Towards confidence estimates in Cascade Networks using the SelectBoost package

Frédéric Bertrand

Université de Strasbourg and CNRS
frederic.bertrand@math.unistra.fr

2019-05-21

#Introduction

Aims of the vignette

Extending results from the Cascade package: reverse engineering with selectboost to compute confidence indices for a fitted model. We first fit a model to a cascade network using the Cascade package inference function then we compute confidence indices for the inferred links using the Selecboost algorithm.

If you are a Linux/Unix or a Macos user, you can install a version of SelectBoost with support for doMC from github with:

devtools::install_github("fbertran/SelectBoost", ref = "doMC")

References

Code

Code to reproduce the datasets saved with the package and some the figures of the article Aouadi et al. (2018), <arXiv:1810.01670>

Data simulation

We define the F array for the simulations.

We set the seed to make the results reproducible

We simulate gene expression according to the network Net

Network inference

We infer the new network.

Heatmap of the coefficients of the Omega matrix of the network. Run the code to get the graph.

Compute the confidence indices for the inference

By default the crossvalidation is made subjectwise according to a leave one out scheme and the resampling analysis is made at the .95 c0 level. To pass CRAN tests, use.parallel = FALSE is required. Set use.parallel = TRUE and select the number of cores using ncores = 4.

Use cv.subject=FALSE to use default crossvalidation

Analysis of the confidence indices

Use plot to display the result of the confidence analysis.

Run the code to plot the other results.

Run the code to plot the remaning graphs.

Distribution of non-zero (absolute value > 1e-5) coefficients

Plot of confidence at .95 resampling level versus coefficient value for non-zero (absolute value > 1e-5) coefficients

Plot of confidence at .5 resampling level versus coefficient value for non-zero (absolute value > 1e-5) coefficients

Plot of confidence at .95 resamling level with groups created by thresholding the correlation matrix versus coefficient value for non-zero (absolute value > 1e-5) coefficients.

Plot of confidence at .95 resampling level versus coefficient value for non-zero (absolute value > 1e-5) coefficients using standard cross-validation.

Further improvements

Either decrease percent selected or choose c0 as a quantile.