A Peak Is Not a Score

In the FELIX trial of obecabtagene autoleucel — the CD19 CAR-T therapy approved for adult B-cell ALL — thirty-one relapsing patients had their blasts checked for CD19 at the time of relapse. Fifteen of them had lost the antigen. The number sits in Supplementary Table S15 of the NEJM report and is otherwise unremarked. A new modelling paper out of UCL — Farman, Walker, Pule and Page’s BEAM model — takes it seriously enough to do the arithmetic, and the arithmetic says something the field’s entire measurement apparatus is built to miss. The derivation is short enough to state. Assume CD19-positive blasts lose the antigen at some small rate during division. Then the number of CD19-negative cells created in a time window is Poisson, with mean proportional to the integral of CD19-positive blast divisions over that window — an area under a curve, in units of cell-days. Each such cell founds a lineage that goes extinct with the classical birth–death probability. Compose the two, take the moment-generating function of the Poisson, and you get a closed-form escape probability that depends on the whole trajectory only through that one integral. Which reclassifies antigen-negative relapse entirely. It is not a resistance phenomenon in the usual sense — nothing is being selected for, nothing is learning. It is a lottery, and the number of tickets bought is the cumulative leukemic burden carried through the first weeks of treatment. Depth of response matters only insofar as it shrinks the integral. Duration of remission does not matter at all. What matters is the area under the blast curve before the curve gets low.

Now look at what that implies about the dials. BEAM sorts its parameters into three groups by their effect on the three-month integral: “Initial Response” parameters shaping the early effector wave, “Growth & Burden” parameters setting initial accumulation, and “Long Term Response” parameters — memory lifespan, memory differentiation bias — which, in the authors’ words, “barely affect the 3-month integral.” Persistence, the thing the entire field has converged on as the marker of a good CAR-T product, is essentially orthogonal to half of relapse. And it really is half. The other half — CD19-positive relapse, disease growing back under a shrinking immune shadow — is exactly what persistence governs, and there the clinical data are unambiguous: a companion FELIX pharmacokinetics analysis found loss of CAR T-cell persistence carried a hazard ratio of 2.7 (95% CI 1.4–5.4) for event-free survival. Obe-cel’s whole design premise is built on that finding — an intermediate-affinity, fast-off-rate binder that disengages quickly, avoids overstimulation and exhaustion, and preserves the memory compartment. It works, and the safety numbers show it: grade ≥3 CRS in 2.4% of patients, grade ≥3 neurotoxicity in 7.1%, strikingly low for this disease.

But the two halves of relapse pull the same dial in opposite directions. A gentler receptor that spares the memory pool buys protection against the CD19-positive half by trading away exactly the early cytotoxic intensity that shrinks the integral governing the CD19-negative half. BEAM makes this explicit as a trade-off between immediate cytotoxicity and durable surveillance, and the model’s treatment of memory replenishment is what forces it, since memory is only restored once tumour burden is already low. You cannot tune your way out with receptor affinity. Whichever way the knob turns, one of the two relapse pathways gets worse. Which leaves burden itself as the only lever that improves both at once — and here the clinical data are unusually clean. A Haematologica analysis of FELIX’s split-dose strategy found that a 50-percentage-point increase in bone marrow blasts at lymphodepletion (70% versus 20%) was associated with a 1.9-fold increase (95% CI 1.4–2.6) in maximal CAR T-cell expansion, with overall remission of 85% in the low-burden group against 73% in the high-burden group.

Read those two facts together and something uncomfortable falls out about the field’s favourite pharmacokinetic readout. Peak CAR T-cell expansion is substantially a function of how much leukemia the patient had. It is, in significant part, a thermometer for the disease rather than a scorecard for the drug — and the patients posting the biggest expansion numbers are precisely the ones carrying the largest integral, holding the most tickets in the antigen-loss lottery. Optimising a product toward higher Cmax is optimising toward a variable whose main driver is the thing you most want to eliminate. So, the bet: the next real gain in adult B-ALL will not come from a better receptor, it will come from moving infusion earlier — treating at measurable residual disease rather than at morphologic relapse — because that is the only intervention attacking the integral directly instead of trading one relapse pathway for the other. Concretely, I would expect a head-to-head of two CD19 receptor variants to show a meaningful difference in CD19-positive relapse and essentially none in the CD19-negative fraction, while a burden-stratified comparison would move the CD19-negative fraction and little else. Two caveats worth stating plainly. The 15-of-31 figure covers only relapses in which CD19 was actually assessed, which is a selected subset. And BEAM is a four-compartment ODE model calibrated to a single trial; it earns its keep by making the structure of the problem legible, not by predicting individual patients. But the structural point does not depend on the calibration. It follows from the integral.