ADAGEpath

ADAGEpath provides necessary functions to perform ADAGE-based signature analysis.

ADAGE introduction

ADAGE (or eADAGE) models are neural networks trained in an unsupervised manner on large publicly available gene expression compendia. ADAGE aims at building essential features that can reconstruct a compendium. We derived gene expression signatures from ADAGE neural network nodes and found that these signatures resemble human-annotated biological pathways and cover many existing pathways [1,2]. In addition to signatures that match known pathways, ADAGE also extracts signatures that may group genes in novel data-driven ways.

Please refer to the following papers if you want to learn more about ADAGE:

1. ADAGE-Based Integration of Publicly Available Pseudomonas aeruginosa Gene Expression Data with Denoising Autoencoders Illuminates Microbe-Host Interactions
2. Unsupervised extraction of functional gene expression signatures in the bacterial pathogen Pseudomonas aeruginosa with eADAGE

Preloaded data

ADAGEpath currently supports one organism Pseudomonas aeruginosa. The package is preloaded with a P.a. expression compendium (PAcompendium and its probe quantile distribution probedistribution) and an eADAGE model (eADAGEmodel) trained on the compendium. The package is also preloaded with P.a. gene (geneinfo and PAO1orthologs) and operon (operons) information.

ADAGE signature analysis

Signatures are gene sets derived from an ADAGE model. They are formed because their genes are expressed coordinately in some samples in the compendium. An ADAGE signature analysis aims to identify signatures in which the expression of constituent genes are altered by an experimental treatment. Such signatures may represent biological processes that are perturbed by the treatment. ADAGE signature analysis usually includes the following steps:

Data loading

ADAGEpath currently supports raw microarray data in CEL format and processed microarray or RNAseq data. Use function load_dataset() to load your own dataset or datasets publicly available in ArrayExpress.

Since ADAGE only accepts expression values in the (0,1) range, we linearly transform expression values to be between 0 and 1 using the function zeroone_norm().

Signature activity calculation

We next calculate each signature's activity for each sample in the dataset with the function calculate_activity().

Active signature detection

We next identify signatures whose activities strongly vary with treatments, such as signatures whose activities are significantly different in conditions of interest. We recommend using limma to test differential activation, particularly when sample size is small. To facilitate the most frequent two-group comparison, we wrapped a two-group limma test into the function build_limma(). You can visualize the limma two-group test results using plot_volcano() and get active signatures from the limma test result using get_active_signatures(). We also provide examples of analyzing time-course experiments or factorial-design experiments using limma in the vignettes. You can also use other statistical tests to identify active signatures. plot_activity_heatmap() generates a heatmap showing how signature activity changes across samples.

Signature overlap examination

To be robust to noise, ADAGE models would sometimes construct signatures that have overlapping genes. We can check whether the active signatures identified above overlap with others using plot_signature_overlap(). If there is a group of signatures that largely overlap, to reduce the number of signatures to look at, we can calculate the marginal activity of each signature using calculate_marginal_activity(), which is the remaining activity of a signature after removing overlapping genes of another signature. We can visualize whether a signature is still strongly active after the impact of another signature has been removed using plot_marginal_activation(). Examples of signature overlap examination are in vignettes ArrayExpress-example and Time-course-example.

Signature interpretation and visualization

Finally, to get a detailed view of a signature or a group of signatures, we can retrieve their constituent genes using annotate_genes_in_signatures() and visualize these genes through a gene-gene network using visualize_gene_network(). We can also download existing KEGG pathways using fetch_geneset() and associate signatures to known KEGG pathways using annotate_signatures_with_genesets().

Vignettes

The package comes with 5 vignettes introducing how ADAGE signature analysis can be performed for different types of datasets.

• ArrayExpress-example: provides an example of loading in a dataset available on ArrayExpress and performing a complete ADAGE signature analysis workflow on a experimental design with two phenotype groups.

• User-input-example: provides an example of loading in a local dataset and performing a complete ADAGE signature analysis workflow on a experimental design with two phenotype groups.

• RNAseq-example: provides an example of loading in a processed RNAseq dataset available on ArrayExpress.

• Factorial-design-example: provides an example of detecting differentially active signatures with limma for an experiment with factorial design. Such design typically has two treatment factors and measurements at four treatment combinations.

• Time-course-example: provides an example of analyzing experiments with complex experimental design. The first part looks at signatures that show the largest activity ranges across samples. The second part builds a limma model to detect signatures with differential temporal patterns between treatment and control.

Installation

You can install the latest development version from github with

install.packages("devtools")
devtools::install_github("greenelab/ADAGEpath")

If you want to build the vignettes, run

devtools::install_github("greenelab/ADAGEpath", build_vignettes = TRUE)

You can also install it via BioInstaller

library(BiocInstaller)
biocLite("greenelab/ADAGEpath")

Potential problems during installation:

• make: gfortran-4.8: No such file or directory.

  ADAGEpath depends on the Bioconductor package impute that requires gfortrain-4.8. Install gfortrain-4.8 to solve this problem. You can follow instructions here to install gfortran-4.8 on Mac https://stackoverflow.com/questions/23916219/os-x-package-installation-depends-on-gfortran-4-8

• Error validating server certificate for 'https://hedgehog.fhcrc.org:443'

  Please follow this thread to accept the certificate r-lib/devtools#1401

• BiocInstaller package not available

  You can install it via

  devtools::install_bioc("BiocInstaller")

  or

  source("https://bioconductor.org/biocLite.R")
  biocLite("BiocInstaller")