Predicting enhancer-gene links from single-cell multi-omics data by integrating prior Hi-C information

SCEG-HiC predicts links between genes and enhancers by integrating multi-omics data (either scATAC-seq/RNA-seq or scATAC-seq data alone) with three-dimensional omics data (bulk average Hi-C). The approach employs the weighted graphical lasso (wglasso) model to incorporate average bulk Hi-C data, effectively regularizing the correlation matrix with the prior Hi-C contact matrix as a penalty term.

Here, we generated mouse bulk average Hi-C data by averaging seven mouse Hi-C datasets from different cell types, including two embryonic stem cells (mESC1, mESC2), CH12LX, CH12F3, fiber cells, epithelial cells, and mature B cells. We adopted the activity-by-contact (ABC) model as a reference. The Hi-C matrices for each cell type were scale normalized at 5 kb resolution. To account for the known power-law decay of intra-chromosomal interactions, we corrected for differences in this decay across cell types before averaging. The final averaged Hi-C matrix was used as prior input for SCEG-HiC and is available as mouse_average_hic.tar.gz.

We evaluated the model on five human and five mouse paired scATAC-seq/RNA-seq datasets, and further applied it to COVID-19 PBMC scATAC-seq datasets. The human datasets, generated by 10x Genomics, include PBMC, skin stromal cells, fetal retina, brain gray matter, and developing cerebral cortex. These data are available as: PBMC_multiomic.rds, human_skin_multiomic.rds, human_retinal_multiomic.rds, human_brain_multiomic.rds, and human_cortex_multiomic.rds. The mouse datasets, collected from various platforms, cover embryonic brain, skin, adult cerebral cortex, thymic epithelial cells, and liver, and are available as: mouse_brain_multiomic.rds, mouse_skin_multiomic.rds, mouse_cortex_multiomic.rds, mouse_thymic_multiomic.rds, and mouse_liver_multiomic.rds. In addition, we applied the model to COVID-19 PBMC scATAC-seq datasets from SARS-CoV-2 infected individuals, available as covid_19_multiomic.rds. Data processing was carried out using standard Seurat and Signac workflows. This included quality control, normalization using SCTransform for RNA data, peak calling and matrix construction for ATAC data, dimensionality reduction using LSI, and cell type annotation based on canonical markers or original study metadata.