Geographic atrophy: spatial transcriptomics versus bulk analysis

The border between surviving and absent retinal pigment epithelium is a narrow pathological territory: RPE cells may be senescent, stressed, hypertrophic, or transcriptionally reprogrammed; photoreceptors above them may be degenerating at a different pace; the choriocapillaris beneath them may carry its own endothelial response. Bulk RNA-sequencing compresses these neighboring states into one averaged signal.

That compression is not a technical footnote. It determines what kind of question the tissue can answer.

In geographic atrophy donor tissue, the comparison between spatial transcriptomics and bulk analysis is therefore not a contest between an outdated method and a superior one. It is a comparison between two different views of retinal pathology: one that measures the molecular mixture of a dissected sample, and another that attempts to preserve the position of each signal within the damaged architecture.

Bulk RNA-sequencing sees the lesion after its geography has been removed

Bulk RNA-seq remains valuable because it offers substantial sequencing depth across a tissue fragment and can provide a robust overview of gene-expression differences between defined sample groups. If the researcher has a relatively homogeneous preparation—cultured human RPE cells, a carefully dissected retinal layer, or a pooled donor-tissue cohort—averaging can be an advantage. It reduces some of the noise that appears when individual cells are measured separately.

The problem emerges when the sample is anatomically heterogeneous by nature.

A donor macula containing geographic atrophy can include:

• residual RPE and areas where the RPE monolayer has disappeared;
• photoreceptors at different stages of structural loss;
• inner retinal cells displaced toward the lesion margin;
• choroidal vascular tissue beneath regions with contrasting pathology;
• infiltrating or activated microglia;
• extracellular deposits, including drusen-associated material;
• connective and vascular components from the choroid.

When these compartments are homogenized, their transcripts enter the same molecular pool. A gene may appear elevated because it is genuinely induced in surviving RPE cells, because the sample contains more vascular tissue, or because one donor contributes a larger fraction of inflammatory cells. The resulting differential expression profile can be statistically persuasive while remaining anatomically ambiguous.

This is the central limitation of bulk analysis in geographic atrophy molecular pathology: it can identify that a tissue region is different without preserving which cell population generated the difference.

That distinction becomes especially important near lesion borders. The atrophy core and the surrounding transition zone do not represent the same biological state. The core may be defined by extensive loss of RPE and photoreceptor structures, while the border contains cells under oxidative, metabolic, and inflammatory pressure but not yet absent. If the core and border are combined, the expression profile may describe an average that exists nowhere in the tissue.

A bulk result can therefore obscure the chronology of the lesion. It may register the molecular consequences of cell loss more strongly than the early response of cells that remain.

Bulk sequencing can tell us that the lesion has changed. Spatial profiling asks which cells changed first, which cells followed, and which signals belong to the border rather than the void.

The distinction is not merely semantic. Pathological progression is a sequence, and tissue homogenization can flatten that sequence into a single endpoint.

Spatial transcriptomics keeps the lesion attached to its coordinates

Spatial transcriptomics changes the starting point. Instead of removing the tissue’s architecture before molecular analysis, the method profiles gene expression within an intact section and retains the location of the measured signal. Techniques such as Visium HD can support genome-wide expression profiling directly on fixed frozen human donor macula sections, allowing molecular patterns to be interpreted relative to drusen, geographic atrophy lesions, surviving RPE, and surrounding tissue.

For donor retina spatial transcriptomics, the preserved coordinate is the critical object. A transcript is not only an expression event; it is an expression event occurring above a lesion, beside a deposit, within a surviving RPE region, or near the choriocapillaris.

That context allows several questions to be separated:

1. Is a transcriptional program restricted to the atrophy border, or does it extend through the apparently preserved macula?

2. Does an RPE-associated signal remain present where the RPE layer is structurally compromised?

3. Are vascular and RPE signatures changing together, or are they spatially decoupled?

4. Do drusen-associated regions show a distinct metabolic program compared with tissue farther from the deposits?

5. Does a gene-expression gradient exist between lesion core, transition zone, and apparently unaffected tissue?

Bulk RNA-seq can address some of these questions only if the tissue is dissected into carefully selected pieces before extraction. Even then, the spatial arrangement inside each piece is lost. Spatial transcriptomics retains more of the original scene, although it does not eliminate every ambiguity.

The macula itself is small—approximately 2.5–3 mm in diameter at the posterior pole—so sampling decisions have disproportionate consequences. A section can preserve the broad geometry of the macula while still missing a critical portion of a lesion boundary. Tissue orientation, section thickness, freezing quality, staining, and the alignment of histological features with molecular data all become part of the interpretation. In a donor specimen, the question is never only whether transcripts were detected. It is whether the detected pattern can be reliably assigned to the anatomical structure under investigation.

What each method preserves

The methods are most informative when treated as complementary. Bulk analysis can provide depth and cohort-level statistical power; spatial analysis can determine whether a signal belongs to a specific anatomical compartment. The practical problem is not choosing a winner. It is designing the experiment so that each method is asked a question it can answer.

The RPE is not one molecular population across the macula

Single-cell and spatial RNA-sequencing studies show that human RPE and choriocapillaris endothelial cells exhibit regional transcriptional heterogeneity. Macular and peripheral regions are not molecularly interchangeable, and donor-to-donor variation adds another layer of complexity. A transcriptomic difference may reflect disease state, anatomical location, donor biology, post-mortem interval, tissue preservation, or some combination of these variables.

For human RPE cell spatial mapping in AMD, this regional heterogeneity is not background noise. It is part of the pathology.

The RPE is a polarized, metabolically active epithelium positioned between photoreceptors and the choroidal circulation. Its transcriptional profile is shaped by the demands of phagocytosing photoreceptor outer segments, processing lipids, regulating oxidative stress, maintaining the visual cycle, and communicating with neighboring immune and vascular compartments. A cell overlying drusen is not simply a healthy RPE cell with one additional exposure. It occupies a different microenvironment, with altered substrate, altered extracellular signaling, and potentially altered mitochondrial demand.

In donor eyes, RPE cells overlying drusen show distinct upregulation of oxidative phosphorylation pathway genes and markers including ABCA5, MT-ND4L, VIM, and galactosyltransferases compared with normal RPE. The significance of this pattern lies partly in its location. If these changes are detected in a spatially defined drusen-associated region, they can be interpreted as a local response rather than a general feature of every RPE cell in the specimen.

Bulk analysis may still detect the same genes, particularly when a large fraction of the dissected tissue carries the relevant state. But it cannot establish whether the signal is concentrated directly over drusen, distributed across adjacent RPE, or contributed by another cell population captured in the same fragment.

This is where spatial transcriptomics becomes a form of pathological localization. It does not simply add a map to a list of genes. It changes the causal questions available to the investigator.

A gene associated with mitochondrial metabolism may represent a compensatory response in surviving RPE. A stress-associated marker may reflect senescence, reactive gliosis, vascular contamination, or a mixture of states. A spatially restricted signal narrows the possible explanations. It does not prove mechanism, but it forces the mechanism to remain accountable to anatomy.

The atrophy border is the most informative and most difficult territory

The center of geographic atrophy is visually dramatic but molecularly complicated. Where cells have been lost, absence can dominate the data. The border is less obvious at low magnification, yet it may contain the sequence of events most relevant to disease progression: RPE senescence, altered mitochondrial activity, photoreceptor distress, extracellular deposit accumulation, microglial activation, and vascular remodeling.

This is the region where spatial resolution benefits are most apparent—and where the limitations of the method become hardest to ignore.

A spatial capture area may intersect more than one structure. A surviving RPE cell can sit above a compromised photoreceptor layer and adjacent to an activated microglial process. Beneath it, choriocapillaris endothelial cells may contribute transcripts from the same neighborhood. Even when the molecular signal is spatially anchored, the biological source may remain mixed.

The exact single-cell spot resolution required to distinguish bordering RPE cells from adjacent microvascular endothelial cells in every zone of human geographic atrophy lesions is not fully established. That uncertainty matters because the boundary between cell populations is not always preserved as a clean geometric line in diseased tissue. Cells retract, hypertrophy, disappear, migrate, or become difficult to classify by morphology alone.

Reading a border without overreading it

A credible interpretation of an atrophy-border signal should connect at least three layers of evidence:

• Histology: What structures are physically present at the measured location?
• Molecular identity: Which genes or gene sets indicate RPE, photoreceptor, endothelial, immune, or glial contribution?
• Regional comparison: Does the signal differ between lesion core, border, drusen-associated tissue, and a matched non-lesional region?

Without this triangulation, a spatial map can create an illusion of certainty. The colored spots appear precise, but precision of position is not identical to precision of cell identity.

The same caution applies to microglial activation. A signal near the lesion edge may indicate activated microglia, but it may also reflect altered expression in stressed RPE or other immune-associated transcripts from neighboring cells. Spatial transcriptomics improves localization; it does not automatically resolve every cellular boundary.

For this reason, spatial datasets are strongest when integrated with morphology, immunostaining, cell-type reference profiles, and donor-level metadata. The point is not to make the map look definitive. The point is to make each inference traceable to the tissue.

RPE, choriocapillaris, and the problem of neighboring signals

The choriocapillaris introduces a second interpretive problem. It lies immediately beneath the RPE, and its endothelial cells are exposed to the same pathological neighborhood while maintaining a distinct vascular identity. In geographic atrophy, vascular changes may accompany RPE and photoreceptor degeneration, but the direction and timing of these changes cannot be assumed from spatial proximity alone.

Bulk RNA-seq is particularly vulnerable here. A choroid-containing sample may show increased endothelial or inflammatory transcripts simply because the relative amount of choroidal tissue has changed after RPE loss. The apparent molecular response may partly represent a compositional shift: less RPE, more vascular or connective tissue in the extracted material.

Spatial profiling can separate these layers more effectively by showing whether endothelial signatures intensify beneath the atrophy border, remain stable across the lesion, or vary independently of the overlying RPE state. Yet the physical closeness of the compartments still creates mixed measurements, especially where section geometry, tissue deformation, or disease-related structural loss reduces the distance between recognizable layers.

This is why donor tissue procurement and handling remain inseparable from downstream interpretation. A spatial assay cannot recover architectural information that was destroyed before sectioning. A poorly oriented posterior pole, a damaged macular center, or a section that misses the relevant border can limit the study before sequencing begins. The most sophisticated analytic pipeline cannot compensate for an absent anatomical target.

The same principle applies to post-mortem pathology. Donor eyes are not standardized experimental objects. They arrive with variation in disease duration, systemic history, treatment exposure, post-mortem interval, fixation or freezing conditions, and tissue integrity. Spatial transcriptomics preserves local context within a specimen, but comparisons across specimens still require careful normalization and study design.

Bulk RNA-seq often has an advantage when the study requires larger numbers of samples or deeper sequencing across a defined tissue preparation. Differential expression workflows such as DESeq2 can support rigorous comparisons between groups, provided that tissue composition and donor variation are addressed. Spatial analysis adds anatomical detail, but it may involve fewer usable sections, more complex quality control, and greater sensitivity to section-level technical variation.

The right choice depends on the biological scale of the question:

• If the question concerns broad pathway differences across a well-defined donor cohort, bulk analysis may be efficient and statistically powerful.
• If the question concerns the molecular transition from drusen-associated RPE to atrophy border to lesion core, spatial profiling is the more appropriate primary lens.
• If the question concerns whether a pathway is RPE-intrinsic or driven by neighboring vascular and immune cells, the methods should be combined rather than substituted for one another.

Toward a layered map of geographic atrophy

The projected global burden of AMD—approximately 300 million affected people by 2040—makes molecular resolution more than an academic preference. But the scale of the disease does not simplify the tissue. It makes the need for precise donor models more urgent, because a broad clinical category can contain several cellular trajectories.

Spatial transcriptomics offers a way to divide that category according to anatomy and cell state. In a fixed frozen human macula section, the researcher can place gene-expression data beside the lesion geometry rather than reconstructing geography from a homogenized average. This creates a foundation for comparing:

• the molecular profile of RPE overlying drusen with RPE farther away;
• the atrophy border with the established lesion core;
• macular and peripheral RPE;
• choriocapillaris regions beneath different retinal states;
• donor specimens with similar morphology but divergent transcriptional programs.

The next step is not simply higher resolution. It is better integration.

A more complete donor retina spatial transcriptomics workflow would align histology, spatial gene expression, cell-type references, and donor metadata at the level of individual pathological regions. Bulk RNA-seq would remain useful for deeper cohort-level profiling and confirmation across larger sample sets. The two approaches could then operate as different magnifications of the same pathology: bulk analysis establishing the broad molecular terrain, spatial analysis identifying where the critical events occur.

That layered approach also keeps the limitations visible. Spatial transcriptomics does not restore damaged tissue, reverse geographic atrophy, or convert a post-mortem section into a living disease model. It provides a molecular map of preserved tissue architecture. The map is powerful precisely because it remains constrained by what the donor specimen retains.

The unresolved question sits at the lesion border. When a surviving RPE cell, an activated microglial population, and a compromised choriocapillaris occupy the same narrow pathological zone, how much spatial resolution is enough to assign the initiating signal rather than merely the neighboring response? Until that boundary can be resolved consistently across donor eyes, geographic atrophy will remain a pathology whose molecular sequence is visible—but not yet completely separable.