Ask a patient what they want to know about CAR-T and the answer is rarely "how does it work". It is some version of: will it work for me, and how badly will it hurt. Ten years and seven approved products into the field, those two questions still do not have a general answer, and a review published in Cancer Cell makes a compelling case that they never will — not because the science has failed, but because the field has been asking them the wrong way.
The review, by Deniz Usta and Crystal Mackall at Stanford, is titled plainly: toward the next generation of CAR T cell therapies for cancer. Mackall has been central to this field since before the first approval, and the paper reads like someone taking stock rather than someone selling a direction. It is organised around three lessons from the clinic and three waves coming next, and the sentence that holds the whole thing together is this: no single CAR design can serve all cancers, and every disease needs a purpose-built engineering solution.
That sounds like a platitude until you follow it into the data. What it means in practice is that "we found a new target" is not the news it sounds like, and that a therapy which works brilliantly in lymphoma can be lethal in a solid tumour carrying the same antigen — not because the antigen changed, but because everything around it did.
The short version
The count: seven products approved by the US FDA across fifteen indications, plus more than 2,600 trials worldwide. That is the global regulatory tally, not a count of what any one patient can access.
Lesson one: a target's value is set by its therapeutic window — where it sits, how densely, how stably, and what happens to normal tissue that carries it — not by whether the tumour expresses it.
Lesson two: toxicity is predictable in its mechanisms but not uniform in its degree. Two entities most patients have never been warned about are IEC-HS and TIAN, and neither behaves like ordinary CRS.
Lesson three: every design choice is a trade. Costimulatory domain, signalling strength, binder shape and armouring all buy one property by paying for it with another.
The waves: logic-gated cells, timed regimens rather than single agents, and manufacturing the CAR-T inside the patient's body. Only the second is changing practice today.
Seven, Nine, or Fifteen: Read the Number Carefully
Before anything else, a note on the arithmetic, because it is the most common source of confusion in conversations about this field. The review counts seven approved products and fifteen indications. Those are FDA approvals. China's regulator has approved nine CAR-T products, several of which do not exist in the American market, and one of which — satri-cel, marketed as Kailimei — is the first CAR-T approved anywhere in the world for a solid tumour.
So "seven" and "nine" are both correct and neither is the number a patient needs. What matters is the count in the jurisdiction where treatment will happen, for the specific disease in question, at the specific line of therapy. A global total is a marketing figure. A local approval is a clinical fact.
The efficacy figures in the review are worth having in mind, though, because they show what the technology can do when the match is right. In relapsed or refractory B-cell acute lymphoblastic leukaemia, CD19 CAR-T produces complete responses in roughly 90% of patients. In aggressive B-cell lymphoma the figure falls to 40–54%. In heavily pretreated multiple myeloma, cilta-cel achieves strict complete response in 67%. Move the same therapy earlier and the gap widens further: in KarMMa-3, ide-cel produced complete responses in 39% of patients against 5% on standard regimens; in CARTITUDE-4, cilta-cel reached 73.1% against 21.8%.
Those are not small differences between products. They are differences produced mostly by disease and by timing, with the product held roughly constant. That pattern runs through everything below.
A Target Is Not a Switch. It Is a Window.
The first lesson is the one that changes how you read every CAR-T headline afterwards. Whether a target is good has almost nothing to do with whether the tumour expresses it. Tumours express a great many things. What decides the matter is the therapeutic window: how the antigen is distributed across normal tissue, at what density, how stable that density is, and what the clinical consequence is when a CAR-T cell finds the antigen somewhere it should not have.
CD19 became the gold standard for three reasons that have little to do with cancer. It sits at uniformly high density on B cells at every stage of development. Losing the entire normal B-cell compartment is clinically survivable and, in most patients, reversible with immunoglobulin replacement. And CD19 is not required for the malignant cell's survival — which means the tumour cannot simply downregulate it and walk away. The technical term in the review is a disposable lineage, and it is the reason CD19 CAR-T can be simultaneously very effective and reasonably safe. BCMA follows the same logic in myeloma, being almost exclusive to plasma cells.
Even gold standards have edges. Low CD19 expression on the vascular wall in the brain has been linked to ICANS, the neurological syndrome that follows some CD19 therapies. BCMA appears on neurons in the basal ganglia, and that expression has been associated with the delayed parkinsonian syndrome reported after BCMA-directed treatment. This is the category the field calls on-target, off-tumour toxicity, and it is unavoidable in principle: the CAR-T cell is doing exactly what it was designed to do, in the wrong place.
The second-line targets are now established on the same reasoning. CD22 produced a 68% overall response rate and 53% complete response rate in large B-cell lymphoma that had progressed after CD19 CAR-T, which earned it a settled place as the second lineage target. CD20, CD79b and BAFF-R have followed into the clinic, and GPRC5D has shown activity in myeloma after BCMA failure.
Ten Patients, Ten Bone Marrow Failures
The clearest demonstration of why the window matters is also the grimmest passage in the review. For myeloid leukaemia, the obvious targets — CLEC12A and CLL1 — are shared with normal hematopoietic stem cells. In the first-in-human study, all ten patients developed severe pancytopenia, and two died of infection.
Nothing malfunctioned. The construct worked as specified. When a target is shared with the cell population that produces blood, the safety margin does not shrink — it disappears, and it takes the platelet and neutrophil counts with it. T-cell cancers present the mirror-image problem: CD5 and CD7 CAR-T cells must survive an environment in which they recognise each other, and deep, sustained depletion of the normal T-cell compartment remains unresolved. TRBC1 offers an elegant way around this by targeting only the malignant TRBC1-positive clone while sparing TRBC2-positive normal T cells, but the data are still early.
This is the part of the review a patient should read twice, because it explains a decision that otherwise looks arbitrary: why a drug company would abandon a target that clearly works in the laboratory.
Solid Tumours: The Missing Piece Is Not the Antigen
Solid tumours have resisted CAR-T for a decade, and the usual explanation — that they lack a suitable antigen — is only half right. They lack a safe antigen, which is a different problem with a different solution.
The review lays out the evidence in a way that makes the pattern unmistakable. Four targets, four failures, and four different reasons:
| Target | What the first attempt did | What changed it |
|---|---|---|
| CAIX (kidney cancer) | Four of eight patients developed grade 2–4 liver toxicity, because the CAR-T cells recognised CAIX on bile duct epithelium; treatment was stopped | A blocking antibody given before infusion abolished the toxicity — proof that the mechanism was on-target, off-tumour, and that a cell therapy can be de-toxified pharmacologically |
| HER2 (sarcoma, glioma) | Early high-dose constructs caused fatal lung toxicity | A CD28 costimulatory domain combined with cautious dose escalation made later products safe in sarcoma and glioma |
| Mesothelin (mesothelioma) | Intravenous delivery caused fatal lung toxicity | Delivering into the pleural space instead, together with pembrolizumab, was safe and showed antitumour activity |
| ROR1 (CLL, mantle cell) | Fourteen of twenty-one patients in phase I developed grade 3–4 lung toxicity; an enhanced version caused one fatal respiratory failure | Not solved. This target has not been salvaged |
Read the middle column against the right one and the lesson is hard to miss. HER2 was lethal and then became manageable without the target changing at all — what changed was the costimulatory domain and the dose. Mesothelin was lethal by one route and usable by another. "Is this a good target" is therefore an incomplete question; the complete one is "is this a good target for this construct, at this dose, by this route, in this disease".
The review's own summary is blunt: solid tumours do not lack an expressed antigen, they lack a safe target. That conclusion is what directly produced the engineered solutions in the next section — if no single antigen is safe enough, build cells that require two.
Claudin18.2 and the First Real Win
The most instructive success in the review is not the most publicised one. It is Claudin18.2, and the reason it works is a piece of tumour biology that has nothing to do with immunology.
In normal stomach lining, Claudin18.2 sits buried inside tight junctions — physically folded away where an antibody or a CAR-T cell cannot reach it. When a cell undergoes malignant transformation, it loses its polarity, the tight junctions come apart, and the antigen is suddenly exposed on the surface. The tumour effectively hands over a target that healthy tissue keeps locked in a drawer.
This is the logic behind satri-cel, the Chinese-developed Claudin18.2 CAR-T that became the first approved CAR-T for a solid tumour anywhere. In its phase I study the overall response rate was 38.8%, disease control 91.8%, and median progression-free survival 4.4 months. The randomised phase II trial, CT041-ST-01, is the part that matters historically: median progression-free survival of 3.25 months against 1.77 months for physician's choice of therapy, a hazard ratio of 0.37.
That is the first positive randomised controlled trial of CAR-T in a solid tumour. Not the largest effect ever reported, and not a cure — but a result that survives randomisation, which nothing in this space had previously produced. The ceiling shows up in the toxicity profile: gastric mucosal damage, predictable from where the antigen lives, and a reminder that a window can be narrow without being closed.
Other routes to the same end are less elegant but real. Some targets are carried by the mutation itself: in H3K27M-mutant diffuse midline glioma, the mutation-driven epigenetic disruption makes GD2 uniformly expressed, and GD2 CAR-T given intravenously or into the ventricle produced significant tumour regression, including one durable complete response. In neuroblastoma, GD2-CART01 was tolerable, persisted beyond twelve months in 64% of patients, and produced five-year overall survival of 43% across the cohort and 68% in the dose-optimised group. Oncofetal antigens — GPC2, GPC3, CLDN6 — form a third category: expressed during development, silenced at birth, switched back on in cancer. And a fourth strategy stops targeting the tumour altogether and targets its accomplices, with TREM2 and FOLR2 directed at immunosuppressive macrophages and uPAR at tumour stroma.
One methodological point in the review deserves more attention than it usually gets. CAR-T activation is threshold-dependent, which means a transcriptomic readout of "positive" does not establish that antigen density is high enough to trigger killing. In lymphoma, low baseline CD19 density is already associated with progression. An IHC report saying the target is present is a starting point, not an answer.
Toxicity Is Predictable. It Is Not Uniform.
The second lesson is the most immediately useful at the bedside, and it starts from an uncomfortable observation: toxicity and efficacy travel together, because both are set by the strength of the encounter between product, tumour burden and host state. In B-ALL trials, higher baseline leukaemia burden predicts more severe cytokine release syndrome. Burden-adapted split dosing reduces it. CAR-T expansion kinetics correlate with CRS and ICANS but not necessarily with response — meaning toxicity has its own rules, independent of benefit.
Mechanistically, CRS is not something the CAR-T cell does alone. The CAR-T cell is the trigger; monocytes and macrophages, with the downstream IL-1, IL-6 and nitric oxide pathways, are the amplifier. This distinction is why cytokine-directed treatment can reduce toxicity without entirely sacrificing efficacy — you are turning down the amplifier rather than disarming the cell. Prophylactic steroids, tocilizumab and anakinra have all shown that early intervention can lower severe toxicity without an obvious efficacy cost, though where exactly to place them is still unsettled.
Two entities beyond CRS and ICANS deserve to be known by name, because both are routinely mistaken for something they are not.
IEC-HS: The One That Arrives as CRS Leaves
Immune effector cell-associated haemophagocytic lymphohistiocytosis-like syndrome typically appears just as CRS is improving, which is precisely why it gets missed. The signature is a ferritin spike, persistent cytopenias, rising transaminases and low fibrinogen. Tocilizumab alone is essentially ineffective against it; management requires corticosteroids combined with HLH-directed therapy such as ruxolitinib or emapalumab. A patient whose fever settles and whose blood counts then fail to recover is not on a normal trajectory, and the distinction matters within hours rather than days.
TIAN: Not ICANS, Despite the Address
Tumour inflammation-associated neurotoxicity occurs when CAR-T cells are delivered locally into central nervous system tumours, and it presents as pseudoprogression, oedema and mass effect. It is a different entity from ICANS, with a different mechanism and a different management algorithm. The GD2 experience in diffuse midline glioma makes the contrast concrete: intravenous high-dose delivery was limited by CRS, while intraventricular delivery had no dose-limiting toxicity but produced TIAN frequently, requiring its own monitoring protocol. For families reading scan reports after intraventricular therapy, the difference between "the tumour grew" and "the tumour is inflamed" is the difference between abandoning a working treatment and continuing it.
The field's response is to design safety in rather than manage it afterwards: the iCasp9 elimination switch, validated in GD2-CART01 where activating it in one patient cleared the CAR-T cells rapidly and brought toxicity under control; SNIP-CAR constructs that a drug can dial up or down without permanently terminating the cell; affinity tuning, such as a micromolar-affinity ICAM-1 CAR for thyroid cancer that lets low-density normal tissue pass; and local delivery, which shrinks exposure anatomically.
On late toxicity, one finding is worth flagging for anyone comparing products. Delayed parkinsonian syndrome has been reported more often after cilta-cel than after ide-cel, despite both targeting BCMA — which indicates that the binder architecture itself can produce clinical differences. On secondary malignancies, the molecular analyses reviewed do not support CAR vector insertion as the principal cause, and continued monitoring is the proportionate response rather than a change to the overall risk-benefit judgement.
There Is No Master Key
The third lesson is engineering, and it is where a patient's intuition usually fails. It is natural to assume that a newer product is a better product, or that more engineering means more efficacy. The review's evidence says every design decision is a trade with a price attached, and the price is often paid in a currency the patient cares about.
The costimulatory domain sets the personality of the cell. CD28 pushes toward early expansion and effector differentiation, at the cost of faster exhaustion. 4-1BB produces a more memory-like state and longer persistence, and commercial CD19 products built on 4-1BB report lower rates of ICANS — though cross-product comparisons must be read cautiously, since more than the costimulation domain differs between them. The instructive anomaly is that first-generation CARs without any costimulatory domain failed almost everywhere, with one exception: GD2 in neuroblastoma, where a first-generation construct induced durable remissions confirmed at eighteen years of follow-up. A design that failed across a dozen diseases succeeded in one, because that one disease happened to reward what that design did well.
Signalling strength is tunable. A CD19 construct with reduced ITAM signalling produced promising activity with a low rate of severe toxicity across 28 patients with large B-cell lymphoma. Targeted knock-in at the TRAC locus puts CAR expression under endogenous control and reduces tonic signalling; the allogeneic product vispa-cel, combining TRAC knock-in with PD-1 knockout, achieved a 76.1% overall response rate and 45.7% complete response rate in 46 patients.
Binder architecture matters more than its obscurity suggests. scFv aggregation produces tonic signalling that accelerates exhaustion; single-domain binders such as nanobodies and DARPins aggregate less, and cilta-cel's tandem dual-VHH format adds avidity on top. Dissociation rate shapes clinical behaviour too: obe-cel uses a fast-off-rate CD19 binder and achieved 77% remission in adult B-ALL with grade 3 or higher CRS in only 2.4% and ICANS in 7.1% — a safety profile almost unheard of in that disease.
Armouring is the clearest double-edged sword in the review. Adding IL-15 to a GPC3 CAR-T increased expansion, intratumoural survival and activity — and increased CRS. A CD19 CAR-T engineered to secrete IL-18, given to patients who had progressed after a previous CD19 CAR-T, produced a three-month overall response rate of 81% and complete response rate of 52%, with mostly low-grade toxicity. But a PSMA CAR-T combined with a dominant-negative TGF-beta receptor in prostate cancer produced dose-limiting CRS and fatal neurotoxicity, and the programme was terminated. Enhancement amplifies inflammation as readily as it amplifies killing; controllability has to be built in at the same time, not added later.
The general rule is that design must match the specific obstacle the disease presents. Overexpressing c-JUN improves killing at low antigen density, which is exactly what you want in a tumour and exactly what you do not want if the target is also present at low density on normal tissue — the same modification that rescues efficacy can amplify off-tumour risk. Constructs that drive a strong effector state suit rapid debulking and suit poorly any setting that requires long-term immune suppression. The number of plausible combinations has outgrown the ability to test them one at a time, which is why systematic approaches such as CRISPR screening are becoming a prerequisite rather than a convenience.
Wave One: Cells That Must Be Convinced
The first wave is multifunctional engineering, and its most conceptually interesting branch is logic gating — making the CAR-T cell compute before it kills.
An IF-THEN gate, built on SynNotch, requires recognition of a primer antigen before the cell expresses a CAR against a second one. An AND gate, as in LINK-CAR, splits the activation machinery across two receptors so that full activation requires both antigens simultaneously. A NOT gate, such as Tmod, uses an inhibitory receptor that recognises HLA retained by normal tissue, so the cell stands down wherever healthy cells are present and fires only where tumour loss of heterozygosity has removed that brake. The validation of a CD93 NOT gate in acute myeloid leukaemia, where it protected endothelium, shows what these circuits are for: not making a good target better, but unlocking targets that were previously disqualified as unsafe.
Controllability runs on a parallel track. SNIP-CAR uses protease inhibitors to adjust activity in real time; SUPRA-CAR swaps or titrates targets through a soluble adapter; the AIR platform embeds a CD62L-derived motif in the CAR so that ADAM17 cleaves it on activation, scaling expression in proportion to tonic signalling — reducing exhaustion and increasing potency at once.
Multi-edit platforms are maturing alongside. Base, prime and epigenome editing avoid double-strand breaks; STASH-Select enriches cells carrying several transgenes in a single step; the allogeneic autoimmune products TyU19, with five gene knockouts, and YTS109, which adds TRAC site-specific integration, demonstrate that multiple perturbations and targeted insertion can coexist in one product.
The review's caution here is worth repeating verbatim in spirit: not every enhancement produces a net benefit. Armouring may amplify toxicity without improving efficacy. Multiple enhancements may antagonise each other. Logic gates and remote-control systems buy safety by spending some potency. Enhancement has to be hypothesis-driven and delivered in the context of a specific disease, or it is just complexity.
Wave Two: From a Drug to a Regimen
The second wave is the least glamorous and the only one already changing practice. CAR-T is unlikely to overcome every layer of resistance as a single agent, particularly in solid tumours and in relapsed disease, and the review borrows deliberately from the history of chemotherapy, radiotherapy and surgery: future outcomes will depend on the regimen, not on a single weapon.
Before infusion, fludarabine and cyclophosphamide lymphodepletion remains the foundation; ibrutinib improves T-cell fitness in CLL; bridging radiotherapy provides local control without adding CAR-T-related toxicity. To widen the proportion of tumours that can be targeted, gamma-secretase inhibitors prevent BCMA shedding and raise surface density, bryostatin-1 upregulates CD22, and EZH2 inhibitors raise GD2 in models of low expression — with the standing caveat that inducing expression on normal tissue narrows the very window you are trying to use.
After infusion, the aim is amplification. A CLDN6 CAR-T combined with an RNA vaccine, CARVac, has amplified CAR-T expansion in patients with solid tumours. In paediatric B-ALL, a phase II trial of sequential infusion of CD19-positive cells improved persistence without CRS or neurotoxicity. Long-acting IL-7, the IL-15 agonist NKTR-255, orthogonal IL-2 platforms and the IL-9/IL-9R axis offer more selective support, and CD20/CD3 bispecifics can activate both the CAR-T cells and the patient's own T cells at once.
To reverse suppression, checkpoint combinations have produced mixed results, with intrapleural delivery plus pembrolizumab in mesothelioma the bright spot. Epitope spreading is the more imaginative direction: EGFRvIII CAR-T induced host immunity against antigen-negative tumour in immunocompetent glioma models, and IL-18 or IL-36-gamma armouring connects to endogenous immunity in the same way.
Two findings in this section have direct implications for anyone planning treatment, and neither is about a drug. First, exposure to bendamustine before apheresis is associated with worse outcomes after CD19 CAR-T — an argument for collecting T cells early rather than after several lines of lymphodepleting therapy. Second, high tumour burden before infusion is associated with both worse survival and worse toxicity, a two-way relationship supported by a multicentre analysis of tisa-cel in paediatric and young adult B-ALL and by metabolic tumour volume analysis in axi-cel recipients. That is the evidential basis for moving CAR-T earlier, into MRD-positive and low-burden settings.
It also explains something patients notice and clinicians rarely say aloud: the same product, at the same dose, performs differently at different centres, because a regimen is a sequence of decisions and the sequence is not identical everywhere.
Wave Three: Building the Cells Inside the Body
The third wave is the one with the largest potential effect on access, and the smallest amount of mature data. Current manufacturing is ex vivo: cells are collected, engineered outside the body and returned. That defines the dose, permits complex engineering and enables quality control — and it is slow, expensive and individual. In a rapidly progressing disease the waiting period can be fatal, and prolonged culture pushes T cells toward a terminally differentiated state with poor persistence.
In vivo generation skips the laboratory. Three platform families are in play.
Lipid nanoparticle mRNA delivery is transient, titratable and repeatable: a CD8-targeted LNP carrying anti-CD19 CAR mRNA controlled tumour growth in humanised mice and depleted B cells in non-human primates, and a comparable product, CPTX2309, has entered phase I in B-cell mediated autoimmune disease. Transient expression is a limitation for cancer and a feature for autoimmunity, and circular RNA and self-amplifying RNA may extend the window.
Viral vectors offer persistence instead: ESO-T01, a T-cell-targeted lentiviral BCMA CAR, is the furthest advanced clinically; UB-VV111 adds a rapamycin-activated cytokine receptor to support selective expansion; INT2104 uses CD7 targeting to deliver a CD20 CAR to both T and NK cells. Among the most closely watched data, KLN-1010, a BCMA-directed phase I product, reported at ASCO 2026 that it achieved robust CAR-T expansion and remissions without lymphodepletion, and LB2501, a dual CD19/20 construct, reported dose-dependent expansion and encouraging responses in an EHA 2026 abstract.
A third family performs gene editing in situ, with Cas9 enveloped delivery vehicles achieving T-cell-selective delivery without hepatocyte transduction, EDV/AAV dual vectors completing promoterless CAR knock-in at TRAC, and LNP co-delivery of minicircle DNA with transposase mRNA achieving stable integration.
The open questions are as clearly stated as the progress. Delivery has to avoid hepatocytes, hematopoietic stem cells and tumour cells — and the last of those has a precedent: in the ex vivo era, leukaemic B cells were inadvertently transduced with CAR, producing antigen masking. With integrating systems, accidental tumour transduction is a far more serious prospect. Expansion needs to be sufficient without becoming uncontrolled. The role of lymphodepletion has to be redefined, since excessive depletion removes the very T cells you intend to modify. And whether the multi-component programmes that solid tumours require can be delivered with adequate fidelity remains an open engineering problem.
| Wave | What it does | Where it stands in 2026 | What it changes for a patient |
|---|---|---|---|
| Logic gating | Requires two antigens, or one antigen plus the absence of another, before the cell fires | Preclinical to early phase I; NOT gates validated in leukaemia models | Could open targets currently rejected as too dangerous — mostly a solid-tumour story |
| Regimens | Treats CAR-T as one timed component rather than a single drug | Already altering practice: apheresis timing, bridging, burden-adapted dosing | Explains why identical products give different results at different centres |
| In vivo generation | Builds the CAR-T inside the patient; no apheresis, no ex vivo culture | Phase I running; one product reports responses without lymphodepletion | Reach, speed and cost — not efficacy, at least not yet |
The Questions Patients Actually Ask
Should I wait for in vivo CAR-T?
For almost nobody reading this, no. The in vivo platforms are in phase I, their endpoints are safety and feasibility, and their most distinctive advantage so far is logistical rather than clinical — no apheresis, no weeks of manufacturing, and in one case no lymphodepletion. If you have a disease where an approved CAR-T already produces durable remissions, waiting converts a known benefit into an unknown one. The calculation changes only in specific situations: you cannot tolerate lymphodepletion, your T cells cannot be collected or manufactured, or your disease has no approved option and a trial is the alternative anyway.
If a product is newer, is it better?
Not automatically, and this review is the best available argument for the point. The clearest example is a first-generation construct with no costimulatory domain, which failed across a dozen diseases and still produced remissions durable at eighteen years in neuroblastoma. Meanwhile, a more heavily engineered product can lose ground: adding IL-15 to a GPC3 CAR-T increased activity and increased CRS in the same step, and a PSMA construct with an added TGF-beta trap was terminated for fatal neurotoxicity. Newer means differently traded, not better.
My pathology report says the target is positive. Is that enough?
It is necessary and not sufficient. CAR-T activation is threshold-dependent, so a transcriptomic or immunohistochemical "positive" does not establish that antigen density is high enough to trigger killing; in lymphoma, low baseline CD19 density is already associated with progression. What also matters is where else the antigen appears. Claudin18.2 works because healthy stomach tissue keeps it folded inside tight junctions. CLEC12A does not work because it is shared with the stem cells that make blood. Both are "positive".
Is the severity of my CRS decided before I start?
Largely, yes — which is why baseline assessment is not a formality. Severity tracks tumour burden at infusion, product characteristics and host state, and burden-adapted split dosing is used precisely because the first of those is modifiable. What is not fixed is the outcome: CRS is a multi-cell cascade in which the CAR-T cell is only the trigger, and interrupting the amplifier with IL-1 or IL-6 blockade reduces severity without necessarily reducing efficacy. Two things to know in advance are IEC-HS, which appears as CRS improves and does not respond to tocilizumab alone, and, for central nervous system tumours treated locally, TIAN, which mimics progression on imaging and is not ICANS.
Are any of these next-generation designs available in China now?
The most consequential one already is. Satri-cel — Claudin18.2 directed, approved for gastric and gastroesophageal junction adenocarcinoma — is the first CAR-T approved anywhere for a solid tumour, and it is a Chinese-developed product. Beyond that, access depends on trials rather than approvals: allogeneic and multi-edited products, in vivo platforms, and logic-gated constructs are largely in investigator-initiated studies at major centres. For an international patient the practical question is not which design is most advanced in principle but which approved product or open trial matches your disease, your line of therapy and your fitness for lymphodepletion.
What the Review Actually Concludes
Ten years of clinical experience compress into one sentence: engineering potency is not enough, and treatment has to match the biology of each disease. The review turns that into three requirements — choose the target by therapeutic window, design the receptor around a controllable signalling programme, and use model systems that can predict human behaviour.
Its forecasts are appropriately modest. Logic gating is expected to widen the therapeutic window where no single safe antigen exists, especially in solid tumours, and to produce synergistic benefit if combined with epitope spreading and platforms that recruit endogenous immunity. Bispecific T-cell engagers, being off-the-shelf and pharmacologically controllable, will remain a complementary modality, with the division of labour between cell therapy and antibody therapy decided by clinical trials rather than by preference. In vivo generation could substantially widen access, provided potency is not lost and toxicity is not added.
For a patient, the useful reframing is this. The question "is there a CAR-T for my cancer" has a short and often disappointing answer. The question "is there a construct whose therapeutic window fits my disease, at my disease stage, with a toxicity profile I can survive" is the one that produces a real decision — and it is a question that has to be asked of a specific centre, a specific product and a specific line of therapy, because the review's central finding is that those three things, not the target alone, determine the outcome.
A note on what this article is
This is a reading of a peer-reviewed review, not medical advice, and it deliberately omits dosing, eligibility criteria and product selection, all of which depend on details a review cannot supply. The review itself is: Usta D, Mackall CL. Toward the next generation of CAR T cell therapies for cancer. Cancer Cell. 2026;44(10):1936–1956.
Figures for approvals in China reflect the regulatory position as of late 2026 and change as new products are cleared. If you are weighing treatment, the numbers that matter are the ones your treating centre can confirm for your diagnosis.
Primary source: Usta D, Mackall CL. Cancer Cell. 2026;44(10):1936–1956. Supporting trials cited therein include KarMMa-3 (NEJM 2023), CARTITUDE-4 (NEJM 2023), CT041-ST-01 (Lancet 2025), CD22 CAR-T in LBCL (Lancet 2024), GD2 CAR-T in H3K27M glioma (Nature 2025), long-term GD2-CART01 outcomes (Nature Medicine 2025) and obe-cel in adult B-ALL (NEJM 2024).