95 between Replicate 2 vs Replicate 3. The average viability of shared data points between the primary and secondary screen showed a correlation of R = 0.92 (Fig. 5d). This variation is within the expected variation in high throughput cellular viability assays 11,26 . Synergistic combinations in primary screen. The primary screen led to the identification of 55 synergies according to the chosen cut-off from 1368 total combinations for all cell lines. An analysis of the list showed that 23 unique combinations were classified as synergistic across seven unique cell lines, with no synergistic combination in the MDA-MB-468 cell line. The inhibitors PI (targets PIK3CA/B/D, mTOR and DNA-PK) and SF (target PTEN) were identified to be involved in the majority of synergistic combination. The PI inhibitor was identified as synergistic with BI (target MAPK14) and with D1 (target RPS6KA1/2/3/6) in more than half of the tested cell lines. Further synergistic drug pairs sorted by occurrence across cell lines are G2 (target PDPK1), JN (targets MAPK8/9/10), P5 (target SYK) and AK (targets AKT1, AKT2, AKT3).
Synergistic combinations in secondary screen. As the PI inhibitor was identified as one of the top drugs involved in synergistic combinations, it was chosen as the anchor drug for the secondary screen. We combined the PI inhibitor with the top four synergistic drug pairs judged by occurrence across cell lines. We also included PI in combination with PD (targets MAP2K1/2), to investigate synergistic behaviour of drug concentrations outside the concentration ratios tested in the primary screen. This was motivated by the fact that drugs targeting PI3K and MAP2K1/2 have been previously reported to be synergistic in several cell lines 18,19 . Additionally, we tested PI in combination with 5Z a combination previously identified as synergistic in the AGS cell line in a previous study by our group 15 and Morris et al. 16 . www.nature.com/scientificdata www.nature.com/scientificdata/ Results from the secondary screen confirmed combinations as being synergistic or antagonistic in most cell lines. We confirmed the synergistic effect of PI-G2 in SF-295, SW-620 and UACC-62 cells. Additionally, we observed a synergistic response in A498 and AGS cells in the secondary screen. The combination PI-BI was also redetected as synergistic in five of the cell lines, while for two cell lines the combination did not show a greater effect compared to single drug treatment in the secondary screen. For two combinations, that were not classified as synergistic in the primary screen, PI-5Z and PI-PD, we could observe synergistic responses in the secondary matrix design screen. The PI-5Z combination was synergistic in two cell lines including AGS cells, confirming previous findings 15,16 . The PI-PD combination (targets PIK3CA/B/D, mTOR and DNA-PK -MAP2K1/2) was found to be synergistic in five cell lines and non-synergistic in three cell lines including the two colorectal cancer cell lines SW-620 and COLO 205. While colorectal cancer cell lines have been frequently observed to display synergistic response to combined inhibition of MAP2K1/2 (MEK1/2) and PI3K [27][28][29][30] , weaker synergy has previously been observed in SW-620 and COLO 205 cells 28 .

Usage Notes
Other users may investigate alternative methods for synergy quantification or another threshold using the provided data. To assist other researchers in visualising and analysing the dataset from the primary screen described herein we have prepared files that can be uploaded to the web-service CImbinator (http://rbbt.bsc.es/CImbinator/ CombinationIndex) from which several synergy metrics can be computed and visualised 31 .
All code applied for processing and visualising the data is provided at the figshare deposit 25 .