Mechanical Fingerprints for Mendelian Randomization of Cortical Shape

Genetic studies can now say, with some confidence, that an exposure such as poor lung function causally changes the size and thickness of particular brain regions. They cannot say how. A change in regional cortical thickness could come from a local program that builds thinner tissue. It could also be a by-product of the cortex folding differently because it grew less in surface area. Or it could be adult tissue loss that has nothing to do with development. This proposal builds a physics simulator of cortical folding that includes realistic tissue relaxation. The simulator turns each candidate mechanism into a predicted pattern of changes across all 34 standard brain regions, in both surface area and thickness. We then ask which pattern a genetic causal estimate actually matches.

The method that supplies the causal estimates is Mendelian randomization (MR). MR uses inherited DNA variants as a natural randomized trial: variants that shift an exposure are assigned at conception, so their downstream effects are protected from lifestyle confounding. The physics comes from differential growth simulation, in which a cortical layer expanding faster than the tissue beneath it buckles into folds. Adding tissue relaxation normally makes such simulations cost grow with the square of simulated time. A Markovian embedding avoids this. It replaces the memory integral with a few extra "internal" state variables that are updated step by step, so cost grows linearly. That makes the hundreds of long developmental runs needed for the fingerprints affordable. The two-sentence version is this: simulate what each biological mechanism would do to regional area and thickness, then decompose MR's causal effects onto those simulated fingerprints. Whatever the fingerprints cannot explain is evidence for a non-mechanical or non-developmental pathway.

Core Claim

MR estimates of an exposure's effect on the 34 Desikan-Killiany regions form a 68-dimensional vector (area and thickness per region). Today that vector is read one region at a time, as in the lung-function study COPD, PRISm and lung function reduction affect the brain cortical structure: a Mendelian randomization study. That study reports, for example, a COPD effect on pars orbitalis area of β = −3.351 (P = 0.035) and an FVC effect on global surface area of β = 1576.61 (P = 0.008). It also notes a puzzling increase in area in some regions, which it attributes to compensatory angiogenesis without testing the idea.

We claim the vector has structure that mechanics predicts. Folding drives cortical thickness variations shows that folding alone produces thick gyri and thin sulci, with an adult ratio near 1.16 (2.87 vs 2.47 mm, n = 564). It follows that a genetic shift in tangential growth should move area and thickness together in a region-specific, folding-dependent way. That paper motivates the proposal: it exposes the unresolved tension between "thickness is genetic" and "thickness is mechanical." Our new contribution is to turn this tension into a testable restriction on MR effect vectors. We build a small set of simulated fingerprints and fit the MR vector as a combination of them. The fingerprints are reduced developmental tangential growth, reduced radial growth, altered stiffness or relaxation time, and uniform post-folding atrophy. The fit has formal error bars. The residual is an explicit estimate of effects that no mechanical pathway explains.

Research Program

The first component is the forward model. We start from the regional growth simulator of Role of data-driven regional growth model in shaping brain folding patterns. That work fits closed-form expansion curves (R² above 0.85) to dHCP and BCP scans from 29 postmenstrual weeks to 2 years. It drives a hyperelastic bilayer finite-element model with those curves, and it supplies both the growth laws and the reference geometry. We replace its rate-independent material with a finite-strain growth model in which the elastic part of the deformation relaxes over time.

Relaxation is written as a memory kernel approximated by a short sum of exponentials. Following the auxiliary-variable construction in Markovian Embedding Procedures for Non-Markovian Stochastic Schrödinger Equations and the Prony-series approach of Data-driven molecular modeling with the generalized Langevin equation, each exponential mode becomes one internal stress variable per integration point. That variable obeys a local first-order equation. A mode with relaxation rate $\gamma_k$ contributes an internal stress $q_k$ with

$$\dot q_k = -\gamma_k q_k + g_k \dot S_e,$$

where $S_e$ is the elastic stress computed from the elastic part of the deformation after the growth part is removed, and $g_k$ is the mode's weight. Because each $q_k$ carries all the memory it needs, the model never stores the deformation history. The technical work is making this stable under continuously changing reference configurations and self-contact of folds. Existing proofs of exact finite-dimensional embedding exist only in one-dimensional linear viscoelasticity Learning Markovian Homogenized Models in Viscoelasticity, so the three-dimensional growth case needs its own convergence checks. Measuring relaxation kernels from sparse MRI is ill-posed, so relaxation times are treated as swept nuisance parameters rather than fitted.

The second component is the fingerprint library. For each Desikan-Killiany region and each mechanism, we perturb the growth or material parameter by a small amount in that region only. We then record the change in regional area and mean thickness at the end of the simulated window, and extend the run with adult atrophy where the fingerprint requires it. This gives a sensitivity matrix $J$, with 68 rows for the region-phenotype outcomes and one column per mechanism-region pair. The matrix is computed at several relaxation times, including the hyperelastic limit.

This step requires hundreds of multi-month runs. It is the concrete reason the linear-time embedding matters. It is also where we test the finding of Role of data-driven regional growth model in shaping brain folding patterns that growth magnitude dominates and trajectory barely matters. That finding holds for elastic tissue. With relaxation, slow growth may dissipate stress that fast growth would release as a fold. If so, timing matters, and MR's assumption of a constant, lifetime-cumulative effect becomes a real limitation rather than a harmless simplification.

The third component is the causal decomposition. Let $\hat\beta$ be the 68-vector of MR estimates from inverse-variance-weighted two-sample MR, with covariance $\Sigma$. We estimate how strongly each mechanism is engaged, $\theta$, by generalized least squares of $\hat\beta$ on $J\theta$, with a sparsity penalty across regions. The weighted residual $r = \hat\beta - J\hat\theta$ is compared to its null distribution under $\Sigma$, with $\Sigma$ built from the MR standard errors and sample overlap. A small residual means the exposure's footprint is consistent with purely mechanical developmental pathways. A large, structured residual, most plausibly concentrated in thickness, is evidence for direct cytoarchitectural or non-developmental effects. Pleiotropy-robust versions, such as MR-Egger and weighted-median estimates of $\hat\beta$, are run through the same decomposition.

Baselines

The naive reference point is region-by-region two-sample MR with inverse-variance weighting plus the usual Egger, weighted-median, and outlier-removal sensitivity analyses, as in COPD, PRISm and lung function reduction affect the brain cortical structure: a Mendelian randomization study. It uses the same FinnGen, UK Biobank, SpiroMeta and ENIGMA summary statistics, the same instrument thresholds (P < 5×10⁻⁸, LD R² < 0.001, F > 10), and the same exposures. Its reported numbers are the effects quoted above. It yields a list of significant regions but no mechanism. Our method must reproduce those estimates exactly as its input. It beats the baseline only by making correct out-of-sample predictions the baseline cannot make, specifically of held-out regions and of infant data.

The strongest statistical competitor is multivariable MR or MR mediation, which treats regional surface area as a mediator of thickness effects. It is in the spirit of Mediation Analysis with Mendelian Randomization and Efficient Multiple GWAS Integration. It will be run on the same summary statistics, using ENIGMA area instruments, with the same pleiotropy corrections and no simulator. We found no reported result on Lacuna for this mediation design applied to cortical area and thickness. The comparison is at matched model complexity: mediation with one free coefficient per region versus our decomposition with a matched number of mechanism weights. Our decomposition beats mediation if it predicts held-out regions' thickness effects from their area effects with lower error, with both methods scored on predictive log-likelihood under $\Sigma$.

The strongest mechanical competitor is the hyperelastic regional growth model of Role of data-driven regional growth model in shaping brain folding patterns. It reports a sulcal-depth mean absolute percentage error of 13.41% in the supramarginal region, versus over 50% for isotropic growth. We rerun it with identical meshes, growth laws and perturbation sizes to build a hyperelastic fingerprint library. This serves as the ablation isolating relaxation. The viscoelastic model must first match its 13.41% sulcal-depth error on dHCP and BCP. To earn its cost, it must then improve either the decomposition fit on real MR vectors or the reproduction of the 1.16 gyral-to-sulcal thickness ratio. A second ablation removes the atrophy fingerprint, which tests whether a developmental story is needed at all.

Why not just…?

Why not just use the relaxation models already built into commercial finite-element software instead of a custom embedding? This is a fair objection, and we largely accept it. The Prony-series internal variables above are the same construction that such software uses for ordinary viscoelastic solids, so we use analytic modes as the primary method. We do not claim a new learned embedding. What remains open is whether those modes stay stable and accurate when the reference configuration itself grows, which standard implementations do not address. We check this by comparing against a brute-force memory-integral solver on short windows. We also try the recurrent-network internal variables of Learning Markovian Homogenized Models in Viscoelasticity only as a secondary check. That work warns that learned hidden states drift from physical internal variables under distribution shift, which is exactly what large-deformation buckling would cause.

Why not just run multivariable MR with area as a mediator of thickness? This is the most serious threat. If mediation predicts held-out thickness effects as well as the fingerprints do, the simulator adds nothing and we will say so. Our reason to expect it falls short is that mechanics predicts a region-dependent and possibly sign-changing coupling between area and thickness. In a homogeneous bilayer, folding thickens crests and thins troughs Folding drives cortical thickness variations, so the net change in regional mean thickness depends on how deeply that region folds. A single linear mediation slope cannot express this coupling. Mediation also cannot distinguish a developmental area effect from adult atrophy, which the fingerprints separate by their different spatial patterns. Experiment 2 below settles the question.

Why not just skip the physics and learn the mapping from exposures to brain shape with a deep generative model trained on imaging? A learned model can reproduce observed shapes, but it cannot produce counterfactual fingerprints for mechanisms that were never observed in isolation. Role of data-driven regional growth model in shaping brain folding patterns found that fitted growth curves alone did not yield realistic folds until coupled to mechanics. The decomposition needs one column per hypothesized mechanism, and only a mechanistic model can supply those columns.

Why not just use hyperelastic simulations, since magnitude dominates trajectory? That finding was obtained without relaxation. We test it directly by comparing the relaxed and hyperelastic libraries. If the libraries agree to within MR standard errors across the relaxation-time sweep, we drop viscoelasticity and report that MR's static assumption is mechanically justified for this problem. That would itself be a useful result.

Proposed Experiments

The first experiment is semi-synthetic recovery, and it is the precondition for everything else. We plant known mechanism weights in the simulator, generate regional phenotypes for synthetic individuals whose genotypes shift those weights, and add noise and horizontal pleiotropy at the levels implied by ENIGMA standard errors. We then run the full MR-plus-decomposition pipeline. Success means recovering the planted mechanism with calibrated coverage and a residual test with correct false-positive rate. We also measure how quickly identifiability degrades when two fingerprints are nearly collinear. We expect developmental radial-growth and post-folding atrophy fingerprints to be the hardest pair to separate, and the results will be reported per pair.

The second experiment is the real decomposition and the decisive test against mediation. It uses the lung-function and COPD exposures on the 34-region area and thickness outcomes. We hold out regions in a leave-one-lobe-out scheme, fit mechanism weights on the remaining regions, and predict the held-out regions' 68-dimensional effects. This asks whether the fingerprints explain the paradoxical area increases reported in COPD, PRISm and lung function reduction affect the brain cortical structure: a Mendelian randomization study as a mechanical redistribution, without invoking angiogenesis. It also asks whether the thickness residual is significant after mechanical explanation. The mediation and hyperelastic baselines are scored on the same held-out log-likelihood.

The third experiment is an independent developmental check that does not use genetics. Abnormal microstructural development of the cerebral cortex in neonates with congenital heart disease is associated with impaired cerebral oxygen delivery ties reduced oxygen delivery to disrupted cortical microstructure in neonates, and the preterm expansion maps of Dynamic patterns of cortical expansion during folding of the preterm human brain come from 30 preterm infants scanned between 28 and 38 weeks. If the MR decomposition assigns the lung-function effect to reduced developmental tangential growth, then infants with lower oxygen delivery or earlier birth in dHCP should show the same regional area and thickness fingerprint at 2 years in BCP. If adult atrophy is the assigned mechanism, that developmental pattern should be absent.

Two sanity checks gate the physics. The viscoelastic model must reproduce the roughly 1.16 gyral-to-sulcal thickness ratio from Folding drives cortical thickness variations, and it must match the 13.41% sulcal-depth error of Role of data-driven regional growth model in shaping brain folding patterns before any fingerprint is trusted. Wall-clock scaling against simulated time, compared with the brute-force memory solver, will confirm the linear-versus-quadratic cost that justifies the embedding. The project is falsified in a useful way if either of two things happens: the fingerprints fail to beat mediation on held-out regions, or relaxation leaves the fingerprints unchanged. Either outcome tells the MR community whether mechanics belongs in its interpretation of brain-structure effects.