Prediction of interferon stimulated and repressed genes using sequence composition

Code for the machine learning component of Shaw, Rihn, Mollentze, et al. (2021) "The ‘antiviral state’ has shaped the CpG composition of the vertebrate interferome". Raw data for this publication are available from DOIs 10.5525/gla.researchdata.1159 and 10.5281/zenodo.5035606, and will be downloaded automatically as needed by this analysis pipeline.

The aim of this part of the analysis was to investigate to what extent interferon-stimulated genes, interferon-repressed genes, and remaining genes are compositionally distinct, and to identify the most important compositional features allowing them to be distinguished.

Usage

Analyses were run using R version 3.5.1. Required R libraries are managed by packrat. A full run requires approximately 15GB of disk space, and by default will run in parallel across 8 threads.

To re-run the entire pipeline, use

Rscript packrat/init.R --args --bootstrap-packrat
make all

Run make help for further details.

Folder structure

[*] Indicates folders which will be created while running the pipeline

└─isg_composition/
   ├─CalculatedData/...................... [*] Final pre-processed data used to train models
   │ 
   ├─Data/ ............................... Raw data (most files downloaded from the main data  
   │   │                                   repositories as needed)
   │   ├─ISG_identities.csv                Interferome data, from Shaw et al. 2017 (downloaded
   │   │                                   from https://isg.data.cvr.ac.uk)
   │   ├─consistent_irgs.csv               IRGs consistently found across all experiments
   │   ├─MouseHoldout/                     [*] Holdout data describing the mouse interferome
   │   │                                       (from Dölken et al., 2008)
   │   ├─Other_Experiments/                [*] Data from the knockdown- and other experiments
   │   │                                       described in the manuscript
   │   ├─Part1_Dinucs/                     [*] Expression data and dinucleotide biases (see 
   │   │                                       "Data" below)
   │   ├─Part2_Codons/                     [*] Codon and codon-pair biases, etc
   │   └─SelectedGenes/                    [*] Genes analyzed in the manuscript
   │
   ├─Output/ ............................. [*] Generated output files
   │
   ├─packrat/ ............................ Record of R libraries used
   │
   ├─Scripts/ ............................ Main pipeline scripts
   │
   └─Makefile ............................ Record of workflow and dependencies between files

Data

Data will be downloaded as needed while running the pipeline, but can also be obtained using make download_data (but note that not all data from the manuscript will be extracted, and most files in the "Other_Experiments" folder will be renamed - refer to the original repositories linked above for the canonical versions of these files). The core genome composition data retreived are split across several files:

• In Data/Part1_Dinucs/:
  • This folder contains the expression data and dinucleotide composition summaries for the longest transcript of each gene
  • Files named [species]_cds_new_dat_fpkm_dups.txt contains data for coding sequences only
  • Files named [species]_cdna_new_dat_fpkm_dups.txt contains the same calculations across the entire coding sequence
  • Most files are derived from cds/cdna datasets provided by the Ensembl FTP server
    • Where [species] is replaced with 'biomart': For humans only, contains all human genes, downloaded manually from Ensembl Biomart
• In Data/Part2_Codons/:
  • Contains codon and codon-pair biases for the same transcripts as above
  • File names have the format [species]_cds_new_cpb_dat_dups.txt
  • Note that there is no corresponding cdna version here, since these features are valid for coding regions only