In the next 60 seconds, a small mob of immune cells inside you will check molecular ID cards, ignore your law-abiding tissues, and go hunting for anything that looks sketchy. Red blood cells will keep cruising like nothing is happening. T cells will do security work with all the glamour of airport staff at 5 a.m. Cancer, naturally, tries to slip past with a fake mustache.
That is the backdrop for this new paper on CAR T cells, which are your own T cells re-engineered to recognize cancer more aggressively. Think of them as immune cells with a custom search warrant. They have already changed the game for several blood cancers, but the system is still messy, expensive, and annoyingly hard to tune just right [2]. The big question is not whether CAR T can work. It can. The question is how to design better versions without brute-forcing our way through a biochemical IKEA warehouse of receptor parts.
CAR T, but make it less guessy
The study by Driessen and colleagues tackles a very modern problem: there are too many possible CAR designs, and T cells do not all behave the same way even when you give them the exact same gadget [1]. Some cells become energetic tumor-killers. Some get tired. Some seem to wander off into the cellular equivalent of replying "per my last email" and then doing very little.
To deal with that chaos, the authors used single-cell transcriptomics, which measures gene activity one cell at a time. That matters because averages can lie. If ten cells are thriving and ten are exhausted, the average says, "eh, medium vibes." Biology is rarely that polite.
Their model tries to predict how individual cells will respond when a specific CAR is expressed. Not just CARs already seen in the training data, but entirely new variants too. That is the ambitious part. Plenty of models look smart when they are quizzed on yesterday's homework. Fewer stay smart when handed a new exam.
A GPS for cellular chaos
The key method here is conditional optimal transport. Optimal transport is a math framework for figuring out the most plausible way to transform one distribution into another. In plain English: if you measure one cloud of cells before engineering and another cloud after engineering, how do you estimate which kinds of changes likely happened, even though single-cell sequencing destroys the cells and you cannot track the same one twice? [5,6]
This paper uses that idea to model CAR-driven changes in gene expression at the single-cell level [1]. Then it adds another layer: protein language models, which encode CAR protein sequences so the system can make educated predictions about unseen CAR designs. Yes, we now have AI reading protein sequences like they are extremely cursed sentences. Biology in 2026 has fully entered its weird crossover season.
The result is a framework that beat a baseline model on known CAR variants and could generalize to out-of-distribution designs, meaning new CARs that had not been experimentally tested [1]. That does not mean the computer has become your new principal investigator. It means the computer may be useful for telling you which experiments are worth your weekends.
Why this matters outside the spreadsheet
CAR T therapy already delivers striking remissions in some blood cancers, but it still runs into relapse, toxicity, manufacturing headaches, and weaker performance in solid tumors [2-4]. Researchers have been using multi-omics and single-cell tools to understand why some CAR T cells persist while others flame out like a New Year's resolution by January 6 [3,4]. This new paper fits right into that push.
If this kind of modeling holds up across larger datasets and more biological settings, it could help researchers prioritize CAR designs with better odds of expanding, persisting, and attacking tumors without burning patients with extra toxicity. It could also shrink the trial-and-error burden in protein engineering, which is good for science and even better for the graduate students currently living inside that trial-and-error burden.
There is also a deeper point here. Cancer is an evolutionary arms race. Tumors diversify. Immune cells adapt. Therapies apply pressure, and cells respond like tiny opportunists with no moral compass whatsoever. A model that respects heterogeneity at the single-cell level is not just technically fancy. It is philosophically aligned with the problem. Cancer is not one enemy. It is a squabbling population of them.
So the real appeal of this paper is not that it makes CAR T simple. Nothing about CAR T is simple. It is that it gives researchers a better map of the battlefield. And when the battlefield is made of millions of cells all improvising at once, a better map is a pretty good place to start.
References
-
Driessen A, Born J, Castellanos Rueda R, Reddy ST, Rapsomaniki M. Modeling chimeric antigen receptor response at the single-cell level with conditional optimal transport. Cell Systems. 2026. doi:10.1016/j.cels.2026.101591
-
Brudno JN, Maus MV, Hinrichs CS. CAR T Cells and T-Cell Therapies for Cancer: A Translational Science Review. JAMA. 2024;332(22):1924-1935. doi:10.1001/jama.2024.19462
-
Yang J, Chen Y, Jing Y, et al. Advancing CAR T cell therapy through the use of multidimensional omics data. Nature Reviews Clinical Oncology. 2023;20:211-228. doi:10.1038/s41571-023-00729-2
-
Barboy O, Katzenelenbogen Y, Shalita R, Amit I. In Synergy: Optimizing CAR T Development and Personalizing Patient Care Using Single-Cell Technologies. Cancer Discovery. 2023;13(7):1546-1555. doi:10.1158/2159-8290.CD-23-0010
-
Bunne C, Stark SG, Gut G, et al. Learning single-cell perturbation responses using neural optimal transport. Nature Methods. 2023;20:1759-1768. doi:10.1038/s41592-023-01969-x
-
Dong M, Viñas R, Abdelmoula WM, et al. Causal identification of single-cell experimental perturbation effects with CINEMA-OT. Nature Methods. 2023;20(11):1769-1779. doi:10.1038/s41592-023-02040-5
Disclaimer: The image accompanying this article is for illustrative purposes only and does not depict actual experimental results, data, or biological mechanisms.