rnapy 3.2.1

Unified RNA Analysis Toolkit - ML-powered RNA sequence analysis and structure prediction

RNAPy — Unified RNA Analysis Toolkit

RNAPy is a unified Python toolkit that wraps several powerful RNA models with a consistent, easy-to-use API. It currently integrates:

• RNA-FM for sequence embeddings and 2D secondary structure prediction
• RhoFold for 3D structure prediction
• RiboDiffusion for inverse folding (sequence generation from structure)
• RhoDesign for inverse folding (structure-to-sequence, optional 2D guidance)
• RNA-MSM for MSA-based embeddings, attention, consensus, and conservation

Key Features

• Consistent high-level API via RNAToolkit
• Extract sequence embeddings (RNA-FM, mRNA-FM)
• 2D structure prediction (RNA-FM)
• 3D structure prediction (RhoFold)
• Inverse folding (RiboDiffusion, RhoDesign)
• MSA analysis and features (RNA-MSM: embeddings, attention, consensus, conservation)

Project Structure

RNAPy
├── rnapy/                    # Library source
│   ├── core/                 # Base classes, factory, config, exceptions
│   ├── providers/            # Model providers (rna_fm/mrna_fm, rhofold, RiboDiffusion, rhodesign, rna_msm)
│   ├── interfaces/           # Public interfaces
│   └── utils/                # Utilities
├── configs/                  # Global and model configs (YAML)
├── demos/                    # Ready-to-run examples
│   ├── models/               # Put pretrained weights here
│   ├── results/              # Default output location for demos
│   └── demo_*.py             # Demo scripts
├── requirements.txt
├── setup.py
└── README.md

Installation

Recommended: Python 3.12+ and a recent PyTorch build compatible with your CPU/GPU.

pip install rnapy --extra-index-url  https://download.pytorch.org/whl/cpu 

Documentation

• Toolkit usage guide: docs/RNAToolkit_Usage_Guide.md

Model Weights

• You can download pretrained weights from the original repositories which will be mentioned in the Acknowledgements section.
• Or you can find weights used in RNAPy on Hugging Face: https://huggingface.co/Linorman616/rnapy_models/
• Actually if you don't provide model-path when loading a model, RNAPy will try to download the weights from this repo automatically.

Quick Start

1) RNA-FM (2D structure + embeddings)

from rnapy import RNAToolkit

sequence = "AGAUAGUCGUGGGUUCCCUUUCUGGAGGGAGAGGGAAUUCCACGUUGACCGGGGGAACCGGCCAGGCCCGGAAGGGAGCAACCGUGCCCGGCUAUC"

# Initialize
toolkit = RNAToolkit(device="cpu")

# Load model (choose one)
toolkit.load_model("rna-fm", "./models/RNA-FM_pretrained.pth")

# 2D structure prediction
result = toolkit.predict_structure(
    sequence,
    structure_type="2d",
    model="rna-fm",
    save_dir="./results/rna_fm/demo.ct",
)

# Embeddings
embeddings = toolkit.extract_embeddings(
    sequence,
    model="rna-fm",
    save_dir="./results/rna_fm/embeddings.npy",
)

print(result.get("secondary_structure"))
print(result.get("confidence_scores"))

2) RhoFold (3D structure prediction)

from rnapy import RNAToolkit

sequence = "GGAUCCCGCGCCCCUUUCUCCCCGGUGAUCCCGCGAGCCCCGGUAAGGCCGGGUCC"

toolkit = RNAToolkit(device="cpu")

# Load RhoFold
toolkit.load_model("rhofold", "./models/RhoFold_pretrained.pt")

# Predict 3D
result = toolkit.predict_structure(
    sequence,
    structure_type="3d",
    model="rhofold",
    save_dir="./results/rhofold",
    relax_steps=500,
)

pdb_file = result.get("structure_3d_refined", result.get("structure_3d_unrelaxed"))
print("3D structure:", pdb_file)

3) RiboDiffusion (inverse folding from PDB)

from rnapy import RNAToolkit

structure_file = "./input/R1107.pdb"

toolkit = RNAToolkit(device="cpu")

# Load RiboDiffusion
toolkit.load_model("ribodiffusion", "./models/exp_inf.pth")

# Generate sequences from structure
result = toolkit.generate_sequences_from_structure(
    structure_file=structure_file,
    model="ribodiffusion",
    n_samples=2,
    sampling_steps=100,
    cond_scale=0.5,
    dynamic_threshold=True,
    save_dir="./results/ribodiffusion",
)

print("Generated count:", result.get("sequence_count", 0))
print("Output dir:", result.get("output_directory"))

4) RhoDesign (inverse folding with optional 2D guidance)

from rnapy import RNAToolkit

pdb_path = "./input/2zh6_B.pdb"
ss_path = "./input/2zh6_B.npy"  # optional numpy file with secondary-structure/contact info

toolkit = RNAToolkit(device="cpu")

# Load RhoDesign (with-2D variant checkpoint)
toolkit.load_model("rhodesign", "./models/ss_apexp_best.pth")

# Generate one sequence from structure (RhoDesign samples one sequence per call)
res = toolkit.generate_sequences_from_structure(
    structure_file=pdb_path,
    model="rhodesign",
    secondary_structure_file=ss_path,  # omit or set None to run without 2D guidance
    save_dir="./results/rhodesign"
)

print("Predicted sequence:", res["sequences"][0])
print("Recovery rate:", res.get("quality_metrics", {}).get("sequence_recovery_rate"))
print("FASTA:", res.get("files", {}).get("fasta_files", [None])[0])

5) RNA-MSM (MSA features, consensus, conservation)

from rnapy import RNAToolkit

# Initialize
toolkit = RNAToolkit(device="cpu")

# Load RNA-MSM
toolkit.load_model("rna-msm", "./models/RNA_MSM_pretrained_weights.pt")

# Prepare an example MSA (aligned sequences)
msa_sequences = [
    "AUGGCGAUUUUAUUUACCGCAGUCGUUACCAACAUACUCGACUUUAAAUGCC",
    "AUGGCAAUUUUAUUUACCGCAGUCGUUACCAACAUACUCGACUUUAAAUGCC",
    "AUGGCGAUUUCAUUUACCGCAGUCGUUACCAACAUACUCGACUUUAAAUGCC",
    "AUGGCGAUUUUAUUUACCGCAGUCGUUACCAGCAUACUCGACUUUAAAUGCC",
]

# Extract embeddings (per-position, last layer by default)
features = toolkit.extract_msa_features(
    msa_sequences,
    feature_type="embeddings",
    model="rna-msm",
    save_dir="./results/rna_msm",
)

# Analyze MSA for consensus and conservation
msa_result = toolkit.analyze_msa(
    msa_sequences,
    model="rna-msm",
    extract_consensus=True,
    extract_conservation=True,
    save_dir="./results/rna_msm",
)

print("Consensus:", msa_result.get("consensus_sequence"))
print("Conservation (first 10):", (msa_result.get("conservation_scores") or [])[:10])

Evaluation Metrics

RNAPy ships with common structural evaluation metrics, available via both the Python API and the CLI.

LDDT (Local Distance Difference Test)

• Python:

from rnapy.toolkit import RNAToolkit

toolkit = RNAToolkit(device="cpu")
res = toolkit.calculate_lddt(
    reference_structure="./demos/input/2zh6_B.pdb",
    predicted_structure="./demos/input/R1107.pdb",
    radius=15.0,
    distance_thresholds=(0.5, 1.0, 2.0, 4.0),
    return_column_scores=True,
)
print(res["lddt"])            # Global LDDT
print(res.get("columns", [])[:5])  # Optional: first 5 per-residue column scores

• CLI:

rnapy metric lddt \
  --reference ./demos/input/2zh6_B.pdb \
  --predicted ./demos/input/R1107.pdb \
  --radius 15.0 \
  --thresholds 0.5,1.0,2.0,4.0 \
  --return-column-scores

Example script: demos/demo_lddt.py

RMSD (Root Mean Square Deviation)

• Python:

from rnapy.toolkit import RNAToolkit

toolkit = RNAToolkit()
rmsd = toolkit.calculate_rmsd(
    "./demos/input/rmsd_tests/resources/ci2_1.pdb",
    "./demos/input/rmsd_tests/resources/ci2_2.pdb",
    file_format="pdb",
)
print("RMSD:", rmsd)

• CLI (common flags only; see rnapy metric rmsd --help for details):

rnapy metric rmsd \
  --file1 ./demos/input/rmsd_tests/resources/ci2_1.pdb \
  --file2 ./demos/input/rmsd_tests/resources/ci2_2.pdb \
  --file-format pdb \
  --rotation kabsch

Other options include: --reorder, --reorder-method inertia-hungarian, --use-reflections, --only-alpha-carbons, --ignore-hydrogen, --output-aligned-structure, --print-only-rmsd-atoms, --gzip-format, etc.

Example script: demos/demo_rmsd.py

TM-score

• Python:

from rnapy.toolkit import RNAToolkit

toolkit = RNAToolkit(device="cpu")
result = toolkit.calculate_tm_score(
    structure_1="./demos/input/2zh6_B.pdb",
    structure_2="./demos/input/R1107.pdb",
    mol="rna",
)
print(result["raw_output"])     # Raw TM-score tool output
print(result["tm_score_1"])     # TM-score normalized by length 1
print(result["tm_score_2"])     # TM-score normalized by length 2

• CLI:

rnapy metric tm-score \
  --struct1 ./demos/input/2zh6_B.pdb \
  --struct2 ./demos/input/R1107.pdb \
  --mol rna

Example script: demos/demo_tm_score.py

Sequence Recovery & Structure F1

Sequence recovery and secondary-structure F1 are common quality metrics for design and prediction.

• Python:

from rnapy import RNAToolkit

toolkit = RNAToolkit()

# Structure F1 (dot-bracket)
f1 = toolkit.calculate_structure_f1("(((...)))", "(((.....)))")
print(f1)  # {precision, recall, f1_score}

# Sequence recovery rate
recovery = toolkit.calculate_sequence_recovery("AUGCUAGCUAGC", "AUGCUAGCUUGC")
print(recovery["overall_recovery"])  # overall recovery

• CLI:

# Structure F1
rnapy struct f1 \
  --struct1 "(((...)))" \
  --struct2 "(((.....)))"

# Sequence recovery
rnapy seq recovery \
  --native  AUGCUAGCUAGC \
  --designed AUGCUAGCUUGC

Example script: demos/demo_f1_recovery.py

Command Line Interface (CLI)

The package installs a console script named rnapy (via setup entry point). After installation, you can run rnapy from your shell.

• Show top-level help:
  • rnapy --help
• Show help for a subcommand:
  • rnapy seq embed --help

Global options

These options are shared by all subcommands:

• --device {cpu,cuda}: Computing device (default: cpu)
• --model {rna-fm,mrna-fm,rhofold,ribodiffusion,rhodesign,rna-msm}: Model provider (required)
• --model-path PATH: Path to the model checkpoint (required)
• --config-dir PATH: Configuration directory (default: configs)
• --provider-config PATH: Optional provider-specific config file
• --seed INT: Random seed
• --save-dir DIR: Output directory
• --verbose or -v: Verbose logs and full tracebacks on errors

Input conventions:

• Use exactly one of --seq or --fasta
  • --seq accepts a single RNA sequence or multiple sequences separated by commas
  • --fasta accepts a .fasta/.fa/.fas file path

Subcommands

1. Sequence embeddings

Extract embeddings from RNA-FM/mRNA-FM:

rnapy seq embed \
  --model rna-fm \
  --model-path ./models/RNA-FM_pretrained.pth \
  --seq "AGAUAGUCGUGGGU...UCGGCUAUC" \
  --layer -1 \
  --format mean \
  --save-dir ./results/rna_fm

• --layer: which layer to use (default: -1, i.e., last layer)
• --format {raw,mean,bos}: output format (default: mean)
• You can also pass --fasta path/to/input.fasta instead of --seq

2. Structure prediction

Predict 2D RNA-FM or 3D (RhoFold) structure:

# 2D with mRNA-FM
rnapy struct predict \
  --model rna-fm \
  --model-path ./models/RNA-FM_pretrained.pth \
  --seq "AGAUAGUCGUGGGU...UCGGCUAUC" \
  --structure-type 2d \
  --save-dir ./results/rna_fm_struct

# 3D with RhoFold (structure-type will auto-infer to 3d)
rnapy struct predict \
  --model rhofold \
  --model-path ./models/RhoFold_pretrained.pt \
  --seq "GGAUCCCGCGCCC...GCCGGGUCC" \
  --save-dir ./results/rhofold_3d

• If --structure-type is omitted: rhofold -> 3d; rna-fm/mrna-fm -> 2d

3. Inverse folding (generate sequences from structure)

RiboDiffusion and RhoDesign take a PDB as input:

# RiboDiffusion: generate multiple sequences
rnapy invfold gen \
  --model ribodiffusion \
  --model-path ./models/exp_inf.pth \
  --pdb ./input/R1107.pdb \
  --n-samples 2 \
  --save-dir ./results/ribodiffusion

# RhoDesign: optional 2D guidance via NPY
rnapy invfold gen \
  --model rhodesign \
  --model-path ./models/ss_apexp_best.pth \
  --pdb ./input/2zh6_B.pdb \
  --ss-npy ./input/2zh6_B.npy \
  --save-dir ./results/rhodesign

• --pdb: required
• --ss-npy: optional; only used by RhoDesign (2D guidance)
• --n-samples: number of sequences to sample (RhoDesign samples one per call; RiboDiffusion supports many)

4. MSA features (RNA-MSM)

Extract embeddings/attention from an aligned MSA:

rnapy msa features \
  --model rna-msm \
  --model-path ./models/RNA_MSM_pretrained_weights.pt \
  --fasta ./input/example_msa.fasta \
  --feature-type embeddings \
  --layer -1 \
  --save-dir ./results/rna_msm_features

• --feature-type {embeddings,attention,both} (default: embeddings)
• --layer: which layer to extract (default: -1)

5. MSA analysis (RNA-MSM)

Compute consensus and/or conservation from an MSA:

rnapy msa analyze \
  --model rna-msm \
  --model-path ./models/RNA_MSM_pretrained_weights.pt \
  --fasta ./input/example_msa.fasta \
  --extract-consensus \
  --extract-conservation \
  --save-dir ./results/rna_msm_analyze

• If you pass a single --seq (not multiple), this subcommand will error because it requires multiple sequences or a FASTA file

6. Metrics (structure evaluation)

• LDDT: see examples above, or run rnapy metric lddt --help
• RMSD: see examples above, or run rnapy metric rmsd --help
• TM-score: see examples above, or run rnapy metric tm-score --help

7. Sequence utilities

• Structure F1: rnapy struct f1 --struct1 ... --struct2 ...
• Sequence recovery: rnapy seq recovery --native ... --designed ...

Outputs and logging

• When --save-dir is provided, results are written under that directory. The exact filenames depend on the provider/task (e.g., .npy for embeddings, .ct for 2D, .pdb/folder for 3D, .json for analysis summaries). The CLI prints a brief summary and (when applicable) a path hint.
• Exit codes: 0 on success; non-zero on errors. Add -v/--verbose for full tracebacks.

Common pitfalls

• Do not pass both --seq and --fasta at the same time.
• Ensure the --model-path points to the correct checkpoint for the chosen --model.
• rhofold defaults to 3D; RNA-FM/mRNA-FM default to 2D if --structure-type is omitted.
• msa analyze requires multiple sequences (comma-separated via --seq) or a FASTA file.

Run the Demos

From the repository root:

# mRNA-FM / RNA-FM demo
cd .\demos
python .\demo_rna_fm.py

# RhoFold demo
python .\demo_rhofold.py

# RiboDiffusion demo
python .\demo_ribodiffusion.py

# RhoDesign demo
python .\demo_rhodesign.py

# RNA-MSM demo
python .\demo_rna_msm.py

# LDDT demo
python .\demo_lddt.py

# RMSD demo
python .\demo_rmsd.py

# TM-score demo
python .\demo_tm_score.py

# Sequence recovery & Structure F1 demo
python .\demo_f1_recovery.py

Additional examples may be available: rna_fm_demo.py, rhofold_demo.py, ribodiffusion_demo.py.

Datasets

You can download example datasets via API or CLI (e.g., Rfam, RNA Puzzles, CASP15, etc.).

• Available dataset names: Rfam, Rfam_original, RNA_Puzzles, CASP15, RNAsolo2

• CLI:

# List available datasets
rnapy dataset list

# Download Rfam (from the HF mirror) with parallel workers
rnapy dataset download --dataset Rfam --max-workers 8

• Python：

from rnapy.toolkit import RNAToolkit

toolkit = RNAToolkit()
print(toolkit.list_available_datasets())
toolkit.download_dataset("Rfam", max_workers=8)

Configuration

YAML configs are provided under ./configs/ and ./demos/configs/. You can:

• Pass config_dir to RNAToolkit to use custom defaults
• Override per-call parameters in load_model(...) and task methods

License

MIT License

Acknowledgements

• RNA-FM: https://github.com/ml4bio/RNA-FM
• RhoFold: https://github.com/ml4bio/RhoFold
• RiboDiffusion: https://github.com/ml4bio/RiboDiffusion
• RhoDesign: https://github.com/ml4bio/RhoDesign
• RNA-MSM: https://github.com/yikunpku/RNA-MSM