dREG-apptainer

A prebuilt Apptainer image of dREG — the reference R/CUDA implementation, frozen with the exact R 4.0.5 / CUDA 11.1 stack it was validated against. Download it, bind your data, call peaks.

Most people should use pydreg instead

uv tool install pydreg     # or: pip install pydreg

pydreg is a Python reimplementation that installs in seconds instead of needing a multi-GB CUDA container, runs substantially faster, works on Apple Silicon, and agrees with this image to >0.999 Jaccard on peak calls at default parameters.

Use this image when you specifically need the R reference: reproducing published dREG 20200515 results, or validating pydreg against R. For calling peaks on your own data, use pydreg.

1. Get the image and the model

Both live in the same Zenodo record: 10.5281/zenodo.22225970

apptainer pull dreg.sif \
  https://zenodo.org/records/22225970/files/dreg.sif
curl -LO https://zenodo.org/records/22225970/files/asvm.gdm.6.6M.20170828.rdata

That's 3.6 GB for the image and 570 MB for the model. Verify if you like:

md5sum dreg.sif asvm.gdm.6.6M.20170828.rdata
# dreg.sif                      2cb1ca16fb4c880b96faedaa90a094bd
# asvm.gdm.6.6M.20170828.rdata  da81e96de4988021fc1e53d5f79b77ad

The model is deliberately not baked into the image — keep it beside your data and bind it in. It is also mirrored at https://dreg.dnasequence.org/themes/dreg/assets/file/asvm.gdm.6.6M.20170828.rdata.

10.5281/zenodo.10113378 is the concept DOI for the record and always resolves to the newest version; cite that one.

Or build it yourself

On a Linux x86_64 host (Apptainer doesn't run on macOS, and the base image is amd64):

apptainer build --fakeroot --ignore-fakeroot-command dreg.sif dreg.def

--ignore-fakeroot-command is not optional on most hosts: without it Apptainer injects a host faked binary that needs newer glibc than Ubuntu 20.04 provides, and %post fails with GLIBC_2.33/GLIBC_2.34 not found. sudo apptainer build and singularity build also work. Expect a multi-GB image and ~10 GB of scratch during the build. Details in docs/MAINTENANCE.md.

2. Check the GPU works

Do this once, before a long run. It trains a tiny SVM through Rgtsvm — the first thing in the stack that touches the GPU:

apptainer exec --nvccli --cleanenv dreg.sif dreg gpu_check

It should print an SVM summary and OK: Rgtsvm created a CUDA context. If it doesn't, go to Troubleshooting — do not skip ahead, since apptainer test passes even when the GPU path is completely dead.

3. Call peaks

apptainer exec --nvccli --cleanenv --bind /path/to/data:/data dreg.sif \
  dreg run_dREG /data/sample.pl.bw /data/sample.mn.bw /data/sample \
                /data/asvm.gdm.6.6M.20170828.rdata 16 0

Argument	Meaning
sample.pl.bw	plus-strand PRO-seq/GRO-seq bigWig
sample.mn.bw	minus-strand bigWig
sample	output prefix, not a filename
asvm...rdata	the model file
16	CPU cores (optional, default 1)
0	GPU id (optional)

Two things that catch people:

• The GPU id is what enables GPU mode. Omit it and dREG silently runs on CPU, which is orders of magnitude slower. Pass 0 to use the default GPU; pass a higher number to select a specific one if nvidia-smi shows GPU 0 is busy.
• bigWigs must be raw read counts, not normalized, with the minus strand negative. dREG validates this and aborts with "bigWig files maybe not meet the requirements"; see upstream data preparation.

(The upstream usage text mentions a 7th GPU.threads argument. The script never reads it — don't bother passing it.)

4. Output files

With prefix sample, you get these next to your data. Each .bed.gz is bgzip-compressed and tabix-indexed (.bed.gz.tbi alongside):

File	Contents
sample.dREG.peak.full.bed.gz	the main result — peaks with score and probability
sample.dREG.peak.prob.bed.gz	peaks with probability only (chrom, start, end, prob)
sample.dREG.peak.score.bed.gz	peaks with raw SVR score only
sample.dREG.infp.bed.gz	scores at every informative position
sample.dREG.raw.peak.bed.gz	raw peak calls before filtering
sample.dREG.peak.prob.bw	probability as a bigWig, for genome browsers
sample.dREG.peak.score.bw	score as a bigWig
sample.dREG.infp.bw	informative-position scores as a bigWig
sample.chrom.info	chromosome sizes derived from your bigWigs

Most downstream work uses *.dREG.peak.full.bed.gz or *.dREG.peak.prob.bed.gz.

Troubleshooting

CUDA driver version is insufficient — use --nvccli, not --nv

If you launch with --nv, Rgtsvm aborts at its first CUDA call:

terminate called after throwing an instance of 'GTSVM::CUDA::Exception'
  what():  svm.cpp:735: Failed to allocate space for found keys on host
           (CUDA driver version is insufficient for CUDA runtime version)

The message is wrong — your driver is fine. --nv bind-mounts a fixed library list from Apptainer's nvliblist.conf, which is incomplete for recent drivers. --nvccli routes through nvidia-container-cli instead and works. Always use --nvccli.

It implies --writable-tmpfs and sets NVIDIA_VISIBLE_DEVICES=all; both are harmless. If your cluster has no nvidia-container-cli, --nv may still work on older drivers — confirm with dreg gpu_check before committing to a long run. docker run --gpus all on the same image also works.

Confirmed broken with --nv on driver 580.167.08 / CUDA 13.0 with Pascal GPUs. The full diagnostic record, including what was ruled out, is in docs/MAINTENANCE.md.

R loads packages from my home directory

If R picks up ~/R/x86_64-pc-linux-gnu-library/4.0 and breaks the pinned stack, --cleanenv normally prevents it. To force the issue:

apptainer exec --nvccli --cleanenv \
  --env R_LIBS_USER=/tmp \
  --env R_PROFILE_USER=/dev/null \
  --env R_ENVIRON_USER=/dev/null \
  --bind /path/to/data:/data dreg.sif \
  dreg run_dREG /data/sample.pl.bw /data/sample.mn.bw /data/sample \
                /data/asvm.gdm.6.6M.20170828.rdata 16 0

Other notes

• Don't load a host CUDA toolkit module before launching. The image supplies the CUDA runtime; only the driver should come from the host.
• Peak calling is memory-hungry on top of the GPU. If it dies partway, retry with fewer CPU cores before assuming a GPU problem.

Other commands

Everything runs through one wrapper, /usr/local/bin/dreg:

Command	Purpose
dreg run_dREG PL.bw MN.bw PREFIX MODEL.rdata [CORES] [GPU]	peak calling (above)
dreg run_predict PL.bw MN.bw PREFIX MODEL.rdata [CORES] [GPU]	legacy score prediction
dreg writeBed THRESHOLD FILE.bedGraph.gz	peaks from legacy bedGraph output
dreg gpu_check	prove the GPU works
dreg cuda_check	print nvidia-smi, LD_LIBRARY_PATH, CUDA library resolution
dreg shell	interactive bash in the container
anything else	run directly in the container

Anything unrecognised is executed as-is, so the image doubles as a frozen R 4.0.5 environment:

apptainer exec dreg.sif Rscript -e 'library(dREG); sessionInfo()'

What's inside

dREG 1.4.0, Rgtsvm 0.55, rphast 1.6.9, bigWig 0.2-9 on R 4.0.5 / CUDA 11.1 / Ubuntu 20.04, from the Cornell BioHPC image biohpc/dreg pinned by digest. Exact versions and commits for all 84 R packages are in docs/reference-stack-manifest.md.

Building the image, why it wraps a prebuilt Docker image, and how it is archived are covered in docs/MAINTENANCE.md.