AI-driven discovery of synergistic drug combinations against pancreatic cancer

Abstract

Pancreatic cancer treatment often relies on multi-drug regimens, but optimal combinations remain elusive. This study evaluates predictive approaches to identify synergistic drug combinations using a dataset from the National Center for Advancing Translational Sciences (NCATS). Screening 496 combinations of 32 anticancer compounds against the PANC-1 cells experimentally determined the degree of synergism and antagonism. Three research groups (NCATS, University of North Carolina, and Massachusetts Institute of Technology) leverage these data to apply machine learning (ML) approaches, predicting synergy across 1.6 million combinations. Of the 88 tested, 51 show synergy, with graph convolutional networks achieving the best hit rate and random forest the highest precision. Beyond highlighting the potential of ML, this work delivers 307 experimentally validated synergistic combinations, demonstrating its practical impact in treating pancreatic cancer.

Fig. 1. Workflow illustrating the combination of computation and experiment to discover synergistic drug combinations.

High-throughput screening of 1785 single-agent compounds from the MIPE4 library in PANC-1 cells identified 32 active compounds. All-vs.-all combinations of these 32 compounds generated synergy data for 496 combinations. The synergy data served as a training dataset for three machine learning teams. NCATS used Random Forest (RF), XGBoost, and Deep Neural Network (DNN); UNC used RF, gradient boosting, DNN and graph convolutional networks (GCN); MIT used GCN. Training features included molecular descriptors, such as Avalon and Morgan fingerprints, RDKit descriptors, and biological features. Models predicted synergies across 1.6 million virtual drug combinations, and the top 30 combinations from each team were experimentally tested in PANC-1 cells, achieving an average 60% hit rate.

Fig. 2. Analysis of the variation and reproducibility of PANC-1 training dataset.

a Variation of activities of 32 compounds in single-agent dose−response curves. Log(IC₅₀) varies between −8.7 and −5.5 ( ≈ 2 nM−3 µM). b Reproducibility of experiments. Gamma values of 496 combinations in two replicates have a Pearson’s coefficient of 0.83. Combinations with Gamma <0.95 are considered synergistic. Similar plots for other synergy metrics are in Supplementary Fig. 1. Source data are provided as a Source Data file.

Fig. 3. NCATS’ random forest models to predict Gamma.

Training set included 496 combinations constituted from 32 drugs (all vs. all). Cross validation strategy was one-compound-out, hence 32 folds. Average of Avalon 2048 fingerprints of each pair featurized the combination as a single 2048-dimensional vector. a Regression. The plot overlays predictions for 32 validation sets (withheld within each of 32 folds) with a Pearson’s coefficient of 0.52. b Classification. Model yields 32 ROC plots (not shown for clarity) with average AUC and standard deviation of 0.78 ± 0.09. Blue line, average ROC plot; grey area, ±1 standard deviation; red dashed line, random prediction baseline; Gamma < 0.95, synergistic; Gamma ≥ 0.95, non-synergistic. Similar plots for Morgan 2048 and RDKit descriptors are in Supplementary Figs. 2 and 3. Source data are provided as a Source Data file.

Fig. 4. Predicted synergy of unseen combinations by three independent models.

a NCATS, (b) UNC, (c) MIT. Each institute nominated 30 combinations, resulting in 60 synergy values per a symmetric matrix. Voxels represent scores between compounds on Compound 1 and Compound 2 axes. Color bars represent ranges of different synergy scores adopted by each team. Compound names, excluded for clarity, are accessible in the source data, provided as a Source Data file.

Fig. 5. Distribution of Gamma for 307 experimentally-validated synergistic combinations.

Most scores are close to 0.95. 26 combinations exhibit strong synergy with scores below 0.5. Source data are provided as a Source Data file.

Fig. 6. Number of occurrence of compounds and MoAs among 307 experimentally-validated synergistic combinations colored by average synergy.

a Compounds, (b) MoA. Each MoA in (b) is abbreviated for simplicity; for example, “HDAC” refers to “HDAC Inhibitor”, where HDAC is a protein name. Supplementary Fig. 8 provides details of abbreviations. Source data are provided as a Source Data file.

Fig. 7. Network analysis of MoAs in 26 strongly synergistic combinations (gamma < 0.5).

Node sizes and edge widths are proportional to MoA and MoA–MoA frequency, respectively. Each MoA is abbreviated for simplicity; for example, “HDAC” refers to “HDAC Inhibitor”, where HDAC is a protein name. Supplementary Fig. 8 provides details of abbreviations. Source data are provided as a Source Data file.