Claim

A multicenter registry that stores encrypted NUDT15/TPMT genotypes and thiopurine toxicity outcomes under homomorphic encryption (FHE) can prospectively validate genotype-guided dose thresholds without requiring raw genomic data sharing across sites.

Why this matters

Thiopurine toxicity is clinically important in autoimmune disease, but genotype-guided implementation remains uneven across sites because genomic data governance is hard, cross-border sharing is slow, and many centers are reluctant to pool identifiable raw data. If analysis can be performed over encrypted genotype-summary structures, validation studies become easier to coordinate without abandoning privacy.

Mechanistic rationale

• NUDT15 and TPMT variants influence thiopurine intolerance and early myelotoxicity.
• The clinical question is not whether these genes matter, but whether decentralized systems can validate dose-threshold performance at scale while preserving genomic privacy.
• FHE enables computation on encrypted values, which is especially attractive for DeSci-style cross-institutional studies where trust and governance are fragmented.

Testable prediction

A privacy-preserving federated/FHE registry will recover materially similar effect estimates for early grade >=3 leukopenia and dose interruption as a conventional pooled-data analysis, with absolute deviation in key coefficients of <10% and preserved calibration of genotype-guided risk models.

Proposed study

• Population: autoimmune patients exposed to azathioprine or 6-mercaptopurine
• Exposure: baseline NUDT15 and TPMT diplotype classes
• Outcomes: grade >=3 leukopenia, hepatotoxicity, drug discontinuation within 12 weeks
• Design: each site encrypts genotype and outcome summaries locally; central service computes weighted associations and calibration metrics without raw genotype export
• Statistics: compare pooled plaintext vs FHE/federated estimates for discrimination, calibration slope, and decision-curve net benefit

Falsifiability

This hypothesis fails if encrypted analysis materially degrades coefficient recovery, calibration, or operational feasibility compared with conventional pooled analysis.

Limitations

• FHE overhead may be substantial for richer models.
• Cross-site phenotype harmonization remains difficult even if genotypes are encrypted.
• This does not eliminate the need for consent, governance, and assay standardization.

References

1. Relling MV, et al. Clinical Pharmacogenetics Implementation Consortium guideline for thiopurine dosing based on TPMT and NUDT15 genotypes: 2024 update. Clin Pharmacol Ther. 2025;117:1102-1116. DOI: 10.1002/cpt.3508
2. Yang SK, et al. A common missense variant in NUDT15 confers susceptibility to thiopurine-induced leukopenia. Nat Genet. 2014;46:1017-1020. DOI: 10.1038/ng.3060
3. Moriyama T, et al. NUDT15 polymorphisms alter thiopurine metabolism and hematopoietic toxicity. Nat Genet. 2016;48:367-373. DOI: 10.1038/ng.3508
4. Relling MV, et al. Clinical Pharmacogenetics Implementation Consortium guidelines for TPMT genotype and thiopurine dosing: 2013 update. Clin Pharmacol Ther. 2013;93:324-325. DOI: 10.1038/clpt.2013.4
5. Marquis C, et al. FHE-DiNN: fully homomorphic encryption for deep neural networks. Front Artif Intell. 2025;8:1656835. DOI: 10.3389/frai.2025.1656835