Integration of clinical and proteomic risk factors enhances prognostic modelling of incident vascular complications in type 2 diabetes

TL;DR

Adding blood protein data to standard clinical factors improves prediction of vascular complications in type 2 diabetes, especially for macrovascular outcomes.

Contribution

This study identifies plasma protein signatures that enhance vascular complication prediction in type 2 diabetes beyond traditional clinical markers.

Findings

Proteomics-integrated models improved prediction of macrovascular complications (C-index 0.72 vs. 0.60 with clinical data alone).

Selected proteins like LRRC37A2 and NT-proBNP showed strong predictive value for macrovascular outcomes.

Predictive gains were stable over 10 years for macrovascular outcomes but less consistent for microvascular complications.

Abstract

Type 2 diabetes complications manifest across various organs, but are fundamentally rooted in vascular dysfunction. This study aims to identify plasma protein signatures that improve prediction of macrovascular and microvascular complications in type 2 diabetes over classical clinical factors, assess the stability of their prognostic performance over time, and explore the cross-ancestry generalizability of the developed models.

We analysed 2,923 plasma proteins in 917 European-ancestry UK Biobank participants with prevalent type 2 diabetes but no prior vascular disease at baseline. The primary outcomes were time to first macrovascular or microvascular complication, identified through ICD-10 codes during a mean follow-up of 10.41 years. Protein selection was performed using clinical-variables-prioritized LASSO Cox regression across 100 resamples to identify proteins offering predictive value beyond established clinical markers. Stably selected proteins were then integrated with and evaluated against clinical-only models using optimism-corrected C-index, time-dependent AUC and Brier score. We also conducted exploratory analyses to assess model generalizability in 116 European genetic outliers and in 80 Asian and 54 African ancestry participants within the UK Biobank.

For macrovascular outcomes, 37 proteins were selected, led by LRRC37A2, NT-proBNP, CHGA, APOD and STAB2. For microvascular complications, 9 proteins were selected, led by IL15, FAM3C and TNFSF11, with overall more moderate stability across resampling. The proteomics-integrated models significantly improved prediction of type 2 diabetes vascular complications beyond clinical markers (Harrell’s C: macrovascular 0.72 vs. 0.60; microvascular 0.67 vs. 0.62) and demonstrated stable prognostic accuracy over 10 years for macrovascular outcomes. In exploratory generalizability analyses, predictive gains of proteomics integration were maintained in European genetic outliers but diminished in African and Asian participants.

Integrating proteomics with clinical data enhances risk prediction of type 2 diabetes vascular complications, especially for macrovascular outcomes. However, less precise prediction for microvascular complications and preliminary evidence of limited cross-ancestry generalizability highlight the need to expand targeted biomarker panels and quantification in larger, more ancestry-diverse cohorts to ensure effective and equitable clinical implementation of proteomics.

The online version contains supplementary material available at 10.1186/s12933-026-03083-6.

Figures