Department of Biostatistics
UNC Gillings School of Global Public Health
& Lineberger Comprehensive Cancer Center
2026-02-05
Luyendijk M et al. JNCI 2023
Despite Precision Oncology Advances
Despite more drugs, we haven’t fundamentally changed how trials learn.
Question: Why aren’t we seeing bigger gains?
Karagiannis D, Rampias T. Cancers 2022
Resistance Mechanisms Emerge Over Time
Traditional trials fail to capture this
Current Standard of Care
Result: Patients may remain on ineffective therapy while resistance mechanisms consolidate
What Patients Need
ADAPT Goal: Enable real-time treatment optimization based on evolving tumor biology
March 2024: ARPA-H launches ADAPT program to address these gaps
Gaps in Current Trials
| Gap | Current → Needed |
|---|---|
| Design | Fixed → Adaptive |
| Biomarkers | Baseline → Continuous |
| Infrastructure | Retrospective → Real-time |
| Translation | Siloed → Embedded |
EVOLVE-BDT: A Biomarker-Stratified Platform
Bayesian adaptive design enables early stopping + continuous learning (~700 patients)
Traditional vs Adaptive
| Aspect | Fixed | Adaptive |
|---|---|---|
| Enrollment | All pts | Stop early |
| Analysis | End only | Continuous |
| Arms | Static | Add/drop |
| Savings | None | 30-50% |
Enables: Early futility/efficacy stopping, adaptive randomization, biomarker-driven arm selection
Traditional Fixed Trial
Mrs. Johnson, 58, ER+/HER2-, ESR1 mutation
EVOLVE-BDT Adaptive Trial
Mrs. Johnson, 58, ER+/HER2-, ESR1 mutation
Bottom line: 19 fewer patients on ineffective therapy. Answers months sooner.
Once we established the scientific framework, we encountered a fundamental obstacle:
We could design the science. We could not design the trial.
Bayesian adaptive trials require 8+ interacting parameters—efficacy thresholds, futility boundaries, interim timing, sample size constraints—each affecting regulatory compliance and patient outcomes.
Manual calibration proved intractable: weeks of iteration, inconsistent results across statisticians, no guarantee of optimality.
We needed a systematic approach to navigate this complexity.
Designing adaptive trials requires achieving ALL FOUR simultaneously
Out of 1,000 possible designs, only ~70 satisfy all four requirements → AI finds the needle in the haystack
70% fewer evaluations — minutes instead of days
Manual Approach
| Aspect | Limitation |
|---|---|
| Process | Trial-and-error |
| Duration | Days to weeks |
| Scope | One at a time |
| Reproducibility | Person-dependent |
| Documentation | Often incomplete |
| Regulatory risk | Variable |
BATON Approach
| Aspect | Capability |
|---|---|
| Process | Systematic |
| Duration | Minutes to hours |
| Scope | 1000s evaluated |
| Reproducibility | Deterministic |
| Documentation | Full audit trail |
| Regulatory risk | Documented |
Designs that would have taken months to discover—or never been found at all
BATON = Bayesian Adaptive Trial OptimizatioN — uses Bayesian optimization to efficiently search complex design spaces
Bottom line: BATON finds designs that spare ~25 patients/arm from ineffective treatment
Each point = one candidate design evaluated (~5,000 simulated trials per design)
BATON revealed a hidden trade-off with real patient consequences:
Minimizing enrollment caps can prevent patients from learning sooner that their treatment works.
Caps set too low sacrifice early success declarations—patients wait longer for answers.
Why this matters: BATON surfaced this before launch, recommending a slightly higher cap that preserves early efficacy stopping while still protecting against futility.
Design calibration: minutes, not days | Documentation: complete FDA audit trail | Reproducibility: 100%
Resources
BATON R packageevolveTrial R packageContact
naim@unc.edu
EVOLVE-BDT: A True Testament to Team (Data) Science