How can deep learning architectures (variational autoencoders or graph neural networks) be effectively designed to integrate heterogeneous multi omics data such as ATAC-seq and RNA-seq, and methylation while maintaining interpretability of latent representations??????
"Integration" is fairly straightforward - TotalVI does this (CITE + RNA) as a very basic example. Most of the foundation model papers (epibert, scmeformer, evo2, borzoi) take the approach of having a single latent represtation with multiple "output heads"
"Interpretability" is a whole different ball game. But I would caution that you may be asking models to do something more than you ask of other methods. If you've integrated data together and have batch-corrected PCs (or TF-IDF PCs) - are these features "interpretable"? I would venture that to understand latent features of just about any sort, one would have to do the same kind of analysis - regardless of whether they come from deep learning or other methodology.
Use of this site constitutes acceptance of our User Agreement and Privacy Policy.
version 2.3.6
This paper I'm reading right now use a VAE-based architecture for multi-modal integration and has explainable latent variables representation. It contains an example of ATAC and RNA-seq integration (Fig 6) in mouse immune cells. https://www.nature.com/articles/s41467-023-37477-x