ARACNe3 is an implementation of ARACNe which presents computational improvements and theoretical changes to the recent ARACNe-AP implementation. Given a list of regulators and a gene expression profile, ARACNe3 is used to infer irreducibly dependent regulatory interactions and output a Gene Regulatory Network (GRN). The mainstream analysis generates many GRNs of subsamples of the profile (called "subnetworks") and then consolidates the subnetworks into a robust GRN. The consolidated output is a GRN whose edge strengths can be quantified by several statistical metrics.
Lachmann A, Giorgi FM, Lopez G, Califano A. ARACNe-AP: gene network reverse engineering through adaptive partitioning inference of mutual information. Bioinformatics. 2016 Jul 15;32(14):2233-5. doi: 10.1093/bioinformatics/btw216. Epub 2016 Apr 23.
Margolin AA, Nemenman I, Basso K, Wiggins C, Stolovitzky G, Dalla Favera R, Califano A. ARACNE: an algorithm for the reconstruction of gene regulatory networks in a mammalian cellular context. BMC Bioinformatics. 2006 Mar 20;7 Suppl 1:S7. doi: 10.1186/1471-2105-7-S1-S7
ARACNe3 supports multithreading using OpenMP. Here is one example of how you might download OpenMP libraries on the latest version of MacOS, with homebrew already installed.
brew install libomp # Install OpenMP
cmake is used to build ARACNe3 and simplify cross-platform compatibility. After installing cmake, follow the instructions below to build ARACNe3.
# Clone the repo
git clone https://github.com/califano-lab/ARACNe3
# Install dependencies
cd ARACNe3/
git submodule update --init --recursive
# Build executable
mkdir build
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build # Build ARACNe3
The executable should be built in ./build/src/app/ARACNe3_app_release. If you are having issues, try using cmake3 instead of cmake in the instructions above.
- Normalize a gene expression profile for sequencing depth in each sample (CPM, TPM, etc.).
- Run ARACNe3 according to command line instructions below.
ARACNe3 input files (see input file format below)
- A
tsvthat contains theG+1 x N+1normalized expression profile, with genes as rows and samples as columns. Note: the "+1" is for row names & column names, respectively. - List of regulators (e.g., transcription factors).
ARACNe3 outputs in the directory provided by the user (e.g., -o outputdir/). If ARACNe3 is run with --runid abc, within outputdir/ the subnetworks are in outputdir/subnets/, the consolidated network is the file outputdir/consolidated-net_abc.tsv, and log information for each subnetwork is in outputdir/subnets_log/. Log information for the overall ARACNe3 instance is the file outputdir/log_abc.txt.
The subnetworks directory outputdir/subnets/ contains a file for each subnetwork named subnet#_abc.tsv. Each subnetwork file describes significant interactions in three columns:
- The regulator.
- The target.
- The APMI (Mutual Information estimated by Adaptive Partitioning) of the pair, based on the subsampled gene expression profiles.
The consensus network outputdir/consolidated-net_abc.tsv is a file that consolidates subnetworks and summarizes each edge in five columns:
- The regulator.
- The target.
- The APMI of the pair, based on the full gene expression profiles.
- The Spearman's Rank Correlation Coefficient of the pair, based on the full gene expression profile.
- The p-value of the edge, based on how many subnetworks in which it was identified (see the manuscript for methodology).
-e is the expression file.
-r is the list of regulators (e.g., TFs).
-o is the output directory.
-x adjusts the stopping criteria when generating or using multiple subnetworks. By default (-x 1) specifies the fixed number of subnetworks to generate, in this case just 1.
--adaptive changes the meaning of -x to specify regulon occupancy, instead of number of subnetworks to generate. Regulon occupancy is defined as the minimum number of unique targets observed per regulator, when all subnetworks are consolidated into one (default: -x 30 if --adaptive is specified). Note that different subnetworks find different targets because of different subsamples, but in the consolidated network there are metrics for confidence in these targets.
--alpha is the cutoff for False Discovery Rate (FDR) pruning (default: --alpha 0.05). The FDR cutoff is the first pruning step, which rejects the null hypothesis for edges based on the Benjamini-Hochberg Procedure.
--seed fixes the seed for deterministic behavior (e.g.: --seed 9001).
--threads sets the number of threads to use (default: --threads 1).
--runid allows you to pass an identifier to replace defaultid in log_defaultid.txt. Does not affect each modulator's log, only the instance log (default: --runid defaultid).
--FWER tells ARACNe3 to prune by control of Family Wise Error Rate (FWER) alpha, instead of the default control for FDR.
--FPR tells ARACNe3 to prune by control of False Positive Rate (FPR) alpha (a.k.a., p-value), instead of the default control for FDR.
--subsample is the population percentage to subsample (--subsample 0.63212).
--noAlpha tells ARACNe3 to skip subnetwork pruning by FDR/FPR/FWER (same as: --alpha 1).
--noMaxEnt tells ARACNe3 to skip subnetwork pruning by Principle of Maximum Entropy, the second pruning step.
--noConsolidate tells ARACNe3 not to consolidate subnetworks, only keeping the final log, the log/ subdirectory, and all subnetworks generated in subnets/.
--consolidate tells ARACNe3 to skip generating subnetworks and consolidate existing subnetworks. An expression file and a list of regulators must still be provided with -e and -r, respectively. -o specifies the directory location of an ARACNe3 output. Finally, -x specifies how many subnetwork files to use in consolidate (default: -x 1). Note that output directory -o must contain the subdirectories subnets/ and log/ that follow the exact conventions as an ARACNe3 output (including numbering). Each subnetwork used must be mapped 1:1 with its log file because consolidation generates p-values for edges strictly based on parameters used during the subnetwork generation, which are stored in the log files.
Note: the examples have been written based on the provided test sets: test/exp_mat.txt (the normalized expression matrix) and test/regulators.txt (the list of regulators).
Example 1: generate one ARACNe3 subnetwork, subsampling $1-e^{-1}%$ of expression profiles, controlling for FDR < 0.05, using only one CPU core
./ARACNe3 -e test/exp_mat.txt -r test/regulators.txt -o test/output
Example 2: generate one ARACNe3 subnetwork with no pruning steps, no subsampling, and seed equal to 343, using 10 CPU cores
./ARACNe3 -e test/exp_mat.txt -r test/regulators.txt -o test/output --subsample 1.00 --noAlpha --noMaxEnt --seed 343 --threads 10
Example 3: generate one ARACNe3 subnetwork with all pruning steps, subsampling 33.3% of profiles, controlling for FDR < 0.01
./ARACNe3 -e test/exp_mat.txt -r test/regulators.txt -o test/output --subsample 0.333 --alpha 0.01
Example 4: generate thirty ARACNe3 subnetworks, subsampling $1-e^{-1}%$ of expression profiles, controlling for FDR < 0.05
./ARACNe3 -e test/exp_mat.txt -r test/regulators.txt -o test/output -x 30
Example 5: generate ARACNe3 subnetworks adaptively, until at least 50 targets are observed per regulon in the consensus network, controlling for FWER < 0.10, and skipping the MaxEnt pruning step
./ARACNe3 -e test/exp_mat.txt -r test/regulators.txt -o test/output -x 50 -FWER --alpha 0.10 --adaptive --noMaxEnt
A text file, containing one symbol per line. E.g.,
g_10_
g_10011_
g_1_
A G+1 x N+1, normalized expression profile, with genes on rows and samples on columns (Note: the +1 is extra, from the row names and column names). This should be in tsv format and have both row and column names. Row names are essential for ARACNe3 to store a gene's expression. Column names are used to count the number of samples. For example, the 5+1 x 3+1 matrix below is a compliant tsv:
col names NOT important
g_1_ 4.99 2.93 0.39
g_10_ 0.58 0.18 2.65
g_9432_ 3.00 1.27 7.63
g_10006_ 1.30 0.05 0.68
g_10011_ 0.055 0.73 4.64
Please contact Aaron Griffin (theory) or Andrew Howe (codebase) for questions regarding this project.
Aaron T. Griffin - atg2142@cumc.columbia.edu
Andrew R. Howe - arh2207@columbia.edu