Using hierarchical variational autoencoders to incorporate conditional independent priors for paired single-cell multi-omics data integration

Ping-Han Hsieh, Ru-Xiu Hsiao, Tatiana Belova, Katalin Ferenc, Anthony Mathelier, Rebekka Burkholz, Chien-Yu Chen, Geir Kjetil Sandve, Marieke Lydia Kuijjer

Keywords: multi-modal variational autoencoder, hierarchical variational autoencoder, single-cell multi-omics data integration

TL;DR: We propose a novel probabilistic learning framework that explicitly incorporates the conditional independent relationships between multi-modal data using a generalized hierarchical variational autoencoder.

Abstract: Recently, paired single-cell sequencing technologies have allowed the measurement of multiple modalities of molecular data simultaneously, at single-cell resolution. Along with the advances in these technologies, many methods based on variational autoencoder have been developed aiming at integrating paired single-cell multi-omics data. However, how to incorporate prior biological understanding of data properties into such models remains an open question in the field. Here, we propose a novel probabilistic learning framework that explicitly incorporates conditional independence relationships between multi-modal data as a directed acyclic graph using a generalized hierarchical variational autoencoder. Applying our approach to single-cell ATAC and RNA-seq data, we find that our method can identify cell clusters with distinct expression profiles that are not driven by chromatin state. We anticipate that our proposed framework can help construct flexible graphical models that reflect biological hypotheses with ease and unravel the interactions between different biological data types, such as different modalities of paired single-cell multi-omics data. The implementation of the proposed framework can be found in the repository https://github.com/kuijjerlab/CAVACHON.