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

scHiR learns joint cellular co-embeddings from paired scRNA-seq and scHi-C, capturing cell-type-specific structure and improving clustering versus single-omics baselines across six paired datasets.

• Why it matters: Matched gene expression and 3D genome contacts can reveal cellular state, but integrative modeling is difficult because the modalities differ in structure and signal characteristics.
• What they did: The method uses a hyperspherical variational autoencoder (S-VAE) to co-embed paired gene-expression profiles and chromatin contact maps, with modality-specific decoders; it was tested on six datasets from HiRES, GAGE-seq, and LiMCA.
• The result: scHiR resolves fine-grained subclusters, preserves trajectory-related structure and improves pseudotime in developmental datasets, and reconstructs transcriptomic and chromatin contact profiles with reduced noise and better recovery of held-out contacts.

AI summary of the abstract · may contain errors · not medical advice · read the paper.