CAR T-cell therapy and tumour-infiltrating lymphocyte (TIL) therapy are both forms of adoptive cell therapy: a patient's own immune cells are collected, prepared in a manufacturing facility, and infused back to attack the cancer. They are "living drugs" given once, and their clinical trials look very different from trials of conventional medicines.
This article explains how the two therapies work, how their trials are designed, which endpoints are used, and how the data are analysed. It then works through a complete example on a dummy dataset with SAS code, output and interpretation.
How CAR T and TIL therapy work
CAR T cells are T cells taken from the patient's blood and genetically engineered to carry a chimeric antigen receptor (CAR). The receptor recognises a specific protein on the surface of cancer cells, such as CD19 on B-cell lymphomas and leukaemias or BCMA on myeloma cells.
TILs are T cells that have already found their way into the patient's tumour. A piece of tumour is surgically removed, the lymphocytes inside it are grown to very large numbers in the laboratory, and they are infused back. They are not genetically modified, and they recognise many different tumour targets through their natural T-cell receptors.
| Feature | CAR T-cell therapy | TIL therapy |
|---|---|---|
| Source of cells | Blood, collected by leukapheresis | Tumour tissue, collected by surgical resection |
| Genetic engineering | Yes: a CAR gene is inserted, usually with a viral vector | No: naturally occurring cells are selected and expanded |
| Target recognition | One defined surface antigen (for example CD19, BCMA) | Many tumour antigens, including patient-specific neoantigens |
| Main use so far | Blood cancers: lymphoma, leukaemia, multiple myeloma | Solid tumours: first approved in advanced melanoma |
| Treatment regimen | Lymphodepleting chemotherapy, then a single infusion | Lymphodepleting chemotherapy, a single infusion, then interleukin-2 |
| Characteristic toxicities | Cytokine release syndrome (CRS), neurotoxicity (ICANS), prolonged low blood counts | Effects of lymphodepletion and interleukin-2: low blood counts, fever, low blood pressure, capillary leak |
| Response criteria | Lugano (lymphoma), IMWG (myeloma) | RECIST 1.1 |
How cell therapy trials differ
- Single-arm pivotal trials. Many approvals have been based on single-arm Phase II trials in patients with no remaining standard options, where a randomised control is difficult. The result is compared with a pre-specified historical response rate.
- A gap between enrolment and treatment. Manufacturing takes weeks. Some enrolled patients never receive the product because of manufacturing failure, disease progression or death. This creates two analysis populations: all enrolled patients and all infused patients.
- One-time treatment. There is no dosing period in the usual sense. Efficacy and safety are measured from the date of infusion.
- Durability matters as much as response. A high response rate is of limited value if responses are short. Duration of response is a key endpoint.
- Long-term follow-up. Recipients of genetically modified cells are followed for up to 15 years for delayed risks.
Endpoints
| Endpoint | Definition | Role |
|---|---|---|
| Objective response rate (ORR) | Proportion of patients with a best overall response of complete response (CR) or partial response (PR) | Primary endpoint in most single-arm trials |
| Complete response rate | Proportion of patients with a best overall response of CR | Key secondary; primary in some blood cancer trials |
| Duration of response (DOR) | Time from first response to progression or death, in responders only | Key secondary; shows whether benefit lasts |
| Progression-free and overall survival | Time from infusion (or enrolment) to progression or death, or to death | Secondary; primary in randomised trials |
| CRS and ICANS | Incidence and worst grade per patient, graded by ASTCT consensus criteria | Key safety endpoints for CAR T |
Statistical concepts behind the analysis
1. Testing against a historical control rate
With no control arm, the hypothesis is framed against a fixed benchmark: the response rate expected from existing treatment, p0. The null hypothesis is that the true ORR is no better than p0. The trial succeeds if the lower limit of the confidence interval for the ORR is above p0. The benchmark must be justified from published data and agreed in advance, because the whole conclusion depends on it.
2. Exact binomial methods
The number of responders follows a binomial distribution. With 30 to 100 patients, large-sample approximations are unreliable, so the Clopper-Pearson exact confidence interval and the exact binomial test are standard. A one-sided test at the 2.5% level corresponds to the lower limit of the two-sided 95% confidence interval.
3. Analysis populations
The infused population (often called modified intention-to-treat) includes patients who received the product and answers the question "how well does the product work once given?". The enrolled population (intention-to-treat) includes everyone who underwent cell collection, counting patients who were never infused as non-responders. It answers "how well does the whole treatment strategy work?" and is always lower. Regulators expect both.
4. Best overall response
Each patient is assessed repeatedly, and the best response across all assessments is used. Patients with no evaluable post-infusion assessment are counted as non-responders, not excluded.
5. Duration of response is a responder-only analysis
DOR is estimated with the Kaplan-Meier method among responders. Responders who are still in response at the data cut-off are censored. Because the group is defined by an outcome that occurs after treatment, DOR describes the responders but cannot be used for a formal comparison between treatments.
6. Follow-up maturity
Median DOR can be estimated only when enough responders have progressed. Early analyses often report the median as "not reached" or give a confidence interval with no upper limit. Landmark estimates, such as the proportion still in response at 6 and 12 months, are then more informative than the median.
7. Safety by worst grade
For CRS and ICANS, each patient is counted once at their worst grade. The incidence of any-grade and of grade 3 or higher events is reported with an exact confidence interval, along with time to onset and duration.
Study design and dummy data
The example is a single-arm Phase II trial of an autologous cell therapy in patients with relapsed or refractory disease. Forty patients were enrolled and underwent cell collection; 34 received the infusion. The primary endpoint is ORR in infused patients, tested against a historical rate of 20%. The same analysis applies to a CAR T or a TIL product; only the response criteria and the relevant toxicities change. The values are simulated and are for illustration only.
| Variable | Meaning |
|---|---|
| USUBJID | Unique subject identifier |
| INFFL | Infused flag: Y = received the cell product, N = enrolled but not infused |
| BOR | Best overall response: CR, PR, SD (stable disease), PD (progressive disease), NE (not evaluable) |
| DOR | Duration of response in months (responders only) |
| CNSR | Censoring flag for DOR: 0 = progression or death, 1 = censored (still in response) |
| CRSGR | Worst grade of cytokine release syndrome (0 = none) |
| ICANSGR | Worst grade of neurotoxicity (0 = none) |
data adcell;
length usubjid $8 inffl $1 bor $2 respfl crfl $1;
input usubjid $ inffl $ bor $ dor cnsr crsgr icansgr;
/* Responder flags: patients without CR/PR count as non-responders */
if bor in ('CR', 'PR') then respfl = 'Y'; else respfl = 'N';
if bor = 'CR' then crfl = 'Y'; else crfl = 'N';
/* Safety flags (infused patients only) */
if inffl = 'Y' then do;
crs_any = (crsgr ge 1);
crs_g3 = (crsgr ge 3);
icans_any = (icansgr ge 1);
icans_g3 = (icansgr ge 3);
end;
datalines;
CT-001 Y SD . . 0 0
CT-002 Y CR 9.3 1 0 0
CT-003 Y CR 7.0 1 1 0
CT-004 N . . . . .
CT-005 Y PD . . 0 0
CT-006 N . . . . .
CT-007 Y SD . . 2 1
CT-008 Y PR 8.5 0 2 0
CT-009 Y PD . . 1 3
CT-010 Y PR 3.2 0 0 1
CT-011 N . . . . .
CT-012 Y PR 13.0 1 1 0
CT-013 Y PD . . 0 1
CT-014 Y PD . . 4 1
CT-015 Y SD . . 1 2
CT-016 Y CR 5.6 1 1 0
CT-017 Y PD . . 2 0
CT-018 Y CR 8.3 1 2 0
CT-019 Y SD . . 1 0
CT-020 N . . . . .
CT-021 Y CR 6.6 1 2 0
CT-022 Y CR 5.5 1 2 0
CT-023 Y CR 5.4 1 1 0
CT-024 N . . . . .
CT-025 Y SD . . 3 0
CT-026 Y PR 3.4 0 1 0
CT-027 Y PR 2.0 0 1 2
CT-028 Y PR 12.3 1 3 0
CT-029 Y NE . . 3 0
CT-030 Y SD . . 1 0
CT-031 Y PR 0.9 0 2 3
CT-032 Y SD . . 1 1
CT-033 Y PD . . 0 0
CT-034 Y CR 5.2 0 1 0
CT-035 Y SD . . 2 0
CT-036 Y PR 4.9 0 1 0
CT-037 Y CR 7.1 1 2 3
CT-038 Y PR 4.7 0 3 1
CT-039 N . . . . .
CT-040 Y PR 6.9 0 1 2
;
run;
SAS code, step by step
Step 1: Best overall response
proc freq data=adcell;
where inffl = 'Y';
tables bor / nocum;
run;
Step 2: Primary analysis of ORR in infused patients
The BINOMIAL option requests the proportion of responders; CL=EXACT gives the Clopper-Pearson interval, P=0.20 sets the historical rate, and the EXACT statement requests the exact binomial test.
proc freq data=adcell;
where inffl = 'Y';
tables respfl / binomial(level='Y' p=0.20 cl=exact) alpha=0.05;
exact binomial;
run;
Step 3: Complete response rate
proc freq data=adcell;
where inffl = 'Y';
tables crfl / binomial(level='Y' cl=exact) alpha=0.05;
run;
Step 4: Sensitivity analysis in all enrolled patients
Removing the WHERE statement includes the six patients who were not infused. They have no response, so RESPFL is N.
proc freq data=adcell;
tables respfl / binomial(level='Y' p=0.20 cl=exact) alpha=0.05;
exact binomial;
run;
Step 5: Duration of response
ods graphics on;
proc lifetest data=adcell method=km conftype=loglog
plots=survival(atrisk(outside)=0 to 15 by 3)
timelist=6 12 reduceout outsurv=dor_rates;
where respfl = 'Y';
time dor*cnsr(1);
run;
Step 6: CRS and neurotoxicity
/* Distribution by worst grade */
proc freq data=adcell;
where inffl = 'Y';
tables crsgr icansgr / nocum;
run;
/* Any-grade and grade 3 or higher with exact confidence intervals */
proc freq data=adcell;
where inffl = 'Y';
tables crs_any crs_g3 icans_any icans_g3
/ binomial(level='1' cl=exact) alpha=0.05;
run;
Output
Patient disposition
| Population | n |
|---|---|
| Enrolled (cell collection performed) | 40 |
| Infused | 34 |
| Not infused | 6 |
Best overall response (infused patients, N=34)
| Best overall response | n | % |
|---|---|---|
| Complete response (CR) | 9 | 26.5 |
| Partial response (PR) | 10 | 29.4 |
| Stable disease (SD) | 8 | 23.5 |
| Progressive disease (PD) | 6 | 17.6 |
| Not evaluable (NE) | 1 | 2.9 |
Response rates
| Endpoint | Population | n/N | Rate (%) | Exact 95% CI (%) | Exact one-sided p vs 20% |
|---|---|---|---|---|---|
| ORR (primary) | Infused | 19/34 | 55.9 | 37.9 to 72.8 | < 0.0001 |
| Complete response rate | Infused | 9/34 | 26.5 | 12.9 to 44.4 | Not tested |
| ORR (sensitivity) | Enrolled | 19/40 | 47.5 | 31.5 to 63.9 | < 0.0001 |
Duration of response (responders, N=19)
| Statistic | Value |
|---|---|
| Responders with progression or death, n (%) | 9 (47.4) |
| Responders censored (still in response), n (%) | 10 (52.6) |
| Median DOR, months (95% CI) | 8.5 (4.7, not estimable) |
| Proportion in response at 6 months, % (95% CI) | 63.2 (37.9, 80.4) |
| Proportion in response at 12 months, % (95% CI) | 41.4 (13.9, 67.5) |
CRS and neurotoxicity (infused patients, N=34)
| Event | Grade 1 | Grade 2 | Grade 3 | Grade 4 | Any grade, n (%) [95% CI] | Grade 3 or higher, n (%) [95% CI] |
|---|---|---|---|---|---|---|
| CRS | 14 | 9 | 4 | 1 | 28 (82.4) [65.5, 93.2] | 5 (14.7) [5.0, 31.1] |
| ICANS | 6 | 3 | 3 | 0 | 12 (35.3) [19.7, 53.5] | 3 (8.8) [1.9, 23.7] |
Interpretation
- Primary endpoint. Among the 34 infused patients, 19 responded: an ORR of 55.9% with an exact 95% confidence interval of 37.9% to 72.8%. The lower limit is well above the historical rate of 20%, and the exact binomial test gives p < 0.0001. The trial meets its primary objective.
- Depth of response. Nine patients (26.5%) achieved a complete response. In cell therapy, complete responses tend to be the most durable, so this rate is watched closely.
- All enrolled patients. Counting the six patients who were never infused as non-responders, the ORR is 47.5% (95% CI 31.5% to 63.9%). The lower limit is still above 20%, so the conclusion holds for the whole treatment strategy, not only for patients who reached infusion.
- Durability. The median duration of response is 8.5 months, with an estimated 63.2% of responders still in response at 6 months and 41.4% at 12 months. The upper confidence limit of the median is not estimable because more than half of the responders are still in response and censored. The 12-month estimate has a very wide interval (13.9% to 67.5%) and is based on few patients still at risk.
- Who is still responding. In the data, 8 of the 9 complete responders are censored, while 8 of the 10 partial responders have progressed. Durable benefit is concentrated in patients who achieve a complete response.
- Safety. CRS of any grade occurred in 82.4% of patients, but most was grade 1 or 2; grade 3 or higher CRS occurred in 14.7%. Neurotoxicity occurred in 35.3%, with 8.8% at grade 3 or higher. With 34 patients the intervals are wide: the true rate of severe CRS could plausibly be as low as 5% or as high as 31%.
- Overall conclusion. In this dummy example the cell therapy shows a response rate far above the historical benchmark, with a meaningful complete response rate. Durability data are not yet mature, and longer follow-up is needed before the duration of benefit can be stated with confidence.
Why the two ORR values differ. The ORR falls from 55.9% to 47.5% when all enrolled patients are included. The difference reflects the 15% of patients (6 of 40) who did not reach infusion. A long or unreliable manufacturing process widens this gap, which is why both figures are reported and why manufacturing time is itself an important outcome.
A caution on single-arm evidence. The comparison with 20% is only as good as the historical data behind it. Differences in patient selection, prior treatments and response assessment between this trial and the historical studies can bias the comparison. This is the main reason randomised trials are increasingly expected as cell therapies move into earlier lines of treatment.
Practical points
- Independent review. Response is usually assessed by an independent review committee for the primary analysis, with investigator assessment as a sensitivity analysis.
- Time origin. State clearly whether time-to-event endpoints start at enrolment, cell collection or infusion. The choice changes the result.
- Bridging therapy. Treatment given between cell collection and infusion can itself produce a response. Record it and repeat the baseline assessment before infusion.
- Minimum follow-up. Pre-specify a minimum follow-up for all responders (for example 6 months) before the primary analysis, so that durability can be judged.
- Product characteristics. Cell dose, cell composition and expansion in the patient are often explored as predictors of response and toxicity.
- Visualisation. Swimmer plots (one bar per patient showing response and duration) and waterfall plots (best change in tumour size) are standard displays for these trials.
References
- US FDA. Considerations for the Development of Chimeric Antigen Receptor (CAR) T Cell Products, Guidance for Industry. 2024.
- US FDA. Clinical Trial Endpoints for the Approval of Cancer Drugs and Biologics, Guidance for Industry. 2018.
- US FDA. Long Term Follow-Up After Administration of Human Gene Therapy Products, Guidance for Industry. 2020.
- Lee DW, Santomasso BD, Locke FL, et al. ASTCT consensus grading for cytokine release syndrome and neurologic toxicity associated with immune effector cells. Biology of Blood and Marrow Transplantation. 2019;25(4):625–638.
- Clopper CJ, Pearson ES. The use of confidence or fiducial limits illustrated in the case of the binomial. Biometrika. 1934;26(4):404–413.
The dataset in this article is simulated and is provided for training purposes only.




