We were testing protoXell's drug similarity search and wanted to try it on a case where we already had some idea of the answer. If the tool is useful, it should be able to find a repurposing story that has been tested in the clinic, without us pointing it there.

Verapamil was a good fit. It is a calcium channel blocker that has been used for decades for blood pressure and heart rhythm problems, and over the past several years it has been tested in diabetes. A small 2018 trial in adults with newly diagnosed Type 1 diabetes found it helped preserve insulin-producing β-cells [1], and a larger trial in children (CLVer) found the same [2]. The largest adult trial so far, Ver-A-T1D with 136 people, narrowly missed its main goal in 2025, partly because the placebo group lost β-cell function more slowly than expected [3]. On the Type 2 diabetes side, R-verapamil lowered blood glucose when added to metformin [4], and a large population study linked verapamil use to fewer new diabetes diagnoses [5]. More trials are underway, including WAVE T1D, which is testing verapamil after an immune therapy in young people with new Type 1 diabetes [6].

So the evidence is promising but not settled, which made it a fair test. Our question was simple: if we give protoXell only Verapamil, does anything diabetes-related come back?

What we ran

All the data in this post comes from Tahoe-100M, a public dataset of about 100 million single cells from 50 cancer cell lines treated with 379 drugs at three doses [7]. Tahoe-100M is public, but at that size it is not easy to work with directly. protoXell handles the underlying heavy lifting such as data preparation, comparing differential expression, gene set enrichment, and computing similarity scores in real-time across the cell lines and doses, so every result in this post came from a simple query rather than a pipeline we had to build and wait for.

protoXell's drug-like perturbation search takes a drug's gene expression response in cells and summarizes it as enrichment scores across generic Gene Ontology gene sets. Then it ranks other drugs by how closely their gene set scores line up with those of the query drug.

To measure "how closely," we used Lin's concordance correlation coefficient (CCC). We picked it up from transcriptomics benchmarking work, where it is used to check how well predicted values agree with true ones [8]. Unlike a plain correlation, CCC only scores high when two profiles agree in both direction and size, not just in overall trend. We applied it to the gene set scores rather than to individual genes.

We compared Verapamil against a reference library of about 200 drugs. The search needs only the drug’s name - no disease or target required; the comparison itself runs on each drug’s gene-set enrichment profile. 

We used one rule for gene and pathway results throughout. A result counts if it is statistically significant (FDR < 0.05) and moves in the same direction in at least two cell lines. Anything that does not meet that bar is still interesting, but we keep it separate in a "leads" section near the end rather than mixing it in.

What came back

Figure 1. Drugs with the most similar response to Verapamil in protoXell. The perturbation table ranks about 200 reference drugs by how closely their gene set enrichment profiles match Verapamil's, scored with Lin's concordance correlation coefficient (CCC). The histogram shows the spread of similarity scores across the library. Dapagliflozin (highlighted) ranks 6th (CCC = 0.271) and is the highest-ranked approved diabetes drug. The panel on the right lists shared pathway signatures with their similarity scores.

Dapagliflozin, an SGLT2 inhibitor and a standard Type 2 diabetes drug, came in 6th.

On its own, 6th out of 200 does not mean much. What made us look twice is how few diabetes drugs are in the library. There are only two: Dapagliflozin and Rosiglitazone (which ranked 25th). If you picked two drugs at random out of 200, the chance that at least one of them lands in the top 6 is about 6%. So it is not impossible that this happened by luck, but it is unlikely.

The scores themselves are modest and close together. All six sit between 0.27 and 0.30, so Dapagliflozin is one of a cluster rather than a standout. The five drugs above it have no known link to diabetes. They might reflect biology nobody has looked at yet, or they might be noise.

The match also does not tell us which mechanism the two drugs share. Dapagliflozin is known to lower glucose mainly by making the kidneys get rid of more sugar, which has nothing to do with how Verapamil is thought to protect β-cells. So the match points toward the broader metabolic area, not a specific Type 1 or Type 2 mechanism.

Checking the mechanisms we already knew about

From here we switched from a blind search to a guided one. Earlier studies tie Verapamil's β-cell effect to two genes. It lowers TXNIP, which drives glucose-induced stress and β-cell death [9], and it raises CaMK4, which supports β-cell survival [10]. We looked up both genes in protoXell and filtered to Verapamil's experiments.

Both passed our bar. TXNIP went down in a colon and a pancreatic cell line (log2FC of -0.14 and -0.31, FDR < 0.001 in both). CaMK4 went up in three unrelated cell lines (FDR between 0.001 and 0.046).

Figure 2. Verapamil lowers TXNIP expression in two cell lines. (1) Search for the gene TXNIP. (2) protoXell returns every perturbation experiment in which TXNIP expression changes. (3) Filter Verapamil's experiments. (4) Per-cell-line results: TXNIP goes down in LS180 (colon, log2FC = -0.14) and MIA PaCa-2 (pancreas, log2FC = -0.314), with FDR < 0.001 in both. In both lines, the glycolysis and gluconeogenesis gene set also goes down (nES = -1.511 and -1.563, respectively).
Figure 3. Verapamil raises CaMK4 expression in three cell lines. (1) Search for the gene CaMK4. (2 and 3) View experiments in which CaMK4 expression changes and filter to Verapamil. (4) CaMK4 goes up in J82 (urinary tract, log2FC = 0.677), HS578T (breast, log2FC = 0.705) and LOXIMVI (log2FC = 0.754), with FDR between 0.001 and 0.046. Glycolysis and gluconeogenesis related gene sets go down in all three (nES = -1.487, -1.394 and -1.426, respectively).

So the genes behind Verapamil's proposed β-cell effect move the way the literature reported. The TXNIP changes are small, and since Tahoe-100M only includes cancer cell lines, none of these are β-cells (the pancreatic line, MIA PaCa-2, is a pancreatic cancer line), so this shows the signal is visible in the data rather than reproducing β-cell protection itself.

Leads we have not confirmed

The similarity score tells us Verapamil and Dapagliflozin look alike overall, but not which gene sets drive it. So we checked a few signals by hand. None met our bar for both drugs, usually because a result was only in one cell line or was not significant. We are including them because they line up with published work and are low-cost to follow up.

TNF-α signaling. This is a well-known driver of insulin resistance [11], and both drugs lowered it. Dapagliflozin's result is solid: it passed our bar in A549 (nES -2.23, FDR < 0.001) and BT474 (nES -1.75, FDR 0.014). Verapamil's is not: it lowered the pathway in one colon line (nES -2.87, FDR 0.228). Separate studies report anti-inflammatory effects for both drugs, Verapamil in the liver [12] and Dapagliflozin in endothelial cells and macrophages [13], so this is the lead we check first.

Figure 4. TNF-α signaling goes down with both Verapamil and Dapagliflozin. (1) Search for TNF-α signaling without specifying a drug. (2) Results for the Reactome pathway "TNFs bind their physiological receptors." Verapamil: SW1417 (colon), nES = -2.866, FDR = 0.228. Dapagliflozin: A549 (lung), nES = -2.225, FDR < 0.001; BT474 (breast), nES = -1.748, FDR = 0.014. Dapagliflozin's result meets our evidence bar; Verapamil's does not.

Other immune pathways. Verapamil lowered IL-27 related signaling in one colon line (nES -2.05, FDR 0.018) and IL-17 related signaling in another (nES -2.84, FDR 0.224). A mouse study found Verapamil reduced inflammation and joint damage in arthritis [14]. Dapagliflozin has been linked to anti-inflammatory effects involving the NLRP3–IL-1β axis in preclinical models [15] while the DAPA-LVH trial clinically reported a significant reduction in CRP relative to placebo [16].

Figure 5. Other immune signatures for Verapamil and Dapagliflozin. Verapamil lowers IL-27 signaling in LS180 cells (nES = −2.049, FDR = 0.018), and IL-17 production in SW1417 cells (nES = −2.841, FDR = 0.224), which is below our evidence bar. Dapagliflozin’s profile points toward IL-1 and IL-12 pathways: the interleukin 1 mediated signaling pathway goes down in C3A cells (nES = −1.613, FDR = 0.027), and positive regulation of interleukin-12 production goes down in SW48 cells (nES = −1.516, FDR = 0.167), also below our evidence bar.

What we take from this

We gave protoXell one drug and nothing else. It came back pointing at diabetes, the same area where Verapamil is being tested in the clinic today, along with five other drugs with similar profiles that nobody has linked to diabetes yet.

The link is not just a ranking, either. When we followed up, the genes behind Verapamil's proposed β-cell effect, TXNIP and CaMK4, were there in the same data, moving the way earlier studies reported, and the inflammation signals add a few more clues worth testing. These are cell line results, so they are starting points for the lab rather than conclusions. But they do provide real molecular clues, not just a similarity score.

This post only used Tahoe-100M, but protoXell is not tied to one dataset. It can work across several perturbation datasets at once, covering both chemical perturbations (drugs) and genetic ones. Right now, Tahoe-100M and X-Atlas/Orion, about 8 million cells in which every human protein-coding gene was knocked down one at a time [17], are both free to try in protoXell.

Every result in this post was done by hand: we ran the searches, chose which genes and pathways to check, and read the results ourselves. Since then, we have built an agent that uses the same tools in protoXell to go from a search result to an insight like this one, so scientists can get there faster. More on that in our next post.

References

  1. Ovalle F, et al. (2018). Verapamil and beta cell function in adults with recent-onset type 1 diabetes. Nature Medicine, 24(8), 1108-1112. doi:10.1038/s41591-018-0089-4 (Trial: NCT02372253)
  2. Forlenza GP, et al. (2023). Effect of verapamil on pancreatic beta cell function in newly diagnosed pediatric type 1 diabetes: a randomized clinical trial. JAMA. PubMed 36826844 (Trial: NCT04233034)
  3. Wych, J., Brunner, M., Stenson, R., Chmura, P. J., Danne, T., Mander, A. P., Mathieu, C., Dayan, C., & Pieber, T. R. (2024). Investigating the effect of verapamil on preservation of beta-cell function in adults with newly diagnosed type 1 diabetes mellitus (Ver-A-T1D): protocol for a randomised, double-blind, placebo-controlled, parallel-group, multicentre trial. BMJ open, 14(11), e091597. https://doi.org/10.1136/bmjopen-2024-091597   
  4. Wang CY, et al. (2022). A randomized controlled trial of R-form verapamil added to ongoing metformin therapy in patients with type 2 diabetes. Journal of Clinical Endocrinology & Metabolism, 107(10), e4063-e4071. doi:10.1210/clinem/dgac436
  5. Yin T, et al. (2017). Verapamil use is associated with reduction of newly diagnosed diabetes mellitus. Journal of Clinical Endocrinology & Metabolism, 102(7), 2604-2610. doi:10.1210/jc.2016-3778
  6. WAVE T1D: A randomized phase 1/2 trial of low dose anti-thymocyte globulin (ATG) with subsequent adalimumab or verapamil in new onset type 1 diabetes. City of Hope Medical Center. Trial: NCT07061574
  7. Zhang, J., Ubas, A. A., Svensson, V., Thomas, N., Subramanyam, V., Carpenter, C., Winters, A., de Borja, R., Teyssier, N., Thakar, N., Khan, U., Akella, P. V., Green, H. J., Thompson, J. D., Tran, V., Pangallo, J., Papalexi, E., Sapre, A., Nguyen, H., Sanderson, O., … Yu, J. (2026). Tahoe-100M: Mapping drug-induced molecular phenotypes at single-cell resolution. Cell, 189(19), 5945–5961.e8. https://doi.org/10.1016/j.cell.2026.08.035  
  8. Malik, A. A., Alif, M. N., Bratton, A., Sun, J., Li, Q., & Zhang, W. (2026). Benchmarking reference-based cellular deconvolution algorithms to predict cell proportions. Computational biology and chemistry, 126(Pt 1), 109370. Advance online publication. https://doi.org/10.1016/j.compbiolchem.2026.109370  
  9. Borowiec AM, et al. (2022). TXNIP inhibition in the treatment of diabetes. Verapamil as a novel therapeutic modality in diabetic patients. Medicine and Pharmacy Reports, 95(3), 243-250. doi:10.15386/mpr-2187
  10. Arefanian H, et al. (2023). Verapamil chronicles: advances from cardiovascular to pancreatic β-cell protection. Frontiers in Pharmacology, 14, 1322148. doi:10.3389/fphar.2023.1322148
  11. Akash MSH, et al. (2018). Tumor necrosis factor-alpha: role in development of insulin resistance and pathogenesis of type 2 diabetes mellitus. Journal of Cellular Biochemistry, 119(1), 105-110. doi:10.1002/jcb.26174
  12. Zhou F, et al. (2018). Verapamil ameliorates hepatic metaflammation by inhibiting thioredoxin-interacting protein/NLRP3 pathways. Frontiers in Endocrinology, 9, 640. doi:10.3389/fendo.2018.00640
  13. Abdollahi E, et al. (2022). Dapagliflozin exerts anti-inflammatory effects via inhibition of LPS-induced TLR-4 overexpression and NF-κB activation in human endothelial cells and differentiated macrophages. European Journal of Pharmacology, 918, 174715. doi:10.1016/j.ejphar.2021.174715
  14. Wang W, et al. (2016). Chronic calcium channel inhibitor verapamil antagonizes TNF-α-mediated inflammatory reaction and protects against inflammatory arthritis in mice. Inflammation, 39(5), 1624-1634. doi:10.1007/s10753-016-0396-1
  15. Quagliariello, V., Canale, M. L., Bisceglia, I., Iovine, M., Paccone, A., Maurea, C., Scherillo, M., Merola, A., Giordano, V., Palma, G., Luciano, A., Bruzzese, F., Zito Marino, F., Montella, M., Franco, R., Berretta, M., Gabrielli, D., Gallucci, G., & Maurea, N. (2024). Sodium-glucose cotransporter 2 inhibitor dapagliflozin prevents ejection fraction reduction, reduces myocardial and renal NF-κB expression and systemic pro-inflammatory biomarkers in models of short-term doxorubicin cardiotoxicity. Frontiers in cardiovascular medicine, 11, 1289663. https://doi.org/10.3389/fcvm.2024.1289663  
  16. Dihoum, A., Brown, A. J., McCrimmon, R. J., Lang, C. C., & Mordi, I. R. (2024). Dapagliflozin, inflammation and left ventricular remodelling in patients with type 2 diabetes and left ventricular hypertrophy. BMC cardiovascular disorders, 24(1), 356. https://doi.org/10.1186/s12872-024-04022-7  
  17. Huang et al. (2025). X-Atlas/Orion: Genome-wide Perturb-seq datasets via a scalable fix-cryopreserve platform for training dose-dependent biological foundation models. bioRxiv. doi:10.1101/2025.06.11.659105