RFdiffusion3 开源了｜安装及操作指南

git clone https://github.com/RosettaCommons/foundry.gitcd foundry

conda create -n rfd3 python=3.12 -yconda activate rfd3

pip install rc-foundry[all] -i https://pypi.org/simple

foundry install rfd3 --checkpoint-dir ./checkpoints

foundry --help # foundry list# foundry list-installed

RFdiffusion3 的使用

from lightning.fabric import seed_everythingfrom rfd3.engine import RFD3InferenceConfig, RFD3InferenceEngine# Set seed for reproducibilityseed_everything(0)# Configure RFD3 inferenceconfig = RFD3InferenceConfig(    specification={        'length': 80,  # Generate 80-residue proteins    },    diffusion_batch_size=2,  # Generate 2 structures per batch)# Initialize engine and run generationmodel = RFD3InferenceEngine(**config)outputs = model.run(    inputs=None,      # None for unconditional generation    out_dir=None,     # None to return in memory (no file output)    n_batches=1,      # Generate 1 batch)# Inspect RFD3 outputs and extract the generated structuresfor idx, data in outputs.items():    print(f"Batch {idx}: {len(data)} structure(s)")    print(f"  Output type: {type(data[0]).__name__}")    print(f"  AtomArray: {data[0].atom_array}")# Extract the first generated structure for downstream usefirst_key = next(iter(outputs.keys()))atom_array = outputs[first_key][0].atom_array# Visualize the generated structureview(atom_array)

from mpnn.inference_engines.mpnn import MPNNInferenceEngine
# Configure MPNN inference engine# See mpnn.utils.inference.MPNN_GLOBAL_INFERENCE_DEFAULTS for all optionsengine_config = {    "model_type": "ligand_mpnn",  # or "protein_mpnn" for vanilla ProteinMPNN    "is_legacy_weights": True,    # Required for now for ligand_mpnn and protein_mpnn    "out_directory": None,        # Return results in memory    "write_structures": False,    "write_fasta": False,}
# Configure per-input inference options# See mpnn.utils.inference.MPNN_PER_INPUT_INFERENCE_DEFAULTS for all optionsinput_configs = [    {        "batch_size": 10,         # Generate 10 sequences per structure        "remove_waters": True,    }]
# Run sequence design on the RFD3-generated backbonemodel = MPNNInferenceEngine(**engine_config)mpnn_outputs = model.run(input_dicts=input_configs, atom_arrays=[atom_array])
from biotite.structure import get_residue_startsfrom biotite.sequence import ProteinSequence
# Extract and display the designed sequencesprint(f"Generated {len(mpnn_outputs)} designed sequences:\n")
for i, item in enumerate(mpnn_outputs):    res_starts = get_residue_starts(item.atom_array)    # Convert 3-letter codes to 1-letter using Biotite    seq_1letter = ''.join(        ProteinSequence.convert_letter_3to1(res_name)        for res_name in item.atom_array.res_name[res_starts]    )    print(f"Sequence {i+1}: {seq_1letter}")

from rf3.inference_engines.rf3 import RF3InferenceEnginefrom rf3.utils.inference import InferenceInput

# Initialize RF3 inference engineinference_engine = RF3InferenceEngine(ckpt_path='rf3', verbose=False)
# Create input from the MPNN-designed structure (first design)# This re-folds the sequence to validate it adopts the intended structureinput_structure = InferenceInput.from_atom_array(atom_array, example_id="example_protein")rf3_outputs = inference_engine.run(inputs=input_structure)
# Outputs: dict mapping example_id -> list[RF3Output] (multiple models per input)print(f"Output keys: {rf3_outputs.keys()}")print(f"Number of models for 'example_protein': {len(rf3_outputs['example_protein'])}")# Extract the top-ranked predictionrf3_output = rf3_outputs["example_protein"][0]

核心参数说明

from biotite.structure import rmsd, superimposefrom atomworks.constants import PROTEIN_BACKBONE_ATOM_NAMESimport numpy as np
# Get structures for comparisonaa_generated = atom_array              # Original RFD3 backbone (from Section 1)aa_refolded = rf3_output.atom_array    # RF3-predicted structure
# Filter to backbone atoms (N, CA, C, O)bb_generated = aa_generated[np.isin(aa_generated.atom_name, PROTEIN_BACKBONE_ATOM_NAMES)]bb_refolded = aa_refolded[np.isin(aa_refolded.atom_name, PROTEIN_BACKBONE_ATOM_NAMES)]
# Superimpose structures and calculate RMSDbb_refolded_fitted, _ = superimpose(bb_generated, bb_refolded)rmsd_value = rmsd(bb_generated, bb_refolded_fitted)
print(f"Backbone RMSD: {rmsd_value:.2f} A")print(f"\nInterpretation: {'Excellent' if rmsd_value < 1.0 else 'Good' if rmsd_value < 2.0 else 'Moderate'} designability")

小Z总结