GeoMx Digital Spatial Profiler Enables Layer-Specific Spatial Transcriptomic Profiling of the Human Retina: A Proof-of-Concept Study from Post-Coronovirus Disease 2019 Donor Tissue

Abstract

Purpose

To generate layer resolved spatial transcriptomic profiles of the human retina using NanoString GeoMx Digital Spatial Profiler (DSP) and evaluate whether this approach can be successfully applied to resolve spatially localized transcriptional changes across retinal compartments in diseased retina.

Design

Study of human postmortem retinal tissue using spatial transcriptomics with cross-platform computational validation.

Subjects

Postmortem human retinal samples from donors with and without history of coronavirus disease 2019 (COVID-19).

Methods

Anatomically defined retinal layers were profiled using GeoMx DSP with layer-resolved region of interest selection. Data were validated using retinal layer markers, and concordance with published single-cell RNA sequencing (scRNA-seq) reference signatures was established. Differential expression and pathway-level comparisons were performed between donors with and without a history of COVID-19, with emphasis on spatial localization of changes across retinal layers.

Main Outcome Measures

(1) Concordance of GeoMx layer resolved expression with known photoreceptor and retinal pigment epithelium marker patterns and scRNA-seq reference signatures; (2) exploratory identification of candidate layer restricted gene expression differences in donors with a history of COVID-19, as a proof-of-concept application of the platform.

Results

GeoMx DSP captured biologically expected, layer-specific transcriptional changes across the human retina, including compartmentalized layer expression, and showed cross-platform concordance with published scRNA-seq reference signatures. In an exploratory comparison of donors with and without a history of COVID-19, candidate gene expression differences were observed across specific retinal layers, suggesting spatially nonuniform transcriptional patterns that warrant further investigation in larger independently validated cohorts.

Conclusions

Layer-resolved GeoMx DSP provides a robust approach for spatial transcriptomic analysis of human retina and enables scRNA informed interpretation through cross platform benchmarking. Exploratory application to donors with COVID-19 history identified spatially restricted transcriptional shifts across retinal compartments, illustrating the utility of layer resolved retinal spatial transcriptomics as a discovery framework for studying clinically relevant perturbations and informing future ocular disease and therapeutic research.

Financial Disclosure(s)

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

The human retina is a highly stratified neural tissue that performs essential visual processing functions. It consists of multiple distinct layers, each containing specialized cell types such as photoreceptors, bipolar cells, Müller glia, horizontal cells, amacrine cells, and retinal ganglion cells (RGCs).¹ Understanding the spatial organization of gene expression within these layers is critical for elucidating the molecular basis of normal retinal function, as well as the mechanisms underlying various retinal diseases.

Conventional transcriptomic techniques, such as bulk RNA sequencing, aggregate signals across all cell types and layers, thereby losing crucial spatial information. While single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cell-type-specific gene expression, it requires tissue dissociation, which can introduce technical artifacts and eliminate the spatial context necessary to interpret complex tissue biology.²

Spatial transcriptomics has become a powerful method to bridge this gap by maintaining tissue architecture while allowing high-resolution gene expression analysis. Among the available platforms, the NanoString GeoMx Digital Spatial Profiler (DSP) stands out for its compatibility with paraffin or cryosectioned tissue sections, its ability to target user-defined regions of interest (ROIs), and its high-plex profiling capacity.³ This makes it particularly suitable for postmortem human tissue studies, including rare or archived samples that are otherwise challenging to analyze.

Despite its increasing application in oncology and neuroscience, the use of GeoMx DSP in ocular research, particularly in human retina, remains limited.⁴ Given the laminar structure and cellular complexity of the retina, it is an ideal tissue to demonstrate the capabilities of spatial transcriptomic platforms.⁵

In this study, we sought to establish and validate the use of NanoString GeoMx DSP for spatial transcriptomic profiling of the human retina. We focused on identifying known layer-specific gene expression patterns and cross-validating our results with publicly available scRNA-seq data sets. To demonstrate the platform's utility in disease settings, we included retinal tissues from deceased patients with a history of coronavirus disease 2019 (COVID-19) to test the platform's reproducibility and performance under altered biological conditions. Our data demonstrate that the NanoString GeoMx DSP platform offers reliable, spatially resolved transcriptomic profiling in fixed human retinal tissues. Validation through comparisons with scRNA-seq and gene markers for RGC, inner nuclear layer (INL), photoreceptors, and retinal pigment epithelium (RPE) confirms its accuracy. Using tissue from patients with a history of COVID-19 as a representative biological perturbation highlights the platform’s potential for disease focused spatial transcriptomics studies. This study provides a generalizable framework for leveraging spatial transcriptomics in future retinal research and disease modeling.

Methods

Sample Collection and Processing

Pathology analysis was conducted with the approval of the Cleveland Clinic Institutional Review Board (IRB #20-755) as Exempt Research and the Institutional Biosafety Committee (IBC #2018). All work was completed in compliance with the Declaration of Helsinki. Research consented deceased donor posterior globes were purchased from the nonprofit eye bank Eversight and used according to the Eye Bank Association of America regulations and US Food and Drug Administration guidelines. Donor information and medical history are summarized in Figure 1B’. The eyes were fixed in 2% paraformaldehyde in Dulbecco phosphate-buffered saline (D-PBS). All eyes were acquired and preserved within ∼9.5 h after death. A 4-mm punch from an area close to the optic nerve within the macula (termed central) and from the periphery was flash-frozen in liquid nitrogen and processed for cryosectioning using an HM 505E cryostat (Microm) equipped with a CryoJane Tape-Transfer system (Leica). Sections from control eyes and donors with a history of COVID-19 (referred to as post-COVID-19 eyes) were cut and collected on the same slide; one slide contained central retinas, and the other slide contained peripheral retinas.

Figure 1.

Macroscopic, histological analysis, and experimental overview of the NanoString platform. (A’) Representative fundus images from control donors showing normal vascular patterns, fovea, and optic nerve (A-F). Fundus images from donors with prior COVID-19 frequently showed focal and diffuse hemorrhagic spots (black arrowheads) (G-L). General differences in retinal integrity across groups are attributed to postmortem imaging variables and are not considered a specific clinical feature of the infection. (B’) Summary table of donor demographics and clinical characteristics, including age, gender, ocular history, systemic conditions, and cause of death (COD), for control and post-COVID-19 cohorts. (C’) Representative retinal cryosections immunostained for calbindin (green), red/green opsin (red), and SYTO13 (blue), highlighting the RGC, INL, photoreceptor, and RPE layers. (D’) Schematic of the NanoString GeoMx DSP workflow showing layer guided ROI selection, UV tag release, and downstream transcriptomic quantification. (E’) A schematic image of the section through the eye to illustrate the representative ROIs within retinal layers. COVID-19 = coronavirus disease 2019; DSP = Digital Spatial Profiler; INL = inner nuclear layer; RGC = retinal ganglion cell; ROI = region of interest; RPE = retinal pigment epithelium; UV = ultraviolet.

Immunohistochemistry

In brief, the slides were washed in D-PBS, incubated with 0.1 μg/ml proteinase K (Qiagen, 19131) for 15 min at 37°C, followed by blocking with 1% bovine serum albumin+0.1% Triton-X in D-PBS for 1 h. The slides were then incubated with the primary antibody (mouse calbindin, Abcam, ab82812, dilution, 1:100, Rabbit R/G opsin, Millipore, AB5405, 1:400) in blocking solution overnight at 4°C, washed 3 times in D-PBS, and incubated with the secondary Antibody (Alexa Fluor 488 anti-rabbit and Alexa Fluor 594 anti-mouse, Thermofisher) for 1 hour at room temperature, followed by GeoMx Nuclear Staining kit (121300303).

GeoMx DSP Workflow

Labeled retinal sections were imaged for morphology markers, after which an ROI ranging from 3000 to 90 000 μm² was drawn to define the retinal layer of interest. RNA probes targeting the Whole Transcriptome Atlas were hybridized. Barcoded oligonucleotides were ultraviolet-cleaved, collected from selected ROIs, and sequenced. Finally, raw counts were Q3-normalized for analysis.⁶

GeoMx Sample Information

A total of 144 areas of illumination (hereafter referred to as ROIs) were profiled using the NanoString GeoMx DSP platform. Each ROI corresponded to one of six anatomically defined retinal layers: the RGC layer, inner plexiform layer, INL, outer plexiform layer, photoreceptor layer, and RPE, collected from both the central and peripheral regions of the retina in six control and six donors with prior COVID-19 infection. This design enables layer-, region-, and disease-specific comparison of spatial gene expression.

Differential Expression Analysis

To account for technical artifacts and intraexperimental variability across multiple regions profiled per tissue section, Q3 normalized expression data were analyzed using a linear mixed-effects model. The model evaluated group-level fixed effects while treating slide as a random intercept to control slide to slide technical variation. Random slopes were omitted to prevent overparameterization and preserve statistical power. To control the type I error rate across the data set, both nominal and false discovery rate (FDR)–adjusted P values are reported, with multiple-testing corrections calculated via the Benjamini–Hochberg procedure. Comparisons between control and post-COVID-19 specimens were conducted independently for each retinal layer. Due to the limited number of samples per group and the exploratory nature of this study, nominal P values (P < 0.05) were used to identify suggestive candidate genes that require independent validation. Fold-change thresholds were also applied to prioritize biologically relevant changes. Given the cohort composition and sample size, inclusion of additional covariates such as age and sex was not feasible without overfitting the model. A complete list of candidate genes is provided in the Supplementary Material, available at www.ophthalmologyscience.org.

Gene Set Enrichment Analysis

Gene Set Enrichment Analysis was performed using the clusterProfiler package (version 4.14) in R. A ranked gene list generated from the differential expression analysis was used as input for enrichment analysis. Gene Ontology biological process, cellular component, and molecular function databases were examined using the gseGO function, while over-representation analysis was performed using the enrichGO function. Statistical significance was assessed using the Benjamini–Hochberg multiple-testing correction, and both nominal and FDR-adjusted P values are reported. The complete enrichment results, including normalized enrichment scores, nominal P values, and FDR-adjusted P values, are provided in the Supplementary Material.

RPE Validation by Cross-Platform Integration of Single-Cell RNA Sequencing

Log normalized scRNA-seq expression matrices from four human retinal donors (GSM4037981, GSM4037982, GSM4037984, and GSM4037985) representing macular and peripheral regions were used to validate RPE expression profiles in GeoMx DSP data. Cells annotated as RPE were filtered from each donor data set, and log-normalized expression values were averaged per donor to create RPE pseudobulks (gene × donor matrices). GeoMx DSP expression data and metadata were imported, and ROIs corresponding to RPE tissue from control samples on slides were selected. GeoMx data were analyzed in log-transformed counts per million space, while scRNA-seq data were analyzed in log-normalized space, ensuring cross-platform comparability without additional normalization. Shared genes were identified, and expression matrices were standardized per gene (row wise Z-score) to remove scale differences. A curated RPE signature gene list (BEST1, RPE65, RLBP1, KRT8, KRT18, TYR, RGR, and CRALBP) was used for targeted validation. For each gene, GeoMx and scRNA-seq expression distributions were compared. Expression concordance was visualized using violin and dot plots, and cross-platform concordance was visualized using Pearson raw correlation heatmaps.

Photoreceptor (Rod and Cone) Validation by Cross-Platform Integration of scRNA-seq and GeoMx DSP Data

Raw scRNA sequence counts from perifovea (GSM5175796-99) and fovea (GSM5175792-95) were processed in R to generate rods and cones photoreceptor pseudobulks, respectively. Cells annotated as rods and cones were filtered, and gene counts were summed per donor, normalized to counts per million (CPM), and log transformed using log_e (CPM + 1). GeoMx DSP expression data (log-transformed counts per million) and GeoMx metadata were imported, and ROIs corresponding to photoreceptor cells from control samples were selected. GeoMx data, Q3 normalized and log-scaled, were harmonized by converting gene names to uppercase and collapsing duplicate probe entries. Shared genes between data sets were identified, and expression matrices were standardized per gene (row wise Z-score) to remove platform dependent scale differences. A curated rod photoreceptor gene set (RHO, GNAT1, NR2E3, PDE6A, PDE6B, PDE6G, SAG, GNGT1, RBP3, CNGA1, CNGB1, ROM1, and NRL) and cone (OPN1LW, OPN1MW, OPN1SW, GNAT2, PDE6C, PDE6H, ARR3, GNB3, GUCA1A, THRB, RXRG, GUCA1C, and ISL2) were used for targeted comparison. Expression distributions were visualized with violin and dot plots, and cross-platform concordance was assessed using Pearson correlation heatmaps on both raw data.

scRNA-seq Reference Data Sets, Anatomical Mapping and Cell Deconvolution

Q3 normalized GeoMx DSP expression matrices and associated ROI metadata were imported into R. Regions of interest were aligned to metadata using the ROI label field. Deconvolution reference signatures were derived from healthy scRNA-seq data sets and applied equally to both the control and post-COVID-19 GeoMx ROIs. As a single cell reference, we combined four human retinal scRNA-seq donor data sets (foveal retina GSM5175792-GSM5175795) and four RPE scRNA-seq donor data sets (macular and peripheral RPE: GSM4037981/GSM4037982 and GSM4037984/GSM4037985, respectively). For each donor file, we harmonized gene symbols and collapsed duplicated gene columns. GeoMx expression data (genes × ROIs) were then intersected with the full-gene-matched signature matrix and subjected to deconvolution using the SpatialDecon function with the scRNA derived signature matrix. For photoreceptor ROIs, in addition to the full-gene-matched signature matrix, we applied a curated panel of 15 rod enriched and 15 cone enriched marker genes to enhance photoreceptor specificity. For each cell type, we visualized the estimated proportions using violin plots by sample type (control vs. post-COVID-19) and performed two-sided Wilcoxon rank sum tests to compare cell-type proportions between conditions. Resulting P values were adjusted for multiple testing across cell types using the Benjamini–Hochberg FDR. These analyses were performed across all ROIs, pooling central and peripheral regions, and all retinal layers represented in the data set. SpatialDecon (v1.13.2), Matrix (v1.6.5), ggplot2 (v3.5.2), and ggpubr (v0.6.1) were used for the analysis.

Results

Fundus Imaging and Spatial Characterization of Postmortem Human Donor Eyes

Fundus macroscopic imaging provides an overview of retinal morphology in human donor eyes prior to histopathological and molecular analyses. Representative fundus photographs from postmortem control donor eyes and from patients with prior COVID-19 infection were obtained to assess structural integrity and gross morphology (Fig 1A–F). Anatomical landmarks, including the optic nerve head and foveal region (black arrows), were clearly visible in all donors. Several eyes from post-COVID-19 donors exhibited hemorrhagic spots (Fig 1G–L, black arrowheads), which were not detected in the control fundus images (Fig 1A–F). Retinal sections from central and peripheral regions were stained with calbindin (green) and with red/green opsins (red) to mark cone photoreceptors; nuclei were counterstained with SYTO13 (blue). Autofluorescence due to lipofuscin granules was used to highlight the RPE layer. The cone markers, calbindin and R/G opsin, were used to identify the cones, while the nuclear labeling was used to define the RGC (RGC), INL, and outer nuclear layer. The inner plexiform layer and outer plexiform layer were spatially demarcated (Fig 1C’). Regions of interest corresponding to the RGC, INL, inner plexiform layer, outer plexiform layer, Photoreceptor, and RPE layers were selected. Following ROI selection, ultraviolet-mediated oligonucleotide tag collection was performed to quantify gene expression within each layer. This integrative workflow linking fundus morphology, histological layer identification, and spatially resolved molecular profiling enabled a comprehensive evaluation of retinal architecture and transcriptional changes between post-COVID-19 and control human eyes.

Quality Assessment and Normalization Strategy for GeoMx Spatial Transcriptomic Data

To ensure reproducibility and comparability of spatial transcriptomic profiles, we conducted rigorous quality control and normalization. Q3 normalization was selected for downstream analysis because it consistently outperformed other methods, such as negative probe and housekeeping normalization. (Fig 2A). To assess sequencing depth and sensitivity, sequencing saturation was calculated as 1 (percent unique reads). Sequencing saturation exceeded 50% for most ROIs, confirming sufficient sequencing depth for low abundance transcripts (Fig 2B). Across 18 676 assayed genes and 144 areas of illumination, 11 483 genes were expressed in at least 10% of areas of illumination and 5054 genes in at least 50%, illustrating strong detection sensitivity across most regions despite biological heterogeneity in expression levels among retinal layers (Fig 2C).

Figure 2.

Quality control metrics and normalization performance of GeoMx spatial transcriptomic data. (A) Workflow illustrates data normalization. Q3 normalization was selected for downstream analysis because it provided consistent performance across heterogeneous retinal layers compared to negative probe and housekeeping normalization. (B) Sequencing saturation for each ROI. (C) Number of genes expressed above LOQ² (second limit of quantitation, calculated as geometric mean × geometric standard deviation² of negative probes) across all the retinal cells, illustrating expression detection breadth. (D) Signal-to-noise ratio (=counts/LOQ) distribution by sample type, reflecting assay sensitivity and specificity across ROIs. (E) Principal component analysis demonstrated moderate clustering by sample group, highlighting notable differences between the molecular profiles of control and post-COVID-19 samples, despite some within-group variance. (F) Pearson correlation analysis of central and peripheral control RPE ROIs, labeled by donor identifier (ID), demonstrates conserved RPE transcriptomic profiles across retinal regions. Scale reflects the full range of observed values. Self-correlation values (r = 1.0) are shown for reference but were excluded from interpretation of regional conservation analyses, as they do not provide biologically informative comparisons. AOI = area of illumination; COVID-19 = coronavirus disease 2019; IPL = inner plexiform layer; OPL = outer plexiform layer; ROI = region of interest; RPE = retinal pigment epithelium; PCA = principal component analysis; QC = quality control.

Additionally, the strong signal-to-noise ratio and broad gene detection across layers further validated assay sensitivity (Fig 2D). Finally, principal component analysis revealed modest clustering trends, particularly for the RPE and photoreceptors, but otherwise similar global expression profiles across principal components between post-COVID-19 and control samples (Fig 2E), indicating subtle disease-associated transcriptomic differences. To further evaluate regional effects, principal component analysis was performed separately for central and peripheral retinal ROIs. Across most layers, control and post-COVID-19 samples showed partial overlap, suggesting that disease effects are modest relative to interdonor and layer-specific variability. Samples are labeled by donor identity to match the clinical information in Figure 1B (Supplementary Figure 4, available at www.ophthalmologyscience.org). Pearson correlation analysis of control RPE ROIs across central and peripheral retina showed moderate similarity between regions, indicating partial preservation of RPE transcriptional profiles with regional variability (Fig 2F). Regions of interest are labeled with donor identity, allowing direct correspondence with clinical metadata (Fig 1B).

Cross-Platform Validation of GeoMx Spatial Transcriptomic Data Using Available scRNA-seq Data Sets

To validate the GeoMx spatial transcriptomic profiles, we compared them with publicly available scRNA-seq data sets from healthy human eyes (RPE/choroid data [GSE135922] and retinal data [GSE169047]). Retinal pigment epithelium and photoreceptor ROIs were matched to corresponding scRNA-seq pseudobulk profiles. Figure 3A, B outline the workflow for cross-platform comparison, including data set normalization, gene selection, and correlation analysis. Both data sets were independently normalized and transformed using gene-wise Z-scores to harmonize expression scales before correlation analyses. Heatmaps were generated using raw normalized expression values (GeoMx Q3 normalized and scRNA CPM) to preserve quantitative signal intensity. Cross-platform comparison showed modest concordance between central and peripheral RPE profiles. This supports partial agreement in broad transcriptional features but also highlights expected differences attributable to region, donor heterogeneity, differences in sampled cellular composition, and platform-specific sensitivity (Fig 3C). Photoreceptor ROIs showed high correlation with cone pseudobulk profiles (Fig 3D) and rods (Supplementary Figure 1, available at www.ophthalmologyscience.org), validating accurate layer-specific expression without additional Z scaling. Marker gene analysis further demonstrated consistent expression patterns across platforms for RPE and cone photoreceptor signature genes (Fig 3E, F). Collectively, these results demonstrate that GeoMx ROIs preserve authentic spatial and cell-type-specific transcriptional profiles, supporting the accuracy of layer annotation and biological relevance of the data.

Figure 3.

Cross-platform validation of GeoMx spatial transcriptomic data using scRNA-seq. (A-B) Workflow illustrates normalization and integration workflow between GeoMx and scRNA-seq data sets for RPE and photoreceptor layers. (C-D) Raw scale Pearson correlation heatmaps comparing GeoMx RPE ROIs from control eyes (central and peripheral regions) with scRNA-seq pseudobulk RPE data sets (C), GeoMx photoreceptor ROIs with cone pseudobulk profiles from four scRNA-seq donors (D). Regions of interest are labeled by donor identity, demonstrating modest expression concordance for cone-specific genes. (E and F) Comparison of RPE and cone-specific gene expression following Z normalization across data sets, remains consistent, thus confirming layer-specific expression and cross-platform reproducibility. BH = Benjamini–Hochberg; CPM = counts per million; DSP = Digital Spatial Profiler; FDR = false discovery rate; ROI = region of interest; RPE = retinal pigment epithelium; scRNA-seq = single-cell RNA sequencing.

To assess whether differential expression results reflect true transcriptional regulation rather than cell composition differences, we performed cell-type deconvolution using scRNA-seq derived reference signatures. Deconvolution was performed separately for the ganglion cell layer, INL, photoreceptor layer, and RPE ROIs. The deconvolution profiles closely matched the cellular identities. Ganglion cell layer ROIs showed predominant contributions from RGCs, with secondary representation from Müller glia and vascular endothelial cells (Fig 4A). The INL ROIs were enriched for amacrine, bipolar, and Müller glial cells (Fig 4B), while photoreceptor ROIs were enriched for rod and cone signatures (Fig 4C). We performed an additional deconvolution using 15 rod- and 15 cone-specific marker genes. This showed a higher proportion of photoreceptors (>80–90%), consistent with the expected composition of these regions (Supplementary Figure 3, available at www.ophthalmologyscience.org). Retinal pigment epithelium ROIs showed strong enrichment for RPE cells, consistent with accurate segmentation (Fig 4D). Minor contributions from adjacent cell types were occasionally detected in each layer, which is expected given the multicell composition of each ROI, tissue curvature, and partial overlap of neighboring layers.

Figure 4.

Cell-type proportion analysis across retinal layers in control and post-COVID-19 eyes. (A-D) Cell-type deconvolution comparing control and post-COVID-19 eyes across layers, showing stable cell-type proportions and expected enrichment of layer-specific cell populations. BC = bipolar cells; COVID-19 = coronavirus disease 2019; INL = inner nuclear layer; RGC = retinal ganglion cell; ROI = region of interest; RPE = retinal pigment epithelium.

Importantly, no significant differences in cell-type proportions were observed between control and post-COVID-19 ROIs within any retinal layer. Together, these cross-platform comparisons and layer-specific deconvolution analyses confirm that GeoMx ROIs accurately represent the intended retinal layers and that disease associated transcriptional changes arise from molecular dysregulation.

Differential Expression and Pathway Enrichment in Post-COVID-19 Retina

Building on cross-platform validation of GeoMx with scRNA-seq data, we next examined whether long-term transcriptional changes were detected in the post-COVID-19 eyes. The analysis was conducted across retinal layers in the central region (Fig 5) and the peripheral region (Supplementary Figure 2, available at www.ophthalmologyscience.org). Differential expression analysis was performed for the ganglion cell layer, INL, photoreceptor layer, and RPE. Volcano plots for each layer, highlighting significantly altered transcripts (P < 0.05), are shown, with upregulated transcripts in post-COVID-19 retinas indicated in orange and downregulated transcripts in green (Figs 5A–D). Differential expression analysis was performed using nominal P values. Adjusted P values were not used as primary cutoffs because of the limited sample size and the exploratory nature of the analysis. In the ganglion cell layer, candidate gene changes associated with mitochondrial and ribosomal activity, including COX4I1, SERP2, RPL24, and RPS4X, were upregulated in post-COVID-19 eyes, suggesting increased metabolic and translational activity. Gene Set Enrichment Analysis confirmed enrichment of mitochondrial respirasome and nicotinamide adenine dinucleotide dehydrogenase complex pathways, while semaphorin receptor complex was significantly downregulated, suggesting potential disruption of synaptic signaling (Fig 5A, E). In the INL, candidate upregulated transcripts included MYLK4, CXCR4, ZNF720, and CRAMP1, indicating activation of stress response and signaling pathways in interneuronal and glial populations (Fig 5B, F). The photoreceptor layer showed pronounced transcriptional changes linked to mitochondrial and endoplasmic reticulum stress. Upregulated candidate genes, such as ATP5MJ, PDIA3, PABPC1, VBP1, and PLCL2, are associated with protein folding, mitochondrial function, and photoreceptor homeostasis. Gene Set Enrichment Analysis revealed upregulation of pathways related to mitochondrial adenosine triphosphate synthesis, oxidative phosphorylation, electron transport, and adenosine triphosphate biosynthetic processes, accompanied by downregulation of processes involved in sensory perception of light, mitotic nuclear division, and cilium organization and assembly, collectively suggesting that early molecular signatures of photoreceptor stress and mitochondrial dysfunction can be detected in post-COVID-19 retinas (Fig 5C, G). The RPE layer showed increased expression of FAM107A, RPS4X, ADAM9, OTX2, TMED10, and CD63. Pathway enrichment similarly indicated increased mitochondrial respirasome activity, suggesting enhanced metabolic demand and potential oxidative stress in RPE cells from post-COVID-19 eyes (Fig 5D, H).

Figure 5.

Differential gene expression and pathway enrichment within retinal layers in post-COVID19 and control eyes of central retina. (A-D) Volcano plots for central region ganglion cell layer, INL, photoreceptor, and RPE layers, respectively. Upregulated genes in post-COVID-19 (Disease, orange) and downregulated genes (green) are shown. (E-H) Gene Set Enrichment Analysis highlighting upregulation of mitochondrial pathways and downregulation of semaphorin receptor signaling, cilium organization, and sensory perception pathways. (I) Cell-type deconvolution comparing control and post-COVID-19 eyes across layers, showing stable overall cell-type proportions and expected enrichment of layer-specific cell populations. ATP = adenosine triphosphate; BC = bipolar cells; COVID-19 = coronavirus disease 2019; FC = fold change; GSEA = Gene Set Enrichment Analysis; INL = inner nuclear layer; NADH = nicotinamide adenine dinucleotide; RGC = retinal ganglion cell; RPE = retinal pigment epithelium.

To confirm that these transcriptional differences were not due to differences in cellular composition between post-COVID-19 and control retinas, we performed cell-type deconvolution across all layers. Control and post-COVID-19 ROIs showed highly overlapping distributions of cell-type proportions, with no significant differences after multiple-testing correction (all FDR-adjusted P > 0.05), except for a change in microglia (P < 0.032). This analysis confirms that the observed transcriptional changes, such as mitochondrial upregulation, photoreceptor stress, and ciliary dysfunction, reflect true gene regulation rather than shifts in the underlying cellular population.

Discussion

Our study establishes the NanoString GeoMx DSP platform as a robust and reproducible method for spatial transcriptomic profiling of the human retina. By distinguishing layer-specific gene expression profiles and validating them against independent data sets, we demonstrate its biological relevance and technical reproducibility. The primary contribution of this work lies in the methodological validation of GeoMx DSP for human retinal tissue, offering a spatially informed framework applicable to a broad range of retinal diseases. To illustrate the platform’s utility in disease-relevant tissue, we included post-COVID-19 donor samples as proof of concept. These findings should be interpreted as hypothesis-generating rather than definitive evidence of COVID-19-induced retinal pathology, given the small cohort size and inherent heterogeneity of postmortem tissue.

Coronavirus disease 2019 has been increasingly recognized to affect multiple organ systems, including the eye.⁷ Histopathological studies of postmortem eyes have reported microvascular damage, microthrombosis, and inflammatory infiltrates in patients with long-COVID-19.^8,9 However, little is known about whether severe acute respiratory syndrome coronavirus 2 infection leads to persistent long-term molecular changes in the retina. Using spatial transcriptomics, we identified persistent transcriptional changes in the retina and RPE from donors with prior COVID-19. We emphasize that, given the exploratory nature of these findings and the limitations of the study design, these results represent candidate molecular associations rather than established consequences of COVID-19 infection.

Spatial transcriptomics offers a distinct advantage over conventional bulk or scRNA-seq approaches when studying heterogeneous tissues like the retina, where pathological changes are often crucial and adjacent to unaffected regions.10, 11, 12 Unlike single-cell sequencing approaches, which require tissue dissociation, spatial profiling enables guided region selection and maintains the anatomical context of gene expression. Although the ROI diameter used here does not reach single-cell resolution, it captures sufficient spatial details to resolve layer-specific transcriptomic signatures, as confirmed by marker localization and cell-type deconvolution. Consistent with previous spatial retinal studies,13, 14, 15, 16 we successfully identified molecular signatures corresponding to RGCs, photoreceptors, and RPE.

Previous studies have reported structural, vascular, and molecular abnormalities in the eyes of patients with long-COVID-19. Human postmortem retinal analyses have described microvascular damage,¹⁷ endothelial activation,¹⁸ perivascular inflammation,¹⁹ and mitochondrial dysregulation^20,21 in long-COVID-19 eyes. Our layer-specific spatial transcriptomics data reveal suggestive upregulation of mitochondrial respiratory chain components, oxidative phosphorylation genes, and nicotinamide adenine dinucleotide dehydrogenase complex members across the ganglion cell, photoreceptor, and RPE layers. These exploratory candidate genes were identified using a nominal P-value threshold without multiple-testing correction and therefore need to be independently validated in a larger cohort of samples. Prior investigations have also documented outer retinal defects including photoreceptor outer segment disruption^5,22, 23, 24 in long COVID-19 cases, which parallels our observation of down regulation of genes involved in cilium organization, photoreceptor sensory perception, and mitotic regulation.

Neuroinflammatory changes like increased expression of interferon stimulated genes and ribosomal/endoplasmic reticulum stress pathways have been detected previously in both retinal and brain tissues of patients with COVID-19.^25,26 Our current study expands on these findings by showing layer-specific transcriptional signatures, including upregulation of genes such as RPS4X, RPL24, SERP2, and COX4L1, which are associated with cellular stress responses. Importantly, most prior COVID-19 ocular studies could not determine whether molecular alterations reflected changes in gene expression or the loss of specific retinal cell types, because traditional analyses lacked spatial resolution. By integrating GeoMx spatial profiling with single cell-derived deconvolution, our study demonstrates that cell-type proportions remain stable across all retinal layers in post-COVID-19 versus control eyes. Nevertheless, interpretation of deconvolution outputs should remain semiquantitative, reflecting inferred relative enrichment rather than absolute cell counts. This cautious approach accounts for the technical variability and reliability inherent to the deconvolution of mixed cell spatial transcriptomic data. This finding indicates that the transcriptional changes identified, such as mitochondrial upregulation and ciliary downregulation, may reflect underlying molecular dysregulation rather than shifts in cellular composition.

This study provides the first spatially resolved, layer-specific transcriptomic analysis of human retinas from donors with a history of COVID-19, serving as an exploratory demonstration of the GeoMx DSP platform’s utility in disease relevant tissue. Our findings indicate that persistent molecular differences can be detected even after clinical recovery from COVID-19. However, these results should be interpreted with caution. The current data demonstrate an association between a history of COVID-19 and layer-specific transcriptional changes and are best viewed as hypothesis-generating rather than as definitive evidence of long-term COVID-19-induced retinal dysregulation. While these results do not establish local viral persistence as the primary driver of the observed retinal expression patterns, direct detection of severe acute respiratory syndrome coronavirus 2 RNA remains an important next step. As spatial profiling technologies advance toward subcellular resolution,27, 28, 29 future studies incorporating larger, sex-balanced cohorts and longitudinal sampling may further elucidate how systemic viral infection induces localized neurovascular and metabolic stress within the retina.

This study has several limitations, including a small sample size and cohorts spanning a broad age range, although comparisons were age-matched. Additionally, the imbalance between male and female participants limits our ability to exclude sex as a potential confounding factor. Due to the limited sample size, age was not included as a covariate in the analysis of transcriptional changes. Variability in COVID-19 disease severity, systemic involvement, and postmortem interval may also have contributed to transcriptional heterogeneity. The differential expression analysis was performed using nominal P-value thresholds; therefore, the resulting gene lists are candidates and require further validation. Future investigations using larger, sex-balanced cohorts with well characterized clinical histories are needed to validate these exploratory post-COVID-19 changes and to establish whether the observed transcriptional differences represent true long-term molecular consequences of severe acute respiratory syndrome coronavirus 2 infection.

Manuscript no. XOPS-D-26-00185.