As 1998 drew to a close, researchers at Merck were excited about their potential new painkiller, a drug called Vioxx. It belonged to a class of molecules called COX-2 inhibitors. While traditional nonsteroidal anti-inflammatory drugs (NSAIDs) including ibuprofen and naproxen inhibited both COX-1 and COX-2 enzymes to decrease inflammation and reduce pain, Vioxx only targeted COX-2. Because of its increased specificity, researchers thought that it would likely match the pain relief provided by traditional NSAIDs without some of their severe gastrointestinal side effects, such as increased risk of bleeding, stomach tears, and ulcers.
Instead, Merck and the FDA, which approved Vioxx in 1999, soon realized that Vioxx had an even more dangerous side effect: It significantly increased a person’s risk for heart attacks. In 2004, Merck withdrew Vioxx from the market, but not before at least 88,000 Americans experienced drug-related heart attacks and approximately 38,000 of them died (1).
“It is a real good lesson in how little we know about pharmacology. It seems like a really advanced field, and in some aspects, it really is. But the body is so much more complex than the drugs that we've designed are,” said Nicholas Tatonetti, a bioscientist at Cedars-Sinai Medical Center who uses data science to investigate drug interactions and side effects.
Using spatial and single cell tools to get down to the extreme specifics about what drugs are doing in the body and how to design really specific drugs has a real potential for great efficacy and safety.
- Nicholas Tatonetti, Cedars-Sinai Medical Center
One way that scientists can better understand the human body and all of its intricacies is to profile every single cell in it. With recent advances in systems biology and single cell technology, this kind of analysis is now possible. To create a map of the human body with single cell resolution, Sarah Teichmann, a biologist at the Wellcome Sanger Institute, and her fellow computational biologist colleague Aviv Regev at the Broad Institute founded the collaborative international consortium the Human Cell Atlas. Using single cell and spatial computational analysis techniques, researchers at institutions around the world are working their way through every tissue and organ system to create a cellular map of the body.
“Mapping the healthy human body really has a very close relationship to also understanding human disease because the healthy Human Cell Atlas provides that reference framework for what changes in disease,” said Teichmann. But as she and her team began profiling cells in different tissues, she wondered if they could use this single cell data to identify how a particular drug acts on different kinds of cells and to find drugs that act on one specific cell type.
Now, Teichmann and her team have developed a new computational tool called drug2cell to help researchers answer those questions. As a Python package, drug2cell takes single cell transcriptomics data and integrates it with a drug-target database to find cells targeted by a specific drug, drugs that target a specific cell type, and even molecules expressed in a particular cell that may interact with a drug. With this new tool, drug discovery and development researchers can predict new drug targets, identify potential therapeutic candidates for repurposing, or identify side effects that their treatment may cause.
“Things like this are really exciting,” said Tatonetti, who was not involved in developing drug2cell. “Using spatial and single cell tools to get down to the extreme specifics about what drugs are doing in the body and how to design really specific drugs has a real potential for great efficacy and safety.”
To the heart of it
Teichmann’s idea for drug2cell sparked when she and her team were profiling cells from heart donors to create a human heart cell atlas. At the time, Kazumasa Kanemaru, a clinician and postdoctoral researcher, and James Cranley, a cardiology clinician and PhD student led the effort. With his clinical interest in heart rhythm problems, Cranley was particularly interested in characterizing the cells of the cardiac conduction system, the electrical system in the heart that controls the heartbeat.
To form a detailed picture of the heart, they combined their previously published single cell and single nucleotide (sn) RNA sequencing data of the heart with paired snRNA sequencing plus sn transposase-accessible chromatin using sequencing (snATAC-seq) and spatial transcriptomics (2,3). With samples from 25 donors in an age range of 20 to 75 years, the researchers profiled cells from eight regions of the heart and discovered multiple cellular niches within those regions.

As they examined the gene expression information from the different cells in these niches, one cell type in particular jumped out as having quite a unique profile: the pacemaker cells, which control heart rate.
“We could see that there was a very rich repertoire of G-protein coupled receptors on pacemaker cells, and they had quite a specific expression profile compared to other cell types,” said Cranley. In fact, G-protein coupled receptors (GPCR) help regulate heart rate, and Tatonetti explained that because many GPCR have cell type specific expression, drugs that target particular GPCR are some of the safest ones (4).
Soon after Cranley and Kanemaru made this discovery about the pacemaker cells, Teichmann headed to Oxford University to give a presentation about her group’s cardiac cell atlas research. Kanemaru and Cranley gave her some slides with their pacemaker cell expression data to include in her talk.
“I did a heat map of different expression of different GPCR, and someone in the audience noticed that the GLP-1 [glucagon-like peptide-1] receptor was enriched in the pacemaker cells. And they said, ‘Oh, that's quite interesting,’” said Cranley. At the time of Teichmann’s presentation, he added, “there were advanced trials with GLP-1 agonists, and I think there had been this signal noted that there was a slight increase in heart rate in some of the people in these trials” (5).
Because the cardiac conduction system, and pacemaker cells specifically, are vital for maintaining a proper heart rate, Teichmann wondered how cardiac and other drugs affect these heart cells.
“One day, [Teichmann] came to us and said, ‘there's a database called ChEMBL, which is this drug and target pair database. Is there any way somehow we could use this using our single cell data?’” said Kanemaru. “Then we started developing this package.”
More targets, new mechanisms
Other tools that integrate drug-target information with expression data rely on RNA sequencing data without any single cell information or RNA sequencing of individual cell lines in vitro, which lack many important features and cues present in vivo. By using drug2cell, users can get that much needed cellular level resolution.
“That's really the special, novel thing about it,” said Teichmann. “It gives you that higher precision. Because then even cells that are quite rare or sparse in a bulk data set, that can be resolved at the single cell level, and then if they are targeted, you'll see it with drug2cell.”
Tatonetti agreed: “As far as the chemical informatics types of work they're doing or the drug data science that they're doing, these all are really standard methods, and I think that's good because the real innovation here is the way to use spatial and single cell transcriptomics data to inform pharmacology. That's at a level of precision that we barely ever see.”
Teichmann and her team designed drug2cell to use the ChEMBL database because it is open source, and having formerly worked at the European Bioinformatics Institute which hosts the database, Techimann was familiar with it. But she added that any compound-target database will work with drug2cell, including in-house or proprietary databases.
To use drug2cell, researchers input their in vivo single cell RNA sequencing dataset of interest, and drug2cell will apply the ChEMBL database of drug-target pairs to it. The user can filter the drugs in the ChEMBL database by a number of factors such as clinical trial phase of the drug and target molecule class. Researchers can also filter by a drug’s category as defined by its Anatomical Therapeutic Chemical (ATC) classification, which delineates the organ or system the drug acts on. They can also filter on what ChEMBL calls bioactivity, which is a measure of the level of confidence in the evidence for the drug-target relationship: for example, if the drug-target relationship was validated via an in vitro binding assay, a docking prediction calculation, or other functional assay.
From there, drug2cell calculates a “drug score” for each cell and drug queried. Researchers can then use the drug scores for the following applications: 1) find cells targeted by specific drugs, 2) find drugs that target specific cells, and 3) find target molecules expressed by cells targeted by a specific drug.
When Teichmann, Kanemaru, Cranley, and their colleagues used drug2cell on their cardiac cell dataset, they filtered for drugs in ChEMBL that were clinically approved and present in all ATC categories. They then looked for the drugs that had the strongest predicted effect on pacemaker cells compared to all other cardiac cell types.

As expected, they identified cardiac drugs including ivabradine, which slows the heart rate, and quinidine, an antiarrhythmic drug. But they also identified noncardiac drugs including the epilepsy medication perampanel and the diabetes drug liraglutide, which is a GLP-1 analogue that hearkened back to the high expression of the GLP-1 receptor identified in the team’s single cell analysis of pacemaker cells. Researchers knew that both perampanel and liraglutide affected heart rate, but the mechanisms by which they did that were unknown.
“There's actually a mini debate in the field about what the mechanism of this is,” said Cranley. “The review articles debate whether the drugs are acting on the autonomic nervous system and affecting the heart indirectly, or whether it's a direct effect on the heart.”
To determine whether liraglutide, the GLP-1 analogue, did indeed act directly on the heart to change heart rate, the researchers treated cardiomyocytes derived from human induced pluripotent stem cells with liraglutide. They also treated one batch of these cells with ivabradine to serve as a negative control. Ivabradine slowed the beating rate of the cardiomyocytes while liraglutide increased it, suggesting that GLP-1 analogues act directly on pacemaker cells to affect heart rate.
“The nice thing about the unbiased profiling and mapping of the cells in the human body is that you can make these unexpected discoveries that are hypothesis free in a sense, where you're just using data science to look for all relationships in a completely unbiased way and then discover things that you didn't expect,” said Teichmann.
Teichmann sees two major applications of drug2cell: one for the healthy human cell atlas and another for cell atlases of disease states. With healthy human cell atlases, researchers can look for side effects of their drug of interest on all cell types in the atlas. This feature is something that Tatonetti is particularly excited about.
“This one is a great example [of a] landmark paper, that the types of methods that we've been advocating for to study drug safety are valid,” said Tatonetti. “I'm going to use [it as] more of an example of the types of experiments that I'm going to argue my collaborators should be doing so that we can get more data in different tissues. We can study liver toxicities, which are super critical. Kidney toxicity is super critical, muscle toxicities, so really go above and beyond just the cardiovascular system.”
On the other side of the coin, as more research groups collect single cell transcriptomic data in disease states as diverse as SARS-CoV-2 infections, inflammatory diseases, and cancer, drug2cell can help researchers find drugs that could be repurposed for a specific disease.
Teichmann has already heard from researchers interested in using drug2cell, and she can’t wait to have them try it out for themselves.
“We had a lot of excitement around the new insights and understanding phenomena now at the molecular mechanistic level in terms of in the cellular context,” said Teichmann. “But there's also a satisfaction in developing a tool that's useful to the community and having other people pick it up.”
References
- Graham, D.J. et al. Risk of acute myocardial infarction and sudden cardiac death in patients treated with cyclo-oxygenase 2 selective and non-selective non-steroidal anti-inflammatory drugs: nested case-control study. Lancet 365, 475-81 (2005).
- Kanemaru, K., Cranley, J., Muraro, D. et al. Spatially resolved multiomics of human cardiac niches. Nature 619, 801-810 (2023).
- Litvi?uková, M. et al. Cells of the adult human heart. Nature 588, 466-472 (2020).
- MacDonald. E.A., Rose, R.A., and Quinn, T.A. Neurohumoral Control of Sinoatrial Node Activity and Heart Rate: Insight From Experimental Models and Findings From Humans. Front Physiol 11, 170 (2020).
- Heuvelman, V.D., Van Raalte, D.H., and Smits, M.M. Cardiovascular effects of glucagon-like peptide 1 receptor agonists: from mechanistic studies in humans to clinical outcomes. Cardiovasc Res 116, 916-930 (2020).