Retinal cell heterogeneity: mapping donor tissue via scRNA-seq

The largest integrated single-cell retinal datasets have now profiled nearly 4 million individual cells from 125 human donors and identified more than 130 retinal cell types.

That scale changes the procurement problem. A donor eye is no longer assessed only by anatomical integrity or suitability for transplantation. Its value for ophthalmic multiomics depends on a chain of variables: anatomical origin, post-mortem interval, handling sequence, dissociation strategy, RNA preservation, and the ability to connect molecular output to donor metadata. Single-cell RNA-seq can resolve cellular heterogeneity, but it cannot erase degradation kinetics introduced before sequencing begins.

The scale of cellular diversity

Bulk RNA sequencing averages expression across the tissue. That average is useful for detecting broad disease-associated signals, but it suppresses the distribution that generates them. A gene may appear moderately expressed in a bulk retinal sample because it is highly active in one rare population and nearly absent elsewhere. The average does not reveal that structure.

Single-cell RNA-seq and single-nucleus RNA-seq separate the signal into cellular or nuclear profiles. This makes it possible to distinguish retinal ganglion cells from photoreceptors, Müller glia from astrocytes, vascular populations from perivascular cells, and multiple interneuron states that would otherwise collapse into broad categories.

The current cell-type inventory is not a fixed biological census. It is a function of sampling depth, donor composition, anatomical region, tissue condition, platform chemistry, and annotation strategy. Still, several large datasets establish the scale:

• An integrated Human Cell Atlas dataset analyzed nearly 4 million retinal cells from 125 human donors and identified more than 130 distinct cell types.
• A separate integrated atlas spanning 48 donors identified more than 90 retinal cell types.
• The rarest population in that analysis represented approximately 0.01% of total retinal cells.
• Profiling of foveal and peripheral retina from seven adult human donors mapped 58 cell types across six major cellular classes.
• A fresh-versus-post-mortem comparison profiled 106,829 cells and identified a novel ELF1-Cone subtype in the analyzed material.

The rare-population figure is operationally important. A cell type present at approximately 0.01% is not merely a statistical footnote. It is a throughput requirement. If the target population is sparse, low cell recovery, uneven dissociation, or selective loss of fragile cells can remove it from the final matrix entirely. The resulting atlas may remain technically clean while being biologically incomplete.

An atlas is constrained less by the number of clusters it produces than by the populations the procurement pipeline fails to preserve.

This is why donor tissue procurement and sequencing design cannot be treated as separate workstreams. The sequencing platform observes the material delivered to it. It does not observe the original retina.

What single-cell profiling actually maps

The term cell type is often used as if it describes a stable object. In practice, single-cell profiling maps several overlapping dimensions:

1. Lineage identity. The broad biological class to which a profile belongs, such as photoreceptor, retinal ganglion cell, interneuron, glial, vascular, or immune.

2. Molecular subtype. A narrower transcriptional program within that class.

3. Cell state. A condition associated with stress, inflammation, injury, metabolic change, maturation, or technical handling.

4. Spatial origin. The anatomical region from which the material was procured, including foveal, macular, or peripheral retina.

5. Technical state. Features introduced or amplified by dissociation, nuclear isolation, post-mortem delay, or RNA degradation.

These dimensions can overlap. A post-mortem stress program may look like a disease-associated state. A region-specific subtype may be mistaken for a donor-specific effect if anatomical provenance is not recorded. A damaged photoreceptor may cluster separately because of RNA loss rather than because it represents a new biological population.

For donor eye multiomics, the metadata pipeline therefore has the same analytical significance as the sequencing matrix. At minimum, a usable dataset needs to preserve the relationship between molecular profiles and:

• donor identity and relevant clinical context;
• anatomical region;
• procurement and processing timestamps;
• fresh or post-mortem status;
• tissue dissociation or nuclear isolation workflow;
• sequencing modality;
• cell or nucleus recovery;
• filtering and annotation decisions.

Without that linkage, downstream users may still obtain a visually coherent embedding. They will have less ability to determine whether the observed structure is biological, regional, donor-specific, or pre-analytical.

Fovea and periphery are different sampling grids

The retina is spatially organized. A foveal sample is not interchangeable with a peripheral retinal sample, even when both are labeled simply as retina.

In profiling from seven adult human donors, foveal and peripheral tissue yielded 58 cell types across six major classes. The dataset also showed that more than 90% of human retinal cell types had transcriptomic counterparts in cynomolgus macaque. Those findings support cross-species comparison, but they do not justify collapsing all human retinal regions into a single reference profile.

The anatomical question comes first: what tissue was actually procured? A broad label such as macula or peripheral retina is not sufficient if the study depends on regional gene expression. The distance from the foveal center, inclusion of adjacent layers, and separation of neural retina from other ocular structures can alter the cellular composition of the sample.

This has direct consequences for retinal functional genomics. A gene associated with cone specialization, synaptic transmission, or metabolic support may show different apparent activity depending on the proportion of foveal and peripheral material. A disease study that does not control for spatial origin can convert anatomical variation into a false disease signal.

Regional composition versus molecular subtype

Several analytical errors recur when regional sampling is poorly resolved:

• Region is interpreted as donor effect. If one donor contributes foveal tissue and another contributes peripheral tissue, clustering may reflect anatomy rather than individual biology.
• Rare populations are diluted. A cell type concentrated in one region may disappear when samples are pooled without region-aware balancing.
• Cell proportions are treated as expression changes. A higher bulk signal may result from more cells of a given type rather than increased transcription within those cells.
• Disease and location are confounded. If diseased samples come disproportionately from one anatomical zone, the atlas cannot cleanly separate pathology from spatial organization.

A robust atlas must therefore track the sample as a coordinate in a larger grid: donor, region, layer, processing state, and assay modality. The cell barcode is not enough. The tissue origin behind the barcode determines how the profile can be interpreted.

Post-mortem delay changes the transcriptomic signal

Post-mortem interval is not a passive timestamp. It is an active experimental variable. RNA degradation, cellular stress, membrane disruption, and loss of recoverable material can alter the dataset before library preparation begins.

A study of 20,009 cells from three donor eyes observed a marked reduction of MALAT1 transcript expression in rod photoreceptors as post-mortem intervals increased. MALAT1 is commonly abundant and is often used as a broad indicator of transcriptomic content. A reduction in its signal is therefore not a minor technical fluctuation. It indicates a shift in the molecular state of the recovered cells.

The consequences extend beyond individual genes. A longer interval can alter:

• the number of cells that survive dissociation;
• the number of transcripts captured per cell;
• the relative visibility of fragile populations;
• mitochondrial and stress-associated signals;
• the apparent separation between closely related subtypes;
• trajectory and RNA velocity calculations.

Fresh and post-mortem samples are not equivalent inputs for dynamic inference. In a comparison of fresh human donor eyes and post-mortem material, post-mortem processing significantly altered trajectory and RNA velocity calculations. This is a critical limitation for studies that attempt to infer developmental, regenerative, or disease-progression paths from static tissue.

RNA velocity is especially sensitive to assumptions about transcript abundance and processing states. When post-mortem handling changes those states globally or selectively, the direction and strength of inferred transitions can become artifacts of procurement latency. A computational trajectory may still be mathematically valid. It is not automatically a biological timeline.

The absence of a universal procurement time limit across all single-cell retinal biobanks is also important. Different facilities may use different workflows, tissue priorities, and quality thresholds. A dataset should therefore not be treated as transportable merely because both studies use the same sequencing modality.

scRNA-seq and snRNA-seq are complementary, not interchangeable

Single-cell RNA-seq profiles intact cells after tissue dissociation. Single-nucleus RNA-seq profiles nuclei isolated from tissue. The distinction affects the molecular content captured, the tissue-processing burden, and the types of populations most likely to be retained.

Intact-cell workflows can provide rich transcriptomic information from cellular material, but retinal tissue is structurally complex and contains populations that may be fragile during dissociation. Nuclear workflows can be advantageous when intact-cell recovery is difficult or when archived, frozen, or otherwise compromised tissue must be profiled. They also produce a different molecular representation, with a greater emphasis on nuclear transcripts and pre-mRNA-related signal.

That difference matters during atlas integration. A cell atlas and a nucleus atlas may use similar labels while observing non-identical transcriptomic states. Their clusters can be aligned computationally, but alignment is not proof that the assays measure the same biological layer.

The choice should be driven by the target question:

• For broad cellular composition, either modality may be informative if tissue provenance and quality are well controlled.
• For fragile retinal populations, the workflow must be evaluated for selective loss during dissociation.
• For post-mortem or structurally compromised material, nuclear profiling may offer a more stable route to molecular recovery, but the resulting signal must be interpreted as a nuclear transcriptomic profile.
• For isoform-level questions, neither modality should be assumed to provide complete coverage across all cell types.
• For spatial comparisons, the loss of anatomical context during dissociation or nuclei isolation must be recorded and compensated through sampling design.

The term “single-cell” is therefore insufficient as a quality descriptor. A study needs to specify whether the analytical unit is a whole cell or a nucleus, how the material was released from the tissue, and which populations were likely to be affected by that release process.

The dissociation bottleneck

Retinal tissue is not a homogeneous suspension waiting to be sequenced. It is a layered neural structure with different cell sizes, membrane properties, physical resilience, and RNA content. Dissociation creates a selection pressure.

The resulting dataset is shaped by at least four linked processes:

1. Mechanical release. Cells are separated from the tissue architecture. Excessive force can damage fragile populations; insufficient release can lower recovery.

2. Enzymatic exposure. Dissociation conditions can change membrane integrity and alter the duration for which cells remain outside their native environment.

3. Filtering and cleanup. Debris, aggregates, dead cells, and damaged material are removed, but aggressive cleanup can also remove biologically relevant populations.

4. Library capture. The final matrix reflects only the material that survives all upstream steps and generates a usable library.

These stages create a difference between biological abundance and observed abundance. A population may be genuinely rare, technically under-recovered, or both. The 0.01% population identified in an integrated atlas illustrates the statistical edge of the problem: rare-cell detection requires enough total throughput to distinguish a true low-frequency population from contamination, doublets, or random capture.

A practical tissue-procurement record should distinguish between:

• cells received from the tissue;
• cells passing initial viability or integrity filters;
• cells entering the library workflow;
• cells retained after quality filtering;
• cells assigned to a final annotated population.

Without those counts, a final atlas can conceal attrition. The number of cells in the published matrix is not the same as the number of cells originally available in the donor tissue.

Annotation is a data pipeline, not a final label

Cell-type annotation is often presented as the final interpretive stage. In reality, it is another transformation of the data. Marker-based labels depend on reference atlases, gene detection, batch structure, and the resolution selected by the analyst.

A broad label such as retinal ganglion cell may be reliable at one level and insufficient at another. A disease study may require separation of ganglion-cell subtypes, stress states, or regional programs. If the source data do not preserve enough transcripts, the analyst may either over-cluster technical noise or under-cluster meaningful biology.

The presence of a novel subtype, such as the reported ELF1-Cone population in fresh-versus-post-mortem profiling, should therefore be evaluated against several questions:

• Does the population recur across donors?
• Is it present across compatible anatomical regions?
• Does it persist across processing conditions?
• Is its signal driven by a coherent transcriptional program?
• Could it reflect a stress, degradation, or dissociation-associated state?
• Can it be connected to independent molecular or spatial evidence?

A cluster is an observation. It becomes a biological category only after its stability and provenance are established.

Cross-species conservation has a defined scope

The comparison between human and cynomolgus macaque retinal profiles found that more than 90% of human retinal cell types had transcriptomic counterparts in the non-human primate dataset. This is a strong basis for comparative retinal biology. It supports the use of macaque models for studying conserved cellular organization and for prioritizing candidate pathways.

It does not establish complete equivalence.

A corresponding cell type can differ in:

• expression magnitude;
• regional distribution;
• disease susceptibility;
• developmental timing;
• splice-isoform usage;
• immune interaction;
• response to injury or treatment.

The comparison is most useful when it is framed as a mapping problem. Which human populations have a plausible primate counterpart? Which genes are conserved within those populations? Which states are species-specific? Which regional differences remain unresolved?

This distinction is particularly relevant to glaucoma genetics, inherited retinal dystrophy sequencing, and optic neuropathy bioinformatics. A conserved cell class can support mechanistic translation, while a species-specific regulatory program may determine whether a model reproduces the human disease phenotype.

Cross-species atlas integration also depends on consistent anatomical sampling. If human foveal tissue is compared with macaque material from a different retinal region, the apparent species effect may include a hidden spatial effect. Region must be controlled before conservation can be quantified cleanly.

Designing a donor tissue dataset that remains usable

A high-value ocular biospecimen is not defined only by whether it yields a library. It is defined by whether later users can determine what the library represents.

For single-cell RNA-seq donor retinal cell heterogeneity, the procurement and data architecture should preserve several linked layers:

• Donor layer: identity, clinical metadata where available, and relevant biological context.
• Clock layer: timing from death or tissue availability through enucleation, dissection, fixation, freezing, or nuclei isolation.
• Anatomical layer: foveal, macular, peripheral, retinal, or other ocular origin with clear boundaries.
• Processing layer: intact-cell or nuclear workflow, dissociation conditions, and major handling transitions.
• Assay layer: sequencing modality, library preparation, and analytical unit.
• Quality layer: recovery, filtering, transcriptomic content, and evidence of degradation or stress.
• Annotation layer: cell-type labels, confidence, reference atlas, and unresolved populations.

This is not administrative overhead. It is the minimum structure required to distinguish biology from logistics.

A donor eye can generate a technically successful library and still produce a low-value dataset if the processing timeline is incomplete or the anatomical origin is ambiguous. Conversely, a post-mortem sample with known latency, well-defined region, and transparent quality metrics may be highly useful for comparative analysis, even if it is unsuitable for questions that depend on fresh-state transcriptional dynamics.

Tissue quality is not a binary property. It is a vector of provenance, latency, anatomical resolution, molecular preservation, and assay compatibility.

The correct decision is therefore not whether a specimen is simply good or bad. It is whether the specimen is fit for the intended inference.

Where the field is still constrained

The existing atlases have expanded the known retinal cellular landscape, but they do not close the catalog. The complete diversity of splice isoforms across all identified cell types remains unresolved, particularly when comparing living-state biology with post-mortem material. This limits conclusions about transcript architecture that go beyond gene-level expression.

The field also lacks a universally adopted procurement window for all single-cell retinal biobanks. That does not make post-mortem datasets unusable. It means that latency must remain an explicit covariate rather than being hidden inside a general sample-quality label.

Several conclusions are already stable:

1. Human retinal tissue contains substantially more cellular diversity than bulk assays can resolve.

2. Large-scale single-cell atlases can identify more than 130 retinal cell types across major neuronal, glial, vascular, and immune lineages.

3. Rare populations require high throughput and careful attrition tracking.

4. Foveal and peripheral retina cannot be treated as interchangeable sampling locations.

5. Post-mortem delay changes transcript abundance and can distort trajectory or RNA velocity inference.

6. Human and cynomolgus macaque retinal profiles show extensive cellular correspondence, but conservation is not equivalence.

7. Procurement metadata is part of the biological result, not an accessory record.

The practical assessment is strict. Single-cell RNA-seq can map donor retinal cell heterogeneity at unprecedented resolution, but resolution is conditional on the upstream logistics. A dataset with millions of cells is not automatically more informative than a smaller, better-documented cohort. The decisive variables remain traceability, spatial provenance, processing latency, and transparent separation of biological signal from technical state.

In ophthalmic multiomics, the bottleneck has moved upstream. Sequencing reveals the structure that the eye bank and tissue-processing pipeline have preserved. No downstream embedding can reconstruct a population that was lost before capture, and no computational correction can fully recover a trajectory distorted by undocumented post-mortem delay.