Why CAR T-cell remissions fade: the two defences that hold cancer at bay

On the surface, CAR T-cell therapy looks like a miracle. A patient's own immune cells are genetically rewired to hunt down a blood cancer, and in relapsed or refractory B-cell acute lymphoblastic leukaemia — a disease that has already shrugged off chemotherapy — more than 80% of patients go into remission. Yet that miracle slides off a cliff. Up to half of those who remit relapse within a year, and the stubborn recurrence reasserts a hard clinical truth: inducing remission and keeping it are two entirely different problems.

The puzzle that has nagged oncologists is a strange one. Patients who maintain long-lived CAR T-cells in their blood tend to stay cancer-free. Those whose CAR T-cells fade, relapse. It sounds obvious — but it isn't, because the two candidate explanations point in opposite directions. Maybe the residual disease, whittled down to a handful of cells, needs a sustained army of effector killers to finish it off. Or maybe the memory pool — the quiet reserve of stem-like cells — matters more, and relapse happens when that pool is ground down by repeated activation. You cannot easily tell which from the clinic, because both are tangled together in every patient.

A new mathematical model from University College London has prised these two pathways apart. The researchers built a computational simulation of CAR T-cell dynamics called BEAM — Blast, Effector, Activated, Memory — calibrated against real patient data from the FELIX trial of the CAR T-cell product obecabtagene autoleucel (obe-cel). The result is a finding that reframes how we should think about why remission holds. Memory persistence and effector persistence both prevent relapse, but they do so through different biological traps — and that distinction has direct consequences for how next-generation CAR T-cells should be engineered.

The Science

To see why the memory-versus-effector question is so hard to answer clinically, consider what a CAR T-cell does. It is a living drug. Engineer a patient's T-cells to carry a receptor (the "chimeric antigen receptor," or CAR) that recognises a protein, CD19, plastered on leukaemic B-cells, then infuse them back, and you have created a predatory population that hunts, expands, and — if all goes well — lingers for months or years.

The UCL team, led by Alexis Farman, Benjamin Walker, Martin Pule and Karen Page, formalised this into a predator–prey system of differential equations. But they added a crucial layer of biological realism: the CAR T-cells are not a single homogeneous blob but three functional states. Resting memory cells (M) can be activated by contact with blasts into activated cells (A), which then commit to one of two fates — differentiating into the frontline effector killers (E) that actually lyse tumour cells, or dividing back into more memory cells. The leukaemic blasts themselves follow logistic growth, expanding up to a carrying capacity of a trillion cells — the mass that constitutes a lethal disease.

The model's elegant lever is the variable $B_{1/2}$, a threshold. When blast burden is high, activated cells predominantly differentiate into effectors — you want maximum firepower when the enemy is all around you. When burden is low, activated cells predominantly become memory — you bank soldiers for the long war. This burden-dependent fate choice mirrors real biology: a T-cell's decision between an effector and a memory identity is shaped by cumulative antigenic signalling, as prior work has shown (Arroyave et al., 2023; cited in the paper).

Because the model is built from coupled ordinary differential equations, the team could analyse it rigorously — hunting for stable steady states, and mapping how outcomes flip as parameters shift. But there was a problem lurking in the differential-equation approach. When treatment works, the tumour can crash to absurdly low numbers — a handful of cells in a body of trillions. Differential equations, which treat populations as smooth continuous fluids, break down in that regime: they can never actually reach zero. So the team built a hybrid algorithm that runs the fast, large-population dynamics deterministically, then switches to discrete stochastic simulation — treating individual cells as lottery tickets — once populations shrink below a threshold. This hybrid matters because it distinguishes two very different kinds of "success": deterministic dormancy, where a tiny blast population is held in a stable standoff, versus true stochastic elimination, where the last cell dies by chance.

What They Found

The central theoretical result is a bifurcation that separates the model's behaviour into regimes. Analysing the steady states, the team found that a dormant — persistently present but non-growing — tumour state exists only below a threshold in the memory-cell death rate $ϵ$, given by

$$ϵ<\frac{k_3{B}_{1/2}}{3+2\sqrt{2}}.$$

Below this threshold, the system becomes bistable: depending on the trajectory, the same patient could either escape (tumour regrows to carrying capacity) or fall into dormancy (tumour held at a tiny burden). The remarkable twist is that the size of that dormant blast population, $B_{-}^{\ast }$, is determined entirely by memory-related parameters — the activation rate $k_3$, the phenotype threshold $B_{1/2}$, and the memory death rate $ϵ$. The effector parameters, $\gamma$ and $\delta$, control how many CAR T-cells are needed to hold the standoff, but they do not change the dormant blast burden itself.

The bifurcation diagram shows this fold: as the memory death rate drops (memory cells live longer), the dormant tumour burden falls toward zero — and the window in which dormancy is even possible widens.

So memory persistence is the gatekeeper of dormancy. A long-lived memory pool sustains the surveillance that pins residual disease at a low, harmless plateau. That in itself is not shocking — long memory persistence is clinically associated with durable remission. But the model separates this from a second, independent protective mechanism. When blasts can hide in immune-privileged niches — shielded sanctuaries in the bone marrow where circulating effectors cannot reach them — they can emerge later as isolated ambushes. The model shows that it is the effector pool, not the memory reserve, that sweeps up these stragglers as they leak out of hiding.

This dual mechanism produces a genuine design trade-off. The team found that parameters favouring immediate cytotoxicity (strong effectors) can come at the cost of durable surveillance, and vice versa — the model captures a tension between maximal upfront killing and long-term vigilance. It is the quantitative version of a biological intuition: an immune system that burns every activated cell into a short-lived effector wins the battle but exhausts its army for the war.

The most actionable result concerns tumour burden — and it slices against intuition. One might assume that a lower starting tumour burden is always better. The model complicates this. Because the effector-vs-memory fate decision is burden-dependent through $B_{1/2}$, the size of the tumour before infusion shapes which protective mechanism dominates. The researchers identify initial tumour burden as a key modifiable factor for reducing antigen-negative relapse — suggesting that cytoreduction (shrinking the tumour with other agents before CAR T infusion) could shift the balance of the differentiation decision toward more durable surveillance.

How the killing rate k2 decides between escape, transient response and elimination

Label	Value
k2 = 1.5e-10 (low)	0
k2 = 2.5e-10 (intermediate)	1
k2 = 4e-10 (baseline)	1

Source: Farman et al. 2026, Fig. 3. Blast counts B(t) over 300 days for baseline parameters and varying killing rate k2. Data are schematic of simulated trajectories — low k2 stays above MRD (escape), intermediate k2 transiently crosses MRD then regrows, baseline k2 is eliminated.

The numerical simulations lay the stakes bare. Holding all else fixed, sweeping the killing rate $k_2$ produces three strikingly different fates. At low killing efficiency, the tumour never even dips below the minimal residual disease (MRD) detection threshold of $10^6$ cells — the therapy fails outright. At intermediate efficiency, the tumour dips below detection but doesn't die; it lingers and later regrows. Only at high efficiency does the blast count genuinely hit zero — and even then, the model computes a probability of elimination, not a guarantee.

Four fates, four blast burdens: the outcomes the model predicts

Label	Value
Escape	100
Transient response then relapse	70
Dormancy	45
Stochastic elimination	30

Source: Farman et al. 2026. Blast burden as % of carrying capacity K=10^12 cells for the four possible treatment outcomes. Dormant burden B*_- is set by memory parameters only (k3, B1/2, epsilon); escaping tumours regrow toward K.

Across a wide range of $k_2$, both elimination and relapse are live possibilities, which is exactly what clinicians see: the same drug, the same dose, different patients, opposite fates.

Why This Changes Things

The deepest implication is that "persistence" is not one thing. Clinicians have long treated durable CAR T-cell persistence as a single favourable correlate. The BEAM model argues it is really two distinct defences working through different routes — and that conflating them obscures how to engineer better therapies.

Consider the two relapse pathways the model distinguishes. In the first, low-burden residual disease rebounds because the memory pool that was holding it dormant eventually fades. In the second, blasts sheltering in immune-privileged niches re-emerge and are only cleared by available effectors. The therapeutic fixes are different. For the memory-limited pathway, you want memory-biased products — cells engineered to skew toward the self-renewing memory fate, or repeated dosing to replenish the pool. For the effector-limited pathway, you want longer-lived effectors that persist at the frontline. A "one more dose" strategy might salvage one pathway while doing nothing for the other.

The trade-off between immediate cytotoxicity and durable surveillance also reframes how we evaluate CAR T-cell products. A therapy that looks spectacular in the first month — massive effector expansion, dramatic tumour clearance — might be quietly mortgaging its long-term durability. Conversely, a product with a more modest first strike might build the memory reserve that prevents late relapse. The model gives a framework for weighing those costs quantitatively rather than relying on intuition or individual trial anecdotes.

The prediction about tumour burden is perhaps the most clinically immediate. Antigen-negative relapse — where the cancer stops expressing CD19 so the CAR T-cells no longer recognise it — accounts for a substantial fraction of treatment failures. The model suggests that reducing initial burden before infusion could make antigen-negative relapse less likely, by shifting the fate decisions in a direction that favours durable immune coverage. This is testable, and it dovetails with existing clinical practice of "bridging" therapy to control disease before CAR T infusion — but now with a mechanistic rationale rather than a vague sense that less cancer is better.

The physics of the finding also matters methodologically. By showing that deterministic dormancy and stochastic elimination are genuinely different regimes — one a stable standoff, the other a roll of the dice — the model warns against trusting purely deterministic simulations of these therapies. If a continuous model says "remission," it may be describing a dormant state that in reality can tip either way through chance. That distinction has consequences for how trial data are interpreted and how "cure" versus "control" is defined.

What's Next

The BEAM model is a scaffold, not a finished cathedral, and the authors are appropriately honest about its simplifications. The core model is "well-mixed" — it assumes CAR T-cells and blasts encounter each other uniformly through the body, which is a fiction for a disease that hides in anatomical niches. The paper's niche model addresses this in an extension, but the two-body problem of circulating versus hidden disease is only beginning to be mapped. The lumped killing rate $k_2$ is tuned to reproduce success, not measured from first principles; several parameters — activation rate, fate-switching threshold — carry wide plausible ranges precisely because they cannot be read directly off clinical data.

Memory persistence is the gatekeeper of durable tumour control

Label	Value
Memory lifespan ~100 days, slow growth	3
Memory lifespan ~100 days, fast growth	2
Memory lifespan ~50 days, slow growth	2
Memory lifespan ~50 days, fast growth	1

Source: Farman et al. 2026, Fig. 5a. Outcome over T=365 days classified from 500 hybrid realisations per grid point; escape shown in red, dormancy in orange, elimination in green. Longer-memory lifespans push the system toward dormancy and elimination.

The stochastic element is where the model earns its keep. Because it treats the death of the last blast as a genuine random event, it produces probabilistic predictions: elimination is a probability, not a certainty. This opens the modelling door to asking sharper questions than "does it work?" — for instance, how many cells of residual disease can a given memory lifespan tolerate, and what dose schedule maximises the odds that a patient crosses from dormancy into true elimination.

The natural next steps are to bind the model tighter to longitudinal patient data — fitting it to individual trajectories from FELIX and other trials rather than aggregate curves — and to extend the niche and antigen-escape mechanisms into the core framework, testing how the dual persistence requirement shifts when blasts can hide and mutate away their target simultaneously. Ultimately, the aim is the one the authors state plainly: a framework for designing more durable, individually tailored CAR T-cell therapies, where the dose, the product composition, and the timing are chosen not to maximise the first-month response but to make remission hold.

That is the quiet revolution here. The field's remarkable success in inducing remission is now the baseline, not the goal. The hard problem — the one that decides whether a patient lives or relapses — is durability. BEAM gives that problem its first clear mathematics: two distinct mechanisms, one trade-off, and a map of the parameters that a clinician or engineer can actually pull.