Measuring how target antigen levels drive CAR T-cell efficacy in AML

CAR T-cell therapy has produced some of the most striking remissions in modern medicine against B-cell leukemias and lymphomas, yet acute myeloid leukemia has stubbornly resisted the same success. A new study now offers one of the most detailed quantitative explanations yet for why: the effectiveness of a CAR T-cell construct against AML depends intimately on which target antigen it recognizes, and that dependence changes in complex, nonlinear ways when the leukemia cells lack the tumor suppressor TP53. By combining laboratory co-culture experiments with mathematical modeling and Bayesian statistics, an international research team has built a framework that can predict, construct by construct, how engineered T-cells will expand, kill, exhaust, and ultimately fail or succeed against resistant leukemia variants.

The work, published in Cancer Cell International, was led by Saumil Shah and Philipp M. Altrock of the Max Planck Institute for Evolutionary Biology in Plön, Germany, together with Jan Mueller, Emanuel Vogel, Steffen Boettcher, and Markus G. Manz of the University of Zurich and University Hospital Zurich, along with Michael Raatz and Arne Traulsen. The team’s central insight is that CAR T-cell expansion is not a simple function of how many cancer cells are present. Instead, it is governed by a delicate balance of attack rate, handling time, T-cell crowding, and effector cell death, and each of these parameters shifts differently depending on the target antigen and the genetic background of the leukemia.

To capture these dynamics, the researchers engineered TP53-wildtype and TP53-knockout versions of the MOLM-13 acute myeloid leukemia cell line using CRISPR/Cas9 gene editing, tagging both with green fluorescent protein and luciferase so that cell populations could be tracked by flow cytometry. Against these targets they pitted second-generation CAR T-cells carrying five different single-chain variable fragments, directed against CD33, CD117, two variants of CD123 (strong and weak binders), and CD371, each paired with a CD8 stalk, transmembrane domain, and the 4-1BB intracellular costimulatory domain. Untransduced donor T-cells served as controls. The co-cultures were set up at multiple initial effector-to-target ratios and followed longitudinally over days, producing a rich dataset of interacting cell populations under twenty distinct experimental conditions.

The mathematical heart of the study is a two-compartment system of ordinary differential equations describing the abundance of leukemia cells and T-cells over time. Target cells grow exponentially and die through mass-action killing by effectors, while the effector population expands according to a function that the researchers allowed to take several candidate forms drawn from classical predator-prey ecology. These included Lotka-Volterra expansion, Holling type 2 saturation, ratio-dependent models, and the Beddington-DeAngelis formulation, which accounts both for the time a T-cell spends engaged in killing a target—its handling time—and for interference among T-cells themselves as densities rise. Before fitting anything, the team verified that every parameter in every candidate model was structurally identifiable, using differential algebra techniques to confirm that the data could, in principle, pin the parameters down.

Bayesian inference then did the heavy lifting. Rather than hunting for a single best-fit point estimate, the researchers assigned log-normal prior distributions to each parameter, constrained by independent measurements of tumor growth and physical plausibility, and sampled full posterior distributions using the No-U-Turn Sampler, a Hamiltonian Monte Carlo algorithm. Five independent chains of ten thousand samples were run for each condition, and convergence was checked before any conclusions were drawn. When the candidate models were ranked by corrected Akaike and Bayesian information criteria, both of which penalize model complexity, the Beddington-DeAngelis expansion won across all experimental settings—despite carrying the largest number of parameters. Its victory carried a biological message: T-cell expansion in these assays is genuinely limited both by the time individual CAR T-cells spend conjugated to leukemia targets and by crowding among the expanding T-cell population itself.

The posterior parameter distributions then revealed something the field had largely been unable to articulate quantitatively: resistance in TP53-deficient AML is not a uniform wall but a mosaic of antigen-specific kinetic changes. CD33-targeting CARs, for example, showed reduced attack rates against TP53-knockout cells compared to wildtype targets, along with elevated death rates. CD123- and CD371-targeting CARs, by contrast, mounted moderately increased attack rates against the mutant cells, but paid for this aggression in other currencies: the CD123 construct suffered higher effector death rates, while CD371 CARs exhibited longer handling times, meaning more of each cell’s lifespan was consumed in the killing process rather than in proliferation. Untransduced T-cells, meanwhile, behaved similarly against both genotypes with low attack rates and high death rates, confirming that the antigen-specific differences were genuinely CAR-driven.

Perhaps the most consequential single finding concerned CD123s, the strong-affinity CD123 construct. Model simulations using median posterior parameters predicted that only this construct would sustain suppression of TP53-knockout leukemia over extended timescales. The apparent reason is its comparatively short handling time, which mitigates the built-in cost of killing: every hour a CAR T-cell spends bound to a target is an hour it cannot divide. CD117-targeting CARs proved largely ineffective in this system, consistent with prior clinical observations, while CD123w and CD371 constructs eventually lost control of the knockout cells despite strong initial cytotoxicity. In four of the five constructs, the leukemia cells did not simply become invisible to the CAR T-cells; rather, the CARs attacked more aggressively against TP53-deficient targets, but that aggression was throttled by faster exhaustion or longer conjugation times.

Stability analysis of the winning model sharpened this picture further. The model permits a stable state of tumor control under specific parameter relationships, but it also predicts a counterintuitive failure mode: a very high attack rate that is not balanced by a short handling time can actually promote tumor escape, because CAR T-cells locked into constant target engagement cannot contribute to population expansion. This mathematical result reframes a long-standing assumption in the field. Killing, the analysis suggests, is not the rate-limiting step in CAR T-cell therapy against AML—expansion is. Sustained effector proliferation with short handling times may be the property that most reliably translates into durable clinical responses.

To guard against overfitting, the researchers validated their trained model on a completely separate dataset from a smaller assay format that extended follow-up to sixteen days, beyond the training window of ten days, and used different initial effector-to-target ratios. Predictions generated from 250 random draws of the parameter posteriors were compared with the unseen data using distance correlation statistics, with values above 0.6 considered adequate and above 0.9 excellent. Every model-data correlation cleared the 0.6 threshold across all experimental scenarios, with leukemia compartment predictions typically outperforming T-cell compartment predictions. The authors are candid about limitations: the system uses a single AML cell line, does not model antigen density as an explicit state variable, does not capture antigen mixtures or the evolutionary selection of TP53-depleted clones, and cannot reflect the full clinical heterogeneity of patients. Antigen abundance per cell and construct-specific binding affinities also varied implicitly rather than being measured directly.

Even so, the framework represents a template for how preclinical CAR T-cell development might be conducted in the future. Instead of declaring a construct “effective” based on a single endpoint like percent killing at 24 hours, the Bayesian workflow extracts a full kinetic fingerprint—attack rate, baseline expansion, handling time, crowding coefficient, and death rate—that explains why a construct works or fails, and under what dosing conditions. The authors note that because the attack rate enters both the killing term and the expansion function in the fitted model, its dual role is statistically identifiable, allowing the framework to separate the immediate benefit of cancer cell destruction from its long-term cost to T-cell fitness. This kind of mechanistic bookkeeping is precisely what has been missing in the search for a suitable CAR T-cell target in myeloid malignancies.

The clinical implications are tantalizing. Since different antigens confer complementary kinetic advantages—some favoring sustained expansion, others favoring potent initial killing—the authors suggest that future therapeutic strategies could combine multiple target-antigen constructs in a single infusion or exploit their distinct properties in sequential dosing. For patients with TP53-mutated AML, an aggressive disease with dismal prognosis and few effective options, such rational, model-guided combination strategies may represent one of the more credible paths forward. The study’s data and code are publicly available through a Zenodo repository, underscoring the authors’ intent that the framework be adopted, adapted, and stress-tested by the wider community. As immunotherapy increasingly becomes a quantitative science, this work demonstrates that the tools of ecology and statistical inference, applied rigorously to the petri dish, can illuminate the rules by which engineered immune cells and resistant cancers wage their microscopic war.