Stress Inference in Retinal Pigment Epithelium in C57BL/6J Mouse

Abstract

Purpose

To quantitatively map the spatial architecture of mechanical stress and pressure in the retinal pigment epithelium (RPE) and to test the central hypothesis that the mechanical state of normal mouse RPE is spatially patterned from center to periphery and that this spatial pattern is maintained during adult aging.

Methods

Whole-mount flat images of C57BL/6J mouse RPE spanning multiple age groups were analyzed using a tissue-scale stress inference framework under the assumption of mechanical equilibrium. Individual RPE cells were segmented, intercellular junctions were skeletonized, and the resulting geometric networks were used to estimate local intracellular pressure and junctional tension. Radial profiles of these mechanical fields were computed from the tissue center to the periphery and systematically compared across age groups.

Results

Inferred intracellular pressure exhibited a robust and monotonic decrease from the central to peripheral RPE across all age groups, whereas junctional tension displayed no consistent radial dependence. We did not detect substantial age-associated changes in the inferred pressure and tension profiles among the usable flatmounts analyzed, consistent with apparent preservation of the large-scale mechanical organization across the ages examined.

Conclusions

Tissue stress inference revealed robust spatial patterning in inferred RPE mechanics, characterized by a pronounced decline in inferred intracellular pressure toward the periphery and relatively uniform junctional tension across the tissue. No strong age-associated changes were detected in the inferred mechanical profiles, suggesting that the spatial mechanical signatures of the RPE are largely preserved across the ages studied. Together, these data establish a quantitative biomechanical baseline against which AMD- or genotype-associated remodeling of RPE mechanics can be evaluated.

The retina is continuously exposed to a wide range of potentially damaging stressors, including environmental insults, genetic mutations, and age-related structural and functional degradation.^1,2 The retinal pigment epithelium (RPE) is a multifunctional monolayer of cells situated between the retinal photoreceptors and the choroid, and it is a central site of age-related macular degeneration (AMD) pathogenesis.^3–7 AMD affects nearly 20 million adults over the age of 40 years in the United States, making it one of the most common causes of irreversible central vision loss.⁸ In healthy eyes, the RPE forms a dense, highly organized epithelial sheet that maintains and metabolically supports the photoreceptors and contributes to the integrity of the outer blood–retinal barrier.^9–11 With aging, RPE structure and function become increasingly perturbed and contribute to a range of retinal disorders, including AMD.^12,13 Distinct stages of AMD progression, from lipofuscin accumulation and drusen formation to geographic atrophy and choroidal neovascularization, impose differing mechanical and biochemical stresses on the RPE.^13–16 At the same time, RPE tissue architecture is maintained through a balance of intra- and extracellular forces, including adhesion, tension, and contraction, acting within a mechanically constrained retinal environment.^17–19

RPE cells exhibit regular polygonal, predominantly hexagonal, packing, a geometry that is common across both biological and physical systems. Hexagonal tiling appears in columnar basalt formations such as the Giant's Causeway in Ireland,²⁰ in honeycombs,²¹ and in two-dimensional soap foams.²² Similar polygonal and often hexagonal-dominated patterns are observed in numerous epithelial monolayers, where they are thought to arise from an interplay among cell–cell adhesion, cortical tension, cell proliferation, and external geometrical and mechanical constraints.^23–25 Vertex and related mechanical models have shown that modest changes in junctional tension, preferred cell area, or adhesion can drive transitions between ordered hexagonal and disordered polygonal packings, linking cell-scale mechanical parameters to emergent tissue architecture.^24,25

As retinal and macular diseases progress, mechanical stresses induced by drusen accumulation, RPE physiological alterations, cell death, and extracellular matrix remodeling lead to quantifiable deformation and reorganization of both individual RPE cells and the global tissue pattern.^26–28 Quantitative analysis of RPE morphology in both mouse models and human tissue have demonstrated that cell area, shape, and topological descriptors discriminate age and genotype and differ between AMD and healthy eyes.^29–31 More recently, RPE curvature and thickness metrics derived from clinical imaging have been associated with risk of progression to late AMD, underscoring the diagnostic and prognostic value of structural RPE measurements.^13,31 These findings together suggest that RPE morphometry encodes information about local tissue stress, disease stage, and susceptibility to degeneration. However, although morphometric changes have been extensively characterized, the underlying mechanical stress and pressure fields in the RPE remain largely unexplored. We hypothesize that the mechanical state of normal mouse RPE is spatially patterned from center to periphery and that this spatial pattern is maintained during adult aging.

As a first step toward linking RPE morphometry to its mechanical state, we applied a tissue-scale stress-inference framework to mouse RPE flatmounts spanning six ages in the C57BL/6J strain. Using a fully automated pipeline for radial strip selection, cell segmentation, skeletonization, and force inference, we estimated relative intracellular pressure and junctional tension across the tissue and characterized how these mechanical fields vary from the retinal center to the periphery and across age groups. By establishing a quantitative baseline of RPE mechanical organization in normal aging, this work provides a foundation for future studies of how AMD- or genotype-associated perturbations remodel RPE mechanics.

Methods

Sample Preparation

We analyzed RPE flatmount images from our previously described reference database.^27,32 A total of 45 flatmount images from C57BL/6J mice across six age groups (postnatal day 30 [P30], P45, P60, P180, P330, and P720) were screened; 29 yielded usable strips, and 25 of them passed all quality-control criteria and were included in the final stress-inference analyses. Images are publicly available at the Cell Image Library (http://cellimagelibrary.org/groups/52245). All animal procedures adhered to the ARVO Statement for the Use of Animals in Ophthalmic and Vision Research and were approved by the Emory University School of Medicine Institutional Animal Care and Use Committee.

Analysis Pipeline

We developed a scalable pipeline to perform stress inference³³ on flatmount RPE images (Fig. 1). Cell segmentation was performed using Cellpose-2,³⁴ skeletonization in ImageJ 1.54f (64-bit, Java 1.8.0_322; National Institutes of Health, Bethesda, MD, USA)³⁵ via the integrated Jython (Python) interpreter, and stress inference using a Mathematica implementation of CellFIT³⁶ (https://github.com/mshutson/CellFIT). After flatmount images were manually annotated to mark usable regions (Fig. 1B, strip selection), all subsequent stages were run in a largely automated, batch processing mode. A directory of marked-up flatmount images served as input and yielded (1) a complete stack of valid cropped radial strips, (2) corresponding cell and nuclear segmentations, and (3) skeletonized junctional networks. These skeletons were then passed to CellFIT via a Mathematica batch controller that repeatedly executes the core stress-inference routine.

Figure 1.

Tissue-stress inference pipeline for mouse RPE: (A) Representative C57BL/6J RPE flatmount stained for tight junctions (ZO-1, green) and nuclei (propidium iodide, red). (B) Same flatmount after manual markup, with holes, folds, tears, and edge artifacts masked in white. (C) Example radial strip automatically cropped from the marked image, extending from the center of optic nerve head (left) to the periphery (right) while avoiding masked regions. (D) Cellpose-2 “cellprob” output showing the cell boundary probability map. (E) Skeletonized cell-edge network generated in ImageJ from the Cellpose probability map. (F) CellFIT output visualizing relative cell pressure as a color map on reconstructed cell skeleton, with warmer colors indicating higher pressure and cooler colors indicating lower pressure.

Strip Selection

CellFIT requires watershed-segmented or skeletonized input images with limited size to avoid excessive run times, numerical instabilities, and failures associated with broken junctions. Consequently, whole flatmounts cannot be processed in a single pass. Because CellFIT outputs relative, not absolute, tension and pressure values, measurements are directly comparable only within a given image. Arbitrary cropping into small, independent squares would therefore destroy the global spatial context. To preserve meaningful spatial gradients, we extracted elongated radial strips of cells (Fig. 1C) extending from the flatmount center to the periphery. This design permits direct comparison of relative tension and pressure between widely separated radial positions, including the center and outer edge.

To avoid subjective and laborious manual cropping, strip extraction was fully automated. The flatmount center was first approximated by searching near the geometric center for the brightest region in the green channel, which was defined as the tissue center. The cropping algorithm then swept around this point in polar coordinates, extracting radial strips as a function of angle.

This procedure was complicated by frequent imaging and tissue artifacts such as broken junctions, folds, and surface ridges (Fig. 2). When applied to unprocessed flatmounts, the algorithm often returns strips that are unsuitable for stress inference; distorted cells yield distorted mechanical estimates, and broken cells can cause CellFIT to fail or return unreliable output. To mitigate this, each flatmount was manually inspected, and artifact-containing regions were masked (whitened). During automated cropping, any candidate strip containing white pixels was discarded, ensuring that only artifact-free strips entered the stress-inference stage.

Figure 2.

Identification and masking of artifacts in RPE flatmounts. (Left) Example RPE flatmount region with several artifacts: (A) fold in the tissue, (B) large hole from tissue loss, and (C) broken cells near the cut edge. (Right) Same region after manual markup, with artifacts masked in white. The strip selection algorithm subsequently discards any candidate strips containing white regions, ensuring that only intact, undistorted tissue is used for stress inference.

At each angular position θ, the algorithm evaluated multiple starting radii r, allowing it to bypass proximal artifacts and recover acceptable strips at more distal positions when necessary. Two versions of the cropping algorithm were implemented: one with fixed strip length and one that tested multiple candidate lengths at each (r, θ), retaining the longest valid strip. In all cases, strip width was held constant across the dataset. Only flatmounts with sufficiently clear segmentation and long continuous strips were used.

Image Segmentation

Cellpose-2³⁴ was used for automated segmentation of both RPE cell bodies and nuclei. For cell boundary extraction, we used the cellprob output, which provides a probabilistic map of the likelihood of each pixel belonging to a cell. This representation is crucial for generating high-quality skeletons of the radial strips. Because of substantial variability in green-channel staining intensity, direct skeletonization of raw fluorescence images produced numerous broken junctions, precluding reliable stress inference. In contrast, skeletonization of the cellprob output yielded continuous, well-defined cell boundaries suitable for downstream mechanical analysis. Default Cellpose parameters were sufficient to generate cellprob images for skeletonization, but additional parameter tuning was required for accurate nuclear segmentation. We performed a grid search over the Cellpose parameter space using a representative subset of strips and selected optimal parameters based on visual assessment of segmentation fidelity. The performance of this nuclear identification approach is illustrated in Figure  3.

Figure 3.

Validation of geometric reconstruction and nuclei assignment. From top to bottom: cropped radial strip from a flatmount image; Cellpose 2 nuclei centroids (red) overlaid on cell walls reconstructed from CellFIT output; CellFIT-derived cell centroids (blue) overlaid on reconstructed cell walls; and the restricted set of nuclei used for assignment superimposed on the reconstruction. These overlays illustrate that segmentation, skeletonization, and CellFIT reconstruction faithfully reproduce the underlying tissue geometry and enable robust mapping of nuclei to individual cells.

Skeletonization

The cellprob output was processed using a custom ImageJ macro to generate a skeletonized representation of each strip (Fig. 1E). The macro first binarized the probabilistic image and then applied the skeletonization operation. For successful skeletonization, binarization must assign white pixels to cell walls rather than interiors; however, local intensity variations can occasionally invert this assignment. When white pixels correspond to cell interiors, skeletonization yields scattered, disconnected line fragments rather than contiguous cell outlines. To resolve this ambiguity, the macro generated two candidate skeletons: one from the original binarized image and another from its inverted version. The number of connected components in each skeleton was computed, and the skeleton with fewer components was selected as the valid representation; the other was discarded. Finally, a 10-pixel border around the perimeter of the skeleton was trimmed to remove spurious boundary artifacts consistently produced at strip edges. The cleaned skeleton served as the direct input to the stress-inference step.

Stress Inference Analysis

CellFIT took the skeletonized images as input, performed stress-inference computations, and returned relative junctional tension and intracellular pressure values for each strip. In addition, CellFIT outputted a full topological and geometric description of the tissue, including mappings between cells and junctional edges, cell centroid coordinates, and node coordinates along cell boundaries. Because the skeletonized image can be reconstructed from the CellFIT output, this provides a robust internal consistency check and enables direct registration of Cellpose-derived nuclear masks with the inferred mechanical fields (Fig. 3). Details of the stress-inference framework and the specific CellFIT implementation were described previously.^36,37 Each edge-level tension estimate returned by CellFIT is accompanied by a standard error estimate. Cell-level pressure estimates, which are derived from the tensions of the surrounding edges, likewise have associated standard errors. We propagated these uncertainties to compute standard errors for averaged tension and pressure values for each image strip and for each flatmount.

Nuclei Assignment

Before downstream analysis, nuclei from the Cellpose-derived masks were mapped onto the CellFIT cell outputs. As shown in Figure  3, this mapping requires restricting candidate nuclei to regions where valid stress-inference solutions exist. Although the nuclei masks include all detected nuclei within each strip, CellFIT does not return solutions for all segmented cells, and cells cut off at strip boundaries do not yield valid measurements. Assigning nuclei outside the spatial domain of valid CellFIT outputs leads to spurious matches. Accordingly, only centrally located nuclei within the valid CellFIT bounds were retained for assignment.

For each remaining nucleus, the nearest CellFIT edge was identified. Each edge is associated with exactly two adjacent cells. Straight-line segments were then drawn from the nucleus centroid to the centroids of each of these two candidate cells. If a segment could be drawn from the nucleus to a given cell centroid without intersecting the intervening cell wall, the nucleus was assigned to that cell. If both candidate paths intersected the edge, the algorithm moved to the next closest edge and repeated the procedure. This geometric assignment strategy yields robust correspondence between nuclear identity and mechanically inferred cellular properties.

Results

RPE Cell Geometry Obeys Lewis and Aboav–Weaire Laws

The geometric organization of epithelial tissues has long been known to obey simple statistical “laws” that link cell area and topology in polygonal tiling, both in biological epithelia and in non-biological systems such as soap bubbles and metallic grains. In particular, Lewis first reported a linear relationship between mean cell area and the number of neighbors in cucumber epidermis.³⁸ Aboav and Weaire later showed that cells with many neighbors tend to be surrounded by cells with fewer neighbors, generating a characteristic negative correlation between the neighbor number of a cell and the average neighbor number of its neighbors, as in liquid foams³⁹ and arsenic–selenium glass,⁴⁰ Drosophila epithelia,⁴¹ and other biological tissues.⁴²

To test whether mouse RPE follows these universal geometric principles, we grouped all segmented RPE cells first by age group and then by polygon class (neighbor number). For each age–polygon combination, we computed the mean apical cell area (Fig. 4, top). Across all ages, mean cell area increased approximately linearly with neighbor number, indicating that RPE obeys Lewis’ law. We next grouped cells by neighbor number and, for each group, computed the average neighbor number of adjacent cells (Fig. 4, bottom). In all age groups, cells with many neighbors were preferentially surrounded by cells with fewer neighbors, yielding a negative, approximately linear relationship consistent with the Aboav–Weaire law. Taken together, Lewis’ law and the Aboav–Weaire law suggest that RPE organization is not random but follows conserved packing principles that balance cell area and local connectivity.

Figure 4.

RPE obeys reproducible geometric and topological rules. (Top) Mean RPE cell area as a function of neighbor number (Lewis law). Cells are grouped by polygon class, and the average apical area in each class is plotted. Lines indicate approximately linear increases in area with neighbor number. (Bottom) Mean neighbor count of neighboring cells as a function of the neighbor number of each cell (Aboav–Weaire law). Cells with many neighbors tend to be surrounded by cells with fewer neighbors, producing a negative, approximately linear relationship that is preserved across age groups.

Junctional Membrane Tension Lacks a Consistent Radial Gradient

CellFIT returns relative junctional membrane tension values for all resolved cell–cell interfaces within each radial strip. To assess whether tension varies systematically from the retinal center to periphery, we assigned each junction a radial coordinate based on the centroid of its adjacent cells and, for each flatmount, averaged tension within radial bins to obtain a one-dimensional profile (Fig. 5).

Figure 5.

Radial distribution of membrane tension in RPE. Each colored curve represents one eye (one flatmount), and shaded bands denote propagated standard errors after averaging strip-level measurements within each flatmount. Average relative cell wall tension is plotted as a function of radial distance from the center of the flatmount to the periphery of the RPE, where 0 corresponds approximately to the optic nerve head. For each strip, a moving-average radial profile was computed and then averaged across strips within the same flatmount. Across ages, tension shows substantial variability but no consistent radial gradient. From top to bottom, the number of flatmounts represented are P30, 2; P45, 4; P60, 4; P180, 7; P330, 4; P720, 4.

Across all animals and ages, radial tension profiles displayed substantial strip-to-strip variability but no consistent monotonic trend. Linear regression of tension versus radial position yielded slopes scattered around zero, with no systematic shift toward positive or negative gradients in any age group (Supplementary Table S3). These analyses indicated that, within the sensitivity of our method, junctional membrane tension does not exhibit a robust radial gradient and does not change appreciably with age.

Intracellular Pressure Decreases From Center to Periphery

In contrast to junctional tension, CellFIT-inferred intracellular pressure showed a clear spatial trend. Using the same binning and regression strategy, we computed radial profiles of relative cell pressure for each flatmount (Fig. 6). For most strips and across all age groups, pressure decreased with increasing distance from the retinal center, yielding predominantly negative slopes in pressure–distance regressions (Fig. 7). Although the magnitudes of these slopes varied across animals, pairwise comparisons of slope values between age groups yielded P > 0.1 in all cases. Thus, the RPE exhibits a robust radial gradient in inferred intracellular pressure, with higher pressure centrally and lower pressure peripherally, and we did not detect a systematic age-associated change in the strength of this gradient.

Figure 6.

Radial distribution of relative cell pressure in RPE. Each colored curve represents one eye (one flatmount), and shaded bands denote propagated standard errors after averaging strip-level measurements within each flatmount. Average relative cell pressure is plotted as a function of radial distance from the center of the flatmount to the periphery of the RPE, where 0 corresponds approximately to the optic nerve head. For each strip, a moving-average radial profile was computed and then averaged across strips within the same flatmount to obtain the curves shown. Across age groups, relative pressure decreased from the center toward the periphery, revealing a robust negative radial gradient. From top to bottom, the number of flatmounts represented are P30, 2; P45, 4; P60, 4; P180, 7; P330, 4; P720, 4.

Figure 7.

Linear regression slopes of pressure–distance relationships by age. Boxplots summarize the distribution of slopes obtained from linear fits of relative intracellular pressure versus radial distance for individual flatmounts within each age group. Slopes are predominantly negative, confirming a center-to-periphery decrease in pressure, but their magnitudes do not vary systematically with age. Pairwise t-tests comparing slopes between age groups yielded P > 0.1 in all cases, indicating no statistically significant age-dependent differences in the radial pressure trend within the sensitivity of this analysis. Because the number of biological replicates per age group was small (n = 2–7), we elected to analyze flatmount-level summary statistics rather than fit mixed-effects models, which would be poorly constrained in several age groups.

Distributions of Membrane Tension and Cell Pressure Show No Major Differences Across Age

To determine whether the overall structure of the inferred mechanical fields changes with age, we compared the distributions of junctional tension and intracellular pressure across age groups. Because CellFIT yields relative rather than absolute values, we normalized tension and pressure in each image so that the mean pressure and mean tension were constrained to 0 and 1, respectively. We then pooled normalized measurements within each age group and estimated their probability densities (Fig. 8).

Figure 8.

Distributions of relative cell pressure and membrane tension across age groups. Kernel density estimates of CellFIT-derived measurements are shown for each age group. (Left) Empirical density curves for relative intracellular pressure. (Right) Empirical density curves for junctional membrane tension. CellFIT constrains the mean pressure and tension to 0 and 1, respectively, for each image; thus, only the spread and shape of the distributions are free to vary. Neither pressure nor tension exhibits systematic age-dependent broadening, narrowing, or skewing of the distributions. See Supplementary Table S1 for sample counts.

Density curves for relative junctional tension were nearly identical across ages, showing similar breadth and shape. Likewise, pressure distributions overlapped strongly between age groups, with no age-dependent broadening, narrowing, or skewing. These observations suggest that the statistical structure of the RPE mechanical fields is preserved with age: Although cell size, shape, and nuclearity change over time, the normalized distributions of tension and pressure appear broadly similar from P30 to P720.

Intracellular Pressure Does Not Scale With Cell Area

Previous work in other tissues has reported a positive relationship between cell pressure and cell size or volume.^43,44 To test whether a similar scaling holds in RPE, we examined the joint distribution of relative intracellular pressure and apical cell area. Within each age group, we pooled all cells and plotted pressure versus area along with the corresponding marginal density estimates (Fig. 9).

Figure 9.

Intracellular pressure does not scale with cell size. Scatterplots show relative intracellular pressure versus cell area for RPE cells aggregated within each age group. Marginal kernel density distributions for pressure and area are displayed along the corresponding axes. See Supplementary Table S2 for cell counts and ratios relevant to this figure.

Across all ages, the RPE did not exhibit a consistent linear relationship between pressure and area. Instead, scatterplots showed a broad cloud of points with only weak apparent structure, and regression fits did not reveal a robust positive or negative trend. Some age groups displayed hints of multiple clusters or modes, suggesting the possible emergence of distinct subpopulations in the pressure–area plane, but these patterns were subtle and not easily generalized across ages. Overall, our data indicate that, in the mouse RPE, intracellular pressure is not simply predicted by cell apical area.

Multi-Nucleated Cells Are Larger and Centrally Enriched Across Age

Finally, we examined how nuclearity relates to spatial position within the RPE. Visual inspection of nuclear maps revealed a pronounced spatial segregation: Multinucleated cells are concentrated near the retinal center, whereas mononucleated cells become increasingly prevalent toward the periphery. This pattern is apparent when individual strips are color coded by nuclearity (Fig. 10A). To quantify this segregation, we computed the radial distributions of mono- and multinucleated cells for each age group (Fig. 10B). Violin plots of nuclearity as a function of radial position showed that, across all ages, multinucleated cells remain biased toward central regions, whereas mononucleated cells are enriched peripherally. An apparent upward shift in the P720 distribution reflects a sampling bias (a disproportionate number of strips captured more central regions in this age group) rather than a true reversal of the spatial trend. Consistent with previous reports, the proportion of multinucleated cells increased with age (Supplementary Table S2). Together, these findings show that multinucleated RPE cells—which tend to be larger—are preferentially located near the retinal center and become more common with aging, whereas smaller, mononucleated cells dominate the periphery.

Figure 10.

Nuclear distribution within RPE radial strips by age. (A) Representative radial strip oriented with the retinal center on the left and the periphery on the right. Multinucleated cells are highlighted in red and mononucleated cells in black, illustrating the central enrichment of multinucleated cells. (B) Violin plots of radial position for mono- and multinucleated cells stratified by age group. Violins display the full distributions with first quartile, median, and third quartile indicated; sample sizes range from approximately 1200 to 10,000 cells per group. Across all ages, multinucleated cells remained concentrated toward the center, whereas mononucleated cells were enriched in the periphery.

Discussion

This study was motivated by the long-standing observation that RPE cells undergo pronounced structural remodeling in aging and AMD, but the mechanical state of the RPE has remained largely unexplored. To address this gap, we developed an end-to-end pipeline that automates radial strip selection, segmentation, skeletonization, and stress inference on mouse RPE flatmounts spanning six age groups. This framework allowed us to quantify how junctional tension and intracellular pressure vary as a function of radial position from center to periphery and to assess whether these mechanical fields remodel with age. We found that RPE cell geometry obeys the Lewis and Aboav–Weaire laws, that inferred intracellular pressure exhibits a consistent, monotonic decline from center to periphery and that junctional membrane tension shows no consistent radial trend. Importantly, both the spatial profiles and normalized distributions of tension and pressure were conserved across all ages examined. We did not detect a strong age-associated change in the large-scale planar mechanical organization of the RPE.

Intracellular pressure in the RPE likely reflects the combined influence of external mechanical loading and intrinsic physiological transport. Externally, the RPE is located between the overlying photoreceptors and the underlying Bruch's membrane and choroid, forming a mechanically coupled sandwich that may impose region-specific compressive and tensile forces. Internally, the active transport by the RPE of ions, solutes, and water between the neural retina and choroidal circulation generates hydrostatic and osmotic forces that contribute to cell pressure and tissue tension. The central-to-peripheral decline in inferred pressure we observe is therefore plausibly shaped by both tissue-level constraints, such as globe curvature and outer-retinal architecture, and local differences in transport or barrier properties. Although our measurements are in relative units, the consistency of this negative pressure gradient across animals and ages suggests that it is a robust feature of normal RPE mechanics. We emphasize that the higher central pressure is not itself a demonstrated functional mechanism but rather evidence for region-dependent mechanical homeostasis. This idea is supported by at least three known central-to-peripheral trends in the outer retina/RPE complex: region-specific enrichment of multinucleated RPE cells near the center, central–peripheral differences in Bruch's membrane composition, and spatial heterogeneity of the choroid/choriocapillaris, each of which could alter how RPE cells bear load, anchor to the substrate, or exchange fluid with surrounding tissues. Because AMD preferentially affects the macular region and involves coupled dysfunction of the RPE, Bruch's membrane, and choriocapillaris, these observations raise the possibility that region-specific mechanical regulation contributes to the particular vulnerability of the macula. At the same time, our study did not directly test these mechanisms, so we regard them as biologically motivated hypotheses for future work rather than established causal conclusions.

Previous work has demonstrated that joint statistics of RPE cell area and shape can robustly discriminate age and genotype and that central–peripheral differences in morphology and multinucleation are pronounced. In contrast, we were unable to distinguish age groups based on stress-inferred tension and pressure alone. This divergence suggests that planar mechanical stress fields may be more tightly regulated or buffered than apical morphology during normal aging. One interpretation is that the RPE accommodates age-related changes in cell size, shape, and nuclearity through subcellular or extracellular remodeling (for example, alterations in cell height, cytoskeleton, or Bruch's membrane properties) while maintaining a relatively stable planar stress architecture. Consistent with prior reports, we likewise observed multinucleated, larger cells concentrated near the tissue center and mononucleated cells dominating the periphery. However, we found no simple relationship between inferred intracellular pressure and either cell area or nuclear content, indicating that local cell size and nuclearity do not directly predict local pressure in mature RPE. More broadly, the age-related increase in mean RPE cell area reported in earlier work is not generally interpreted as evidence of continued cell division in adult tissue, because mature mammalian RPE is largely postmitotic. Instead, the prevailing interpretation is that local cell loss or damage is compensated by expansion and rearrangement of neighboring cells, allowing the epithelial sheet to remain continuous while average cell area increases. Our observation of a robust negative pressure gradient therefore suggests that large-scale tissue-level stress organization may coexist with, and potentially contribute to, those age-associated morphometric changes.

Taken together, these findings indicate that, despite substantial age-associated alterations in RPE morphology reported in the literature, the planar mechanical stress patterns measured here appear broadly similar across the ages examined in C57BL/6J RPE. Rather than viewing this as a lack of diagnostic signal, we interpret it as establishing a stable biomechanical baseline. Junctional tension and intracellular pressure, as inferred by our approach, may not be sensitive markers of normal, strain-matched aging, but deviations from this conserved baseline could prove informative in disease or under experimental perturbation. As an internal methodological cross-check, Bayesian force inference⁴⁵ reproduced the same qualitative center-to-periphery decrease in inferred pressure (Supplementary Fig. S2), whereas a preliminary analysis of arc-shaped crops did not reveal a consistent angular trend (Supplementary Fig. S3). Together, these supplementary analyses support the robustness of the radial pressure pattern while also suggesting that future work could examine anisotropy more systematically. In principle, the same pipeline could be applied to RPE from AMD models, other mouse strains, or genetically altered backgrounds to test whether specific insults or genotypes produce detectable reorganization of planar stress fields.

This study has several important limitations. First, the mechanical quantities reported here are inferred from tissue geometry using a force-balance framework and are therefore model-dependent rather than direct physical measurements. To assess robustness, we compared CellFIT-derived results with an independent Bayesian force-inference approach and obtained qualitatively similar spatial patterns, including the center-to-periphery pressure gradient. However, this comparison should be viewed as a consistency check between computational inference methods rather than an independent experimental validation. Previous studies have demonstrated agreement between force-inference methods and independent mechanical measurements, including laser-ablation recoil and photoelasticity-based approaches in epithelial tissues,^45,47 but these validations were performed in other biological systems. Because the present study is based on fixed RPE flatmounts, direct mechanical perturbation experiments or live-tissue validation are not possible. Consequently, we interpret the observed pressure gradient as an inferred mechanical signature derived from tissue geometry rather than a directly measured intracellular pressure field.

Second, tissue stress inference assumes that cell boundaries are accurately segmented and can be locally approximated as arcs of constant curvature under quasi-static mechanical equilibrium. These geometric and mechanical assumptions substantially restrict the contexts in which the method can be applied. In degenerative settings such as the rd10 mouse retina, RPE cell borders become highly irregular and tortuous, violating the smooth-curvature assumption and precluding reliable tension estimation.^27,46 Similarly, in AMD patient samples, cell borders are frequently fragmented or discontinuous, so the underlying geometry no longer satisfies the mathematical requirements for solving the inverse force-balance problem. Accordingly, stress inference based on cell borders is unlikely to yield meaningful mechanical measurements in late-stage or severely disrupted RPE tissues.

Third, because CellFIT provides only relative, not absolute, measures of tension and pressure, comparisons are strictly valid within individual images; we mitigated this limitation by analyzing continuous radial strips, but we could not directly quantify absolute stress magnitudes. Fourth, our analysis is cross-sectional, limited to one mouse strain, and restricted to planar projections of a three-dimensional tissue; future work integrating three-dimensional imaging, alternative force-inference methods, or longitudinal designs could reveal additional aspects of RPE mechanics that are not captured here.

Fifth, although force-inference methods have shown good sensitivity in other epithelial systems,³³ it remains possible that the RPE is a comparatively challenging tissue in which the present approach is less sensitive than expected to subtle age-dependent changes. Finally, we agree that validation in genetically altered or experimentally perturbed RPE models would be valuable, but identifying a suitable model is not straightforward. Many available injury models, including sodium iodate, produce acute, severe, or heterogeneous disruption with fragmented boundaries and loss of local mechanical equilibrium, conditions that can invalidate the assumptions required for reliable stress inference. A more appropriate validation system would have to produce reproducible, progressive, and sufficiently mild remodeling while preserving enough tissue integrity for morphology-based mechanical analysis.

Despite these limitations, our results demonstrate that tissue stress inference can be robustly applied to normal RPE, yielding reproducible patterns of intracellular pressure and junctional tension that are conserved across age. By providing a quantitative description of the baseline mechanical architecture of the mouse RPE, this work lays the foundation for future studies into how disease, genotype, or therapeutic intervention remodel RPE mechanics and how such mechanical changes intersect with the morphologic and functional hallmarks of retinal degeneration and AMD.