Fig. 2: The representation-level results of the SHARE-seq mouse skin hair follicle dataset.

The decoupled representations contain the chromatin potential information in the latent space. A The UMAP constructed from concatenated RNA and ATAC representations, colored by cell type. B The UMAP constructed from RNA coupled representations, colored by cell type. C The UMAP of RNA decoupled representations, colored by cell type. D The UMAP constructed from ATAC decoupled representations, colored by cluster membership (cluster 4 and the rest of the clusters). E The distribution of cluster 4 and other clusters, along with chromatin potential vector field, shown on previously published UMAP ^4,75. F The UMAP constructed from ATAC coupled representations, colored by cell type. G The UMAP constructed from ATAC decoupled representations, colored by cell type. H The UMAP constructed from ATAC decoupled representations, colored by the value of ATAC decoupled latent representation 14, which mainly characterizes cluster 4 (see D). I The original UMAP of SHARE-seq data⁴, colored by the cosine distance between decoupled RNA and decoupled ATAC representations. J The cosine similarity between decoupled RNA and decoupled ATAC representations of cluster 4 (novel root) and rest of the cells. The boxplot shows the median (center line), interquartile range (box), and data within 1.5 × IQR from the quartiles (whiskers); points outside this range are plotted as outliers. The p-values are calculated using Welch’s two-sided t-test. K Expression levels of marker genes in the novel root and expected roots. L The Pearson’s correlations matrix of latent representations. Note the strong one-to-one mapping between coupled RNA and ATAC representations, which correspond to those in the red box ("Highly Coupled RNA-ATAC Pairs") in (M). M The Pearson’s correlations between latent representations for HALO (coupled representations), HALO (decoupled representations), GLUE, scMVP, MIRA, and MultiVI. Boxplots show the median (center line), interquartile range (box), and data within 1.5 × IQR from the quartiles (whiskers). All individual data points are overlaid on the boxplots. N The enriched transcription factors (TFs) for coupled ATAC latent representation 3 and decoupled ATAC latent representation 14. O The genomic matching score of coupled representations with different genomic distances from genes' TSS locations. Genomic matching score calculates the fraction of important ATAC representation peaks that lie within the specified genomic distance from the TSS of significant RNA representation genes. P The gene ontology (GO) enrichment of top genes from RNA coupled representation 3. Fisher’s exact test is used to calculate p-values for gene set enrichment analysis. SHARE-seq data are available at the MultiVelo website. Source data are provided as a Source Data file.