Background

Systemic lupus erythematosus (SLE) exhibits complex temporal interdependencies among autoantibody species (anti-dsDNA, anti-Sm, anti-RNP, anti-Ro, anti-La, anti-C1q, antiphospholipid antibodies). Current clinical practice treats these as independent biomarkers, ignoring directional causal relationships in their temporal evolution. Treatment-refractory SLE remains identifiable only after multiple therapeutic failures, incurring cumulative organ damage.

Hypothesis

We hypothesize that multivariate Granger causality analysis applied to serial autoantibody panels (≥6 timepoints over 12–24 months) will reveal directed acyclic influence networks where specific causal cascade motifs—particularly feed-forward loops from anti-C1q → anti-dsDNA → complement consumption, and reciprocal amplification between anti-Sm and anti-RNP—predict treatment-refractory phenotypes 3–6 months before clinical failure of standard-of-care immunosuppression.

Specifically:

1. Granger-causal density (proportion of significant pairwise causal links at FDR < 0.05) exceeding 0.4 in the autoantibody network predicts multi-drug refractoriness with AUC > 0.80
2. Causal motif enrichment for feed-forward loops (vs. feedback or isolated nodes) distinguishes refractory from responsive patients with sensitivity > 75%
3. The temporal lag structure (optimal VAR model order) correlates with time-to-refractoriness, enabling individualized therapeutic window estimation

Methodology

• Vector autoregressive (VAR) models of order p (selected via BIC) on log-transformed, z-scored autoantibody titers
• Granger causality F-tests with Bonferroni-Holm correction across all pairwise directions
• Network motif census using mfinder algorithm on the resulting directed graph
• Penalized VAR (LASSO-VAR) for high-dimensional settings with >10 autoantibody species
• Validation via time-series cross-validation (expanding window) to prevent information leakage
• Comparison against static correlation networks and univariate biomarker thresholds

Testable Predictions

1. Granger-causal network density at month 6 predicts refractoriness at months 9–12 (AUC > 0.80)
2. Feed-forward loop motifs are ≥2× enriched in refractory vs. responsive patient networks (permutation test p < 0.01)
3. Optimal VAR lag order is significantly higher in refractory patients (reflecting slower, more entangled immune dynamics)
4. LASSO-VAR identifies a sparse set of ≤5 key causal edges sufficient for >70% classification accuracy

Limitations

• Requires frequent serial sampling (≥bimonthly), which is uncommon in routine clinical care
• Granger causality captures predictive but not necessarily mechanistic causation
• VAR models assume stationarity; non-stationary autoantibody dynamics may require VECM or regime-switching extensions
• Confounding by concurrent treatment changes is difficult to fully address without randomization
• Sample size requirements for stable VAR estimation grow quadratically with the number of autoantibody species
• External validation across ethnically diverse cohorts is essential given known HLA-autoantibody associations

Clinical Significance

Early identification of treatment-refractory SLE phenotypes would enable proactive escalation to belimumab, voclosporin, or anifrolumab before cumulative organ damage accrues. The directed network approach provides mechanistic insight into which autoantibody cascades drive refractoriness, potentially identifying patient-specific therapeutic targets. Integration with pharmacogenomic data (e.g., Fcγ receptor polymorphisms affecting autoantibody clearance) could further refine predictions.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology