Skip to content

Latest commit

 

History

History
97 lines (73 loc) · 3.73 KB

README.md

File metadata and controls

97 lines (73 loc) · 3.73 KB

vcocena in Docker

Reproducible container for the Vertical CoCena² pipeline.

Build

From the repository root (the directory containing DESCRIPTION), not from docker/:

docker build -f docker/Dockerfile -t vcocena:1.0.0 .

The build is layered so that editing package source under Vcocena/R/ only re-runs the fast package install, not the ~30-minute dependency install. .dockerignore keeps the build context at ~1 MB by excluding the 2.1 GB of .RData workspaces and the datasets.

Run

Smoke test (self-contained, seconds — uses the 100-gene example baked into the package):

docker run --rm vcocena:1.0.0 Rscript /opt/docker/run_smoke.R

Full run on a real two-layer dataset:

mkdir -p out_docker
docker run --rm \
  -v "/path/to/your/dataset:/data:ro" \
  -v "$PWD/out_docker:/out" \
  vcocena:1.0.0 Rscript /opt/docker/run_full.R

Expect 10–30 min and at least 8 GB of container memory: correlation_actions() materializes a data frame of all choose(5000,2) ≈ 12.5M gene pairs per layer.

The dataset mount must contain data/ (count matrices + reference_files/) and sample_info/ (annotations). Paths are overridable with VCOCENA_DATA and VCOCENA_OUT.

Why the inputs are copied out of the mount

run_vcocena() creates its save_folder inside working_directory and writes plots there, so the analysis directory can never be a read-only mount or the installed package library. Both run scripts copy the (small) inputs into /work first. The same applies to vcocena_example(), which returns a path inside the installed library.

What the full run emits

Alongside the pipeline's PDFs, run_full.R writes three TSVs to the output mount:

file contents
integrated_cluster_information.tsv integrated module assignments
integrated_GFC_per_cluster.tsv Group Fold Change per module per condition
integrated_edgelist.tsv the integrated network

The pipeline has no tabular export path of its own — every native output is a PDF — so these exist to give a future regression baseline something numeric to compare against.

Notes

  • Base image bioconductor/bioconductor_docker:RELEASE_3_22-R-4.5.2 (R 4.5.2), chosen to match the R 4.5.0 that produced the existing out_stepwise/ output.
  • MCDA is installed from the CRAN Archive: it was removed from CRAN on 2025-11-20 but is still used on the main path.
  • Only hard dependencies are installed. devtools is not present, so the vCoCena_stepwise_*.R drivers will not run in this image — they also depend on plot_sample_distributions(), which was never ported into the package. Use run_vcocena().

Reproduction harness

run_paper.R and run_paper_integration.R compare this package against the published vCoCena analysis in Carraro et al., eLife 2022 (doi:10.7554/eLife.78012), using the authors' own saved R environment as ground truth (Zenodo 10.5281/zenodo.6984701, CC-BY-4.0).

Two findings from that exercise are worth knowing before using either script:

  • run_vcocena() cannot reproduce that paper end to end. Its published layers were hCoCena integrations across two cell lines, not count matrices. run_paper_integration.R therefore injects the published layer networks directly into the vertical-integration stage, which reproduces the published network exactly (3,405 nodes / 186,723 edges / 35,453 cross-layer edges).
  • Module counts will not match a pre-2024 analysis. igraph 2.0 rewrote community detection; on the identical graph, cluster_louvain returns 6 communities where igraph 1.2.6 returned 147. This affects any re-run of an older CoCena analysis, not just this one.