Omics-based Machine Learning Model to Predict AD dementia

Alzheimer’s disease (AD) dementia is a heterogeneous neurodegenerative disease among older adults. Early detection of AD dementia remains challenging due to heterogeneity in disease onset and progression. Our goal is to develop a genetic variants-based VAE model to predict AD dementia. We will use GWAS data collected in Health and Retirement Study (HRS). HRS has genotyped saliva DNA samples collected since 2006 at multiple time points during field visits yielding a total of 19004 unique participants. We will use the most recent version of genotype data from 2006 to 2015 for those samples that have both epigenetic and transcriptomic data available. The genotyping was performed by NIH Center for Inherited Disease Research, using the Illumina HumanOmni2.5-4v1/8v1 array, and genotyping QC analysis was performed at the University of Michigan using HumanOmni2.5-4v1 H for SNP annotation. We will use the quality and minor allele frequency (MAF) filters specified in the HRS QC report for genotypic data to filter out poor quality SNPs. We will use cognition measures collected in HRS 2016 survey to classify participants into 'Dementia' and 'Normal' using the Langa Weir Classification algorithm.

We will employ two main feature selection processes:
1. Based on the association with dementia, the top 50% of associated SNPs will be selected to input to the VAE model and filter out low-frequency SNPs.
2. We will also train a VAE model with a more comprehensive list of SNPs.

We will employ the model regularization by incorporating biological knowledge as constraints in the model using the gene-gene interaction network from REACTOME/ STRING. We will also evaluate the biological interpretability of latent features that are representative of input genetic variants. We will evaluate the distribution of weights of all encoded features to select the positive high and negative high features based on 2 SD above or below the mean weight. These selected features will be input for the pathway analysis to identify pathways associated with AD. The candidate genes identified can be used to develop blood-based biomarkers for early identification of AD

We will develop a genetic variants-based VAE model to predict dementia. We will employ various feature selection processes based on the complexity of data to evaluate the VAE model performance and to identify a representative list of genes. In addition, we will evaluate the biological interpretability of latent features obtained from the VAE encoder layer by extracting their decoder weights that captures the input feature contribution to the learned latent feature. This will also allow us to evaluate if the VAE model has learned novel features known to be associated with AD dementia. We will then utilize the learned features to identify the biological pathways associated with AD dementia.