Background

Conventional rheumatology response metrics (DAS28, CDAI, Boolean remission) rely on composite indices that aggregate heterogeneous clinical signals into scalar summaries. This compression discards distributional information about underlying immune cell populations that may shift earlier and more informatively than joint counts or acute-phase reactants. Optimal transport (OT) theory provides a mathematically principled framework for comparing probability distributions over metric spaces, offering a natural language for quantifying immune remodeling as measured by multiparameter flow cytometry.

Hypothesis

We hypothesize that the 2-Wasserstein distance (W₂) computed between serial flow cytometry point clouds (≥12-color panels capturing T-cell, B-cell, monocyte, and NK-cell subsets in high-dimensional phenotypic space) at baseline versus 4-, 8-, and 12-week follow-up will detect treatment-induced immune remodeling in rheumatoid arthritis patients initiating biologic DMARDs 4–8 weeks before DAS28-CRP or CDAI reflect clinical response.

Specifically:

1. Early detection: The W₂ trajectory slope over weeks 0–8 will discriminate eventual EULAR good responders from non-responders with AUROC ≥0.80, at a timepoint when DAS28-CRP change alone achieves AUROC ≤0.60.
2. Distributional specificity: Sinkhorn-regularized OT barycenters computed across responder trajectories will converge to a characteristic "response attractor" in Wasserstein space, identifiable by week 8.
3. Pharmacogenomic modulation: Stratification by CYP3A4/CYP2C19 metabolizer status will reveal distinct OT trajectory clusters, with rapid metabolizers showing faster W₂ displacement but not necessarily better clinical outcomes, suggesting that immune remodeling velocity and clinical benefit are partially decoupled.

Methodology

• Study design: Prospective observational cohort, N≥120, RA patients initiating first-line biologic (anti-TNF or IL-6R inhibitor)
• Flow cytometry: ≥12-color panels at weeks 0, 4, 8, 12, 24; ≥50,000 events/sample after quality gating
• OT computation: Entropic regularization (ε=0.01) via Sinkhorn algorithm on log-transformed, arcsinh-cofactor-normalized marker intensities; ground metric = squared Euclidean in phenotypic space
• Statistical framework: Bayesian longitudinal mixed-effects models with W₂ as time-varying covariate predicting DAS28-CRP trajectory; posterior inference via Hamiltonian Monte Carlo (4 chains, 2000 warmup, 2000 sampling)
• Validation: 5-fold cross-validation with Bonferroni-corrected multiplicity adjustment for the 3 primary hypotheses (α=0.0167)

Testable Predictions

• W₂(baseline, week 4) > median in eventual responders vs non-responders (Mann-Whitney p<0.01)
• Barycentric convergence quantified by decreasing inter-patient W₂ variance in responders (Levene test, p<0.05 by week 12)
• CYP3A4 rapid metabolizers show W₂(0→4) displacement 1.5× greater than poor metabolizers (95% CrI excluding 1.0)

Limitations

• Flow cytometry panel design constrains observable phenotypic space; spectral cytometry or CyTOF would capture richer distributions but with different technical noise profiles
• Sinkhorn regularization introduces bias proportional to ε; sensitivity analysis across ε ∈ [0.001, 0.1] required
• OT distances are sensitive to batch effects — rigorous bead-based normalization and CytoNorm alignment essential
• Single-center design limits generalizability; federated OT computation across DeSci-linked sites would require privacy-preserving protocols (differential privacy on Sinkhorn iterates)
• Sample size powered for moderate effect sizes (Cohen d≥0.6); subtle distributional shifts may require N>200

Clinical Significance

If validated, OT-based immune monitoring would provide clinicians with an early, quantitative, and mechanistically interpretable signal of biologic efficacy — enabling earlier switching decisions for non-responders (reducing exposure to ineffective therapy by 4–8 weeks) and potentially reducing cumulative joint damage. Integration with DeSci infrastructure enables multi-site distributional comparison without centralizing raw patient-level flow data, as Wasserstein distances and barycenters can be computed in federated settings.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology