This page aims to address some implementation and use case questions. We strongly recommend first reading the tutorial for an overview of COGENT’s functionality and intended use.
COGENT can be installed directly from R, using the devtools package. You can then load COGENT like any other R package.
# Install
require("devtools")
install_github("lbozhilova/COGENT")
# Load
require("COGENT")
For more detailed instructions please refer to the tutorial.
COGENT is compatible with R v.3.6.2 and above. However, we recommend always using the most recent version of R (R v.4.0.2 at the time of writing) and regularly updating your packages.
COGENT has been tested on Windows, MacOS and Linux (Ubuntu and Fedora). You can access all of its functionality on MacOS and Linux. However, Windows does not support forking, and so the paralellised functions of COGENT cannot be used to improve your computation speed on Windows. We therefore recommend that either MacOS or Linux is used for computationally intensive analysis.
propShared) should I use when running COGENT?A COGENT iteration works by randomly splitting your gene expression samples in two subsets, constructing a network from each subset, and comparing the resulting networks. These sample subsets can overlap, and the overlap is controlled by the propShared parameter (see ?cogentSingle for details).
The default value propShared=0 will always result in disjoint subsets. If your dataset consists of relatively few samples (say less than 50), we recommend introducing a set overlap. The best amount of overlap will depend on your data, but we recommend setting propShared=0.25 or propShared=0.5 in the first instance.
To illustrate, given twenty samples and propShared=0, cogentSingle() will generate two disjoint subsets of ten samples each. Ten samples may not be enough to construct an informative co-expression network. Setting propShared=0.5 will result in two overlapping subsets, each consisting of fifteen samples. Of these, ten will be shared and five will be unique to each subset.
repCount) should I run?Multiple independent iterations of COGENT are required to provide the most robust results. The number of iterations is controlled by the repCount parameter (see ?cogentLinear for details).
We generally recommend using COGENT with the default value of repCount=100 iterations. Please note that cogentParallel can be used to reduce the computational time if needed.
getEdgeSimilarityCorrected)?Density correction, also known as density adjustment, is introduced in the getEdgeSimilarityCorrected() function. It can be used to compare network construction methods which result in networks of significantly different density.
We recommend that you always use density correction if you are trying to find an optimal threshold for network construction, i.e. if you are trying to determine score cut-offs using COGENT. The tutorial contains a worked example of this.
We also recommend you use density correction when comparing two network construction methods which are not guaranteed to result in networks of similar density.
You do not need to use density correction if you are using node metric consistency as your primary consistency measure. In this case, it is important that your node metric of choice is independent of network density.
We recommend that you do not use density correction for methods resulting in signed and/or weighted networks.
There are two density correction methods available through the type parameter of the getEdgeSimilarityCorrected() function. These are the semi-random correction (type="expected") and the fully-random correction (type="random"). Refer to the Supplementary Information of our paper for details on how they are computed.
By default, the semi-random correction is carried out. This is significantly faster than the fully-random correction, and we recommend using it.
COGENT is designed for unsigned networks. In particular, edge similarity calculated using getEdgeSimilarity() and getEdgeSimilarityCorrected() will be entirely non-informative when calculated for signed networks.
Nevertheless, there are several ways in which you can use COGENT for signed networks.
You can convert signed networks to unsigned by rescaling or using quantiles. The buildPearson() function in the tutorial illustrates one possible way of doing this - Pearson correlation coefficients can be negative, but if their quantiles are used instead, thresholding and weighted comparisons can be readily performed.
You can also use node metric consistency, either through the main COGENT functions or through getNodeSimilarity(), to evaluate signed network construction methods. Note that in this case, your node metric of choice should produce non-negative values. This can be achieved via rescaling, for example.
Please feel free to contact us with any questions by emailing Charlotte Deane at deane@stats.ox.ac.uk.
COGENT was originally developed and is currently maintained by Lyuba Bozhilova. You can email her directly at lyubabozhilova@gmail.com.