spatialDWLS

Cell type deconvolution of spatial transcriptomics spots with dampened weighted least squares.

spatialDWLS estimates what fraction of each spot is made up of each cell type, given a mean-expression reference built from single-cell data. Each spot's result is a composition: the proportions of a solved spot sum to one.

The method (Dong and Yuan 2021) runs in two rounds:

Cluster round. For each cluster of spots, cell types enriched in it (PAGE z-score above cutoff) become candidates, and every spot is solved against them.
Spot round. Per spot, cell types whose round-1 proportion is below 1 / n_cell are dropped and the spot is solved again against the rest.

Each solve is DWLS (Tsoucas et al. 2019): non-negative weighted least squares whose weights are iterated and capped by a cross-validated dampening constant.

This package is the standalone engine. It works on plain matrices, so it has no dependency on Giotto; Giotto's runDWLSDeconv() calls it.

Website: https://drieslab.github.io/spatialDWLS/, with two worked examples: plain matrices and through Giotto.

Installation

# install.packages("remotes")
# remotes::install_github("drieslab/spatialDWLS")
install.packages("spatialDWLS", repos = c('https://drieslab.r-universe.dev', 'https://cloud.r-project.org'))

Usage

library(spatialDWLS)

# reference: mean normalised (not logged) expression per cell type
sig <- dwls_signature(sc_norm, sign_gene = markers, cell_type = sc_labels)

# spots: log-normalised expression, features x spots; one cluster per spot
prop <- dwls(spot_lognorm, sig, clusters = spot_clusters, n_cell = 10)

prop is a cell types x spots matrix. Spots with no signal on the signature features cannot be solved and are NA (na_as_zero = TRUE returns 0).

Sparse input stays sparse; only the signature features are densified, one cluster at a time.
Parallel: pass a BiocParallel param, e.g. BPPARAM = BiocParallel::MulticoreParam(8). Results are identical to the serial run.
Larger than memory: dwls_stream() takes a read_window(i, features) callback and reads spots one window at a time. This is how Giotto's disk-backed stores (GiottoDisk) run it.
Reproducible: the dampening constant's cross-validation is seeded internally; your session's RNG state is left untouched.

Citation

Dong R, Yuan GC (2021). SpatialDWLS: accurate deconvolution of spatial transcriptomic data. Genome Biology 22, 145. https://doi.org/10.1186/s13059-021-02362-7

Tsoucas D, Dong R, Chen H, Zhu Q, Guo G, Yuan GC (2019). Accurate estimation of cell-type composition from gene expression data. Nature Communications 10, 2975. https://doi.org/10.1038/s41467-019-10802-z