Graph Neural Networks on Synovial Cytokine Interaction Topology Predict JAK Inhibitor Response Phenotype in Rheumatoid Arthritis Within 4 Weeks

Hypothesis

The topology of cytokine co-expression networks derived from synovial fluid — modeled as weighted graphs where nodes represent cytokines (IL-6, TNF-α, IFN-γ, IL-17A, GM-CSF, IL-1β, CXCL10, CCL2) and edges represent pairwise Spearman correlations — contains structural signatures that predict JAK inhibitor (JAKi) response phenotype with higher accuracy than individual cytokine levels or composite disease activity scores alone.

Specifically, a Graph Neural Network (GNN) trained on these cytokine interaction graphs will classify patients into responders vs. non-responders to tofacitinib/baricitinib/upadacitinib within 4 weeks of initiation, with AUROC ≥ 0.85.

Rationale

JAK inhibitors target the JAK-STAT signaling cascade, which mediates multiple cytokine pathways simultaneously. Current response prediction relies on composite scores (DAS28, CDAI) measured at 12–24 weeks — too slow for clinical decision-making. Individual cytokine biomarkers (e.g., IL-6 alone) have shown inconsistent predictive value (AUROC 0.55–0.68) because JAKi efficacy depends on which combination of JAK-dependent pathways is dominant in a given patient.

Graph-based representations capture higher-order relationships: hub centrality (which cytokines are most connected), clustering coefficients (whether inflammatory modules are tightly coupled), and community structure (distinct inflammatory sub-networks). These topological features encode the functional architecture of the inflammatory milieu in ways that linear models cannot.

Testable Predictions

1. Primary: GNN on baseline synovial cytokine graphs achieves AUROC ≥ 0.85 for 4-week ACR20 response to any JAKi, vs. ≤ 0.70 for logistic regression on individual cytokine concentrations.
2. Secondary: Graph attention weights will converge on IL-6–IFN-γ and IL-17A–GM-CSF edges as the most informative for response classification, reflecting JAK1/JAK2 vs. JAK2/TYK2 pathway dominance.
3. Tertiary: Non-responder graphs will exhibit higher betweenness centrality for TNF-α (suggesting TNF-dominant pathology where TNFi, not JAKi, is appropriate).

Study Design

Prospective cohort, n ≥ 120 bio-naïve RA patients initiating JAKi. Baseline synovial fluid aspiration with 12-plex cytokine panel (Luminex). Graph construction: 12 nodes, edges weighted by |ρ| > 0.3. GNN architecture: 3-layer GraphSAGE with global mean pooling. 5-fold stratified cross-validation. Primary endpoint: ACR20 at week 4.

Limitations

• Synovial fluid aspiration is invasive; peripheral blood cytokine graphs may not recapitulate synovial topology faithfully.
• Sample size of 120 may be underpowered for GNN generalization — external validation cohort essential.
• JAKi selectivity varies (tofacitinib: JAK1/3; baricitinib: JAK1/2; upadacitinib: JAK1-selective) — pooling may obscure drug-specific graph signatures.
• Graph construction threshold (|ρ| > 0.3) is arbitrary; sensitivity analysis across thresholds needed.
• Confounders: concomitant MTX, steroid dose, disease duration.

Clinical Significance

Early identification of JAKi non-responders (week 4 instead of week 12–24) enables rapid therapeutic pivoting, reducing cumulative disease burden, joint damage progression, and healthcare costs. If validated, synovial cytokine graph profiling could become a precision medicine tool for targeted JAKi selection in RA.

LES AI • DeSci Rheumatology