Uncovering hypergraphs of cell-cell interaction from single cell RNA-sequencing data

Key results

1. SuggestedOn simulated data with three cell types, scTensor matched a label-permutation test when all interactions were one-to-one. When interactions were many-to-many, it still detected all three ground-truth L-R sets, while the permutation test missed two of them.

   Simulations: 3 cell types × 50 cells, 500 L-R pairs (three true sets of 50, 350 non-differential). Case I: no AUC difference between methods. Case II: the label-permutation test (the authors' reimplementation of CellPhoneDB's approach) missed the blue and green sets and its AUC fell; scTensor separated each set as its own CaH with AUCs still high. AUC values appear only in figures. · Accuracy of the detection of CCIs and the…; Simulation datasets; Label permutation method

2. ShownIn four published scRNA-seq datasets (female and male human fetal germ cells (FGCs) with gonadal niche cells, human metastatic melanoma, and mouse heart non-myocytes), scTensor recovered the cell-cell interactions that the original studies reported.

   Germline Female: BMP-pathway Soma-FGC interactions as CaH(4,2,5). Germline Male: AMH-BMPR (with AMHR2) as CaH(1,1,10). Melanoma: interactions between T cells and cancer-associated fibroblasts as CaH(1,1,3)/CaH(1,1,9), with complement-related C3, CXCL12, CFB, C4A. NonMyocyte: pericyte/fibroblast-macrophage Il34/Csf1-Csf1r as CaH(2,2,19). Agreement is judged qualitatively. · Biological interpretation of real datasets; Real datasets

3. SuggestedscTensor also proposed interactions that the original studies did not report, including autocrine and chemokine-based patterns. These are putative and were not tested experimentally here.

   Germline Female: autocrine FGC CaH(3,4,14) with WNT5A/WNT6 and growth-factor genes (NGF/IGF/FGFR/VEGF), and FGC-Soma CaH(1,4,22) pairing neuropeptide ligands with mostly G protein-coupled receptors. Melanoma: B cell to T cell antigen presentation via HLA genes and CD8 as CaH(2,2,6). NonMyocyte: macrophage autocrine CaH(1,2,17) and NK cell-macrophage chemokine CaH(3,2,21). · Biological interpretation of real datasets

4. SuggestedIn zebrafish larval habenular neurons, scTensor paired neuron subtypes whose RNA-FISH positions in the original study fit the detected pairings. It also proposed dorsal partners for a subtype that RNA-FISH could not place.

   CaH(3,3,4) (La Hb01/03/07 with La Hb02/08) and CaH(2,1,3) (La Hb09 with La Hb02/08) pair cell types that RNA-FISH places dorsally and close together; CaH(1,3,1) spans the whole habenula. La Hb03, unresolved by RNA-FISH in the original study, was paired with La Hb02/08 in the dorsal region. · Application to minor organism

5. ShownThe method is released as R packages: nnTensor (CRAN) for non-negative Tucker decomposition, scTensor (Bioconductor) for annotated HTML reports, and LRBase.XXX.eg.db L-R databases for 12 organisms. The databases are built by LRBaseDbi, which also accepts user-supplied L-R lists.

   LRBase packages use a pipeline almost identical to FANTOM5's for putative L-R lists and store NCBI Gene ID pairs read from CSV files. Reports add annotation and enrichment via biomaRt, reactome.db, GOstats, meshr, ReactomePA and DOSE, plus links to RefEx, Expression Atlas, PanglaoDB and Connectivity Map. Analyses of the five empirical datasets are posted in CellCelldb. · nnTensor and scTensor packages; LRBase.XXX.eg.db-type packages; LRBaseDbi package; CellCelldb