Do Brain-Specific Gene Aging Programs Predict Where the Connectome Ages?

Brain networks change with age, but not uniformly. Some regions lose functional coupling faster than others, and some structural pathways degrade earlier. Separately, post-mortem gene expression atlases show that aging is tissue-specific: the genes that switch on or off with age in the liver are not the ones that change in the brain cortex or the caudate. Nobody has asked whether these two facts line up. This proposal tests one question: does the age-related change in gene expression measured in brain tissue predict the spatial pattern of connectome aging in living people? It also asks whether that prediction is specifically neural, or whether any tissue's aging signature, from blood, muscle, or aorta, predicts it equally well.

The obstacle has been that the Human Connectome Project (HCP), the largest high-quality brain imaging resource, has no molecular data on its participants. The proposal works around this without pretending to have matched subjects. It builds a predicted regional "aging load" map by combining two population-level resources. The first is the regional distribution of each gene's expression across the brain, from the Allen Human Brain Atlas (AHBA). The second is each gene's age slope, from the Genotype-Tissue Expression project (GTEx; The Genotype-Tissue Expression (GTEx) project), which sequences RNA from more than 40 tissues of deceased donors of varying ages. The predicted map is then compared with the observed map of per-region connectome change across age in HCP-Aging. Because GTEx provides age slopes for every tissue, the comparison has a built-in control group: about 40 non-brain tissues whose aging signatures should not predict brain network aging if the link is tissue-specific.

Core Claim

We claim that the regional pattern of connectome aging in HCP-Aging is partly predicted by brain-tissue transcriptomic aging, and that this prediction goes beyond two confounds. The first confound is the brain's dominant spatial expression gradient. The second is generic, organism-wide aging genes. Concretely, let $E_{g,r}$ be the AHBA expression of gene $g$ in brain region $r$, standardized across regions, and let $\beta^{(t)}_g$ be the GTEx age slope of gene $g$ in tissue $t$. The predicted aging load of region $r$ under tissue $t$ is

$$ L^{(t)}_r = \sum_g \beta^{(t)}_g \, E_{g,r}, $$

which weights each region by how strongly it expresses the genes that rise or fall with age in tissue $t$. The observed map $A_r$ is the per-region age slope of a connectome property, such as functional connectivity strength (the summed correlation of a region's fMRI signal with all others) or structure–function coupling, estimated across HCP-Aging participants. The claim is that $\mathrm{corr}(L^{(t)}, A)$ is significant only when $t$ is a brain tissue, that it ranks brain tissues near the top of all GTEx tissues, and that it survives after removing the principal AHBA gradient from both maps.

The closest prior work is A structural equation model for imaging genetics using spatial transcriptomics. That paper uses AHBA to group 112 regions into latent variables that link SNPs to regional volume and thickness in ADNI. It uses brain expression only as a clustering prior. It involves no age information from any tissue, no functional or structural connectomes, and no cross-tissue contrast. Our proposal differs on all three counts, and it needs no subject-level genotypes, which HCP-Aging does not supply in the material we have.

Why This Pairing, and What the Literature Supplies

Several papers motivate the tissue-contrast design. An analysis of aging-related genes derived from the Genotype-Tissue Expression project (GTEx) regressed GTEx expression on donor age across more than 40 tissues. It found 1,573 age-associated genes that split into two classes. Downregulated genes are conserved intracellular housekeeping hubs. Upregulated genes are extracellular, immune, and signaling genes. That paper never touched brain imaging, and its split gives us a natural mechanistic sub-question: does connectome aging track the loss of metabolic housekeeping, or the rise of inflammatory signaling? Transcriptomic landscape profiling of metformin‐treated healthy mice profiled ten mouse tissues and showed that a single systemic drug produces protective programs in liver and harmful ones in aorta. This is direct evidence that one tissue's signature cannot stand in for another's, which is the premise our negative-control tissues test. That study treated brain as an undifferentiated homogenate. We do the opposite and resolve the brain into hundreds of regions.

On the imaging side, three papers constrain the design rather than supplying the core idea. Krakencoder fuses structural and functional connectomes into a 128-dimensional latent space and predicts age with Pearson r ≥ 0.5. This confirms that HCP connectomes carry strong age signal, but its latent space contains no molecular information. Defining and Benchmarking a Data-Centric Design Space for Brain Graph Construction showed that rank-based edge definitions beat Pearson correlation. We therefore build regional maps with both Spearman and Pearson and require the result to hold under both. A simple but tough-to-beat baseline for fMRI time-series classification found that time-order-invariant models match complex temporal ones. This justifies starting with static connectivity summaries rather than dynamic ones. Finally, Deconvolving Complex Neuronal Networks into Interpretable Task-Specific Connectomes deliberately omitted subcortex. GTEx's brain samples are mostly subcortical (caudate, putamen, accumbens, hippocampus, amygdala, hypothalamus, cerebellum, substantia nigra), so our design lets subcortical structures enter a connectome–molecular analysis directly.

Research Program

The first component is the observed aging maps. We will process HCP-Aging resting-state fMRI and diffusion MRI into Glasser 360-region cortical connectomes plus HCP subcortical structures. For each region we compute functional connectivity strength, structural connectivity strength, and structure–function coupling (the correlation between a region's structural and functional connection profiles). We regress each on age with sex, head motion, and scanner covariates. Motion rises with age and inflates or deflates connectivity, so we also rerun the analysis on a motion-matched subsample. We estimate the reliability of each aging map by split-half resampling of participants. Only maps with high split-half correlation go forward.

The second component is the predicted maps. We compute GTEx age slopes per gene for every tissue, using identical models and covariates. This gives each tissue the same statistical treatment, so any brain-specific advantage is not an artifact of better modeling. We map AHBA samples to Glasser regions and form $L^{(t)}$ for every tissue. Subcortex needs a separate treatment. There, GTEx gives region-specific slopes directly, so we can additionally form a "matched" prediction: caudate slopes for caudate, hippocampus slopes for hippocampus. We then ask whether matched subcortical regions beat mismatched assignments.

The third component is the statistics. With spatially smooth maps, naive correlations are badly inflated. All cortical tests will therefore use spin permutations, which randomly rotate the cortical sphere to preserve spatial autocorrelation. We will also use gene-set permutations, which draw random gene sets of equal size and equal mean expression. The subcortical test has only about nine matched regions, so we will treat it as confirmatory and use exact permutation over region assignments.

Baselines

The naive reference point is the observed aging map correlated with itself after spin rotation. This gives the spin-null distribution that any predicted map must exceed. A predicted map that sits inside this distribution is uninformative.

The strongest competitor is the principal AHBA expression gradient: the first principal component of $E_{g,r}$ across genes, with no age information at all. Cortical properties of many kinds follow this sensorimotor-to-association axis. The connectome aging map may simply follow it too, and then any gene-weighted map would "work." To keep the comparison fair, it uses the same AHBA preprocessing, the same regions, and the same spin tests. We count the proposal as beating it only if $L^{\text{brain}}$ explains significant variance in $A$ after both maps are residualized on this component. No reported result was found on Lacuna for spatial correlation between AHBA gradients and HCP-Aging connectome age slopes, so this number must be measured here.

The key ablation isolates the age information. It keeps $E_{g,r}$ fixed and shuffles the brain-cortex age slopes $\beta_g$ across genes. This preserves which genes are expressed where, but destroys which ones change with age. If shuffled-slope maps correlate with $A$ as well as the real map, the claim fails.

The tissue baseline replaces brain slopes with slopes from each of the roughly 40 non-brain GTEx tissues and with a pooled all-tissue slope. Brain tissues must rank in the top tail of this distribution. As an external reference for how much age signal the imaging carries, Krakencoder reports age prediction of Pearson r ≥ 0.5 from its fused latent space (Krakencoder). We will report the same predictive r for a simple regression on our regional features to confirm that our preprocessing retains comparable age signal.

Why not just…?

Why not just correlate individual AHBA genes with the connectome aging map, as standard imaging transcriptomics does? That approach finds genes whose spatial pattern resembles the map. It cannot say whether those genes change with age, and with about 15,000 genes and smooth maps it produces many false hits. Our shuffled-slope ablation directly measures what the age information adds beyond spatial resemblance. If that ablation matches the full model, the simpler approach is sufficient and we will say so.

Why not just use the principal expression gradient, since so much follows it? This is the real threat. The residualized test in the Baselines section settles it. The UAG/DAG split from the GTEx aging-gene analysis adds a sharper check. If up- and down-regulated gene sets yield maps of opposite sign relative to $A$, that structure cannot come from a single gradient.

Why not just train a multimodal model that adds GTEx features into a connectome embedding like Krakencoder? GTEx donors and HCP participants are different people. Subject-level fusion would learn nothing real, because there is no pairing to learn from. The only valid link is at the level of regions and genes. Graph Representation Learning on Tissue-Specific Multi-Omics also cautions against extra complexity. There, adding methylation to RNA moved link prediction only from 71.4% to 71.6% balanced accuracy. A linear projection of known slopes is the appropriate model until it demonstrably fails.

Proposed Experiments

The primary experiment computes $\mathrm{corr}(L^{(t)}, A)$ for each tissue and each connectome property, with spin and gene-set nulls. The decisive result is the rank of brain-cortex tissue among all GTEx tissues, after residualizing on the principal gradient. There are three possible outcomes. Top rank with a significant residual correlation supports tissue-specific molecular aging as a correlate of network aging. Equal performance from peripheral tissues such as whole blood would indicate a shared systemic aging program. That would be a different, still reportable finding about body–brain coupling, and we would not claim neural specificity. Failure of the shuffled-slope ablation to degrade performance falsifies the core claim.

The secondary experiments are as follows. We run the matched-versus-mismatched subcortical test using GTEx's regional brain samples. We separately project upregulated immune-signaling genes and downregulated housekeeping genes, to see which class tracks functional versus structural aging. We repeat the full pipeline with Spearman and Pearson connectivity, following the data-centric brain graph benchmark and its code (GeQinwen/DataCentricBrainGraphs). We repeat it again on the motion-matched subsample. A result that survives the gradient residualization, the slope shuffle, the tissue ranking, and both robustness reruns is the evidence this proposal is designed to produce.