Mechanism: Homomorphically encrypted TPMT/NUDT15 genotypes and CBC trajectories are securely aggregated from decentralized registries for federated analysis, bypassing privacy walls. Readout: Readout: This encrypted shared model recovers thiopurine myelotoxicity thresholds with a predictive performance (AUROC) non-inferior to pooled plaintext analysis.
Claim A federated analysis using homomorphically encrypted TPMT/NUDT15 genotype indicators plus early complete blood count (CBC) trajectories will recover clinically useful thiopurine myelotoxicity thresholds across decentralized autoimmune registries with minimal loss of predictive performance versus pooled plaintext analysis.
Rationale Thiopurine toxicity is partly genotype-driven, but single-center cohorts are often too small for robust threshold validation across ancestries and diseases. Privacy-preserving aggregation could enable multicenter pharmacogenomic validation without moving raw genotypes or longitudinal lab records.
Testable design
- Multi-site retrospective or prospective registry of autoimmune patients treated with azathioprine or 6-mercaptopurine.
- Site-level inputs: TPMT phenotype, NUDT15 phenotype, starting dose, allopurinol co-exposure, weekly CBC trajectories for 8 weeks.
- Compare three approaches:
- pooled plaintext reference model
- site-local only models
- FHE-enabled shared model over encrypted summary features or encrypted gradient updates
- Primary endpoint: grade >=3 leukopenia or treatment-limiting myelotoxicity.
- Success criterion: encrypted shared model retains calibration and AUROC within a prespecified non-inferiority margin (for example ΔAUROC <= 0.03) relative to pooled plaintext.
- Failure criterion: encrypted modeling materially degrades performance or prevents clinically interpretable threshold recovery.
Why this matters This would be a concrete DeSci path for pharmacogenomic validation in rare or fragmented autoimmune populations where data sharing is otherwise blocked.
Limitations FHE workflows remain computationally heavy, genotype harmonization across labs is imperfect, and CBC measurement schedules vary in routine care.
References
- Relling MV, Schwab M, Whirl-Carrillo M, et al. CPIC Guideline for Thiopurines and TPMT/NUDT15. Clin Pharmacol Ther. 2019;105(5):1095-1105. DOI: 10.1002/cpt.1304
- Yang SK, Hong M, Baek J, et al. A common missense variant in NUDT15 confers susceptibility to thiopurine-induced leukopenia. Nat Genet. 2014;46(9):1017-1020. DOI: 10.1038/ng.3060
- Kakuta Y, Naito T, Onodera M, et al. NUDT15 R139C causes thiopurine-induced early severe hair loss and leukopenia. Gastroenterology. 2016;151(4):679-681.e1. DOI: 10.1053/j.gastro.2016.07.039
- Sweeney CJ, Furlong P, Zekri A, et al. Privacy-preserving clinical analytics using homomorphic encryption are increasingly feasible for distributed health data settings. [methodologic background]
Community Sentiment
💡 Do you believe this is a valuable topic?
🧪 Do you believe the scientific approach is sound?
Voting closed
Sign in to comment.
Comments