Background

Methotrexate (MTX) remains the anchor drug in rheumatoid arthritis (RA) management, yet hepatotoxicity risk varies dramatically across patients. Current pharmacogenomic stratification relies on candidate gene approaches (MTHFR C677T, SLC19A1, ABCB1) with modest predictive power, partly because they assume discrete, pre-defined subgroups rather than allowing the data to reveal latent population structure.

Hypothesis

We hypothesize that Dirichlet Process Gaussian Mixture Models (DP-GMMs) applied to longitudinal pharmacokinetic-pharmacodynamic (PK-PD) profiles — including serial MTX-polyglutamate concentrations, ALT/AST trajectories, and red cell folate dynamics — will identify latent pharmacometabolic subgroups that:

1. Emerge without pre-specification of the number of clusters, allowing the nonparametric prior to discover the true population heterogeneity
2. Correlate post hoc with polygenic pharmacogenomic scores (PGS) constructed from MTHFR, DHFR, FPGS, GGH, TYMS, and folate transporter variants at genome-wide significance
3. Predict grade >=2 hepatotoxicity (CTCAE) within 6 months with >85% specificity and >70% sensitivity, outperforming both fixed-genotype stratification and standard logistic regression by >=15% AUROC

Mathematical Framework

Let each patient i have a multivariate time series x(t) = [MTX-PG(t), ALT(t), AST(t), RBC-folate(t)]. We model:

• G ~ DP(alpha, G0) where G0 is a Normal-Inverse-Wishart base measure
• theta_i | G ~ G (cluster assignment via stick-breaking)
• x(t) | theta_i ~ GP(mu_theta(t), K_theta(t,t')) (Gaussian process within each cluster)

The concentration parameter alpha governs cluster proliferation. We use slice sampling for posterior inference, marginalizing over the infinite-dimensional mixing measure.

Testable Predictions

1. The DP-GMM will identify 4-8 distinct pharmacometabolic clusters (95% credible interval) versus the 2-3 assumed by candidate gene approaches
2. At least one discovered cluster will have no correspondence to known MTHFR/SLC19A1 genotypes, representing a novel pharmacogenomic subgroup
3. The hepatotoxicity-predictive cluster will show a characteristic "delayed polyglutamation" PK signature detectable within the first 8 weeks of therapy
4. Bayesian model comparison (WAIC) will favor the DP-GMM over finite mixture models with K=2,...,10 components

Study Design

• Population: Prospective cohort, n>=300 MTX-naive RA patients, 12-month follow-up
• Sampling: Serial MTX-PG (RBC), hepatic panels biweekly x8 then monthly, genotyping array + imputation
• Validation: 5-fold cross-validation with held-out temporal windows; external replication in independent cohort
• Software: Stan/PyMC with custom DP-GMM module; convergence via R-hat <1.01, ESS >400

Limitations

• DP-GMMs assume exchangeability within clusters, which may not hold if disease activity modifies PK independently
• Polyglutamate measurement requires specialized LC-MS/MS not available in all centers
• The nonparametric nature means cluster number varies across posterior samples — clinical translation requires summarization (e.g., modal clustering)
• Confounders: alcohol use, concomitant hepatotoxic drugs, and NAFLD prevalence must be carefully controlled
• Sample size of 300 may be insufficient for rare subgroups (<5% prevalence)

Clinical Significance

If validated, this approach would replace the current one-size-fits-all MTX dosing paradigm with data-driven, pharmacogenomically informed subgroup-specific protocols. The "delayed polyglutamation" signature could serve as an early pharmacodynamic biomarker for hepatotoxicity risk, enabling dose adjustment or drug switching within the first 8 weeks rather than waiting for liver enzyme elevation. The Bayesian nonparametric framework is directly deployable on DeSci federated infrastructure, enabling multi-center discovery without centralizing patient-level PK data.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology