Fig 1.
scTGCN learns a hybrid graph of both scRNA-seq data and scATAC-seq cell mappings, in the hybrid graph, transfer learning is used to transfer cell type labels from scRNA-seq data to scATAC-seq data. a. Broad schematic of scTGCN workflow. The input of scTGCN consists of two modaities. One is scRNA-seq data and the other is scATAC-seq, scATAC-seq data is converted to gene activity scores calculated from the accessibility peak matrix. scTGCN is constructed upon the transfer learning, comprising three key modules: omics-specific autoencoder module, domain adaption transfer learning module and Graph convolutional network model. b. Graph convolutional network captures inter and intra modalities information by two stage of semisupervised learning. The graph convolutional layer aggregates information from neighboring nodes to update the features of each node. c. the input of scTGCN omics-specific autoencoders comprises n sub-vectors, characterized by dimensions of (n, s), the output also consists of n sub-vectors with dimensions of s.
Fig 2.
Integration and analysis of overlapping cell types from scRNA-seq and scATAC-seq modalities in mouse cell atlas subset data.
a, t-SNE Visualization of scTGCN, Seurat, Conos and scJoint with cell types defined in Cusanovich et al. [46] as Color Labels. b, t-SNE Visualization of scTGCN, Seurat, Conos and scJoint with three protocols. c, Predicted cell types and fractions of agreement with Cusanovich et al. [46] for scTGCN, Seurat, Conos and scJoint. A clearer diagonal structure indicates a higher level of agreement. d, Comparison of modality silhouette coeffcient and cell-type silhouette coeffcient of different methods. e, Comparison of F1 scores of different methods. f, Comparison of ASW of different methods. g, Comparison of SAS of different methods. h, Comparison of NC of different methods. i, Comparison of foscttm of different methods.
Fig 3.
Analysis of large scale full mouse cell atlas data.
a, t-SNE plots generated from the top 100 dimensions resulting from singular value decomposition of TF-IDF transformed ATAC-seq data, with data points colored according to their original labels. b, Refining scATAC-seq annotations in heterogeneous atlas data. c, The transferred labels accuracy of each cell type identified by different methods based on large-scale full atlas data. d, Gene expression levels of cd19 in B cells, Eno2, Snap25, Rbfox3, Calb1 in neuron cells, and Col1a1, Fn1, Vim in stromal cells. e, Marker expressions in stromal cells: Col1a2, Col1a1, and pdgfra. The left column displays the high-level gene activity scores and the right column exhibits the gene expression levels in endothelial, stromal cells, and others from scRNA-seq.
Fig 4.
Integration of full mouse atlas data.
a, t-SNE Visualization of scTGCN, Seurat, Conos and scJoint with cell types defined in Cusanovich et al. [46] as Color Labels. b, t-SNE Visualization of scTGCN, Seurat, Conos and scJoint with three protocols. c, Comparison of NC and SAS of different methods. d, Comparison of cell type ASW and modality ASW of different methods. e, Comparision of foscttm of different methods. f, Comparison of modality silhouette coeffcient and cell-type silhouette coeffcient of different methods.
Fig 5.
Integration of multimodal PBMC data.
a, t-SNE visualizations of PBMC data, comparing scTGCN, Seurat, Conos and scJoint results. b, The plot represents different technology of scTGCN, Seurat, Conos and scJoint. c, Predicted cell types and fractions of agreement with true label. For scTGCN, Seurat, Conos and scJoint. A clearer diagonal structure indicates a higher level of agreement. d, Comparison of F1 scores of scTGCN, Seurat, Conos and scJoint. e, Comparison of modality silhouette coefcient and cell-type silhouette coefcient of different methods. f, Comparision of ASW of different methods. g, Comparision of foscttm of differet methods. h, Comparision of NC of differet methods. i, Comparision of SAS of different methods.