Optimizing Dosing Strategies in Cell Therapy With Machine Learning and Exposure-Response Integration

In cell therapy product development, cell expansion is highly correlated with response and safety. Significant heterogeneity in patient and product characteristics contributes to variability in cell expansion, persistence and response. Integrating exposure assessments enables us to use comprehensive information to make informed dose selection decisions, aligning with the expectations outlined in FDA's Project OPTIMUS. We propose a seamless phase I/II design that integrates data from toxicity, efficacy, cellular kinetics (CK), and baseline patient/product characteristics for optimal dose selection. Utilizing random forest (RF) with all available data, we guide dose escalation and subsequently narrow down the dose options. Additional patients are then randomly assigned to promising doses for further investigation. Throughout this process, interim analyses based on RF estimations are conducted to discontinue doses that are deemed futile or toxic. The optimal dose (OD) is ultimately chosen based on its safety profile and highest efficacy rate. The proposed RF-based seamless phase I/II design, incorporating exposure data, is entirely data-driven for dose determination. Simulation studies show that the proposed design has desirable operating characteristics, including high accuracy in selecting the optimal dose and effectively allocating patients to potentially therapeutic doses while minimizing exposure to toxic doses. On average, the RF-based algorithm selects 0-3 doses for further exploration.