Riboformer¶

Predict and refine codon-level ribosome densities from ribo-seq data.

Paper: Nature Communications 2024
Upstream: https://github.com/lingxusb/Riboformer
License: Upstream repository license
Device: CPU or GPU. Two image variants:
- rnazoo-riboformer:latest — CUDA-enabled (default, used with -profile gpu)
- rnazoo-riboformer-cpu:latest — CPU-only (smaller, used with -profile cpu)

What it does¶

Riboformer is a transformer model that takes existing ribo-seq data (ribosome profiling WIG coverage) along with genome sequence and annotation, and predicts refined codon-level ribosome densities. It can transfer learned patterns from a reference condition to a target condition.

Pre-trained models are available for: yeast (mono/disome), E. coli, C. elegans, and SARS-CoV-2.

Input format¶

A directory containing:

WIG files (forward + reverse strands): ribosome profiling coverage for reference and target conditions
<reference>_f.wig, <reference>_r.wig
<target>_f.wig, <target>_r.wig
FASTA file: genome sequence
GFF3 file: gene annotation

The bundled datasets are in /opt/Riboformer/datasets/ inside the Docker image (e.g., GSE152850_yeast/).

Output format¶

model_prediction.txt: codon-level predicted ribosome density values (one value per line per codon)
pause_indices.txt (optional): ribosome pause indices per codon

Run with Docker¶

See the Direct Docker guide for the shared docker run recipe (UID, HOME, USER env vars, and GPU flag). Below are the model-specific parts.

Using the bundled yeast disome dataset (CPU shown; for GPU swap rnazoo-riboformer-cpu → rnazoo-riboformer and add --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=all):

docker run --rm \
  -v /path/to/output:/out \
  ghcr.io/ericmalekos/rnazoo-riboformer-cpu:latest \
  bash -c "cd /opt/Riboformer/Riboformer && \
    python transfer.py -i GSE152850_yeast -m yeast_disome && \
    cp /opt/Riboformer/datasets/GSE152850_yeast/model_prediction.txt /out/"

With your own data (two-step pipeline):

docker run --rm \
  -v /path/to/your/data:/opt/Riboformer/datasets/my_data \
  -v /path/to/output:/out \
  ghcr.io/ericmalekos/rnazoo-riboformer:latest \
  bash -c "cd /opt/Riboformer/Riboformer && \
    python data_processing.py -d my_data -r reference_wig_name -t target_wig_name -p 14 -w 40 -th 25 && \
    python transfer.py -i my_data -m yeast_disome && \
    cp /opt/Riboformer/datasets/my_data/model_prediction.txt /out/"

Run with Nextflow¶

# CPU
nextflow run main.nf -profile docker,cpu \
  --riboformer_input /path/to/data_dir \
  --riboformer_reference_wig reference_name \
  --riboformer_target_wig target_name \
  --riboformer_model yeast_disome

# GPU
nextflow run main.nf -profile docker,gpu \
  --riboformer_input /path/to/data_dir \
  --riboformer_reference_wig reference_name \
  --riboformer_target_wig target_name \
  --riboformer_model yeast_disome

Only models with input provided will run — no ignore flags needed.

Parameters¶

Parameter	Default	Description
`--riboformer_model`	`yeast_disome`	Pre-trained model to use
`--riboformer_psite`	`14`	P-site offset
`--riboformer_wsize`	`40`	Window size
`--riboformer_threshold`	`25`	Minimum read threshold
`--riboformer_bundled_dataset`	`null`	Use a dataset already inside the image (skips external staging). When set, `--riboformer_input` is ignored. See "Bundled datasets" below.

Bundled datasets¶

The Docker image ships with several upstream datasets at /opt/Riboformer/datasets/. Pointing --riboformer_bundled_dataset at one of these lets you run end-to-end without supplying external files:

nextflow run main.nf -profile docker,cpu \
  --riboformer_bundled_dataset GSE119104_Mg_buffer \
  --riboformer_reference_wig GSM3358138_filter_Cm_ctrl \
  --riboformer_target_wig GSM3358140_freeze_Mg_ctrl \
  --riboformer_model bacteria_cm_mg

Datasets that have all the files needed for end-to-end inference (WIG + FASTA + GFF):

Bundled dataset	Organism	Reference / target WIG names	Suggested model
`GSE119104_Mg_buffer`	E. coli (~146 MB)	`GSM3358138_filter_Cm_ctrl` / `GSM3358140_freeze_Mg_ctrl`	`bacteria_cm_mg`
`GSE139036_disome`	Yeast disome (~244 MB)	`GSM4127880_end3SM015Fd` / `GSM4127896_SM015M`	`yeast_disome`

The default -profile test uses the external-input path against a ~2 MB subsample of GSE119104_Mg_buffer committed to tests/data/riboformer/ (generated with scripts/subsample_mg_buffer.py — first 100 kb of NC_000913.2 + matching WIG slices). The bundled-mode path is exercised by the recipes above but not by CI.

Available pre-trained models¶

Dataset	Description
`GSE152850_yeast`	Yeast monosome/disome
`GSE139036_disome`	Disome profiling
`GSE152850_celegans`	C. elegans
`GSE119104_Mg_buffer`	E. coli
`GSE165592_trmD`	E. coli trmD
`GSE77617_miniORF`	Mini-ORF
`GSE152664_circuit`	Synthetic circuit

Example output¶

1.629764437675476074e+00
1.895173668861389160e+00
2.439188957214355469e+00
4.431646347045898438e+00
5.384204864501953125e+00

Each line is the predicted ribosome density for one codon position.