/network_collection

A collection of gene regulatory networks, mostly but not entirely from chromatin data & ChIP or motif analysis.

Primary LanguageR

This is a collection of pre-built gene regulatory networks, accompanied by the code used to acquire and clean the data. This part of our benchmarking project.

Using the networks

For the Python API, consult the companion package.

There is also an R API included as a script in this folder. Usage:

source("R/load_networks.R")
# Set this to point to the "networks" folder adjacent to this README. 
options(GRN_PATH = "networks")
# What networks are available?
View(load_grn_metadata())
# What tissues do they cover, or how many?
list_subnetworks("gtex_rna") %>% head
lapply(load_grn_metadata()[[1]], list_subnetworks) %>% sapply(length)
# Show me the edges for a tissue. 
load_grn_by_subnetwork("gtex_rna", "Adipose_Subcutaneous.txt.gz") %>% head
# Show me the edges for all tissues.
iterate_within_grn("gtex_rna", load_grn_by_subnetwork) %>% sapply(dim)
# Count the edges for all tissues for all networks. Takes a while to run.
# iterate_over_grns(load_grn_by_subnetwork) %>% sapply(nrow)

Installation

The networks themselves are too big to put on GitHub. Download them from Zenodo (DOI: 10.5281/zenodo.10363068).

Storage format

Metadata are stored in networks/published_networks.csv. networks also contains a set of published GRN's stored uniformly. Each network is stored as <source_name>/networks/<subnetwork_name>.csv.gz. (UPDATE: now .parquet.) The file format has three columns called 'regulator', 'target', and 'weight'. An edge weight of -1 means the network is unweighted.

regulator,target,weight
Pou5f1,Sox2,1
...

There is an optional fourth column, cell_type.

Adding new networks

To add a new collection of networks, add a row to networks/published_networks.csv and save three-column Parquet files in the above format. For examples, look in the setup folder. To test your network, try out the loader commands for the R or python API's mentioned above. You should be able to load the networks into R or Python and query them to verify any individual edge.

Source data

Source URL's, citations, and descriptions are given in networks/published_networks.csv. The setup is not fully automated, but large parts of it are, and scripts are given in the setup folder (start with main.R). A few notes on specific networks:

  • For CellOracle, I installed the package, which ships with a default base network. Then there's a setup script that does the formatting and gzipping. No point-and-click downloads needed.

  • Networks from FNTM and Humanbase were subsetted to retain only edges with posterior probability over 50%. See setup script; no point-and-click downloads needed.

  • The least straightforward setup was for the CellNet networks, but it's also mostly automated. Starting from the r package page, I copied these files:

    cnProc_Hg1332_062414.R
cnProc_mogene_062414.R
cnProc_Hugene_062414.R
cnProc_mouse4302_062414.R

    I installed cellnetr using this tip to solve the affy version problem. I then read the data and exported GRN's using cellnet.R in the setup folder.

Layout

  • networks: Networks stored in triplet format. This is the only folder accessed by our benchmarking project.
  • not_ready: Datasets that we may in the future process into triplet format, or raw downloads that we did already process.
  • setup: code we used to assemble and format this collection.

Networks that we may ingest in the future