This repository is organized as follows. It has five subdirectories:

Code

This contains source code and scripts used in the work. The files executeDBN.m, pycausal.py, and tigramite.py are the ones used to learn the different networks. The files unrolling.R and deconfounding.R are the ones used to compute the unrolling and deconfounding results, respectively.

Data

Each file contains a single omic data set, in either the aligned or unaligned formats, with one of five different constraint matrices (HEGTM, HEGTM_Skeleton, TM, TG, GM), and for both the Crohn's disease subset and the subset with all disease statuses are available here.

Networks

All networks created in this work are made available here. They are all in GRAPHML format, with a naming convention explained below.

Unrolling

All results from the unrolling approach are provided here in csv formats for each combination of parameters, as explained below. Note that we are using all the omic subsets in order to generate one combined file for the results of unrolling. The column headers of the file are as follows:

T and T_i are the source and target taxon of the unrolling operation.
Tx, G, and M represent the taxon, gene, or metabolite intermediary discovered by unrolling for the edge from T to T_i.
T-> (G|Tx)Bs, T-> (G|Tx){Bs}, (T|G)-> M_Bs, (M|G|Tx)-> T_i_Bs and T-> T_i_Bs are bootstrap scores from unrolling of the row
$overallScore$ is the multiplication of the three bootstrap scores.
$ChainInTTxTi$, $ChainInTGMTi$, $ChainInTGTi$, and $ChainInTMTi$ are booleans indicating whether that interaction also appears in the data subsets that are mentioned.

The files statsAllAugmented.csv and overallAVGStatsForEachMethod.csv are statistics for each network execution, and the overall statistics for all methods, respectively.

De-confounding

All results from the de-confounding approach are provided here in csv formats for each combination of parameters, as explained below. Note that we are using all the omic subsets in order to generate one combined file for the results of de-confounding. The column headers of the file are as follows:

$confounder$, $confoundedParent$, and $confoundedChild$ are the names of the three nodes involved in the confounding, and are therefore self-explanatory.
$BootscoreConfounderP$, $BootscoreConfounderC$, and $BootscoreConfounded$ are the bootstrap scores of the nodes from the item above. $overalScore$ is the result of their product.
$method$, $sourceLevel$, and $deconfoundedWith$ are the names of the network-learning method, the dataset where the confounded relationship was found, and the dataset used to de-confound it, respectively.

The additional files mostCommonConfounders.csv, overallDeconfoundingStats.csv, and overallDeconfoundingStatsByMethod.csv contain the list of all confounders found in this experiment, ordered by the number of times they were found, and the overall statistics of the results.

Details on the networks created are given below.

Dynamic Bayesian networks computed using PALM

DBNs were learned using palm for all subsets of the omics datasets mentioned in the manuscript (T, TG, TM, GM, TGM), alignment option (yes, no), sampling rate (7 days), number of bootstrap repetitions (100), number of allowable parents ({3, 4, 5, 6}), and the different types of constraint matrix used (HEGTM, HEGTM_Skeleton, TM, TG, GM). Some combination of parameters yielded empty networks and were not saved, making a maximum of 5 * 2 * 1 * 1 * 4 * 5 = 200 potential networks.

The networks were then combined, averaging the regression coefficient (weight) of the edges as long as they appeared in at least 10% of the repetitions. Each edge was also labeled with the bootstrap score or support (proportion of times that edge appears). Each repetition was set to run independently on a separate processor using Matlab's Parallel Computing Toolbox. All the networks from these experiments are made available.

The network files are named using the following convention:

DBN_{Data Subset}_{Dataset}_{Alignment Option}_sr{Sampling Rate}d_nboots{Bootstrap Repetitions}_nParents[Maximum Number Of Parents]_matrix{Constraint Matrix}_s{Computation Time (seconds)}.graphml

Causal Networks using the tetrad Suite

Networks computed with tetrad that are made available were learned for each omics dataset (T, TG, TM, GM, TGM), alignment option (yes, no), sampling rate (7 days), significance threshold α ∈ {0.0001, 0.001}, number of bootstrap repetitions (10), the FisherZScore network score, type of CI test (PositiveCorr, FisherZ, DegenerateGaussianLRT, SemBIC), maximum number of parents (3), and the different types of constraint matrix used (HEGTM, HEGTM_Skeleton, TM, TG, GM). Some combination of parameters yielded empty networks and were not saved, which makes for a maximum of 5 * 2 * 1 * 2 * 1 * 1 * 4 * 1 * 5 = 400 potential networks.

The files are named using the following convention:

PyCausal_{Data Subset}_{Dataset}_{Alignment Option}_sr{Sampling Rate}d_a{α}_nboots{Bootstrap Repetitions}_score{ScoreName}_test{CITestName}_nParents{Maximum Number Of Parents}_matrix{matrixName}_s{Computation Time (seconds)}.graphml

Causal Networks with Tigramite

Networks computed with Tigramite that are made available were learned for each omics dataset (T, TG, TM, GM, TGM), alignment option (yes, no), sampling rate (7 days), bootstrap threshold (0.1), significance threshold values α ∈ {0.0001, 0.001, 0.01, 0.1}, Tau max of 1, type of CI test (GPDC, CMIknn, ParCorr), and the type of constraint matrix used (HEGTM, HEGTM_Skeleton, TM, TG, GM). Some combination of parameters yielded empty networks and were not saved, which makes for a maximum of 5 * 2 * 1 * 1 * 4 * 1 * 3 * 5 = 600 potential networks.