ODesignBench

ODesignBench is a multimodal benchmark toolkit for structure-based biomolecular design. It standardizes data contracts, inverse-folding/refolding workflows, and evaluation metrics across multiple design settings.

Note: Additional modules are still being updated and will be released in upcoming updates.

The repository is designed to evaluate generated structures from external models under a consistent protocol, so different generators can be compared fairly with shared preprocessing, folding, and metric pipelines.

Environment Setup

This project integrates multiple heavy-weight structural biology models (Chai-1, ESM, Protenix, and AlphaFold3). Due to complex C++ dependencies, environment setup is supported via Conda.

A unified environment.yml is provided in the repository root. It combines the RosettaCommons channel, NVIDIA CUDA packages, and the PyTorch ecosystem to match the CUDA and runtime assumptions used by the benchmark pipelines.

Installation Steps

Clone the repository:

git clone https://github.com/OTeam-AI4S/ODesignBench.git
cd ODesignBench

Create the Conda environment:

conda env create -f environment.yml

Activate the environment:

conda activate designbench

Notes

fair-esm==2.0.0 is used and is compatible with modern Python 3.10 and PyTorch 2.x.
openbabel is included in environment.yml for ligand reconstruction/docking metrics (mol_rec, rdkit_utils, docking_vina).
environment.yml also includes the extra ligand-evaluation stack used by run_pbl_pipeline.py, including EFGs, meeko==0.1.dev3, pdb2pqr, vina, and AutoDockTools_py3.
gRNAde requires PyTorch Geometric compiled extensions. Install them after conda activate designbench with the PyG wheel index matching your torch/cuda versions (see command below).
If your system has multiple CUDA installations, verify runtime visibility with nvidia-smi and python -c "import torch; print(torch.cuda.is_available())".

Install gRNAde / PyG Runtime (Required for RNA pipeline)

For the default environment in this repo (torch==2.5.1, CUDA 12.1), run:

conda activate designbench
pip install --upgrade --no-cache-dir \
  pyg_lib torch-scatter torch-sparse torch-cluster torch-spline-conv \
  -f https://data.pyg.org/whl/torch-2.5.1+cu121.html

Verifying Ligand Docking Dependencies

After environment creation, it is helpful to verify the docking stack before launching a large PBL run:

conda activate designbench
python -c "import AutoDockTools, vina, meeko, easydict, EFGs; print('PBL docking dependencies are available')"
pdb2pqr30 --help

If your environment was created before these dependencies were added to environment.yml, recreate the environment or install the missing packages into the existing designbench environment.

Inversefold Model Parameters (ProteinMPNN / LigandMPNN)

Pipelines that use inversefold=ProteinMPNN require the ProteinMPNN checkpoint at:

inversefold/LigandMPNN/model_params/proteinmpnn_v_48_020.pt

If this file is missing, you may see:

FileNotFoundError: ProteinMPNN checkpoint not found: .../proteinmpnn_v_48_020.pt

Download model params into the expected repo-local directory:

bash inversefold/LigandMPNN/get_model_params.sh inversefold/LigandMPNN/model_params

By default, the config reads:

PROTEINMPNN_CHECKPOINT_PATH (environment variable), or
inversefold.checkpoint_path (Hydra override, defaulting to the path above)

Optional explicit override examples:

export PROTEINMPNN_CHECKPOINT_PATH=/absolute/path/to/proteinmpnn_v_48_020.pt
python scripts/run_pbp_pipeline.py design_dir=/path/to/pbp_designs gpus=0

python scripts/run_pbp_pipeline.py \
  design_dir=/path/to/pbp_designs \
  gpus=0 \
  inversefold.checkpoint_path=/absolute/path/to/proteinmpnn_v_48_020.pt

OInvFold Checkpoints

nbl and pbn now use OInvFold inside ODesignBench, so inputs can be direct ODesign design outputs (no pre-applied inversefold required).

Expected checkpoint filenames under OINVFOLD_CKPT_ROOT (default: ./ckpt):

oinvfold_protein.ckpt
oinvfold_ligand.ckpt
oinvfold_dna.ckpt
oinvfold_rna.ckpt

Download example:

mkdir -p ckpt
wget -c -P ckpt -O ckpt/oinvfold_protein.ckpt "https://huggingface.co/The-Institute-for-AI-Molecular-Design/OInvFold/resolve/main/oinvfold_protein.ckpt"
wget -c -P ckpt -O ckpt/oinvfold_ligand.ckpt "https://huggingface.co/The-Institute-for-AI-Molecular-Design/OInvFold/resolve/main/oinvfold_ligand.ckpt"
wget -c -P ckpt -O ckpt/oinvfold_dna.ckpt "https://huggingface.co/The-Institute-for-AI-Molecular-Design/OInvFold/resolve/main/oinvfold_dna.ckpt"
wget -c -P ckpt -O ckpt/oinvfold_rna.ckpt "https://huggingface.co/The-Institute-for-AI-Molecular-Design/OInvFold/resolve/main/oinvfold_rna.ckpt"

If checkpoints are stored elsewhere:

export OINVFOLD_CKPT_ROOT=/absolute/path/to/ckpt

ESMFold Weights

To run ESMFold for refolding, you need to download the ESMFold model weights from Hugging Face. The scripts expect the weights to be located in refold/esmfold/weights.

Download using the Python API (recommended):

# Make sure you are in the ODesignBench directory
mkdir -p refold/esmfold/weights
python -c "from huggingface_hub import snapshot_download; snapshot_download('facebook/esmfold_v1', local_dir='refold/esmfold/weights')"

Or manually download the weights from Hugging Face.

Chai-1 Weights

By default, the chai_lab library attempts to download its model weights and ESM weights automatically during the first run. To bypass this, you may download the weights before running ODesignBench pipeline:

Method 1: Fast Download via `aria2` (Recommended for unstable networks)

You can use a multi-threaded download tool like aria2 to download all weights directly from their source URLs. This is much faster and more reliable than the default Python script.

# 1. Define where you want to store the weights
export CHAI_DOWNLOADS_DIR=$(pwd)/refold/chai1/weights
mkdir -p $CHAI_DOWNLOADS_DIR/models_v2
mkdir -p $CHAI_DOWNLOADS_DIR/esm

# 2. Download Chai-1 main components
for comp in feature_embedding.pt token_embedder.pt trunk.pt diffusion_module.pt confidence_head.pt; do
    aria2c -x 16 -s 16 -d $CHAI_DOWNLOADS_DIR/models_v2 -o $comp "https://chaiassets.com/chai1-inference-depencencies/models_v2/$comp"
done

# 3. Download Conformers
aria2c -x 16 -s 16 -d $CHAI_DOWNLOADS_DIR -o conformers_v1.apkl "https://chaiassets.com/chai1-inference-depencencies/conformers_v1.apkl"

# 4. Download ESM weights
aria2c -x 16 -s 16 -d $CHAI_DOWNLOADS_DIR/esm -o traced_sdpa_esm2_t36_3B_UR50D_fp16.pt "https://chaiassets.com/chai1-inference-depencencies/esm2/traced_sdpa_esm2_t36_3B_UR50D_fp16.pt"

Method 2: Pre-download via Python script

Run the following Python script:

import os
from chai_lab.utils.paths import chai1_component, cached_conformers

# Define where you want to store the weights
download_dir = os.path.abspath("./refold/chai1/weights")
os.environ["CHAI_DOWNLOADS_DIR"] = download_dir
os.makedirs(download_dir, exist_ok=True)

components = [
    "feature_embedding.pt", 
    "token_embedder.pt", 
    "trunk.pt", 
    "diffusion_module.pt", 
    "confidence_head.pt"
]

print(f"Downloading Chai-1 weights to {download_dir}...")
for comp in components:
    print(f"Fetching {comp}...")
    chai1_component(comp)

print("Fetching conformers_v1.apkl...")
cached_conformers.get_path()

print("Fetching ESM weights...")
from chai_lab.data.dataset.embeddings.esm import ESM_URL, esm_cache_folder
from chai_lab.utils.paths import download_if_not_exists
local_esm_path = esm_cache_folder.joinpath("traced_sdpa_esm2_t36_3B_UR50D_fp16.pt")
download_if_not_exists(ESM_URL, local_esm_path)
    
print("✅ Chai-1 and ESM weights successfully downloaded!")

Next Step: Set the Environment Variable

Regardless of whether you used Method 1 or Method 2, before running the ODesignBench pipeline on your compute node, export the CHAI_DOWNLOADS_DIR variable to tell the pipeline where the offline weights are located:

# Replace with your actual absolute path
export CHAI_DOWNLOADS_DIR="/path/to/ODesignBench/refold/chai1/weights"

# Now you can safely run the pipeline without auto-download
python3 scripts/run_ame_pipeline.py design_dir=examples/tip_atom_scaffolding/ gpus=0

AlphaFold3 Setup

Some benchmark tasks use refold=af3 and run AlphaFold3 through a wrapper script. Default wrapper: refold/run_af3.sh (Docker). HPC wrapper: refold/run_af3_singularity.sh (Singularity/Apptainer).

Download required artifacts

Follow the official AlphaFold3 instructions in the upstream repo to download:

AF3 model parameters (see: Obtaining model parameters)
AF3 public databases (see the database download instructions in the upstream AlphaFold3 repo: https://github.com/google-deepmind/alphafold3).

Directory layout expected by `refold/run_af3.sh`

The wrapper expects:

$AF3_BASE/models -> mounted to /root/models inside the container
$AF3_PUBLIC_DB -> mounted to /root/public_databases inside the container

Configure environment variables

Set these before running any pipeline that uses AF3 refolding:

export AF3_BASE=/path/to/af3              # must contain: $AF3_BASE/models
export AF3_PUBLIC_DB=/path/to/public_databases
# Optional: if your docker image tag differs
export AF3_DOCKER_IMAGE=alphafold3

Singularity/Apptainer mode

Singularity/Apptainer is commonly allowed in HPC and can run containerized AF3 workloads without Docker daemon privileges. If your cluster does not support Docker on compute nodes, switch AF3 execution to:

export AF3_EXEC=/absolute/path/to/ODesignBench/refold/run_af3_singularity.sh
export AF3_SIF_IMAGE=/absolute/path/to/alphafold3.sif
export AF3_BASE=/path/to/af3              # must contain: $AF3_BASE/models
export AF3_PUBLIC_DB=/path/to/public_databases

run_af3_singularity.sh accepts both singularity and apptainer commands. If your command name differs, ensure it is available in PATH.

Note on PBP target MSA injection: PBP tasks inject pre-computed MSA for the target chain via a runtime patch (AF3_DIALECT_PATCH=true, default). The assets are expected at /assets. Set AF3_DIALECT_PATCH=false to disable the patch.

Run

After exporting the variables above, run the normal pipeline commands (e.g. scripts/run_pbp_pipeline.py with refold=af3).

Benchmark Content

Nucleic Acid Monomer (DNA / RNA)

run_rna_pipeline.py: RNA free generation (gRNAde inversefold)

run_dna_pipeline.py: DNA free generation (OInvFold inversefold)

Input: nucleic-acid structure files (.cif/.pdb) in design_dir
Inverse fold:
- RNA free generation: generate 8 RNA sequences per backbone using gRNAde
- DNA free generation: generate 8 DNA sequences per backbone using OInvFold
Config:
- RNA: config_rna.yaml + inversefold: gRNAde_rna
- DNA: config_dna.yaml + inversefold: OInvFold (inversefold.data_name=dna, inversefold.oinvfold_topk=8)
Refold: AlphaFold3
Evaluation: C4' RMSD and TM-score

Minimal run commands:

python scripts/run_rna_pipeline.py \
  design_dir=/path/to/rna_designs \
  gpus=0

python scripts/run_dna_pipeline.py \
  design_dir=/path/to/dna_designs \
  gpus=0

Optional: set root=/path/to/output_dir to change output location (default: results).

gRNAde Checkpoint Download

This repo's bundled inversefold/gRNAde code expects the legacy-compatible checkpoint below:

cd /path/to/ODesignBench
mkdir -p inversefold/gRNAde/checkpoints

# Recommended checkpoint for this repo
aria2c -x 16 -s 16 -k 1M \
  -d inversefold/gRNAde/checkpoints \
  -o gRNAde_ARv1_1state_all.h5 \
  "https://huggingface.co/genbio-ai/AIDO.RNAIF-1.6B/resolve/main/other_models/gRNAde_ARv1_1state_all.h5?download=true"

For RNA without precomputed MSA in AF3 input JSON, set refold.run_data_pipeline=true to let AF3 run MSA/template search during refold. Keep global default run_data_pipeline=false for other tasks.

Atomic Motif Scaffolding / Enzyme Design (AME)

AME evaluates the scaffolding of atomic motifs, which are crucial for enzyme design and small molecule binding:

Input: scaffold PDB files + ame_info.csv (or ame.csv)
Inverse fold: LigandMPNN (motif-constrained design)
Refold: Chai-1
Evaluation: catalytic heavy atom RMSD, ligand clash count, overall success rate

Minimal run command:

python scripts/run_ame_pipeline.py \
  design_dir=/path/to/ame_scaffolds \
  gpus=0

Optional: set root=/path/to/output_dir to change output location (default: results).

ame.csv (or ame_input.csv) should include 3 columns (no header required):

id: The filename of the scaffold (e.g., m0024_1nzy_seed_46_bb_9_seq_0-1.pdb)
task: The AME task name (one of the 41 standard tasks, e.g., M0024_1nzy)
motif_residues: Comma-separated list of motif residues to keep fixed (e.g., "A114,A137,A145,A64,A86,A90")

Example:

m0024_1nzy_seed_46_bb_9_seq_0-1.pdb,M0024_1nzy,"A114,A137,A145,A64,A86,A90"

Protein Binding Protein (PBP)

PBP evaluates designed protein-protein complexes with a fixed interface role definition:

Input: complex PDB files in design_dir
Required metadata: pbp_info.csv in design_dir
Inverse fold: LigandMPNN
Refold: AlphaFold3 (sequence-only)
Evaluation: interface and structure quality metrics in the benchmark pipeline

Minimal run command:

python scripts/run_pbp_pipeline.py \
  design_dir=/path/to/pbp_designs \
  gpus=0

Optional: set root=/path/to/output_dir to change output location (default: results).

pbp_info.csv must provide one row per design with at least:

design_name
design_chain
target_chain

Example:

design_name,target_chain,design_chain
CD3d_0,B,A
CD3d_1,B,A

Ligand Binding Protein (LBP)

LBP evaluates ligand-binding protein designs, where the ligand is retained during inverse folding and the designed protein is refolded with Chai-1.

Input: ligand-containing design structures (.cif recommended)
Required metadata: lbp_info.csv in design_dir
Recommended layout: nested case folders such as design_dir/FAD/FAD_seed_1_bb_0_seq_0.cif
Inverse fold: LigandMPNN
Refold: Chai-1
Evaluation: plddt, ipae, min_ipae, iptm, ptm_binder, plus Foldseek-based diversity/novelty

LBP uses the CCD components database during ligand handling. By default the repo expects:

preprocess/ccd_component/components.cif

If this file is missing, download it from wwPDB and place it there:

cd /path/to/ODesignBench/preprocess/ccd_component
wget https://files.wwpdb.org/pub/pdb/data/monomers/components.cif.gz
gunzip -f components.cif.gz

Minimal run command:

python scripts/run_lbp_pipeline.py \
  design_dir=/path/to/ligand_binding_protein_designs \
  gpus=0

Optional: set root=/path/to/output_dir to change output location (default: results).

lbp_info.csv must provide one row per design with at least:

design_name
target_chain
design_chain

Example:

design_name,target_chain,design_chain
FAD_seed_1_bb_0_seq_0,A,B

Recommended example layout:

examples/ligand_binding_protein/
|-- lbp_info.csv
`-- FAD/
    `-- FAD_seed_1_bb_0_seq_0.cif

Example run:

python3 scripts/run_lbp_pipeline.py \
  design_dir=examples/ligand_binding_protein \
  gpus=0 \
  root=results/examples/ligand_binding_protein

Successful runs will write the final evaluation table to raw_data.csv.

Interface Design

Interface design is the pocket-constrained variant of ligand-binding protein design used in the ODesign paper. It reuses the same protein-ligand inputs as LBP, but only redesigns protein residues within 3.5A of the ligand.

Input: ligand-containing design structures (.cif recommended)
Required metadata: lbp_info.csv in design_dir
Recommended layout: nested case folders such as design_dir/FAD/FAD_seed_1_bb_0_seq_0.cif
Pocket definition: protein residues within 3.5A of any non-water ligand atom
Inverse fold: LigandMPNN with pocket-only redesign
Refold: Chai-1
Evaluation: sc_rmsd, pocket_rmsd, pocket plddt, plus global_plddt, ipae, min_ipae, iptm, ptm_binder

The unified interface pipeline now has its own config:

python3 scripts/run_interface_pipeline.py \
  design_dir=examples/ligand_binding_protein \
  gpus=0 \
  root=results/examples/interface_design

interface reads the same lbp_info.csv as lbp to determine the ligand target_chain and redesigned design_chain, then computes the pocket only on design_chain.

Protein Binding Nucleic Acid (PBN)

PBN evaluates designed protein-nucleic acid complexes, where the protein chain is treated as the fixed conditioning partner and the nucleic-acid chain is the designed partner.

Input: sequence-assigned complex structures (.cif recommended)
Required metadata: pbn_info.csv
Current input contract: direct ODesign design outputs are accepted
Inverse fold: OInvFold (integrated in ODesignBench)
Refold: AlphaFold3
Evaluation: protein-aligned nucleic-acid C4' RMSD

pbn_info.csv must provide one row per design with at least:

design_name
target_chain
design_chain

Example:

design_name,target_chain,design_chain
prot_binding_rna_demo_seed_2_bb_0_seq_0,A,B

Recommended example layout:

examples/protein_binding_nuc/
|-- pbn_info.csv
`-- prot_binding_rna_demo_seed_2_bb_0_seq_0.cif

Example run:

python3 scripts/run_pbn_pipeline.py \
  design_dir=examples/protein_binding_nuc \
  gpus=0 \
  root=results/examples/protein_binding_nuc

Successful runs will write:

preprocessed inputs to formatted_designs/
OInvFold outputs to inverse_fold/
AlphaFold3 outputs to refold/af3_out/
final metrics to raw_data.csv

The evaluator aligns on shared protein CA residues and reports RMSD on shared nucleic-acid C4' atoms.

Protein Binding Ligand (PBL)

PBL evaluates ligand-containing protein structures and reports geometry, chemistry, and optional Vina docking metrics.

Input: .cif files in design_dir
Accepted layout: either design_dir/*.cif or nested case folders such as design_dir/2vt4/*.cif
Evaluation: automatic ligand extraction, pocket extraction, ligand geometry metrics, chemistry metrics, and Vina docking metrics

If your input CIF files are already inverse-fold outputs, the current PBL pipeline skips the inverse-fold stage and evaluates them directly after preprocessing.

Minimal run command:

python scripts/run_pbl_pipeline.py \
  design_dir=/path/to/protein_binding_ligand_designs \
  gpus=0

Optional: set root=/path/to/output_dir to change output location (default: results).

Recommended example layout:

examples/protein_binding_ligand/
`-- 2vt4/
    `-- 2vt4-1_seed_42_bb_0_seq_0.cif

Example run:

python scripts/run_pbl_pipeline.py \
  design_dir=examples/protein_binding_ligand \
  gpus=0 \
  root=results/examples/protein_binding_ligand

Successful runs will write:

preprocessed CIFs to formatted_designs/
evaluation inputs to inversefold_formatted_designs_for_evaluation/
ligand metrics to inversefold_formatted_designs_for_evaluation_metrics/

The final CSV and summary JSON are:

inversefold_formatted_designs_for_evaluation_metrics/evaluation_results.csv
inversefold_formatted_designs_for_evaluation_metrics/evaluation_summary_metrics.json

MotifBench (Motif Scaffolding)

MotifBench evaluates whether generated scaffolds preserve motif geometry while remaining structurally plausible and diverse.

Input: scaffold PDB files + scaffold_info.csv
Inverse fold: ProteinMPNN (motif-constrained design)
Refold: ESMFold
Evaluation: motif RMSD/scaffold RMSD, novelty, diversity

Foldseek Installation (Required for MotifBench Evaluation)

MotifBench uses Foldseek for diversity clustering and novelty evaluation. Follow these steps to install:

1. Install Foldseek via conda:

conda install -c conda-forge -c bioconda foldseek

2. Download the Foldseek PDB database:

export FOLDSEEK_DATABASE=/path/to/foldseek_pdb_database
mkdir -p $FOLDSEEK_DATABASE
cd $FOLDSEEK_DATABASE
foldseek databases PDB pdb tmp

Note: The database download may take 30-60 minutes depending on your connection speed. The PDB database is approximately 60GB uncompressed.

3. Set environment variables:

# Add to your ~/.bashrc or ~/.zshrc for persistence
export FOLDSEEK_DATABASE=/path/to/foldseek_pdb_database
export FOLDSEEK_BIN=$(which foldseek)  # or explicitly: /path/to/conda/bin/foldseek

4. Verify installation:

foldseek --version
foldseek createdb --help

5. Running MotifBench evaluation with Foldseek:

python scripts/run_motif_scaffolding_pipeline.py \
  design_dir=/path/to/motif_scaffolds \
  gpus=0 \
  motif_scaffolding.foldseek_database=$FOLDSEEK_DATABASE/pdb

Or via environment variable:

export FOLDSEEK_DATABASE=/path/to/foldseek_pdb_database
python scripts/run_motif_scaffolding_pipeline.py \
  design_dir=/path/to/motif_scaffolds \
  gpus=0

Minimal run command:

python scripts/run_motif_scaffolding_pipeline.py \
  design_dir=/path/to/motif_scaffolds \
  gpus=0

Optional: set root=/path/to/output_dir to change output location (default: results).

scaffold_info.csv should include:

sample_num
motif_placements

Example:

sample_num,motif_placements
0,34/A/70
1,30/A/25/B/30

Examples

We provide ready-to-use examples in the examples/ directory. You can run them directly to verify your installation and understand the pipeline workflow.

1. Motif Scaffolding

python3 scripts/run_motif_scaffolding_pipeline.py design_dir=examples/motif_scaffolding/01_1LDB/ gpus=0 root=results/examples/motif_scaffolding

2. Protein Binding Protein (PBP)

python3 scripts/run_pbp_pipeline.py design_dir=examples/protein_binding_protein/ gpus=0 root=results/examples/protein_binding_protein

2. Ligand Binding Protein

cd /path/to/ODesignBench
python3 scripts/run_lbp_pipeline.py \
  design_dir=examples/ligand_binding_protein \
  gpus=0 \
  root=results/examples/ligand_binding_protein

If preprocess/ccd_component/components.cif is missing, download it first:

cd preprocess/ccd_component
wget https://files.wwpdb.org/pub/pdb/data/monomers/components.cif.gz
gunzip -f components.cif.gz

Example metadata:

design_name,target_chain,design_chain
FAD_seed_1_bb_0_seq_0,A,B

3. Interface Design

python3 scripts/run_interface_pipeline.py \
  design_dir=examples/ligand_binding_protein \
  gpus=0 \
  root=results/examples/interface_design

This task uses the same examples/ligand_binding_protein/lbp_info.csv metadata file as LBP.

4. Atomic Motif Scaffolding / Enzyme Design (AME)

python3 scripts/run_ame_pipeline.py design_dir=examples/tip_atom_scaffolding/ gpus=0 root=results/examples/tip_atom_scaffolding

5. Protein Binding Ligand (PBL)

python3 scripts/run_pbl_pipeline.py design_dir=examples/protein_binding_ligand/ gpus=0 root=results/examples/protein_binding_ligand

5. Nucleic Acid (RNA free generation example)

python3 scripts/run_rna_pipeline.py design_dir=examples/nuc/rna/ gpus=0 root=results/examples/rna

6. Protein Binding Nucleic Acid (PBN)

python3 scripts/run_pbn_pipeline.py \
  design_dir=examples/protein_binding_nuc \
  gpus=0 \
  root=results/examples/protein_binding_nuc

Repository Layout

scripts/: task-level pipeline entry points
configs/: Hydra/OmegaConf configuration groups
preprocess/: input standardization and conversion utilities
inversefold/: sequence design backends
refold/: structure prediction/refolding backends
evaluation/: task-specific metrics and evaluators
assets/: benchmark assets and reference metadata

General Usage Pattern

Most tasks follow the same lifecycle:

Provide standardized input structures and task metadata.
Run inverse folding to generate sequences.
Refold generated sequences to structures.
Compute benchmark metrics and export CSV results.

Running Specific Pipeline Steps

The unified pipeline consists of five main stages: preprocess, inversefold, refold_prepare, refold, and evaluation. By default, all stages run sequentially.

You can skip specific stages by setting them to false via Hydra overrides using the +unified.steps.<stage_name>=false syntax. This is particularly useful if you only want to re-run evaluation or skip a time-consuming step that has already completed.

Example: Skip preprocessing and inverse folding

python scripts/run_ame_pipeline.py design_dir=examples/tip_atom_scaffolding/ gpus=0 root=results/examples/tip_atom_scaffolding \
  +unified.steps.preprocess=false \
  +unified.steps.inversefold=false

Example: Run ONLY evaluation (skip all other steps)

python scripts/run_ame_pipeline.py design_dir=examples/tip_atom_scaffolding/ gpus=0 root=results/examples/tip_atom_scaffolding \
  +unified.steps.preprocess=false \
  +unified.steps.inversefold=false \
  +unified.steps.refold_prepare=false \
  +unified.steps.refold=false

License and Citation

Please cite the corresponding benchmark release and model/tool dependencies used in your run (for example, PyRosetta, ESM, Chai-1, and AlphaFold3 where applicable).

Name		Name	Last commit message	Last commit date
Latest commit History 21 Commits
assets		assets
configs		configs
evaluation		evaluation
examples		examples
inversefold		inversefold
preprocess		preprocess
refold		refold
scripts		scripts
tests		tests
.gitignore		.gitignore
README.md		README.md
environment.yml		environment.yml

Folders and files

Latest commit

History

Repository files navigation

ODesignBench

Environment Setup

Installation Steps

Notes

Install gRNAde / PyG Runtime (Required for RNA pipeline)

Verifying Ligand Docking Dependencies

Inversefold Model Parameters (ProteinMPNN / LigandMPNN)

OInvFold Checkpoints

ESMFold Weights

Chai-1 Weights

Method 1: Fast Download via aria2 (Recommended for unstable networks)

Method 2: Pre-download via Python script

Next Step: Set the Environment Variable

AlphaFold3 Setup

Download required artifacts

Directory layout expected by refold/run_af3.sh

Configure environment variables

Singularity/Apptainer mode

Run

Benchmark Content

Nucleic Acid Monomer (DNA / RNA)

gRNAde Checkpoint Download

Atomic Motif Scaffolding / Enzyme Design (AME)

Protein Binding Protein (PBP)

Ligand Binding Protein (LBP)

Interface Design

Protein Binding Nucleic Acid (PBN)

Protein Binding Ligand (PBL)

MotifBench (Motif Scaffolding)

Foldseek Installation (Required for MotifBench Evaluation)

Examples

1. Motif Scaffolding

2. Protein Binding Protein (PBP)

2. Ligand Binding Protein

3. Interface Design

4. Atomic Motif Scaffolding / Enzyme Design (AME)

5. Protein Binding Ligand (PBL)

5. Nucleic Acid (RNA free generation example)

6. Protein Binding Nucleic Acid (PBN)

Repository Layout

General Usage Pattern

Running Specific Pipeline Steps

License and Citation

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Method 1: Fast Download via `aria2` (Recommended for unstable networks)

Directory layout expected by `refold/run_af3.sh`

Packages