New arχiv pre-print
We have posted a new arχiv pre-print by Ping-Han. Ping-Han developed CAVACHON, or Cell cluster Analysis with Variational Autoencoder using Conditional Hierarchy Of latent representatioN, an approach to integrate multi-modal single-cell data.
CAVACHON is a hierarchical variational autoencoder (VAE) that can incorporate information on data dependencies in the form of a directed acyclic graph (DAG). For example, it can be applied to SNARE-Seq data, conditioning the RNA layer on ATAC signals. Through integrative multi-facet clustering during training, CAVACHON can find clusters in the joint and distinct latent spaces. Clustering after training is not needed, which allows for speedy clustering, as well as clustering of new cells onto an existing latent space. An interesting aspect of CAVACHON is that, as a generative model, it can construct “chimeric cells,” combining e.g. chromatin states from one cell type with expression from another. Thus, differential signal specific to each data modality (e.g. differential expression driven by the expression of transcription factors) can be identified. CAVACHON is highly adaptable and not limited to single-cell multi-modal data. It can, for example, also be used in time series analysis. We anticipate it will facilitate the construction of flexible graphs that capture the complexity of biological data, and hope its applications will contribute to our understanding of regulatory mechanisms.
More information on the pre-print can be found here.
