PII-035 - MACHINE LEARNING-BASED DOSE OPTIMIZATION DESIGN INCORPORATING EXPOSURE-RESPONSE RELATIONSHIP IN CELL THERAPY.

S. Wang¹, Y. Zhao¹, J. Li², R. Liu², G. Mugundu²; ¹Takeda Development Center Americas Inc., Lexington, MA, USA, ²Takeda Development Center Americas, Inc., Lexington, MA, USA.

Background: In cell therapy product development, cell expansion has been highly correlated with response/safety. Patients’ heterogeneity and product characteristics contribute to variabilities in cell expansion, persistence & response. Incorporating exposure assessments enables us to use full spectrum information for dose selection decision-making, which is also aligned with FDA’s Project OPTIMUS. Therefore, we propose a seamless phase I/II design integrating data from toxicity, efficacy, cellular kinetics (CK), and baseline patient/product characteristics for optimal dose selection.

Methods: Eight scenarios were generated to cover potential exposure-response relationships and test the robustness of the proposed design. 1000 simulations were conducted for each scenario to evaluate the performance of the proposed design. Random Forest (RF) with the totality of data was used to guide dose escalation. Then we narrowed down dose levels (DLs) and randomly assigned patients to these DLs for further exploration. During the process, RF-based interim analysis was conducted to discontinue toxic or futile DLs. Finally, the optimal dose (OD) that met predefined efficacious and safety criteria was selected by the algorithm.

Results: In most scenarios, percentages of correctly expanding OD and selecting OD are both more than 65%. In all scenarios, less than 20% of patients were allocated to toxic DLs, and the percentage of not terminating sub-optimal doses until final analysis was less than 20%. 0-3 DLs were selected for further exploration by the RF-based algorithm.

Conclusion: The proposed RF-based seamless phase I/II design incorporating exposure data is entirely data-driven for dose determination, has high accuracy in selecting the OD, and can allocate patients more on potential therapeutic effective DLs and less on toxic DLs.