MMP9 Genotype and Systemic T-Cell Subsets Correlate With Structural and Functional Outcomes in Neovascular Age-Related Macular Degeneration

Abstract

Purpose

Genetic studies implicate the matrix metalloproteinase-9 (MMP9) locus in neovascular age-related macular degeneration (nvAMD) risk but genotype–phenotype associations of MMP9 with nvAMD are lacking. This study aimed to investigate the influence of MMP9 genotype and T-cell subset frequency on structural and functional treatment outcomes in nvAMD.

Methods

We reanalyzed single-cell RNA sequencing data and used ELISA and flow cytometry in THP-1–derived monocytes to measure immune cell expression of MMP9 within human choroids. In a clinical nvAMD cohort of 38 patients, we genotyped the nvAMD risk single nucleotide polymorphism (SNP; rs4810482) and quantified retinal fluid using deep-learning–based optical coherence tomography (OCT) image analysis. On a subset of nine patients, we performed high-dimensional immunophenotyping.

Results

MMP9 is predominantly expressed in mature THP-1–derived dendritic-like cells (ELISA, P = 0.009; flow cytometry, P = 0.001). Patients with the TC genotype of MMP9 exhibited greater disease severity compared to CC or TT genotypes with a significantly higher total retinal fluid volume (P = 0.009). Immunophenotyping revealed that higher proportions of circulating CD8⁺ effector memory T cells re-expressing CD45RA (T_EMRA) were associated with increased residual subretinal fluid (P = 0.03), indicating persistent disease activity.

Conclusions

MMP9 genotype affects structural and functional outcomes in patients with nvAMD. Along with the observed systemic immune dysregulation, these findings support the role of a MMP9–dendritic–T-cell axis in nvAMD immunopathogenesis and highlight this as a potential therapeutic target.

Age-related macular degeneration (AMD) is the leading cause of irreversible vision loss worldwide, projected to affect 288 million people by 2040.¹ In neovascular AMD (nvAMD), macular neovascularization (MNV) leads to intraretinal and subretinal fluid accumulation with rapid visual decline. Intravitreal anti–vascular endothelial growth factor (VEGF) therapy remains the standard of care, but the response to treatment is highly variable among individuals, with too many patients responding poorly for unclear reasons.² Emerging evidence suggests that systemic and ocular immune dysregulation may contribute to the disease process, potentially accounting for part of the variability in response to anti-VEGF therapy.^3,4

Recent advances have facilitated the study of clinical, immunological, and genetic markers that may influence treatment response. For example, imaging analysis using deep learning enables precise, automated, and computationally efficient optical coherence tomography (OCT) biomarker quantification in AMD, and fluid volume has emerged as a more sensitive disease activity metric than central retinal thickness. These tools also allow large-scale structure–function analyses in clinical datasets.⁵ In parallel, genetic studies have identified immune-related AMD risk variants—for example, complement gene polymorphisms (CFH, CFI, C2, and C9)⁶—and, more recently, a variant in matrix metalloproteinase-9 (MMP9; rs4810482) has been specifically associated with nvAMD.^7,8

Patients with nvAMD have elevated levels of MMP9 in plasma and aqueous humor.^9,10 MMP9 is produced by immune cells, including neutrophils, macrophages, and dendritic cells (DCs) with choroidal macrophages in AMD showing notable MMP9 activity.^11–13 MMP9 encodes an extracellular matrix (ECM) enzyme that degrades key matrix components (elastin, fibrinogen, and types I and IV collagen) that are critical for Bruch's membrane integrity.^14,15 By breaking down the ECM, MMP9 also promotes immune cell extravasation.^16–18 Collectively, these observations suggest that MMP9 links structural alterations to immune dysregulation in nvAMD.

Although chronic inflammation plays a significant role in AMD pathophysiology, the contribution of systemic T cells and their subsets is poorly understood. One study found no significant difference in circulating total T cells (CD4⁺ and CD8⁺) between nvAMD and controls,¹⁹ yet other work showed that T cells accumulate in the choroid of nvAMD eyes, especially in early-stage AMD.^20,21 Peripheral T-cell populations were also recently found to be associated with genetic risk score and AMD disease stage.²² T cells in nvAMD show phenotypic alterations. For example, changes in CD35 and CD46—both key regulators of the complement system—have been reported on lymphocytes,^23,24 and the anti-angiogenic chemokine receptor C-X-C motif chemokine receptor 3 (CXCR3) is expressed at lower levels on both CD4⁺ and CD8⁺ T cells in nvAMD.²⁵ Given that MMP9 can modulate T-cell surface receptors such as CXCR3, further study is needed to determine its role in T-cell–mediated nvAMD pathology.

In addition to T cells, mononuclear phagocytes (macrophages and DCs) are also key cells in nvAMD. Macrophages produce angiogenic factors (VEGF, inflammatory cytokines) and regulate tissue inhibitors of metalloproteinases (TIMPs), contributing to MNV lesions and subretinal fibrosis.^26,27 However, the role of DCs in nvAMD is less well defined. Notably, single-cell RNA sequencing (scRNA-seq) shows a predominance of MMP9 and vascular endothelial growth factor A (VEGFA) expression in DCs.¹² Because DCs present antigens and activate T cells, MMP9 expression by DCs could foster a pro-angiogenic immune milieu in nvAMD through DC–T-cell interactions and lead to altered anti-VEGF therapy responses.

Despite evidence implicating MMP9 in MNV immunopathogenesis, its impact on nvAMD progression and treatment response is not understood. Here, we investigated the role of MMP9 and the MMP9–DC–T-cell axis in nvAMD and examined whether genetic variation in MMP9 influences immune phenotypes or treatment outcomes using clinical data and deep-learning OCT segmentation.

Methods

Re-Analysis of scRNA-seq Data

To investigate the role of MMP9-derived immune dysregulation in nvAMD, we assessed cell type–specific expression using a publicly available dataset that investigated gene expression in donor eyes with AMD using scRNA-seq.¹² This dataset contains samples from 21 human choroids, including 11 with AMD. The complete sequence data are available in GSE183320, and the processed data are available for interactive use at singlecell-eye.org.²⁸

To generate robust evaluations of gene expression, we used scRNA-seq differential expression with the pseudo-bulking protocol in Section 4.10 of the edgeR User's Guide²⁹ to summarize expression values for each cell type in each sample and to evaluate the statistical significance of differential gene expression. In brief, the Seurat2PB function from edgeR was used to generate the pseudo-bulk expression profiles for each cell type in each sample. From that matrix, samples with low representation and genes with low expression were removed, followed by gene expression normalization using trimmed mean of M values (TMM).³⁰ The stability of gene expression patterns within each cell type was verified using multidimensional scaling to verify that pseudobulk expression profiles derived from the same cell type clustered together. T-cell expression of several immune genes (B2M, NKG7, ITGB2, HLA-A, HLA-B, HLA-C, HLA-DRB1, CD154, CD69, IL2RA, and TNFRSF9) were evaluated. The statistical significance of changes in gene expression was assessed using the Wilcoxon test comparing normalized expression values between the samples with nvAMD and those with early AMD.

MMP9 Measurement in Differentiated Immune Cells From the THP-1 Cell Line

Culture and Differentiation of THP-1 Cells

THP-1 cells were thawed and cultured in RPMI 1640 (10% fetal bovine serum [FBS], penicillin–streptomycin). After recovery, ∼2 × 10⁶ cells were seeded in T-75 flasks (15 mL medium) and expanded to ∼8 × 10⁵ cells/mL before inducing differentiation. THP-1 differentiation protocols were as follows: (1) immature DC-like cells, 5 days in serum-free RPMI with granulocyte–macrophage colony-stimulating factor (GM-CSF; 50 ng/mL) + IL-4 (200 ng/mL); (2) mature DC-like cells, 3 days in serum-free RPMI with IL-4 (200 ng/mL), GM-CSF (20 ng/mL), tumor necrosis factor (TNF; 20 ng/mL), and ionomycin (200 ng/mL); (3) M0 macrophages, phorbol 12-myristate 13-acetate (PMA) for 1 to 2 days, then 3 days rest; (4) M2 macrophages, PMA (100 ng/mL) for 1 day, then IL-4 + IL-13 (20 ng/mL each) for 2 days. After each differentiation protocol, cells were placed in serum-free medium for 24 hours before analysis. Three biological replicates were performed on each cell line for analysis. Creation of these DC-like cells from THP-1 cells has been validated by multiple groups as a robust surrogate model for human DCs.³¹

MMP9 Measurements by ELISA and Flow Cytometry

Cell culture supernatants were clarified (12,000 rpm for 10 minutes at 4°C), and MMP9 concentration was measured using a Quantikine ELISA kit (DMP900; R&D Systems, Minneapolis, MN, USA), normalized to cell number. Differentiated cells were also analyzed for MMP9 by flow cytometry: Cells were surface stained in 100 µL fluorescence-activated cell sorting (FACS) buffer (PBS + 0.5% azide + 2% FBS) with antibodies to human leukocyte antigen–DR isotype (HLA-DR), CD206, CD86, and CD68 (30 minutes at room temperature in the dark), washed, then fixed/ permeabilized (Cytofix/Cytoperm; BD Biosciences, Franklin Lakes, NJ, USA) and stained intracellularly with CD206 and CD68 antibodies (30 minutes at room temperature in BD Perm/Wash buffer). Data were acquired on a Aurora cytometer (Cytek Biosciences, Fremont, CA, USA) and analyzed using SpectroFlo/FlowJo software. Three biological replicates were performed on each cell line.

Cohort of Subjects With nvAMD Genotyped for MMP9

Subjects with MNV consistent with nvAMD in at least one eye were recruited from University of Iowa Healthcare.⁷ The study was approved by the University of Iowa Institutional Review Board and adhered to the tenets of the Declaration of Helsinki. Informed written consent was obtained from all subjects prior to intervention. Patients underwent full ocular examination, including fundoscopy, best-corrected visual acuity assessment, and OCT on each visit. Those diagnosed with MNV and active exudation by an experienced retina specialist via OCT were included in the study as nvAMD. If there was uncertainty about the cause of neovascularization, participants further underwent OCT angiography and/or fluorescein and/or indocyanine green angiography.

Exclusions were made for eyes with inactive exudation, a history of vitrectomy, those with pathology limiting OCT analysis, or comorbid conditions such as diabetic retinopathy, venous occlusions, central serous chorioretinopathy, angioid streaks, presumed ocular histoplasmosis, and inherited retinal degenerations. Patients were also excluded from systemic immune analyses if they were taking systemic immunosuppressive medications, had cancer within the prior 6 months, and/or had a viral illness within the prior 2 weeks. One subject with both nvAMD and concurrent venous stasis retinopathy in the same eye was excluded from immune analyses due to the potential for non–AMD-related retinal pathology to confound interpretation of systemic immune phenotypes. Although venous stasis retinopathy is not expected to directly alter peripheral T-cell profiles, the presence of dual pathology and its known impact on visual acuity and intraocular fluid measures warranted exclusion to preserve the specificity of immune associations with AMD.

We collected the best-corrected visual acuity (from pinhole check when applicable) and injection interval at the initial (baseline), fourth injection, second year, and final visits. All visual acuity measurements were converted to a logarithm of the minimum angle of resolution (logMAR). Injection intervals were determined by chart review based on the physician's recommendation for the next follow-up appointment, and, if a range was given, the average value was used.

Genotyping was performed on DNA extracted from blood samples at the rs4810482 single nucleotide polymorphism (SNP) associated with the MMP9 locus by real-time quantitative polymerase chain reaction (PCR) with a CFX96 PCR machine (Bio-Rad Laboratories, Hercules, CA, USA) following the manufacturer's protocol for Applied BioSystems TaqMan assays (Thermo Fisher Scientific, Waltham, MA, USA), as previously described.⁷ Patients were categorized based on the rs4810482 SNP as those with CC (low risk), TC, or TT (high risk) alleles based on the risk-related prevalence of nvAMD, not risk due to severity of nvAMD phenotype. Heterozygotes were included in the high-risk group for analysis due to the presence of a high-risk allele, similar to what has been observed for CFH risk for nvAMD.

Automated Fluid Quantification of nvAMD From OCT Scans

To provide objective and efficient quantification of retinal fluid metrics, we developed a deep-learning–based automated segmentation framework. A total of 100 6 × 6-mm macular OCT volume (SPECTRALIS; Heidelberg Engineering, Heidelberg, Germany) were utilized for artificial intelligence (AI) model development. Ground-truth segmentations for pigment epithelial detachment (PED), subretinal fluid (SRF), and intraretinal fluid (IRF) were generated by two retina specialists; the initial annotations from one specialist were reviewed and refined by a second masked specialist to establish a definitive consensus. The segmentation architecture employed an ensemble of U-Net architectures³² with three distinct backbones (resnet34,³² inceptionv3,³² and vgg16³³) to simultaneously segment PED, SRF, and IRF regions from OCT scans in nvAMD. Three consecutive B-scan images, downsampled to 128 × 128 pixels, were used as an input for the U-Net architectures which were trained using the following hyperparameters: The optimizer was Adam with a learning rate of 0.0001; the loss function incorporated categorical cross entropy and dice loss; and the maximum number of epochs was 200. The resnet34, inceptionv3, and vgg16 neural networks utilized the weights trained on the 2012 ImageNet Large Scale Visual Recognition Challenge dataset.³⁴

To validate the framework, 10-fold cross-validation was performed with training OCT scans from 90 subjects and a validation dataset of 10 subjects per fold. Prediction performance for each fluid compartment was assessed via correlation between automated results and manual ground truth, reported as R² values. Two scans containing extreme fluid volumes, which exceeded the detection threshold of the model, were identified as outliers and excluded from the final regression analysis to ensure stability of the performance metrics.

Immune Profiling

We analyzed peripheral blood mononuclear cells (PBMCs) from nine nvAMD patients to characterize T-cell subsets. PBMCs were isolated by density-gradient centrifugation (Lymphoprep/SepMate; StemCell Technologies, Vancouver, BC, Canada) and cryopreserved in 90% FBS/10% dimethyl sulfoxide (DMSO) in liquid nitrogen. On the day of flow cytometry, PBMC vials were thawed, washed with warm RPMI, and resuspended in RPMI at ∼1 to 2 × 10⁷ cells/mL.

For the 31-color steady state T-cell panel (see gating strategy in Supplementary Fig. S1), cells were stained as previously described.³⁵ Briefly, cells were stained in PBS with LIVE/DEAD Blue (Thermo Fisher Scientific) for 15 minutes at room temperature. Cells were centrifuge at 500g for 5 minutes at room temperature. The supernatant was discarded, and the cells were then resuspended in an antibody cocktail containing the following antibodies for 30 minutes at room temperature: anti-CD3 (SK7), anti-CD4 (SK3), anti-CD8 (SK1), anti-CD45RA (5H9), anti-CD45RO (UCHL1), anti-CD11a (HI111), anti-CD49d (L25), anti-CCR4 (1G1), anti-CD73 (AD2), anti-CD27 (O323), anti-CCR7 (G043H7), anti-CCR6 (G034E3), anti-CD28 (CD28.2), anti-PD-1 (EH12.2H7), anti-TIM-3 (7D3), anti-CXCR5 (RF8B2), anti-CD38 (S17015F), anti-CD226 (11A8), anti-CXCR3 (G025H7), anti-CD95 (DX2), anti-TCRgd (B1), anti-CD39 (A1), anti-KLRG1 (SA231A2), anti-CD127 (A019D5), anti-CD25 (BC96), anti-CD45 (2D1), anti-CD56 (NCAM16.2), and anti-HLA-DR (L243). Following staining, cells were washed twice with FACS buffer and then fixed with Foxp3 fixation/permeabilization buffer for 45 minutes at 4°C. Following fixation, cells were washed twice with Foxp3 permeabilization/wash buffer and resuspended in permeabilization/wash buffer containing the following antibodies for 30 minutes at room temperature: anti-Foxp3 (PCH101) and anti-granzyme B (QA16A02). Cells were then washed twice with permeabilization/wash buffer and once with PBS, resuspended in PBS, and stored at 4°C until analysis.

Statistical Analyses

Data were analyzed using RStudio 2024.12.1 (R Foundation for Statistical Computing, Vienna, Austria) and SPSS Statistics 27 (IBM, Chicago, IL, USA). Continuous data are reported as mean ± SD or median (interquartile range), as appropriate. THP-1 ELISA results were compared by Kruskal–Wallis test with Dunn's post hoc without Bonferroni corrections. The THP-1 flow groups passed the Shapiro–Wilk normality test and were computed with a one-way ANOVA with Tukey post hoc without Bonferroni corrections. Clinical outcomes by MMP9 genotype were compared using generalized estimating equations (GEEs) to account for two-eye correlation, adjusting for age, sex, body mass index, and smoking. P values for multiple comparisons were Bonferroni corrected, with significance at P < 0.05 (adjusted). Flow cytometry medians (immune phenotypes) in low-risk versus high-risk genotypes were compared by Mann–Whitney U tests. Spearman's rank correlation was used to assess associations between immune markers and clinical measures (using the mean of the values for both eyes for visual acuity [VA] and fluid) without Bonferroni corrections with marginal significance at P < 0.05. We used ggplot2 in R to create correlation plots.

Results

MMP9 Expression by scRNA-seq and In Vitro Quantification

A prior scRNA-seq study demonstrated higher choroidal MMP9 expression in AMD patients compared to controls.¹² We reanalyzed that dataset (21 human choroids, 11 with AMD) to probe the role of MMP9 in nvAMD. MMP9 expression was primarily localized to DCs with lower expression in inflammatory macrophages and almost none in resident macrophages (Fig. 1).

Figure 1.

A heatmap of MMP9 expression from cell clusters of resident or inflammatory macrophages (MACROPHAGE-RES, MACROPHAGE-INF), dendritic cells, and B cells from scRNA data of 21 human donor choroids with AMD.^12,28 UMAP, uniform manifold approximation and projection.

To demonstrate variation in the expression of MMP9 among immune cells, we designed experiments using the THP-1 cell line (Supplementary Fig. S2) which mirrored the scRNA-seq results (Fig. 2). Only the mature DC-like cells had a significantly higher concentration of MMP9 than control medium (P = 0.009, Kruskal–Wallis test with the post hoc Dunn's test); immature DC-like cells and M0 and M2 macrophages showed no significant difference. The increase in MMP9 expression in mature DC-like cells was further confirmed by flow cytometry analysis. Almost all mature DC-like cells in culture expressed MMP9 (Figs. 2B, 2C) and this expression was significantly higher compared to other THP-1 differentiated cells that exhibited little to no MMP9 expression (P ≤ 0.001 for THP-1, M0, M2, and immature DC-like cells).

Figure 2.

Mature dendritic-like cells show high MMP9 expression across THP-1–derived immune cell types. (A) ELISA of conditioned media demonstrates extracellular MMP9 levels across unpolarized macrophages (M0), IL-4–polarized macrophages (M2), immature dendritic-like cells (iDCs), and mature dendritic-like cells (mDCs). Results were compared with the Kruskal–Wallis test with Dunn's post hoc without Bonferroni correction. (B) Flow cytometry using an intracellular anti-MMP9 antibody confirmed intracellular expression. (C) Quantification of geometric mean fluorescence intensity (gMFI) from B. The flow cytometry passed the Shapiro–Wilk normality test and was compared with a one-way ANOVA with Tukey post hoc without Bonferroni corrections. Data represent mean ± SEM (n = 3 biological replicates for each cell type in both experiments). *P < 0.05, **P < 0.01, ***P < 0.001.

To assess T-cell markers in the choroid of patients with nvAMD, we compared nvAMD versus non-nvAMD samples. Due to the relatively low sample size, several canonical markers were not significantly different by pseudobulk scRNA-seq analysis (IL2RA, CD69, CD40L, TNFRSF9; P ≈ 0.07–0.3l) (Supplementary Fig. S3). However, the trends observed for antigen-presentation and cytotoxic-associated genes are consistent with altered immune-retinal interaction biology (B2M, NKG7, ITGB2, HLA-A, HLA-B, HLA-C, and HLA-DRB1; P ≈ 0.06–0.2 (Supplementary Fig. S4). These findings support that MMP9 expression by DCs is increased while markers of T-cell antigen presentation and cytotoxicity may be dysregulated in nvAMD. Together, these experiments support the concept that DCs are the main source of MMP9 within the choroid and a potential for these DCs to alter the way T cells interact with the retina, contributing to nvAMD pathophysiology.

Deep-Learning–Based Segmentation of Retinal Fluid Components

With previous data demonstrating a connection between MMP9 expression and T-cell marker dysregulation, we then set out to determine if the MMP9 genotype could be connected to AMD phenotypes in patients. To determine accurate and efficient anatomical outcomes, we developed a deep learning model to automatically segment retinal fluid on OCT scans of nvAMD patients. The model, trained on a separate set of annotated OCT scans, was evaluated on an independent test set not seen during training. Using 100 expert-labeled three-dimensional OCT scans (95 patients), the predictions showed excellent agreement with manual segmentations (Fig. 3). Predicted versus expert-measured fluid volumes were highly correlated for PED, SRF, and IRF (R² = 0.89, R² = 0.95, and R² = 0.85, respectively; P < 0.001 for each) (Fig. 4), after excluding two extreme outlier scans. These results validate the model as a reliable tool for automated fluid quantification in nvAMD. Beyond confirming strong agreement with expert graders, the model enables consistent, large-scale measurement of IRF, SRF, and PED fluid volumes—metrics that are more sensitive indicators of disease activity than central retinal thickness. This provides a robust foundation for linking structural OCT biomarkers with genetic and immune phenotypes in our cohort.

Figure 3.

Representative OCT scans illustrate AI-based segmentation of fluid in nvAMD. Raw OCT scans (left column) from three patients (P1–P3) are compared with expert manual tracings (middle) highlighting fluid compartments: intraretinal fluid (red), subretinal fluid (yellow), and pigment epithelial detachment (green). Automated segmentation results after AI training are shown on the right.

Figure 4.

Automated AI segmentation correlates with ground truth for retinal fluid compartments in nvAMD. Correlation plots from 98 OCT scans show 10-fold cross-validated agreement between AI predictions and manual tracings for PED, SRF, and IRF. These were analyzed with the Pearson correlation test yielding P < 0.001 for all fluid types.

MMP9 Genotype Affects Total Retinal Fluid Volume and Treatment Response in nvAMD

We genotyped 38 nvAMD patients (56 eyes; mean age, 75.6 years; 62.5% female) for the MMP9 SNP rs4810482, where low-risk (CC) versus high-risk (TC or TT) groups indicate nvAMD susceptibility⁷ not phenotypic severity (Table). The genotype groups did not differ in baseline age or in mean treatment duration (Supplementary Table S1).

Table.

Comparison of Structural and Functional Outcomes of Different MMP9 Variants of the High-Risk SNP (rs4810482)

	MMP9 Genotype (rs4810482), Mean ± SD				Pairwise Comparison
	CC (N = 10)	TC (N = 26)	TT (N = 20)	Significance CC Vs. TC Vs. TT	CC Vs. TC	CC Vs.TT	TC Vs. TT
Age (y) at baseline	72.08 ± 10.54	77.53 ± 9.26	74.89 ± 6.9	0.57
Female sex	50%	54%	80%	0.12
Baseline VA (logMAR)	0.42 ± 0.31	0.32 ± 0.26	0.4 ± 0.39	0.216
2-year VA (logMAR)	0.46 ± 0.49	0.26 ± 0.22	0.27 ± 0.23	0.322
Final VA (logMAR)	0.52 ± 0.58	0.42 ± 0.28	0.3 ± 0.33	0.863
Baseline IRF (mm3)	0.02 ± 0.04	0.04 ± 0.12	0.02 ± 0.05	0.817
2-year IRF (mm3)	0.01 ± 0.02	0.02 ± 0.05	0.03 ± 0.07	0.003	0.016	0.002	0.050
Final IRF (mm3)	0.001 ± 0.001	0.01 ± 0.04	0.01 ± 0.03	0.233
Baseline SRF (mm3)	0.17 ± 0.28	0.3 ± 0.57	0.15 ± 0.18	0.182
2-year SRF (mm3)	0.002 ± 0.004	0.05 ± 0.14	0.03 ± 0.07	0.308
Final SRF (mm3)	0.001 ± 0.001	0.07 ± 0.16	0.02 ± 0.04	0.102
Baseline PED (mm3)	0.17 ± 0.15	0.8 ± 1.12	0.29 ± 0.32	0.041	0.062	1	0.65
2-year PED (mm3)	0.15 ± 0.15	0.35 ± 0.39	0.29 ± 0.34	0.285
Final PED (mm3)	0.15 ± 0.11	0.38 ± 0.41	0.28 ± 0.28	0.041	0.054	0.123	1
Final total fluid (mm3)	0.15 ± 0.1	0.46 ± 0.46	0.3 ± 0.3	0.009*	0.012*	0.065	1
ΔVA (logMAR), final vs. baseline	0.11 ± 0.45	0.1 ± 0.28	−0.1 ± 0.24	0.029	1	0.52	0.026
ΔSRF (mm3), final vs. baseline	−0.17 ± 0.28	−0.24 ± 0.61	−0.13 ± 0.18	0.252
ΔPED (mm3), final vs. baseline	−0.02 ± 0.13	−0.42 ± 1.01	−0.01 ± 0.2	0.077
ΔTotal fluid (mm3), final vs. baseline	−0.21 ± 0.38	−0.68 ± 1.38	−0.16 ± 0.28	0.033	0.317	0.201	0.03*

GEE analysis, adjusted for age, gender, body mass index, and smoking history, was used to compare mean values, and χ² tests were used to compare percentages.

^*Bonferroni-corrected significance with alpha set at 0.05.

To explore the association between MMP9 genotype and clinical phenotype in our cohort we applied GEE analysis on both anatomical and visual outcomes (Supplementary Fig. S5). Baseline VA, injection intervals, and total injections were similar across the three genotypes (P > 0.05), indicating that all groups received comparable treatment. After adjusting for covariates (age, sex, smoking, body mass index), VA at each follow-up point did not differ by genotype. However, the overall change in VA from baseline to final visit was significantly worse in the TC group compared to TT (Table). Using our previously mentioned automated segmentation model for OCT analysis, we observed higher volumes of IRF at the second year of treatment with each additional risk allele, although this was not found at any other time point. The baseline total combined fluid burden was highest for the TC genotype group, and these patients showed the greatest decrease in total fluid from initial visit to final, whereas the TT group had the least amount of fluid improvement. There were no significant differences in the change in SRF or PED volume between genotypes from initial to final visit. Although our sample size is modest, these findings suggest the possibility of more severe disease in the TC genotype due to total fluid volume and change in visual acuity from baseline to the final visit.

Systemic Immune Dysregulation in nvAMD

To understand how different T-cell phenotypes are associated with clinical outcomes, we used flow cytometry to profile peripheral T cells in nine nvAMD patients (Supplementary Table S2), seven of whom had OCTs available for analysis by our deep learning model. Subretinal fluid volume at the final visit correlated strongly with specific T-cell subsets (Fig. 5A). Higher frequencies of CD8⁺ effector memory T cells re-expressing CD45RA (T_EMRA) cells were associated with more SRF (R = 0.81, P = 0.03), whereas higher CD4 T_EMRA and CD4 early effector memory (T_EM) frequencies (R = −0.85 and R = −0.79, respectively) were associated with less SRF (R = −0.85 and R = −0.79, respectively; P < 0.05 for both). Interestingly, higher frequencies of CD4⁺ Th17 and Th1/Th17 cells were each associated with lower SRF volume (R ≈ –0.8, P ≈ 0.03 for both) (Fig. 5B). In addition, we found that those with the MMP9 TT and TC genotypes had a significantly lower Th17/Treg ratio compared to the CC group (P = 0.02) (Supplementary Table S3).

Figure 5.

Systemic T-cell subsets correlate with retinal structure and function in nvAMD. Scatterplots show Spearman rank correlations between circulating T-cell subtypes and quantitative OCT or visual acuity readouts at the final study visit. (A) Correlation plots of systemic T-cell subsets with subretinal fluid volume. (B) Negative correlations of functional CD4 subtypes with subretinal fluid. (C) Negative correlations of functional CD4 subtypes with total retinal fluid volume (sum of subretinal, intraretinal, and pigment epithelial detachment). (D) Correlations between CD8 subtypes and best-corrected visual acuity in logMAR units, where higher values represent worse vision. Each dot represents an individual patient (n = 7–9 per correlation). T_EM, effector memory; T_EMRA, effector memory re-expressing CD45RA; Int., intermediate. A summary of additional non-significant or significant correlations not depicted is provided in Supplementary Table S2.

We also observed negative correlations between functional subtypes and total fluid measurement, including cytotoxic CD4⁺ T cells (R = −0.82, P = 0.02) and CD4 T follicular helper cells (R = −0.79, P = 0.04) (Fig. 5C). It is important to note that the cytotoxic CD4⁺ T-cell correlation with total fluid appears to be driven by its association with final PED (R = −0.82, P = 0.02), whereas no individual fluid volume had a significant correlation for the CD4 T follicular helper cell subtype. We saw correlations between CD8 subtypes and final corrected distance VA (Fig. 5D), with CD8 T_EM intermediate effectors displaying a negative correlation (R = −0.71, P = 0.03), indicating better VA at higher frequencies. However, increased CD8 T_EMRA early-like frequencies were associated with worse VA at the final visit (R = 0.71, P = 0.03). Additionally, higher frequencies of CD4 naïve T cells were associated with greater SRF volumes (R = 0.88, P = 0.009). These analyses suggest that there are numerous significant positive and negative associations between T-cell subsets and retinal fluid components derived from the automated AI-based OCT image analyses. In addition to these correlations, we found that our cohort had a CD8⁺ T-cell population that was markedly skewed toward memory cells compared to CD4⁺ T cells, with a memory-to-naïve ratio of 5.59 versus 0.84 (P = 0.005, Mann–Whitney). Although our sample size is small, our findings set the groundwork for utilizing a combination of immunophenotyping and OCT biomarkers in larger studies to predict long term outcomes in nvAMD.

Discussion

Neovascular AMD is a complex retinal disease with limited therapies, and genetic influences on treatment response have been studied for variants in VEGFA, CFH, CFI, C3, ARMS2, and HTRA1.^36,37 Our study is the first, to our knowledge, to examine how MMP9 genotype affects disease phenotype and treatment outcomes in nvAMD. The rs4810482 SNP is a noncoding T>C variant in the promoter region for MMP9. The rs4810482 T allele is associated with increased prevalence of nvAMD, but associations of various MMP9 genotypes to phenotypic severity in those who already have nvAMD have not been published to our knowledge.⁷ Although our cohort did not display differences in visual acuity at any of the selected post-treatment time points for patients genotyped for rs4810482, those with the TC genotype demonstrated a greater decrease in VA from the initial to final visit compared to those with the TT genotype. The heterozygotes also had the highest amount of residual final total fluid compared to the CC group, despite a similar baseline retinal fluid volume. Although the heterozygotes had the greatest decrease in fluid after treatment, they also had the highest volume of residual fluid. In accordance with our findings, vitreous levels of MMP9 have been shown to be elevated in nvAMD and have also been positively correlated with the amount of subretinal fluid.³⁸ The positive feedback loop between MMP9 and VEGF may explain the increased angiogenic activity and accumulation of total retinal fluid in those who have at least one high-risk allele.^39–41

The observed clinical severity in heterozygotes—characterized by significantly greater residual subretinal fluid and worse visual outcomes compared to TT homozygotes—presents a nonlinear genotype–phenotype relationship that challenges standard gene-dosage models. However, this phenomenon is not without precedent in other human pathologies. For example, studies on the MMP9 promoter -1562 C/T polymorphism in papillary thyroid carcinoma have demonstrated that the presence of the T allele leads to the loss of a transcriptional repressor binding site, significantly upregulating promoter activity and driving aggressive clinical phenotypes.⁴² We speculate that, in nvAMD, the non-coding rs4810482 variant may similarly disrupt specific transcription factor binding sites within an enhancer or promoter region. In the TC heterozygous state, the imbalance between wild-type and variant alleles may reflect nonlinear regulation of MMP9 expression or feedback signaling that is otherwise triggered by the higher MMP9 concentrations produced in TT homozygotes. These findings must be confirmed in a larger cohort of patients with additional laboratory work for verification.

Macrophages are well studied in nvAMD, but the role of DCs is less clear. DCs have been found in choroidal neovascularization (CNV) lesions, and their presence—especially in an immature state—can enlarge lesions in a laser model.⁴³ Correspondingly, depleting DCs via Flt3 deficiency led to smaller CNV lesions in one study.⁴⁴ However, another study noted that, even with increased DC recruitment after laser injury, Flt3-knockout mice did not develop smaller CNV areas, suggesting that DCs may influence lesion severity but not initiate nvAMD.⁴⁵ Notably, AMD-related factors such as VEGF and hypoxia can inhibit DC maturation while inducing MMP9 expression.^46,47 Taken together, our findings and prior studies support a modulatory role for DCs in nvAMD immunopathogenesis.

Our analyses confirm that DCs, rather than macrophages, are the chief source of MMP9 in the choroid (with inflammatory and resident macrophages showing minimal MMP9). Although there are no lymphatics in the choroid, extensive work by McMenamin et al.^48,49 and others have demonstrated that the human choroid contains DCs and an extensive network of resident major histocompatibility complex (MHC) class II⁺ cells. Our THP-1 experiments demonstrated significantly higher MMP9 expression in mature DC-like cells than in undifferentiated cells or M0 macrophages (we did not assess M1 macrophages, which are less angiogenic).⁵⁰ Our findings are consistent with the study by Hollender et al.,⁵¹ which found expression of both active and inactive MMP9 in mature DCs but expression of only inactive MMP9 in immature DCs. Similarly, Zhou et al. showed that PMA-driven THP-1 differentiation markedly upregulated MMP9.⁵² MMP9 expression in DCs affects their functionality. MMP9 indirectly regulates DC maturation through a positive feedback loop with VEGFs.^39–41,46 Expression of MMP9 from immune cells is affected by local mediators and is crucial for their respective functions. For example, MMP9 expression is decreased in hypoxia and greatly increased during oxidative stress.⁵³

MMP9-mediated degradation of the ECM is in part regulated by TIMPs.^54–56 Patients with TIMP3 mutations (Sorsby fundus dystrophy) have a higher incidence of MNV lesions.^57,58 In experimental models, MMP9 is upregulated in CNV lesions, and MMP9 knockout mice develop smaller CNV lesions.⁵⁹ However, it is important to note that MMP9 not only is an extracellular protein degrading the ECM but also exhibits significant intracellular expression, as shown in Figure 2C. Dendritic cells also rely on MMP9 for the expression and modulation of surface molecules that are required for cellular migration,^13,52 and a high-risk MMP9 status could allow for subretinal migration of DCs and activation of the adaptive immune system.

By integrating automated deep-learning retinal fluid measurements and the use of flow cytometry in our nvAMD cohort, we revealed novel correlations between systemic T-cell subtypes and treatment outcomes. We found that terminally differentiated CD8⁺ T cells and undifferentiated CD4 naïve T cells were correlated with increased SRF (Fig. 5A). This is supported by a previous study that showed increased maturation of CD8⁺ T cells in nvAMD compared to healthy controls, although CD4⁺ T-cell maturation frequencies were not significantly different.⁶⁰ Interestingly, we found that distinct CD8⁺ T-cell subsets demonstrated divergent relationships with both retinal fluid and visual function. CD8 T_EMRA intermediate cells showed a strong positive correlation with final visit SRF, suggesting a potential pathogenic role in maintaining or exacerbating exudation. In contrast, CD8 T_EM intermediate cells did not correlate with SRF but were negatively correlated with logMAR corrected distance visual acuity (CDVA), indicating better visual function at higher frequencies. These contrasting patterns likely arise from functional differences among CD8⁺ T-cell subsets. Less differentiated CD8 T_EM cells may be more regulated or beneficial, whereas an accumulation of highly differentiated (T_EMRA early-like) CD8 cells—associated with worse vision—could indicate chronic stimulation or immune exhaustion leading to tissue damage. Consistent with this idea, the CD8 T_EMRA early-like subset showed no correlation with exudative fluid (SRF/IRF), suggesting that it might impair vision through non-exudative mechanisms. CD4 naïve T cells were previously thought of as quiescent cells awaiting activation. However, Yoon et al.⁶¹ found that these cells are not homogenous, as a subpopulation was shown to exhibit inflammatory Th1 characteristics, suggesting a possible explanation for our findings. These changes in T-cell subtypes of nvAMD patients could reflect senescence of the immune system. Naïve CD8⁺ T cells have been reported to decline more sharply, leading to a predominance of memory and senescent cells, whereas naïve CD4⁺ T cells are less prone to senescence.⁶²

The immunophenotype of our nvAMD cohort demonstrated a negative relationship between Th17 and Th1Th17 hybrid cell frequencies and subretinal fluid at the final visit (Fig. 5B). We also observed negative correlations between CD4 cytotoxic cells and follicular helper T cells with the final total combined fluid (Fig. 5C). IL-17 is known to be important in nvAMD pathogenesis^63–65 and is involved in the upregulation and activation of MMP9.⁶⁶ Despite this, the literature on the level of Th17 cells in AMD is mixed.^67–69 Although Th17 cells are known for their production of IL-17, they are not the only source, as it is expressed by other innate lymphoid cells and γδ T cells which may explain these negative correlations.⁶⁴ The CD4 cytotoxic subset has not previously been studied in AMD but has been detected in cancers and autoimmune diseases.⁷⁰ On the other hand, follicular T helper cells are known to activate B cells to produce antibodies and may play a role in the initial stages of AMD as retinal autoantibodies have been found in early stages⁷¹ and follicular T helper cells are increased in nvAMD compared to controls.⁷² Our findings showed higher levels of fluid at lower follicular T-cell concentrations that can be explained either by these cells having a less important role in the later stages of the disease, and/or because we measured systemic cells and do not have data on the composition of local T cells. On a broader level, although our pseudobulk data are limited by a small sample size, we observed non-significant trends in the transcription of genes responsible for antigen presentation and cytotoxic phenotypes that are consistent with altered immune–retinal interaction biology. The significant correlations observed with nvAMD outcomes and T-cell populations suggest that the adaptive immune system may modulate disease severity.

Our results revealed a significantly lower ratio of TH17 to Treg cells in those with the TT or TC genotypes (Supplementary Table S3). Modulation of the T-cell differentiation pathway could be expected based on changes in the expression of MMP9, as it regulates cytokines such as TGF-β which are key in the activation of CD4 naïve cells to TH17 and Treg cells.⁷³ MMP9 itself is regulated by many cytokines and is involved in regulating other immune cells by cleaving cytokines and cytokine receptors, leading to a wide array of possible inflammatory and anti-inflammatory effects.^{16,17,74–76} Notably, MMP9 has been shown to degrade the CXCR3 receptor, further contributing to a pro-angiogenic environment.¹⁷ Together with our previous implication of T cells in treatment outcomes, this creates a new mechanistic hypothesis for increased nvAMD risk and altered anti-VEGF response by MMP9 genotype due to immune dysregulation.

This study has several limitations. Our sample size was modest, and the multitude of comparisons (despite Bonferroni correction) means some findings were only marginally significant; thus, validation in a larger cohort is needed. Also, we analyzed peripheral T cells even though nvAMD is primarily an ocular condition. Nonetheless, the systemic immune alterations we observed are echoed by epidemiological links between AMD and systemic autoimmune diseases (e.g., higher AMD rates in Crohn's disease and lupus).⁷⁷ Moreover, our finding that peripheral T-cell profiles correlate with retinal fluid is in line with recent work by Lin et al.,⁷⁸ which found altered systemic immune signaling associated with AMD severity (including nvAMD). This growing evidence supports a broader role for systemic immunity in AMD. Finally, because our study was observational, we cannot infer causality between the immune changes and nvAMD. More extensive and longitudinal studies are warranted to clarify the role of the MMP9–DC–T-cell axis in AMD pathogenesis and therapy response. Additionally, although the model has been validated, the dendritic cell experiments used THP-1–derived DC-like cells as an in vitro model, which may not fully recapitulate the phenotype and functional diversity of primary ocular dendritic cells.³¹ Examining the ocular T-cell compartment directly (e.g., through ocular samples) and conducting longitudinal studies will be important to determine whether the immune changes are a cause or a consequence of nvAMD.

Conclusions

We propose that MMP9 contributes to nvAMD not only through ECM remodeling but also as a crucial immune modulator. Dendritic cells appear central to this process, linking localized ocular inflammation with systemic immune dysregulation. Moreover, although rs4810482 in MMP9 was known as a genetic risk factor for nvAMD, our results are the first, to our knowledge, to show that the MMP9 genotype correlates with differences in patients’ visual and fluid outcomes, implicating this genotype as a potential marker of disease severity and treatment responsiveness. Although very preliminary, our findings have clinical implications, including that more aggressive treatment of patients with nvAMD with the TC variant may be warranted. Additional research with larger numbers of patients is needed to confirm this hypothesis. Collectively, our findings provide a foundation for developing personalized, immune-targeted therapies for nvAMD.