Proteomic clocks combined with deep learning phenotypes track eye aging and diseases Authors: Shaopeng Yang, Zhuoyao Xin, Huangdong Li, Zhuoting Zhu, Ziyu Zhu, Yanping Chen, Yongjie Zhang, Ruilin Xiong, Minwen Zhou, Mingguang He, Wei Wang Published: 2026 ABSTRACT AbstractProteomics represents a powerful but underutilized approach for characterizing eye aging. Here, leveraging data from three large-scale, cross-national cohorts of over 55,000 transethnic participants, we demonstrate the ability of high-throughput proteomics combined with deep learning (DL) phenotyping to track eye aging and disease in both discovery and external validation settings. Proteomic aging driven by machine learning modeling closely aligns with signals of eye aging and DL aging phenotypes. We identifiy and validate premature proteomic aging in individuals with major age-related eye diseases (AREDs), including cataract, diabetic retinopathy, age-related macular degeneration, and glaucoma, and propose evidence supporting proteomic aging acceleration as a robust biomarker for predicting these conditions beyond chronological age, with adaptability across sexes and ethnicities. We also develop a streamlined, cost-effective proteomic aging clock that preserves predictive performance while reducing assay complexity. By integrating advanced tomographic and angiographic imaging, we derive structural and functional biomarkers through DL-driven pipelines and link accelerated proteomic aging to both neuroretinal degeneration and microvascular rarefaction in the Guangzhou Diabetic Eye Study (GDES) and the High-definition Oculo-Phenomic Evaluation (HOPE) study, highlighting coupled neural–vascular decline in eye aging. Our findings position proteomic aging combined with AI as a scalable tool for tracking eye health and disease, and provides new insights into shared aging pathways underlying multiple ocular pathologies.Subject terms: Biomarkers, Computational biology and bioinformatics, Diseases, Medical research Introduction With the global population aging, visual impairment has become a major public health concern.1 As the leading cause of visual impairment, age-related eye diseases (AREDs) not only diminish quality of life, but also increase the risk of depression, dementia, cardiovascular disease, and lung cancer, while exacerbating health inequities.1–9 However, eye aging is highly heterogeneous, whereby older chronological age (CA) does not necessarily align with physiological decline.10 For instance, while some centenarians maintain clear lenses and healthy retinas, others experience vision deterioration before age 50.11,12 Should the aging process be quantified, monitoring of individuals at risk for ARED would be facilitated, especially since over 90% of these conditions are preventable with cost-effective interventions.3 Aging clocks built on omics approaches have advanced the molecular characterization of the organismal aging at molecular resolutions.13 While initially based on DNA methylomics,14,15 proteomic approaches offer distinct advantages in elucidating aging biology by directly linking to life function,16,17 placing proteostasis along with other processes pointing to proteomic changes as primary hallmarks of aging.18 Recent advances in proteomics and machine learning (ML) algorithms have led to proteomic clocks that capture aging heterogeneity and the mechanistic connections of aging across multiple organs, such as the brain, heart, and kidney.19–21 However, though preliminary evidence has linked accelerated aging to retinal vascular leakage and macular degeneration,22,23 the role of proteomic aging in eye health and AREDs remains largely unexplored. Parallel progress in AI algorithms has enabled aging signal extraction from ocular imaging.24–26 Deep learning (DL)-driven models trained on fundus photographs have shown promise in predicting AREDs and systemic diseases,27–29 but most frameworks are optimized for disease-specific risk classification and offer limited insight into broader, shared aging processes. Using the same imaging modality, we previously developed a DL-driven eye aging clock30–34 and improved its dynamic performance through longitudinal pre-training and label distribution learning.35 Other high-throughput DL pipelines facilitate scalable segmentation and quantification of ocular endophenotypes, complementing eye aging signals with finer-grained, interpretable aging biomarkers.25,36 However, it remains unknown whether proteomic aging aligns with or diverges from these eye aging signals. Given the minimally invasive nature of plasma sampling and the proteomics’ potential to provide mechanistic insights into aging biology,13 we hypothesize that a proteomic clock can track eye aging and disease. To test this, we developed an ML-driven proteomic aging clock within a biobank-scale cohort from Europe to characterize eye aging and four major AREDs (cataract, diabetic retinopathy [DR], age-related macular degeneration [AMD], and glaucoma) and externally validated these findings in two independent cohorts from East Asia (Fig. 1). We first assessed whether proteomic aging aligns with eye aging signals and fine-grained fundus biomarkers by DL phenotyping, while comparing aging across participants with and without prevalent AREDs. Next, leveraging over 780,000 person-years of follow-up, we examined whether accelerated proteomic aging exacerbates the risk of AREDs. Lastly, we provided evidence of neurodegeneration and microvascular rarefaction dynamics in the retina across proteomic aging. Overview of the study design. a A total of 53,014 UKB participants, 1584 GDES participants, and 430 HOPE participants who underwent proteomic profiling were included in this study. b The machine learning (ML)-based proteomic clock model was trained and evaluated in five-fold nested cross-validation. Accelerated proteomic aging was associated with neural degeneration and other eye aging signals derived from deep learning (DL) phenotyping. Participants with prevalent AREDs exhibited significant proteomic aging, while accelerated proteomic aging was predictive of incident AREDs. c Findings were validated in an independent cohort of distinct ethnicity (GDES). d Accelerated proteomic aging was associated with microvascular rarefaction in the GDES and HOPE. The retinal microvasculature was segmented into SCP and DCP slabs, and micro-perfusion from each slab was analyzed within three concentric ETDRS circles. Parts of (a–c) were created from BioRender (biorender.com) and Flaticon (flaticon.com). GDES Guangzhou Diabetic Eye Study, HOPE High-definition Oculo-Phenomic Evaluation, DR diabetic retinopathy, AMD age-related macular degeneration, ARED age-related eye disease, MAE mean absolute error, SCP superficial capillary plexus, DCP deep capillary plexus. Results Study population and proteomic clocks A total of 55,028 transethnic participants from the UK Biobank (UKB), the Guangzhou Diabetic Eye Study (GDES), and the High-definition Oculo-Phenomic Evaluation (HOPE) study, who underwent proteomic profiling using the antibody-based Olink Explore proximity extension assay (PEA) were included in this study to model proteomic aging and track eye health and disease (Methods and Supplementary Fig. S1). The mean (SD) CA of the UKB population was 56.8 (8.2) years, with 53.9% female and 93.2% Caucasian ethnicity (Table 1). Among these participants, 2082 (3.9%) had at least one baseline diagnosis of ARED, demonstrating an older CA, male sex, non-Caucasian ethnicity, lower education and income, and higher body mass index (BMI) (Supplementary Tables S1). Cataract and AMD were more prevalent in women, reflecting typical sex disparities37,38 (Supplementary Tables S2–S5). During follow-up, a total of 778.1 thousand person-years were accumulated. Among participants free of AREDs at baseline, 6,623 (12.6%) developed cataract, 217 (0.4%) developed DR, 1042 (2.0%) developed AMD, and 888 (1.7%) developed glaucoma, with similar characteristic patterns in prevalent cases (Supplementary Tables S6–S10). Data from the GDES and HOPE served as external replication. Table 1Baseline characteristics of all study participantsCharacteristicsaUKBGDESHOPENo. of participants53,0141584430Age, years56.8 ± 8.264.4 ± 7.564.2 ± 7.2Sex Female28,580 (53.9%)904 (57.1%)250 (58.1%) Male24,434 (46.1%)680 (42.9%)180 (41.9%)Ethnicity Caucasian49,435 (93.2%)0 (0.0%)0 (0.0%) East Asian986 (1.9%)1584 (100.0%)430 (100.0%) Other/mixed2333 (4.4%)0 (0.0%)0 (0.0%) Unknown260 (0.5%)0 (0.0%)0 (0.0%)Educational attainment No post-compulsory educationb23,086 (43.5%)539 (34.0%)158 (36.7%) A/AS levels or equivalent5806 (11.0%)590 (37.2%)170 (39.5%) College/university16,923 (31.9%)445 (28.1%)102 (23.7%) Others7199 (13.6%)10 (0.6%)0 (0.0%)Smoking Never28,670 (54.1%)1394 (88.0%)380 (88.4%) Ever/current24,089 (45.4%)190 (12.0%)50 (11.6%) Missing255 (0.5%)0 (0.0%)0 (0.0%)Body mass index, kg/m227.5 ± 4.824.6 (3.2)24.4 (3.2)aContinuous data are presented as mean ± standard deviation and categorical variables as number (percentage).bO levels/GCSEs/CSE or equivalent. Baseline characteristics of all study participants aContinuous data are presented as mean ± standard deviation and categorical variables as number (percentage). bO levels/GCSEs/CSE or equivalent. The participants were randomly divided into five equal-sized, non-overlapping datasets for the ML training and cross-dataset validation of the proteomic aging clock (ProAge; Methods), with baseline characteristics balanced across all sets (Supplementary Table S11). Prior to all fold training, individuals with any reported medical history or diagnosed conditions were excluded to ensure that the model captures aging rather than disease signals (Methods). The five-fold training process generated consistent results, with non-zero coefficients for 2139–2190 proteins across folds (Fig. 2a). The ProAge exhibited a robust correlation with CA in both the training (mean absolute error [MAE] = 2.23 years; r = 0.94, 95% confidence interval [CI]: 0.93–0.94 [P < 0.001]) and test datasets (MAE = 2.36 years; r = 0.93, 95% CI: 0.92–0.93 [P < 0.001]) of the first fold, as well as in the out-of-fold assessment (MAE = 2.38 years; r = 0.93, 95% CI: 0.92–0.93 [P < 0.001]) (Fig. 2c, e). Given that the cardiometabolic panel captured most of the top-ranked proteins and that each Olink panel operates independently, a simplified clock was trained on the Olink cardiometabolic panel (cProAge; Methods). The cProAge also demonstrated a strong correlation with CA (out-of-fold MAE = 3.25 years; r = 0.87, 95% CI: 0.87–0.88 [P < 0.001]), performing on par with the ProAge trained on full panels (Fig. 2b, c, f). Proteomic clock performance and heterogeneity across sexes, ethnicities, and prevalent AREDs. a–h Proteomic clocks trained on the full protein panel (ProAge) (a, e) and the cardiometabolic protein panel (cProAge) (b, f) demonstrated robust associations with chronological age in the out-of-fold assessment (e, f), with individual protein contributions shown in a, b. Distributions of chronological age and proteomic age, and distributions of ProAge among participants with prevalent AREDs are shown in (c, g). Men and Caucasians exhibited older proteomic age compared to women and other ethnicities (d, h). i–l Participants with prevalent AREDs exhibited older proteomic age compared to those without. Source data are provided in Supplementary Data. MAE mean absolute error, CA chronological age, ARED age-related eye disease, DR diabetic retinopathy, AMD age-related macular degeneration. Acceleration of proteomic aging supports eye aging Acceleration of proteomic aging (ProAgeAccel; Methods) provides molecular support for eye aging signals and their fine-grained features, independent of CA and other covariates (Fig. 3 and Methods). We developed a DL-driven aging clock (RetinAge) derived from fundus photographs (Fig. 3a and Methods), previously shown to capture eye aging signals.30–35 Notably, both ProAgeAccel (β = 0.04, 95% CI: 0.02 to 0.06 [PFDR = 0.001]) and cProAgeAccel (β = 0.06, 95% CI: 0.03–0.09 [PFDR < 0.001]) demonstrated significant associations with accelerated eye aging (RetinAgeAccel), indicating that the robust signals captured by proteomic aging extend to the eye (Fig. 3d). Further extraction of fine-grained phenotypes from fundus photographs using a DL-driven pipeline36,39 (Fig. 3b and Methods) revealed that each 1-SD increase in ProAgeAccel was associated with greater vessel tortuosity density (β = 0.09, 95% CI: 0.05–0.14 [PFDR < 0.001]), alongside reductions in vessel density (VD) (β = −0.11, 95% CI: −0.15 to −0.07 [PFDR < 0.001]), fractal dimension (β = −0.12, 95% CI: −0.16 to −0.08 [PFDR < 0.001]), central retinal artery equivalent (CRAE) (β = −0.06, 95% CI: −0.10 to −0.01 [PFDR = 0.026]), and central retinal vein equivalent (CRVE) (β = −0.07, 95% CI: −0.12 to −0.02 [PFDR = 0.005]). These associations were consistently observed with cProAgeAccel (Fig. 3d). Acceleration of proteomic aging supports eye aging. a A DL-driven eye aging clock (RetinAge) was employed to quantify eye aging. b A DL-driven pipeline (AutoMorph) was employed to segment vessels and extract fine-grained phenotypes from fundus photographs, including density (VD), tortuosity (tortuosity density), complexity (fractal dimension), and caliber (CRAE and CRVE). c Tomographic imaging combined with a graph search–based segmentation algorithm was employed to assess neuroretinal degeneration and segment retinal layer. Thickness measurements were obtained for the RNFL, GC-IPL, PL, and RPE as these layers are known to be affected by aging and AREDs. d Associations of proteomic aging acceleration with signals of eye aging, adjusted for chronological age, sex, ethnicity, educational attainment, household income, smoking status, body mass index, hypertension, and diabetes. Data are presented as per-SD change of ProAgeAccel or cProAgeAccel. Parts of a were created from BioRender (biorender.com). *P values were adjusted for multiple testing using the Benjamini–Hochberg method within each analysis group. Bold indicates statistical significance. Source data are provided in Supplementary Data. VD vessel density, CRAE central retinal artery equivalent, CRVE central retinal vein equivalent, RNFL retinal nerve fiber layer, GC-IPL ganglion cell-inner plexus layer, PL photoreceptor layer, RPE retinal pigment epithelium, CI confidence interval. Moreover, ProAgeAccel was significantly associated with structural degeneration in neuroretinal layers, independent of CA. Among the various layers segmented and examined via a validated graph search-based algorithms (Fig. 3c and Methods), the retinal nerve fiber layer (RNFL) demonstrated the most prominent association, with a −0.24 (95% CI: −0.34 to −0.15 [PFDR < 0.001]) and a −0.14 (95% CI: −0.23 to −0.06 [PFDR = 0.002]) μm thinning in thickness per 1-SD increase in ProAgeAccel and cProAgeAccel, respectively. Photoreceptor layer (PL) thinning was more strongly associated with increases in cProAgeAccel (β = −0.16, 95% CI: −0.25 to −0.07 [PFDR = 0.001]) than in ProAgeAccel (β = −0.09, 95% CI: −0.17 to −0.01 [PFDR = 0.033]), which may reflect the metabolically active nature of photoreceptors that navigate enrichment of functionally relevant proteins in the cardiometabolic panel assay. While increased ProAgeAccel was associated with thinning of the ganglion cell-inner plexus layer (GC-IPL) (β = −0.15, 95% CI: −0.25 to −0.04 [PFDR = 0.017]), no significant association was found for cProAgeAccel (PFDR = 0.849). Neither ProAgeAccel (P = 0.677) nor cProAgeAccel (PFDR = 0.997) showed significant associations with the retinal pigment epithelium (RPE), though trends towards layer thinning were consistently observed for both proteomic age accelerations (Fig. 3d). Acceleration of proteomic aging predicts future AREDs Participants with prevalent AREDs exhibited a significant population-level shift toward older proteomic age compared with those without (Fig. 2g, i–l). Specifically, cataract patients were 4.93 (95% CI: 4.45–5.42 [PFDR < 0.001]) and 4.42 (95% CI: 3.93–4.92 [PFDR < 0.001]) years older in ProAge and cProAge, respectively, compared to their non-cataract counterparts (Fig. 2i). This discrepancy persisted even after adjusting for CA, sex, ethnicity, educational attainment, household income, smoking status, body mass index (BMI), systolic blood pressure (SBP), and HbA1c (βProAge = 0.52, 95% CI: 0.41–0.62 [PFDR < 0.001]; βcProAge = 0.33, 95% CI: 0.26–0.41 [PFDR < 0.001]). Similar proteomic aging was also observed for prevalent DR, AMD, and glaucoma, revealing biological aging captured by proteomics beyond CA (Fig. 2i–l and Supplementary Table S12). Subgroup analyses confirmed consistent patterns of proteomic aging that were independent of CA in prevalent AREDs across sex and ethnic strata (Supplementary Tables S13, S14), though men and Caucasians generally exhibited older ProAge and cProAge compared to women and other ethnicities (Fig. 2d, h). Acceleration of proteomic aging (ProAgeAccel) was predictive of incident AREDs over the follow-up period, independent of CA (Fig. 4). The most prominent association per 1-SD increase in ProAgeAccel was observed for incident DR (hazard ratio [HR] = 1.30, 95% CI: 1.20–1.40 [PFDR < 0.001]), followed by incident AMD (HR = 1.18, 95% CI: 1.08–1.27 [PFDR < 0.001]) and cataract (HR = 1.11, 95% CI: 1.04–1.19 [PFDR = 0.003]), with incident glaucoma holding the least yet significant elevated risk (HR = 1.09, 95% CI: 1.06–1.13 [PFDR < 0.001]) (Fig. 4a). Though cProAge showed slightly inferior performance in correlating with CA, the cProAgeAccel revealed strong associations with incident AREDs on par with ProAgeAccel (Fig. 4b). A 1-SD increase in cProAgeAccel was associated with a 32% elevated risk of DR (HR = 1.32, 95% CI: 1.22–1.42 [PFDR < 0.001]), 17% for AMD (HR = 1.17, 95% CI: 1.07–1.27 [PFDR < 0.001]), 10% for cataract (HR = 1.10, 95% CI: 1.03–1.18 [PFDR = 0.008]), and 10% for glaucoma (HR = 1.10, 95% CI: 1.07–1.14 [PFDR < 0.001]). These findings remained robust in Fine–Gray competing-risk analyses with mortality treated as a competing event (Supplementary Table S15). Acceleration of proteomic aging predicts future AREDs. a, b Hazard ratios for incident AREDs per 1-SD increase in ProAgeAccel (a) and cProAgeAccel (c), adjusted for chronological age, sex, ethnicity, educational attainment, household income, smoking status, body mass index, systolic blood pressure, and HbA1c. c–j Cumulative hazard for incident AREDs over the follow-up period, stratified by deciles of ProAgeAccel (c, e, g, i) and cProAgeAccel (d, f, h, j). Asterisk indicates statistical significance. P values were adjusted for multiple testing using the Benjamini–Hochberg method within each analysis group. Source data are provided in Supplementary Data. ARED age-related eye disease, DR diabetic retinopathy, AMD age-related macular degeneration. Cumulative hazard trajectories revealed a pronounced risk stratification across deciles of proteomic aging acceleration for all incident AREDs (Fig. 4c–j). Compared to participants in the bottom decile of ProAgeAccel and cProAgeAccel, those at the top had a 28.68-fold (95% CI: 22.15–37.15 [PFDR < 0.001]) and 20.94-fold (95% CI: 16.66–26.32 [PFDR < 0.001]) higher risk of developing cataract, respectively, over a span of 15 years (Fig. 4c, d). For AMD, the top decile of ProAgeAccel and cProAgeAccel conferred a 28.68-fold (95% CI: 22.15–37.15 [PFDR < 0.001]) and 20.94-fold (95% CI: 16.66–26.32 [PFDR < 0.001]) higher risk, respectively (Fig. 4g, h). For both DR and glaucoma, the ProAgeAccel decile captured a twelvefold increased risk (HRDR = 12.83, 95% CI: 4.16–35.67 [PFDR < 0.001]; HRglaucoma = 12.28, 95% CI: 7.77–19.43 [PFDR < 0.001]), while the cProAgeAccel decile provided even more distinct risk separation for DR (HR = 14.36, 95% CI: 5.19–39.76 [PFDR < 0.001]) than ProAgeAccel, though it was less informative for stratifying glaucoma (HR = 8.36, 95% CI: 5.59–12.51 [PFDR < 0.001]) (Fig. 4e, f, i, j). These stratified patterns were supported by tertile-based analyses that showed consistent risk separation beyond extreme tail contrasts (Supplementary Table S16 and Supplementary Fig. S2) and were broadly consistent across sex and ethnicity subgroups (Supplementary Figs. S3–S8). Validation in the GDES with rigorous eye assessment A total of 1,584 GDES participants who underwent the Olink Explore cardiometabolic PEA were included for internal domain adaptation and external validation (Fig. 5 and Methods). The adaptation (20%, n = 316) and withheld test (80%, n = 1268) datasets share comparable baseline characteristics (Supplementary Table S17). When applied to the GDES, the cProAge clock demonstrated a robust correlation with CA in both the adaptation (MAE = 3.39 years; r = 0.85, 95% CI: 0.82–0.88 [P < 0.001]) and the withheld test (MAE = 3.59 years; r = 0.80, 95% CI: 0.78–0.81 [P < 0.001]) datasets (Fig. 5a). Sensitivity analyses applying the cProAge clock directly to the withheld test set without any adaptation yielded consistent performance (Supplementary Table S18). A total of 166 out of the 226 (73.4%) proteins from the cardiometabolic panel with non-zero coefficient replicated individual protein contributions as observed in the UKB (Fig. 5b). All GDES participants underwent rigorous eye assessment at baseline and each follow-up visit for ARED determination (Methods). Among the withheld test set, 728 had at least one baseline diagnosis of AREDs (Supplementary Table S19), where advanced proteomic aging with prevalent AREDs was replicated (Fig. 5c–f). These associations also remained significant after adjusting for CA, sex, educational attainment, household income, smoking status, BMI, SBP, and HbA1c (Fig. 5g). External validation demonstrates generalization to a population with distinct demographic profile. a, b The proteomic clock (cProAge) demonstrated robust association with CA in the fully-withheld external validation in the GDES (a), with individual protein contributions shown in (b). c–f Participants with prevalent AREDs exhibited older proteomic age compared to those without. g Coefficients for prevalent AREDs per 1-SD increase in cProAge, adjusted for CA, sex, educational attainment, household income, smoking status, body mass index, systolic blood pressure, and HbA1c. h Hazard ratios for incident AREDs per 1-SD increase in cProAgeAccel, adjusted for the same covariates as above. i Coefficients for neuroretinal thinning derived from OCT imaging per 1-SD increase in cProAgeAccel, adjusted for the same covariates as above. j Hazard ratios for rapid microvascular rarefaction derived from OCT angiography imaging per 1-SD increase in cProAgeAccel, adjusted for the same covariates as above. k–l Examples of microvascular rarefaction between two participants of the same CA but differing cProAgeAccel. Note that the larger FAZ area and more deficient VD in participant B (CA = 54.3 years, cProAgeAccel = 4.7 years), who shared the same baseline CA as participant A (CA = 54.6 years, cProAgeAccel = 1.2 years) but exhibited more accelerated proteomic aging. Source data are provided in Supplementary Data. CA chronological age; ARED age-related eye disease, DR diabetic retinopathy, AMD age-related macular degeneration, OCT optical coherence tomography, SCP superficial capillary plexus, DCP deep capillary plexus. During follow-up, a total of 9.6 thousand person-years were accumulated. Among participants free of baseline AREDs, 206 (18.5%) GDES participants developed cataract, 228 (17.1%) developed DR, 260 (20.4%) developed AMD, and 37 (2.4%) developed glaucoma. After adjusting for the same covariates as in the UKB, the associations between cProAgeAccel and the incidence of the four major AREDs were confirmed in the GDES (Fig. 5h), where the corresponding HRs for these associations were 1.25 (95% CI: 1.05–1.49 [PFDR = 0.034]) for incident cataract, 1.22 (95% CI: 1.03–1.47, PFDR = 0.042) for incident DR, 1.19 (95% CI: 1.01–1.42 [PFDR = 0.045]) for incident AMD, and 1.37 (95% CI: 1.06–1.76 [PFDR = 0.034]) for incident glaucoma, per 1-SD increase in cProAgeAccel, respectively. In addition, the associations of increased cProAgeAccel with RetinAgeAccel (β = 0.13, 95% CI: 0.05 to 0.21 [PFDR = 0.003]), VD (β = −0.31, 95% CI: −0.40 to −0.22 [PFDR < 0.001]), tortuosity density (β = 0.14, 95% CI: 0.06–0.21 [PFDR < 0.001]), fractal dimension (β = −0.24, 95% CI: −0.31 to −0.16 [PFDR < 0.001]), CRAE (β = −0.22, 95% CI: −0.30 to −0.13 [PFDR < 0.001]), CRVE (β = −0.23, 95% CI: −0.31 to −0.15 [PFDR < 0.001]), and PL (β = −0.21, 95% CI: −0.28 to −0.13 [PFDR < 0.001]) and RNFL thinning (β = −0.67, 95% CI: −0.95 to −0.39 [PFDR < 0.001]) were validated, whereas the associations with GC-IPL (PFDR = 0.128) and RPE thinning (P = 0.440) did not reach statistical significance, similar to the findings in the UKB (Fig. 5i). cProAgeAccel signals longitudinal microvascular rarefaction in the GDES and HOPE Leveraging annual retinal microvascular phenotyping data from the GDES (Methods), cProAgeAccel emerged as a predictor of rapid microvascular rarefaction in the macula (Figs. 5j and 6b). Rapid enlargement of the foveal avascular zone (FAZ) area at both the superficial capillary plexus (SCP) and the deep capillary plexus (DCP) was associated with deviations in cProAgeAccel, conferring HRs of 1.09 (95% CI: 1.01–1.18 [PFDR = 0.044]) and 1.20 (95% CI: 1.10–1.30 [PFDR < 0.001]) per 1-SD increase in cProAgeAccel, respectively. Across a broader macular region (6 × 6 mm), increased cProAgeAccel was associated with a rapid decline in average VD at the SCP (HR = 1.16, 95% CI: 1.07–1.25 [PFDR < 0.001]), whereas the association did not reach statistical significance at the DCP layer (PFDR = 0.293). Further geographic subdivision of the retina revealed consistent rarefaction in both the parafoveal (1–3 mm; HR = 1.17, 95% CI: 1.08–1.26 [PFDR < 0.001]) and perifoveal VD (3–6 mm; HR = 1.17, 95% CI: 1.09–1.27 [PFDR < 0.001]) at the SCP layer per 1-SD increase in cProAgeAccel, and a significant DCP rarefaction in the perifoveal VD (HR = 1.12, 95% CI: 1.03–1.21 [PFDR = 0.006]). Figure 5k–l shows representative images of microvascular rarefaction from two participants with the same CA but differing cProAgeAccel. Acceleration of proteomic aging signals microvascular rarefaction. a Ultra-widefield microvascular imaging was performed in the HOPE study and multiple DL-driven pipelines were employed for phenotyping microvascular rarefaction. b Ultra-widefield microvascular phenotyping in the HOPE study replicated the findings from the GDES. Analyses in both cohorts were adjusted for CA, sex, educational attainment, household income, smoking status, body mass index, systolic blood pressure, and HbA1c, and data are presented as per-SD change of cProAgeAccel. * In the GDES, participants underwent annual OCT angiography and were evaluated for longitudinal rapid microvascular rarefaction. † In the HOPE study, participants underwent ultra-widefield OCT angiography at baseline and were evaluated for cross-sectional microvascular rarefaction. ‡ P values were adjusted for multiple testing using the Benjamini–Hochberg method within each analysis group. Bold indicates statistical significance. Source data are provided in Supplementary Data. SCP superficial capillary plexus, DCP deep capillary plexus, FAZ foveal avascular zone, GDES Guangzhou Diabetic Eye Study, HOPE High-definition Oculo-Phenomic Evaluation, CI confidence interval. Ultra-widefield microvascular phenotyping in the HOPE study replicated findings from the GDES (Fig. 6). This ultra-widefield modality extends visualization of the peripheral microvasculature beyond the reach of conventional imaging (Fig. 6a and Methods). Of 430 HOPE participants who underwent both Olink Explore cardiometabolic PEA and ultra-widefield microvascular phenotyping at baseline, their characteristics were comparable to those in the GDES (Supplementary Table S20). When applied to the HOPE study, the cProAge clock again showed a robust correlation with CA (MAE = 3.15 years; r = 0.84, 95% CI: 0.81–0.86 [P < 0.001]). Per 1-SD increase in cProAgeAccel was associated with a larger FAZ area at both the SCP (β = 0.18, 95% CI: 0.07–0.28 [PFDR = 0.002]) and the DCP (β = 0.19, 95% CI: 0.08–0.29 [PFDR = 0.002]) layers, reflecting consistency across both imaging modalities (Fig. 6b). For a much wider area around the macula (12 × 12 mm), an increase in cProAgeAccel was associated with a reduction in the average SCP VD (β = −0.16, 95% CI: −0.26 to −0.06 [PFDR = 0.003]). This was accompanied by a decrease in both the parafoveal (1–6 mm; β = −0.15, 95% CI: −0.26 to −0.05 [PFDR = 0.004]) and the perifoveal (6–12 mm; β = −0.17, 95% CI: −0.27 to −0.06 [PFDR = 0.001]) SCP VD. DCP rarefaction in the perifoveal VD was also replicated (β = −0.12, 95% CI: −0.23 to −0.02 [PFDR = 0.039]). Discussion This study represents the first comprehensive investigation into the role of proteomic aging in eye health and AREDs. Leveraging data from three large-scale, cross-national cohorts combined with DL phenotyping, we present comprehensive evidence of signal concordance between ML-driven proteomic aging and eye aging, while establishing proteomic age acceleration as a robust biomarker for predicting incident AREDs independent of CA. We further developed a parsimonious aging model with streamlined proteomic assays, demonstrating competent performance and adaptability across sex and ethnic groups, both within and beyond the UKB. By integrating tomographic imaging and graph search-based segmentation, we show that accelerated proteomic aging is associated with neuroretinal thinning. These findings were validated in an independent cohort of distinct ethnicity, where rigorous and standardized multimodal imaging and diagnostic procedures were applied. In addition, microvascular phenotyping in the GDES and HOPE provides additional evidence of microvascular rarefaction across proteomic aging. Together, these findings highlight the potential of ML-based proteomic aging as a powerful and scalable tool for tracking eye aging and AREDs, while offering novel insights into the shared aging processes that underline diverse ocular pathologies. Circulating proteomics provides a complementary approach to characterizing biological aging.13 This is particularly compelling as minimally invasive blood sampling reflects proteins from virtually all organs and cell types, which directly regulates lifespan and overall health.13,16 In a landmark study involving plasma proteomic profiling of ~5000 participants, Oh et al.19 developed organ-specific aging models using least absolute shrinkage and selection operator (LASSO) algorithm across 11 organs and uncovered mechanistic connections between proteomic aging and the brain, heart, and kidneys. Leveraging a multicohort study spanning the UK, China, and Finland, Argentieri et al.23 demonstrated that ML-driven proteomic aging is strongly associated with mortality, multimorbidity involving 26 chronic diseases across multiple systems, and their effects on physical function, frailty, and cognitive status. Goeminne et al.21 further showed that elastic net (ElasticNet)-based proteomic models, whether trained on chronological age or mortality, epitomize disease onset through accelerated aging of organismal systems. Building upon these contemporary findings, our study extends this framework to the eye and provides the first comprehensive evidence linking ML-driven proteomic aging to AREDs. These findings underscore shared aging mechanisms across organs and highlighting the relevance of whole-organism aging processes for eye health and disease. We further show that proteomic aging signals align with DL-derived eye aging phenotypes. Using a DL-driven model applied to the fundus photographs, we previously developed and validated the RetinAge clock as a biomarker of eye aging.30–35 Here, we demonstrate that the acceleration of proteomic aging is associated with accelerated RetinAge, indicating that signals in the circulating proteome support eye aging. Another interesting observation is that the cardiometabolic-focused proteomic clock showed an even stronger association with RetinAge. This was consistently observed when analyzing fine-grained vascular phenotypes extracted from fundus photographs using a DL-driven pipeline,36,39 which is likely attributed to the vascular information inherently captured by fundus photographs. Retinal vascular features such as VD, tortuosity density, fractal dimension, CRAE, and CRVE are among the most informative features derived from fundus photographs, representing the density, tortuosity, complexity, and caliber of the retinal vasculature. Previous evidence has demonstrated that reduced VD, complexity, and caliber, coupled with increased tortuosity are implicated in the development of AREDs,40–44 in line with our observed associations with accelerated proteomic aging. These results provide structural and functional interpretability for proteomic aging signals. Proteomic aging was also associated with neuroretinal degeneration and microvascular rarefaction. Leveraging tomographic imaging and graph search-based segmentation from both the UKB and the GDES, we explored and validated the association between proteomic age acceleration and layer-specific retinal thinning. While the UKB’s spectral-domain optical coherence tomography (OCT) permits satisfactory in vivo retinal imaging, its primary limitation is the signal attenuation with increasing depth due to the sensitivity roll-off effect.45 In contrast, the swept-source OCT employed in the GDES offers deeper penetration attributed to its longer wavelengths.45 The faster scanning speed combined with an eye-tracking system further minimizes motion artifacts, a major challenge degrading OCT image quality.46 This is crucial, as improved raw image quality enables more accurate layer segmentation and quantification, where thinning of the neuroretina rather than the epithelium was replicated. As we and others have previously demonstrated the independent association of PS and ganglion cell complex thinning with mortality,47–49 accelerated proteomic aging may increase the susceptibility of the neuroretina to damage due to a reduced ability to cope with oxidative stress. This parallels findings in other neurodegenerative diseases such as Alzheimer’s and Parkinson’s disease,50–52 supporting generalized neuronal damage throughout the brain and retina with biological aging. Through OCT angiography in the GDES and HOPE, we demonstrated that accelerated proteomic aging is linked to retinal microvascular rarefaction. As an emerging interferometric-based angiographic imaging modality, OCT angiography noninvasively maps the three-dimensional angioarchitecture to millimeter depth and has proven effective in revealing functional changes in the microvasculature across ocular and neurodegenerative diseases.45 Our study is the largest to date linking proteomic data with longitudinal OCT angiography, addressing a previous gap posed by small sample sizes (<~500) due to the relatively recent development of this angiographic modality.53 Interestingly, the most consistent findings in the SCP microvascular rarefaction anatomically corresponded to the most prominent effect on RNFL thinning, highlighting the interconnected coupling of neural and vascular degeneration. These findings align with earlier studies on neurodegenerative diseases, where RNFL degeneration and microvascular rarefaction at the SCP were confirmed.54,55 For instance, while most studies have noticed RNFL thinning and SCP rarefaction in Parkinson’s disease, there is considerable variability in findings related to the DCP.55,56 This discrepancy may be explained by the stronger autoregulatory capacity of the DCP that buffers the impact of aging-related pathologies.55 Such findings were further corroborated by ultra-widefield microangiographic vascular phenotyping in the HOPE study, emphasizing the coupled neural and vascular decline in aging. We also developed a streamlined cardiometabolic proteomic clock using the same ElasticNet algorithm, which reduced assay costs while maintaining comparable performance. This simplified approach reduced experimental and computational costs by ~90%, with only a modest 5% compromised in performance when compared to the full model’s correlation with CA. Encouragingly, the acceleration of the parsimonious proteomic clock proved to be a robust biomarker that revealed consistent associations with both prevalent and incident AREDs, on par with the full proteomic clock. For certain outcomes such as DR, acceleration of the parsimonious clock provided an even more distinct risk stratification. Given that DR is a vascular complication of diabetes, one of the most prototypical cardiometabolic disorders,57 the superior performance of the cardiometabolic-based proteomic clock in stratifying DR risk is expected. Moreover, PL thinning was more strongly associated with the cardiometabolic-based proteomic clock than with the full model, aligning with the metabolically active nature of photoreceptors and their role in DR pathophysiology.58 Validation in an independent cohort with standardized ocular grading systems adds robustness to our findings. A major limitation of previous studies in biobank-scale settings has been their reliance on crude disease ascertainment based on inpatient medical records.19–21,23 While the large sample sizes increase statistical power, this approach tends to underestimate the prevalence and incidence of diseases, particularly because many ophthalmic conditions are managed in outpatient and primary eye-care settings.59 Other studies have relied on self-reported questionnaires, which introduce recall bias and contribute to poor reproducibility in large-scale research.60 In this context, GDES provided a standardized phenotyped validation setting in which AREDs were assessed through comprehensive ocular examinations and grading following the Lens Opacities Classification System II (LOCS II), modified Early Treatment Diabetic Retinopathy Study (ETDRS) system, Wisconsin Age-Related Maculopathy Grading System (WARMGS), and established glaucomatous optic neuropathy (GON) criteria.61–64 This approach allowed for more reliable effect size estimation, especially given the insidious nature of these conditions where early cases were barely noticeable for a patient to seek medical attention and thus remain unrecorded.65–68 Cross-cohort comparisons should therefore be interpreted with appropriate caution, particularly when comparing absolute effect sizes derived from different ascertainment strategies. UKB provided statistical power for discovery in a biobank-scale prospective cohort, whereas GDES allowed us to examine whether the same proteomic aging signal could be reproduced when outcomes were defined through standardized ophthalmic assessment. External validation was therefore intended to assess consistency in the direction and pattern of associations, rather than to assume directly interchangeable effect sizes across cohorts. While medical record-based definitions may underestimate disease associations when outcome misclassification is largely nondifferential, they nevertheless introduce uncertainty arising from health-care utilization, referral pathways, clinical documentation, and coding practices. Replication in the GDES where outcomes were defined through standardized clinical grading therefore strengthens confidence that the observed proteomic aging signal extends beyond UKB-specific ascertainment patterns and remains detectable in clinically defined AREDs. Another interesting observation warrant careful interpretation is the aging disparities across sex and ethnicity. In this study, males and Caucasians tended to exhibit greater proteomic aging even after adjustment for chronological age in the UKB cohort. A similar trend was observed for sex differences in the GDES cohort, although statistical significance was not reached, likely reflecting reduced statistical power due to the smaller sample size. These findings are concordant with prior evidence from other molecular aging biomarkers. DNA methylation-based clocks have consistently reported greater epigenetic aging in men than in women, even after accounting for major demographic and clinical factors.69,70 Complementary evidence from telomere biology further indicates that women have longer telomeres than men.71 Evidence regarding ethnic differences in biological aging is more heterogeneous, but studies have reported lower epigenetic aging in Hispanic and African Americans compared with Caucasians.69 However, despite these intriguing observations, it remains difficult to disentangle subtle population-specific aging trajectories from effects related to training data composition at the current stage. Future work leveraging larger and more diverse cohorts will be essential to clarify how systemic biological aging manifests across different populations. Several mechanisms may underlie the link between proteomic aging and AREDs. Experimental models and human studies have identified loss of proteostasis as a key hallmark of aging,18 driving structural and functional degeneration in the eye.72–74 This in turn heightens susceptibility to oxidative stress, mitochondrial dysfunction, and metabolic damage that activate age-related pathophysiology and conditions. While the conceptual framework of rejuvenation strategies for age-related disorders was proposed over a decade ago,75 only recently has a landmark trial demonstrated the potential of targeting aging by inhibiting BCL-2,73 a key protein that also promoted proteomic aging in our study, in advanced stages of DR. Their preclinical and phase 1 trials provided promising evidence of reduced local inflammation and vascular regeneration with improved barrier function by inhibiting this age-related protein.73 Moreover, other proteins contributing to proteomic aging are linked to established mechanisms in AREDs.76–78 For instance, EphA receptors contribute to glaucoma by inhibiting axon regeneration and promoting glial scar formation,79 while the imbalance between MMPs and TIMPs in AMD leads to extracellular matrix degradation and vessel wall pathologies,80 consistent with their roles in proteomic aging observed here. This work has important clinical implications when proteomic aging is interpreted as a population-level biomarker rather than an individual diagnostic test. First, the proteomic clock may serve as a biological reference to contextualize eye health within broader systemic aging processes, thereby informing lifestyle modifications aimed at promoting eye health and facilitating timely ophthalmic evaluation for the early detection of AREDs. Second, given its potential to predict future diseases, the proteomic clock may contribute to risk stratification at the population level that supports public health policies and preventive strategies. Third, the proteomic clock may complement ophthalmic assessments by providing biological context on systemic aging status. For instance, among individuals with similar chronological age and ocular phenotype, differences in systemic biological aging may reflect heterogeneity in physiological reserve, with potential relevance for treatment tolerance and clinical outcomes. Fourth, the proteomic clock may offer value in the design and conduct of clinical trials. Given the slow progression of many AREDs and the challenges of detecting meaningful changes over short trial periods, stratifying patients based on molecular aging markers may improve trial efficiency by enriching individuals at the optimal disease stage for intervention. Importantly, proteomic aging is not intended to replace imaging-based measures such as RetinAge, nor to serve as a superior diagnostic test for eye aging. Instead, it complements ocular imaging by providing molecular context that imaging-derived phenotypes cannot directly resolve. Ocular imaging summarizes the downstream, tissue-level manifestations of aging, such as structural and vascular patterns that reflect cumulative stress and remodeling over the lifespan. In contrast, the circulating proteomic clock quantifies upstream systemic aging processes that are distributed across organs and cell types, encompassing inflammatory, metabolic, vascular, and extracellular matrix-related pathways that may shape ocular susceptibility. Linking proteomic aging with RetinAge and other imaging phenotypes therefore enables a cross-modal interpretation of eye aging, where similar retinal phenotypes may arise from distinct systemic aging states, and conversely, accelerated systemic aging may be detectable before overt ocular changes become apparent. Together, imaging and proteomic clocks characterize complementary yet interconnected layers of aging—phenotypic expression within the eye and systemic molecular state across the organism—thereby supporting more biologically grounded interpretation of eye health and disease. In contrast to analyses based on small sample sizes with limited insights into aging biology, large-scale, high-throughput proteomics combined with AI offers a comprehensive view of protein heterogeneity across individuals in varying physiological states. Additional strengths of this study include extensive follow-up, a multinational and multiethnic perspective, a pioneering approach linking proteomic aging to eye diseases, a comprehensive ocular endophenotype collection, and external validation with rigorous diagnostic procedures. However, certain limitations should be acknowledged. First, while the Olink platform offers a broad assessment of proteomics, it does not encompass the entire human proteome. Although we expect that more expansive protein coverage would further improve the model, the proof-of-concept is platform agnostic. Second, the proteomic age models were trained on a Caucasian-skewed population. Although our findings were validated in two independent East Asian cohort, future studies should assess the applicability across more ethnically and geographically diverse populations. Third, participants in both cohorts were middle-aged or older adults, and a long-term cohort spanning the entire lifespan is warranted to clarify the impact of accelerated proteomic aging on AREDs before age 50. Fourth, while rigorous adjustment for confounders strengthens our confidence in the findings, causality should not be inferred at this stage. Potential residual confounders, as with any observational study, cannot be fully ruled out. Finally, while an economical model was developed, high-throughput proteomics remains costly ($500–1000 per assay). Overall, future studies should expand to broader proteome coverage, more diverse populations, younger cohorts, and incorporate cost-reducing technological advances. In summary, large-scale proteomics combined with ML modeling and DL phenotyping provide novel insights into eye aging and AREDs within a broader framework of systemic biological aging. Our findings reveal a concordance between proteomic aging and eye aging signals, establishing proteomic age acceleration as an effective biomarker for tracking eye aging and AREDs. By integrating structural tomographic and angiographic imaging with multiple AI-driven pipelines, we further show how accelerated proteomic aging is linked to neuroretinal degeneration and microvascular rarefaction, revealing a coupling of neural and vascular degeneration throughout the aging process. Proteomic aging thus represents a promising tool for tracking ocular health and informing anti-aging strategies across multiple AREDs. Methods Study population The UKB is an ongoing, prospective, multicenter cohort study that recruited over 500,000 participants aged 40–69 years from across England, Scotland, and Wales between 2006 and 2010.81 A random subset of 53,014 UKB participants underwent proteomic profiling at baseline and were included in the current analysis. This subset was highly representative of the overall UKB population.17 The GDES (ISRCTN15853192) is an ongoing, prospective cohort study that recruited ~3000 participants aged 35–85 in Guangzhou, China, between 2017 and 2019.82,83 A random subset of 1584 GDES participants underwent proteomic profiling at baseline visit. The HOPE study (ISRCTN47133398) is an ongoing, prospective cohort study that recruited over 2000 participants aged 18–85 in Guangzhou, China, between 2023 and 2024. A random subset of 430 HOPE participants underwent proteomic profiling at baseline visit. This study adhered to the Declaration of Helsinki and was approved by the Northwest Multicenter Research Ethics Committee (11/NW/0382) and the Ethics Committee of Zhongshan Ophthalmic Centre (ZOC; 2017KYPJ094 and 2022KYPJ247). Written informed consent was obtained from all participants prior to their enrolment. Proteomic profiling Baseline blood samples were collected from UKB participants at 22 assessment centers across the UK between 2007 and 2010, from GDES participants at ZOC between 2017 and 2021, and from HOPE participants at ZOC between 2023 and 2024. The proteomics assay, data processing, and quality control procedures have been previously detailed.17,84,85 In brief, the Olink Explore PEA was used to quantify specific proteins through matched antibody pairs conjugated to complementary oligonucleotides. Upon proximity, the probes hybridize, facilitating amplification and subsequent quantification via next-generation sequencing, with minimal cross-reactivity.86 This method enabled the capture of 2923 unique proteins across eight panels in the UKB and 361 proteins from the cardiometabolic panel in the GDES and the HOPE. Sample controls ensured precision within plates, and plate controls were used to standardize protein levels across plates. For all three cohorts, the interplate coefficients of variation were < 20% and the intraplate coefficients were < 10%. Raw assay counts were normalized using extension controls and log-transformed to generate Normalized Protein eXpressions (NPXs), minimizing technical variability and enabling comparability across samples. Proteomic aging clock The UKB participants were randomly split into five nearly equal-sized, non-overlapping partitions. The model was evaluated in five-fold nested cross-validation, where one partition was set aside as a test set, and the aggregated remaining partitions were used for training. Within each training fold, participants with more than 50% missing proteomic measurements and proteins with more than 10% missing NPX values were excluded.21 Multiple imputations were then performed using chained equations with random forests on each of the five-fold pairs. Imputation models were fitted on each training set and applied to the respective paired test set. To ensure that the model captured aging rather than disease-driven signals,87 participants with any reported medical history or baseline-diagnosed diseases were excluded from all training sets before model training, with the number of excluded participants varying slightly across folds (fold 1: 3574; fold 2: 3541; fold 3: 3412; fold 4: 3423; fold 5: 3611). However, all individuals, including those with prevalent diseases, were retained in the test sets to maintain representativeness and minimize healthy volunteer bias. This design allowed the model to learn from disease-free individuals while enabling evaluation of its ability to detect deviations from normative aging trajectories. In addition, all predictions were generated from strictly withheld test sets, thereby preventing any risk of data leakage. The ProAge and cProAge models were trained using ElasticNet, a robust ML algorithm widely recognized for its efficacy in modeling proteomic data and estimating biological age.21,88,89 Accelerated proteomic aging was defined as the residual from regressing CA on proteomic age, in accordance with prior work on aging clocks,15,89,90 with positive values indicating accelerated aging (ProAgeAccel or cProAgeAccel). Compared to a simple difference between predicted and CA, this method accounts for systematic prediction bias (e.g., under- or overestimation at age extremes) and isolates individual-level deviations from normative aging trajectories by centering predictions around the expected value for a given CA. By definition and construction, AgeAccel is statistically uncorrelated with CA. External validation was performed on the GDES and HOPE study, where missing data were imputed using the same approach, and model features and parameters derived from the UKB were applied without retraining. To enable application of a single, fixed predictor to external cohorts, we constructed a unified predictor by averaging the coefficients across the five UKB cross-validation folds, with non-selected features assigned a weight of zero. To account for potential distributional shifts, both GDES and HOPE participants were randomly split at a ratio of 1:4, respectively, and domain adaptation was performed by post hoc calibrating of the predictions using a linear regression model fitted on the 20% internal adaptation set, which was then applied to the remaining 80% withheld test set. The withheld test set therefore represents external validation of the UKB-derived model and was used for all subsequent analyses. Sensitivity analyses were performed by applying the UKB-derived predictor directly without domain adaptation. Fundus photography and RetinAge clock In the UKB, 45-degree, non-mydriatic, non-stereoscopic retinal fundus photographs centered on the macula were acquired under dark room conditions using the Topcon 3D OCT-1000 Mk II (Topcon, Japan). In the GDES, stereoscopic fundus photographs of seven standard fields were taken using a Canon CX-1 fundus camera (Canon, Japan) under dark room after pupil dilation. Image quality control protocols, including the exclusion of images with blur, uneven illumination, low contrast, or artifacts, have been detailed in our previous studies.30–35 Quality was further assessed using a validated DL-model with EfficientNet-B4 architecture, with images classified as “reject quality” excluded from all downstream analyses.36 Anatomical segmentation of the retinal vasculature was performed using AutoMorph, a validated DL-driven pipeline for retinal vascular morphology quantification.36 Binary vessel segmentation uses a U-Net variant optimized for multi-scale feature learning, consisting of a segmenter and a discriminator trained with an adversarial learning strategy. For artery/vein segmentation, the model decomposes the multi-class classification into multiple binary tasks, followed by binary-to-multi-class information fusion to correct ambiguities around vessel intersections. Representative vascular features including density (VD), tortuosity (tortuosity density), complexity (fractal dimension), and caliber (CRAE and CRVE) were extracted for quantification. The RetinAge clock is a validated DL-driven biological age predictor that we previously developed on fundus images from 32,282 healthy individuals, leveraging longitudinal pre-training and progressive label distribution learning to better capture its dynamic nature.35 Details of model development and performance are described elsewhere.30–35 Accelerated eye aging was defined as the residual from regressing CA on the RetinAge, in accordance with prior work on biological aging clocks,15,89,90 with positive values indicating accelerated eye aging (RetinAgeAccel). Ascertainment of AREDs The electronic health record-based combined with DL-based diagnoses from the UKB provided an initial, albeit crude, exploration into the relevance of proteomic aging to AREDs, while the standardized ocular assessments conducted in the GDES served to validate these findings. By linking to the Hospital Episode Statistics database, the Scottish Morbidity Record, and the Patient Episode Database, AREDs in the UKB were identified using predefined criteria adapted from prior studies.91–94 Cataract was defined by International Classification of Diseases-10 code H25–H26 and/or a history of cataract surgery;91 DR by ICD-10 code H36.0;92 AMD by ICD-10 code H35.3;93 and glaucoma by ICD-10 code H40, excluding H40.0, and/or history of glaucoma surgery or laser therapy.94 To supplement EHR-based classification, a DL-driven system with an adapted VGGNet architecture that we previously developed for detecting AREDs was applied.27 This system was trained and validated using 494,661 fundus photographs from the Singapore National Diabetic Retinopathy Screening Program and ten additional multiethnic cohorts, with model development and performance described in detail elsewhere.27 This DL system was applied uniformly to both the UKB and GDES. In the UKB, ARED cases were determined by the union of electronic health record-based and DL-derived diagnoses. In the GDES, DL-based outputs served as an initial screen, with final diagnoses adjudicated by standardized protocols incorporating slit-lamp and imaging-based criteria, as detailed below. In the GDES, lens opacities were assessed under slit-lamp following maximum dilation and classified according to the LOCS II.61 Cataracts were defined as any opacity grade ≥2 and/or undergone cataract surgery in either eye.61 Stereoscopic color fundus photographs of seven standard fields and OCT imaging were employed to ascertain DR, AMD, and GON. DR was graded based on the modified Airlie House classification and the ETDRS severity scheme, with the severity levels for both eyes concatenated and greater weight assigned to the eye with greater severity.62 AMD lesions were graded according to the WARMGS, which includes assessment of drusen size, type, and area of involvement, along with pigmentary abnormalities, geographic atrophy, or exudative signs.63 Glaucoma was ascertained by predefined signs of GON, including a vertical cup-disc ratio of the optic nerve head ≥0.9 or inter-eye differences ≥ 0.3, rim width at the superior (11–1 o’clock) or inferior (5–7 o’clock) <0.05 × disc diameter, or nerve fiber layer defects corresponding to rim narrowing or localized notches.64 Structure imaging and segmentation of neuroretina Structure imaging of the neuroretina was obtained using OCT in both cohorts. The OCT scanning protocol for the UKB has been previously described.95–97 In brief, participants underwent OCT scanning using the Topcon 3D OCT-1000 Mk II (Topcon, Japan) in darkened rooms, with a scan density of 512 A-scans × 128 B-scans centered at the fovea. In the GDES cohort, a swept-source OCT (DRI OCT Triton; Topcon, Japan) was employed after pupil dilation.47,48,98 This system offers enhanced penetration (with a 1050 nm wavelength), faster scanning speed (within 1.3 s), and higher scanning density (512 A-scans × 512 B-scans). A built-in fundus camera and eye-tracking system were used to monitor fixation and minimize motion artifacts. Image quality scores (IQSs), inner limiting membrane indicator, validity count, and motion indicators were recorded for quality control, whereby images with an IQS < 45 or those exhibiting poor segmentation or centration (the worst 20% of each indicator) were excluded.95–97 Retinal layer segmentation was performed using the Topcon Boundary Segmentation algorithm, a fully automated, graph search-based method based on multiscale gradient information, with validated accuracy and reproducibility.95–97 The algorithm adopts a two-step segmentation framework that integrates both local and complementary global gradient information to delineate layer boundaries. In brief, local edges are initially detected using a Canny edge detector, while a complementary global gradient map is computed along the axial direction using a larger kernel. These two gradient sources are then combined to construct a graph representation of the image, on which a shortest-path graph search, implemented via dynamic programming, is applied to extract the optimal boundary path. This process is repeated for each anatomical boundary, ensuring consistent and precise layer delineation across the full retinal profile. Measurement of RNFL, GC-IPL, PL, and RPE were obtained for analysis, as these layers are known to be affected by aging and AREDs.58,99–101 Microvascular phenotyping GDES participants underwent interferometric-based OCT angiography using a swept-source OCT (DRI OCT Triton; Topcon, Japan) at each follow-up visit to visualize the retinal microvascular system.82,98 The Angio Macula 6 × 6 mm scan mode centered on the fovea was employed. Ultra-wide field OCT angiography imaging was performed in HOPE participants using a 400-kHz swept-source OCT (TowardPi BMizar; TowardPi, China) at baseline assessment. These scans covered a 24 × 20 mm area centered on the fovea, corresponding to a 120° angular field of view.102 For both dataset, segmentation of the macular region into the SCP and DCP slabs was performed in accordance with prior work.82,103 The SCP was segmented from 2.6 μm below the internal limiting membrane to 15.6 μm below the junction between the inner plexiform and inner nuclear layers, and the DCP from 15.6 μm below the inner plexiform and the inner nuclear layers to 70.2 μm below them.82,103 Following quality control based on predefined criteria and magnification correction using Littman and Bennett’s methods,82,98,104–107 a DL-based pipeline with modified U-Net architecture was employed to segment the FAZ, with validated accuracy and robustness reported elsewhere.108 Quantitative analysis was performed to calculate FAZ area and VD using Fiji (v2.9.0). VD was calculated as the ratio of white pixels representing vessels/flow to the total pixels in the image, while FAZ area is the total number of pixels within the FAZ region, measured on a scaled basis. In the GDES, parafoveal VD was determined by averaging the VD across the four ETDRS inner-ring quadrants (1–3 mm), while perifoveal VD was determined by averaging the VD of the four outer-ring quadrants (3–6 mm). In HOPE, parafoveal and perifoveal regions were defined as 1–6 mm and 6–12 mm concentric ring quadrants, respectively. Rapid microvascular rarefaction was defined as participants exhibiting a top quantile rate of longitudinal FAZ area enlargement or VD decline in the corresponding subregion.82,98 Covariates Face-to-face interviews and questionnaires were administered to all UKB and GDES participants, collecting information on CA, sex, ethnicity, educational attainment, household income, and smoking status, and medical history. Physical and biochemical measurements, including height, weight, blood pressure, and HbA1c levels, were also taken. BMI was calculated as weight in kilograms divided by height in meters squared. Hypertension was defined as an average blood pressure ≥140/90 mmHg, a previous diagnosis, and/or current use of antihypertensive medication.109 Diabetes was defined as a previous diagnosis, current use of insulin or hypoglycemic medication, and/or an HbA1c ≥ 6.5% (48 mmol/mol), as adapted from prior studies.110 Statistical analysis Data analyses were performed using R (v4.2.2). Continuous variables were presented as means with SDs, while categorical variables were presented as numbers with percentages. The Student’s t test and Pearson’s χ2-test were used to compare continuous and categorical variables, respectively. Associations between proteomic aging and prevalent AREDs were evaluated using multiple logistic models, adjusting for CA, sex, ethnicity, educational attainment, household income, smoking status, BMI, hypertension, and diabetes. The effect of a 1-SD increase in ProAgeAccel or cProAgeAccel on incident AREDs, RetinAgeAccel, fine-grained phenotypes on fundus photographs each retinal layer thickness, and macular microvascular rarefaction was assessed using multiple Cox proportional hazards (CPH) models and linear models, where applicable, with the same covariates adjusted as above. As a sensitivity analysis, Fine–Gray competing-risk models were fitted for incident ARED, with death treated as a competing event. For the GDES and HOPE, which are ethnically homogeneous, ethnicity was not included as a covariate. Next, participants were stratified based on ProAgeAccel or cProAgeAccel and cumulative event rates were compared across the top, middle, and bottom deciles for each outcome.111 An alternative tertile-based analysis was performed to assess whether risk separation persisted beyond extreme tail contrasts. The proportional hazards assumption was tested using the Schoenfeld residual method and was satisfied for each CPH model. Subgroup analyses stratified by sex and ethnicity were conducted, using the same approach but excluding sex or ethnicity as covariates. When multiple testing was performed, P values were adjusted using the Benjamini–Hochberg method within each analysis group. A two-sided P < 0.05 was considered statistically significant. Supplementary information 20260429-1730 Supplementary materials Supplementary Data Acknowledgements This study was funded by the National Key R&D Program of China (2023YFC2506100, M.Z.), the GBRCE for Major Blinding Eye Diseases Prevention and Treatment, the Hainan Province Clinical Medical Center, the National Natural Science Foundation of China (82371086, W.W.; 82401298, R.X.), and the Chinese Postdoctoral Science Foundation (2024M763790, R.X.). We thank all the participants and staff of the UKB, GDES, and HOPE. Author contributions Study concept and design: W.W.; Acquisition, analyses, or interpretation: S.Y., W.W.; Drafting of the manuscript: S.Y., W.W.; Critical revision of the manuscript for important intellectual content: Z.X., H.L., ZH.Z., ZI.Z., Y.C., Y.Z., M.Z., M.H.; Statistical analyses: S.Y., R.X., W.W.; Obtained funding: W.W., R.X.; Administrative, technical, or material support: S.Y., M.H., W.W.; Study supervision: W.W., M.H., R.X. Data availability All data used in this study are available from UKB via data access procedures (http://www.ukbiobank.ac.uk). Permission to use the UKB resource was obtained via a material transfer agreement as part of Application 105658. The data from the GDES and HOPE are available upon reasonable request to the corresponding author (wangwei@gzzoc.com). Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Shaopeng Yang, Zhuoyao Xin. Contributor Information Ruilin Xiong, Email: xiongruilin@gzzoc.com. Minwen Zhou, Email: zmw8008@163.com. Mingguang He, Email: mingguang.he@polyu.edu.hk. Wei Wang, Email: wangwei@gzzoc.com. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41746-026-02805-0. References Associated Data Supplementary Materials 20260429-1730 Supplementary materials Supplementary Data Data Availability Statement All data used in this study are available from UKB via data access procedures (http://www.ukbiobank.ac.uk). Permission to use the UKB resource was obtained via a material transfer agreement as part of Application 105658. The data from the GDES and HOPE are available upon reasonable request to the corresponding author (wangwei@gzzoc.com).