foundry

name	foundry
description	Foundry protein design toolkit reference. Auto-invoked when the user discusses protein design, structure prediction, inverse folding, RFdiffusion3, RosettaFold3, ProteinMPNN, LigandMPNN, SolubleMPNN, Enhanced MPNN, or running Foundry tools on GPUs.
user-invocable	false
allowed-tools	Bash, Read, Write, Glob, Grep

Foundry Protein Design Toolkit Reference

Foundry provides core protein design and prediction tools. All tools require GPU compute and are designed to run on Vast.ai instances using a pre-built Docker image that includes all tools, checkpoints, and dependencies.

Tools Overview

Tool	CLI Command	Purpose	GPU VRAM
RFdiffusion3 (RFD3)	rfd3 design	All-atom generative protein structure design	24-48GB+
RosettaFold3 (RF3)	rf3 fold	Biomolecular structure prediction	24-48GB+
ProteinMPNN/LigandMPNN	mpnn	Fixed-backbone inverse folding sequence design	16-24GB
Enhanced MPNN	mpnn	Designability-optimized inverse folding (fine-tuned LigandMPNN)	16-24GB

Docker Image

A pre-built Docker image based on vastai/base-image:cuda-12.4.1-auto is available with everything pre-installed:

• Python 3.12, PyTorch with CUDA 12.4
• Foundry with all model extras (rc-foundry[all])
• All checkpoints pre-downloaded to /root/.foundry/checkpoints
• Enhanced MPNN weights included

When launching Vast.ai instances, use this image with:

vastai create instance <OFFER_ID> --image <FOUNDRY_DOCKER_IMAGE> --disk 64 --ssh

No --onstart-cmd is needed — the image is ready to use immediately. The working directory is /workspace.

Available Checkpoints

All checkpoints are pre-installed in the Docker image at /root/.foundry/checkpoints:

Name	File	Tool
rfd3	rfd3_latest.ckpt	RFdiffusion3
rf3	rf3_foundry_01_24_latest_remapped.ckpt	RosettaFold3 (latest, recommended)
rf3_preprint_921	rf3_foundry_09_21_preprint_remapped.ckpt	RF3 (benchmark, 09/21 cutoff)
rf3_preprint_124	rf3_foundry_01_24_preprint_remapped.ckpt	RF3 (preprint)
proteinmpnn	proteinmpnn_v_48_020.pt	ProteinMPNN
ligandmpnn	ligandmpnn_v_32_010_25.pt	LigandMPNN
solublempnn	solublempnn_v_48_020.pt	SolubleMPNN
enhanced_mpnn	enhanced_mpnn_step_80000.pt	Enhanced MPNN

Checkpoint env: FOUNDRY_CHECKPOINT_DIRS=/root/.foundry/checkpoints

---

Tool 1: RFdiffusion3 (RFD3)

All-atom generative model for designing protein structures under complex constraints: protein binders, enzyme active sites, nucleic acid binders, small molecule binders, symmetric assemblies.

CLI

rfd3 design out_dir=<OUTPUT_DIR> inputs=<INPUT_JSON> [OPTIONS]

Key Parameters

Parameter	Default	Description
out_dir	required	Output directory
inputs	required	Input JSON/YAML file
ckpt_path	rfd3	Checkpoint (auto-resolved from registry)
skip_existing	True	Skip existing outputs
diffusion_batch_size	8	Designs per batch
n_batches	1	Number of batches
dump_trajectories	False	Save denoising trajectory (large files)
prevalidate_inputs	False	Validate inputs before loading model
low_memory_mode	False	Memory-efficient tokenization

Sampler Parameters

Parameter	Default	Description
inference_sampler.num_timesteps	200	Diffusion steps
inference_sampler.step_scale	1.5	Diversity vs designability tradeoff
inference_sampler.noise_scale	1.003	Noise scaling
inference_sampler.use_classifier_free_guidance	False	Enable CFG
inference_sampler.cfg_scale	1.5	CFG scale (if enabled)
inference_sampler.kind	default	default or symmetry

Input JSON Format

The input is a JSON file mapping example names to design specifications:

{
    "design_name": {
        "input": "./path/to/target.pdb",
        "contig": "40-120,/0,A1-100",
        "length": "140-160",
        "ligand": "NAI,ACT",
        "unindex": "A108,A139",
        "select_fixed_atoms": {
            "A108": "ND2,CG",
            "A139": "OG,CB,CA"
        },
        "select_hotspots": {
            "E64": "CD2,CZ"
        },
        "is_non_loopy": true,
        "infer_ori_strategy": "hotspots",
        "dialect": 2
    }
}

Key input fields:

• input: Path to target structure (PDB/CIF)
• contig: Contig specification — defines which residues to keep/design and chain breaks
• length: Length range for designed protein (e.g., "140-160")
• ligand: Comma-separated ligand residue names to include
• unindex: Residues to unindex (make designable in position)
• select_fixed_atoms: Per-residue atom selections to fix
• select_hotspots: Target hotspot residues for interface design
• is_non_loopy: Disable loop-only design mode
• infer_ori_strategy: How to determine orientation (hotspots)
• dialect: Input dialect version (use 2 for latest)
• partial_t: Partial diffusion timestep (for refinement)
• ori_token: Orientation token indices

Output

• {name}.cif — Designed all-atom structures
• {name}.json — Full design metadata
• Trajectory files (if dump_trajectories=True)

---

Tool 2: RosettaFold3 (RF3)

All-atom biomolecular structure prediction for proteins, nucleic acids, ligands, and complexes.

CLI

rf3 fold inputs=<INPUT_FILE_OR_DIR> [OPTIONS]

Key Parameters

Parameter	Default	Description
inputs	required	JSON, CIF, PDB file, list, or directory
out_dir	./	Output directory
ckpt_path	auto	Checkpoint path
n_recycles	10	Number of recycling iterations
diffusion_batch_size	5	Number of output structures
num_steps	200	Diffusion sampling steps (50 is faster, similar quality)
early_stopping_plddt_threshold	0.5	Skip low-confidence predictions
seed	null	Random seed
dump_trajectories	False	Save denoising trajectories
skip_existing	False	Skip existing predictions
one_model_per_file	False	Separate files per model
annotate_b_factor_with_plddt	False	pLDDT as B-factors
template_noise_scale	1e-5	Template noise

Structural Control

• template_selection — AtomSelection syntax for template regions (e.g., "[A, B/*/1-42]")
• ground_truth_conformer_selection — Fix ligand conformations (e.g., "[C, D]")
• cyclic_chains — List of chain IDs to cyclize

Input JSON Format

[
    {
        "name": "example_prediction",
        "components": [
            {
                "seq": "MTSENPLLALREK...",
                "chain_id": "A",
                "msa_path": "path/to/protein.a3m"
            },
            {
                "ccd_code": "MG"
            },
            {
                "smiles": "[nH]1cc[nH+]c1"
            },
            {
                "path": "path/to/ligand.sdf"
            }
        ],
        "template_selection": ["A"],
        "ground_truth_conformer_selection": ["C"]
    }
]

Component types:

• seq + optional msa_path — Protein/nucleic acid sequence (supports non-canonical: (PTM))
• ccd_code — CCD compound code (e.g., MG, NAG)
• smiles — Small molecule SMILES string
• path — Structure file (CIF, PDB, SDF)

AtomSelection syntax: CHAIN/RES_NAME/RES_ID/ATOM_NAME

• A — all atoms in chain A
• A/*/5-10 — residues 5-10 in chain A
• B/*/1-42, B/*/49-63 — multiple regions (CDR framework)

Output

• {name}_metrics.csv — Overall confidence metrics (pTM, pLDDT, ipTM)
• {name}.score — Granular per-atom metrics
• {name}_model_0.cif.gz ... {name}_model_N.cif.gz — Predicted structures

---

Tool 3: ProteinMPNN / LigandMPNN / SolubleMPNN / Enhanced MPNN

Lightweight inverse-folding models for fixed-backbone protein sequence design.

CLI

mpnn --model_type <MODEL_TYPE> --structure_path <STRUCTURE> [OPTIONS]

Or from JSON config:

mpnn --config_json config.json

Model Variants

Variant	model_type	Checkpoint	Use Case
ProteinMPNN	protein_mpnn	proteinmpnn_v_48_020.pt	Standard protein sequence design
LigandMPNN	ligand_mpnn	ligandmpnn_v_32_010_25.pt	Design around small molecules, DNA, ions
SolubleMPNN	protein_mpnn	solublempnn_v_48_020.pt	Solubility-optimized design
Enhanced MPNN	ligand_mpnn	enhanced_mpnn_step_80000.pt	Designability-optimized (highest success rate)

Enhanced MPNN

Enhanced MPNN is a fine-tuned LigandMPNN trained with ResiDPO (Residue-level Designability Preference Optimization). Instead of optimizing for native sequence recovery like standard MPNN, it directly optimizes for designability — whether designed sequences fold into the target structure with high confidence (as measured by AlphaFold2 pLDDT scores).

Performance improvements over standard LigandMPNN:

• Enzyme design: ~2.7x higher success rate (17.6% vs 6.6%)
• Binder design: ~2.3x higher success rate (16.1% vs 7.1%)
• Makes previously "undesignable" backbones designable — doubles the fraction of viable backbone scaffolds

How it works:

• ResiDPO decouples preference learning from KL regularization at the residue level
• Residue-level Preference Learning (RPL) improves positions where pLDDT is low
• Residue-level Constraint Learning (RCL) preserves knowledge at positions already working well
• Trained on 19k PDB structures with AF2 pLDDT as the reward signal

When to use Enhanced MPNN (recommended for most de novo design):

• Binder or enzyme design pipelines (RFD3 → MPNN → RF3 validation)
• Any scenario where standard MPNN yields low designability/validation rates
• Maximizing the fraction of designs that pass structure validation

When to prefer standard MPNN/LigandMPNN:

• When sequence recovery (similarity to native) matters more than designability
• Benchmarking against published results using original models

Usage — same CLI as LigandMPNN, just swap the checkpoint:

mpnn --model_type ligand_mpnn \
  --checkpoint_path /root/.foundry/checkpoints/enhanced_mpnn_step_80000.pt \
  --structure_path backbone.cif \
  --temperature 0.1 \
  --number_of_batches 8

Key Parameters

Global:

• --model_type — protein_mpnn or ligand_mpnn
• --checkpoint_path — Path to model weights
• --is_legacy_weights — Set True for original repository weights (not needed for Enhanced MPNN)
• --out_directory — Output directory
• --write_fasta — Write FASTA output (default: True)
• --write_structures — Write designed structures (default: True)

Per-input:

• --structure_path — Input structure (CIF/PDB)
• --batch_size — Sequences per batch (default: 1)
• --number_of_batches — Number of batches (default: 1)
• --temperature — Sampling temperature, controls diversity (default: 0.1)
• --seed — Random seed
• --designed_chains — Chains to redesign
• --fixed_chains — Chains to keep fixed
• --designed_residues — Specific residues to design
• --fixed_residues — Specific residues to fix

Advanced:

• --omit — Amino acids to exclude (default: ["UNK"])
• --bias — Per-residue logit bias
• --structure_noise — Noise level (default: 0.0)
• --symmetry_residues — Residues for symmetric design
• --homo_oligomer_chains — Homo-oligomer chains

JSON Config Format

{
    "checkpoint_path": "enhanced_mpnn_step_80000.pt",
    "model_type": "ligand_mpnn",
    "out_directory": "./outputs/",
    "inputs": [
        {
            "structure_path": "complex.pdb",
            "name": "example",
            "seed": 42,
            "batch_size": 1,
            "number_of_batches": 5,
            "temperature": 0.1,
            "fixed_chains": ["A"],
            "designed_chains": ["B"]
        }
    ]
}

Output

• {name}_sequences_*.fasta — Designed sequences
• {name}_*.cif — Designed structures (if write_structures=True)

---

Common Design Workflows

Protein Binder Design Pipeline

1. RFD3: Generate binder backbones targeting a protein → outputs CIF structures
2. Enhanced MPNN: Design sequences for the generated backbones (~2.3x better success) → outputs FASTA sequences
3. RF3: Validate designed sequences fold correctly → outputs confidence metrics

Enzyme Design Pipeline

1. RFD3: Design enzyme scaffolds around a ligand/substrate
2. Enhanced MPNN: Design ligand-aware sequences (~2.7x better designability) → outputs FASTA sequences
3. RF3: Validate predicted structure matches design

Sequence Optimization

1. Start with an existing structure (PDB/CIF)
2. ProteinMPNN/SolubleMPNN: Redesign sequences for stability/solubility
3. RF3: Predict structure of redesigned sequences

---

Vast.ai GPU Recommendations

Tool	Min VRAM	Recommended GPU	Notes
RFD3	24 GB	A100 40GB, RTX 4090	Large designs need 48GB+
RF3	24 GB	A100 40GB, RTX 4090	Multi-chain complexes need more
MPNN (all variants)	8 GB	RTX 4090, RTX 3090	Very lightweight

Use the pre-built Foundry Docker image — no setup commands needed. Instance is ready to run immediately after launch.