An extended Bayesian semi-mechanistic dose-finding design for phase I oncology trials using pharmacokinetic and pharmacodynamic information

We propose a model-based, semi-mechanistic dose-finding (SDF) design for phase I oncology trials that incorporates pharmacokinetic/pharmacodynamic (PK/PD) information when modeling the dose-toxicity relationship. This design is motivated by a phase Ib/II clinical trial of anti-CD20/CD3 T cell therapy in non-Hodgkin lymphoma patients; it extends a recently proposed SDF model framework by incorporating measurements of a PD biomarker relevant to the primary dose-limiting toxicity (DLT). We propose joint Bayesian modeling of the PK, PD, and DLT outcomes. Our extensive simulation studies show that on average the proposed design outperforms some common phase I trial designs, including modified toxicity probability interval (mTPI) and Bayesian optimal interval (BOIN) designs, the continual reassessment method (CRM), as well as an SDF design assuming a latent PD biomarker (SDF-woPD), in terms of the percentage of correct selection of maximum tolerated dose (MTD) and average number of patients allocated to MTD, under a variety of dose-toxicity scenarios. When the working PK model and the class of link function between the cumulative PD effect and DLT probability is correctly specified, the proposed design also yields better estimated dose-toxicity curves than CRM and SDF-woPD. Our sensitivity analyses suggest that the design's performance is reasonably robust to prior specification for the parameter in the link function, as well as misspecification of the PK model and class of the link function.

Automated summary

We propose a model-based, semi-mechanistic dose-finding (SDF) design that incorporates pharmacokinetic/pharmacodynamic (PK/PD) information when modeling the dose-toxicity relationship. This design outperforms common phase I trial designs in terms of selecting the maximum tolerated dose (MTD) and allocating patients to MTD. By jointly modeling PK, PD, and dose-limiting toxicity (DLT) outcomes using Bayesian methods, the proposed design yields better estimated dose-toxicity curves and is robust to certain model specifications.