Drug combinations have emerged as the standards of care for difficult-to-treat conditions including for cancers and infectious diseases. "Once our drugs are targeting parallel pathways, we can overcome resistance mechanisms in a multitude of ways,” said Craig Thomas, a drug discovery chemist at the National Center for Advancing Translational Sciences. The activities of individual drugs can add up when they attack a disease from multiple directions. And drug combinations may show potent synergistic responses that are greater than the sum of their parts, potentially enabling a lower effective dose and a larger therapeutic index.
Thomas’ team uses sophisticated high-throughput screening technology to analyze up to tens of thousands of drug combinations in a single week. Computational models that predict hit drug combinations could help expedite this process. The goal is that “instead of having to search through 10,000 combinations at random, we could search through 10 or 20 or 100 to find the one that's going to work,” said Nicholas Tatonetti, a biomedical data scientist at Columbia University. Developing these models requires training them to recognize effective drug combinations using existing examples, Tatonetti explained. “You'll show it a bunch of pictures of cats, for example. And then after a while, it learns what a cat looks like. But you have to start with those pictures,” he said.
In a recent paper in Scientific Data, other researchers at Ben-Gurion University of the Negev in Israel reported the Continuous Drug Combination Database (CDCDB), a continuously updated database that assembles examples of effective combinatorial drugs to serve as “cat pictures” when developing predictive computational models (1). The researchers wrote a code that automatically retrieves drug combinations under investigation in clinical trials, approved in the FDA Orange Book®, and protected by patents and compiles them all into a single database. The database features the drug names, disease indications, clinical trials, FDA Orange Book® entries, or patents (which may provide details about how the drugs were tested). It also includes identification numbers that can be used to look up a drug’s chemical structure, drug target, and drug-drug interactions, said Guy Shtar, a data scientist and former graduate student at Ben-Gurion University of the Negev who helped develop the CDCDB.
While the database’s ability to synthesize disparate sources of data is impressive, it also presents challenges when comparing the efficacy of drug combinations that have been tested under different conditions, said Thomas. “Combination outcomes are highly contextual. They’re concentration dependent; they're time dependent,” he said. “There is going to be some necessity to make sure we take into account all of those contextual differences from one dataset to the next.”
With efforts to normalize the database for these differences, it could be used alongside public functional genomics datasets like the Broad Institute’s Cancer Dependency Map (DepMap) to better understand complex cellular networks and potentially provide a starting point for combinatorial drug discovery, Thomas said (2). “You have the data coming out of this uniformed combination database, and that tells you that two drugs that target two different enzymes might be combining in a favorable setting,” he said. “And you go to DepMap, and you see that the targets of those drugs are both dependencies for the same disease. Now, you might have a hypothesis that you can build upon to look for new treatments.”
Researchers can locate information such as dosage when it is available in the source material, Shtar said. “We didn't focus on that. But still, we think that the actual fact that these drugs participated together in a clinical trial, even if you don't know the dosage, is a big piece of information to learn from.”
Other researchers agreed. “I do appreciate this paper, and I think I will definitely be using this new database for my own research,” said Jing Tang, a computational biologist at the University of Helsinki who wasn’t involved in the project.
Tang develops models to predict effective drug combinations using the results of high-throughput screening experiments performed in cells. These experiments provide standardized, quantitative datasets that are well-suited for training models. They also include both positive and negative results, allowing the model to better recognize a “cat” by also showing it pictures of something that’s definitely not a cat, like a dog.
Tang sees value in the CDCDB as a source of successful drug combinations to validate the predictions made by machine learning models trained using high-throughput in vitro experimental results. The inclusion of patented drug combinations provides an interesting new type of data, but it’s unclear if they represent examples of effective therapeutics, Tang said. And while drugs investigated in clinical trials also haven’t yet been proven successful, they have passed several tests, meaning “we can be relatively confident that those combinations are showing good promise and may be more likely to be considered a valid combination,” Tang said. “So that’s why we can use this database as a ground truth or validation data. And then we can evaluate those machine learning models to see how likely they can predict these kinds of combinations.”
However, Christian Meyer, a drug synergy and drug-drug interaction researcher at the University of Colorado who wasn’t involved in the research, questions whether the clinical trials listed in the database actually provide effective drug combinations because the majority are in early phases and some have already been terminated. “A caveat that's really critical to consider as you try to build models with this data is to recognize that it does not represent success stories only,” he said.
But information on clinical trials that have been terminated holds its own value, Meyer said. Clinical trial failures due to toxicity “are not well-modeled by the in vitro systems that we tend to try to discover with. This is a huge gap in our current pipeline,” he said. “This [database] would give you a resource to double check and make sure that you did not stumble upon a combination or even a single drug that, when combined with lots of other things, always has a certain toxicity emerge.”
While clinical trials may be terminated for other reasons, Shtar agreed that identifying trends in toxicity is an important application for the CDCDB. “I hope that the combinations that people will find using this database will be both more effective and safer for the patients,” he said.
- Shtar, G., Azulay, L., Nizri, O., Rokach, L., & Shapira, B. CDCDB: A large and continuously updated drug combination database. Sci Data 9, 263 (2022).
- Broad Institute. Explore the Cancer Dependency Map. At <https://depmap.org/portal/>.