Applying AI To Identify Synergistic Drug Combinations For Cancer Treatment

By Tim Sandle, Ph.D.

Synergistic drug combinations refer to a development where two or more drugs are brought together to produce an effect greater than the sum of their individual effects. This approach is beneficial for treating diseases that require a multifaceted approach rather than monotherapy, including cancer or infectious diseases.¹ By combining drugs with different mechanisms of action, medics can target different pathways in cancer cells, potentially leading to improved patient outcomes.²

Assessing which drug combinations might be effective is conventionally time-consuming, and months of work can lead to fruitless outcomes. With n medications, there are n(n−1)/2 possible paired drug combinations.²

Hence, cost, feasibility, and complexity have placed limitations on advances within the field.³ The application of artificial intelligence can reduce these assessment times from months to days or even to hours. The form of AI most successfully applied is deep learning-based methods in terms of predicting synergistic drug combinations.

This article presents some background on the models and highlights four recent examples of AI’s success in selecting drug synergies for cancer treatments.

Technological And Knowledge Basis

The application of AI has been made possible through advances in forms of machine intelligence and the accumulation of large-scale pharmaceutical data sets (of which around a dozen Big Data sets are available to researchers, including the U.S. government’s ClinicalTrials.gov).⁴ The twinning of computational power and data has helped to advance the screening and prioritizing of candidate drug pairs.

Deep learning-based methods are based on different architectures, such as fully connected neural networks (FCNN), which are sets of learned weights and biases; convolutional neural networks (CNNs) seeking local patterns via convolutional operations created through the use of data filters; transformer networks, where knowledge is gained through sequential data analysis; and graph neural networks (GNNs), where graph-structured data is processed in a series of iterative steps.⁵ Each process is capable of extracting patterns from raw data and running models based on complex dimensional data sets.

Multiple models need to be run. Common models include the Loewe additivity model (developed in 1926, this is a model that assumes no self-interaction between drugs and interchangeability within a combination), the Bliss independence model (which assumes the relative effect of each drug is independent of the other), the highest single agents (which calculates the maximum effect of each drug), and zero interaction potency (the concept that two drugs do not potentiate each other).

Prerequisites

Important prerequisites in using a deep learning model so that accurate synergy effects are assessed are detailed drug and cell features. This requires an understanding of biological attributes (such as the inherent characteristics of drugs and cells, including gene expression) and biological knowledge (including drug-target interactions).

Recent Advancements

Recent advances in the use of AI to identify synergistic drug combinations can be illustrated in terms of cancer research. In the first case, researchers from the Mount Sinai School of Medicine have developed a powerful computational tool named iDOMO. This tool was designed to improve the prediction of drug synergy and accelerate the development of combination therapies for complex diseases using gene expression data (a measure of the activity levels of genes in a given biological sample). The model also examines gene signatures (the diverse patterns of gene activity associated with a specific illness, assessing both the disease state and the drug response). The iDOMO platform can rapidly compare and contrast gene signatures of drugs and diseases to predict the beneficial and detrimental effects of differing drug combinations.

To illustrate the usefulness of the platform, the researchers applied iDOMO to examine triple-negative breast cancer. This is a particularly aggressive and difficult-to-treat form of cancer. The platform identified a promising drug combination, bringing together trifluridine and monobenzone. These two drugs in combination were subsequently evaluated in in vitro experiments. The experimental findings confirmed that this combination inhibited triple-negative breast cancer cell growth more effectively than either drug alone, confirming iDOMO’s prediction and paving the way for a new treatment option.⁶

The second example involves a similar application that was undertaken at Daraya University, Egypt. This study revealed that kinase inhibitors combined with mTOR inhibitors, DNA damage-inducing drugs, or HDAC inhibitors displayed anti-cancer effectiveness, particularly against ovarian, melanoma, prostate, lung, and colorectal carcinomas. Specifically, the analysis highlighted that the drugs gemcitabine, MK-8776, and AZD1775 produced frequent patterns in interaction with other drugs across these cancer types.

The Daraya approach was divided into three phases:

1. Processing input data (taken from the O'Neil medication combination database, which consisted of 22,737 medication combinations in 39 cancer cell lines covering seven different tissue types).
2. Classification of the combined drugs in terms of synergism, antagonism, and additive.
3. Output: The prediction of the best combination drugs.

The following forms of AI were deployed: naive Bayes (which measures the probability of each input feature (attribute) for a predictable state); random forests (an ensemble learning method for classification and regression); k nearest neighbor (estimating the nearest neighbor and looking for comparability); and logistic regression (this measures the relationship between the response variable and one or more explanatory variables). This enabled a ranking of different drug combinations to be produced and for the scientists to select from the list the most effective ones.

The third illustration involves the DrugComboExplorer, which weighs different omics data types to analyze cancer pathways (combining single omics with supplementary multi-omics data). Researchers from Cornell University used AI to generate driver signaling networks by processing DNA sequencing, gene copy numbers, DNA methylation, and RNA-seq data from individual cancer patients. This was achieved through an integrated pipeline of algorithms. The algorithms deployed included bootstrap aggregating-based Markov random field (enabling the classification and regression algorithms); weighted co-expression network analysis (which studies biological networks based on pairwise correlations between variables); and supervised regulatory network learning (which predicts the expression of a gene using the expression levels of transcription factors as the features).⁷ These algorithms combined in the form of DrugComboExplorer as a network-based drug efficacy screening tool.⁸

The researchers were able to assess the combinatorial drug efficacies and synergy mechanisms of candidate compounds through drug functional module-induced regulation of target expression analysis. This led to clinical trials to treat diffuse large B-cell lymphoma and prostate cancer; in both cases, the drug combinations selected inhibited multiple driver signaling pathways.

A fourth illustration is seeking cell therapy combinations for cancer treatment. Researchers at Stanford University harnessed AI to create an extreme gradient boosting (XGBoost)-based drug–drug cell line prediction model (XDDC) to predict synergistic drug combinations. This is a machine learning iterative method for finding the roots of a “differentiable function,” which examines the sensitivity to change of individual variables. XDDC looks at drug features (drug chemical structures, adverse drug reactions, and target information); cell therapy features (gene expression, methylation, mutations, copy number variations, and RNA interference data); and links the two together using cell signal pathway information.⁹

Future State

Models are improving at a steady rate, although the range of models available makes selecting the most effective ones difficult (a recent assessment notes over 100 different machine learning base models available for drug synthesis assessment).¹⁰ While models have advanced in terms of accuracy in predicting efficacy, a limitation remains for predicting toxic side effects.

Artificial intelligence will continue to grow in computational power, significantly expanding treatment options for clinicians to improve outcomes for patients, especially those who do not respond to standard therapies and for which drug combinations offer a better solution.

References

1. Kim Y, Zheng S, Tang J. et al. Anticancer drug synergy prediction in understudied tissues using transfer learning. J Am Med Inform Assoc 2021; 28(1): 42-51
2. Alruwaili MM, Zonneville J, Naranjo MN, et al. A synergistic two-drug therapy specifically targets a DNA repair dysregulation that occurs in p53-deficient colorectal and pancreatic cancers. Cell Rep Med. 2024; 5(3):101434
3. DiMasi, J. A., Hansen, R. W. & Grabowski, H. G. The price of innovation: new estimates of drug development costs. J. Health Econ. 2003; 22(2), 151–185
4. Shtar G, Azulay L. Nizri O. et al. CDCDB: a large and continuously updated drug combination database. Sci Data. 2022: 9(1): 263
5. He H, Chen G, Yu-Chian Chen C. 3DGT-DDI: 3D graph and text based neural network for drug–drug interaction pre- diction. Brief Bioinf. 2022; 23(3):bbac134
6. Xianxiao Zhou, Ling Wu, Minghui Wang, Guojun Wu, Bin Zhang. iDOMO: identification of drug combinations via multi-set operations for treating diseases. Briefings in Bioinformatics, 2025; 26 (1) DOI: 10.1093/bib/bbaf054
7. Huang, L., Brunell, D., Stephan, C. et al. Driver network as a biomarker: systematic integration and network modeling of multi-omics data to derive driver signaling pathways for drug combination prediction, Bioinformatics, 2019; 35 (19): 3709–3717
8. Huang L, Brunell D, Stephan C, et al. Driver network as a biomarker: systematic integration and network modeling of multi-omics data to derive driver signaling pathways for drug combination prediction. Bioinformatics. 2019; 35(19):3709-3717
9. Chen J, Han H, Li L, et al.. Prediction of cancer cell line-specific synergistic drug combinations based on multi-omics data. PeerJ 2025; 13:e19078
10. Kong, W. et al. Systematic review of computational methods for drug combination prediction. Comput. Struct. Biotechnol. J. 2022; 20, 2807–2814

About The Author:

Tim Sandle, Ph.D., is a pharmaceutical professional with wide experience in microbiology and quality assurance. He is the author of more than 30 books relating to pharmaceuticals, healthcare, and life sciences, as well as over 170 peer-reviewed papers and some 500 technical articles. Sandle has presented at over 200 events and he currently works at Bio Products Laboratory Ltd. (BPL), and he is a visiting professor at the University of Manchester and University College London, as well as a consultant to the pharmaceutical industry. Visit his microbiology website at https://www.pharmamicroresources.com.