Clinical Proteomics: Predicting Diabetic Retinal Nerve Loss
- Jun 23
- 4 min read
Most people with type 2 diabetes are checked for eye disease only after the retina already shows visible harm, which leaves a narrow window for prevention. A clinical proteomics study in PLoS Medicine took a different route. The team measured thousands of circulating blood proteins and asked which ones forecast retinal nerve fiber loss years before it surfaces on imaging. Working across a Guangzhou cohort and UK Biobank, they tied 71 plasma proteins to diabetic retinal neurodegeneration and turned that signal into a usable risk score.
Key Takeaways
Clinical proteomics can flag people with type 2 diabetes who are likely to lose retinal nerve tissue before standard eye imaging detects the change.
A panel of 71 plasma proteins combined with machine learning predicted diabetic retinal neurodegeneration with a C-index of 0.860, rising to 0.908 alongside clinical variables.
Three proteins, ACTA2, COL6A3, and HSPG2, drove most of the prediction, pointing toward extracellular matrix and microvascular biology.
The protein signals held up from a Chinese discovery cohort to UK Biobank, which matters for any marker meant to travel between populations.
What Clinical Proteomics Revealed About Retinal Nerve Loss
The discovery work ran inside the Guangzhou Diabetic Eye Study. There, 1,492 participants gave baseline plasma and macular optical coherence tomography scans, and 1,218 stayed in follow-up over six years. Researchers profiled plasma with a proximity-extension assay, an affinity-based form of protein profiling that reads thousands of targets per sample, then defined neurodegeneration by the annualized rate of retinal nerve fiber layer thinning. After adjusting for age, sex, smoking, blood pressure, HbA1c, and diabetes duration, 71 proteins tracked with both onset and progression. They clustered into immune recruitment, extracellular matrix remodeling, and microvascular upkeep, which lends the association a believable mechanism rather than a bare statistical hit.
Key Findings
71 proteins tracked nerve fiber thinning: Each was linked to the OCT-derived thinning rate after adjustment for six standard clinical confounders.
The model outperformed clinical scores: Pro-DRN reached a 0.860 C-index in the held-out test set and 0.908 once clinical variables were added, beating six conventional risk models by 0.137 to 0.159.
ACTA2, COL6A3, and HSPG2 led the panel: SHAP analysis ranked these as the most consistent contributors across the eight machine learning algorithms tested.
Replication across ethnicity: Core protein signals and effect directions reproduced in 502 UK Biobank participants with type 2 diabetes.
Figure 1. Study design and workflow. The schematic traces the four analysis modules: plasma proteomic profiling and optical coherence tomography in the Guangzhou discovery cohort, longitudinal association testing against retinal nerve fiber layer thinning, machine learning model development with SHAP-based interpretation, and cross-cohort replication in UK Biobank. Adapted from Li et al. (2026), PLoS Medicine.
Building a Risk Model From Blood Proteins
The Guangzhou data were split 8:2 into training and an untouched test set, and four families of learners were trained, from generalized-linear to neural-network designs. The locked winner, Pro-DRN, was read with SHAP values so each protein's contribution stayed inspectable. The authors then wrapped the frozen model in a web-based calculator for real-time risk assessment. That kind of deployment is rare in discovery papers, and it signals a group that cares about clinical use, not just a publishable AUC.
Why These Proteins Make Biological Sense
ACTA2, COL6A3, and HSPG2 sit at the junction of vascular smooth muscle tone, matrix scaffolding, and basement membrane integrity, all of which shape how small retinal vessels and neurons hold up under chronic hyperglycemia. The plasma proteome read here echoes the large Olink-based reference work from Sun and colleagues (2023), where circulating protein levels mapped cleanly onto organ-specific disease risk. One caveat deserves weight: proteins were sampled once, so the study shows association, not the dynamic trajectory a repeated-measures design would capture.
What This Means for Biomarker Programs
For biomarker teams, the useful lesson is that a locked, interpretable protein panel can outrank clinical risk factors without adding a new imaging step. In client work at Dalton, we see the same behavior when an affinity panel is anchored to a hard longitudinal endpoint rather than a cross-sectional label, which is the discipline behind dependable plasma proteomics biomarker services. Results like these are why we treat panel lock-down and external replication as non-negotiable steps in any multi-omics biomarker discovery effort running at cohort scale.
Frequently Asked Questions
What is clinical proteomics? Clinical proteomics is the large-scale measurement of proteins in patient samples such as plasma or serum to find markers of disease, risk, or treatment response. It pairs high-throughput protein profiling with statistics or machine learning to turn raw protein readouts into usable clinical signals.
How does clinical proteomics predict diabetic retinal neurodegeneration? The study linked 71 circulating plasma proteins to the rate of retinal nerve fiber thinning, then trained a model called Pro-DRN on those signals. It predicted neurodegeneration with a 0.860 C-index, climbing to 0.908 when clinical variables were included.
What samples does a clinical proteomics biomarker study need? Most need only a standard blood draw, with plasma or serum profiled on an affinity platform like Olink or by mass spectrometry. Paired clinical endpoints and an independent validation cohort matter more for credibility than raw sample count.
Conclusion
Circulating proteins clearly carry enough information to rank diabetic patients by their risk of early retinal nerve loss, and that ranking survived the move from Guangzhou to UK Biobank. Whether acting on a high score improves vision outcomes is still open, since the proteins were measured once in an observational design. The sensible next step is a prospective test of the locked panel before it guides any neuroprotective decision.
Related Reading
See how we run these analyses in one lab: Dalton's multi-omics CRO services.
Citation
Li, H., Zhu, Z., Yang, S., Cheng, W., Tan, S., Xin, Z., Zhang, L., Zhu, Z., Chen, S., Huang, W., & Wang, W. (2026). Proteomic signatures of early retinal neurodegeneration in type 2 diabetes mellitus. PLoS Medicine, 23(6), e1004868. https://doi.org/10.1371/journal.pmed.1004868
Note
This blog post summarizes findings from the above-cited research. Figures are adapted from the original publication. For full details, please refer to the source article.
By Seungjun Yeo, CEO at Dalton Bioanalytics. Specializing in multi-omics mass spectrometry for drug discovery and biomarker research.