Assessing Alzheimer’s disease risk and heterogeneity using multimodal machine learning approaches

The objective of this study is to develop machine learning models using genetic and phenotype data from the NIAGADS database https://dss.niagads.org/. We will develop both unsupervised and supervised learning models to characterize the heterogeneity and risk of Alzheimer’s disease (AD). This is an MPI study in collaboration with Dr. Anita DeStefano at Boston University School of Public Health.

For the first aim, we will build an expandable multimodal unsupervised machine learning framework to investigate AD heterogeneity. We will perform AD subtyping by harnessing the rich multimodality information across a wide spectrum of data (e.g., genetics, images and blood biomarkers). A Bayesian kernel network will be built to estimate the relative weight of each individual data modality, which would also allow the addition of new data modalities as they become available. The analyses will be performed both within and between ethnic populations.

For the second aim, we will build an expandable multimodal supervised machine learning framework to quantify AD risk from longitudinal follow-up of clinically normal elders. We will build a separate deep learning network for each data modality in consideration of its unique feature sets. A multiplicative strategy will then be taken to aggregate information from different modalities with weighted contributions. Feature selection will also be performed to identify the most informative features predictive of AD risk.

For the third aim, we will build AD-related gene regulatory networks in post-mortem human brain samples. We will examine the association of multi-omics data with AD, which will be used to assign gene priority based on the combinatorial evidence from each type of omics data. A gene ontology-guided greedy search strategy will then be implemented to build gene regulatory networks, and identify key drivers that might be potential therapeutic targets for AD. The analyses will be stratified by ethnic populations and AD phenotypic clusters.

Alzheimer's disease (AD) is the most common form of dementia characterized by progressive loss of cognitive function. There are very limited treatment options for AD. For the current application, we seek to develop multimodal machine learning models by leveraging the rich collection of AD-related omics data and phenotypical data recently generated from the Alzheimer's Disease Sequencing Project (ADSP). Three aims will be pursued in the current application. For Aim 1, we will build an expandable multimodal unsupervised machine learning framework to investigate AD heterogeneity. For Aim 2, we will build an expandable multimodal supervised machine learning framework to quantify AD risk from longitudinal follow up of cognitively normal elders. For Aim 3, we will build AD-related gene interaction networks in post-mortem human brain samples. The present application represents an innovative approach to identify individuals at high risk of AD. The outlined strategy will provide new insights into the risk stratification and prevention strategies for AD.