Photographs of the back of the eye show tiny blood vessels that resemble the small vessels of the brain. Features of those vessels, such as how narrow the arteries are or how much they twist, predict silent strokes and small-vessel brain damage in people. Preclinical drug testing for the same diseases is done mostly in rats and mice, and nearly every therapy that works there fails in human trials. This proposal asks a question neither community has posed: for each retinal vessel feature that tracks brain disease in humans, which cell type in the vessel wall produces it? And does that cell type behave the same way in a rodent? A feature is only worth measuring in a mouse if the cell lineage that generates it in humans is conserved in mice.
We will give each human retinal-cerebral biomarker a cell-of-origin attribution. We do this by linking the feature's genetic associations to cell-type-specific gene programs in human retinal and brain vascular cells. We then score how well those programs are conserved across species, using alignment methods that do not depend on strict one-to-one gene matches. The output is a per-feature translatability score. We test it against a prospectively collected rodent imaging study and against simpler explanations, including plain image-level similarity and plain gene-level conservation. If the score works, preclinical neurovascular studies get a principled way to choose which eye-based readouts to trust. Some features may be anatomically measurable in rats but biologically unrelated to the human signal, and the score is meant to flag those.
Core Claim
The claim is that a retinal vessel feature's value as a cross-species proxy is set by the lineage that produces it, more than by whether the feature can be measured geometrically in both species. Arterial narrowing might originate mainly in arteriolar smooth muscle. Capillary rarefaction or branching complexity might originate in pericytes, the contractile cells wrapped around capillaries, or in endothelium. Those lineages can diverge between species in their gene programs even when the cell type is nominally present.
The prediction we can falsify is this. Features whose attributed lineages have conserved programs in rodents will show the same direction and relative size of change in rodents and humans under a shared perturbation, namely aging. Features whose lineages have drifted will not.
Why This Gap Is Real
The human proxy literature motivates the work but does not address mechanism. Retinal Biomarkers Discovery for Cerebral Small Vessel Disease in an Older Population extracted 302 morphometric features with VAMPIRE in the Lothian Birth Cohort 1936. It found that narrower arterioles (CRAE, the summary arteriolar caliber) were selected in 99.6% of bootstrap trials for predicting enlarged perivascular spaces, and venular bifurcation coefficients in 96.8%, independently of vascular risk factors. It says nothing about which cells change.
The deep models are stronger predictors but even less interpretable. A deep learning system for detecting silent brain infarction and predicting stroke risk detects silent infarcts at an area under the ROC curve (AUC) of 0.797, against 0.633 for clinical traits. A foundation model for generalizable disease detection from retinal images (RETFound) transfers to systemic outcomes. Neither tells us whether its signal is endothelial, mural or parenchymal. Kim et al. found that age and sex remain predictable even after vessels are digitally erased, so pixel-level models may not be reading vascular biology at all. We therefore build on explicit morphometric features rather than on foundation-model embeddings.
The genetic bridge we need already exists in outline. Lu et al. showed by Mendelian randomization that shared variants causally link retinal caliber with cerebrovascular phenotypes. Cheung et al. tie caliber changes to endothelial inflammation. What is missing is the step from loci to cell lineage.
The cross-species literature constrains the design. Treatment of acute cerebral ischemia using animal models: a meta-analysis documents 145 rodent stroke studies, 89% in rats, with heterogeneity above 80% and strong publication bias. Inhibition of AMD-Like Pathology With a Neurotrophic Compound in Aged Rats and 3xTg-AD Mice concedes that rodents have no macula, so "central retina" is an imperfect stand-in. Raw geometric correspondence is therefore already compromised.
The cell-of-origin literature supplies the logic that a lineage can be human-specific or human-amplified. CLIP cells traced tuberous sclerosis lesions to a progenitor pool largely absent in mice. Immunoregulatory Roles of Tumor‐Originated Pericytes Identified by Single‐Cell Analysis in Glioblastoma showed that brain pericytes can have a dual origin, either host-derived (higher PDGFRB) or tumor-derived (higher Desmin and CD44). Mural identity is therefore not one fixed category. We borrow the "docking" approach that cell-of-origin studies use, where aberrant cells are projected onto healthy developmental trajectories, as in Developmental basis of SHH medulloblastoma heterogeneity. Here we dock feature-associated gene programs onto vascular mural and endothelial maturation trajectories rather than tumor cells.
The method side supplies alignment tools. Towards Universal Cell Embeddings: Integrating Single-cell RNA-seq Datasets across Species with SATURN groups genes into "macrogenes" by protein-language-model similarity, avoiding one-to-one orthologs. It showed that the glaucoma gene MYOC maps to different functional macrogenes in human than in mouse or pig. CACIMAR separates conservation of marker genes, regulatory networks and cell-cell signaling, and found these can disagree for the same cell type. scRetinaDB compiles 3.6 million retinal cells across 17 species but links none of them to imaging.
None of these works connects an imaging feature to a lineage, or a lineage to a translation decision.
Research Program
The pipeline has three stages, and each is useful on its own.
Stage one is standardized features. We compute interpretable vessel features with AutoMorph: fractal dimension, vessel density, CRAE, CRVE, artery-to-vein ratio and tortuosity, by zone. We use UK Biobank fundus images, which come paired with genotypes, and keep AutoMorph's quality gate. That gate's confidence analysis cut falsely gradable images by 76%, and its feature intraclass correlation exceeds 0.9 for fractal dimension, density and width. We run a genome-wide association study (GWAS) per feature.
Stage two is cell-of-origin attribution. For each feature we estimate how much of its heritability concentrates in genes specifically expressed by each candidate cell type. We use a standard partitioned-heritability analysis on cell-type specificity scores. The candidate cell types are retinal endothelium, pericytes, arteriolar smooth muscle, Müller glia, astrocytes and microglia from the human portion of scRetinaDB, plus normal brain pericytes and endothelium from the normal-brain reference used in the glioblastoma pericyte study. Mural cells are sparse in dissociated retina, so we pool across the retinal and brain references. We report attribution only where a cell type has enough cells for a stable specificity estimate. To capture maturation state, not just cell type, we split mural cells along a pseudotime trajectory into immature and mature segments and attribute to segments.
Stage three is conservation and scoring. For each attributed cell type or segment we measure conservation between human and rat, and between human and mouse. We use SATURN macrogene embeddings fed into CACIMAR's three separate scores: markers, regulatory network and signaling. The aqueous-humor outflow atlas in SATURN, which spans human, macaque, mouse and pig vascular-adjacent cells, gives a second source of vascular cells. The translatability score for feature $f$ is
$$T_f = \sum_{c} w_{f,c}\, C_c ,$$
where $w_{f,c}$ is the fraction of feature $f$'s cell-type-attributed heritability that falls in cell type $c$ (summing to one over cell types), and $C_c \in [0,1]$ is the cross-species conservation of cell type $c$'s program. We fix the way CACIMAR's three layers are combined before seeing any rodent data. We also report each layer separately, so the result tells us which layer of conservation matters.
Baselines
The naive reference point assumes every feature is equally translatable, or equivalently ranks features at random. Under that null, the expected rank correlation between score and observed cross-species concordance is zero. Every other method must beat it with a bootstrap confidence interval over features that excludes zero.
The image-level baseline is what a preclinical imaging lab would do by default. We run the same AutoMorph features on human and rodent fundus images and rank features by how similar their between-subject distribution and age slope look once each species is standardized. This baseline is fit on a held-out half of the rodent animals, and the outcome is evaluated on the other half, so it cannot reuse the outcome data. It is the strongest competitor because it uses the most direct evidence there is: the feature itself in the target species.
The gene-level baseline drops the cell types. It scores each feature by the fraction of its top GWAS genes that have one-to-one rat or mouse orthologs, weighted by association strength. This isolates whether cell lineage adds anything beyond gene presence.
The forced-integration baseline replaces SATURN and CACIMAR with Harmony integration, as used in scRetinaDB. It scores conservation as how much human and rodent cells share each integrated cluster. CACIMAR argues that this approach hides real divergence, and this baseline tests that claim for vascular cells.
There are two ablations. The first swaps SATURN macrogenes for one-to-one orthologs inside our full pipeline. The second scores conservation from each CACIMAR layer alone.
Every method receives the same features, the same rodent cohort and the same fixed preregistered specification, with no tuning on outcome animals. For the human proxy itself, the strongest reported numbers are DeepRETStroke's silent-infarct AUC of 0.797 against 0.633 for clinical traits (paper) and the VAMPIRE selection frequencies above. We use these to choose which features matter, not as targets to beat.
For the actual task of predicting cross-species concordance of retinal vascular features, no reported result found on Lacuna. The nearest alignment figure is SATURN's 85.8% median zebrafish-to-frog label-transfer accuracy, which improves on SAMap. That is a different task, and we cite it only to justify the alignment choice. Our method beats the field if its rank correlation with observed concordance exceeds the image-level baseline's, with the paired-bootstrap difference excluding zero.
Why not just…?
Why not just measure the features in rodent fundus images and keep the ones that look similar? This is the real threat, and we treat it as the primary comparison. The P021 study's admission that rodents lack a macula means geometric similarity can be spurious. A tortuosity index can behave alike in both species for anatomical reasons unrelated to brain vessels. Even so, if the image-level baseline matches our score on the rodent aging outcome, the lineage layer adds cost without value. We would report that, and recast the contribution as a mechanistic annotation of human features rather than a translation tool. The two can also be combined. If lineage information adds predictive value beyond image similarity in a joint model, that supports keeping it even when its standalone correlation is not higher.
Why not just check whether the feature's genes have rodent orthologs? The SATURN MYOC result shows that an ortholog can be present while its function has drifted. CACIMAR found nearly 2,500 species-specific markers against 139 conserved ones in mouse and zebrafish retina. Gene presence is therefore a weak proxy for functional equivalence. The gene-level baseline and the ortholog ablation measure this directly. If they tie with the full method, the macrogene machinery is unnecessary.
Why not just ask RETFound or DeepRETStroke which regions matter, and skip genetics? Saliency on a human-trained model tells us where the model looks, not which cell type changes. Kim et al. show that much of the predictive signal survives vessel removal, so saliency could point at parenchyma for reasons unrelated to vascular lineage. As a secondary analysis, we project RETFound's silent-infarct-predictive direction onto AutoMorph features. That tells us which interpretable features the foundation model's signal resembles, but we do not treat it as a cell attribution.
Proposed Experiments
The decisive experiment is a preregistered rodent aging study. We take fundus photographs of young and aged rats, matching the aged Fisher strain used in the P021 work, and of young and aged mice, at sample sizes set by a pilot. AutoMorph segmentation is fine-tuned on a small manually annotated rodent set. Segmentation quality is checked against those annotations before any feature is used, and features failing an intraclass-correlation threshold are excluded blind to their scores.
For each feature we compute concordance. Concordance is the sign agreement and standardized-slope ratio between the rodent age effect and the human age effect in UK Biobank, using age-stratified human estimates and the human retinal aging work (multimodal retinal aging clock) as a sanity check on direction. The primary endpoint is the rank correlation between $T_f$ and concordance across features, compared against each baseline by paired bootstrap.
With roughly thirty usable features, power is modest. We therefore preregister directional hypotheses for the clearest cases. We expect arteriolar caliber, likely smooth-muscle-attributed, to transfer well. We expect capillary-density features, likely attributed to pericytes, to transfer worse. Rat and mouse serve as two semi-independent replications.
Two human-only checks come first and could falsify the premise cheaply. First, attribution must be specific. Caliber and capillary features should load on different lineages, and a shuffled-specificity control should erase the enrichment. Second, the features whose attributed lineages are shared between retinal and brain vascular references should be the ones LBC1936 selected most often for small-vessel disease. That would support the idea that the eye-brain proxy runs through shared lineages.
A secondary check uses the dual-origin finding. If mature, host-type pericyte programs (PDGFRB-high) carry the capillary-feature attribution while immature or CD44-high states do not, the cell-of-origin framing earns its place beyond simple cell-type labels.
If attribution is non-specific in humans, or if the image-level baseline wins in both rodent species, the central claim fails, and we would say so.
References
- Retinal Biomarkers Discovery for Cerebral Small Vessel Disease in an Older Population
- A deep learning system for detecting silent brain infarction and predicting stroke risk
- A foundation model for generalizable disease detection from retinal images
- Effects of Hypertension, Diabetes, and Smoking on Age and Sex Prediction from Retinal Fundus Images
- Phenotypic screening and genetic insights for predicting major vascular-related diseases using retinal imaging
- Influence of secondhand smoke exposure on the retinal vasculature of children in Hong Kong
- Treatment of acute cerebral ischemia using animal models: a meta-analysis
- Inhibition of AMD-Like Pathology With a Neurotrophic Compound in Aged Rats and 3xTg-AD Mice
- Amplification of human interneuron progenitors promotes brain tumors and neurological defects
- Immunoregulatory Roles of Tumor‐Originated Pericytes Identified by Single‐Cell Analysis in Glioblastoma
- Developmental basis of SHH medulloblastoma heterogeneity
- Towards Universal Cell Embeddings: Integrating Single-cell RNA-seq Datasets across Species with SATURN
- CACIMAR: cross-species analysis of cell identities, markers, regulations, and interactions using single-cell RNA sequencing data
- scRetinaDB: A Comprehensive Database of Single-Cell and Spatial Omics from Cross-Species Retinas
- AutoMorph: Automated Retinal Vascular Morphology Quantification Via a Deep Learning Pipeline
- A multimodal retinal aging clock for biological age prediction and systemic health assessment via OCT and fundus imaging