Prediction of context-specific regulatory programs and pathways using interpretable deep learning

Variational autoencoders (VAEs) are being widely adopted for the analysis of single-cell RNA sequencing (scRNA-seq) data. As with any non-linear models, however, they lack interpretability, which is a crucial aspect in the biomedical field where researchers want to be able to trust their model predictions. Our previously developed OntoVAE model addressed this issue by integrating biological ontologies in the decoder, which made the neuronal activations correspond to pathway activities. However, when multiple covariates are present, disentangling their relative contributions is challenging. To address this limitation, we developed COBRA, a VAE tool that combines the interpretable decoder part of OntoVAE with an adversarial approach that separates covariate effects in the latent space. In this work, we demonstrate the use of COBRA on two different scRNA-seq datasets in different contexts. We applied the tool to an interferon stimulated mouse dataset to separate the effects of celltype and treatment on transcription factors and biological pathways. We furthermore showed how COBRA can be used to predict the state of unseen celltypes.