When a genetic association found in European biobank participants looks weaker in people of African, South Asian, or East Asian ancestry, the usual explanation is genetic. The marker SNP no longer sits next to the true causal variant in the other population, or the variant acts differently on a different genetic background. For anatomical shape traits there is a second, mostly unexamined explanation. The trait itself may be measured worse in non-European people. Brain and heart shape phenotypes are built by warping every scan onto a template made from mostly European anatomy, then compressing the result with a principal component analysis fitted to mostly European data. If non-European anatomy fits that template less well, its measured shape carries more registration error. More error weakens every genetic association, and the weakening looks exactly like poor genetic transferability.
This note proposes to measure that confound and to remove it. We would build shape phenotypes whose coordinate system does not depend on any reference population. The tool is landmark-free shape analysis, which compares surfaces as geometric measures and so needs no point-to-point correspondence or template. The test is whether these phenotypes close the gap between ancestry groups in measurement reliability and in replication of genetic effects. The core claim is falsifiable. If a phenotype built without a population-specific template has equal reliability across ancestry groups, and European lead loci replicate more strongly in non-European samples when measured this way, then part of the published "ancestry bias" in shape genetics was a measurement artifact. If replication does not improve, the gap is genetic, and fixing it is a genetics problem rather than a geometry problem.
What exists and what is missing
The closest work is the brain-shape genome-wide association study of Genetic influences on the shape of brain ventricular and subcortical structures. It computes vertex-wise radial distance and surface Jacobian metrics on co-registered subcortical meshes in more than 32,000 White British UK Biobank participants. It reduces each structure to 10 principal components and finds 60 loci, 19 of which volume alone cannot detect. It then checks the findings in non-British and non-European UK Biobank subsets and in the racially diverse ABCD cohort of about 8,500 adolescents. This paper motivates the proposal: shape carries genetic signal that volume misses, and the authors themselves flag that it is unclear whether effect sizes carry over across ancestries. Its pipeline, template registration followed by PCA fitted on the discovery sample, is exactly the construction we suspect of ancestry-dependent error.
Other imaging-genetics studies share the same structural dependence and are tested in even narrower populations. Three-dimensional cardiovascular imaging-genetics regresses SNPs on atlas-registered left-ventricle meshes in 1,124 Caucasian volunteers. Deep learning enabled analysis of cardiac sphericity collapses the ventricle to a single 2D axis ratio. TransferGWAS runs association tests on 10 PCA components of pretrained CNN features. None of them asks whether the phenotype is measured equally well across ancestries. GIANT comes closest in spirit: it reports that its heritability-driven parcellation keeps its validity across ancestries. But it works on voxel grids, not surface shape, and its parcellation was discovered in a European subset.
The landmark-free methods supply our tools. Varifolds treat a surface as a distribution of positions and unoriented tangent planes. Comparing two surfaces then becomes a kernel inner product, with no mesh registration, as used in BaRe-ESA and implemented in Deformetrica 4. The smooth Euler characteristic transform takes a different route. It sweeps a plane across the shape in many directions and records how the topology of the swept part changes, giving a smooth function of direction and height (Randomness of Shapes and Statistical Inference on Shapes via the Smooth Euler Characteristic Transform). The related SINATRA pipeline maps statistical evidence from those curves back onto the mesh, though it has only been used on 59 primate molars. FlowSSM learns non-linear deformation models without correspondences.
Two limitations noted in this literature constrain the design. Projection-based transforms and measure kernels still need a global pose alignment. And arbitrary points in the kernel feature space do not correspond to real shapes. So "landmark-free" by itself is not enough. A varifold embedding summarized by PCA fitted on Europeans would reintroduce the same reference-population dependence through the back door.
Core Claim
A shape phenotype can carry hidden ancestry dependence through three routes. The first is the template used for registration. The second is the data used to fit the low-dimensional basis. The third is the pose alignment. Our claim is that the first two cause a measurable share of the cross-ancestry attenuation in shape genetics, and that a reference-free construction removes it.
The reference-free construction has three parts. Pose is set by rigid principal-axis alignment, which needs no mean shape. The surface is represented by a varifold or smooth Euler characteristic transform. Low-dimensional coordinates are fixed in advance rather than fitted to data. For the Euler transform this means a preset grid of directions and heights. For the varifold it means projecting each shape's measure onto kernel features centered at a fixed spatial grid.
The logic is a standard measurement-error relation. Let $R_g$ be the test-retest reliability of a phenotype in ancestry group $g$, meaning the share of its variance that is not measurement noise. Let $\rho$ be the true genetic transferability of an effect. The ratio of standardized SNP effects observed in a non-European group $A$ versus Europeans $E$ is then approximately
$$\frac{\hat\beta_A}{\hat\beta_E} \approx \rho \sqrt{R_A / R_E}.$$
The square root arises because classical noise inflates the phenotype's variance while leaving its covariance with genotype unchanged. Current pipelines report the left-hand side and read it as $\rho$. We will estimate $R_A$ and $R_E$ separately for each phenotype construction, which lets us separate the geometry term from the genetics term.
Research Program
The first stage is a controlled simulation that isolates the mechanism. We take real subcortical meshes and plant a known shape deformation driven by a simulated genotype, with exactly the same effect in every group. We then vary group-level baseline anatomy, such as ventricle size and hippocampal curvature, by amounts matching the ancestry-group differences observed in the cohorts. Each phenotype pipeline is run on the result. Because $\rho = 1$ by construction, any ancestry gap in recovered effect size comes from the phenotype.
The second stage is empirical, on the structures studied by Zhao et al., namely the hippocampus and lateral ventricles. We use UK Biobank, split by genetic ancestry, and ABCD as the external diverse cohort. Reliability is estimated two ways. One is the UK Biobank repeat-imaging subset, where non-European counts will be small. The other is re-segmentation perturbation, which can be applied to everyone.
For each pipeline we will then estimate four quantities. These are SNP heritability within each ancestry group, the reliability gap, the replication slope of European lead loci in non-European samples, and that slope after reliability correction. We then check whether the corrected slopes agree across pipelines. Agreement would mean the pipelines differ only in measurement quality.
A third, secondary stage asks whether the two-sample Mendelian randomization estimates linking hippocampal and ventricular shape to Alzheimer's disease, reported by Zhao et al., shift when the exposure is measured reference-free. This would show whether phenotype error propagates into causal claims as weak-instrument bias, which is distinct from the LD-tagging failures documented for polygenic scores (Ancestry-specific polygenic scores and SNP heritability of 25(OH)D; On cross-ancestry cancer polygenic risk scores).
Stratification is handled throughout with genetic principal components and head-size covariates. Without them, real anatomical differences between ancestry groups could leak into the associations.
Baselines
The naive reference point is scalar volume of each structure. Volume needs no template, so it is a fair test of whether reliability gaps exist at all when no registration is involved. It is also the comparison Zhao et al. used to show that 19 of 60 loci are invisible to volume.
The strongest existing method is the Zhao et al. pipeline: template-registered vertex metrics with PCA fitted on the White British discovery sample. We will re-implement it on the same scans, the same discovery and test splits, and the same association software. Its reported reference numbers, from Genetic influences on the shape of brain ventricular and subcortical structures, are 60 genome-wide significant loci at $P < 1.09 \times 10^{-10}$ and SNP heritability of 22.1% to 27.3%. We found no reported result on Lacuna for ancestry-stratified replication slopes or reliability for this pipeline, so we will produce that number ourselves. The proposal beats this baseline if three things hold together. The non-European-to-European reliability ratio must be closer to one. The raw replication slope of European lead loci in non-European samples must be higher. And European heritability must stay within a few points of the template pipeline. Losing much heritability would mean we removed signal along with error.
The most threatening baseline repairs the template pipeline directly. It uses a multi-ancestry template, built from an ancestry-balanced subsample, together with PCA fitted on an ancestry-balanced pool. It receives the same subjects and the same tuning budget for template construction.
We also include two learned landmark-free alternatives. One is FlowSSM latent codes. Its reported distal femur generality error is 0.24 ± 0.04 mm in FlowSSM, and it would be trained on the same balanced pool. The other is a TransferGWAS-style embedding of the brain scan from the pretrained UK Biobank SFCN network (UKBiobank_deep_pretrain), reduced to 10 components.
The key ablation keeps the varifold representation but replaces the fixed coordinate basis with PCA fitted on Europeans. Comparing the two isolates whether the benefit comes from dropping correspondences or from dropping the fitted basis.
Why not just…?
Why not just build a multi-ancestry template? This is the real threat. If a balanced template with pooled PCA closes the reliability gap as well as the reference-free pipeline, the honest conclusion is that template composition matters and correspondence-free geometry is unnecessary. The diagnostic stage would still be the contribution. Two arguments suggest it may not be enough. Any single template still sits closer to some anatomies than others. And balancing is limited by how few non-European scans the cohorts contain. The direct comparison of the two pipelines settles the question, and the simulation, where baseline anatomy is varied beyond the training range, will show which construction degrades first.
Why not just use volume, which needs no template? Volume avoids registration error but discards genetic signal. Zhao et al. found that 31.7% of shape loci are invisible to volume. Three-dimensional cardiovascular imaging-genetics found that scalar left-ventricle mass detected none of six known SNPs, where surface modeling detected four. Volume therefore serves as our reliability floor, not our competitor for discovery.
Why not just add genetic principal components? These covariates remove false positives from population stratification. They cannot change how noisy the phenotype is within a group, so they leave $R_A/R_E$ untouched. Every pipeline in the comparison already includes them.
Why not just use pretrained CNN embeddings? The CNN is trained on UK Biobank brains, so it carries a reference population inside its weights. Its embedding is also sensitive to acquisition and pose. We include it as a baseline, and we expect a reliability gap similar to the template pipeline. If it shows no gap, that would count against our mechanism.
Proposed Experiments
The decisive simulation check is simple. With identical planted effects across groups, the template-plus-European-PCA pipeline should recover a smaller effect in the anatomically shifted group, and the reference-free pipeline should not. If both recover equal effects, registration error is not a material confound, and the program stops there.
The decisive empirical check is whether reliability-corrected replication slopes converge across all pipelines while raw slopes differ. Convergence would show that measurement, not genetics, accounts for the gap between pipelines. The size of the remaining gap, where $\rho < 1$, would then be a cleaner estimate of true genetic non-transferability for brain shape than any currently reported.
A last check guards against pose artifacts. We compare principal-axis alignment with alignment to a European mean. If the reliability gap reappears under the second, pose alignment is a third hidden route for ancestry dependence.
References
- Genetic influences on the shape of brain ventricular and subcortical structures
- Three-dimensional cardiovascular imaging-genetics: a mass univariate framework
- Deep learning enabled analysis of cardiac sphericity
- transferGWAS: GWAS of images using deep transfer learning
- A genetically informed brain atlas for enhancing brain imaging genomics
- BaRe-ESA: A Riemannian Framework for Unregistered Human Body Shapes
- Deformetrica 4: An Open-Source Software for Statistical Shape Analysis
- Randomness of Shapes and Statistical Inference on Shapes via the Smooth Euler Characteristic Transform
- A statistical pipeline for identifying physical features that differentiate classes of 3D shapes
- Landmark-free Statistical Shape Modeling via Neural Flow Deformations
- Ancestry-specific polygenic scores and SNP heritability of 25(OH)D in African- and European-ancestry populations
- On cross-ancestry cancer polygenic risk scores