Hypothesis

A contextual multi-armed bandit (CMAB) framework incorporating pharmacogenomic features (MTHFR C677T/A1298C polymorphisms, ABCB1 C3435T, ATIC 347C>G, and polyglutamation enzyme FPGS/GGH expression ratios) as state variables will optimize methotrexate (MTX) dose titration in rheumatoid arthritis (RA) and reduce median time-to-DAS28 target (<2.6) by ≥40% compared to standard empirical escalation protocols.

Background and Rationale

MTX remains the anchor DMARD in RA, yet dose optimization is empirical: typically 7.5–15 mg/week initial, escalated by 2.5–5 mg every 4–6 weeks based on clinical response. This slow titration wastes 3–6 months while patients accumulate joint damage. Known pharmacogenomic variants explain ~30–40% of inter-individual variability in MTX efficacy and toxicity, yet this information is rarely integrated into dosing algorithms.

Contextual bandits are ideally suited to this problem: each dosing decision is a sequential action under uncertainty, the reward signal (DAS28 change) is delayed but measurable, and the context (genomic + clinical features) is patient-specific. Unlike full reinforcement learning, bandits avoid the curse of long horizons and require fewer samples — critical in clinical settings.

Formal Framework

Let the context vector at visit t be x_t = [MTHFR genotype (0/1/2 risk alleles), ABCB1 genotype, ATIC genotype, FPGS/GGH ratio, baseline DAS28, current dose, hepatic transaminase trajectory, folate level]. Arms correspond to dose actions: {maintain, +2.5mg, +5mg, −2.5mg, switch to subcutaneous}.

We employ a Thompson Sampling policy with a Bayesian linear reward model:

r_t = x_t^T β_a + ε_t, where ε_t ~ N(0, σ²)

Posterior updates on β_a incorporate both efficacy (DAS28 reduction) and safety (hepatotoxicity, cytopenias) as a composite reward with safety penalties.

Testable Predictions

1. Primary: CMAB-guided titration achieves DAS28 <2.6 in median 8 weeks vs. 14 weeks with standard care (≥40% reduction), testable in a parallel-arm RCT with N=200.
2. Secondary: Patients with MTHFR 677TT homozygosity are routed to subcutaneous MTX or split-dosing earlier, reducing hepatotoxicity (ALT >2× ULN) by ≥50%.
3. Mechanistic: The learned β_a weight vectors will show FPGS/GGH ratio as the strongest contextual predictor of dose-response (|β| > 2× other features), reflecting polyglutamation as the rate-limiting step.
4. Regret bound: Cumulative Bayesian regret scales as O(d√T log T) where d=8 (context dimension), empirically verified against uniform random exploration.

Limitations

• Delayed reward (DAS28 measured every 4 weeks) creates temporal credit assignment challenges; we assume stationary context between visits.
• Pharmacogenomic testing is not universally available — the framework degrades gracefully to clinical-only features but with reduced optimization gain (~20% instead of 40%).
• Sample size of 200 assumes moderate effect size (Cohen d=0.6); smaller effects require adaptive enrichment designs.
• The linear reward model may miss nonlinear genotype-dose interactions — a kernelized or neural CMAB extension would address this at the cost of interpretability.
• Single-center validation limits generalizability across ethnic pharmacogenomic distributions.

Clinical Significance

If validated, this approach transforms MTX titration from empirical art to precision medicine. Reducing time-to-target by 6 weeks translates to measurable prevention of radiographic progression (estimated 0.5–1.0 Sharp/van der Heijde units saved per patient). The framework is deployable as a clinical decision support tool requiring only a one-time pharmacogenomic panel (~$200) and standard DAS28 monitoring.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology