Representational Learning from Healthy Multi-Tissue Human RNA-seq Data such that Latent Space Arithmetics Extracts Disease Modules

1Developing computational analyses of transcriptomic data has dramatically improved our understanding of complex multifactorial diseases. However, such approaches are limited to small sample sets of disease-affected material, thus being sensitive to statistical biases and noise. Here, we ask if a variational autoencoder (VAE) trained on large groups of healthy, human RNA-seq data of multiple tissues can capture the fundamental healthy gene regulation system such that the learned representation generalizes to account for unseen disease changes. To this end, we trained a multi-scale representation to encode cellular processes ranging from cell types to genegene interactions. Importantly, we found that the learned healthy representations could predict unseen gene expression changes from 25 independent disease datasets. We extracted and decoded disease-specific signals from the VAE latent space to dissect this finding. Interestingly, the gene modules corresponding to this signal contained more disease-specific genes than the respective differential expression analysis in 20 of 25 cases. Finally, we matched genes related to the disease signals to known drug targets. We could extract sets of known and potential pharmaceutical candidates from this analysis and demonstrate the utility in three use cases. In summary, our study showcases how data-driven representation learning using a VAE as a foundational model allows an arithmetic deconstruction of the latent space such that biological insights enable the dissection of disease mechanisms and drug targets. Our model is available at https://github.com/ddeweerd/VAE_Transcriptomics/.