Biomarker Discovery: GDF15 Predicts Diabetic Kidney Risk

HbA1c misleads more often than most clinicians like to admit. Iron deficiency, hemoglobinopathies, and high triglycerides all distort it, so the patient whose kidneys are quietly failing can slip past a single lab value. A recent biomarker discovery study in Frontiers in Endocrinology went hunting for a blood signal that warns earlier, mining plasma proteomics and metabolomics from 50,021 UK Biobank participants to find markers that flag type 2 diabetes complications years before symptoms surface. One protein, GDF15, stood out.

What Multi-Omics Adds to Biomarker Discovery in Diabetes

The team paired affinity-based plasma proteomics (1,463 Olink protein readouts reported as NPX) with an NMR metabolite panel of more than 280 analytes, then followed participants for a median of roughly 12 years. After splitting the cohort evenly into training and test sets of about 25,000 each, they screened candidates with LASSO-Cox regression and gradient-boosted trees, ranked feature importance with SHAP, and rebuilt the strongest signals across nine modeling algorithms. The metabolite layer, it should be said, contributed less than the proteins; almost every top marker that survived was a protein. That is an honest limitation of a study framed as multi-omics, and it shapes how much weight the metabolomics deserves here.

Key Findings

• GDF15 outperformed standard glucose markers: For diabetic kidney disease developing within five years, plasma GDF15 reached an AUC of 0.94, against 0.85 for HbA1c and just 0.68 for fasting glucose, with an adjusted hazard ratio of 1.76 (95% CI 1.62 to 1.91).

• One protein, many complications: GDF15 ranked as a key predictor for nearly every diabetic complication studied, from cardiovascular to neurological, with the lone exception of metabolic-disorder outcomes.

• Pairing GDF15 with routine labs sharpened the call: Adding a standard clinical panel lifted five-year kidney-risk prediction to an AUC near 0.97, beating either the protein or the clinical indicators on their own.

• Other renal proteins tracked alongside it: COL6A3, cystatin C, and HAVCR1 also flagged kidney risk strongly (GDF15 survival HR 6.12), while NELL1 behaved as a protective signal, hinting at a wider tubular-injury program.

Figure 1. Study design and analytical workflow. Panel A traces participant selection in the UK Biobank, narrowing 502,185 enrolled individuals to the 50,021 analyzed after removing those without blood samples, with high trait missingness, with invalid ICD-10 outcome codes, or with type 2 diabetes already present at baseline. Panel B lays out the pipeline: a 50/50 train and test split, coarse filtering of risk factors by statistical and machine learning methods, and performance checks using prediction models, risk trajectories, and an integrated model. Adapted from Hao et al. (2026), Frontiers in Endocrinology.

How the Model Singled Out GDF15

Feature-importance scoring inside the gradient-boosted model kept surfacing the same protein. As markers were added one at a time and the AUC tracked, GDF15 drove the steepest early gains for kidney outcomes, and its SHAP swarm showed the widest spread, a sign of strong, direction-consistent predictive weight. LASSO-Cox regression, run as an independent statistical check, landed on the same protein, which matters because agreement between a tree ensemble and a penalized regression is harder to dismiss as overfitting. In survival analysis the kidney-disease hazard ratio for high versus low GDF15 reached 6.12.

Why a Kidney-Damage Protein Reads as an Early Warning

GDF15 is best understood as a repair signal, not a culprit. Mazagova and colleagues showed that it climbs in kidney tissue before overt nephropathy appears, and that knocking it out leaves diabetic mice with worse tubular damage, evidence that the protein is protective even as its blood level rises with injury. Mechanistically it dampens the AGE/RAGE axis and downstream NF-kB signaling. So the rising plasma concentration is the tissue calling for help, which is exactly what makes it readable as an early, phenotype-anchored marker of kidney stress.

Running Proteins and Metabolites From One Sample

A study like this leans on two platforms, an affinity panel for proteins and NMR for metabolites, and the metabolite arm quietly underdelivered. When proteins and small molecules are measured from the same aliquot in one mass spectrometry workflow, which the Omni-MS approach at Dalton was built to do, you sidestep the cross-platform harmonization and sample-splitting that blunt the metabolite signal in designs like this one. Reading both layers together is also where a marker such as GDF15 can be anchored to the metabolic shifts that surround it, rather than viewed in isolation.

Frequently Asked Questions

What is biomarker discovery in multi-omics research?

Biomarker discovery is the process of finding measurable molecules, such as proteins or metabolites, that signal a disease state or predict its course. In multi-omics work it means screening thousands of candidates across several molecular layers at once, then narrowing to the few that hold up statistically. The goal is a marker specific and sensitive enough to act on clinically.

How did this biomarker discovery study identify GDF15 for diabetic kidney disease?

Researchers analyzed plasma proteomics and metabolomics from 50,021 UK Biobank participants, then used machine learning and Cox regression to rank predictors. GDF15 repeatedly came out on top for kidney complications, reaching an AUC of 0.94 for disease within five years. Two independent methods agreeing on the same protein strengthened the result.

Do biomarker panels like this need mass spectrometry or affinity assays?

Both routes are used. This study read proteins with an Olink affinity assay and metabolites by NMR, while many labs quantify both proteins and metabolites by mass spectrometry from a single sample. The choice affects cost, throughput, and how easily the two omics layers can be compared.

Conclusion

What looks believable here is that a single plasma protein can flag diabetic kidney risk years ahead, and do it more accurately than HbA1c in the near term. What is not yet proven is whether that holds outside the UK Biobank, since the authors ran no external validation and set no clinical cutoff. For a drug-discovery or diagnostics team, GDF15 is worth carrying forward as a stratification marker, with the caveat that a prospective, multi-ethnic check comes first.

Citation

Hao, M., Li, H., Xin, M., Li, J., Sun, R., Liu, Q., Zhang, Y., Shan, X., He, Y., Xu, B., Guo, Q., Kuang, H., & Wang, P. (2026). Plasma protein GDF15 has a good predictive potential for the kidney complications of type 2 diabetes. Frontiers in Endocrinology, 17, 1758267. https://doi.org/10.3389/fendo.2026.1758267

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.