Shear-Mapped Retinal Vessels as a Translatable Readout of Endothelial Mechanotransduction

The endothelial cells lining every blood vessel are mechanical sensors. Where blood drags slowly along the wall — a condition called low wall shear stress — those cells switch on a matrix-degrading program through a now well-mapped cascade: integrin adhesion molecules activate the enzyme Focal Adhesion Kinase (FAK), which triggers ERK and NF-κB signaling and turns on membrane type-1 matrix metalloproteinase (MT1-MMP), the cell's machinery for chewing through surrounding tissue (Membrane type 1 matrix metalloproteinase activation is enhanced by low shear stress through integrin pathway in ApoE-/- mice). At the molecular scale, FAK itself is a force-gated enzyme whose autoinhibitory interface unzips under tensions of 150–450 pN (Mechanism of Focal Adhesion Kinase Mechanosensing). The retina is the one place in the body where microvessels of brain origin can be photographed non-invasively, and deep learning systems already read those photographs to detect silent brain infarctions with an AUC of 0.797 and predict incident stroke at 0.901 (A deep learning system for detecting silent brain infarction and predicting stroke risk). But those systems see pixels, not physics: nothing in them knows that a narrowed, tortuous arteriole implies altered flow drag, and nothing in them can say whether the same drag acts on a mouse's endothelium as on a human's.

This proposal argues that the missing link is a shear map. If we can compute local wall shear stress from retinal vessel geometry alone, then a routine fundus photograph becomes a readout of the mechanotransduction state of the brain-adjacent endothelium — and, crucially, a shared physical currency for comparing that state across species whose hearts beat at 600 versus 70 beats per minute. The cross-species divergence literature tells us the stakes: 145 animal studies of acute cerebral ischemia report enormous effects (statins reduced infarct volume with a standardized mean difference of -27.61) that never reached the clinic, with inter-study heterogeneity above 80% (Treatment of acute cerebral ischemia using animal models: a meta-analysis), and rodent hearts, vessels, and retinas operate in hemodynamic regimes distant from human ones (A Porcine Model of Heart Failure With Preserved Ejection Fraction Induced by Chronic Pressure Overload Characterized by Cardiac Fibrosis and Remodeling, Inhibition of AMD-Like Pathology With a Neurotrophic Compound in Aged Rats and 3xTg-AD Mice). The open problem the file identifies is exact: nobody has linked non-invasive retinal microvascular geometry to hemodynamic shear stress profiles and endothelial mechanotransduction signaling states in any cohort, human or animal, and nobody has built an allometric framework that reconciles shear regimes across rodent, porcine, and human neurovasculature.

Core Claim

Retinal vessel geometry, interpreted through fluid mechanics rather than as static shape, is a quantitative readout of endothelial mechanotransduction state. Concretely: from a segmented retinal vascular tree we can estimate local wall shear stress using branching scaling laws and a small number of calibrated anchors, convert those estimates into predicted signaling states via the empirically measured low-shear versus physiological-shear response of the integrin–FAK–ERK–NF-κB–MT1-MMP axis, and then test two claims. First, shear-derived features predict cerebral small vessel disease and stroke outcomes in human cohorts at least as well as purely geometric features, and transfer across imaging domains and species better than either geometric or pixel-based features. Second, when the same mapping is applied across species, the degree of shear-regime mismatch between an animal model and human hemodynamics predicts when mechanotransduction findings fail to translate — turning "the mouse is not a small human" from a slogan into a computable, testable quantity.

Research Program

The first component is a shear-inversion pipeline. Starting from color fundus photographs or OCT angiography scans, we segment the vascular tree and separate arterioles from venules using the modular components of AutoMorph, whose vessel segmentation reaches an F1 of 0.78 on DR-HAGIS and 0.73 on AV-WIDE and whose artery/vein differentiation reaches 0.66 on IOSTAR-AV (AutoMorph: Automated Retinal Vascular Morphology Quantification Via a Deep Learning Pipeline). Flow cannot be measured everywhere at once, but it can be inferred: under Murray's branching law, volumetric flow $Q$ scales with the cube of vessel radius $r$, so that calibre measurements across a branching tree constrain relative flow, and local wall shear stress follows from

$$\tau = \frac{4\mu Q}{\pi r^3},$$

where $\mu$ is blood viscosity. Absolute calibration comes from three anchors in the literature: the murine carotid measurements of 5.61 ± 0.72 N/m² under constrictive low shear versus 9.76 ± 1.48 N/m² physiological shear (Membrane type 1 matrix metalloproteinase activation is enhanced by low shear stress through integrin pathway in ApoE-/- mice), the matched HUVEC flow-chamber values of 0.4 versus 1.2 N/cm², and the FAK force window of 150–450 pN (Mechanism of Focal Adhesion Kinase Mechanosensing), which we relate to wall shear through an explicit membrane-and-cytoskeletal force transmission model with stated assumptions. The output is a per-vessel, per-patient mechanotransduction state score — the fraction of the measured tree predicted to sit in the matrix-degrading low-shear regime.

The second component validates this score against human cerebrovascular disease. The natural cohorts are those with paired retinal imaging and brain MRI: the Lothian Birth Cohort 1936 analysis, where decreased central retinal arteriolar equivalent was selected in 99.6% of 500 bootstrap trials for enlarged perivascular spaces (Retinal Biomarkers Discovery for Cerebral Small Vessel Disease in an Older Population), the OCTA study tying radial peripapillary capillary density to deep white matter lesions (Fundus Changes Evaluated by OCTA in Patients With Cerebral Small Vessel Disease and Their Correlations: A Cross-Sectional Study), UK Biobank, and the AlzEye record-linkage cohort of 353,157 patients for incident outcomes (AlzEye: longitudinal record-level linkage of ophthalmic imaging and hospital admissions of 353 157 patients in London, UK). The key comparison is not shear features versus nothing, but shear features versus the same features without the physics: tortuosity, fractal dimension, and calibre entered as raw geometry, exactly as VAMPIRE-based elastic-net regression does today.

The third component is the cross-species layer, which is where the proposal becomes more than a biomarker study. Applying the identical pipeline to rodent and porcine retinal and cerebral vascular geometries, we ask a question the literature has never posed: given a mouse heart rate of 500–700 beats per minute and arteriolar radii an order of magnitude smaller than human ones, what fraction of the rodent retinal tree sits in the low-shear MT1-MMP-activating regime compared with the human tree? If the distributions barely overlap, then rodent mechanotransduction experiments are probing a different operating point of the same pathway, and the framework can be tested prospectively against the stroke meta-analysis corpus: we will re-analyze the 145-study database to ask whether interventions whose effective hemodynamic regime is most human-mismatched show the largest effect sizes and heterogeneity — inflated effects being exactly what a regime mismatch plus publication bias would produce (Treatment of acute cerebral ischemia using animal models: a meta-analysis). The porcine pressure-overload model, with its human-scale E/A ratio shift to 2.49 and its documented activation of TGF-β/SMAD, ERK, MMP-2, MMP-9, and NF-κB (A Porcine Model of Heart Failure With Preserved Ejection Fraction Induced by Chronic Pressure Overload Characterized by Cardiac Fibrosis and Remodeling), provides a tissue-level check: predicted signaling states from geometry should agree with measured transcriptomic shifts in GSE167643.

Baselines

The naive reference point is standard clinical risk assessment — age, blood pressure, smoking, glucose, lipids. This is not a strawman: it is what determines most real-world cerebrovascular risk stratification, and the file quantifies its ceiling. On MRI-confirmed silent brain infarction, clinical traits achieve an AUC of 0.633 (A deep learning system for detecting silent brain infarction and predicting stroke risk), and when risk factors are pooled with retinal morphometrics in regularized regression, the retinal features are consistently preferred — only hypertension survives alongside them for white matter hyperintensities (Retinal Biomarkers Discovery for Cerebral Small Vessel Disease in an Older Population). Any shear-informed model must at minimum beat 0.633 on silent infarction detection to justify its existence.

The strongest existing competitor is DeepRETStroke, a Vision Transformer foundation model pretrained on 895,640 retinal photographs, which detects silent brain infarction at an AUC of 0.797 and predicts incident stroke at 0.901 (A deep learning system for detecting silent brain infarction and predicting stroke risk). We will run it, and a RETFound-style masked-autoencoder model (A foundation model for generalizable disease detection from retinal images), on exactly our cohorts with the same label budgets and the same tuning compute; the fair comparison is at matched training labels, since RETFound already shows that pretraining buys label efficiency — its diabetic retinopathy AUROCs of 0.943, 0.822, and 0.884 across three target datasets were achieved with only 10% of the labeled data of baseline pipelines. Beating DeepRETStroke means one of two results: a higher AUC on held-out external sites, or an equal AUC that survives domain transfer to a new country, scanner, and image quality distribution where the pixel model degrades. We consider a within-domain AUC tie with substantially better cross-domain and cross-species stability to be a win, and we say so in advance.

The decisive ablation isolates the physics. We compare three feature sets on identical patients and identical classifiers: raw geometric features (calibre, tortuosity, branching angles, fractal dimension, exactly the VAMPIRE 3.1 parameter set), the same geometry passed through the shear-inversion layer without the mechanotransduction response curve, and the full pipeline with the response curve. If the full pipeline does not beat raw geometry, the mechanotransduction story is decorative and we will report that plainly.

Why not just…?

Why not just fine-tune a bigger foundation model on the same images? This is the most serious threat, and we treat it as such. A sufficiently large pretrained model may match our within-domain AUC without knowing anything about shear, exactly as DeepRETStroke already achieves 0.797 on silent infarction from pixels alone (A deep learning system for detecting silent brain infarction and predicting stroke risk). Three things keep our idea alive against it. First, the same literature documents fragility: pretrained models drop from AUROC 0.943 to 0.822 across geographic cohorts (A foundation model for generalizable disease detection from retinal images), and segmentation pipelines degrade to F1 of 0.73 and 0.66 on external sets before confidence modules rescue them (AutoMorph: Automated Retinal Vascular Morphology Quantification Via a Deep Learning Pipeline); physics-derived features are parameter-poor and should transfer better, and our matched-label transfer experiment will test exactly this. Second, a pixel model cannot be applied to a mouse or a pig at all, so the entire cross-species claim has no competitor in this alternative. Third, there is a live tension the pixel camp has not resolved: when vascular arcades are digitally removed by inpainting, age and sex prediction barely suffers (Effects of Hypertension, Diabetes, and Smoking on Age and Sex Prediction from Retinal Fundus Images), meaning pixel models may be reading diffuse parenchymal texture rather than vascular mechanics. If our inpainting-style ablation shows parenchymal features dominate even for vascular-specific endpoints like white matter hyperintensities, our claim narrows and we will say so.

Why not just run better rodent trials — blinding, randomization, larger samples? The stroke meta-analysis shows that methodological rigor is genuinely absent and that publication bias (p < 0.001 for infarct size) inflates the record (Treatment of acute cerebral ischemia using animal models: a meta-analysis), so this alternative would help. But rigor does not fix physics: a perfectly blinded mouse experiment still operates at 500–700 heartbeats per minute in vessels without a human-scale branching architecture, and rodent "retina" lacks the macula and foveal avascular zone entirely (Inhibition of AMD-Like Pathology With a Neurotrophic Compound in Aged Rats and 3xTg-AD Mice). Our shear-mismatch metric is complementary: it identifies which rigorously conducted experiments are nonetheless in the wrong regime, and our re-analysis of the 145-study corpus gives it a direct empirical test.

Why not just validate FAK forces directly with FRET tension sensors in mice? This is the natural experiment from the mechanotransduction side — the 150–450 pN window is so far validated only in simulation (Mechanism of Focal Adhesion Kinase Mechanosensing) — and we would welcome it, but it validates the molecule, not the imaging readout, and not the human translation. Our pipeline inverts the dependency: it generates vessel-by-vessel predictions of where FAK activation and MT1-MMP expression should concentrate in an intact animal, which is precisely the target map such single-molecule experiments need. The two approaches confirm each other; neither substitutes for the other.

Proposed Experiments and Decisive Checks

Experiment one applies the pipeline to the Lothian-style paired retinal–MRI setting, evaluating prediction of white matter hyperintensities (478 subjects) and perivascular spaces (381 subjects) with area under the receiver operating characteristic curve as the metric, against the raw-geometry, shear-only, and full-pipeline feature sets, with the geometry baseline expected to land near the elastic-net bootstrap performance already reported. Experiment two runs the matched-label external-transfer comparison against DeepRETStroke and a RETFound-style baseline on UK Biobank and AlzEye incident stroke outcomes, with success defined as higher external-site AUC or equal AUC with materially smaller degradation across sites. Experiment three computes shear-regime distributions for rodent, porcine, and human retinal trees and tests the mismatch-predicts-inflation hypothesis on the 145-study meta-analysis corpus, checking predicted signaling states against the porcine transcriptomic shifts in GSE167643 (A Porcine Model of Heart Failure With Preserved Ejection Fraction Induced by Chronic Pressure Overload Characterized by Cardiac Fibrosis and Remodeling) and against OCTA fractal-dimension loss in Alzheimer's disease measured on the ROSE dataset, where the segmentation model reaches an AUC of 0.9453 on superficial vessels (ROSE: A Retinal OCT-Angiography Vessel Segmentation Dataset and New Model). The decisive check is the ablation chain: if the mechanotransduction layer adds nothing beyond raw geometry on human endpoints and nothing beyond random ordering on the species-mismatch test, the hypothesis is wrong in its strong form and the note will say so. If it holds, the payoff is a fundus photograph that reads out not just disease association but the physical mechanism — where in the tree the endothelium is being told to degrade its surroundings, and whether an animal experiment shares that instruction with a human.