Plasma proteomics combined with machine learning identifies nine core proteins (ANG, IL1R1, CXCL14, ESAM, PTGDS, HAVCR1, FGFR2, IGSF8, CCL3) as robust non-fatal MACE predictors in diabetic kidney disease, with a final model achieving AUC of 0.768 for non-fatal MACE, 0.808 for MI, and 0.816 for stroke.
Key Findings
Results
Of 1,463 plasma proteins screened, 561 were associated with non-fatal MACE across Cox regression models, with 14 overlapping proteins identified across all three models.
Study population consisted of 317 DKD patients from the UK Biobank
Three sequential Cox regression models were used: crude, socio-demographic-adjusted, and socio-demographic-metabolic adjusted
14 proteins overlapped across all three Cox models before machine learning feature selection
The cohort was split 70% training and 30% testing to prevent information leakage
Results
Nine core proteins were validated as predictors of non-fatal MACE: ANG, IL1R1, CXCL14, ESAM, PTGDS, HAVCR1, FGFR2, IGSF8, and CCL3.
These nine proteins were identified through a four-step machine-learning pipeline: LASSO-Cox, random survival forest, Boruta, and XGBoost-Cox
All high-expression groups for all nine proteins had elevated non-fatal MACE risk
ANG showed the strongest association with non-fatal MACE (HR = 3.88, 95% CI 2.33–6.48, p < 0.001)
Validation was performed using Kaplan-Meier survival analyses, longitudinal trajectory modeling, and ROC benchmarking
Results
ANG (Angiogenin) demonstrated the strongest individual protein association with non-fatal MACE among all validated proteins.
Hazard ratio for ANG = 3.88 (95% CI 2.33–6.48, p < 0.001)
High ANG expression group had significantly elevated non-fatal MACE risk
ANG was identified as a top predictor through the machine-learning pipeline
Results
The integrated model combining proteins, demographic factors, and clinical variables achieved the highest predictive performance for non-fatal MACE, MI, and stroke outcomes.
AUC for non-fatal MACE = 0.768
AUC for myocardial infarction (MI) = 0.808
AUC for stroke = 0.816
The model demonstrated superior stability in cross-validation compared to alternative approaches
CoxBoost + Elastic Net framework was selected as the optimal framework after benchmarking 101 algorithms
Results
The final predictive model demonstrated favorable calibration in high-risk patients and positive net clinical benefit across a broad range of decision thresholds.
Favorable calibration was observed specifically in high-risk patients
Positive net clinical benefit was demonstrated across decision thresholds of 5% to 45%
Model performance was validated in the held-out testing set (30% of cohort)
Results
GO/KEGG enrichment analysis highlighted inflammatory-immune pathways as key mechanistic drivers of non-fatal MACE in DKD.
Key pathways identified included positive regulation of MAPK cascade, cytokine-cytokine receptor interaction, and PI3K-Akt signaling pathway
Enrichment analysis was conducted on the nine core validated proteins
Results point to inflammatory and immune mechanisms underlying cardiovascular risk in DKD
Results
An interactive web application was developed and deployed to enable clinical implementation of the prediction model.
The web tool is accessible at https://jiangli2941.github.io/MACE-prediction-v2/
The tool accepts input of 28 variables
Outputs include non-fatal MACE risk status, risk probability, and highlights abnormal indicators
What This Means
This research suggests that measuring specific proteins in the blood of patients with diabetic kidney disease (DKD) — a condition where diabetes damages the kidneys — can help predict who is at risk of having a serious but non-fatal heart or stroke event (called a major adverse cardiovascular event, or MACE). The researchers analyzed blood protein data from 317 DKD patients in the UK Biobank, screened over 1,400 proteins, and used a combination of statistical methods and machine learning to narrow down the most predictive proteins. They identified nine key proteins, with one called ANG (angiogenin) being particularly strongly linked to cardiovascular risk — people with high ANG levels had nearly four times the risk of a MACE event. The biological pathways these proteins are involved in point to inflammation and immune system activity as important drivers of heart disease risk in DKD.
The final prediction model, which combined these nine proteins with demographic and clinical information, performed well at distinguishing high-risk from lower-risk patients, with accuracy scores (AUC) of 0.768 for overall MACE, 0.808 for heart attacks, and 0.816 for stroke. The model was selected from benchmarking 101 different algorithm combinations and showed reliable performance across different validation tests. Importantly, it also demonstrated clinical usefulness across a wide range of risk thresholds, meaning it could help doctors make treatment decisions for patients at varying levels of risk.
To make this tool accessible in real clinical settings, the researchers built a freely available interactive website where clinicians can enter 28 patient variables and receive a personalized risk estimate. This research suggests that integrating blood protein measurements with routine clinical data could meaningfully improve how doctors identify DKD patients who are most at risk of cardiovascular complications, potentially enabling earlier or more targeted preventive care.
Jiang L, Chien C, Li T, Zhang H, Zhao T, Wu X. (2026). Discovery and validation of a multi-protein panel for predicting non-fatal major adverse cardiovascular events in diabetic kidney disease.. Frontiers in endocrinology. https://doi.org/10.3389/fendo.2026.1883523