T cell attacking cancer cells* Dose finding designs driven by immunotherapy outcomes with Application to a metastatic melanoma phase I trial Elizabeth Garrett-Mayer, PhD Cody Chiuzan, MS Hollings Cancer Center, MUSC SRCOS June 2012 * Azgad, The Cutting Edge, 2011 Dose Finding in Medical Research In cancer research, usually small studies: 6 – 30 patients is the norm ◦ not ethical to use ‘healthy volunteers’ ◦ few eligible patients ◦ small expectation of efficacy Due to ethical concerns, you must try doses sequentially Algorithmic designs are most common: 3 patients per dose and use predefined escalation and de-escalation rules. Rules are defined by a binary measure of toxicity: bad toxicity vs. no or acceptable toxicity. Model based designs arrived on the scene in 1990 ◦ most use toxicity as the outcome ◦ driven by assumed monotonic association between both dose and toxicity dose and efficacy 1.0 Assumption of classic dose finding designs in oncology DLT = doselimiting toxicity 0.8 0.6 0.4 0.2 0.0 Probability of Outcome Response DLT 1 2 3 4 Dose Level 5 6 7 Immunotherapies in cancer Immunotherapies often are expected to be nontoxic ◦ anecdotally, they are often not There is not strong rationale to assume that the highest tolerable dose is the optimal dose Standard algorithmic and model-based designs for dose finding based on binary measures of toxicity are inappropriate for identifying the optimal dose Efficacy-driven dose finding is more relevant, although safety concerns need to be incorporated Background Immunotherapy approaches can be ridiculously expensive ◦ cost may increase exponentially by dose level Unnecessary overdosing would be costly ◦ actual monetary costs ◦ possible non-monotonicity of association between dose and response ◦ safety needs to be considered Current status Most immunotherapy trials in cancer use a two step approach to dose finding 1. Perform an algorithmic design to identify safe doses 2. Collect immunological data and “explore” it to see if there appears to be an optimal dose Optimal dose? ◦ we imagine there will be a clear plateau in the association between dose and outcome. ◦ unrealistic and simplistic due to small sample sizes at each dose ◦ unrealistic and simplistic due to heterogeneity across patients Challenges with immunotherapy outcomes Usually continuous ◦ target levels often not known ◦ heterogeneity across patients Often not well-defined or described prior to the trial. There is not a clear link between clinical outcomes and the immunology “target” Assumption of monotonicity is not wellfounded Motivating Project “Transfer of Genetically Engineered Lymphocytes in Melanoma Patients - A Phase I Dose Escalation Study” Objective: To establish the recommended phase II dose of autologous T cell receptor (TCR) transduced T cells when administered with low dose IL-2 to stage IV melanoma patients following a non-myeloablative and lymphodepleting chemotherapy preparative regimen. PI’s: Mike Nishimura and David Cole Part of P01 Program Project Grant (funded Aug 1, 2011) Adoptive T-Cell transfer Tumor Infiltrating Lymphocytes (TIL) ◦ Tumor infiltrating lymphocytes are white blood cells that have left the bloodstream and migrated into a tumor. ◦ They are an important prognostic factor in melanoma,higher levels being associated with a better outcome. Adoptive cell transfer uses T cell-based cytotoxic responses to attack cancer cells. ◦ T cells that have a natural or genetically engineered reactivity to a patient's cancer are generated in vitro and then transferred back into the cancer patient ◦ This can be achieved by taking T cells that are found with the tumor of the patient, which are trained to attack the cancerous cells. ◦ These T cells are referred to as tumor-infiltrating lymphocytes (TILs) Expansion: TILs are multiplied in vitro These T cells are then transferred back into the patient along with IL2 to act as a growth factor for T cells Cartoon version* Strategy towards adoptive cell transfer with genetically modified T cells. 1) Extractions of T cells from the patient. 2) Transfection of an rationally optimized TCR in those cells using a viral vector. 3) Optional expansion. 4) Lymphodepletion of the patient. 5) Reinfusion of the modified T cells to the patient. Adapted from Olivier Michielin: http://www.nccr-oncology.ch/scripts/index.aspx?idd=134 Original Trial Design Subjects will receive a single infusion of autologous bulk TIL 1383I TCR transduced T cells supported with low dose IL-2. Four cohorts of 3 patients will be treated with increasing doses of TIL 1383I TCR transduced T cells ◦ ◦ ◦ ◦ cohort 1: 2x108 TIL 1383I TCR transduced T cells cohort 2: 5x108 TIL 1383I TCR transduced T cells cohort 3: 2x109 TIL 1383I TCR transduced T cells cohort 4: 5x109 TIL 1383I TCR transduced T cells Experimental Design Desire to explore each dose level Safety concerns suggested dose escalation necessary Significant accrual concerns: N=18 over 2 years. Single center ‘3+3’ Data analysis at the end to identify best dose based on immunologic parameters Immunologic parameters? Difficult to get them to “commit” to a quantitative definition. Basis was paper by Johnson, Rosenberg et al. (2009) “% Persistence of T cells” In a related trial, the binary endpoint of persistence was defined as “20% or greater TIL 1383I TCR transduced CD8+ T cells in the CD3+ T cell fraction of the subject’s PBMC 30 days post-infusion” Rosenberg results 6 responses in 20 patients “There was no correlation between the number of cells administered and the likelihood of a clinical response, with some responding patients receiving a log fewer cells than others.” Better design?: Assign patients to doses showing more promise Rosenberg shows weak association between dose of cells and response Do not want to assume monotonicity. Selection of optimal dose is not obvious Practical Goals Make it easy to implement ◦ relatively few assumptions ◦ estimation can be done using standard software ◦ flexibility to different outcomes fold change (e.g., genetic marker) % persistence (e.g., immunology) absolute count (e.g., pharmacokinetics; CTCs) Make it easy to understand ◦ clinician ‘buy-in’ ◦ statistician ‘buy-in’ Adaptive randomization approach A basic scenario: K doses of interest outcome is persistence at 2 weeks (or 30 days?) ◦ for accrual reasons: 2 weeks preferred ◦ for link to clinical outcomes: 30 days may be preferred Treat two patients at each dose, escalating from dose 1 to K Implement rules to disallow doses if not safe (e.g., 2 DLTs) Continue enrolling to a total of N patients using adaptive randomization After 2K patients, adapt randomization Estimate % persistence at each dose using data from first 2*K patients = % CD3 cells at follow−up compared to baseline Standard linear regression model*: log( ) = 1 + 2 = 2 + 3 = 3 + … + = Define = estimated persistence (%) at dose j Define = or = * log link here. others could be used. After 2K patients, adaptively randomize For the next patient, randomize to doses j = 1,…,K based on Fit model above based on updated persistence outcome. Repeat until total sample size of N achieved. Theorized benefits More patients will be allocated to doses with higher persistence Better inferences will be made regarding optimal doses Precision estimates for doses with highest persistence will be improved Dose selection for RP2D will be improved compared to balanced design Evaluating the Results Number of patients treated per dose level Estimated persistence per dose level Accuracy of dose selection: what is the best dose? Three types of criteria: ◦ dose with maximum persistence ◦ minimum dose with persistence of at least X% ◦ highest persistence prior to plateau (defined by increase of <P% between doses). Incorporating uncertainty into dose selection ◦ based on median persistence per dose? mean? ◦ select dose so that most patients will have certain level of persistence? Model Comparisons We compared our adaptive model to the equal allocation and doubly adaptive biased coin (DBCD) designs. Equal allocation (balanced) design: randomization to achieve equal sample size per dose For the DBCD, the first 2K patients were equally allocated to K doses; the assignments of the remaining patients were made using the following allocation function and target allocation proportions: y ( y / x ) L Allocation Function: g ( x , y ) , 1, L 2 y ( y / x ) L k k k K j j 1 Target Allocation Proportion: k pˆ j k pˆ k j j , k 1,..., K , K 3 Simulations Total N=25 ◦ 2 at each of five dose levels ◦ 15 allocated by adaptive randomization or balanced allocation Five true models. True Models Considered linear curvilinear flat plateau quadratic Simulation Setup Persistence can range from 0% to 100% (technically can be greater, but very unlikely) Persistence is generated from beta-binomial where between patient heterogeneity is controlled by beta distribution and within patient heterogeneity by N: ~ , ~(100, ) So far, we’ve used two sets of variance assumptions ◦ constant variance across doses vs. larger variance near persistence of 50% ◦ small vs. large variance in Beta (v=0.002, v=0.01) Reasonable assumptions, yet ◦ not completely consistent with fitted model ◦ allows robustness to misfit to be evaluated Results: Allocation to doses (large V) Results: estimated persistence (large V) What we learned (so far) we can allocate more patients to doses with higher persistence estimation at the doses with higher persistence is marginally improved ◦ less bias, greater precision ◦ depends on level of variance assumed (work in progress!) ◦ square root vs. no square root does not have much effect on results. when there is no dose response we maintain essentially the average properties as the balanced design N of 25 is not very big. ◦ We are only considering 15 patients in adaptive portion. ◦ larger sample sizes provide greater improvements compared to balanced design. Dose selection: work in progress Choosing the best dose ◦ ’eyeball test’ vs. a quantitative approach? ◦ incorporating clinical outcomes into dose selection current approach (so far) addresses dose assignment dose selection may incorporate both persistence and clinical outcomes (and the association between persistence and clinical outcome) ◦ defining a plateau is application specific SAFETY CONSTRAINTS ◦ doses may become ‘disqualified’ if there are adverse events at those dose levels ◦ The main cause is represented by the nonimmune systemic toxicities and autoimmunity triggerred by IL2 ◦ relatively easy insertion: will likely have similar effects on balanced and adaptive approaches. Lots more to consider many more scenarios! lag time: ◦ 14 days (or 30 days) to measure persistence in this situation. ◦ if there is rapid accrual, randomization probability will not be updated as frequently and design will lean more towards balanced. transformations: ◦ choice of link function for deriving randomization probabilities will be context specific ◦ dose selection will have a similar issue ◦ Should we consider using ranks? other outcomes drop-outs/inevaluables ◦ there is the reality of patients who drop out or whose follow-up measures are inevaluable accounting for uncertainty in the model: ◦ quite a few ways to go. ◦ shall we be Bayesian? Acknowledgements Cody Chiuzan, MS (PhD candidate) Mike Nishimura, PhD Supported by: ◦ NIH/NCI P01 CA54778-01 ◦ NIH/NCI P30 CA138313-02 References: ◦ Johnson L. A., Morgan R. A., Dudley M. E., Rosenberg S. A., Gene Therapy with human and mouse Tcell receptors mediates cancer regression and targets normal tissues expressing cognate antigen. Blood, 2009, Vol. 114, No. 3. ◦ Duval L., Schmidt K., Fode K., Jensen J., Nishimura M., Adoptive Transfer of Allogeneic Cytotoxic T Lymphocytes Equipped with a HLA-A2 Restricted MART-1 T-Cell Receptor: A Phase I Trial in Metastatic Melanoma. Clinical Cancer Research 12: 1229-1236. Feifang Hu, Li-Xin-Zhang, Asymptotic Properties of Doubly Adaptive Biased Coin Designs for Multitreatment Clinical Trials. The Annals of Statistics, 2004, Vol. 32, No. 1, 268-301. Lanju Zhang, Rosenberger W. F., Response-Adaptive Randomization for Clinical Trials with Continuous Outcomes. Biometrics 62, 2006, 562-569. Liangliang Duan, Feifang Hu, Doubly Adaptive Biased Coin Designs with Heterogeneous Responses. Journal of Statistical Planning and Inference 139, 2009, 3220-3230. ◦ ◦ ◦ Contact info Elizabeth Garrett-Mayer firstname.lastname@example.org Cody Chiuzan email@example.com Brief description of approach T Cell Receptor (TCR) Modified T Cells “Genetically engineered lymphocytes” This approach involves ◦ identifying and cloning the TCR genes from tumor reactive T cell clones (human or mouse). ◦ constructing retroviral vectors capable of introducing these genes into normal cells ◦ genetically modifying the patients PBL-derived T cells or hematopoietic stem cells ex vivo. genes encoding TCRs are engineered into retroviral vectors these are then used to transduce autologous peripheral lymphocytes ◦ these gene modified autologous cells are then returned to the patient. Brief description of rationale Redirect the specificity of normal T cell to recognize a variety of antigens There are several advantages to treating patients with cells that have been engineered to express TCR genes. ◦ The vectors represent an “off the shelf” reagent that could be used to treat any patient that expresses the antigen and MHC molecules recognized by the TCR. ◦ This approach does not rely on the patients TCR repertoire and precursor frequency. ◦ The unique sequences within theTCR enable us to monitor the persistence, localization, and frequency of these genetically engineered cells using clone specific PCR primers. This ability to monitor patients based on the presence of the transduced TCR will enable us to understand more about the behavior of tumor-reactive T cells in cancer patients.