Compromise-Bounded Pooling of Small Alzheimer's Cohorts to Test the GPX3 Antioxidant Axis

One of the more concrete genetic leads linking oxidative damage to Alzheimer's disease (AD) is GPX3. GPX3 is an antioxidant enzyme measurable in the fluid around the brain and spinal cord (cerebrospinal fluid, CSF). A common variant near the TNIP1 gene lowers its CSF level, and low GPX3 tracks tau pathology. The evidence, however, is assembled from small cohorts analyzed one at a time. The discovery cohort had 137 people, and the variant effect reached only P = 0.041. Replication cohorts used different protein-measurement platforms that do not always agree. The analyses that would settle the question include a pooled test of whether the variant acts through GPX3 to affect tau, with confidence intervals that are honest at small sample sizes. These analyses want individual-level data from every cohort at once, and that is exactly what biobanks will not ship.

This note proposes a way to fit those pooled models across cohorts. The idea is to split the computation across three protections. Sites add masks that cancel when summed. Homomorphic encryption, a form of encryption that allows sums to be computed on encrypted numbers, hides the totals from the coordinating server. A hardware enclave is a sealed region of a processor that even the machine's administrator cannot inspect. Here it holds the decryption key, takes the nonlinear model-fitting steps, and enforces what may be released.

The design goal is a stated bound on damage if the enclave is broken. Enclaves have been broken by speculative-execution and interface attacks. Under this design, an attacker who fully controls the enclave learns only cross-site totals, never any single cohort's data. The proposal is equally about an honest empirical question: does pooling actually beat ordinary meta-analysis of per-cohort results for this kind of small, multi-platform oxidative-stress finding? If it does not, the cryptography is unnecessary and we will say so.

Core Claim

Two claims are tested, and the second only matters if the first holds.

The statistical claim is narrow. Pooled individual-level analysis should beat inverse-variance meta-analysis in three specific settings:

  • Logistic models of amyloid-and-tau positivity within small cohorts, where some cohorts have few minor-allele carriers and per-cohort fits are unstable or fail to converge.
  • Mediation analysis, meaning variant → CSF GPX3 → p-tau, where the indirect effect is a product of two coefficients estimated on the same people.
  • Cross-cohort harmonization within a measurement platform, using quantile maps that need the pooled distribution.

For plain linear models with every cohort well powered, meta-analysis and pooling are expected to agree, and we do not claim otherwise.

The systems claim is that every quantity these models need can be written as a sum of per-site quantities. The fits include:

  • Newton–Raphson iterations of logistic regression.
  • A bootstrap in which each patient's resampling weight is drawn independently.
  • Binned histograms for quantile harmonization.

Writing everything as sums lets the enclave see only totals, and lets the transcript of what it saw be audited with reconstruction attacks.

What Prior Work Does and Does Not Settle

Motivation from the biology. The oxidative-stress literature supplies the scientific question rather than the method. Mechanistic work shows reactive oxygen species switching on cell-death signaling (Daphnetin protects oxidative stress-induced neuronal apoptosis via regulation of MAPK signaling and HSP70 expression) and correlating with symptom severity in Parkinson's disease (Potential roles of oxidative distress on neurodegeneration in Parkinson's disease with neuropsychiatric symptoms). The multi-omic TNIP1 study (Post‐GWAS multiomic functional investigation of the TNIP1 locus in Alzheimer's disease highlights a potential role for GPX3) turns this into a testable human readout. It reports CSF GPX3 falling with amyloid and tau positivity (ANOVA P = 1.5 × 10⁻⁵ in discovery), a replication association with p-tau (P = 4.38 × 10⁻⁹), and a weak variant effect (P = 0.041). It also notes disagreement between mass-spectrometry and aptamer-based platforms. That study analyzed cohorts separately in discovery and replication tiers. It did not pool them and did not run a formal mediation test, which is the gap we target.

Why imaging and forecasting work does not cover this. The TADPOLE challenge results (The Alzheimer's Disease Prediction Of Longitudinal Evolution (TADPOLE) Challenge: Results after 1 Year Follow-up) report that better diagnosis prediction was tied to CSF measurements. That supports biochemical rather than purely cognitive endpoints. UG-FedDA (UG-FedDA: Uncertainty-Guided Federated Domain Adaptation for Multi-Center Alzheimer's Disease Detection) federates MRI classifiers across ADNI, AIBL, and OASIS. It sends unprotected model updates and uses no biochemical markers. We borrow its multi-site setting, not its method.

The closest technical prior work. SAFETY (SAFETY: Secure gwAs in Federated Environment Through a hYbrid solution with Intel SGX and Homomorphic Encryption) sums Paillier-encrypted genotype counts on an untrusted server. It then decrypts the totals inside an SGX enclave to run nonlinear tests, and reports 1.7× to 4.82× speedups over pure SGX. We reuse its "sum encrypted, decrypt only the total" pattern but differ in three ways:

  • SAFETY is one-shot over counts, whereas we need iterative fits and resampling.
  • SAFETY gives the enclave the key with no protection if per-site ciphertexts reach a compromised enclave, whereas we add cancelling masks.
  • SAFETY ran on synthetic genotypes, whereas we evaluate on real AD cohorts with a statistical question attached.

Supplying the integrity mechanism. The trustworthy-HE design (Achieving trustworthy Homomorphic Encryption by combining it with a Trusted Execution Environment) has the enclave check authentication tags and perform the encrypted additions so that a malicious host cannot alter ciphertexts undetected. It reports about 1.7× slowdown over either component alone. We adopt its tag checking. Unlike that design, ours lets the enclave decrypt, because Newton steps and quantile maps cannot be computed on Paillier ciphertexts.

The strongest all-cryptographic competitor. Multiparty homomorphic encryption for oncology (Collaborative Privacy-Preserving Analysis of Oncological Data using Multiparty Homomorphic Encryption) shows that survival curves and logistic regression can be fit across hospitals with no trusted hardware at all. It reports 100-iteration logistic regression in about 500 seconds and precision matching cleartext to 7 decimal digits. This shows our models are feasible without an enclave. Its cost is interactive key-share decryption and polynomial approximation of nonlinear functions. We must therefore show that the enclave buys something measurable, such as fewer synchronous rounds, exact nonlinear steps, or easier resampling.

Constraints from the enclave literature. The attack papers (SgxPectre Attacks: Leaking Enclave Secrets via Speculative Execution, SmashEx: Smashing SGX Enclaves Using Exceptions, Interface-Based Side Channel Attack Against Intel SGX) are why we design for a compromised enclave rather than assume it is sound. The memory limits that drive Slalom and Goten (Slalom: Fast, Verifiable and Private Execution of Neural Networks in Trusted Hardware, Goten: GPU-Outsourcing Trusted Execution of Neural Network Training) are largely irrelevant here. Cohort tables of a few thousand patients fit easily in enclave memory, so we make no performance argument against running everything in plaintext inside an enclave.

Research Program

Protocol. Each round of fitting has the same structure:

  1. The enclave broadcasts the current coefficient vector.
  2. Each site computes, in plaintext on its own data, its contribution to the gradient and to the Hessian (the matrix of second derivatives that sets the Newton step size).
  3. Each site adds a mask agreed pairwise with the other sites, by key exchange that does not involve the enclave, so that the masks cancel in the sum.
  4. Each site encrypts the masked contribution additively and attaches an authentication tag.
  5. The untrusted server sums the ciphertexts.
  6. The enclave checks the tags, decrypts the total, takes the Newton step, and repeats.

Bootstrap confidence intervals. These use a Poisson bootstrap. Patient $i$ at any site receives an independent weight $w_{ib} \sim \text{Poisson}(1)$ in replicate $b$, generated locally from a seed that the enclave issues. For logistic regression with design rows $x_i$, outcome $y_i$, and fitted probability $\mu_i(\beta)$, the replicate's sufficient quantities are

$$ g_b(\beta)=\sum_{s}\sum_{i\in s} w_{ib}\,x_i\,(y_i-\mu_i(\beta)), \qquad H_b(\beta)=\sum_{s}\sum_{i\in s} w_{ib}\,\mu_i(\beta)(1-\mu_i(\beta))\,x_i x_i^{\top}, $$

where $s$ indexes sites. Both are sums over sites, so the masking and encryption pipeline applies unchanged. Mediation is fit the same way: two regressions, with the indirect effect and its bootstrap interval computed in the enclave from the per-replicate coefficients.

Within-platform harmonization. Sites on the same platform contribute encrypted histogram counts of GPX3 on a fixed grid. The enclave computes a pooled quantile map and returns it, and sites transform locally. Cohort and platform enter the model as fixed-effect indicator columns, which keeps everything a sum.

Release policy. The enclave releases only the final coefficients, intervals, and p-values. It refuses any fit in which a design cell, such as minor-allele homozygotes among amyloid-and-tau-positive patients pooled across sites, holds fewer than $k$ patients.

Leakage audit. We record the full transcript a compromised enclave would see, meaning every decrypted total across all iterations and bootstrap replicates. We then give it to reconstruction attacks of three kinds:

  • Linear solves that try to separate a single site's contribution.
  • Attempts to recover individual records in small cells from Hessians evaluated at many coefficient values.
  • Membership tests.

We run the same attacks on the transcripts of the baselines.

Baselines

The naive reference is the current practice: separate per-cohort analysis, as in the TNIP1 study. Each cohort fits its own regressions and the results are read side by side in discovery and replication tiers. It is run on identical cohort splits and covariates. It sets the floor that any pooled method must clear in power and in estimate stability. There is no reported result on Lacuna for this approach on a formal mediation test of the GPX3 axis. The reported reference points are the per-tier p-values above.

The strongest statistical competitor is random-effects inverse-variance meta-analysis of per-cohort estimates, including meta-analysis of per-cohort indirect effects. It needs no cryptography and is the standard tool in genetic epidemiology. It is run on the same cohort partitions with the same covariates and the same harmonized outcome, with heterogeneity estimated in the usual way. Pooling "beats" it only if, in simulation with known truth, pooled analysis achieves higher power at matched type-I error or better confidence-interval coverage for the indirect effect. The gain must hold in the small-cohort, sparse-allele regime and must not be explained by a single non-converging cohort. There is no reported result on Lacuna for meta-analysis on these cohorts.

The accuracy ceiling is a plaintext pooled oracle: all data on one machine with the same models. It is not a deployable baseline. Our secure pipeline should match it to numerical precision, and we target agreement at least as tight as the 7 decimal digits reported by the oncology work.

The strongest cryptographic competitor is the multiparty homomorphic encryption pipeline of Collaborative Privacy-Preserving Analysis of Oncological Data using Multiparty Homomorphic Encryption, which reports about 500 seconds for 100-iteration logistic regression and 7-digit precision. We re-implement it for our models, including the bootstrap, on the same machines and site count with the same iteration budget. We compare three things:

  • Wall-clock time.
  • The number of rounds in which every site must be online simultaneously.
  • Bytes communicated.

We beat it if we match its precision with a several-fold reduction in synchronous rounds, or in total time once bootstrap replicates are included. If we do not, the enclave is not earning its place for these models.

Pure-enclave pooling has sites upload encrypted records that are decrypted and analyzed in plaintext inside SGX. It is the fastest secure option and the right reference for what the extra layers cost in speed. Its compromise transcript is the raw data.

The ablations isolate the key ingredient. They remove, in turn:

  • The cancelling masks, which leaves the SAFETY-style arrangement.
  • The encryption, which leaves plain secure aggregation with totals visible to the server.
  • The integrity tags.
  • The release floor.

Each ablation is evaluated only on the leakage audit and on whether a malicious server can shift a reported coefficient without detection.

Why not just…?

Why not just meta-analyze per-cohort results? This is the real threat. For linear models with well-powered cohorts, meta-analysis is essentially as efficient as pooling, so in that regime the whole secure apparatus is pointless. The proposal stands or falls on the regime where it is expected to differ:

  • Logistic fits in cohorts of around 100 with few carriers.
  • Indirect effects that are products of noisy per-cohort coefficients.
  • Platform harmonization that benefits from pooled distributions.

Experiment 1 below is designed to settle this before any systems work is credited. If meta-analysis matches pooling there, we report that negative result and narrow the claim.

Why not just run everything inside an enclave? For tables this size it would be fast, and the memory limits that motivate Slalom and Goten do not bite. The objection is trust, not speed. The SGX attack papers show enclave contents can leak, and in that case every patient record is exposed. Our audit makes this concrete by giving the same attacks the pure-enclave transcript and ours. The answer is only convincing if our transcript yields materially less.

Why not just use multiparty homomorphic encryption with no enclave? The oncology results show it can fit logistic regression with excellent precision, so feasibility is not in question. The costs are its interactive decryption rounds, which need every site online each iteration, and its polynomial approximation of nonlinear steps. Its cost scales with bootstrap replicates and quantile maps. The head-to-head comparison measures whether these costs are large enough to matter for a multi-cohort mediation study with a thousand bootstrap replicates. If they are not, the right recommendation is the pure-cryptography route.

Why not just use secure aggregation and publish intermediate totals? Masking alone hides each site's contribution from the server but reveals every iteration's pooled gradient and Hessian. The encryption layer exists only to keep those intermediate totals away from the coordinator. The ablation that removes encryption and runs the reconstruction attacks on small design cells measures what those totals reveal. If they reveal nothing beyond the final output, the encryption layer should be dropped.

Proposed Experiments

Experiment 1 tests whether pooling helps at all. It uses the TADPOLE D1 set of ADNI, 1,667 subjects with CSF, genetic, and imaging data (TADPOLE Challenge: Prediction of Longitudinal Evolution in Alzheimer's Disease). We partition it by acquisition site into silos sized to mimic the TNIP1 cohorts: one of about 137, others of a few hundred. CSF amyloid and p-tau define positivity. Where an ADNI CSF proteomic measurement of GPX3 is available we use it. Otherwise we plant a GPX3-like variable with known genetic and mediating effects on real genotypes and real p-tau. We impose platform differences as monotone distortions on a subset of silos.

Across a grid of effect sizes, allele frequencies, and cohort sizes, we compare four approaches: separate analysis, meta-analysis, plaintext pooling, and our secure pooling. The metrics are power, type-I error under a planted null, confidence-interval coverage of the indirect effect, and the rate of failed per-cohort fits. The test is decisive: if pooling does not beat meta-analysis in the sparse regime, the remaining experiments are reported only as systems measurements.

Experiment 2 measures exactness and cost. On the same partitions, we compare our pipeline with the plaintext oracle on coefficient and p-value agreement. Against the re-implemented multiparty-encryption baseline and pure-enclave pooling, we measure:

  • Wall-clock time.
  • Synchronous rounds.
  • Bytes sent.
  • Scaling with bootstrap replicates and site count.

The pipeline is deployed on separate cloud instances in different regions, following the SAFETY evaluation style.

Experiment 3 is the compromise audit. For each design and ablation, we give the attacker the complete enclave or coordinator transcript and run the reconstruction and membership attacks described above. We report the smallest design cell for which individual values are recovered, how much of a single site's contribution can be isolated, and membership advantage. The bounded-compromise claim holds only if the full design leaks nothing per-site and nothing per-patient above the $k$ floor, while the ablations measurably do. A separate integrity test has a malicious server tamper with summed ciphertexts. The tags should detect every alteration that would shift a released coefficient.

Experiment 4 is a reproduction with real collaborators, conditional on data-use agreements. We would run the pooled, harmonized mediation analysis across the discovery and replication cohorts of the TNIP1 study. The first check is whether pooling reproduces the published per-tier associations. The scientific payoff is a single pooled estimate, with a bootstrap interval, of how much of the variant's link to tau runs through CSF GPX3. That is the number the oxidative-stress hypothesis actually needs.