Study design and participants

This international multi-cohort prospective observational study was designed to identify plasma protein biomarkers of DRN and to develop, interpret, and externally validate a machine learning based prediction model suitable for clinical translation. Two large-scale prospective cohorts were used: the Guangzhou Diabetic Eye Study (GDES) as the discovery and training cohort, and the UKB as the external validation cohort [32–35]. GDES is a large prospective cohort of adults with type 2 diabetes integrating plasma proteomics and multimodal retinal imaging. Between 2017 and 2019, over 3,000 individuals with type 2 diabetes aged 35–85 years were recruited in Guangzhou, China. Its proteomics sub-cohort (GDES-PPP) served as the development dataset. UKB is a population-based multicenter cohort that enrolled over 500,000 participants aged 40–69 years across the United Kingdom between 2006 and 2010. Its proteomics subproject (UKB-PPP) was used for external validation. Both studies adhered to the Declaration of Helsinki. Ethical approvals were obtained from the Zhongshan Ophthalmic Center Ethics Committee (2017KYPJ094) and the North West Multi-Centre Research Ethics Committee (11/NW/0382). All participants provided written informed consent. The study was reported following STROBE guidelines (S1 Checklist) and TRIPOD statement (S2 Checklist).

Participants from GDES were included in two steps. First, Population 1 comprised 1,492 participants with type 2 diabetes but no diabetic retinopathy (DR) at baseline, all with baseline proteomics and qualified OCT data, and was used for cross-sectional analysis of plasma protein associations with retinal nerve fiber layer thickness. Second, to focus on neurodegeneration independent of vascular or exudative lesions, we identified within Population 1 a subgroup of 1,218 participants who remained free of DR during 6 years of follow-up (Population 2). All individuals in Population 2 had baseline proteomic data and OCT scans, with OCT repeated at years 2, 4, and 6 after enrollment. This population was used for longitudinal RNFL thinning analyses, pathway enrichment, and ML model development. For external validation, Population 3 comprised 502 UKB participants with type 2 diabetes but no baseline DR (S1 Table), all with available proteomics and a single time point OCT measure obtained at the UKB baseline ocular imaging assessment (2009–2010). Across both cohorts, we excluded participants lacking plasma proteomics or analyzable OCT-derived RNFL measures, applied prespecified OCT quality control, and excluded glaucoma, dementia, and other ocular conditions that could confound RNFL measurements (S1 Text).This cohort was used for cross-sectional analyses of associations between baseline protein levels and RNFL thickness to evaluate the robustness and generalizability of protein associations across diverse ethnicities and healthcare systems. The prespecified Phase I–IV objectives are summarized in the Results.

Retinal OCT imaging protocols

High-resolution OCT was used to quantify RNFL thickness, a key biomarker of DRN [36–39]. Standardized acquisition and analysis procedures were applied in both cohorts: (1) GDES cohort: Disc-centered three-dimensional OCT scans were acquired under dark-room conditions using the Topcon DRI OCT Triton swept-source system (Topcon Corporation, Japan), featuring a 1,050 nm light source, 100,000 A-scans/s, and 8 μm axial resolution. Peripapillary RNFL thickness was measured along a 360° circular path, 3.4 mm in diameter centered on the optic disc. (2) UKB cohort: Three-dimensional volume scans were obtained using the Topcon 3D OCT-1000 Mk II (Topcon Corporation, Japan) under similar dark-room conditions, with 18,000 A-scans/s and 6 μm axial resolution.

In GDES cohorts, scans were performed after adequate mydriasis by trained technicians, with internal eye-tracking to ensure accuracy. In UK Biobank, OCT imaging followed the standard UKB protocol and was performed without pharmacologic mydriasis. RNFL segmentation was performed automatically using the Topcon Advanced Boundary Segmentation (TABS) algorithm, defining RNFL thickness as the perpendicular distance from the internal limiting membrane to the inner border of the ganglion cell layer. The peripapillary region was divided into four quadrants (superior, inferior, temporal, and nasal), and both average and quadrant-specific RNFL thicknesses were calculated. All images were visually inspected by two experienced OCT technicians, with manual correction applied as needed. Only high-quality scans were included; images with poor quality, motion artifacts, or signal strength index <60 were excluded. For participants with valid data in both eyes, right-eye measurements were preferentially used (S1 Text).

DRN definition and phenotype quantification

DRN was defined as a pathological state marked by accelerated RNFL loss. Using repeated OCT measurements at baseline, year 2, year 4, and year 6, the annualized RNFL thinning rate was estimated using a linear mixed-effects model (median follow-up, 5.49 years). In the GDES cohort, participants were ranked by this rate: those in the fastest quartile (Q1) were classified as DRN, while the remaining participants (Q2–Q4) served as non-DRN controls. This objective, biologically grounded definition identifies individuals with pronounced neurodegenerative phenotypes, providing a precise foundation for biomarker discovery and predictive modeling. In sensitivity analyses, we additionally defined DRN as (i) the fastest 10% of RNFL thinners (DRN-Top10Slope), (ii) thin/abnormal pRNFL on the OCT report at the final available visit (DRN-ThinLast), and (iii) an annualized RNFL thinning rate below the lower normative reference limit derived from the 6-year longitudinal OCT data of healthy participants in the Chinese Ocular Imaging Project (COIP) cohort (DRN-ExcessLoss; S1 Text).

Plasma proteomics protocols

High-throughput plasma proteomics was performed using Olink Proximity Extension Assay (PEA) technology in both GDES and UKB cohorts, following standardized protocols for sample collection, preprocessing, measurement, and quality control. Laboratory personnel were blinded to clinical data. In GDES, a targeted approach employed the Olink Cardiometabolic Panel, measuring 361 proteins relevant to metabolic regulation, cardiovascular function, and inflammatory pathways, all mechanistically plausible for diabetes and its neuro-complications. This focused panel enabled deep, pathway-specific interrogation of DRN biology. The UKB cohort applied a broader proteomic panel, providing an independent platform for validation and assessment of biomarker robustness and generalizability. All proteins assayed in GDES were included in the UKB-PPP protein set; therefore, all cross-cohort replication and validation were conducted using this shared set, which comprised the full GDES panel. Rigorous quality control was implemented in both cohorts: within-plate CV <10% and between-plate CV <20%, with inter-plate normalization using control samples. Raw counts were normalized by extension controls and log₂-transformed to generate Normalized Protein eXpression (NPX) values, minimizing technical variation [26,27]. All NPX data were standardized and batch-corrected for downstream statistical analyses (S1 Text).

Covariate measurements

Covariates were collected using rigorous, standardized procedures within the prospective GDES cohort. Data encompassed demographics, lifestyle factors, clinical indices, and ophthalmic parameters, ensuring accuracy and comparability. Demographic (age, sex), socioeconomic (education, household income), and lifestyle data (smoking, alcohol, physical activity) were obtained through validated electronic questionnaires combined with face-to-face interviews. Diabetes-related information, including duration, medication use, and complications, was ascertained via medical record review and patient interview, with diagnoses based on international criteria and independently confirmed by two internists.

Anthropometric and biochemical measurements were performed by trained staff using standardized protocols. Height and weight were measured with calibrated instruments; blood pressure was recorded on the non-dominant arm after 5 min of rest, with three consecutive measures averaged. Fasting venous blood was collected for HbA1c (HPLC), lipid panel (total cholesterol, triglycerides, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol via enzymatic colorimetry), renal function markers (serum creatinine, estimated glomerular filtration rate), serum urate, and urinary albumin. All assays incorporated internal quality-control samples and participated in external quality assessment programs.

Ophthalmic evaluations followed standardized protocols. Slit-lamp biomicroscopy (Haag-Streit BQ-900, Switzerland) assessed the anterior segment. Best corrected visual acuity (BCVA) was measured using Early Treatment Diabetic Retinopathy Study (ETDRS) LogMAR charts (Precision Vision, USA) at 4 m. Refractive status was measured with an autorefractor (Topcon KR-8800, Japan); intraocular pressure (IOP) was measured with a non-contact tonometer (Topcon CT-1, Japan); axial length was measured by optical biometry (Lenstar LS900, Haag-Streit, Switzerland). All assessments were conducted by trained personnel following uniform procedures. Baseline covariate protocols for the UKB cohort are detailed in S1 Text.

Functional enrichment analysis

To elucidate the biological mechanisms underlying plasma proteins associated with DRN, we conducted comprehensive functional enrichment analyses. Using the set of plasma protein markers significantly linked to DRN, we applied the clusterProfiler R package (v4.0) to perform enrichment against Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG), and Reactome databases. Analyses encompassed biological process (BP), cellular component (CC), and molecular function (MF) terms, KEGG pathways, and Reactome reaction networks, systematically profiling the biological activities, subcellular localization, molecular functions, and pathway interactions of the target proteins. Enrichment significance was assessed using the hypergeometric (Fisher’s) test, with P values adjusted for multiple comparisons by the Benjamini–Hochberg method (FDR) < 0.05 was considered statistically significant.

Pro-DRN model construction

To develop a proteomics-driven risk prediction model, we selected plasma proteins significantly associated with longitudinal RNFL thinning in the discovery stage as features. The GDES cohort was randomly split into training and test sets at an 8:2 ratio, stratified by DRN status to preserve the proportion of positive cases. All data preprocessing, feature selection, and hyperparameter tuning were confined to the training set; the test set was used solely for final evaluation to prevent information leakage.

To ensure robustness and cross-validate the proteomic signal, we trained eight supervised learning algorithms in parallel: XGBoost, LightGBM, Random Forest, support vector machine (SVM), Neural Network, Logistic Regression, k-nearest neighbors (KNN), and Decision Tree. Stratified 10-fold cross-validation was applied within the training set, using fold-averaged C-index as the primary metric. Simpler models were favored when discrimination was comparable, and in-fold calibration was evaluated. Gradient-boosting models incorporated early stopping to mitigate overfitting, and hyperparameters were optimized via grid search or Bayesian methods based on cross-validation performance (S1 Text).

The final model was chosen based on consistent superiority across cross-validation and the independent test set. Model interpretability was assessed using Shapley values, quantifying each protein’s contribution to predictions. Two metrics were computed: (1) global feature importance based on mean absolute SHAP values, reflecting overall contribution, and (2) individual-level feature attributions, indicating the direction and magnitude of each protein’s effect on specific predictions.

Clinically meaningful outcomes for external validity

To assess clinical relevance beyond the OCT-based DRN phenotype, we examined associations between the baseline proteomic risk score and subsequent vision impairment and diabetic peripheral neuropathy. Vision impairment was defined as a loss of ≥15 ETDRS letters in best-corrected visual acuity during follow-up. Diabetic peripheral neuropathy was defined according to a previously published approach as a dichotomized Michigan Neuropathy Screening Instrument (MNSI) patient questionnaire score ≥4.

Statistical analysis

Statistical analysis protocol is provided in the Supplementary Materials. Analyses were conducted using R (v4.5.1) and Stata/MP (v18.0). Continuous variables are presented as mean ± SD or median (IQR), and categorical variables as counts (percentages). Between-group comparisons used t-tests or χ² tests as appropriate. All tests were two-sided, with P < 0.05 considered statistically significant. Multiple comparisons were controlled using the Benjamini–Hochberg method, with FDR < 0.05 considered significant.

In Population 1, cross-sectional associations between plasma proteins and RNFL thickness were assessed using multivariable linear regression, adjusting for age, sex, smoking, systolic blood pressure, HbA1c, and diabetes duration. Proteins with FDR < 0.05 were deemed significantly associated. In Population 2, the same covariates were used to test longitudinal associations with RNFL thinning rate.

For DRN prediction, participants in the top 25% of RNFL decline were classified as DRN cases, with the remainder as controls. Samples were randomly split 8:2 into training and testing sets. Using selected protein features, Pro-DRN were built with eight ML algorithms. The optimal model was selected based on C-index and generalization performance. Model evaluation included comparisons with nine individual traditional predictors and a baseline age–sex model using C-index and ΔC as discrimination metrics. Pro-DRN was further integrated into multiple existing clinical prediction models (Aspelund, Hippisley, Dagliati, ISDR, JDC, Tarasewicz, and a combined “All” model; variable composition was summarized in S13 Table) to assess incremental predictive value [40–48]. Risk reclassification was quantified using NRI and IDI, calibration assessed with calibration plots. Decision curve analysis (DCA) was used to evaluate clinical utility, and the net-benefit definition and underlying assumptions are described in S1 Text. The final model was implemented as an interactive online tool using the R Shiny [49].

For external validation, UKB-PPP participants with diabetes and available proteomics and OCT data were analyzed as an independent cohort. Cross-sectional analyses were fully replicated using the same covariates and modeling strategy to assess consistency in effect direction and significance of key protein associations. We additionally validated the locked Pro-DRN model using a cross-sectional thin-RNFL outcome, reporting discrimination (C-index) and calibration (Brier score and Hosmer–Lemeshow test).

Study design and framework

This study leveraged two large-scale prospective cohorts: the Guangzhou Diabetic Eye Study (GDES) and the UK Biobank (UKB) [32–35]. The GDES Proteomic Program (GDES-PPP) served as the discovery and model development dataset, and the UKB Proteomic Program (UKB-PPP) provided external validation. Participants with both high-quality Olink proteomic data and RNFL measurements by OCT were included.

The analytical framework comprised four modules (Fig 1). Module I involved cross-sectional protein screening using GDES-PPP baseline data to identify plasma proteins associated with DRN. Module II performed longitudinal analyses among participants who remained DRD-free over six years, evaluating associations between candidate proteins and RNFL thinning rates, and revealing underlying biological pathways via pathways enrichment analyses. Module III built and evaluated ML–based DRN prediction models, assessing performance through interpretability metrics, predictive accuracy, and comparison with conventional models [49]. Module IV validated DRN-associated protein signatures in UKB participants with type 2 diabetes to confirm generalizability across populations.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Study design and workflow.

(A) Cohorts and phenotype. Participants from the Guangzhou Diabetic Eye Study (GDES; discovery cohort with 6-year follow-up; n = 1,492 baseline) and the UK Biobank (UKB; external replication; n = 502) underwent optical coherence tomography (OCT) to quantify retinal nerve fiber layer (RNFL) thickness; accelerated RNFL thinning operationalizes DRN. (B) Plasma proteomics and functional annotation. EDTA plasma was profiled using a proximity-extension assay (Olink) to derive protein fingerprints and functional enrichment. (C) Model development. GDES was randomly split (8:2) into training and an untouched test set. Four families of learners were trained (generalized-linear, instance-based, tree-based, and neural-network models), with internal validation for hyperparameter tuning. (D) Model selection and interpretability. The optimal model was chosen based on validation performance and interpreted with SHAP (Shapley additive explanations) to quantify feature contributions. (E) Clinical frameworks. Candidate clinical variables were organized according to established risk models (Aspelund, Hippisley, Dagliati, ISDR, JDC, Tarasewicz) and a comprehensive “All model.” (F) Performance assessment. Discrimination (C-index/ROC), calibration, and clinical utility (decision-curve analysis) were evaluated in the hold-out test set. (G) Analysis phases. Phase I: cross-sectional protein–RNFL thickness associations; Phase II: longitudinal associations with RNFL thinning rate; Phase III: development and clinical evaluation of Pro-DRN in GDES; Phase IV: cross-cohort replication in UKB. Abbreviations: RNFL, retinal nerve fiber layer; OCT, optical coherence tomography; SHAP, Shapley additive explanations.

https://doi.org/10.1371/journal.pmed.1004868.g001

Baseline characteristics

The baseline characteristics of the three analytical populations are summarized in Tables 1 (Population 2) and S1 (Populations 1 and 3). In GDES-PPP, 1,492 participants with type 2 diabetes, with baseline plasma proteomics and qualified macular OCT scans (Population 1) were included in cross-sectional analyses (mean age 64.4 ± 7.5 years; 57.3% women). A total of 1,218 DRD-free over six years participants were included in longitudinal analyses (Population 2). Based on quartiles of RNFL thinning rate, the fastest-declining quartile (Q1, n = 305, 25.0%) defined the DRN group, and Q2–Q4 (n = 913) the non-DRN group. Participants in the DRN group were older, had a longer duration of diabetes, and higher systolic blood pressure at baseline (Table 1). Baseline pRNFL thickness was comparable between DRN progressors and non-DRN participants (110.07 ± 13.67 μm versus 109.81 ± 12.29 μm; P = 0.756). For external validation, 502 UKB-PPP participants with type 2 diabetes, no DR or other retinal disease at the time of OCT imaging, and available proteomic and OCT data (mean age 58.3 ± 7.6 years; 50.8% women) provided an independent cross-ethnic dataset (S1 Table).

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Baseline characteristics of Guangzhou Diabetic Eye Study (GDES) participants with Non-DR during follow-up.

https://doi.org/10.1371/journal.pmed.1004868.t001

Plasma proteins associated with DRN

In the GDES-PPP discovery cohort, cross-sectional analyses identified 72 plasma proteins significantly associated with average RNFL thickness (FDR < 0.05; Fig 2A and 2B). Among these, the 10 proteins most strongly negatively correlated with RNFL thickness were TFF3, MCFD2, CST3, FAM3C, CGREF1, NECTIN2, NPDC1, CLC, HSPG2, and DEFA1, with higher protein levels associated with thinner baseline RNFL (S2 Table). The adjusted effect sizes (β) ranging from −2.014 μm (95% CI [−2.800, −1.229]) to −0.983 μm (95% CI [−1.702, −0.265]) per 1–SD increase. In longitudinal analyses, 71 of these proteins were confirmed being significantly associated with RNFL thinning rate (FDR < 0.05; S3 Table). The top 10 proteins most strongly linked to accelerated thinning included CST3, HSPG2, NECTIN2, COL6A3, ACTA2, NPDC1, CD59, PTGDS, RNASET2, and FAM3C, with β values ranging from −0.445 μm/year (95% CI [−0.518, −0.372]) to −0.112 μm/year (95% CI [−0.173, −0.050]). Notably, UMOD consistently exhibited a protective effect, with higher levels associated with thicker baseline RNFL (β = 1.019 μm; 95% CI [0.273, 1.764]) and slower thinning over time (β = 0.256 μm/year; 95% CI [0.192, 0.320]). These associations between the RNFL thinning rate and proteins remained robust in sensitivity analyses (S4 Table).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. DRN-associated proteins and functional enrichment.

(A) Volcano plot summarizing DRN-associated proteins. The x-axis represents standardized β coefficients; the y-axis shows statistical significance as −log₁₀(P) values. Highlighted proteins (e.g., TFF3, CST3, MCFD2, FAM3C, NPDC1, UMOD) exceed predefined thresholds. (B) Ranked effect sizes (lollipop plot) for all proteins; point size reflects −log₁₀(P). Most associations are negative (protective), with UMOD showing a positive effect. (C) Over-representation analysis of FDR-significant proteins across Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG), and Reactome databases. Numbers above bars denote the count of proteins contributing to each term. Enriched pathways including inflammatory response, extracellular matrix remodeling, and neurotrophic signaling.

https://doi.org/10.1371/journal.pmed.1004868.g002

Functional annotation of DRN-associated proteins

Functional enrichment analyses using the GO, Kyoto Encyclopaedia of Genes and Genomes (KEGG), and Reactome revealed four interrelated biological processes (Fig 2C): (1) inflammation and immune recruitment (Cytokine–cytokine receptor interaction, Chemokine signaling, Leukocyte migration, Neutrophil degranulation); (2) basement membrane and ECM remodeling (Collagen-containing ECM, ECM organization, Glycosaminoglycan binding, Basement membrane); (3) membrane receptor function and blood–retina barrier integrity (Cell adhesion molecules, Cytokine receptor binding); and (4) microvascular homeostasis and metabolic stress (Complement and coagulation cascades, Hemostasis, AGE–RAGE signaling). These enrichment results implicate chronic inflammation, extracellular matrix dysregulation, receptor-mediated vascular dysfunction, and microvascular-metabolic stress as interlinked molecular correlates of early DRN.

Proteomics-based DRN (Pro-DRN) prediction model

We constructed Pro-DRN using plasma protein markers identified in GDES-PPP, employing eight widely used ML algorithms: XGBoost, LightGBM, Random Forest, Neural Network, Logistic Regression, K-Nearest Neighbors, Support Vector Machine (SVM), and Decision Tree. Models were trained using an 8:2 training/test split, with hyperparameters optimized via cross-validation, and performance evaluated in an independent test set (Fig 3; S5 Table).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Performance of machine-learning models for DRN prediction.

(A) Training set. Left, model discrimination summarized as AUC (C-index) with 95% confidence intervals; right, ROC curves for each model. (B) Held-out test set, displayed as in (A). The dashed diagonal denotes chance performance (AUC = 0.5). Across both sets, gradient-boosting models (XGBoost and LightGBM) show the highest discrimination. Abbreviations: DT, decision tree; KNN, k-nearest neighbors; SVM, support vector machine; MLP, multilayer perceptron; RF, random forest.

https://doi.org/10.1371/journal.pmed.1004868.g003

In the training set, model C-index ranged from 0.760 to 0.984, with Random Forest 0.984 (95% CI [0.977, 0.990]), XGBoost 0.979 (95% CI [0.972, 0.986]), and SVM 0.947 (95% CI [0.933, 0.961]) performing best. In the test set, C-indices ranged from 0.736 to 0.860, with XGBoost 0.860 (95% CI [0.810, 0.911]), LightGBM 0.832 (95% CI [0.777–0.886]), and Random Forest 0.816 (95% CI [0.756–0.876]) leading. XGBoost was selected as the final model on the basis of its superior balance between discrimination, generalisability, and training stability, and was used for all subsequent Pro-DRN scoring, feature-importance interpretation, and clinical utility analyses.

Importance ranking of Pro-DRN plasma proteins

To interpret the model, we applied Shapley values, providing both global and individual-level insights while accounting for protein collinearity and complex interactions (Fig 4). Globally, the top 10 proteins contributing to DRN risk were ACTA2, COL6A3, CCL27, NT-proBNP, CD46, MCFD2, CTSZ, CCL15, RARRES2, and HSPG2 (Fig 4A). Beeswarm plots confirmed consistent effect directions across individual samples, with higher expression of most top proteins increasing predicted risk.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Global and individual explanations of the Pro-DRN model.

(A) XGBoost interpretation. Left: global feature importance ranked by mean |SHAP| for the top 30 proteins (longer bars = greater overall contribution). Right: SHAP beeswarm plot showing the signed effect of each feature on the predicted probability of DRN (x-axis, SHAP value); red/blue indicate higher/lower feature values. High-impact proteins include ACTA2, COL6A3, CCL27, NT-proBNP, CD46, MCFD2, CTSZ, CCL15, RARRES2 and HSPG2. (B) Individual explanation (force plot) for a representative participant with predicted risk f(x)=0.85. Red features push the prediction above the model base value (increase risk), and blue features pull it down; arrow width reflects absolute SHAP contribution.

https://doi.org/10.1371/journal.pmed.1004868.g004

In high-risk individuals (predicted probability = 0.85), COL6A3, ACTA2, CCL16, REG3A, HSPG2, PAM, ART3, and TIMP1 jointly drove risk, demonstrating a multi-protein additive effect (Fig 4B). The concordance between global and individual explanations indicates that the discriminative power of Pro-DRN arises from the integrated contributions of multiple proteins rather than from any single dominant feature, a property that simultaneously enhances interpretability and predictive robustness.

Pro-DRN substantially improves DRN risk discrimination

We systematically evaluated the predictive performance of Pro-DRN in the test set. As a standalone predictor, Pro-DRN achieved excellent discrimination (C-index = 0.860; 95% CI [0.810, 0.911]), far exceeding any single conventional predictor (C-index: 0.500 to 0.688; S6 Table), highlighting the independent and superior value of plasma proteomic features. Integrating Pro-DRN with a baseline model of age and sex increased the C-index from 0.697 to 0.877 (95% CI [0.834, 0.921]), representing a 25.9% relative improvement (Table 2).

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Incremental predictive value of the Pro-DRN across established models.

https://doi.org/10.1371/journal.pmed.1004868.t002

To assess incremental values, Pro-DRN was added to six established clinical models (Fig 5A and 5B). All models showed consistent gains in discrimination, with average ΔC-index ranging from 0.137 to 0.159, and the fully adjusted model (all clinical variables plus Pro-DRN) achieved a C-index of 0.908 (95% CI [0.869, 0.948]; Table 2). Importantly, the Pro-DRN model achieved discrimination that was comparable to, and modestly higher than, the model built using all measured proteins (S7 Table). Risk reclassification analyses further confirmed its value: average IDI increased by 23.3% (range 21.2% to 24.5%) and NRI by 38.7% (range 22.6% to 45.2%), with the All model showing IDI = 18.3% (95% CI [0.120, 0.248]; P < 0.001) and NRI = 35.5% (95% CI [0.119, 0.593]; P = 0.008; Table 3).

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Reclassification and discrimination gains with Pro-DRN across established models.

https://doi.org/10.1371/journal.pmed.1004868.t003

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Pro-DRN enhances conventional DRN prediction models.

(A) Discrimination for eight clinical frameworks with and without the proteomics score (Pro-DRN). Gray points and horizontal bars denote the original models with 95% CIs; colored points indicate models after adding Pro-DRN. Across frameworks, Pro-DRN consistently improves the C-index, with the “All model + Pro-DRN” performing best. (B) ROC curves. Solid lines show models with Pro-DRN and dashed lines the corresponding clinical models. Pro-DRN shifts the curves upward and left. (C) Calibration (observed event probability vs. predicted risk). Models augmented with Pro-DRN (solid) align more closely with the 45° reference line than their baselines (dashed). (D) Decision curve analysis. Net benefit across threshold probabilities is higher after adding Pro-DRN over a broad range of thresholds.

https://doi.org/10.1371/journal.pmed.1004868.g005

To evaluate robustness across algorithms, Pro-DRN was reconstructed using seven alternative ML methods. All versions consistently improved discrimination and reclassification. For instance, LightGBM achieved ΔC-index values ranging from 0.113 to 0.136 (best C-index = 0.891; 95% CI [0.850, 0.932]), with IDI = 17.9% and NRI = 38.7%. Random Forest yielded ΔC-index values ranging from 0.115 to 0.137 (best C-index = 0.897; 95% CI [0.856–0.937]), with IDI = 21.6% and NRI = 29.6% (Tables 2 and 3). These consistent improvements across diverse algorithms underscore the reproducibility, generalizability, and robustness of proteomic markers for enhancing DRN risk stratification.

Pro-DRN demonstrates excellent calibration and clinical utility

Beyond discrimination and reclassification, we systematically evaluated Pro-DRN’s calibration and clinical utility (Fig 5C and 5D). Calibration plots showed that all conventional models incorporating Pro-DRN achieved markedly improved agreement between predicted and observed DRN risk, with curves closely approaching the ideal diagonal and substantially reducing systematic bias, particularly in mid-to-high risk ranges (Fig 5C). Decision curve analysis (DCA) further demonstrated that models with Pro-DRN consistently delivered higher net clinical benefit across relevant threshold probabilities, outperforming both the original clinical models and extreme strategies of “treat all” or “treat none.” Notably, the fully adjusted model combined with Pro-DRN achieved the greatest net benefit across most thresholds (Fig 5D). These results indicate that Pro-DRN substantially enhances calibration and clinical utility, offering a robust, interpretable, and actionable tool for early identification and targeted intervention in individuals at high risk of DRN.

Robustness Analyses of Pro-DRN Integrated Models

To further stress-test the benchmarks, we additionally augmented each conventional model with measures of insulin resistance (fasting insulin, HOMA2-IR, TyG), beta cell function (fasting C-peptide, HOMA2-B), and glycemic indices (fasting glucose, HbA1c); Pro-DRN remained consistently additive across all augmented baselines (S8 Table).

We tested robustness using three alternative DRN definitions (DRN-Top10Slope, DRN-ThinLast and DRN-ExcessLoss). Pro-DRN preserved discrimination across all definitions (C-index 0.880, 0.753 and 0.848, respectively; S9 Table) and consistently improved the fully adjusted clinical framework (All model: 0.856 to 0.935; 0.738 to 0.832; and 0.823 to 0.902; S10 Table).

To establish the clinical relevance of Pro-DRN beyond its OCT-based definition, we examined whether the baseline Pro-DRN score was associated with subsequent patient-centered outcomes. Higher Pro-DRN was associated with subsequent clinically meaningful vision impairment defined by 15 or more ETDRS letters loss (OR = 27.98; 95% CI [1.07, 578.13]; P = 0.036), and was also associated with diabetic peripheral neuropathy (OR = 5.82; 95% CI [1.41, 22.97]; P = 0.013) as defined by the Michigan Neuropathy Screening Instrument patient questionnaire (Table 4). These findings support the broader clinical relevance of Pro-DRN, although they should be interpreted cautiously given the limited number of events and the wide confidence interval.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Association of the Pro-DRN score with vision impairment and diabetic peripheral neuropathy.

https://doi.org/10.1371/journal.pmed.1004868.t004

Shiny-based online tool for clinical translation

To accelerate clinical translation, we deployed the locked Pro-DRN model as an interactive, web-based decision-support tool (https://fmb-gdes2025.shinyapps.io/DRN_predict/). The tool comprises two core modules: (1) Pro-DRN score calculation: users can upload standardized measurements of DRN-associated plasma proteins, and the system computes the Pro-DRN score in real time using the XGBoost model developed in this study; (2) DRN risk prediction: based on the calculated score, the tool generates individualized risk estimates with dynamic visual feedback. The interface supports real-time parameter linkage, input validation, and continuous risk updating as data change. An intuitive risk-stratification display further aids clinical interpretation. Designed with a “left input–right output” layout, the platform enables rapid, reliable risk assessment and documentation, making it suitable for clinical consultations, screening programs, and longitudinal follow-up.

Cross-ethnic external validation in UKB-PPP

To assess generalizability and robustness, we performed cross-ethnic external validation in the independent cohort. Despite limited longitudinal proteome–imaging data and platform differences, we systematically examined cross-sectional associations between plasma proteins and macular RNFL thickness using standardized covariate adjustment and FDR correction (Table 5). In univariate analyses, 52 of 71 proteins (73.2%) were significantly negatively associated with RNFL thickness, with β values ranging from −0.841 μm (95% CI [−1.206, −0.476]) to −0.337 μm (95% CI [−0.618, −0.056]) per 1–SD increase. UMOD was the sole protein with a positive association (β = 0.488; 95% CI [0.171, 0.805]). In multivariable models adjusting for confounders and multiple testing, 34 proteins retained significant negative associations, fully concordant with GDES-PPP findings, including key markers such as ACTA2, COL6A3, CD46, RNASET2, CCL15, NT-proBNP, NECTIN2, and HSPG2, with β values ranging from −0.716 μm (95% CI [−1.077, −0.355]) to −0.399 μm (95% CI [−0.717, −0.082]). UMOD maintained a protective trend (β = 0.322), though not reaching significance (P = 0.088). Core GDES protein signatures are reproducible across populations, with consistent effects and biological relevance, highlighting their robust translational potential for DRN.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Associations of proteins with retinal nerve fiber layer thickness in UK Biobank.

https://doi.org/10.1371/journal.pmed.1004868.t005

We further performed model-level external validation in UKB-PPP using a cross-sectional thin-RNFL phenotype (thinnest quartile of baseline pRNFL thickness). Adding Pro-DRN consistently improved discrimination across all conventional frameworks (e.g., Age and Sex: 0.635 to 0.714; All model: 0.676 to 0.756; S11 Table) and improved calibration as assessed by lower Brier scores and no evidence of lack of fit by the Hosmer–Lemeshow test (S12 Table).

Ethics approval

The study was approved by the Northwest Multicenter Research Ethics Committee (11/NW/0382) and the Ethics Committee of Zhongshan Ophthalmic Center (2017KYPJ094).