Lymphoma Study Links Lasting CAR T Responses to Host CD8 T Cells

Large B-cell lymphoma that returns after chemotherapy is now routinely treated with CAR T cells, which are a patient's own T cells engineered to recognize the surface protein CD19 on B cells. In this preprint's cohort only about a third of patients had a durable response. A natural assumption is that the engineered cells do all the killing, so resistance should come down to CD19 loss or CAR T cell exhaustion. The preprint by Ziccheddu and colleagues asks a different question: whether the patient's own, non-engineered CD8 T cells also take part, and whether tumor genetics decide if they can ^[1].

The finding

Resistance tracked with losing the machinery that shows antigens

The authors sequenced tumor biopsies from patients with relapsed or refractory large B-cell lymphoma who received CD19-directed CAR T cells. Two features stood out: biallelic (both copies) inactivation of the genes for MHC class I, the molecules that display protein fragments to CD8 T cells, and biallelic inactivation of CD58, a partner molecule that helps T cells form a stable contact with their target ^[1].

The preprint reports that all patients with permanent loss of MHC class I or CD58 were refractory to CAR T therapy (p<0.0001, log-rank test). Losing only one copy of either gene, which was common, did not associate with outcome. In a multivariate Cox model that included sex, age, metabolic tumor volume and genomic complexity, loss of MHC class I or CD58 remained independently associated with progression (hazard ratio 5.73, p=0.008) ^[1], Figure 3.

A CAR does not need MHC class I to recognize CD19, so MHC class I loss alone would not be expected to protect a tumor from the engineered cells. Yet tumors with this loss resisted treatment in the clinic. The authors' reading is that the patient's own T cells contribute, which does depend on MHC class I.

Mouse tumors needed MHC class I and host CD8 T cells

The authors then turned to mouse lymphoma models with intact immune systems. In the A20 model, tumors engineered to lack B2m, a required part of MHC class I, were completely insensitive to CAR T treatment in mice, even though they engrafted more slowly and so faced a higher ratio of CAR T cells to tumor cells at infusion. In a dish, the same B2m-knockout cells stayed sensitive to direct CAR T killing, and were somewhat more sensitive at lower ratios of CAR T cells to tumor cells ^[1].

To test the role of host CD8 T cells, the authors used the BL3750 lymphoma line in C57BL/6 mice, comparing wild-type hosts with Cd8-knockout hosts. After CAR T treatment, all Cd8-knockout animals had tumors by imaging and at necropsy. Wild-type hosts stayed free of tumor signal to day 90, although two wild-type mice died during the experiment without detectable tumor by imaging or at necropsy ^[1], Figure 4i-k and Supplementary Fig. 10.

The authors read these results as CAR T cells acting as a catalyst that recruits the host's existing anti-tumor response, in addition to killing directly. That is their interpretation across experiments.

The evidence

Gene loss and outcome in 54 patients

The cohort was 61 tumor biopsies with matched germline samples from 54 patients, treated with CD19-directed CAR T cells. Whole-genome sequencing covered all 61 samples and RNA sequencing covered 54. Median follow-up was 24 months, durable response occurred in 18 patients (33.3%), and median progression-free and overall survival were 7.85 and 16.2 months ^[1].

Biallelic MHC class I inactivation was found in 10 patients (7 before CAR T therapy), and biallelic CD58 inactivation in 5. Patients with a high neoantigen burden (15 or more predicted neoantigens, 12 patients) were enriched for biallelic loss of MHC class I or CD58 genes (p=0.0005) ^[1]. One possible explanation, which the preprint does not test, is that tumors with more visible mutant antigens face stronger pressure to hide them. These are associations in a single treated cohort, and they do not show that gene loss caused the failures.

Do the patient's T cells actually respond after infusion?

To test activation directly, the authors used ELISPOT assays, which count cells releasing the signaling molecule IFN-gamma when challenged with tumor. In 5 patients, post-infusion samples produced significantly more IFN-gamma in response to the patient's own tumor than baseline samples did (p=0.039, two-way ANOVA). In treated mice, endogenous CD4 and CD8 T cells likewise made more IFN-gamma than cells from tumor-only controls (p<0.05) ^[1]. Five patients is a small sample, and the assay shows activation, not that these cells caused the remissions.

Tumor immune makeup and rechallenge in mice

Single-cell RNA sequencing of 46,873 cells from eight A20 tumors compared four wild-type tumors with four B2m-knockout tumors. The MHC class I-competent wild-type tumors had higher frequencies of CD4 central memory and CD8 effector memory cells, while the B2m-knockout tumors skewed toward immunosuppressive states such as Th2/Th17 CD4 cells and showed more monocyte infiltration ^[1] and Figure 4e-h.

In a rechallenge experiment, mice in durable remission received repeated CD19-knockout tumor cells at escalating doses across three independent experiments. These tumor cells lack the CAR target, so protection would have to come from something other than the CAR. Previously treated mice survived significantly longer than controls (p=0.0081, log-rank test) ^[1], Figure 7. This is consistent with an immune memory against the tumor that does not rely on CD19.

Simulating joint recognition through CD19 and MHC class I

The authors also built a computational model in which CAR T cells and tumor-resident endogenous CD8 T cells kill cooperatively. Its parameters came from an independent dataset, and the neoantigen comparison used a separate cohort. Simulated tumors with high CD19 and high MHC class I expression were eliminated, while those lacking MHC class I or neoantigen presentation were more likely to progress. The authors state that the model explores biological plausibility and is not meant to predict outcomes for individual patients ^[1].

Because this is a preprint, none of these results has been peer reviewed. Mouse lymphomas differ from human disease, and the authors note that the specific tumor neoantigens recognized by endogenous T cells need further investigation ^[1].

Future directions

The central unresolved question is whether host CD8 T cells are needed in patients, or only in these mouse models. The patient data are correlative, so a direct test would follow patients treated with CAR T cells and measure tumor-reactive CD8 clones in blood before and after infusion, then ask whether clone expansion predicts durable response independently of CAR T cell expansion. This is a proposal, since the preprint's patient immune measurements were limited to the 5-patient ELISPOT comparison.

A second gap is the identity of the antigens. Mapping which mutant peptides endogenous T cells recognize, and whether those peptides are shared across patients with high neoantigen burden, would show whether the cooperation depends on a few common targets or on each tumor's private mutations.

The authors conclude that new treatments should aim to enhance CAR T cells' ability to stimulate host immune responses, not only to maximize their direct killing. One experiment that could connect this to the MHC class I result, proposed here and not performed in the preprint, is to re-express B2m in the B2m-knockout A20 model and ask whether CAR T sensitivity returns in mice, and how large a fraction of tumor cells must regain it.

Last, whether a pretreatment biopsy readout of MHC class I, CD58 and neoantigen status could sort patients prospectively, in a cohort not used to define those features, remains untested.