Uncovering hidden factors of cognitive resilience in Alzheimer's disease using a conditional-Gaussian mixture variational autoencoder.

Understanding the molecular mechanisms underlying cognitive resilience in Alzheimer's disease (AD) is essential for identifying novel drivers of preserved cognitive function despite neuropathology. Rather than directly searching for individual genetic factors, we focus on latent factors and deep learning modeling as a systems-level approach to capture coordinated transcriptomic patterns and address the problem of missing heritability. We developed a conditional-gaussian mixture variational autoencoder (C-GMVAE) that integrates single-cell transcriptomic data with behavioral phenotypes from a genetically diverse BXD mouse population carrying 5XFAD mutations. This framework learns a structured latent space that captures biologically meaningful variation linked to cognitive resilience. The resulting latent variables are highly heritable and reflect genetically regulated molecular programs. By projecting samples along phenotype-aligned axes in the latent space, we obtain continuous gradients of genetic features associated with AD cognitive resilience. These findings highlight the potential of latent variable approaches not only to model high-dimensional biological data but also to reveal hidden factors driving phenotypic variability in neurodegeneration.