Hypothesis

We propose that equivariant graph neural networks (E(3)-GNNs) operating on autoantigen tertiary structures — specifically citrullinated vimentin, α-enolase, and fibrinogen — combined with learned HLA-DRB1 shared epitope allele embeddings, can predict which specific citrullinated peptide epitopes will elicit autoantibody responses in genetically susceptible individuals 12–36 months before ACPA seroconversion.

Background and Rationale

Current preclinical RA prediction relies on aggregate ACPA positivity, ignoring the sequential, epitope-specific nature of autoantibody spreading. The "epitope spreading" phenomenon — where immune responses progressively target additional citrullinated peptides — follows non-random patterns likely governed by (1) structural accessibility of citrullinated residues on native protein surfaces, (2) HLA-peptide binding affinity determined by shared epitope geometry, and (3) molecular mimicry with mucosal pathogen-derived peptides.

E(3)-equivariant GNNs preserve rotational and translational symmetries of protein structures, enabling physically meaningful representation learning over atomic coordinates. By encoding citrullination sites as node-level perturbations on AlphaFold2-predicted autoantigen structures and conditioning on HLA-DRB1 allele-specific binding groove geometries, the model can learn structure-function relationships governing immunodominant epitope selection.

Testable Predictions

1. Primary: E(3)-GNN epitope vulnerability scores will predict the temporal order of ACPA fine-specificity seroconversion (anti-CCP, anti-CEP-1, anti-cFib, anti-cVim) with Kendall τ > 0.65 in prospective at-risk cohorts (first-degree relatives of RA patients).
2. Secondary: HLA-DRB1*04:01 vs *04:04 vs *01:01 carriers will exhibit statistically distinct predicted epitope hierarchies (permutation test p < 0.01), reflecting allele-specific binding groove geometry.
3. Tertiary: Integration of mucosal microbiome citrullinome data (P. gingivalis PPAD targets) will improve prediction AUC by ≥0.08 over structure-only models, supporting the molecular mimicry pathway.

Proposed Validation

• Discovery cohort: Leiden Early Arthritis Clinic or similar biobank with serial pre-RA sera and HLA genotyping (n ≥ 300 at-risk individuals, ≥50 converters)
• Architecture: SE(3)-Transformer backbone with cross-attention between autoantigen structural embeddings and HLA binding groove representations
• Calibration: Platt-scaled probabilities with Brier score < 0.15; conformal prediction intervals for uncertainty quantification
• External validation: Independent cohort (e.g., Umeå, North American SERA) with Harrell C-index > 0.70 for time-to-seroconversion

Limitations

• AlphaFold2 structures may not capture post-translational citrullination-induced conformational changes; molecular dynamics refinement may be necessary
• Training requires matched longitudinal sera with fine-specificity ACPA profiling — such datasets are rare
• HLA-peptide binding predictions carry inherent uncertainty; NetMHCIIpan benchmarks show AUC ~0.85 but performance on citrullinated peptides specifically is less validated
• Model assumes epitope spreading follows structure-driven rules, but stochastic immunological events (infections, mucosal barrier disruption) introduce irreducible noise

Clinical Significance

Predicting which epitopes will be targeted — and when — transforms preclinical RA interception from binary risk stratification to personalized immune trajectory forecasting. This enables epitope-specific tolerogenic interventions (e.g., citrullinated peptide-loaded tolerogenic dendritic cells) timed to the predicted spreading window, potentially preventing RA onset entirely in genetically susceptible individuals.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology