Fig. 1

Overview of scDART. ascDART takes as input a scRNA-seq data batch, a scATAC-seq data batch, and a pre-defined GAM. It learns the latent embedding of integrated data from the two data batches and a more accurate gene activity function between regions and genes. This gene activity function can be used to predict scRNA-seq data from scATAC-seq data (the predicted scRNA-seq data is also called pseudo-scRNA-seq data). b The neural network structure of scDART. scDART includes two modules: (1) the gene activity function module is a fully-connected neural network. This module encodes the nonlinear regulatory relationship between regions and genes, and generate the pseudo-scRNA-seq data from scATAC-seq data. (2) The projection module takes in the scRNA-seq data and the pseudo-scRNA-seq data and generates the latent embedding of both modalities