This tutorial describes an R-package for finding active metabolic modules based on high throughput data. The pipeline takes as input transcriptional and/or metabolic data and finds a metabolic subnetwork (module) most regulated between the two conditions of interest.

The package relies on the active module analysis framework developed in BioNet package, but extends it to work with metabolic reaction networks. Further, it illustrates the usage of mwcsr package which provides a number of solvers for Maximum Weight Connected Subgraph problem and its variants.

Example of using the pipeline include:

More details on the pipeline are available in Sergushichev et al, 2016 and Emelianova et al, 2022.

Installation

You can install gatom via BiocManager: ```{R eval = FALSE}

if (!require("BiocManager", quietly = TRUE))

install.packages("BiocManager")

BiocManager::install("gatom")

# Example workfow

In this example we will find an active metabolic module based on macrophage 
activation gene expression and metabolomics data 
([Jha et al, 2015](http://dx.doi.org/10.1016/j.immuni.2015.02.005)).
For improved performance here we will consider a simplified version of the data.
See [Example on full data and full network](#example-full) section for the real-scale analysis.


```r
library(gatom)
library(data.table)
library(igraph)
library(mwcsr)

First let's load the example tables with input differential gene expression and metabolite abundance data for LPS-IFNg stimulated macrophages compared to controls:

data("gene.de.rawEx")
print(head(gene.de.rawEx))
data("met.de.rawEx")
print(head(met.de.rawEx))

Next we will load example network related objects: global reaction network (networkEx object), metabolite annotations (met.kegg.dbEx), and organism-specific enzyme annotations for mouse (org.Mm.eg.gatom.annoEx).

data("networkEx")
data("met.kegg.dbEx")
data("org.Mm.eg.gatom.annoEx")

Here networkEx object contain information about r nrow(networkEx$reactions) KEGG reactions, their atom mappings and relation to enzymes:

str(networkEx, max.level=1, give.attr = FALSE)

Object met.kegg.dbEx contains information about r nrow(met.kegg.dbEx$metabolites) KEGG metabolites, including mappings from HMDB and ChEBI:

str(met.kegg.dbEx, max.level=2, give.attr = FALSE)

Object org.Mm.eg.gatom.annoEx contains mouse-specific mapping between enzyme classes and genes, as well as mapping between different types of gene identifiers:

str(org.Mm.eg.gatom.annoEx, max.level=2, give.attr = FALSE)

Then we create a metabolic graph with atom topology from the loaded data. Depending on topology parameter, the graph vertices can correspond either to atoms or metabolites. For metabolite topology, see Using metabolite-level network section.

g <- makeMetabolicGraph(network=networkEx, 
                        topology="atoms",
                        org.gatom.anno=org.Mm.eg.gatom.annoEx, 
                        gene.de=gene.de.rawEx,
                        met.db=met.kegg.dbEx, 
                        met.de=met.de.rawEx)
print(g)

After creating the metabolic graph, we then score it, obtaining an instance of Signal Generalized Maximum Weight Subgraph (SGMWCS) problem.

The size of the module can be controlled by changing scoring parameters k.gene and k.met. The higher the values of scoring parameters are, the bigger the module is going to be.

gs <- scoreGraph(g, k.gene = 25, k.met = 25)

Then we initialize an SMGWCS solver. Here, we use a heuristic relax-and-cut solver rnc_solver for simplicity.

See mwcsr package documentation for more solver options, or Using exact solver section for the recommended way.

solver <- rnc_solver()

Then we find an active metabolic module with chosen solver and scored graph:

set.seed(42)
res <- solve_mwcsp(solver, gs)
m <- res$graph

The result module is an igraph object that captures the most regulated reactions:

print(m)
head(E(m)$label)
head(V(m)$label)

The module can be plotted in R Viewer with createShinyCyJSWidget(). Here, red color corresponds to up-regulation (positive log-2 fold change) and green to down-regulation (negative log-2 fold change). Blue nodes and edges come from data with absent log-2 fold change values. Bigger size of nodes and width of edges reflect lower p-values.

createShinyCyJSWidget(m)

Saving modules

We can save the module to graphml format with write_graph() function from igraph:

write_graph(m, file = file.path(tempdir(), "M0.vs.M1.graphml"), format = "graphml")

Or it can be saved to an interactive html widget:

saveModuleToHtml(module = m, file = file.path(tempdir(), "M0.vs.M1.html"), 
                 name="M0.vs.M1")

We can also save the module to dot format:

saveModuleToDot(m, file = file.path(tempdir(), "M0.vs.M1.dot"), name = "M0.vs.M1")

Such dot file can be further used to generate svg file using neato tool from graphviz suite if it is installed on the system:

system(paste0("neato -Tsvg ", file.path(tempdir(), "M0.vs.M1.dot"),
              " > ", file.path(tempdir(), "M0.vs.M1.svg")), 
       ignore.stderr=TRUE)

Alternatively, the module can be saved to pdf format with a nice layout.

You may vary the meaning of repel force and the number of iterations of repel algorithm for label layout. Note, that the larger your graph is the softer force you should use.

You may also set different seed for different variants of edge layout with set.seed().

set.seed(42)
saveModuleToPdf(m, file = file.path(tempdir(), "M0.vs.M1.pdf"), name = "M0.vs.M1", 
                n_iter=100, force=1e-5)

Example on full data and full network {#example-full}

Let's now look at how the analysis will work with the full dataset and the full network. For this case we will be using the combined network instead of KEGG network (see Networks for details on the network types).

library(R.utils)
library(data.table)

The full macrophage LPS+IFNG-activation dataset can be downloaded from http://artyomovlab.wustl.edu/publications/supp_materials/GAM/:

met.de.raw <- fread("http://artyomovlab.wustl.edu/publications/supp_materials/GAM/Ctrl.vs.MandLPSandIFNg.met.de.tsv.gz")
gene.de.raw <- fread("http://artyomovlab.wustl.edu/publications/supp_materials/GAM/Ctrl.vs.MandLPSandIFNg.gene.de.tsv.gz")

Full pre-generated combined network, corresponding metabolite annotation, and enzyme annotation can be downloaded from http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/:

network.combined <- readRDS(url("http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/network.combined.rds"))
met.combined.db <- readRDS(url("http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/met.combined.db.rds"))

org.Mm.eg.gatom.anno <- readRDS(url("http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/org.Mm.eg.gatom.anno.rds"))

For better work of the combined network we highly recommend using additional supplementary gene files (see Supplementary Genes).

gene2reaction.extra <- fread("http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/gene2reaction.combined.mmu.eg.tsv", colClasses="character")

Running gatom on the combined network and the full dataset:

cg <- makeMetabolicGraph(network=network.combined,
                         topology="atoms",
                         org.gatom.anno=org.Mm.eg.gatom.anno,
                         gene.de=gene.de.raw,
                         met.db=met.combined.db,
                         met.de=met.de.raw,
                         gene2reaction.extra=gene2reaction.extra)

cgs <- scoreGraph(cg, k.gene = 50, k.met = 50)

solver <- rnc_solver()
set.seed(42)
sol <- solve_mwcsp(solver, cgs)
cm <- sol$graph
cm

The result module for combined network:

createShinyCyJSWidget(cm)

Networks {#networks}

We provide four types of networks that can be used for analysis:

  1. KEGG network
  2. Rhea network
  3. Combined network
  4. Rhea lipid subnetwork

KEGG {#kegg-network}

KEGG network consists of network.kegg.rds & met.kegg.db.rds files and is based on KEGG database.

Both metabolites and reactions in KEGG network have KEGG IDs.

This network was generated with the pipeline available here. For extra details on KEGG network you can also reference shinyGatom and GAM articles.

network <- readRDS(url("http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/network.kegg.rds"))
met.db <- readRDS(url("http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/met.kegg.db.rds"))

Running gatom with KEGG network on full dataset:

kg <- makeMetabolicGraph(network=network, 
                        topology="atoms",
                        org.gatom.anno=org.Mm.eg.gatom.anno, 
                        gene.de=gene.de.raw,
                        met.db=met.db, 
                        met.de=met.de.raw)

kgs <- scoreGraph(kg, k.gene = 50, k.met = 50)

solver <- rnc_solver()
set.seed(42)
sol <- solve_mwcsp(solver, kgs) 
km <- sol$graph
km
createShinyCyJSWidget(km)

Rhea

Rhea network consists of network.rhea.rds & met.rhea.db.rds files and is based on Rhea database.

Reactions in Rhea have their own IDs, but unlike KEGG, metabolite IDs come from a separate database -- ChEBI database.

This network was generated with the pipeline available here. For extra details on Rhea network you can also reference shinyGatom article.

To use Rhea network download the following files:

network.rhea <- readRDS(url("http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/network.rhea.rds"))
met.rhea.db <- readRDS(url("http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/met.rhea.db.rds"))

For proper work of the Rhea network we also need a corresponding supplementary gene file (ref. Supplementary Genes).

gene2reaction.extra <- fread("http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/gene2reaction.rhea.mmu.eg.tsv", colClasses="character")

And run gatom on Rhea network:

rg <- makeMetabolicGraph(network=network.rhea,
                         topology="atoms",
                         org.gatom.anno=org.Mm.eg.gatom.anno,
                         gene.de=gene.de.raw,
                         met.db=met.rhea.db,
                         met.de=met.de.raw,
                         gene2reaction.extra=gene2reaction.extra)

rgs <- scoreGraph(rg, k.gene = 50, k.met = 50)

solver <- rnc_solver()
set.seed(42)
sol <- solve_mwcsp(solver, rgs)
rm <- sol$graph
rm

Result Rhea network module:

createShinyCyJSWidget(rm)

Combined network

Combined network comprises not only KEGG and Rhea reactions, but also transport reactions from BIGG database.

This means that reactions in such network have either KEGG or Rhea or BIGG IDs, and metabolite IDs are KEGGs and ChEBIs.

Rhea lipid subnetwork

Rhea lipid subnetwork is subset of Rhea reactions that involve lipids, and it consists of network.rhea.lipids.rds & met.rhea.lipids.db.rds files.

This network was generated with the pipeline available here. For extra details on Rhea lipid subnetwork you can also reference shinyGatom article.

network.lipids <- readRDS(url("http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/network.rhea.lipids.rds"))
met.lipids.db <- readRDS(url("http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/met.lipids.db.rds"))

For proper work of the lipid network we will also need a corresponding supplementary gene file (ref. Supplementary Genes)

gene2reaction.extra <- fread("http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/gene2reaction.rhea.mmu.eg.tsv", colClasses="character")

To test lipid network we will use example lipidomics data for WT mice control vs high fat diet comparison from ST001289 dataset.

met.de.lipids <- fread("https://artyomovlab.wustl.edu/publications/supp_materials/GATOM/Ctrl.vs.HighFat.lipid.de.csv")

For lipid network we recommend setting topology to metabolites (ref. Using metabolite-level network):

lg <- makeMetabolicGraph(network=network.lipids,
                         topology="metabolites",
                         org.gatom.anno=org.Mm.eg.gatom.anno,
                         gene.de=NULL,
                         met.db=met.lipids.db,
                         met.de=met.de.lipids,
                         gene2reaction.extra=gene2reaction.extra)

lgs <- scoreGraph(lg, k.gene = NULL, k.met = 50)

solver <- rnc_solver()
set.seed(42)
sol <- solve_mwcsp(solver, lgs)
lm <- sol$graph
lm

Result lipid subnetwork module:

createShinyCyJSWidget(lm)

If IDs for metabolite differential abundance data are of type Species we can process metabolite labels into more readable ones:

lm1 <- abbreviateLabels(lm, orig.names = TRUE, abbrev.names = TRUE)

createShinyCyJSWidget(lm1)

Misc

Supplementary gene files {#suppl-genes}

For combined, Rhea and lipid networks we provide supplementary files with genes that either come from proteome or are not linked to a specific enzyme. These files are organism-specific and are also available at http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/.

network.combined <- readRDS(url("http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/network.combined.rds"))
met.combined.db <- readRDS(url("http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/met.combined.db.rds"))
gene2reaction.extra <- fread("http://artyomovlab.wustl.edu/publications/supp_materials/GATOM/gene2reaction.combined.mmu.eg.tsv", colClasses="character")

gg <- makeMetabolicGraph(network=network.combined, 
                         topology="atoms",
                         org.gatom.anno=org.Mm.eg.gatom.anno, 
                         gene.de=gene.de.raw,
                         met.db=met.combined.db, 
                         met.de=met.de.raw, 
                         gene2reaction.extra=gene2reaction.extra)
gg

Non-enzymatic reactions

Optionally, we can also preserve non-enzymatic reactions that are found in the network. This can be done by setting keepReactionsWithoutEnzymes to TRUE:

ge <- makeMetabolicGraph(network=network.combined, 
                         topology="atoms",
                         org.gatom.anno=org.Mm.eg.gatom.anno, 
                         gene.de=gene.de.raw,
                         met.db=met.combined.db, 
                         met.de=met.de.raw, 
                         gene2reaction.extra=gene2reaction.extra, 
                         keepReactionsWithoutEnzymes=TRUE)
ge

Using exact solver {#exact-solver}

For proper analysis quality we recommend to use exact SGMWCS solver virgo_solver(), which requires Java (version >= 11) and CPLEX (version >= 12.7) to be installed. If the requirements are met you can then find a module as following:

vsolver <- virgo_solver(cplex_dir=Sys.getenv("CPLEX_HOME"), 
                        threads=4, penalty=0.001, log=1)
sol <- solve_mwcsp(vsolver, gs) 
m <- sol$graph

Edge penalty option there is used to remove excessive redundancy in genes.

Running with no metabolite data

If there is no metabolite data in your experiment assign met.de and k.met to NULL:

g <- makeMetabolicGraph(network=networkEx, 
                        topology="atoms",
                        org.gatom.anno=org.Mm.eg.gatom.annoEx, 
                        gene.de=gene.de.rawEx,
                        met.db=met.kegg.dbEx, 
                        met.de=NULL)
gs <- scoreGraph(g, k.gene = 50, k.met = NULL)

Running with no gene data

If there is no gene data in your experiment assign gene.de and k.gene to NULL:

g <- makeMetabolicGraph(network=networkEx, 
                        topology="atoms",
                        org.gatom.anno=org.Mm.eg.gatom.annoEx, 
                        gene.de=NULL,
                        met.db=met.kegg.dbEx, 
                        met.de=met.de.rawEx)
gs <- scoreGraph(g, k.gene = NULL, k.met = 50)

Using metabolite-level network {#met-topology}

Sometimes it could make sense to work with metabolite-metabolite topology of the network, not atom-atom one. Such network is less structured, but contains more genes.

gm <- makeMetabolicGraph(network=network, 
                        topology="metabolite",
                        org.gatom.anno=org.Mm.eg.gatom.anno, 
                        gene.de=gene.de.raw,
                        met.db=met.db, 
                        met.de=met.de.raw)

gms <- scoreGraph(gm, k.gene = 50, k.met = 50)

solver <- rnc_solver()
set.seed(42)
sol <- solve_mwcsp(solver, gms) 
mm <- sol$graph
mm

Pathway annotation

To get functional annotation of obtained modules by KEGG and Reactome metabolic pathways, we can use hypergeometric test with fora() function from fgsea package.

foraRes <- fgsea::fora(pathways=org.Mm.eg.gatom.anno$pathways,
                       genes=E(km)$gene,
                       universe=unique(E(kg)$gene),
                       minSize=5)
foraRes[padj < 0.05]

Optionally, redundancy in pathways can be decreased with collapsePathwaysORA() function:

mainPathways <- fgsea::collapsePathwaysORA(
  foraRes[padj < 0.05],
  pathways=org.Mm.eg.gatom.anno$pathways,
  genes=E(km)$gene,
  universe=unique(E(kg)$gene))
foraRes[pathway %in% mainPathways$mainPathways]
sessionInfo()


ctlab/gatom documentation built on May 3, 2024, 3:44 p.m.