Multiomics profiling of human plasma and CSF reveals ATN derived networks and highlights causal links in Alzheimer’s disease

Abstract

INTRODUCTION: This study employed an integrative system and causal inference approach to explore molecular signatures in blood and CSF, the Amyloid/Tau/Neurodegeneration [AT(N)] framework, MCI conversion to AD, and genetic risk for AD. METHODS: Using the EMIF-AD MBD cohort, we measured 696 proteins in cerebrospinal fluid (n=371), 4001 proteins in plasma (n=972), 611 metabolites in plasma (n=696) and genotyped data in whole-blood (7,778,465 autosomal SNPs, n=936). We investigated associations: molecular modules to AT(N), module hubs with AD Polygenic Risk scores and APOE4 genotypes, molecular hubs to MCI conversion and probed for causality with AD using Mendelian Randomization (MR). RESULTS: AT(N) framework associated key hubs were mostly proteins and few lipids. In MR analyses, Proprotein Convertase Subtilisin/Kexin Type 7 showed weak causal associations with AD, and AD was causally associated with Reticulocalbin 2 and sphingomyelins. DISCUSSION: This study reveals multi-omics networks associated with AT(N) and MCI conversion and highlights AD causal candidates.

Multiomics machine learning identifies inflammation molecular pathways in prodromal Alzheimer’s Disease

“…Clinical covariates are included in all the models. Statistics of the clinical characteristics of this dataset has been previously published by Shi et al 22 The same pipeline was applied to create both multiclass models for proteins with covariates and the multiclass models for metabolites with covariates. Classes were first balanced using the SMOTE approach 42 with the 'imbalanced-learn' python package.…”

“…The DL model produced an AUC value of 0.85, XGBoost model 0.88 whereas the RF model resulted in a 0.85 AUC value (all values based on a 95% confidence interval). Additionally, the same proteomics data alone was investigated by Shi et al 22 , where ML algorithms were used to classify the donors into amyloid positive and amyloid negative participants. In this study, Lasso Regression and SVM models parsed a predictive panel composed of 44 proteins, age and the risk gene APOE4 achieving an AUC of 0.68 in the replication group.…”