Metabolomic-based risk score informed by neuroimaging biomarkers improves Alzheimer’s disease risk

239 

Abstract Submission 

Erica Suh¹, Kwangsik Nho², Li Shen^1,3, Andrew Saykin², Dokyoon Kim^1,3

¹Department of Biostatistics, Epidemiology & Informatics, University of Pennsylvania, Philadelphia, PA, ²Department of Radiology and Imaging Sciences, Indiana University School of Medicine, Indianapolis, IN, ³Institute for Biomedical Informatics, University of Pennsylvania, Philadelphia, PA

Erica Suh  
Department of Biostatistics, Epidemiology & Informatics, University of Pennsylvania
Philadelphia, PA

Kwangsik Nho  
Department of Radiology and Imaging Sciences, Indiana University School of Medicine
Indianapolis, IN

Li Shen  
Department of Biostatistics, Epidemiology & Informatics, University of Pennsylvania|Institute for Biomedical Informatics, University of Pennsylvania
Philadelphia, PA|Philadelphia, PA

Andrew Saykin  
Department of Radiology and Imaging Sciences, Indiana University School of Medicine
Indianapolis, IN

Dokyoon Kim  
Department of Biostatistics, Epidemiology & Informatics, University of Pennsylvania|Institute for Biomedical Informatics, University of Pennsylvania
Philadelphia, PA|Philadelphia, PA

Early risk prediction and diagnosis of Alzheimer's disease (AD) remain challenging in the clinical setting. Polygenic risk scores (PRS), while common, fall short in specificity and sensitivity for AD risk estimation. Recent developments in neuroimaging techniques, including FDG-PET and AV45-PET, have highlighted changes in glucose metabolism, brain structure, and blood-brain barrier dysfunction, aligning with biofluid biomarker data (Sweeney et al. 2018). Metabolomic technologies, offering a more cost-effective and non-invasive approach, have identified disease-specific biomarkers, enabling the potential for metabolites to shed further light on the pathophysiological cascade of AD (Nho et al. 2021, Quintero et al. 2021). We aim to develop a novel metabolomics-based score that leverages significant correlations between lipid metabolites and PET biomarkers to enhance AD risk prediction and aid in early disease detection.

Serum-based metabolomics data containing 781 lipid species were collected from 997 fasted participants of the Alzheimer's Disease Neuroimaging Initiative (ADNI). We focused on two separate tasks: 1) cognitively normal control (CN) vs AD classification, and 2) mild cognitive impairment (MCI) conversion prediction. For each task, we measured Pearson correlation coefficients between each lipid species and the neuroimaging biomarkers, FDG- and AV45-PET. Significantly correlated lipids (p<0.05) were extracted to estimate a metabolomics-based risk score for each patient using a logistic regression model with an 80:20 train test split and 10-fold cross-validation. Prediction accuracy was measured using the Area Under Receiver Operating Characteristic (AUC) and Area Under the Precision-Recall Curve (AURPC). Risk-based stratification and interpretation analysis was also performed to further assess the potential clinical utility of the generated risk score. Results were compared to those of conventional PRS, as well as a baseline model which uses all lipid species as features. PRS was calculated using pruning and thresholding (PRS-pT, p<1e-5) with GWAS summary statistics from IGAP.

The metabolomics-based score outperformed PRS and the baseline model in classifying CN vs AD, achieving an AUC of 0.797 with 187 lipid features (PRS AUC=0.605, and baseline AUC=0.678). It also showed superior performance in predicting MCI conversion, with an AUC of 0.726 using 174 lipid features (PRS AUC=0.525, baseline AUC=0.657). With the addition of covariates APOE, sex, and age, the AUC increased to 0.838 and 0.756 for each task, respectively (Figure 1). The metabolomics-based score also demonstrated lower risk in controls and, conversely, higher risk for AD patients. Stratification analysis revealed improved calibration with a smooth increase in predicted AD patients as the risk score rose. Many of the significant metabolites belonged to choline-containing phospholipids, such as lysophosphatidylcholine and phosphatidylcholine, in which pronounced increases of plasma levels have been observed in AD patients (Whiley et al. 2014, Tomioka et al., 2017).

We developed and evaluated a novel metabolomics-based risk score which leverages PET neuroimaging biomarkers to robustly identify individuals with high or low risk of developing AD. Compared to conventional PRS, our risk score improved prediction performance and risk stratification of AD patients. Further study is required to functionally validate the selected metabolites and their roles in AD-related pathophysiology, to ultimately identify risk or progression-related biomarkers that can aid in the downstream development of therapeutic treatments for AD.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ¹

Genetic Association Studies

Classification and Predictive Modeling ²

Physiology, Metabolism and Neurotransmission Other

Aging

Degenerative Disease

Machine Learning

Neurological

Phenotype-Genotype

Positron Emission Tomography (PET)