Implicit Hierarchical Tensor Decomposition of Single-Cell Four-Omics Data Reveals Cell-Type-Linked Candidate Enhancer--Promoter-Associated Programs

An implicit hierarchical tensor decomposition of CHARM mouse brain four-omics data links an Inh_Ndnf/Lamp5–Ex_L3/4_IT cell axis to contrasting enhancer–promoter programs, using 730,969 E–P pairs across 4,258 cells.

• Why it matters: Integrating single-cell modalities into enhancer–promoter programs that reflect 3D genomic context is difficult, and it is unclear which regulatory structure is shared versus modality-specific.
• What they did: The method jointly analyzed ATAC, H3K27ac, RNA, and reconstructed 3D enhancer–promoter distances in GSE303006, retaining 730,969 E–P pairs, then decomposed an implicit 730,969×4,258×20×4 tensor without materializing it.
• The result: Cell type explained more variance than biological replicate for 12 cell components in the largest Tucker core elements, and the recovered E–P and cell factors were reproducible across replicate splits, with modality-specific effects differing by module.

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