Volume-Silent Loci: Genetic Effects on Brain Shape That Cancel Out in Volume

Most large brain-imaging genetics studies reduce each brain structure to one number, its volume, and most genetic interpretation starts from protein-coding sequence. This proposal argues that these two habits hide the same class of signal. A genetic variant that makes one part of the hippocampus swell and an adjacent part shrink leaves total volume unchanged, so a volume-based genome-wide association study (GWAS, a scan testing each genetic variant against a trait) cannot see it. We hypothesize that such spatially restricted, opposite-signed effects come mostly from non-coding regulatory variants. These are DNA changes outside genes that alter where and when a gene is switched on. Coding variants, by contrast, change the protein wherever it is expressed and so should tend to produce more uniform, volume-visible effects.

We propose to look for these hidden effects directly. The brain is aligned to a template in full 3D, and the local expansion or compression at every voxel is kept as the phenotype. Each structure's deformation is then split into the part that changes its volume and the part that cannot. A variant significant on the second part but not the first we call "volume-silent." The deliverable is a catalog of these loci in UK Biobank brain MRI, with a test of whether they are enriched for predicted regulatory disruption compared with loci that volume does detect. If they are, the fields' two blind spots reinforce each other: the regulatory genome is hidden twice, once by gene-centric annotation and once by volume-centric phenotyping.

Core Claim

The cancellation argument is well established for disease. Bansal et al. showed that atrophy in one subregion and compensatory expansion in a neighbor can sum to a normal total volume. Kuijf et al. found diabetic white-matter lesions that differ in number and elongation at identical total volume. On the prediction side, PASCS-MP predicts amyloid positivity from hippocampal surface shape statistics and outperforms volume, surface area and spherical harmonics. These papers motivate the proposal, but none of them is about genetics.

At biobank scale, imaging genetics has escaped volume only by giving up interpretability. TransferGWAS ran GWAS on 20 deep latent dimensions from a downsampled 3D variational autoencoder or a slice-wise 2D ResNet50 across 36,311 UK Biobank scans. It found 289 independent loci, but its features have no anatomical meaning and cannot tell us whether a locus acts by cancellation. The regional gene-expression TWAS found 1,065 associations, but against aggregate gray-matter volumes of 10 regions, which is exactly the phenotype that cancels. The partially functional linear regression model keeps the hippocampal surface as a continuous function, but uses 606 ADNI subjects and about 10,000 pre-filtered SNPs. It also reports that the surface explains only 1% to 4.6% of cognitive variance. That figure constrains us: per-variant shape effects will be small, so the test must be well powered and low-dimensional.

What is new here is the claim that shape associations are not only "more" than volume associations but specifically include a cancelling class with a definite genomic signature. The phenotype decomposition and the enrichment test are built to falsify that claim.

Research Program

Native 3D phenotype. We register each T1 scan to a population template with 3D non-rigid registration. For each voxel $v$ inside a structure $s$, we record the log-Jacobian $J_i(v)$, the log of the local volume change needed to map subject $i$ onto the template. This follows the template-deformation logic that Dubb et al. used for the corpus callosum. Doing it in 3D addresses the volumetric context loss documented in this literature. Roth et al. showed slice-wise prediction breaking continuity between adjacent slices, and Wolf et al. showed vision-language models performing at chance on relative position in medical images. A slice-wise encoder cannot represent "this end grows while that end shrinks" as a single coherent effect.

Volume-null decomposition. For structure $s$ with voxel set $V_s$, we expand $J_i$ in a smooth basis (for example, Laplacian eigenfunctions of the template structure) and split it into two parts. The first is the projection onto the constant function, which to first order is log-volume. The second is the remaining coefficients $z_i \in \mathbb{R}^K$, constrained to integrate to zero over $V_s$. Given variant dosage $g_i$, we fit an estimated effect map $\hat\beta(v)$ and define a cancellation index

$$ C = \frac{\left|\sum_{v \in V_s} \hat\beta(v)\right|}{\sum_{v \in V_s} |\hat\beta(v)|}, $$

which is near 1 for a uniform effect and near 0 for an effect whose positive and negative parts balance.

A locus is called volume-silent if it meets two conditions. The first is a joint multivariate test of $g$ against $z$ (with $K$ around 20 to 50 per structure, matching TransferGWAS's dimensionality) that is genome-wide significant after correcting for the number of structures. The second is that its marginal volume p-value exceeds 0.05. Leakage between the two parts comes from nonlinearity and residual registration error. We will estimate it by simulating injected uniform effects, to confirm that pure volume effects do not leak into $z$.

Genomic interpretation. Each locus's credible-set variants are scored for disruption of transcription-factor binding sites and splice sites with DeepVRegulome. This tool is DNABERT fine-tuned on more than 700 regulatory datasets, with reported 94–96% splice-site and over 85% transcription-factor model accuracy. Its 300-bp window means it cannot see distal looping. To link distal regulatory variants to target genes we therefore use HUGin Hi-C contact weights, in the manner of the Hi-C informed kernel association test. That test turns chromatin-contact confidence into weights for borrowing strength across physically looped loci, and recovered an RCL1 platelet association at $p = 1.36 \times 10^{-7}$ that linear tests missed. Here it supplies the gene-linking method; we do not reuse it as the association test.

Cellular heterogeneity: analogy and caveat. Cellular heterogeneity enters in two limited ways. First, it is the mathematical analogy behind the method: volume is a "bulk" sum over voxels in the same way bulk RNA-seq is a sum over cells. The volume-null decomposition is a deconvolution in this sense, recovering signed components that the aggregate erases, which is the problem Count Bridges and POPPY address for counts. Second, it limits interpretation. HUGin Hi-C and GTEx-based expression are bulk-tissue measurements, so a regulatory element active only in one cell type may be linked to the wrong gene or diluted.

We test one consequence that the available data can support. Using the 10-region genetically regulated expression from the regional TWAS, we ask whether genes near volume-silent loci show more heterogeneity in predicted expression effects across regions, including sign flips, than genes near volume-detected loci. Single-cell resolution of these effects is out of scope. We list it as the main follow-up rather than claim it.

Baselines

Naive reference: scalar-volume GWAS. We run a GWAS of FreeSurfer-style total structure volumes on the identical UK Biobank subjects, covariates and quality control. This defines the "volume-detected" set. The proposal needs volume-silent loci that replicate. A count at or below the null expected from leakage simulations would mean failure.

Finer scalar volumes. We repeat the GWAS on the finest available subfield and nucleus volumes, such as hippocampal subfields and thalamic nuclei. This is the cheapest fix for cancellation, because parcellating finely enough breaks opposing effects into separately visible pieces. It runs on the same subjects, with multiple-testing correction scaled to its larger number of phenotypes. The proposal beats it if a substantial fraction of replicated volume-silent loci (we pre-register "more than half") remains non-significant at every subfield.

Strongest existing method: TransferGWAS. TransferGWAS reports 289 independent loci, 14 of them in regions with no previously reported association, on 36,311 T1 scans. We re-run its pipeline (3D VAE and slice-wise ResNet50 features, 20 principal components) on our exact subject set, matching the total phenotype dimensionality. Beating it on locus count alone is not the goal, since the deep features may find more loci. The proposal beats it on its stated limitation if two things hold. Our volume-null loci should include replicated loci that TransferGWAS misses on the same data. And every locus we report should come with an anatomical effect map and cancellation index, which TransferGWAS's features cannot supply. Where TransferGWAS finds loci that we do not, we report them as a ceiling on what our basis captures.

Ablation: the 3D phenotype itself. We compute the same volume-null decomposition from features built slice by slice, by stacking 2D in-plane Jacobians without through-plane registration. If volume-silent discovery survives unchanged, native 3D context is not needed for this question, and we will say so.

Morphometry quality. As a check that our deformation features carry shape information at least comparable to established surface statistics, we compare hippocampal amyloid classification against PASCS-MP. It reports ACC 0.89 in MCI and 0.79 in cognitively unimpaired subjects in ADNI, and 0.81 in OASIS cognitively unimpaired subjects. The comparison uses the same subject splits.

Why not just…?

Why not just use finer subfield volumes? This is the most serious alternative, and the decisive experiment is built around it. If cancellation happens only at boundaries that existing atlases already draw, subfield GWAS will recover the volume-silent loci, and the right conclusion is "parcellate more finely." Zhu et al. and Wan et al. found Alzheimer's and APOE effects concentrated in sub-portions of structures that do not follow atlas borders, such as the hippocampal head. That is suggestive, but these are disease and risk-factor effects, not genome-wide evidence. Our pre-registered threshold settles the question.

Why not just take TransferGWAS's deep features? They already escape volume and find many loci. But they cannot say whether a locus acts by cancellation, and the enrichment hypothesis needs that distinction. TransferGWAS's own stated limitations are uninterpretable latent dimensions and loci without functional mapping. Our direct overlap analysis on identical data will show whether its loci split cleanly into high-$C$ and low-$C$ groups when projected onto our basis.

Why not just stay with coding variants, or test enrichment with off-the-shelf annotations? Most common GWAS hits are non-coding anyway, so a naive enrichment would be trivially "true." The comparison is therefore internal: volume-silent loci against volume-detected loci from the same scan, matched on minor allele frequency, linkage-disequilibrium score and distance to the nearest gene. The coding-centric view predicts no difference. Annotation-only approaches without DeepVRegulome scores are a cheaper version, and we will report them alongside. If they give the same answer, the deep regulatory model is unnecessary for this claim.

Why not just run voxel-wise mass-univariate GWAS? It would find shape effects but faces a huge multiple-testing burden. It also does not separate volume-changing from volume-null signal, so it cannot define the class of loci we are claiming exists. The decomposition is what makes the claim testable.

Decisive Experiments

All discovery uses UK Biobank T1 scans split into discovery and replication halves. The regional TWAS reports 45,549 participants with imaging and the TransferGWAS cohort is 36,311, so each half has roughly 18,000 or more. ADNI and OASIS are used for method development and the morphometry sanity check, not for genetic replication, which they are too small to power.

The first check is existence. We count volume-silent loci that replicate in the held-out half, against the leakage-calibrated null, and report how their cancellation indices are distributed. If the proposal is right, they should cluster near zero instead of spreading across the range.

The second check is the subfield test described in the Baselines. If more than half of the replicated loci survive it, the phenomenon is real. If they do not, the honest outcome is a negative result favoring finer parcellation.

The third check is the central scientific test. We compare the fraction of credible-set variants with high DeepVRegulome disruption scores, and the fraction of exome-captured variants, between matched volume-silent and volume-detected loci. Predicted direction: volume-silent loci have more regulatory disruption and less coding tagging. No difference, or the reverse, falsifies the shared-blind-spot hypothesis even if volume-silent loci exist.

The fourth check is a consistency test against bulk-tissue expression. Genes linked to volume-silent loci through proximity or Hi-C contacts should show higher cross-region heterogeneity in predicted expression effects in the 10-region data. A null result is ambiguous: given the bulk-tissue caveat, it would argue for cell-resolved follow-up rather than refute the main claim.

The fifth check is the 3D-versus-slice ablation. It tells us whether native volumetric context matters for finding genetic cancellation or only for segmentation.