Multimodal Probabilistic Machine Learning for Uncovering Latent Factors of Resilience in Alzheimer’s Disease

PROJECT SUMMARY The complex etiology underlying cognitive aging, Alzheimer’s disease (AD), and AD-related dementias (ADRD) involves factors across biological scales that are highly dynamic and interdependent. A deeper mechanistic understanding of individual differences in disease is likely to accelerate the design of new therapies. This requires examining data across multiple genetic and environmental contexts, and across multiple data modalities simultaneously (molecular, anatomic, functional imaging, cognitive assays under stress and throughout aging). Identifying targetable disease factors from such high-dimensional and multimodal datasets is a big data challenge, in part due to the immense search space of possible underlying factors to cognitive resiliency. And even once a latent factor or set of factors is hypothesized, such claims must be causally investigated – a slow, expensive undertaking. Here, we aim to develop novel multimodal variational autoencoder (VAE) models that learn factors associated with cognitive resilience to AD. We will employ generative methods to not just discover new factors from existing data but also identify dynamic relationships among factors in the learned latent space. Our model will enable us to predict the effects of perturbations to the latent factors and offer a means of testing complex genetic or environmental interventions prior to investing effort experimentally. In Aim 1, we will develop a multimodal VAE to shape a resiliency-susceptibility latent space and model dynamical features contributing to both the onset and progression of cognitive aging across short and long timescales. In Aim 2, we will incorporate Bayesian hierarchical layers to enable iterative updates of new data and employ generative sampling to produce testable hypotheses of how time or causal interventions change the latents towards resiliency. Finally, in Aim 3 we will validate learned factors via genetic engineering in our diverse panel of AD-BXD mice that better model heterogeneity of human AD, spanning the spectrum from highly resilient to highly susceptible. We anticipate that our methods will drastically shorten the model-experiment validation cycle and will lead to novel findings of latent factors that drive resilience in AD and ADRD and of essential mechanisms for designing new therapies.