Many people who carry a heavy genetic burden for a disease never develop it. Genetics usually explains this away as noise or as "missing heritability", the gap between how heritable a trait looks in family studies and how much variance measured genetic variants actually explain. Computational neuroscience suggests a more specific answer. Physical structure can make a system's function insensitive to variation in its parts. In crab neurons, the shape and electrical coupling of the cell keep rhythmic bursting intact while all fourteen ion-channel densities are perturbed at once (Neuronal morphology enhances robustness to perturbations of channel densities). If human tissue structure behaves the same way toward genetic perturbation, then genetic risk should matter a great deal in people with little structural reserve and very little in people with ample reserve. A model that averages over both groups would systematically understate how much genetics matters, and it would assign the wrong uncertainty to every individual forecast.
We propose to test this claim in the Alzheimer's Disease Neuroimaging Initiative (ADNI), using a forecasting model whose output is a full probability distribution. In that model, structure (for example amygdala and hippocampal volume) is allowed to shrink both the size of the polygenic effect and the spread of outcomes. Two complications make the test hard and worth doing. First, in a single snapshot of data, small brain structure can be a cause of decline, a consequence of it, or simply a marker of how far the disease has progressed. Second, genetic effect estimates in cohorts of ADNI's size carry wide, non-Gaussian uncertainty. Our design treats both problems as part of the question rather than as caveats to mention afterward.
Core Claim
Existing work that links Alzheimer's genetics to brain structure treats structure as a mediator: genes change the amygdala, and the amygdala changes disease risk. The module-level polygenic score framework of A Module-Level Polygenic Risk Score-Based NetWAS Framework is the clearest example. It projects Alzheimer's genome-wide association signals onto an amygdala-specific protein-interaction network and shows that the resulting module scores separate clinical stages better than single imaging-linked variants.
Our claim is different. Structure may also act as a moderator. It changes the slope and the variance of the genotype-to-outcome map, which is what buffering means in the neuron models. A moderator effect and a mediator effect predict different things. A mediator shifts the expected outcome. A buffer flattens the polygenic slope and narrows the predictive distribution as reserve grows.
We write the forecast for person $i$ as
$$y_i \sim \mathcal{D}\big(\mu_i, \sigma_i\big), \quad \mu_i = \alpha + \beta_0 e^{-\kappa S_i} G_i + \gamma S_i + \delta^\top x_i, \quad \log \sigma_i = \eta_0 + \eta_S S_i + \eta_G G_i,$$
where the symbols are defined as follows:
- $y_i$ is the future outcome: the change in a cognitive score, or the time to conversion from mild cognitive impairment to Alzheimer's disease.
- $G_i$ is a polygenic score, computed either from external summary statistics or at the module level.
- $S_i$ is baseline structural reserve, measured as standardized regional volume.
- $x_i$ are covariates, including a latent disease stage described below.
- $\mathcal{D}$ is a location-scale distribution.
- $\kappa$ is the buffering coefficient.
The proposal stands or falls on whether $\kappa > 0$ and $\eta_S < 0$ hold prospectively, after disease stage has been accounted for, and on whether modeling them improves calibrated forecasts. The exponential form is a simple monotone choice. We will also check a spline version of the slope to see whether buffering saturates.
A saturating slope would also bear on an unexplained tension in the cross-sectional literature. Insulin sensitivity relates linearly to heart failure (Association between estimated glucose disposal rate and heart failure) but shows a threshold for cognitive impairment, with an inflection point at 6.36 mg/kg/min (Association between estimated glucose disposal rate and cognitive impairment). A reserve that saturates would produce exactly such a threshold. We treat this only as motivation, not as evidence.
Research Program
The first component is a ground-truth testbed. Before touching human data, we need to know whether our estimator can recover buffering at all from snapshot data, and how reverse causation corrupts it. We will reuse the roughly 1,300 valid lateral pyloric neuron models from the morphology study (PNAS version; Neuronal morphology enhances robustness to perturbations of channel densities). The fourteen channel conductances play the role of "genotype", axial resistivity of the axon plays the role of "structure", and rebound bursting (or a continuous burst-quality measure) plays the role of "phenotype".
In that work, buffering has a known optimum. Resistivity of about 95–125 Ω·cm preserves bursting, and both 2 and 500 Ω·cm destroy it. We sample synthetic "populations" of perturbed models, fit the location-scale model, and check whether the estimated $\kappa$ and $\eta_S$ recover the known window. We then inject the three cross-sectional problems on purpose:
- Reverse causation: structure drifts in response to phenotype.
- Survival selection: non-bursting models are dropped from older "strata".
- Misclassified phenotypes: the outcome label is randomly flipped at a set rate.
This tells us which failure each design choice guards against. It is the only setting in the proposal where the right answer is known.
The second component is the main human test in ADNI. Baseline genotypes, structural MRI, amyloid measures and repeated clinical visits are all available there. Polygenic scores are computed from external Alzheimer's summary statistics, which is how MPRS-NetWAS did it. Disease stage is the hard confounder: a person with a small hippocampus may simply be further along. To handle it, we borrow the latent time-shift model from Probabilistic disease progression modeling. That paper fits a Gaussian-process progression curve, a smooth nonparametric trajectory learned jointly across many biomarkers, and gives each person an inferred position along that curve. In our model, that position, together with its uncertainty, enters $x_i$ as a covariate. Structure therefore has to explain variation in genetic sensitivity beyond what stage explains.
We run the primary analysis in people who are cognitively normal at baseline. In this group, reduced structure is least likely to be a consequence of the outcome being forecast.
Inference is fully Bayesian, because the regime is small-sample. Quantifying the uncertainty in heritability showed that at N=892, and even more so at N=160, posteriors for heritability in ARIC become broad, skewed or flat, and normal approximations give misleading intervals. We use the same reasoning for $\kappa$ and report its full posterior.
As a secondary analysis, we estimate SNP heritability separately within strata of low and high reserve, using that paper's eigendecomposed mixed model. Buffering predicts higher heritability where reserve is low. We state in advance that ADNI per-stratum posteriors may well be nearly flat. If they are, we report that honestly rather than reading signal into them. The polygenic-slope test is the primary test because it borrows strength from external summary statistics and needs far less data.
The third component is a direction-of-effect check. Following Cross-fitted instrument: A blueprint for one-sample Mendelian randomization, we use cross-fitted genetic instruments for regional volume in a one-sample design. These instruments are sensible because brain structure is itself heritable (Multidimensional heritability analysis of neuroanatomical shape). The question is whether genetically predicted volume, which cannot be caused by later disease, reproduces the attenuation of the polygenic slope. This check addresses the reverse-causality objection that every cross-sectional study in the file leaves open.
Baselines
The naive reference is a forecaster that uses only age, sex, education and baseline cognition. It emits a predictive distribution with constant spread. Every richer model must beat it on the continuous ranked probability score (CRPS), a proper scoring rule that rewards both accuracy and honest spread (Predictive Uncertainty Quantification via Risk Decompositions for Strictly Proper Scoring Rules), and on calibration error.
The mediator baseline adds the polygenic score and regional volume as additive terms. This is the statistical content of the MPRS-NetWAS view, with module scores in place of a genome-wide score. It is the most directly competing account of how structure and genetics combine. No reported result found on Lacuna for MPRS-NetWAS AUC or regression accuracy, so we reimplement it on our splits rather than compare against a quoted number.
The strongest probabilistic competitor is the Gaussian-process progression model of Probabilistic disease progression modeling with polygenic score and volume added as covariates. It already produces calibrated, individualized trajectories under irregular sampling. No reported result found on Lacuna for its numerical forecasting accuracy (the file reports only that time-shift recovery on synthetic data had "high" R²).
The key ablation is our model with $\kappa$ and $\eta_S$ fixed at zero, every other component unchanged. This isolates whether the buffering terms themselves carry predictive and calibration value.
All models share the same subject-level splits and covariates. Longitudinal visits are held out by subject. Each model gets the same budget for hyperparameter search and the same inference compute. Beating the baselines means two things together. The full model must reduce held-out CRPS relative to the zero-buffering ablation, with a paired bootstrap interval that excludes zero. And interval coverage must stay near nominal in the low-reserve stratum and the high-reserve stratum separately.
Why not just…?
Why not just add a polygenic-score-by-volume interaction term to a logistic regression? This is the most serious threat, and it captures half of our claim. A mean interaction can represent a flattening slope. It cannot represent narrower outcome spread at high reserve, and it does not handle disease stage as an uncertain latent variable. The experiment that settles it is a variant of our model where the interaction enters only the location: $\kappa$ is free and $\eta_S$ is fixed at zero. If that variant matches the full model on CRPS and coverage, the variance part of buffering is unnecessary. We would then report that buffering is a slope effect only, which is still a finding that the mediator framing misses.
Why not just use the Gaussian-process progression model with genetics as a covariate? It handles time and uncertainty well. As usually specified, though, covariates shift trajectories rather than rescale genetic sensitivity, so it averages over the reserve strata. If it matches our model, buffering adds nothing beyond flexible trajectory modeling. It is one of our baselines precisely so that this possibility is tested directly.
Why not just run Mendelian randomization on volume and disease? Mendelian randomization estimates one average causal effect of structure. It answers whether structure matters, not whether structure changes how much genes matter. We use it only as a direction check on the attenuation, not as the forecasting model.
Why not just stratify heritability with standard confidence intervals? At ADNI's sample size, the ARIC results in Quantifying the uncertainty in heritability show that normal-approximation intervals can be badly wrong. They can imply negative heritability or overconfident precision near the boundary. Bayesian posteriors are required, and they may turn out uninformative.
Decisive Checks
In the neuron testbed, the estimated buffering curve must peak inside the published 95–125 Ω·cm window. Injecting reverse causation must visibly bias a purely cross-sectional fit. That bias is the evidence we would cite to justify the longitudinal and instrument-based checks in humans.
In ADNI, buffering is supported only if four conditions all hold:
- The posterior for $\kappa$ excludes zero among cognitively normal participants.
- It remains away from zero after conditioning on latent disease stage.
- The instrumented volume reproduces its sign.
- The full model improves held-out CRPS over the zero-buffering ablation without losing coverage.
If $\kappa$ vanishes once stage is controlled, then "reserve" was merely a proxy for disease severity. In that case the mediator account stands, and the missing-heritability explanation we propose is wrong for this phenotype.
References
- Neuronal morphology enhances robustness to perturbations of channel densities
- A Module-Level Polygenic Risk Score-Based NetWAS Framework for Identifying AD Genetic Modules Mediated by Amygdala: An ADNI Study
- Association between estimated glucose disposal rate and heart failure in patients with diabetes or prediabetes: a cross-sectional study
- Association between estimated glucose disposal rate and cognitive impairment in elderly patients with type 2 diabetes mellitus: a cross-sectional study
- Probabilistic disease progression modeling to characterize diagnostic uncertainty: Application to staging and prediction in Alzheimer's disease
- Quantifying the uncertainty in heritability
- Cross-fitted instrument: A blueprint for one-sample Mendelian randomization
- Multidimensional heritability analysis of neuroanatomical shape
- From Risk to Uncertainty: Generating Predictive Uncertainty Measures via Bayesian Estimation