Project Optimus Has a Menu Problem: Too Many Designs, Not Enough Regulatory Signal
The simulation-evidence burden has shifted to sponsors — and a crowded design landscape makes the justification conversation harder, not easier.
- Dose optimization (Project Optimus)
- Regulatory
- Methodology Frontier
Three hundred and fifteen methodology articles. That is the tally in a new systematic review in Pharmaceutical Statistics (PMID 42157027) mapping the adaptive dose-finding landscape — and it understates the current pace, because the 2025–2026 publication surge had not fully landed when the authors closed their search. The review’s headline finding is directional rather than surprising: the field has pivoted from MTD-seeking to optimal biological dose (OBD) jointly targeting toxicity and efficacy, and model-assisted interval designs (BOIN, TPI) are pulling ahead of purely model-based approaches like CRM. Project Optimus is cited explicitly as a regulatory driver. FDA finalized that guidance in August 2024, converting a pilot initiative into a standing expectation.
The methodological response has been vigorous — and slightly unwieldy. BLRM-EffTox (Journal of Biopharmaceutical Statistics) grafts efficacy-toxicity utility modeling onto the BLRM backbone for seamless Phase I operation. PRO-ADD (Statistical Methods in Medical Research) incorporates patient-reported outcomes as a formal decision input into the escalation algorithm. BPCT (Statistical Methods in Medical Research) jointly models binary toxicity and time-to-event efficacy within a platform architecture — filling the binary-endpoint gap that most Phase I/II designs leave open. Each is methodologically coherent. None has regulatory acceptance yet.
Backfill is where the friction lives
The sharpest immediate tension is in backfill strategy. Two independent simulation studies — one in Statistics in Biopharmaceutical Research and one in the Journal of Biopharmaceutical Statistics comparing backfill against randomized expansion cohorts — both address a gap the FDA guidance leaves open: what backfill rules should actually look like, and when randomization is preferable. Both studies provide simulation-based operating characteristics for RP2D selection, patient allocation, and overdose risk; sponsors writing Phase I protocols now have two independent references to cite, though neither has been adopted as a standard. Backfill rules must be pre-specified, and regulators are scrutinizing them — having simulation evidence in hand, even from non-definitive literature, is materially better than a protocol footnote.
The CAR T problem is structurally different
Autologous CAR T therapies sit outside the standard dose-optimization frame entirely. A new analysis in the Journal of Biopharmaceutical Statistics characterizes why: the administered dose is not a controlled independent variable but a patient-specific random variable driven by manufacturing variability. The paper focuses on characterizing this dose uncertainty and its downstream consequences for dose-response inference — it does not resolve the problem with a new design — but even the characterization is practically useful, particularly for estimand construction in CAR T submissions where the choice between nominal dose, infused dose, and biological potency metric is a live analytical question.
The cumulative implication is near-term and practical. Design choice is now only half the argument sponsors need to make. The other half is simulation evidence — operating characteristics under a range of plausible scenarios — and the burden of producing that evidence has shifted decisively to the sponsor. A menu of competing designs does not simplify that conversation with reviewers; it makes choosing deliberately, and documenting why, more consequential than ever.
Protocol read: Project Optimus has produced more design options than regulatory absorption capacity — choosing deliberately and documenting why now matters more than any single design’s elegance, because simulation evidence is what reviewers will actually evaluate.
What to do now:
- Pre-specify Phase I designs with simulation operating characteristics across at least three plausible toxicity-efficacy scenarios; reviewers expect the evidence, not just the algorithm name.
- For backfill rules, cite the SBR and JBS simulation studies even where they disagree — pre-specified evidence beats a protocol footnote.
- For CAR T programs, decide the dose variable (nominal / infused / biological potency) at the protocol stage and treat dose uncertainty as a first-class estimand input, not a footnote.