Skip to content

RiboNN

Predict translation efficiency (TE) from mRNA sequence across 82 human cell types/tissues.

  • Paper: Nature Biotechnology 2025
  • Upstream: https://github.com/Sanofi-Public/RiboNN
  • License: Apache 2.0
  • Device: CPU or GPU (CPU works, GPU faster)

What it does

RiboNN is a multi-task neural network that predicts translation efficiency from mRNA sequence (5'UTR + CDS + 3'UTR). It outputs TE predictions for 82 human (or mouse) cell types/tissues simultaneously.

Input format

Tab-separated text file with 4 columns (no header):

tx_id    utr5_sequence    cds_sequence    utr3_sequence

Example (tests/data/ribonn_prediction_input.txt): 

ENST00000215375.7   AGACGTCCCTGCGCGTCGTCCTCCTCGCCCTCCAGGCCGCCCGCGCCGCGCCGGAGTCCGCTGTCCGCCAGCTACCCGCTTCCTGCCGCCCGCCGCTGCC    ATGCTGCCCGCCGCGCTGCTCCGCCGCCCGGG... GCGGTGCGTACCCGGTGTCCCGAGGCCCGGCC...

Transcripts with combined CDS + 3'UTR length exceeding 11,937 nt are automatically removed.

Output format

Tab-separated text file (prediction_output.txt) with columns:

tx_id  utr5_sequence  cds_sequence  utr3_sequence  predicted_TE_108T  predicted_TE_12T  ...  mean_predicted_TE

Each predicted_TE_* column is the predicted log2 translation efficiency for that cell type. The final column is the mean across all cell types.

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.

docker run --rm \
  -v /path/to/input.txt:/app/data/prediction_input1.txt \
  -v /path/to/output:/out \
  ghcr.io/ericmalekos/rnazoo-ribonn-cpu:latest \
  bash -c "cd /app && python3 /opt/bin/ribonn_predict.py -i /app/data/prediction_input1.txt -o /out --species human"

For mouse: replace --species human with --species mouse.

To predict using a fine-tuned checkpoint:

docker run --rm \
  -v /path/to/input.txt:/app/data/prediction_input1.txt \
  -v /path/to/checkpoint.ckpt:/app/checkpoint.ckpt \
  -v /path/to/output:/out \
  ghcr.io/ericmalekos/rnazoo-ribonn-cpu:latest \
  bash -c "cd /app && python3 /opt/bin/ribonn_predict.py -i /app/data/prediction_input1.txt -o /out --checkpoint /app/checkpoint.ckpt --target TE_MyCondition"

Run with Nextflow

nextflow run main.nf -profile docker,cpu \
  --ribonn_input /path/to/input.txt

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

Results appear in results/ribonn/ribonn_out/prediction_output.txt.

Predict with a fine-tuned checkpoint

After fine-tuning (see below), use the saved checkpoint for prediction on new data:

nextflow run main.nf -profile docker,cpu \
  --ribonn_input /path/to/input.txt \
  --ribonn_checkpoint results/ribonn_finetune/fold0/*.ckpt \
  --ribonn_finetune_target TE_MyCondition

Parameters

Parameter Default Description
--ribonn_input null Tab-separated input file
--ribonn_species human Species for pretrained models (human or mouse)
--ribonn_checkpoint null Path to a fine-tuned .ckpt checkpoint for prediction

Example output (truncated)

tx_id   predicted_TE_HEK293 predicted_TE_HeLa   ... mean_predicted_TE
ENST00000215375.7   1.0705105   1.0321615   ... 1.0564895
ENST00000231004.5   0.47455412  0.41119576  ... 0.54676294

Fine-tuning on your own data

RiboNN supports transfer learning from the pretrained 78-cell-type human model to a new cell type or condition. The pretrained convolutional layers are frozen initially, then the full model is fine-tuned with a lower learning rate.

Input format

Tab-separated file with columns: tx_id, utr5_sequence, cds_sequence, utr3_sequence, and your target TE column (numeric values). The target column name is specified via --ribonn_finetune_target.

Run fine-tuning

nextflow run main.nf -profile docker,cpu \
  --ribonn_finetune_input my_training_data.tsv \
  --ribonn_finetune_target TE_MyCondition

Parameters

Parameter Default Description
--ribonn_finetune_target (required) Column name containing TE values
--ribonn_finetune_phase1_epochs 50 Epochs for head-only training (conv layers frozen)
--ribonn_finetune_phase2_epochs 150 Epochs for full model training (all layers)
--ribonn_finetune_patience 50 Early stopping patience
--ribonn_finetune_folds 5 Cross-validation folds

Output

  • ribonn_finetune/predictions.tsv — cross-validated predictions on held-out test folds
  • ribonn_finetune/fold*/ — saved model checkpoints per fold