Abstract A106: Plasma Proteomic Signatures Shared Across Cerebrovascular Diseases and Cognitive Decline

Background: Cerebrovascular diseases and cognitive decline are major public health challenges in aging populations, with epidemiological evidence showing vascular contribution to increased dementia risk. However, the molecular mechanisms connecting these 2 disease domains remain incompletely understood. We aimed to identify shared plasma proteomic signatures to uncover common molecular pathways and inform strategies for intervention across both disease domains. Methods: We collected plasma protein data from the UK Biobank Pharma Proteomics Platform (UKB-PPP) in White British participants linked to clinical and death records. The following 7 incident binary outcomes after enrollment were identified from diagnosis codes: acute ischemic stroke (AIS), intracerebral hemorrhage (ICH), transient ischemic attack (TIA), small vessel disease (SVD), vascular dementia (VasD), Alzheimer's disease (AD), and mild cognitive impairment (MCI). A multi-task least absolute shrinkage and selection operator (LASSO) jointly modeled all the outcomes and leveraged shared structure across outcomes to enhance predictive strength and yield a more stable set of shared protein predictors. Results: We included 43,073 participants (median age 59 years [IQR 51-64], 54% female) and identified 2,911 protein markers each with less than 20% missingness. The joint model revealed a compact and interpretable coefficient pattern ( Figure ). The glial injury marker GFAP and the axonal injury marker NEFL were clear cross-domain predictors. GFAP was most predictive of AD and bridged to the cerebrovascular domain through SVD and VasD, while NEFL bridged through AIS, TIA, and SVD. The cardiac strain marker NT-proBNP was predictive mainly of cerebrovascular outcomes. A synaptic protein group (VGF, SNAP25, SYT1, NPTXR) and the lipid transport protein APOE were primarily predictive of AD and bridged to the cerebrovascular domain through SVD and VasD. For each protein, the predictive directions were broadly consistent across outcomes. Conclusions: Multi-task learning across cerebrovascular and cognitive outcomes identified shared protein predictors, highlighting glial and axonal injury, synaptic integrity, and lipid transport as convergent pathways that connect both disease domains. These findings may inform targets relevant to prevention and treatment for both types of diseases. Future work will compare alternative multi-task learning frameworks and assess generalizability across different subpopulations.