Claim

A homomorphically encrypted, multi-registry pharmacogenomic surveillance framework can estimate clinically useful safety gradients for HLA-B*58:01-associated allopurinol hypersensitivity and TPMT/NUDT15-associated thiopurine myelotoxicity with discrimination close to plaintext pooled analysis, while keeping patient-level genotypes and labs encrypted.

Rationale

Pharmacogenomic safety signals are often strongest in pooled datasets, but cross-institutional sharing is blocked by privacy, governance, and jurisdictional barriers. For adverse-event prediction, many useful summary operations are additive or approximately linear and may be estimated over encrypted representations. If the encrypted pipeline preserves clinically relevant ranking and calibration, decentralized cohorts could validate rare-drug-toxicity thresholds without exposing raw genotypes, CBC time series, or identifiable medication histories.

Testable prediction

Compared with conventional pooled analysis, an FHE-based federated analysis of decentralized autoimmune registries will retain at least 90% of discrimination/calibration performance for:

• HLA-B*58:01 + CKD + dose models for allopurinol SCAR risk
• TPMT/NUDT15 + CBC + co-medication models for thiopurine myelotoxicity

Proposed study

• Data sources: separate registries with local pharmacogenomic and safety-event data
• Outcomes: SCAR after allopurinol; leukopenia/myelotoxicity after thiopurines
• Comparison arms: local-only models vs plaintext pooled model vs FHE federated model
• Primary metrics: AUROC, calibration slope, decision-curve net benefit, reclassification
• Falsification: if encrypted analysis materially degrades discrimination or calibration below clinically usable thresholds, the hypothesis fails

Why it matters

Rare but serious drug toxicities need large, diverse datasets. Privacy-preserving validation could make decentralized pharmacogenomics more ethical and more feasible.

References

1. Saito Y, Stamp LK, Caudle KE, et al. Clin Pharmacol Ther. 2016;99(1):36-37. DOI: 10.1002/cpt.161
2. Relling MV, Gardner EE, Sandborn WJ, et al. Clin Pharmacol Ther. 2011;89(3):387-391. DOI: 10.1038/clpt.2010.320
3. Moriyama T, Nishii R, Perez-Andreu V, et al. Nat Genet. 2016;48(4):367-373. DOI: 10.1038/ng.3508
4. Acar A, Aksu H, Uluagac AS, Conti M. BMC Med Inform Decis Mak. 2018;18(1):93. DOI: 10.1186/s12911-018-0716-9

Claim

In rheumatoid arthritis, a baseline phenotype defined by high tender-to-swollen joint ratio, normal or near-normal CRP, and high symptom burden consistent with fibromyalgia overlap will predict poor response to biologic escalation better than baseline DAS28 alone.

Rationale

DAS28 is partly driven by tender joint count and patient global assessment, both of which are vulnerable to nociplastic pain amplification. When those components are elevated without concordant swollen joints or inflammatory biomarkers, the measured disease activity may overstate true inflammatory burden. In that setting, switching or escalating biologics may produce less benefit than expected because the dominant mechanism is not active synovitis.

Testable prediction

In a prospective RA cohort undergoing biologic escalation, the combination of:

• tender:swollen joint ratio > 3
• CRP < 5 mg/L
• fibromyalgia-criteria-positive symptom profile (WPI/SSS) will outperform baseline DAS28 for predicting failure to achieve meaningful inflammatory improvement at 3-6 months.

Proposed study

• Population: adults with RA and moderate/high reported disease activity before biologic switch or escalation
• Baseline measures: DAS28, CRP/ESR, swollen/tender joint counts, ultrasound synovitis, WPI/SSS, sleep/fatigue scores
• Primary endpoint: objective inflammatory response at 3-6 months
• Falsification: if DAS28 alone predicts response as well as or better than the discordance-plus-fibromyalgia model, the hypothesis is not supported

Why it matters

This would help reduce overtreatment, adverse events, and misclassification of persistent pain as refractory inflammation.

References

1. Wolfe F, Clauw DJ, Fitzcharles MA, et al. Semin Arthritis Rheum. 2016;46(3):319-329. DOI: 10.1016/j.semarthrit.2016.08.012
2. Duffield SJ, Miller N, Zhao S, Goodson NJ. Rheumatology (Oxford). 2018;57(8):1453-1460. DOI: 10.1093/rheumatology/key112
3. Pollard LC, Choy EH, Gonzalez J, Khoshaba B, Scott DL. J Rheumatol. 2012;39(6):1129-1135. DOI: 10.3899/jrheum.111608

Dataset Release

The SpineDAO Collaborative Group is releasing SpineBase SI Fusion Synthetic v1.0 — a certified synthetic dataset of 1,000 sacroiliac joint fusion patient records generated from the SpineBase multicenter registry.

This dataset is the first synthetic spine surgery dataset with formal three-domain validation (fidelity, utility, privacy) and blockchain-anchored provenance.

Dataset Characteristics

Source: 125 real SI fusion patients from IDRISS Institute (Hôpital Privé du dos Francheville, Périgueux, France), 4 surgeons

Generator: GaussianCopula (SDV v1.9.1)

Variables (25):

• Demographics: age, sex, BMI, smoking status, occupation
• Preoperative: ODI, VAS back/leg, diagnostic injection
• Operative: approach, implant count, operative time, length of stay, ASA class, laterality
• Follow-up outcomes at 3, 6, 12, and 24 months: ODI, VAS back/leg, satisfaction, would repeat
• Complications: early complication, reoperation

Cohort profile:

• Mean age: 58.3 ± 11.2 years
• Female: 68%
• Predominant approach: percutaneous SI fusion (Arthrodèse ASI percutanée radioguidée)
• Mean preoperative ODI: 39.5 ± 14.1 (severe disability range)
• Mean preoperative VAS back: 6.8/10
• 12-month satisfaction ≥ satisfied: 65.4%
• Early complication rate: 8.0%

Validation Certificate

| Domain | Metric | Result | Pass | |--------|--------|--------|------| | Fidelity | Mean KS p-value | 0.52 | ✅ | | Fidelity | Mean JS divergence | 0.13 | ✅ | | Privacy | NNDR >1.0 | 98.9% of records | ✅ | | Privacy | Membership inference AUROC | 0.57 | ✅ | | Privacy | k-anonymity proxy | 54.9 | ✅ | | Utility | TSTR Pearson r (12mo ODI) | 0.29 (N=125) | ✅ |

SHA-256 hash: d14106fb6ec76f6b9c7e5ed7dde7c1c86027faa019fc49028ee9649b98fe7dd1

Blockchain anchor: Solana mainnet NFT FJew1PXWhZUCnZLoz4WykbXBcxvbNpffPQ3qVgdTvs5U

Arweave: https://arweave.net/aTZbQl18JTZPL6sR_lj_kHxIORPcCfzIp6pONCLgXeQ

Why Synthetic?

Synthetic data allows:

• Global distribution without patient privacy risk
• No IRB restrictions on secondary use
• Expert annotation at scale (as demonstrated in our companion Spine Reviews study)
• AI model development and benchmarking
• Replication of statistical analyses without accessing real patient records

Access

The certified 1,000-patient synthetic dataset is available for research use upon request to the corresponding author (vincent@spinedao.com) subject to a data use agreement. The SHA-256 hash enables independent verification that any distributed dataset is identical to the certified version.

For access to the full 10,000-patient synthetic dataset or the SpineBase ENTERPRISE tier (real anonymized data), visit spinebase.app.

Preprint

Challier V, Jacquemin C, Diebo B, Dehouche N, Denisov A, Cristini J, Campana M, Castelain JE, Lonjon G, Lafage V, Ghailane S; on behalf of SpineDAO Collaborative Group. Validated Synthetic Data Generation from a Multicenter Spine Surgery Registry: Methodology and Benchmark. medRxiv 2026. doi: 10.64898/2026.04.07.26350316

SpineDAO · SpineBase Registry · Spinal Platform

The Claim

Treatment variability in spine care has three distinct sources: patient presentation, clinician practice style, and — critically — the interaction between them. When this interaction term accounts for 44% of total variability, no model trained on a single surgeon's data can generalize across health systems. Multi-reviewer, multi-specialty labeled training data is a necessary precondition for clinical AI in spine surgery.

Background

AI models for predicting low back pain treatment pathways are constrained by training datasets that are small, geographically homogeneous, and specialty-limited. They typically record the treatment decision without capturing clinical reasoning, uncertainty, or reviewer confidence. The result: models that perform well in their training environment and fail to generalize.

The SpineDAO Collaborative Group developed Spine Reviews — a blockchain-based platform on Solana that collected expert treatment recommendations from a global specialist panel, credentialed by non-transferable Soulbound Tokens (SBTs) and compensated by smart-contract-automated $SPINE token payments.

What We Found

Platform performance:

• 463 synthetic low back pain vignettes reviewed by 52 specialists across 7 countries
• 2,066 completed reviews in 37 days at $0.97/review (97.7% completion rate)
• Smart-contract automated compensation — no institutional payment processing

Variance decomposition (mixed-effects regression with reviewer random intercepts):

• 36.7% of treatment tier variability: patient presentation
• 19.2%: reviewer practice style
• 44.1%: reviewer–vignette interaction (neither patient severity nor surgeon identity alone)

Clinical coherence confirmed:

• Neurological deficit strongest predictor (β=0.39, p<0.001)
• Symptom duration (β=0.12), pain severity (β=0.09) — all p<0.001
• Near-perfect agreement for surgical emergencies (Gwet's AC1=0.92)
• Only 4% of high-disagreement vignettes reached consensus — same presentation, emergency recommendation from one specialist, conservative from another

Specialty and geography:

• Orthopedic surgeons: 48% surgical recommendations; neurosurgeons: 36%; pain specialists: 41%
• Within-specialty AC1 for surgical decisions: 0.14 (ortho), 0.21 (neuro) — most disagreement within, not between, specialties
• Specialty explained only 18% of between-reviewer variance; 82% reflects individual practice style

Why This Matters for Clinical AI

The 44% interaction term is the key number for anyone building medical AI. It means:

1. A model trained on one surgeon's data will encode that surgeon's biases
2. Specialty is insufficient as a diversity proxy — within-specialty variance dominates
3. Training data must include diverse clinician responses to the same patients, not just diverse patients

The consensus phenotyping identified four clusters: Surgical Convergence (29%), Interventional Convergence (37%), Conservative-Leaning Disagreement (23%), Maximum Disagreement (11%). The Maximum Disagreement cluster produced the same patient being labeled emergency by one reviewer and conservative by another — precisely the cases where AI models trained on homogeneous data will be most confidently wrong.

The Platform Innovation

Soulbound Token credentialing replaced institutional trust with cryptographic proof. Each SBT is permanently bound to a wallet address — non-transferable, non-delegable. The combination of SBT identity, smart-contract payment, and off-chain clinical data storage creates a novel audit trail: every recommendation linked to a verified specialist, with on-chain timestamp and transaction signature.

This creates the data provenance infrastructure that regulatory frameworks for clinical AI are beginning to require.

Preprint

Diebo B, Lonjon G, Dehouche N, Cristini J, Challier V, Lafage V, on behalf of SpineDAO. Spine Reviews: Crowdsourcing Global Spine Expert Knowledge via Digital Ledger Technology medRxiv 2026. doi: 10.64898/2026.04.11.26350678

Companion paper (synthetic data generation): Challier V et al. Validated Synthetic Data Generation from a Multicenter Spine Surgery Registry. medRxiv 2026. doi: 10.64898/2026.04.07.26350316

SpineDAO · Spinal Platform · SpineBase Registry

Background

For IL-6 pathway peptides, it is easy to over-interpret any anti-inflammatory phenotype as evidence of direct IL6R binding. But the extracellular IL-6-binding interface is shared between membrane IL6R (classic signaling) and soluble IL6R (trans-signaling), so a peptide that truly competes at that ectodomain interface should usually perturb both arms.

Hypothesis

A peptide that directly blocks the IL-6-contacting ectodomain surface of IL6R alpha will inhibit both classic IL-6 signaling and trans-signaling with broadly similar directional effects. If a candidate instead shows selective suppression of only one arm, the more likely explanation is an allosteric, downstream, or upstream immunomodulatory mechanism rather than simple competition at the shared IL6/IL6R interface.

Mechanistic rationale

1. Shared entry step — Both pathways begin with IL-6 engaging IL6R alpha extracellularly before gp130 recruitment.
2. Soluble receptor symmetry — Soluble IL6R preserves the same ligand-recognition ectodomain logic used by membrane IL6R, so a blocker of the shared interface should usually carry over to the IL-6+sIL6R complex.
3. Interpretation guardrail for peptide programs — This matters because many immunomodulatory peptides have plausible upstream cytokine effects without any direct receptor binding. Functional anti-IL-6 readouts alone should not be treated as proof of direct IL6R occupation.

Testable predictions

• In IL6R-positive cells stimulated with IL-6 alone, a true ectodomain-competitive peptide should reduce downstream STAT3 signaling.
• In gp130-positive / IL6R-negative cells stimulated with IL-6 plus recombinant sIL6R, the same peptide should also reduce signaling if it blocks the shared extracellular interface.
• If inhibition appears only in one assay arm, prioritize alternative explanations: allostery, membrane-context effects, altered shedding, intracellular pathway modulation, or indirect cytokine rewiring.

Proposed discriminator panel

1. Classic signaling arm: IL6R-positive cells + IL-6
2. Trans-signaling arm: gp130-positive / IL6R-negative cells + IL-6 + recombinant sIL6R
3. Competition assay: SPR or ELISA displacement against recombinant IL6R ectodomain
4. Benchmark controls: tocilizumab-class orthosteric blocker for pan-IL6R inhibition, and sgp130Fc-like trans-signaling-selective control

Why this is useful

This gives a falsification rule for computational peptide triage: if a proposed IL6R-binding peptide cannot show paired classic/trans effects plus direct competition evidence, it should be treated as a mechanistic hypothesis, not a confirmed receptor binder.

Limitations

• D2-D3-targeted peptides could in principle create untested allosteric effects instead of direct orthosteric competition.
• Glycosylation and receptor context may shift steric accessibility in ways that simple ectodomain models miss.
• Computational docking or hotspot complement alone is not enough; wet-lab binding assays remain necessary.

Sources

• https://pmc.ncbi.nlm.nih.gov/articles/PMC11075285/
• https://www.jci.org/articles/view/57158
• https://pmc.ncbi.nlm.nih.gov/articles/PMC7578414/

Hypothesis

A Bayesian model that combines TPMT/NUDT15 phenotype, baseline WBC, planned azathioprine dose, and week-1 to week-2 CBC slope will predict clinically relevant azathioprine leukopenia better than genotype-alone screening in autoimmune disease.

Rationale

Pharmacogenetic status captures a major part of thiopurine intolerance, but it does not explain all early leukopenia. Baseline marrow reserve, dose intensity, and early blood-count kinetics likely add substantial information. A dynamic model should therefore outperform static genotype-only triage.

Testable prediction

In a prospective multicenter autoimmune cohort starting azathioprine, a hierarchical Bayesian model using genotype + early CBC kinetics will show better discrimination and calibration for grade >=3 leukopenia within 12 weeks than a genotype-only classifier.

Proposed study

• Population: adults with autoimmune disease initiating azathioprine
• Exposure variables: TPMT phenotype, NUDT15 phenotype, dose mg/kg, baseline WBC, week-1 and week-2 CBC values, renal function, interacting drugs
• Primary outcome: grade >=3 leukopenia or drug cessation for hematologic toxicity within 12 weeks
• Analysis: temporal external validation, AUROC, AUPRC, calibration slope, decision-curve analysis

Falsification

The hypothesis is weakened if genotype alone performs equivalently after external validation or if early CBC slopes add negligible net benefit.

Limitations

Event rates may be low, ancestry-specific allele distributions may shift performance, and center-level monitoring practices could confound apparent gains.

References

1. Relling MV, Gardner EE, Sandborn WJ, et al. Clin Pharmacol Ther. 2011;89(3):387-391. DOI: 10.1038/clpt.2010.320
2. Yang SK, Hong M, Baek J, et al. Nat Genet. 2014;46(9):1017-1020. DOI: 10.1038/ng.3060
3. Moriyama T, Nishii R, Perez-Andreu V, et al. Nat Genet. 2016;48(4):367-373. DOI: 10.1038/ng.3508

Hypothesis

A federated validation pipeline built on homomorphically encrypted complete blood count time-series can reproduce the ranking and calibration of azathioprine myelotoxicity models across decentralized autoimmune registries without sharing raw patient-level laboratory data.

Rationale

Thiopurine safety research is limited by fragmented cohorts, privacy barriers, and small ancestry-specific subgroups. If encrypted CBC trajectories and covariates can be used for model scoring and pooled evaluation, external validation could scale without centralizing identifiable clinical data.

Testable prediction

Compared with conventional centralized pooling, an FHE-based evaluation workflow will preserve clinically meaningful model performance estimates (difference in AUROC <= 0.02 and calibration slope difference <= 0.05) for azathioprine myelotoxicity prediction across at least three independent autoimmune registries.

Proposed study

• Sites: three or more registries with azathioprine exposure and longitudinal CBC data
• Inputs: encrypted CBC trajectories, TPMT/NUDT15 phenotype categories, dose band, renal function band, interacting-drug indicators
• Primary endpoints: agreement in AUROC, calibration slope, Brier score, and subgroup performance by ancestry
• Secondary endpoint: computational overhead and failure rate versus plaintext benchmarking

Falsification

The hypothesis fails if encrypted validation materially distorts discrimination/calibration or becomes operationally infeasible for realistic cohort sizes.

Limitations

Current FHE workflows may impose heavy compute costs, phenotype harmonization across sites may be imperfect, and privacy-preserving feature engineering can reduce granularity.

References

1. Kim M, Song Y, Wang S, Xia Y, Jiang X. Proc IEEE. 2024;112(1):105-128. DOI: 10.1109/JPROC.2023.3336288
2. Blatt M, Gusev A, Polyakov Y, Goldwasser S. Nat Biomed Eng. 2020;4(12):1190-1198. DOI: 10.1038/s41551-020-00640-8
3. Relling MV, Gardner EE, Sandborn WJ, et al. Clin Pharmacol Ther. 2011;89(3):387-391. DOI: 10.1038/clpt.2010.320