Clinical trial outcomes for cancer drug combinations can be predicted using cancer cell line monotherapy screens and a model of independent drug action Alexander Ling and R. Stephanie Huang Experimental and Clinical Pharmacology, University of Minnesota MSI Research Exhibition, 2020

Introduction In vitro validation of IDACombo Clinical validation of IDACombo In many cancer settings, it is evident that combination drug therapies show We utilized three high-throughput cancer cell line drug combination screens to Figure 4. Trial selection Figure 5. Clinical trial validation results increased efficacy compared to monotherapies. However, it is infeasible to validate whether or not IDACombo’s predicted drug combination efficacies match pipeline for clinical show accurate efficacy predictions for trials experimentally evaluate the vast number of possible drug combinations when with measured drug combination efficacies. For each screen, IDACombo was validation. Flowchart in previously untreated patients but not for designing new therapies. To overcome this problem, many computational used to predict drug combination efficacies using the monotherapy data from the detailing how completed, trials in previously treated patients. methods have been developed to estimate drug combination efficacy prior to screen, and then these predictions were compared to the measured drug phase III cancer clinical trials IDACombo’s predicted powers for PFS/TTP experimental testing using existing pre-clinical datasets. These methods have combination efficacies also available in each screen (Figure 2). were selected for the clinical correctly classify 88.5% of clinical trials in which focused on estimating drug synergy and additivity, but they have so far shown trial validation analysis. patients had not received cancer drug treatment limited potential to be translated to independent pre-clinical datasets or to clinical NCI-ALMANAC7 Searches of ClinicalTrials.gov prior to trial entry (Figure 5A), with >85% drug development.1,2 A B C and PubMed.gov were sensitivity and specificity. For OS powers in performed via web scraping treatment-naïve trials (Figure 5B), accuracy, Recent evidence has suggested that independent drug action, where the effect of to identify published results sensitivity, and specificity were >90%, but it is a drug combination is equal to the effect of the single most effective drug in the for trials that may meet our difficult to confidently assess the suitability of combination with no synergy or additivity, may explain the effectiveness of many inclusion criteria, and the IDACombo for predicting OS benefit, because 3 clinical cancer drug combinations. In light of these findings, we developed identified clinical trial we only identified 3 clinical trials in treatment- IDACombo, a method which estimates drug combination efficacy based on publications were then naïve patients which detected a statistically independent drug action using the GDSC and CTRPv2 monotherapy cancer cell manually inspected to identify significant improvement in OS. Unfortunately, line screen datasets4,5. AstraZeneca-Sanger Drug Combination DREAM Challenge8 trials that met our study’s the model performed much more poorly for Using resources from the Minnesota Supercomputing Institute (MSI), we have D E F inclusion criteria. clinical trials in patients who had undergone validated our method both against measured drug combination efficacies from cancer drug treatment prior to entering the trial cancer cell line screens and against published clinical trial results for cancer drug (Figures 5C and 5D). combinations. We have also generated prospective predictions of efficacy for thousands of drug combinations to aid researchers in identifying new candidates for clinical development. Prospective analyses with IDACombo

Purpose of this work O’Neil et al., 20169 IDACombo was used to predict drug combination efficacies for thousands A more focused analysis was also performed to identify candidate drugs G H I of 2-drug combinations consisting of pairs of clinically advanced drugs for to combine with navitoclax in EGFR-WT lung cancer. This analysis Our work seeks to address two challenges in the field of cancer drug combination which monotherapy data is available in CTRPv2 or GDSC. These suggests that navitoclax can be efficaciously combined with taxanes (i.e. development: predictions are plotted in a heatmap in Figure 6. The clear clustering in docetaxel or paclitaxel) (Figure 7A). Notably, IDACombo predicts that this heatmap indicates that drugs which have the same mechanism of this combination will have greater efficacy than either monotherapy even Challenge 1: It is infeasible to experimentally test all possible drug action are not predicted to combine well together via IDA. if the concentrations of each drug in the combination must be reduced to combinations allow them to be used in combination (Figures 7B & 7C). • Example: Pre-Clinical testing with 100 drugs in 1000 cell lines

• Monotherapy Figure 2. Agreement between predicted and observed combination viabilities in three 100 drugs x 1000 cell lines = 100,000 experiments published cancer cell line drug combination screens. • 2-Drug Combos A-C) Results using the NCI-ALMANAC drug combination dataset, which tested ~5000 drug 100! combinations in 60 cell lines. D-F) Results using the AstraZeneca-Sanger DREAM challenge 푐표푚푏푖푛푎푡푖표푛푠 × 1000 푐푒푙푙 푙푖푛푒푠 = 4,950,000 experiments 2!(98!) drug combination dataset, which tested ~800 drug combinations in up to 49 cell lines (median of 13 cell lines per combination). G-I) Results using the O’Neil et al., 2016 drug combination • 3-Drug Combos dataset, which tested ~600 drug combinations in 39 cell lines. 100! 푐표푚푏푖푛푎푡푖표푛푠 × 1000 푐푒푙푙 푙푖푛푒푠 = 161,700,000 experiments A,D,G) Scatterplots showing high correlation between predicted average percent viability and 3!(97!) experimentally observed average percent viability for each drug combination in the dataset. Challenge 2: Even with pre-clinical testing, most oncology clinical trials do Predictions were made using monotherapy data from the dataset. The green line is a not currently lead to FDA approval6 reference diagonal with slope = 1 and intercept = 0. Note that predictions were only made for the maximum concentration tested for each drug. B,E,H) Density plots showing that the absolute values of the differences between the predicted percent viabilities and the observed percent viabilities for each drug combination are generally below 10%, with >50% of drug combinations having an absolute prediction error below 5%. The red line marks a difference of ±10% viability between predicted and observed values. C,F,I) Density plots showing that the differences between the predicted percent viabilities and the observed percent viabilities for each drug combination have a slight tendency towards being positive—indicating that IDA- Combo underestimates efficacy more often than it overestimates efficacies.

Figure 7. IDA-Combo predicts strong benefits for combinations of navitoclax and taxanes. A) An ordered bar plot of the IDA-comboscores predicted for combinations of navitoclax with other drugs that have reached late-stage clinical trials. Each bar represents a different combination of navitoclax with another drug. B & C) 3-D plots of measured and Strategy to validate IDACombo’s clinical utility Figure 6. Top IDAcomboscore predictions for late-stage clinical drugs in CTRPv2. predicted average cell viabilities at different concentrations of navitoclax and docetaxel (B) Heatmap of predicted IDAComboscores for 2-drug combinations of clinical drugs in Hypothesis or paclitaxel (C). The transparent plane represents the lowest average viability achievable We used published results from phase III clinical trials to validate the ability of CTRPv2. Higher comboscores (darker blue) indicate higher predicted efficacy. Black with monotherapy. The red arrow represents the difference between the best observed squares represent missing values. Notably, these predictions show clear clustering at least We hypothesized that it is possible to create a computational method capable of IDACombo to predict drug combination efficacy in the clinical setting. monotherapy effect and the best predicted combination effect, which suggests that the in part due to drug mechanism of action. Note: Only drugs with at least one comboscore using data from monotherapy cancer cell line screens to predict the clinical First, raw drug screening data from the CTRPv2 and GDSC monotherapy combination therapy will reduce tumor cell viability below what is achievable with >0.004 are shown here. efficacy of untested drug combinations. cancer cell line screens was processed using Minnesota Supercomputing monotherapy alone. Resources to obtain dose-response curve for ~800,000 monotherapy drug screening experiments (Figure 3A). This allowed us to harmonize the two IDACombo can also be used to make Figure 8. IDACombo predicts that elesclomol will efficaciously combine with datasets so that they could both be used as input data for IDACombo in our predictions for combinations of 3+ cisplatin+gemcitabine in EGFR WT lung cancer. A) IDAcomboscores were calculated for analysis. drugs. Figure 8 demonstrates that the addition of late-stage clinical drugs in GDSC at their clinical concentrations to the control elesclomol, an inducer of oxidative treatment combination of Cisplatin (6.44μM) + Gemcitabine (1.14µM) in EGFR WT lung Next, ClinicalTrials.gov and PubMed.gov were systematically searched for stress, is predicted to efficaciously cancer. Only the top 20 IDAcomboscores are plotted here. B) Predicted IDAcomboscores phase III clinical trials that tested drug combinations that we could predict combine with cisplatin + gemcitabine for the addition of elesclomol to the combination of Cisplatin (6.44μM) + Gemcitabine efficacy for using IDACombo and the CTRPv2 and GSDC datasets (Figures 3B in EGFR-WT lung cancer. As with (1.14µM) across a range of concentrations of elesclomol in EGFR WT lung cancer. C) and 4). The results of these trials were then compared to study power Maximum predicted hazard ratios for the addition of elesclomol to combination of Cisplatin Figure 7, the predicted increase in predictions made using the published study sizes and IDACombo’s efficacy (6.44μM) + Gemcitabine (1.14µM) in EGFR WT lung cancer across a range of efficacy is maintained even if predictions (Figures 3B and 5). concentrations of elesclomol. Maximum hazard ratio is defined as the higher hazard ratio elesclomol concentrations must be (i.e. the hazard ratio that indicates less efficacy improvement from the test treatment vs the Figure 3. Strategy to validate IDACombo’s clinical utility reduced below its maximal clinical control treatment) of either: 1. elsclomol+cisplatin+gemcitabine vs cisplatin+gemcitabine or How the IDACombo algorithm works concentration for use in the 2. elesclomol+cisplatin+gemcitabine vs elesclomol monotherapy. A Data Pre-Processing for Monotherapy Cell Line Screens combination. We designed IDACombo to work on the principle of independent drug action Fit dose-response curves to ~800,000 (IDA), predicting that the efficacy of a drug combination in a given cell line or Harmonized CTRPv2 Raw Screening experiments using the Mesabi compute and GDSC Data from CTRPv2 cluster at the Minnesota Supercomputing patient will be equal to the effect of the single best drug in that combination. Monotherapy Efficacy and GDSC Datasets Institute to obtain harmonized screening Conclusions Data References Importantly, IDACombo predictions are concentration dependent, which allows us results from both datasets 1. Menden, M. P. et al. Community assessment to advance computational prediction of cancer drug combinations in a pharmacogenomic screen. Nat. to predict combination efficacy specifically when each drug is used at its clinically Commun. 10, 2674 (2019). These results demonstrate that IDACombo can be used with monotherapy cell relevant concentration. Furthermore, predictions represent an average response 2. Madani Tonekaboni, S. A., Soltan Ghoraie, L., Manem, V. S. K. & Haibe-Kains, B. Predictive approaches for drug combination discovery in cancer. Brief. line screening data to accurately predict drug combination efficacy both in vitro Bioinform. 19, 263–276 (2016). across populations of cell-lines/patients, which mimics the way treatment and in previously untreated patients. This provides a framework for translating B Clinical Validation Strategy 3. Palmer, A. C. & Sorger, P. K. Combination Cancer Therapy Can Confer Benefit via Patient-to-Patient Variability without Drug Additivity or Synergy. Cell efficacies are measured in clinical trials. 171, 1678-1691.e13 (2017). monotherapy cell line data into clinically meaningful predictions of drug 4. Seashore-Ludlow, B. et al. Harnessing Connectivity in a Large-Scale Small-Molecule Sensitivity Dataset. Cancer Discov. 5, 1210–1223 (2015). ClinicalTrials.gov 1. Identify published phase III clinical Figure 1. Example drug combo efficacy prediction using cell line data and IDA. trial results for drug combinations in Published Clinical Trial 5. Iorio, F. et al. A Landscape of Pharmacogenomic Interactions in Cancer. Cell 166, 740–754 (2016). combination efficacy. Critically, while it is currently infeasible to experimentally cancer. Results 6. Wong, C. H., Siah, K. W. & Lo, A. W. Estimation of clinical trial success rates and related parameters. Biostatistics 20, 237–286 (2018). test the vast number of possible cancer drug combinations, the algorithmic PubMed.gov See Figure 4 for details. 7. Holbeck, S. L. et al. The National Cancer Institute ALMANAC: A Comprehensive Screening Resource for the Detection of Anticancer Drug Pairs with Enhanced Therapeutic Activity. Cancer Res. 77, 3564–3576 (2017). simplicity of IDACombo could allow researchers to computationally predict the 5. Compare 8. Menden, M. P. et al. Community assessment to advance computational prediction of cancer drug combinations in a pharmacogenomic screen. Nat. 4. Use predicted efficacies from Predicted 2. Search literature Clinical Plasma clinical trial Commun. 10, 2674 (2019). efficacies of hundreds of millions of drug combinations in a matter of weeks to IDACombo and study size information Clinical for clinical plasma Drug findings with from published trials to predict trial Trial 9. O’Neil, J. et al. An Unbiased Oncology Compound Screen to Identify Novel Combination Strategies. Mol. Cancer Ther. 15, 1155–1162 drug concentrations. Concentrations predicted powers months. This enables data driven drug combination selection and could power. Powers (2016). for each trial. significantly speed up the rate of novel cancer drug combination discovery.

Harmonized CTRPv2 3. Predict efficacy of clinical drug and GDSC combinations using CTRPv2 and GDSC Predicted Clinical Drug Acknowledgements IDACombo has been released as an R package, and it can be accessed at Monotherapy monotherapy information and the IDA- Combination Efficacies Figure 5 Efficacy Data Combo algorithm. https://github.com/Alexander-Ling/IDACombo. We thank the Minnesota Supercomputing Institute (MSI, http://www.msi.umn.edu) at the University of Minnesota for providing resources that contributed to the research results reported in this poster. This study was supported by an NIH/NCI Grant 1R01CA204856-01A1. R.S.H. also receives support from a research grant from the Avon Foundation for Women and an OACA Faculty Research Development grant.