The original paper used R 3.3.0 on Debian 7 Wheezy, but because CRAN only released R version 3.2.5 for Wheezy, we use a Singularity image to replicate the R executable.
Install the latest version of singularity for your operating system.
Then run the bootstrap.sh script in the cloned directory to:
- Download the supplement data to
downloads/, - Install the R packages sandboxed under
r-lib/in the current directory, and finally - Run the analysis.
The R minor release version matters because Bioconductor package releases are pinned against them.
Singularity produces portable, read-only, containers based on a non-proprietary image format and suited to scientific workflows.
Docker has lately been favored by the academic community, but has several troubling drawbacks in practice:
- A reason many academics use Docker is for
Makefile-like caching long operations. - But without continuous integration, caching does not make the work
reproducible, and there are more sensible language specific
memoization operations available such as
Biobase::cache()inR. - Docker images have also become a crutch for distributing poorly packaged software. In other words, distributing containers has become customary instead of writing the traditional language-specific software package and minimal analysis scripts that rely on package controlled code.
- Images also encourage bad practices of not separating data and code.
- Docker has non-sensical default security, such as root privileged containers without user namespace remapping (see NCC Group whitepaper).
- Docker has poor community support with nearly all questions on the docker forum unanswered.
In industry settings, having the smallest possible infrastructure footprint using containers is important, but less so in academia given the drawbacks above. Preferring language-specific software packaging, and minimizing the user of containers is more extensible, maintainable and reproducibile.