HydraRNA

Track: RNA Foundation Models
Paper: Li et al., Genome Biology 2025 — doi:10.1186/s13059-025-03853-7
Upstream repo: https://github.com/GuipengLi/HydraRNA
License: MIT (code + weights)
GPU required: Yes (Mamba CUDA selective-scan kernel; no CPU fallback)

What it does

HydraRNA is a hybrid bidirectional-SSM + multi-head attention RNA language model trained on full-length transcripts (up to 10,000 nt). The model stack alternates between Hydra (bidirectional Mamba) blocks and MHA blocks (12 layers total), producing per-token and mean-pooled sequence embeddings (1024-d). It outperforms comparable models on mRNA-related benchmarks (MRL, half-life, TE) despite fewer parameters, attributed to the full-length input window and bidirectional SSM context.

The zoo ships the base model (HydraRNA_model.pt, 337 MB from Zenodo 10.5281/zenodo.20481998). The secondary structure fine-tuned variant (HydraRNA_model_V2.pt + HydraRNA_SS_model.pt) is not currently wired.

Input

A FASTA file of RNA sequences (A/C/G/U or A/C/G/T). Sequences should ideally be complete mature transcripts; partial or very short sequences will produce out-of-distribution embeddings. No hard length limit is enforced — sequences longer than 10,240 nt are split into 10,240-nt chunks and the final embedding is the mean over chunk embeddings.

Example test sequence (tests/data/test_rna.fa works):

>my_transcript
AUGCUAGCUAGCUAGCUA...

Output

File	Description
sequence_embeddings.npy	(N, 1024) float32 — one row per sequence
labels.txt	FASTA headers, one per line, matching row order
{header}_tokens.npy	(L, 1024) per-token embeddings (optional; first 10,240 nt only for very long sequences)

Docker run

docker run --rm --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=all \
  -v /path/to/data:/data \
  ghcr.io/ericmalekos/rnazoo-hydrarna:latest \
  hydrarna_predict.py \
    -i /data/sequences.fa \
    -o /data/hydrarna_out

With per-token embeddings:

docker run --rm --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=all \
  -v /path/to/data:/data \
  ghcr.io/ericmalekos/rnazoo-hydrarna:latest \
  hydrarna_predict.py \
    -i /data/sequences.fa \
    -o /data/hydrarna_out \
    --per-token

Nextflow run

nextflow run ericmalekos/RNA-Zoo \
  --hydrarna_input sequences.fa \
  --outdir results \
  -profile docker,gpu

Linear-probe fine-tuning

nextflow run ericmalekos/RNA-Zoo \
  --hydrarna_finetune_input labeled.tsv \
  --hydrarna_finetune_label rl \
  --outdir results \
  -profile docker,gpu

From precomputed embeddings (no GPU needed for head training):

nextflow run ericmalekos/RNA-Zoo \
  --hydrarna_finetune_input labeled.tsv \
  --hydrarna_finetune_label rl \
  --hydrarna_finetune_embeddings sequence_embeddings.npy \
  --outdir results \
  -profile docker,cpu

Parameters

Parameter	Default	Description
--hydrarna_input	null	Path to FASTA; activates the HYDRARNA process
--hydrarna_outdir	<outdir>/hydrarna	Output directory
--hydrarna_per_token	false	Emit per-token .npy files alongside sequence embeddings
--hydrarna_max_len	10000	Warn (don't fail) for sequences exceeding this length
--hydrarna_finetune_input	null	TSV/CSV with name, sequence, and a label column
--hydrarna_finetune_label	null	Column name in the fine-tune TSV to use as the label
--hydrarna_finetune_outdir	<outdir>/hydrarna_finetune	Fine-tune output directory
--hydrarna_finetune_epochs	20	Training epochs for the probe head
--hydrarna_finetune_lr	1e-3	Learning rate for the probe head
--hydrarna_finetune_embeddings	null	Precomputed (N, 1024) .npy; skips embedding extraction
--hydrarna_finetune_head_type	linear	Head architecture: linear, mlp, or xgboost
--hydrarna_finetune_task	auto	auto, regression, or classification

Implementation notes

• The model uses vendored fairseq (in the upstream repo's fairseq/ directory). The Dockerfile installs it as an editable package alongside the main inference code.
• Tokenization: individual characters separated by spaces, prefixed with <s>. T→U conversion is applied before tokenization (model vocabulary is ACGU).
• Embeddings are extracted via model.encoder.extract_features(src_tokens=...); BOS and EOS tokens are stripped before mean-pooling.
• The model is run in float16 for inference.
• No CPU variant is provided — the Mamba SSM kernel requires CUDA.