Hypothesis

Bayesian response-adaptive randomization (RAR) that dynamically re-weights treatment allocation based on real-time pharmacogenomic strata — specifically CYP2C19 metabolizer status for leflunomide and NAT2 acetylator phenotype for sulfasalazine — will achieve equivalent statistical power to conventional 1:1 randomization while requiring 35–50% fewer participants in rheumatoid arthritis DMARD comparison trials.

Background

Current rheumatology RCTs treat pharmacogenomic heterogeneity as noise. CYP2C19 poor metabolizers (*2/*2, *2/*3) produce 60–80% less teriflunomide (active metabolite of leflunomide), while NAT2 slow acetylators exhibit 2–3× higher sulfasalazine plasma concentrations. These known sources of variance inflate required sample sizes when ignored.

Bayesian RAR allows allocation probabilities to shift toward arms showing greater benefit within each stratum, concentrating enrollment where equipoise genuinely exists. The key innovation is coupling RAR with pre-randomization genotyping to define strata, creating a pharmacogenomically-informed adaptive design.

Testable Predictions

1. Primary: A simulated trial of leflunomide vs. sulfasalazine using Thompson sampling with CYP2C19/NAT2 strata will reach 95% posterior probability of superiority with 35–50% fewer patients than a fixed 1:1 design (validated via Monte Carlo simulation, ≥10,000 iterations).
2. Secondary: The adaptive design will assign ≥70% of NAT2 slow acetylators away from high-dose sulfasalazine arms, reducing grade ≥3 adverse events by ≥40% compared to fixed allocation.
3. Falsification: If genotype-stratified RAR fails to reduce sample size by >20% across ≥80% of simulation scenarios, the hypothesis is rejected.

Proposed Methodology

• Design: Two-arm Bayesian RAR with burn-in period (first 30% conventional 1:1)
• Strata: CYP2C19 (extensive/intermediate/poor) × NAT2 (rapid/slow)
• Prior: Weakly informative Beta(1,1) per stratum-arm, updated every 20 patients
• Decision rule: Stop for superiority when P(δ > 0 | data) > 0.975; stop for futility when P(δ > 0 | data) < 0.05
• Calibration: Type I error controlled at 0.025 (one-sided) via simulation-based calibration of decision boundaries
• Software: R package RBesT for Bayesian historical borrowing + custom RAR implementation

Limitations

• Genotyping turnaround time (48–72h) may create enrollment bottlenecks in real-world implementation
• CYP2C19/NAT2 explain only a fraction of PK variability; unmeasured covariates may attenuate efficiency gains
• RAR designs are susceptible to time trends (drift bias) — calendar-time adjustment is essential
• Regulatory acceptance of RAR in confirmatory trials remains inconsistent across FDA/EMA/COFEPRIS
• Simulation results may overestimate real-world gains if treatment effects are more heterogeneous than modeled

Clinical Significance

Rheumatology trials are chronically underpowered due to disease heterogeneity and slow enrollment. A validated pharmacogenomically-stratified RAR framework could: (1) make smaller, faster trials feasible for DMARDs with known pharmacogenomic modulators, (2) reduce exposure of genetically-predicted poor responders to ineffective treatments, and (3) provide a template for precision-medicine trial design in other autoimmune diseases. This directly addresses the ethical imperative to minimize unnecessary participant exposure while maintaining scientific rigor.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology