Multi-Omics Integration of Transcriptomics and Metabolomics with Machine Learning Uncovers Novel Risk Factors for Alzheimer's disease

BackgroundAlzheimers disease (AD) is a neurodegenerative disorder marked by cognitive decline, memory impairment, and functional deterioration. Its complex pathogenesis involves factors such as amyloid plaques, tau tangles, neuroinflammation, and synaptic dysfunction, but the precise mechanisms remain unclear, hindering effective treatment. Genetic, environmental, and lifestyle factors contribute to AD risk, yet their interactions are poorly understood. Recent advances in transcriptomics and metabolomics have shed light on the molecular underpinnings of AD, with gene expression alterations and metabolic disruptions implicated in disease progression. These multi-omics disruptions highlight the need for integrative analytical approaches to better characterize AD-relevant biology and advance biomarker discovery. ObjectivesTo integrate genetically imputed whole blood transcriptomics and plasma metabolomics to predict cognitive performance (PACC3) and to identify risk genes and metabolites contributing to prediction, thereby characterizing molecular signatures associated with cognitive performance in AD. MethodsThis study applies a machine learning algorithm to integrate genetically imputed whole blood transcriptomics and measured plasma metabolomics data to predict cognitive performance, as measured by PACC3 score, using data from the Wisconsin Registry for Alzheimers Prevention (WRAP) cohort (N = 1,046). After training a machine learning model on WRAP, the predictive performance was evaluated using an independent dataset from the Wisconsin Alzheimers Disease Research Center (ADRC) cohort (N = 85). Feature importance was assessed to identify genes and metabolites that may play a role as potential risk factors in AD. ResultsThe machine learning model achieved a normalized root mean squared error (NRMSE) of 0.743 {+/-} 0.037 and an R{superscript 2} of 0.311 {+/-} 0.016 across 5-fold holdout test folds in WRAP (p = 5.93 x 10-30), and an NRMSE of 0.915 and an R{superscript 2} of 0.061 when applied to the Wisconsin ADRC cohort. Feature importance revealed transcriptomic biomarkers such as RIPK1, IL6ST, and BIN1 whose higher imputed expression levels were associated with poorer cognitive performance whereas other potential biomarkers including UGP2, NDUFB5, and TMOD2 were associated with better cognitive performance, reflecting mitochondrial energy metabolism and molecular processes associated with cognitive resilience. Several predictive metabolites including benzoate, 3-phenylpropionate, and imidazolelactate also mapped to AD vulnerability signatures, while acyl-carnitine species such as hexanoylcarnitine (C6) and propionate-related metabolites aligned with metabolic resilience. ConclusionIntegrated analysis of transcriptomics and metabolomics demonstrated potential utility for identifying candidate biomarkers associated with cognition in AD. Genes and metabolites reflecting inflammatory signaling, mitochondrial dysregulation, and lipid metabolism emerged consistently among the most influential contributors. These findings align with well-established AD vulnerability pathways and highlight convergent biology across two omics layers. Collectively, this supports the value of multi-omics integration for improving molecular characterization of AD and advancing biomarker prioritization for future mechanistic and translational studies.