GAM-clustering Tutorial Notebook For Spatial RNA-Seq Data
2023/10/05
Required packages to be installed:
if (!require("devtools", quietly = TRUE))
install.packages("BiocManager")
devtools::install_github("ctlab/GAMclust")
devtools::install_github("ctlab/gatom")
devtools::install_github("ctlab/mwcsr")
devtools::install_github("ctlab/fgsea")library(GAMclust)
library(gatom)
library(mwcsr)
library(fgsea)
library(data.table)
library(Seurat)
set.seed(42)Preparing working environment
First, please load and initialize all objects required for GAM-clustering analysis:
- load KEGG metabolic
network
network.kegg.rdsand its metabolites annotationmet.kegg.db.rds;
network <- readRDS(url("http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/network.kegg.rds"))
metabolites.annotation <- readRDS(url("http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/met.kegg.db.rds"))- load species-specific network annotation:
org.Hs.eg.gatom.annofor human data ororg.Mm.eg.gatom.annofor mouse data;
network.annotation <- readRDS(url("http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/org.Hs.eg.gatom.anno.rds"))- load provided list of metabolites that should not be considered during the analysis as connections between reactions (e.g., CO2, HCO3-, etc);
met.to.filter <- data.table::fread(system.file("mets2mask.lst", package="GAMclust"))$ID- initialize SMGWCS solver:
(4.1.) we recommend to use here either heuristic relax-and-cut solver
rnc_solver from mwcsr
package,
solver <- mwcsr::rnc_solver()(4.2.) either proprietary CPLEX solver (free for academy);
cplex.dir <- "/opt/ibm/ILOG/CPLEX_Studio1271"
solver <- mwcsr::virgo_solver(cplex_dir = cplex.dir)- set working directory where the results will be saved to.
work.dir <- "results"
dir.create(work.dir, showWarnings = F, recursive = T)- TEMPORARY: collecting logs while developing the tool.
stats.dir <- paste0(work.dir, "/stats")
dir.create(stats.dir, showWarnings = F, recursive = T)
log_file <- file(paste0(stats.dir, "/log.txt"), open = "wt")
sink(log_file, type = "output", append = TRUE)
sink(log_file, type = "message", append = TRUE)Preparing objects for the analysis
Preparing data
GAMclust works with bulk, single cell and spatial RNA-seq data.
This vignette shows how to process spatial RNA-seq data.
Let’s load the data and take 10,000 genes for the GAM-clustering analysis.
seurat_object <- readRDS(url("http://artyomovlab.wustl.edu/publications/supp_materials/GAMclust/275_T_seurat.rds"))
seurat_object <- Seurat::SCTransform(seurat_object,
assay = "Spatial",
variable.features.n = 10000,
verbose = FALSE)
E <- as.matrix(Seurat::GetAssayData(object = seurat_object,
assay = "SCT",
slot = "scale.data"))
nrow(E) # ! make sure this value is in range from 6,000 to 12,000# [1] 10000E[1:3, 1:3]# AAACAAGTATCTCCCA-1 AAACACCAATAACTGC-1 AAACAGAGCGACTCCT-1
# NOC2L 1.2429330 0.8244317 3.6945664
# ISG15 2.6233759 -0.7541004 0.5605173
# AGRN -0.6149024 -0.7259493 0.5248512Genes in your dataset may be named as Symbol, Entrez, Ensembl or
RefSeq IDs. One of these names should be specified as value of
gene.id.type parameter in prepareData().
If you analyse singe cell or spatial RNA-seq data, please set
use.PCA=TRUE in prepareData().
E.prep <- prepareData(E = E,
gene.id.type = "Symbol",
use.PCA = TRUE,
use.PCA.n = 50,
network.annotation = network.annotation)
E.prep[1:3, 1:3]# PC1 PC2 PC3
# 885 4.977872 -3.7994982 0.8635909
# 4504 -6.323743 -2.5276675 0.5063040
# 9569 -1.237742 -0.6967359 -0.3546444Preparing network
The prepareNetwork() function defines the structure of
the final metabolic modules.
network.prep <- prepareNetwork(E = E.prep,
network = network,
met.to.filter = met.to.filter,
network.annotation = network.annotation)# > Global metabolite network contains 4214 edges.# > Largest connected component of this global network contains 1062 nodes and 3516 edges.Preclustering
The preClustering() function defines initial patterns
using k-medoids clustering on gene expression matrix. It is strongly
recommended to do initial clustering with no less than 32 clusters
(initial.number.of.clusters = 32).
You can visualize the initial heatmap as shown below.
cur.centers <- preClustering(E.prep = E.prep,
network.prep = network.prep,
initial.number.of.clusters = 32,
network.annotation = network.annotation)# > 836 metabolic genes from the analysed dataset mapped to this component.cur.centers[1:3, 1:3]# PC1 PC2 PC3
# 1 2.2339275 1.394434 -0.9111931
# 2 -0.4371787 1.349472 1.7452293
# 3 1.0833861 1.046135 0.6386994pheatmap::pheatmap(
GAMclust:::normalize.rows(cur.centers),
cluster_rows=F, cluster_cols=F,
show_rownames=F, show_colnames=T)
GAM-clustering analysis
Now you have everything prepared for the GAM-clustering analysis.
Initial patterns will be now refined in an iterative process. The
output of gamClustering() function presents a set of
specific subnetworks (also called metabolic modules) that reflect
metabolic variability within a given transcriptional dataset.
Note, that it may take a long time to derive metabolic modules by
gamClustering() function (tens of minutes).
There is a set of parameters which determine the size and number of your final modules. We recommend you to start with the default settings, however you can adjust them based on your own preferences:
If you consider final modules to bee too small or too big and it complicates interpretation for you, you can either increase or reduce by 10 units the
max.module.sizeparameter.If among final modules you consider presence of any modules with too similar patterns, you can reduce by 0.1 units the
cor.thresholdparameter.If among final modules you consider presence of any uninformative modules, you can reduce by 10 times the
p.adj.val.thresholdparameter.
results <- gamClustering(E.prep = E.prep,
network.prep = network.prep,
cur.centers = cur.centers,
start.base = 0.5,
base.dec = 0.05,
max.module.size = 50,
cor.threshold = 0.6,
p.adj.val.threshold = 1e-5,
batch.solver = seq_batch_solver(solver),
work.dir = work.dir,
show.intermediate.clustering = FALSE,
verbose = FALSE,
collect.stats = TRUE)Visualizing and exploring the GAM-clustering results
Each metabolic module is a connected piece of metabolic network whose genes expression is correlated across all dataset.
The following functions will help you to visualize and explore the obtained results.
Get graphs of modules:
getGraphs(modules = results$modules,
network.annotation = network.annotation,
metabolites.annotation = metabolites.annotation,
seed.for.layout = 42,
work.dir = work.dir)# Graphs for module 1 are built# Graphs for module 2 are built# Graphs for module 3 are built# Graphs for module 4 are builtExample of the graph of the fourth module:
Get gene tables:
The table contains gene list. Each gene has two descriptive values: i) gene’s correlation value with the modules pattern and ii) gene’s score. High score means that this gene’s expression is similar to the module’s pattern and not similar to other modules’ patterns.
m.gene.list <- getGeneTables(modules = results$modules,
nets = results$nets,
patterns = results$patterns.pos,
gene.exprs = E.prep,
network.annotation = network.annotation,
work.dir = work.dir)# Gene tables for module 1 are produced# Gene tables for module 2 are produced# Gene tables for module 3 are produced# Gene tables for module 4 are producedExample of the gene table of the fourth module:
head(GAMclust:::read.tsv(paste0(work.dir, "/m.4.genes.tsv")))Get plots of patterns:
for(i in 1:length(m.gene.list)){
plot(fgsea::plotCoregulationProfileSpatial(m.gene.list[[i]],
seurat_object,
title = paste0("module ", i)))
}
Get plots of individual genes expression (example for fourth module):
Seurat::DefaultAssay(seurat_object) <- "SCT"
i <- 4
Seurat::SpatialFeaturePlot(seurat_object,
features = m.gene.list[[i]],
ncol = 8)