Beyond amyloid-ß and tau: A cell-based multivariate blood test (iCAP) for early Alzheimer's detection

Abstract: Background: With amyloid-ß (Aß) therapies for Alzheimer's disease (AD) under active debate, there is a need for diagnostic tools that reflect disease biology beyond Aß and tau for early detection and for patient stratification. The indicator cell assay platform (iCAP) is a tool for blood-based diagnostics that uses standardized cells as biosensors to transduce complex circulating signals into gene-expression readouts, which are then used to train multivariate disease classifiers for precision medicine. We developed an iCAP for early detection of Alzheimer's disease (AD-iCAP). Methods: In a retrospective study, AD-iCAP was developed by incubating banked plasma samples from patients with early-stage AD (mild cognitive impairment or mild dementia) and age-matched controls with standardized neurons; whole-transcriptome responses were measured and used train disease classifiers by machine learning. The assay was optimized and analytically validated. Patient AD-iCAP data were separated into a training set, external validation set and an independent test set and used to parameterize and test models. To minimize bias, modeling features were selected from a predefined (a priori) 84-gene panel derived from an independent AD-iCAP dataset generated using plasma from 5xFAD mice. Results: We developed AD-iCAP using 191 banked plasma samples across three cohorts. Performance was assessed in two held-out sets: an external validation set (n=82; AUC 0.64, 95% CI 0.51–0.78) and an independent test set (n=23, AUC 0.77, 95% CI 0.57–0.96). Systems biology analyses of differential response profiles showed concordance with postmortem AD brain transcriptomes and enrichment of AD-relevant pathways beyond amyloid, including cholesterol biosynthesis, synaptic structure/neurotransmission and NGF/TrkA signaling (FDR < 0.05). The final model's features included AD-linked genes AKT3, GPR50, PALLD and RGMA, related to neuronal signaling and cytoskeletal/axon-guidance processes. Conclusions: AD-iCAP is a blood-based diagnostic for early-stage detection of AD. It outputs cell-based transcriptional profiles with disease-relevance that are induced by circulating factors. In retrospective, multi-cohort testing, it showed modest-to-good discrimination, supporting prospective confirmation in larger cohorts. Because its readout captures biology beyond amyloid and tau, it may provide complementary information for combined testing. The multivariate readout supports further development for patient stratification as Aß-targeted therapies and alternatives evolve. Overall design: Overview. The goal of the study was to develop a biosensor-based diagnostic assay whereby cultured cells (iCell neurons) detect and respond to Alzheimer's disease (AD) signals in human blood samples and machine learning tools are used to develop disease classification models based on the cellular gene expression profiles of the cells. We exposed human indicator cells (iCell Neurons (Fujifilm) to plasma from patients and measured their gene expression response. Plasma samples were from patients who were either cognitively normal or who had early stage Alzheimer's disease. The differential gene expression pattern between the two conditions was used to predict presence of Alzheimer's disease in the patients using machine learning tools. The study included retrospective analysis of 470 human plasma samples from 3 cohorts (See data file 1 of teh manuscript for complete description). Plasma was from either affected patients (preclinical AD (preAD) or MCI stage) or from unaffected control subjects (normal). In addition, a limited number of plasma samples from patients with non-AD dementia were also included as controls in some experiments. Samples were from three different cohorts, each with a ~1:1:1 ratio of samples from each preAD, MCI and normal class, with gender and age were roughly balanced between classes. Plasma samples were exposed to indicator cells, and total RNA was isolated and analyzed by RNA-seq. RNA-seq data from affected and unaffected samples were used to train and test machine learning classifiers to detect early-stage AD. Before classifier development, sample QC and RNA-seq data QC were performed and samples were excluded that failed quality check. For classifier development, only samples of MCI vs normal class were used. For feature selection for model development, plasma samples from a mouse model of Alzheimer's disease (5xFAD) was analyzed in the AD-iCAP and genes with AD versus normal differential expression were selected as candidate features. For the mouse study: Longitudinal AD-iCAP data were generated from mice at 3, 5, and 7 months, spanning pre-symptomatic and MCI stages,34,35 under nine assay conditions (3 mouse ages x 3 neuron pre-culture times). After RNA-seq and data processing using two different GC bias correction methods, we obtained 18 differential expression datasets. From these data, we derived two feature sets: 398 genes detected in =1 dataset (Set 1) and 84 genes detected in =3 datasets (Set 2), all with adjusted p = 0.05 (see Data File 4 of the manuscript). Human plasma samples were processed in the iCAP in 24 experimental batches and each batch included four technical replicates of "reference control plasma samples", which were technical replicates of plasma from a non-affected individual at 40 years of age. These referece control samples were used for quality control and normalization only.