Installation¶
Requirements¶
- Nextflow >= 23.04.0 (install guide)
- Docker or Singularity (for running model containers)
- NVIDIA GPU (recommended) — required for seq2ribo and Orthrus; foundation/structure models also run much faster on GPU. CPU works for the other 13 models, just slower.
- NVIDIA Container Toolkit (for GPU runs) — needed so Docker can expose the GPU to containers. See the official install guide. Without it, GPU profiles fail with
unknown runtime: nvidia.
Install Nextflow¶
Or via conda:
Clone the pipeline¶
Docker images¶
All model containers are hosted on GitHub Container Registry (ghcr.io/ericmalekos/rnazoo-*). Nextflow pulls them automatically on first run — no manual docker pull needed. The per-model image sizes are listed on the home page so you can pick a subset rather than pulling everything.
If you'd rather skip Nextflow entirely and drive the containers yourself, see the Direct Docker guide for per-model docker run recipes.
If you want to warm the cache up front, pick the loop that matches your hardware. If you have an NVIDIA GPU, use the GPU loop below — that's the project's default and the path the test suite is tuned for. The CPU loop is the fallback for hardware without CUDA.
# --- GPU setup (~70 GB total compressed download)
# Pulls the CUDA-enabled variants for every tool. seq2ribo and orthrus
# are GPU-only by design (mamba-ssm requires CUDA at import time).
for img in ribonn riboformer tristan seq2ribo saluki translationai \
codontransformer rnafm rinalmo ernierna orthrus rnaformer \
rhofold spotrna multirm utrlm; do
docker pull ghcr.io/ericmalekos/rnazoo-${img}:latest
done
VRAM caveat: running 15+ models in parallel on a single GPU needs more memory than a typical consumer card (a 4 GB device will OOM partway through). If you have a small GPU, run models one-at-a-time with explicit
--<model>_inputflags instead of using the test profile.
# --- CPU-only setup (~28 GB total compressed download)
# Selects the smaller -cpu variants where they exist. ernierna and
# rnaformer only ship a single image (built with CUDA-enabled framework
# but runs on CPU too) — included here for completeness. seq2ribo and
# orthrus are GPU-only and not in this list.
for img in ribonn-cpu riboformer-cpu tristan-cpu saluki-cpu translationai-cpu \
codontransformer-cpu rnafm-cpu rinalmo-cpu ernierna rnaformer \
rhofold-cpu spotrna-cpu multirm-cpu utrlm-cpu; do
docker pull ghcr.io/ericmalekos/rnazoo-${img}:latest
done
CPU performance note: foundation models (RNA-FM, RiNALMo, ERNIE-RNA, UTR-LM) take ~30–60 s per inference call on CPU and sub-second on GPU. Use the GPU loop above if you have an NVIDIA card.
Profiles¶
| Profile | Description |
|---|---|
docker |
Use Docker as container engine |
singularity |
Use Singularity as container engine |
gpu |
Enable GPU with NVIDIA runtime (default for the test suite) |
cpu |
Force CPU mode (GPU-only models auto-skip with a warning) |
test |
Use bundled minimal test inputs (CPU-friendly subset) |
test_gpu |
Use bundled minimal test inputs including GPU-only / GPU-recommended models (seq2ribo, RhoFold, Orthrus) |
Combine profiles with commas:
Verify installation¶
If you have an NVIDIA GPU, run the GPU test suite — this is the canonical end-to-end check:
# Run the GPU test suite (16 models, requires NVIDIA Container Toolkit)
nextflow run . -profile test_gpu,docker,gpu
Expected output — all 16 models should pass:
RNAZOO:RIBONN (ribonn) | 1 of 1 ✔
RNAZOO:RIBOFORMER (riboformer:bacteria_cm_mg) | 1 of 1 ✔
RNAZOO:RIBOTIE (ribotie) | 1 of 1 ✔
RNAZOO:TRANSLATIONAI (translationai) | 1 of 1 ✔
RNAZOO:SALUKI (saluki:human) | 1 of 1 ✔
RNAZOO:CODONTRANSFORMER (codontransformer:...) | 1 of 1 ✔
RNAZOO:RNAFM (rnafm) | 1 of 1 ✔
RNAZOO:RINALMO (rinalmo) | 1 of 1 ✔
RNAZOO:ERNIERNA (ernierna) | 1 of 1 ✔
RNAZOO:ORTHRUS (orthrus) | 1 of 1 ✔
RNAZOO:RNAFORMER (rnaformer) | 1 of 1 ✔
RNAZOO:SPOTRNA (spotrna) | 1 of 1 ✔
RNAZOO:MULTIRM (multirm) | 1 of 1 ✔
RNAZOO:UTRLM (utrlm:mrl) | 1 of 1 ✔
RNAZOO:SEQ2RIBO (seq2ribo:te:hek293) | 1 of 1 ✔
RNAZOO:RHOFOLD (rhofold) | 1 of 1 ✔
Succeeded : 16
Riboformer's bundled-dataset test runs faster on GPU (~1.5 min vs ~2.5 min on CPU). Models with both CPU and GPU image variants automatically use their GPU images under -profile gpu.
CPU fallback¶
If you don't have an NVIDIA GPU, the smaller test profile runs the 13 of 16 models that work on CPU in a reasonable time:
Expected output — all 13 should pass:
RNAZOO:RIBONN (ribonn) | 1 of 1 ✔
RNAZOO:RIBOFORMER (riboformer:bacteria_cm_mg) | 1 of 1 ✔
RNAZOO:RIBOTIE (ribotie) | 1 of 1 ✔
RNAZOO:TRANSLATIONAI (translationai) | 1 of 1 ✔
RNAZOO:SALUKI (saluki:human) | 1 of 1 ✔
RNAZOO:CODONTRANSFORMER (codontransformer:...) | 1 of 1 ✔
RNAZOO:RNAFM (rnafm) | 1 of 1 ✔
RNAZOO:RINALMO (rinalmo) | 1 of 1 ✔
RNAZOO:ERNIERNA (ernierna) | 1 of 1 ✔
RNAZOO:RNAFORMER (rnaformer) | 1 of 1 ✔
RNAZOO:SPOTRNA (spotrna) | 1 of 1 ✔
RNAZOO:MULTIRM (multirm) | 1 of 1 ✔
RNAZOO:UTRLM (utrlm:mrl) | 1 of 1 ✔
Succeeded : 13
Three of the 16 models are excluded from the default CPU test:
| Model | Reason | Covered by test_gpu? |
|---|---|---|
| seq2ribo | GPU-only (mamba-ssm requires CUDA at import) | Yes |
| Orthrus | GPU-only (mamba-ssm requires CUDA at import) | Yes |
| RhoFold | Too slow on CPU (~5 min, 10 recycling iterations) | Yes |
Riboformer uses the in-image E. coli dataset (
GSE119104_Mg_buffer) via--riboformer_bundled_dataset— no external test data needed. See the Riboformer page for how to point it at a bundled dataset.
How profiles select container images¶
Most models ship two image variants on GHCR:
ghcr.io/ericmalekos/rnazoo-<tool>:latest— CUDA-enabled GPU build (default)ghcr.io/ericmalekos/rnazoo-<tool>-cpu:latest— CPU-only build (smaller)
-profile gpu selects the unsuffixed (GPU) image; -profile cpu selects the -cpu image. Nextflow handles the pull automatically — you do not invoke Docker directly. Exceptions: seq2ribo and Orthrus ship only as a single GPU image (mamba-ssm requires CUDA at import time and they auto-skip under -profile cpu).