Dose optimization in multi-indication oncology basket trials: Leveraging Bayesian borrowing to identify OBD

Introduction: Project Optimus encourages early identification of the Optimal Biological Dose (OBD) by balancing efficacy and safety, moving beyond the maximum tolerated dose paradigm. For programs spanning multiple indications, a key challenge is that the OBD may differ by indication. We propose a basket trial strategy that formally borrows information across indications to improve dose selection efficiency, aiming to reduce per indication sample size and time to decision while maintaining robust inference when OBDs are heterogeneous. Methods: We assume a basket trial design to address dose optimization up front within a single trial rather than multiple separate cohorts. We apply a utility based, three outcome decision rule (recommend low dose, consider, or high dose) integrating efficacy-toxicity tradeoffs. We compare three strategies: fully exchangeable Bayesian hierarchical modeling (BHM-EX) that borrows across all indications, latent cluster hierarchical modeling (BHM-LC) that borrows within data driven clusters, and no borrowing (NB) analyzing indications independently. We evaluate operating characteristics and sample size efficiency via simulations across varying sample sizes and true OBD heterogeneity. Results: When indications share the same true OBD, borrowing improves correct dose selection and enables fewer patients per indication to reach decisions; BHM-EX performs best, particularly under uneven enrollment. NB shows higher misselection risk with small samples. When true OBDs differ, BHM-EX underperforms due to full pooling, whereas BHM-LC is more robust, maintaining improved accuracy without overborrowing. Conclusions: Bayesian borrowing in basket trials can reduce sample size while maintaining the same probability of correct OBD selection and time to decision, especially when indications share OBDs or have imbalanced enrollment. Clinical, preclinical, and clinical pharmacology insights should guide assumptions about similarity across indications and the choice of borrowing strategy, with prespecified sensitivity analyses to support regulatory confidence.