Background

Azathioprine remains a cornerstone of SLE maintenance therapy, yet hepatotoxicity occurs in 5–15% of patients and is often indistinguishable from disease-related transaminase elevation until clinically significant. Current pharmacogenomic approaches use TPMT phenotyping alone, missing the substantial contribution of NAT2 acetylator status and the competing risk structure inherent in treatment outcomes.

Hypothesis

Bayesian Model Averaging (BMA) across an ensemble of competing risk models — incorporating Fine-Gray subdistribution hazards, cause-specific Cox models, and multi-state Markov transition models — with patient-specific NAT2 rapid/slow acetylator and TPMT activity covariates will achieve superior discriminative performance (ΔC-statistic >0.08) over any single-framework approach in simultaneously predicting hepatotoxicity onset and therapeutic response (defined as SLEDAI reduction ≥4 points) within 24 weeks of azathioprine initiation.

Mechanistic Rationale

Azathioprine undergoes complex metabolism: xanthine oxidase, TPMT, and HGPRT pathways generate competing metabolite pools. NAT2 acetylator status modulates downstream clearance of reactive intermediates. The competing risk structure is genuine — hepatotoxicity and therapeutic response share metabolic precursors but diverge at branch points determined by enzyme kinetics. No single survival model correctly captures both the subdistribution (population-level prognostic) and cause-specific (mechanistic) perspectives simultaneously. BMA assigns posterior model probabilities via marginal likelihood integration, weighting each framework proportional to its fit, thereby propagating model uncertainty into the final prediction interval.

Testable Predictions

1. BMA posterior model weights will assign >0.3 probability to the multi-state Markov framework in NAT2 slow acetylators (reflecting the importance of state transitions through intermediate hepatic stress states) versus >0.3 to Fine-Gray in rapid acetylators (where competing events dominate prognosis)
2. Calibration assessed by integrated calibration index (ICI) will be <0.03 for the BMA ensemble versus >0.06 for any individual model across both pharmacogenomic strata
3. Net reclassification improvement (NRI) for clinically actionable risk categories (low/intermediate/high hepatotoxicity risk) will exceed 0.15 when adding NAT2 to TPMT-only models within the BMA framework
4. Posterior predictive intervals from BMA will achieve 90% empirical coverage of observed event times, whereas single-model 90% CIs will show undercoverage (<80%) in at least one pharmacogenomic stratum

Proposed Validation

Retrospective cohort: ≥500 SLE patients initiating azathioprine with available TPMT phenotype, NAT2 genotype (rs1801280, rs1799930, rs1208), serial hepatic panels, and SLEDAI scores at 0/4/8/12/24 weeks. BMA implementation via BayesSurv or rstanarm with bridge sampling for marginal likelihoods. Internal validation via 10-fold cross-validation with blocked randomization preserving pharmacogenomic strata proportions. External validation in an independent lupus cohort.

Limitations

• NAT2 genotyping is not standard of care; implementation requires prospective genotyping infrastructure
• BMA computational cost scales with model ensemble size; MCMC convergence for multi-state models may require >50,000 post-warmup iterations
• Competing risk structure assumes non-informative censoring for events other than hepatotoxicity and response — intercurrent illness or treatment discontinuation for non-hepatic reasons may violate this
• Retrospective cohort design cannot establish causality; confounding by indication (sicker patients receiving azathioprine) must be addressed via propensity score weighting

Clinical Significance

If validated, BMA with dual pharmacogenomic stratification would provide the first uncertainty-quantified, model-robust prediction tool for azathioprine outcomes in SLE. The framework is generalizable to any setting where pharmacogenomic variation creates branching metabolic pathways with competing clinical endpoints — applicable to methotrexate, mycophenolate, and cyclophosphamide decisions across rheumatology.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology