Previously: binGetSpatialGenes. BinSpect (Binary Spatial Extraction of genes) is a fast computational method that identifies genes with a spatially coherent expression pattern.
binSpect(
gobject,
bin_method = c("kmeans", "rank"),
expression_values = c("normalized", "scaled", "custom"),
subset_genes = NULL,
spatial_network_name = "Delaunay_network",
spatial_network_k = NULL,
reduce_network = FALSE,
kmeans_algo = c("kmeans", "kmeans_arma", "kmeans_arma_subset"),
nstart = 3,
iter_max = 10,
extreme_nr = 50,
sample_nr = 50,
percentage_rank = 30,
do_fisher_test = TRUE,
adjust_method = "fdr",
calc_hub = FALSE,
hub_min_int = 3,
get_av_expr = TRUE,
get_high_expr = TRUE,
implementation = c("data.table", "simple", "matrix"),
group_size = "automatic",
do_parallel = TRUE,
cores = NA,
verbose = T,
knn_params = NULL,
set.seed = NULL,
bin_matrix = NULL,
summarize = c("p.value", "adj.p.value"),
return_gobject = F
)giotto object
method to binarize gene expression
expression values to use
only select a subset of genes to test
name of spatial network to use (default = 'spatial_network')
different k's for a spatial kNN to evaluate
default uses the full network
kmeans algorithm to use (kmeans, kmeans_arma, kmeans_arma_subset)
kmeans: nstart parameter
kmeans: iter.max parameter
number of top and bottom cells (see details)
total number of cells to sample (see details)
percentage of top cells for binarization
perform fisher test
p-value adjusted method to use (see p.adjust)
calculate the number of hub cells
minimum number of cell-cell interactions for a hub cell
calculate the average expression per gene of the high expressing cells
calculate the number of high expressing cells per gene
enrichment implementation (data.table, simple, matrix)
number of genes to process together with data.table implementation (default = automatic)
run calculations in parallel with mclapply
number of cores to use if do_parallel = TRUE
be verbose
list of parameters to create spatial kNN network
set a seed before kmeans binarization
a binarized matrix, when provided it will skip the binarization process
summarize the p-values or adjusted p-values
data.table with results (see details)
We provide two ways to identify spatial genes based on gene expression binarization. Both methods are identicial except for how binarization is performed.
1. binarize: Each gene is binarized (0 or 1) in each cell with kmeans (k = 2) or based on rank percentile
2. network: Alll cells are connected through a spatial network based on the physical coordinates
3. contingency table: A contingency table is calculated based on all edges of neighboring cells and the binarized expression (0-0, 0-1, 1-0 or 1-1)
4. For each gene an odds-ratio (OR) and fisher.test (optional) is calculated
Three different kmeans algorithmes have been implemented:
1. kmeans: default, see kmeans
2. kmeans_arma: from ClusterR, see KMeans_arma
3. kmeans_arma_subst: from ClusterR, see KMeans_arma,
but random subsetting the vector for each gene to increase speed. Change extreme_nr and sample_nr for control.
Other statistics are provided (optional):
Number of cells with high expression (binary = 1)
Average expression of each gene within high expressing cells
Number of hub cells, these are high expressing cells that have a user defined number of high expressing neighbors
By selecting a subset of likely spatial genes (e.g. soft thresholding highly variable genes) can accelerate the speed. The simple implementation is usually faster, but lacks the possibility to run in parallel and to calculate hub cells. The data.table implementation might be more appropriate for large datasets by setting the group_size (number of genes) parameter to divide the workload.