scHiR learns joint cellular representations from paired single-cell transcriptome and 3D genome data

Simultaneously sequenced scRNA-seq and scHi-C technologies provide matched measurements of gene expression and three-dimensional (3D) genome organization within individual cells, offering complementary insights into cellular state. However, integrative analysis remains challenging due to the heterogeneous structures and characteristics of these modalities. To realize the full potential of such data, we need effective computational methods to learn joint cellular representations. Here, we present scHiR, a framework based on a hyperspherical variational autoencoder (S-VAE) for learning co-embeddings from paired gene-expression profiles and chromatin contact maps. The framework facilitate downstream analyses such as cell clustering, fine-grained cell-state characterization, and developmental trajectory inference, while modality-specific decoders enable the signal recovery of input profiles. We evaluated scHiR across six paired datasets generated by HiRES, GAGE-seq, and LiMCA. scHiR captured cell-type-specific information and achieved improved clustering performance compared with representative single-omics methods. It was also capable of resolving fine-grained subclusters and revealed discordant and distinct cellular profiles. In developmental biological datasets, scHiR preserved trajectory-related structure and improved pseudotime in agreement with known developmental progression. Finally, scHiR generated reconstructed transcriptomic and chromatin contact profiles with reduced noise, enhanced cell-type-related signals, and improved recovery of held-out chromatin contacts. These results demonstrate that scHiR provides an effective method for learning joint cellular representations from paired scRNA-seq and scHi-C data.