Saluki

Predict mRNA half-life from sequence.

• Paper: Genome Biology 2022
• Upstream: https://github.com/calico/basenji (Saluki model)
• License: Apache 2.0 (weights from Zenodo)
• Device: CPU or GPU (50-member ensemble — GPU helps for large inputs). Two image variants:
  • rnazoo-saluki:latest — CUDA-enabled (default, used with -profile gpu)
  • rnazoo-saluki-cpu:latest — CPU-only (smaller, used with -profile cpu)

What it does

Saluki predicts mRNA half-life from sequence using a convolutional neural network. It takes a single mRNA sequence (with UTR/CDS annotation via case convention) and outputs a log2(half-life) prediction. The model uses an ensemble of 50 cross-validated models (10 folds x 5 replicates) and supports both human and mouse.

Input format

FASTA file where lowercase = UTR and UPPERCASE = CDS:

>test_transcript_1
gcgccgagcggcgccccgctgccctgtccccgcgtgcagaccccgggcccggccccggccATGCTGTGCGGCCGCTGG
AGGCGTTGCCGCCGCCCCGCCCGAGGAGCCCCCGGTGGCCGCCCAGGTCGCAGCCCAAGTCGCGGCGCCGGTCGCTCT
CCCGTCCCCGCCGACTCCCTCCGATGGCGGCACCAAGAGGCCCGGGCTGCGGGCGCTGAAGAAGATGGGCCTGACGGAG
...CACTCCAACAAcgaccgtctagagggggctgagatcgaggagttcctgcggcggctgctgaagcggccggag

• 5'UTR: lowercase at the start
• CDS: UPPERCASE in the middle
• 3'UTR: lowercase at the end

Output format

NumPy file (preds.npy) containing an array of log2(mRNA half-life) predictions.

Shape: (N, 1, 50) where N = number of transcripts and 50 = ensemble members.

To read the output:

import numpy as np
preds = np.load("preds.npy")
mean_pred = preds.mean(axis=-1)  # average across ensemble
print(f"Predicted log2(half-life): {mean_pred}")

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.

# CPU
docker run --rm \
  -v /path/to/input.fa:/data/input.fa \
  -v /path/to/output:/out \
  ghcr.io/ericmalekos/rnazoo-saluki-cpu:latest \
  bash -c "saluki_predict_fasta.py -d 0 -o /out /opt/saluki_models /data/input.fa"

# GPU
docker run --rm --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=all \
  -v /path/to/input.fa:/data/input.fa \
  -v /path/to/output:/out \
  ghcr.io/ericmalekos/rnazoo-saluki:latest \
  bash -c "saluki_predict_fasta.py -d 0 -o /out /opt/saluki_models /data/input.fa"

• -d 0 = human, -d 1 = mouse

Run with Nextflow

# CPU
nextflow run main.nf -profile docker,cpu \
  --saluki_input /path/to/input.fa --saluki_species human

# GPU
nextflow run main.nf -profile docker,gpu \
  --saluki_input /path/to/input.fa --saluki_species human

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

Parameters

Parameter	Default	Description
--saluki_species	human	Species: human or mouse

Example output

For a single transcript, the 50-member ensemble produces:

Shape: (1, 1, 50)
Predictions (log2 half-life): [[[-0.695 -1.680 -1.501 -1.245 -1.145 -0.774 ... ]]]

The mean across the ensemble gives the final predicted log2(half-life). Negative values indicate short-lived transcripts; positive values indicate stable transcripts.