Computationally predicting clinical drug combination efficacy with cancer cell line screens and independent drug action

Abstract

Evidence has recently emerged that many clinical cancer drug combinations may derive their efficacy from independent drug action (IDA), where patients only receive benefit from the single most effective drug in a drug combination. Here we present IDACombo, an IDA based method to predict the efficacy of drug combinations using monotherapy data from high-throughput cancer cell line screens. We show that IDACombo predictions closely agree with measured drug combination efficacies both in vitro (Pearson's correlation = 0.93 when comparing predicted efficacies to measured efficacies for >5000 combinations) and in a systematically selected set of clinical trials (accuracy > 84% for predicting statistically significant improvements in patient outcomes for 26 first line therapy trials). Finally, we demonstrate how IDACombo can be used to systematically prioritize combinations for development in specific cancer settings, providing a framework for quickly translating existing monotherapy cell line data into clinically meaningful predictions of drug combination efficacy.

Fig. 1. IDACombo allows drug combination efficacy predictions to be made using monotherapy cell line screening data, and these predictions can be validated against measured efficacies or used to identify novel efficacious drug combinations.

a Example calculations demonstrating how IDACombo predicts drug combination efficacies based on IDA. In this example, three cell lines (1–3) with measured efficacies for three monotherapies (A–C) at their selected concentrations are used to predict the efficacy of the combination of drugs A + B + C. Highlighted cells indicate the best monotherapy for that cell line (i.e. provides greatest reduction in viability). b Strategy for validating IDACombo efficacy predictions using in vitro measurements of efficacy. Measured and predicted average efficacies for each treatment can be directly compared by calculating their correlation and calculating prediction errors. c Strategy for validating IDACombo efficacy predictions using published clinical trial results. Combination efficacies which have been predicted using cell line data are used to predict study HRs, and these predicted HRs are used, along with the number of events observed in each clinical trial, to predict study powers. Predicted HRs can be compared to reported HRs, and a power threshold (80%) can be set to classify trials as likely or unlikely to detect a significant improvement in a trial outcome (i.e. PFS), which can then be compared to observed trial outcomes. d Analysis techniques available for using IDACombo predictions to identify novel efficacious drug combinations. High-throughput analyses using summary statistics can be used to compare efficacy predictions for many drug combinations at once, or detailed analyses can be used to explore the efficacy of a single drug combination at varying concentrations of each drug in the combination.

Fig. 2. Agreement between predicted and observed combination viabilities in NCI-ALMANAC.

a Scatterplot showing high correlation between predicted average percent viability and experimentally observed average percent viability for each drug combination in NCI-ALMANAC. Predictions were made using monotherapy data from the dataset. The green line is a reference diagonal with slope = 1 and intercept = 0. Note that predictions were only made for the maximum concentration tested for each drug. b Density plot 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 Density plot 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. Source data are provided as a Source Data file.

Fig. 3. Trial selection pipeline for clinical validation.

Flowchart detailing how completed, phase III cancer clinical trials were selected for the clinical trial validation analysis. Searches of ClinicalTrials.gov and PubMed.gov were performed via web scraping (see “Methods” section) to identify published results for trials that may meet our inclusion criteria, and the identified clinical trial publications were then manually inspected to identify trials that met our study’s inclusion criteria.

Fig. 4. Clinical trial validation results show accurate efficacy predictions for trials in previously untreated patients but not for trials in previously treated patients.

IDACombo was used to make efficacy predictions for the control and experimental treatments of the clinical trials selected using the pipeline in Fig. 3. Hazard ratios were then calculated using these predictions, and study powers were calculated for each available comparison of a control therapy vs. an experimental therapy. These comparisons are separated based on whether or not the experimental arm statistically improved either PFS/TTP (panels a and c) or OS (panels b and d) in the published trial results. Predicted powers for each comparison are plotted on the y-axes, and an 80% power threshold (dashed line) is used to classify whether or not a comparison is expected to yield a statistically significant improvement. Comparisons are colored according to the dataset used to make predictions for the compared treatments. Panels a and b show results for trials in which patients had received no previous drug treatments. Panels c and d show results for trials in patients who had received previous treatment. Error bars for each plotted clinical trial power represent mean estimated power ± standard error (bounded between 0% and 100% power). Gray and orange arrows in Panel a indicate misclassified trials that are discussed in the text. P values were calculated using one-tailed t-tests. Blue circles indicate predictions made using the CTRP dataset, and red circles indicate predictions made using the GDSC dataset. Boxplots are plotted so that the lower and upper whiskers indicate the extreme lower and upper values, respectively, the box boundaries indicate the first and third quartiles, and the center line indicates the median. Source data are provided as a Source Data file.

Fig. 5. Top IDAcomboscore predictions for late-stage clinical drugs in CTRPv2.

IDAcomboscores were calculated for all two-drug combinations of late-stage clinical drugs in CTRPv2 using all available cell lines for each drug combination. Darker blue squares represent higher comboscores and, therefore, greater predicted drug combination efficacies relative to the constituent monotherapies. Black boxes represent missing values, where efficacies could not be predicted for a combination. The first bar, farthest left on the right side of the heatmap, indicates whether or not that drug is currently used for cancer treatment, the second bar indicates what stage of clinical trials that drug has reached, and the third bar indicates if the known Csustained concentration for the drug was higher than the maximum tested concentration in CTRPv2 such that predictions had to be made with a lower than clinical concentration. The barplot on top of the heatmap indicates the average viability achieved using that drug as a monotherapy (full bar indicates 100% viability). Drugs that were not predicted to combine well with any other drugs (i.e. with a comboscore < 0.004) were excluded from this plot to improve readability, but a full heatmap with all late-stage clinical drugs can be found in the “IDACombo Paper” project on OSF (see “Methods” section). Colored outlines are used to highlight certain combinations according to the “Combinations of” legend in the bottom right. Source data are provided as a Source Data file.

Fig. 6. IDA-Combo predicts strong benefits for combinations of navitoclax and taxanes in EGFR-WT lung cancer.

a An ordered bar plot of the IDA-comboscores predicted using EGFR-WT lung cancer cell lines 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 and c 3-D plots of measured and predicted average cell viabilities at different concentrations of navitoclax and docetaxel b or paclitaxel c. The gray plane represents the lowest average viability achievable with monotherapy. The red arrow represents the difference between the best observed monotherapy effect and the best predicted combination effect, which suggests that the combination therapy will reduce tumor cell viability below what is achievable with monotherapy alone. Source data are provided as a Source Data file.