Generalized linear regression models with network-regularization in R. Now with TensorFlow.

About

Modelling dependencies using linear regression models is often complicated when the analysed data-sets are high-dimensional and less observations than variables are available (n << p). netReg implements generalized linear models that utilize network penalties for regularization. Network regularization uses graphs or trees to incorporate information about interactions of covariables, or responses, into the loss function of a GLM. Ideally this allows better (i.e., lower variance) estimation of regression coefficients.

For instance, in R, you could fit a network-regularized model like that:

From version v1.9.0 on, we use TensorFlow, instead of custom C++ and Dlib, for estimation of regression coefficients replacing a custom cyclic coordinate descent. This allowed deleting of major parts of the code base. netReg still uses some RcppArmadillo for fast matrix algebra.

In order to estimate the optimal hyperparameters, i.e., the regularization parameters of the network models, we use Powell’s BOBYQA algorithm in a standard cross-validation framework.

For more details, please check out the respective vignettes of the single models.

Installation

Before installing the package, make sure to have these Python dependencies installed:

  • tensorflow==1.14.0,
  • tensorflow-probabiltiy==0.5.0

The easiest way is probably to install TensorFlow from with R and then call:

tensorflow::install_tensorflow(extra_packages = "tensorflow-probability")

That creates a conda environment (in case you use it) called r-tensorflow and installs the Python dependencies automatically.

You can install and use netReg either from Bioconductor, or by downloading the latest tarball.

If you want to use the recommended way using Bioconductor just call:

> if (!requireNamespace("BiocManager", quietly=TRUE))
>   install.packages("BiocManager")
> BiocManager::install("netReg")
  
> library(netReg)

Installing the R package using the downloaded tarball works like this:

I do not recommend using devtools, so please prefer tarballed releases over installing from the main branch.

Documentation

  • Load the package using library(netReg). We provide vignettes for the package that can be called using: vignette(package="netReg").
  • You can also use the online vignette.

Citation

If netReg was useful for you or your work, it would be great if you cited it like this:

@article{,
  title={netReg: network-regularized linear models for biological association studies},
  author={Dirmeier, Simon and Fuchs, Christiane and Mueller, Nikola S and Theis, Fabian J},
  journal={Bioinformatics},
  volume={34},
  number={5},
  pages={896--898},
  year={2017},
  publisher={Oxford University Press}
}

Author

Simon Dirmeier simon.dirmeier@gmx.de