A proteomics- and ML-based precision prediction system for diabetic retinal neurodegeneration (Pro-DRN) substantially enhanced early risk stratification beyond conventional clinical factors and may support targeted screening and timely neuroprotective interventions.
Key Findings
Results
Seventy-one plasma proteins were associated with the development and progression of diabetic retinal neurodegeneration (DRN) in multivariable analyses.
Discovery cohort comprised 1,492 participants with baseline plasma proteomics and OCT from the Guangzhou Diabetic Eye Study (GDES).
1,218 participants were followed with repeated OCT over 6 years.
DRN was quantified by the annualized OCT-derived retinal nerve fiber layer thinning rate.
Analyses were adjusted for age, sex, smoking, systolic blood pressure, HbA1c, and diabetes duration.
The proteomics-based DRN prediction model (Pro-DRN) achieved a C-index of 0.860 in the independent test set, rising to 0.908 when integrated with clinical variables.
Pro-DRN was developed using eight machine learning algorithms including XGBoost and LightGBM.
C-index of 0.860 was achieved on the independent test set using proteomics alone.
Integration with clinical variables improved the C-index to 0.908.
The proteins most consistently driving model performance included ACTA2, COL6A3, and HSPG2.
Results
Pro-DRN significantly improved discrimination and reclassification compared with six conventional prediction models.
Delta C-index improvement ranged from 0.137 to 0.159 versus conventional models (all P < 0.001).
Integrated Discrimination Improvement (IDI) ranged from 0.212 to 0.245 (all P < 0.05).
Net Reclassification Improvement (NRI) ranged from 0.226 to 0.452 (all P < 0.05).
Pro-DRN also demonstrated higher net benefit compared to conventional models.
Results
Adding Pro-DRN to the Hippisley model substantially improved its predictive performance.
The C-index of the Hippisley model increased from 0.739 (95% CI [0.670, 0.808]) to 0.898 (95% CI [0.858, 0.937]).
IDI was 0.245 (95% CI [0.177, 0.318]) (P < 0.001).
NRI was 0.452 (95% CI [0.222, 0.673]) (P < 0.001).
Net benefit was also higher with the integrated model.
Results
Cross-ethnic external validation in the UK Biobank reproduced core protein signals with consistent effect directions.
External validation cohort comprised 502 participants from the UK Biobank, recruited 2006–2010.
Core protein signals identified in the GDES discovery cohort were reproduced in this independent, ethnically distinct population.
Effect directions were consistent across both populations, confirming robustness.
This validation supports the generalizability of the Pro-DRN model across different ethnic groups.
Results
The locked Pro-DRN model was deployed as an interactive, web-based risk-assessment tool for clinical translation.
The tool was designed to support early DRN screening and longitudinal monitoring.
Deployment as a web-based tool was intended to facilitate clinical usability.
The principal methodological limitation identified was single time point proteomic assessment.
What This Means
This research suggests that specific proteins circulating in the blood can predict early nerve damage in the retinas of people with type 2 diabetes, a condition known as diabetic retinal neurodegeneration (DRN). The study followed over 1,200 people with diabetes for six years, measuring both their blood proteins and the thickness of retinal nerve layers using eye scans (OCT). Researchers identified 71 proteins linked to retinal nerve thinning, related to inflammation, tissue remodeling, and blood vessel health. Using these proteins alongside machine learning, they built a prediction model (Pro-DRN) that was substantially more accurate than existing clinical models at identifying who is at risk.
The Pro-DRN model performed impressively, with accuracy scores (C-index) of 0.860 using proteins alone and 0.908 when combined with standard clinical information — compared to 0.739 for one of the best existing models. Three proteins in particular — ACTA2, COL6A3, and HSPG2 — were the most important drivers of these predictions. The findings were validated in an independent group of 502 participants from the UK Biobank, showing the model works across different ethnic populations.
This research suggests that blood protein testing combined with machine learning could enable earlier and more precise identification of people with diabetes who are losing retinal nerve tissue, before vision problems become apparent. The researchers have made their model available as a web-based tool, which could help clinicians screen patients and monitor them over time, potentially allowing earlier use of neuroprotective treatments to slow or prevent diabetic eye disease.
Li H, Zhu Z, Yang S, Cheng W, Tan S, Xin Z, et al.. (2026). Proteomic signatures of early retinal neurodegeneration in type 2 diabetes mellitus.. PLoS medicine. https://doi.org/10.1371/journal.pmed.1004868