A Human-Anchored Retinal Readout for Rodent Stroke Trials

Drugs that protect the brain in rodent stroke experiments almost never work in human patients. One reason is that the animal and human studies measure different things. Rodent studies weigh dead tissue: they stain the brain after sacrifice to measure infarct size, which is the volume of tissue killed by the stroke. Human studies score function and recurrence months later. Nothing measured in the animal sits on a scale that is known to predict human outcomes. Over the last few years the retina, the light-sensitive tissue at the back of the eye, has become that kind of scale in humans. The retina is wired into the brain's blood supply and develops as part of the central nervous system, and photographs of its vessels now predict silent brain infarcts, incident stroke and recurrent stroke better than clinical risk factors do. The retina is also imageable in rodents, repeatedly and without killing the animal.

This note proposes to build one retinal vascular readout that is learned from human stroke outcomes and then applied unchanged to rodent stroke experiments. The readout has two uses. The first is to rank which animal stroke models produce retinal vascular changes that most resemble those of human stroke-prone patients. The second is to re-express each candidate drug's effect in the animal as a shift along a risk axis that is known to matter in humans, instead of as a reduction in post-mortem infarct volume. The central bet is that this shared, human-calibrated readout will shrink the inflated preclinical effect sizes that meta-analyses have documented. It should also keep the effect of interventions that do work in humans, such as timely reperfusion.

Core Claim

The claim is that a retinal vascular score with three properties can serve as a translational endpoint that post-mortem infarct volume cannot. It must be trained only on human outcome data. It must be built from species-comparable vessel geometry. It must be measured longitudinally in living animals. The evidence motivating this comes from both directions.

On the human side, Jiang et al. built DeepRETStroke, a foundation model for fundus photographs (2D colour photographs of the retina). It reached an AUC of 0.797 for detecting MRI-confirmed silent infarcts against 0.633 for clinical traits, and 0.901 for incident stroke. Lu et al. used Mendelian randomization, which treats genetic variants as natural experiments, and found that retinal microvascular traits causally precede stroke and myocardial infarction rather than merely correlating with them. This causal direction is what licenses using a retinal axis as more than a convenient correlate.

On the preclinical side, Wang et al. pooled 145 rodent ischemia studies. They found very large standardized effects for statins (−27.61 on infarct size), fingolimod (−18.34) and insulin (−11.20 on neurobehaviour). They also found heterogeneity above 80%, significant publication bias for infarct size (p < 0.001), and routine absence of blinding and allocation concealment.

Nobody has put these two literatures on the same measurement. The closest methodological precedent is SystemMatch. It embeds in vitro macrophage systems and human tumour macrophages from single-cell RNA sequencing into one latent space, ranks preclinical systems by Earth Mover's Distance to the human target, and found that the best of 18 in vitro models was only 20% closer than the worst. We reuse its logic of ranking models by distance to a human target distribution. We replace transcriptomes with retinal vessel networks and replace cell culture with whole living animals, where a stroke drug's effect actually unfolds.

Research Program

The work has three linked components. Each can fail independently, which is what makes the program testable.

Human-anchored encoder. We will build the human-anchored encoder on vessel graphs, following Huang et al.. That work converts a vessel segmentation into a graph, with bifurcations and crossings as nodes and vessel segments as edges, and encodes the graph with a graph attention network. On UK Biobank it reached 71.92% AUROC for stroke with 3.4M parameters, against 71.54% for a 28.2M-parameter ResNet50 on raw pixels. We depart from Huang et al. in one deliberate way. Edge and node features are restricted to quantities that are dimensionless or can be normalized across species: arteriole-to-venule caliber ratio, caliber normalized to the vessel's own order in the tree, branching angles, tortuosity, and fractal dimension. Absolute microns, image texture and optic-disc appearance are excluded, because a rat retina differs from a human one on all of them for reasons unrelated to disease.

The encoder is trained on UK Biobank alone to predict incident stroke. The scalar it outputs, which we call the human risk axis, is then frozen. Segmentation uses existing tools: AutoMorph and SIVA-DLS as described by Ullah et al. and Cheung et al., with SGL vessel segmentation adapted to rodent fundus images through a small hand-labelled set.

Cross-species bridge. Rodent fundus images will be segmented, converted to graphs with the same feature definitions, and passed through the frozen encoder. A frozen encoder will not be enough if rodent graphs fall outside the human training distribution. We therefore add a light alignment step, fitted only on sham-operated and pre-stroke animals, that matches the healthy rodent graph distribution to healthy human graphs. This step is in the spirit of the heterogeneity-aware alignment in THERAPI, which moved preclinical cell-line response to patient tumours (0.775 AUROC on TCGA against 0.669 without adaptation). Because the alignment never sees a stroked animal, any stroke- or drug-induced shift along the risk axis is not something the alignment was fitted to produce.

Animal programme. The third component is an imaging protocol in rats. Fundus photography and OCT (optical coherence tomography, which gives micrometre-resolution cross-sections of retinal layers) will be acquired at baseline, 24 hours, 3, 7 and 28 days. These are paired with MRI infarct volume, terminal staining, and standard neurobehavioural scores.

One confound needs naming up front. In intraluminal filament occlusion, the eye's own arterial supply branches close to the occlusion site, so ipsilateral retinal changes may reflect direct retinal ischemia rather than a brain proxy. We treat this as something to measure, not assume away. Both eyes are imaged, and embolic models are included because they spare the ocular supply differently. A contralateral-eye analysis is reported separately.

Baselines

Status-quo rodent endpoints. The naive reference point is infarct volume and neurobehavioural score, which are the endpoints behind every number in Wang et al.. They are run on exactly the same animals, randomized and blinded identically. The proposal beats them if the retinal axis produces drug effect sizes with lower between-lab heterogeneity and smaller inflation relative to known human outcomes, while still separating the reperfusion positive control from sham.

Classical caliber measurements. The strongest simple competitor is central retinal artery and vein equivalents (CRAE and CRVE) measured with SIVA-DLS. Cheung et al. showed these are sensitive enough to detect sub-micron dilation from secondhand smoke in children. Ullah et al. report a hazard ratio of 1.28 for stroke with wider venules. We will fit a logistic model on the arteriole-to-venule ratio to the same UK Biobank stroke labels and carry it through the same cross-species pipeline with the same alignment step. If it matches the graph axis on the decisive checks below, the graph encoder is unnecessary and we will say so.

Strongest human-side model. For the human anchor itself, the bar is the 71.92% AUROC that Huang et al. report for graph-based stroke prediction on UK Biobank. Lu et al. report 0.737 for a related 10-year composite of stroke, MI and CKD. DeepRETStroke's 0.901 is on non-public cohorts with different labels and is not a fair comparator. Our species-comparable restriction discards information, so we do not expect to beat 71.92%. The acceptance criterion is staying within roughly two points of it on the same split. A larger drop would mean the comparable features lose too much human signal to serve as an anchor.

Pixel foundation model applied directly. We run RETFound Zhou et al., fine-tuned on UK Biobank stroke and applied to rodent images, with the same alignment budget.

Ablations. Two ablations isolate the key ingredients. The first removes human anchoring by training the same graph encoder self-supervised on rodent data and using its principal shift axis. This tests whether the human labels, rather than the graph representation, do the work. The second adds parenchymal texture back in, testing the claim of Kim et al. that the non-vascular background of the retina carries systemic signal. That claim is in direct tension with Huang et al.'s choice to discard pixels.

For rodent retinal readouts of any kind, there is no reported result found on Lacuna.

Why not just…?

Why not just keep measuring infarct volume, but run better-designed, blinded, multi-lab studies? Better design is necessary, and our drug pilot adopts it. It does not, however, solve the problem that infarct volume in a rodent brain sits on no human-calibrated scale. Tang et al. note that the smooth, white-matter-poor rodent brain responds differently from the folded human brain. Their review of Annexin A1 shows strong rodent plaque and infarct effects with no human efficacy trial to date. The experiment settles this directly. Both endpoints are run under identical rigorous design, and we ask which one's effect sizes sit closer to human results.

Why not just apply DeepRETStroke or RETFound to rodent fundus images zero-shot? Pixel-level models learn camera, pigment and disc appearance that differ wholesale between species. Ullah et al. and Zhou et al. already report sensitivity to camera vendor and modality shift within humans alone. This is nonetheless included as a baseline rather than dismissed. If it tracks rodent MRI infarct as well as the graph axis does, the simpler route wins.

Why not just use CRAE and CRVE? This is the real threat. Caliber is interpretable, transfers across species as a ratio, and has decades of human validation. Huang et al.'s gain over a pixel CNN was small (71.92% against 71.54%), which hints that the advantage of graph structure may be modest. We keep caliber as a full pipeline baseline. The graph encoder is justified only if it beats caliber on rodent infarct tracking or on between-lab heterogeneity.

Why not just image the rodent brain with MRI longitudinally? MRI is the better direct measure of rodent infarct, and we use it as the reference. The point of the retina is not resolution but a shared scale: the same measurement exists in millions of humans with outcome labels, and that does not hold for rodent-style MRI protocols.

Proposed Experiments

Step one: human anchor. We train the comparable-feature graph encoder on UK Biobank, using the split and cohort definition of Huang et al. with 82,569 participants. We report AUROC for incident stroke against their 71.92% and against the caliber model. As a secondary check, we apply the frozen axis to the Moorfields AlzEye linkage described by Wagner et al., with 11,735 stroke cases, if access permits.

Step two: model-fidelity ranking. Rats receive filament occlusion of 60 or 90 minutes, embolic occlusion, or sham. We use young and aged animals and normotensive and hypertensive strains. For each condition, we compute the Earth Mover's Distance between its day-7 graph distribution along the encoder's embedding and the distribution of UK Biobank participants who later had a stroke, following the SystemMatch procedure. We predict that aged, hypertensive, embolic conditions will rank closest, and that young filament models, which supply most of the 89%-rat literature in Wang et al., will rank furthest. A flat ranking, comparable to SystemMatch's 20% spread, would itself be an informative negative.

Step three: drug pilot. We test statins, fingolimod, insulin and the ANXA1 mimetic Ac2-26 against vehicle, which are the standouts in Wang et al. and Tang et al. Reperfusion serves as the positive control and sham as the null. The study runs across two labs with central randomization and blinded imaging analysis. Each drug's effect is reported on infarct volume, on neurobehaviour, and on the human risk axis, all as standardized effect sizes with between-lab heterogeneity.

Decisive check. The proposal succeeds if three things hold. The reperfusion control shifts the retinal axis clearly. The four non-translating agents show substantially smaller retinal effects than infarct effects, relative to that positive control. Between-lab heterogeneity on the retinal axis is lower than on infarct volume.

It fails if any of three things happens. The retinal axis does not move under reperfusion. The axis moves only in the ipsilateral eye of filament models, which would be the direct-ischemia confound. Caliber alone reproduces every result. Four drugs and one positive control cannot validate a surrogate endpoint. What they can show is whether this readout deserves the larger, prospective multi-drug study that a real validation would need.