A new study identifies 71 plasma proteins tied to diabetic retinal neurodegeneration, leading to a machine learning model that significantly outperforms current risk prediction tools.
A Proteomic Leap Forward in Predicting Diabetic Eye Disease
A study published in PLOS Medicine has identified 71 plasma proteins associated with diabetic retinal neurodegeneration (DRN) and developed a powerful new risk prediction model called Pro-DRN. The research, leveraging data from two major prospective cohorts, represents a significant step toward personalized risk assessment for this common diabetes complication.
Study Design and Key Findings
The study utilized two large-scale cohorts: the Guangzhou Diabetic Eye Study (GDES) , with 1,492 participants for discovery and 1,218 for longitudinal analysis over six years, and the UK Biobank, which provided 502 participants for external validation.
DRN was quantified using the annualized rate of retinal nerve fiber layer thinning, measured via optical coherence tomography (OCT). After adjusting for traditional risk factors (age, sex, smoking, systolic blood pressure, HbA1c, and diabetes duration), the research team identified 71 plasma proteins linked to both the development and progression of DRN.
The biological pathways involved include inflammatory immune recruitment, extracellular matrix remodeling, and microvascular homeostasis.
The Pro-DRN Model: Unprecedented Accuracy
The Pro-DRN model, built using machine learning algorithms like XGBoost and LightGBM, demonstrated remarkable predictive power. On an independent test set, it achieved a C-index of 0.860. This performance improved dramatically to 0.908 when combined with standard clinical variables.
Compared to six conventional models, Pro-DRN showed substantial improvement in discrimination (ΔC-index 0.137–0.159) and reclassification (IDI 0.212–0.245; NRI 0.226–0.452).
Key Predictive Proteins and Clinical Application
The study highlighted three key predictive proteins: ACTA2, COL6A3, and HSPG2. The final, locked model has been deployed as an interactive, web-based risk assessment tool for potential clinical use.
Cross-ethnic validation in the UK Biobank successfully reproduced core protein signals and consistent effect directions, supporting the model's broader applicability.
Study Limitations
The authors acknowledge several limitations:
- Proteomic assessment was conducted at a single time point.
- Plasma protein levels may not fully reflect local retinal expression.
- The RNFL measurement region differed between cohorts (peripapillary vs. macular).