spatial_ccc_experiments

This repository contains experiments with performing spatial ccc. Our current aims are to:

• List a few candidate datasets
• Decide on a common input format
• Decide on a common output format
• Implement a selection of spatial CCC methods
• Run all methods on all datasets
• Compare outputs (visually, or with a metric)

Candidate datasets

Input format

Related issue: #2

AnnData object
 obs: 'celltype'
 obsm: 'spatial'
 layers: 'counts'

How to read in Python:

import anndata as ad

adata = ad.read_h5ad('path/to/file.h5ad')

# Access celltype
adata.obs['celltype']

# Access spatial coordinates
adata.obsm['spatial']

# Access counts
adata.layers['counts']

How to read in R:

library(anndata)

adata <- read_h5ad('path/to/file.h5ad')

# Access celltype
adata$obs[["celltype"]]

# Access spatial coordinates
adata$obsm[["spatial"]]

# Access counts
adata$layers[["counts"]]

Output format

Related issue: #1

Spatial CCC methods typically output one of three formats.

Per-spot prediction scores

Examples: SpatialDM, LIANA+, NICHES

Per slide

That is, one summary of interactions for the whole slide / sample

Examples: COMMOT / MISTy / LIANA+

Per source-target

That is, a summary for the whole slide but with respect to the cell type annotations

Examples: Giotto / CellChatv2

Implement a selection of spatial CCC methods

The goal of the group was to run multiple spatial CCC methods, compare evaluations/visualizations and results. We selected the methods from Armingol et al., 2024.

Tool	Data	Input	Output	Synthetic data	Experimental	Visualizations	Benchmarked against	Comments
COMMOT	Mouse MERFISH, mouse placenta & brain STARMap, mouse hippocampus Slide-seqV2, Visium	LR pairs from CellChat DB (Heteromeric/homomeric binding) or manually defined genes, Log1p counts	Sparse matrix (in adata.obsp) with interaction scores for cell to cell	LR binding	Immunostaining of proteins; RNAScope imaging	Spatial signaling direction between spots/cells, heatmaps of DGE in CCC	CellChat, Giotto, CellPhoneDB v.3	Got output; Easy to set up and run, but took a long time to run; Weird output because every interaction is in separate files
stMLNet	Breast cancer ST, seqFISH+, MERFISH, Slide-seq v2, Visium glioma & COVID-19	Counts, location, and cell type annotations; method only provides human DB	Ligand-receptors scores for each sender-receiver-target gene combination	Data following ligand diffusion model	N/A	Sankey plots, CCC network, chord diagram	NicheNet, CytoTalk, MISTy	Ran on the example dataset from their documentation; Tried running on chosen dataset, but the preprocessing is quite slow; Returns folders per cell type pair
SpatialDM	Visium intestine, melanoma, seqFISH subventricular zone	CellChat DB	SVCA (Gaussian process model) simulated ST data	Chord diagram, spatial plots for interaction pattern clustering	CellChat, Giotto, SpaTalk	Ran the global part; The local part, but it's currently error	Apparently LIANA+ already has SpatialDM included so we're running this – ran through and much faster, will check the results	Initial implementation is pretty slow
SpaTalk	Many untargeted and targeted datasets	Counts, location, cell type annotation	Prioritized ligand-receptors for each sender-receiver pair; enriched pathways	Simulated spots for deconvolution	N/A	Chord diagrams, snakey plots, heatmaps	Giotto, SpaOTsc, NicheNet, CytoTalk, CellCall, CellPhoneDB, and CellChat	No Seurat v5 compatibility
HoloNet	Breast cancer & liver cancer Visium, Slide-seq v2	Strategy of Tang et al., generated spatial patterning of two cell types	N/A	CCC network overlaid on spatial data	NicheNet, SpaTalk	N/A	N/A	N/A
CellNeighborEX	seqFISH mouse embryo, slide-seq from mouse embryo, mouse liver, mouse brain	Artificial heterotypic spots	Cell isolation and sorting with specific markers	Spatial visualization of cell neighbor-dependent gene expression	N/A	N/A	N/A	N/A
LIANA+ (cellphonedb)	Visium, Slide-tags, Spatial metabolite-transcriptome (MALDI-MSI x Visium)	-	Cell type specific score of communication events (lr_means)	Intracellular signalling network, ugly dotplots, heatmaps	Slide-tags spatially-informed vs not, classification of malignant spots, classification of conditions	Received dotplots and heatmaps for interactions cell-type specific, not spot specific	Need to investigate the resulting matrix, didn't yet find the specific ligand-receptor pairs
MEBOCOST	scRNAseq	Gene expression of enzymes and knowledge repository of enzyme-metabolite-sensor partners (scFEA, COMPASS, HMDB, literature mining)	Number of communication events and score (sum of -log10(FDR)) of detected metabolite-sensor communications between a pair of cell types	-	-	-	-	Got metabolite-abundance estimation, is not comparable to other outputs
NicheCompass	seqFISH mouse organogenesis; Slide-seq mouse hippocampus	-	Circular plot	CellCharter; STACI; GraphST	Not very easy to run, because we encountered a bug	Not very responsive on GitHub	N/A
Giotto	seqFISH+, merFISH, osmFISH, STARmap, Visium 10X brain & kidney, Slide-seq, t-cyCIF, MIBI, CODEX	a gene-by-cell count matrix and the spatial coordinates for the centroid position of each cell	score based on adjusted p-value and log2 fold change was used to rank a ligand-receptor pair across all identified cells of interacting cell types	Simulated seqFISH+ and spatial pattern	N/A	All types of plots, GiottoVisuals, GiottoViewer	N/A	Not very easy to run; Not user friendly; Working running the method
CellChat v2	Visium 10X mouse brain and human fetal intestine	Normalized and log-transformed gene-by-cell matrix, spatial coordinates,Metadata (cell type labels and sample names), CellChatDB	Ligand, receptor, source, target, interaction and their probabilities, pathways	N/A	N/A	Circular plot, heatmap, spatial plot	N/A	Result is a single dataframe with predicted interactions and corresponing probabilities; Easy to visualize; Need to try different values for scaling (pairwise) distances (between regions/cells) while computing cell communication probabilities for different spatial technologies

Run all methods on all datasets

Compare outputs

Either visually or with a metric