A machine-learning analysis of nearly 3,000 plasma proteins identified a proteomic signature that accurately predicted dementia risk in people with type 2 diabetes up to almost 15 years before diagnosis, according to a study published in the Journal of Advanced Research. 

“Proteomic signatures may offer improved prediction over standard clinical strategies for assessing individuals at high risk of dementia in T2D patients, and may be useful for identification of the potential targets for possible drug therapy to reduce or delay the onset of dementia,” the researchers wrote. 

They analyzed 2,920 plasma proteins from 52,958 dementia-free UK Biobank Pharma Proteomics Project participants, including 3,292 with type 2 diabetes. Proteins were measured with the Olink Explore platform. 

During a median follow-up of 14.6 years, 1,473 people developed dementia, including 713 with Alzheimer disease and 291 with vascular dementia. Among those with type 2 diabetes, 203 developed dementia, including 101 with Alzheimer disease and 52 with vascular dementia. 

Using Cox regression models with interaction terms, researchers identified 471 proteins whose associations with dementia differed significantly according to type 2 diabetes status. Of these, 250 had higher hazard ratios among people with type 2 diabetes than among those without type 2 diabetes, whereas 221 had lower hazard ratios. 

Among people with type 2 diabetes, higher plasma levels of rho guanine nucleotide exchange factor 12 (ARHGEF12) were specifically associated with increased dementia risk, with a hazard ratio of 1.45. Tumor necrosis factor receptor superfamily member 27 (EDA2R) and neurofilament light polypeptide (NEFL) were also associated with future dementia risk, although both showed stronger predictive associations in participants without diabetes. 

The researchers then developed a 51-protein predictive model for incident dementia in type 2 diabetes. In the final least absolute shrinkage and selection operator (LASSO)-based model, latent-transforming growth factor beta-binding protein 2 (LTBP2), glial fibrillary acidic protein (GFAP), and apolipoprotein E (APOE) were the strongest contributors. 

Correlation analyses linked many of the predictive proteins to glucose metabolism, lipid metabolism, kidney function, and inflammatory pathways. 

Patients were classified according to their proteomic risk score. Those with scores above the median had a substantially higher risk for future dementia outcomes than those with lower scores. Hazard ratios ranged from 5.99 for vascular dementia to 9.45 for all-cause dementia, indicating approximately sixfold to ninefold greater risk. 

To build the predictive model, investigators used repeated LASSO feature selection to narrow nearly 3,000 proteins to a final panel of 51 biomarkers. Among five machine-learning approaches tested, the LASSO-based model performed the best.  The model predicted incident dementia with an area under the curve (AUC) of 0.835 and a concordance index of 0.829, outperforming models based only on demographic variables or cognitive testing. It also maintained high predictive performance for Alzheimer disease and vascular dementia. 

Sensitivity analyses addressed concerns about model complexity. A 20-protein model retained 89% of the original model’s AUC performance, while a 13-protein model retained 86%, suggesting the signal was not driven solely by the number of proteins included. 

The researchers noted several limitations, including a smaller number of participants with type 2 diabetes than without, a predominantly White study population, limited data on diabetes severity, medication data collected only at baseline, and the absence of contemporary glucose-lowering therapies, including GLP-1 receptor agonists and SGLT2 inhibitors.  

The researchers reported no conflicts of interest.