Background

Sulfasalazine (SSZ) remains a cornerstone csDMARD for peripheral arthritis in ankylosing spondylitis (AS), yet 15–25% of patients experience serious adverse drug reactions (SADRs) — agranulocytosis, hepatotoxicity, or severe hypersensitivity — often within 8–16 weeks of initiation. Current pharmacogenomic screening (HLA-B*13:01 for DRESS, NAT2 acetylator status for hepatotoxicity) captures only a fraction of at-risk patients, and standard predictive models require large labeled cohorts unavailable for rare ADR phenotypes.

Hypothesis

We hypothesize that a rheumatology foundation model pre-trained on longitudinal clinical sequences (labs, disease activity scores, medications, comorbidities) and subsequently fine-tuned using pharmacogenomic haplotype embeddings — encoding CYP2C9, NAT2, HLA-B, HLA-A, and ABCG2 allele combinations as dense vector representations — can predict SSZ-associated SADRs in AS patients with >90% negative predictive value (NPV) using as few as 50–100 labeled ADR cases via few-shot patient similarity retrieval in the model latent space.

Mechanism

The approach operates in three stages:

1. Pre-training phase: A GPT-2-style decoder model is trained on OMOP-standardized clinical sequences from >100,000 rheumatology encounters using next-token prediction. This captures temporal patterns of disease activity, treatment response, and laboratory trajectories without pharmacogenomic labels.

2. Pharmacogenomic embedding injection: CYP450/HLA/transporter haplotype combinations are encoded via a cross-attention module that projects polygenic pharmacogenomic profiles into the same latent space as clinical sequences. The haplotype encoder uses learned embeddings for each allele with multiplicative interaction terms for known epistatic pairs (e.g., NAT2 slow + CYP2C9*3 compound risk).

3. Few-shot retrieval: Given a new AS patient initiating SSZ, the model generates a latent representation combining their clinical trajectory and pharmacogenomic profile. K-nearest neighbor retrieval in this joint space identifies the most similar patients from a curated ADR registry. A calibrated Bayesian classifier on the retrieved neighbor outcomes produces a posterior SADR probability with uncertainty quantification via conformal prediction intervals.

Testable Predictions

1. The foundation model with pharmacogenomic embeddings achieves NPV >90% and AUROC >0.85 for SSZ SADR prediction using only 50–100 labeled cases, compared to AUROC <0.70 for logistic regression on pharmacogenomic features alone.
2. Few-shot retrieval in the joint clinical-pharmacogenomic latent space outperforms retrieval in either space independently (ablation Δ AUROC >0.08).
3. The model identifies novel pharmacogenomic interaction patterns (beyond HLA-B*13:01 and NAT2) that contribute >15% of predictive signal, discoverable via SHAP analysis on the cross-attention weights.
4. Conformal prediction intervals achieve nominal 95% coverage on held-out external validation cohorts without recalibration.

Limitations

• Foundation model pre-training requires access to large-scale OMOP-formatted rheumatology data, which may carry institution-specific biases in coding practices and laboratory assay calibration.
• Pharmacogenomic haplotype data availability remains limited in non-European populations, potentially reducing NPV in underrepresented ancestries where allele frequencies differ substantially.
• Few-shot retrieval performance degrades when the ADR registry is phenotypically narrow (e.g., only agranulocytosis cases without hepatotoxicity representation).
• SSZ SADRs are heterogeneous mechanisms — immune-mediated hypersensitivity vs. dose-dependent hepatotoxicity may require separate sub-models rather than a unified predictor.
• Retrospective validation cannot fully account for informative censoring (patients discontinued SSZ early due to minor side effects, masking potential SADR progression).

Clinical Significance

SSZ is frequently prescribed in resource-limited settings where biologic access is restricted, making SADR prevention critical. A high-NPV screening tool would enable confident SSZ initiation in >75% of patients while flagging the remaining subset for enhanced monitoring or alternative therapy. Integration into DeSci infrastructure via federated learning across multiple rheumatology registries — with privacy-preserving pharmacogenomic embeddings transmitted rather than raw genotypes — could enable continuous model improvement without centralized data aggregation. This represents a concrete pathway toward precision csDMARD prescribing in spondyloarthritis, an area largely neglected by current pharmacogenomic implementation efforts.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology