RosettaFold3¶

All-atom biomolecular structure prediction network using transformer + diffusion architecture, capable of predicting protein, DNA, RNA, and ligand complexes with confidence scoring.

License: BSD-3-Clause · Molecules: protein, dna, rna, ligand, complex · Tasks: structure_prediction

Actions: fold · Variants: 1

At a glance¶

Use it when

You need the highest accuracy available for multi-molecule biomolecular structure prediction -- RF3 represents the current state of the art from the Baker Lab lineage
You need template-guided structure prediction where partial experimental data (homolog structures, known domain folds) can be incorporated to improve prediction accuracy
You are predicting protein-ligand complex structures for drug discovery, especially when the ligand can be specified via SMILES strings or CCD codes
You need to predict structures of cyclic peptides or proteins with non-canonical amino acids, which require specialized handling beyond standard amino acid vocabularies
You want to sample multiple diverse conformations via diffusion to explore alternative binding modes, allosteric states, or conformational heterogeneity
You are working on multi-component biological systems (protein-DNA, protein-RNA, protein-ligand, protein-protein) and need a single model that handles all entity types
You need interface-specific confidence scores (ipTM) to evaluate the quality of predicted inter-molecular contacts for complex validation

Strengths

Newest-generation all-atom biomolecular structure prediction model from the Baker Lab (RosettaCommons), representing the cutting edge of diffusion-based structure prediction with state-of-the-art accuracy on multi-molecule benchmarks
Handles the broadest range of molecular inputs: proteins, DNA, RNA, small-molecule ligands (SMILES or CCD codes), cyclic peptides, non-canonical amino acids, and covalent modifications -- all in a single unified model
Template conditioning via template_selection allows users to incorporate partial experimental data (known domain structures, homolog structures, crystallographic fragments) to guide and improve predictions
Diffusion-based sampling generates multiple diverse conformations per input, enabling exploration of alternative states and conformational heterogeneity -- critical for drug discovery and allosteric mechanism studies
Comprehensive confidence metrics: pTM (global), ipTM (interface-specific), pLDDT (per-residue), PAE (pairwise aligned error), and composite ranking_score for automated quality triage
BSD-3-Clause open-source license from RosettaCommons allows academic and commercial use with minimal restrictions
Supports cyclic_chains parameter for cyclic peptide structure prediction, a specialized capability not available in most competing models
Built on the foundry framework, pairing a Transformer-based trunk (informed by evolutionary MSA and pairwise information) with a diffusion-based structure generation module for robust all-atom predictions

Limitations

Requires A100 40GB GPU with 64GB system RAM -- expensive infrastructure that limits accessibility for routine use or large-scale screening
No binding affinity prediction -- produces structures and confidence scores but cannot estimate binding energies; a dedicated affinity predictor is required for that
MSA preprocessing is recommended for best accuracy, adding minutes to hours of database search time before structure prediction can begin
Newer model with less published community validation than AlphaFold2 -- fewer independent benchmarks and downstream application studies in the literature
Membrane protein predictions remain weak due to training data bias toward soluble proteins and absence of lipid bilayer modeling
Single-structure outputs from each diffusion trajectory -- while multiple samples provide diversity, the model does not explicitly model thermodynamic ensembles or Boltzmann-weighted populations
Large multi-chain complexes may exceed GPU memory, requiring careful input size management or tiling strategies
Small molecules alone (without biomolecular context) are not well-handled -- RF3 is designed for biomolecular complexes, not standalone molecular structure optimization

Reach for something else when

You need rapid single-chain protein structure screening -- ESMFold provides predictions in seconds without MSA, at a fraction of the cost
You need binding affinity prediction alongside structure -- RF3 predicts structures and confidence scores but does not estimate binding affinities
You need de novo protein design or binder generation -- RF3 is a prediction model; use BoltzGen for end-to-end binder design
GPU resources are limited and you only need protein structure prediction -- AF2 NIM on A100 80GB provides well-validated protein predictions; ESMFold on A10G provides fast screening
You need conformational ensemble modeling for intrinsically disordered proteins -- RF3 produces individual conformations, not thermodynamically weighted ensembles
You need the most extensively validated predictions for publication -- AlphaFold2 has the broadest published benchmark literature and may be preferred by reviewers

Alternatives

Model	Better when	Worse when
chai1	Well-benchmarked AF3-class multi-molecule complex predictor with ESM-embedding single-sequence boost and commercial support from Chai Discovery.	No template conditioning, cyclic-peptide/CCD-code handling, or the diffusion multi-sample conformational diversity RF3 offers; newer RF3 architecture may be more accurate on some targets.
esmfold	100-1000x faster single-sequence protein folding on a cheap A10G GPU with no MSA -- ideal for high-throughput protein screening.	Protein-only (no DNA/RNA/ligand/complex or SMILES), substantially lower accuracy, 768-residue limit, no template conditioning or diffusion sampling.

API & schema¶

Call an action with POST /api/v1/{slug}/{action} — the request envelope is {"items": [...], "params": {...}} and a success returns {"results": [...]}. See the HTTP API page for the base URL, error shape, and full contract.

`fold`¶

Call it

curl -X POST http://127.0.0.1:8000/api/v1/rf3/fold \
  -H "Content-Type: application/json" \
  -d '{
  "items": [
    {
      "name": "string",
      "components": [
        {
          "name": "string",
          "type": "protein"
        }
      ]
    }
  ]
}'

Request — RF3FoldRequest

Field	Type	Required	Constraints	Description
`params`	RF3FoldParams	no		Optional parameters controlling this action (defaults are used when omitted).
`items`	list[RF3FoldRequestInput]	yes	items 1–1	Batch of inputs to process in a single request. Up to 1 input per request.

Nested types

RF3AlignmentDatabase

Allowed values: mgnify, small_bfd, uniref90

RF3Component

Field	Type	Required	Description
`name`	string	yes	Human-readable label for this component; used as an identifier within the prediction job.
`type`	RF3EntityType	yes	Component entity type: protein, DNA, RNA, or ligand.
`sequence`	string \| null	no	A protein, DNA, or RNA sequence in single-letter codes (maximum 2048 characters).
`smiles`	string \| null	no	Ligand structure as a SMILES string.
`ccd_code`	string \| null	no	Chemical Component Dictionary code for a ligand.
`structure_path`	string \| null	no	Container-local path to a template structure file (CIF/PDB/SDF). Not usable by external callers; prefer structure_cif instead.
`structure_cif`	string \| null	no	Template structure in mmCIF format (inline text). Use this to supply a structure template.
`chain_id`	string \| null	no	Chain identifier to operate on (e.g. "A").
`msa_path`	string \| null	no	Container-local path to an MSA file (.a3m). Not usable by external callers; prefer msa_content or alignment instead.
`msa_content`	string \| null	no	Multiple-sequence alignment for the query sequence, in A3M format.
`alignment`	object \| null	no	MSA alignments keyed by sequence database.

RF3EntityType

Allowed values: protein, DNA, RNA, ligand

RF3FoldParams

Field	Type	Required	Constraints	Description
`n_recycles`	integer	no	≥2; ≤20; default `10`	Number of trunk recycling iterations.
`num_steps`	integer	no	≥50; ≤500; default `200`	Number of diffusion sampling steps.
`diffusion_batch_size`	integer	no	≥1; ≤10; default `5`	Number of diffusion samples to generate in parallel.
`seed`	integer \| null	no	default `42`	Random seed for reproducible sampling.
`template_selection`	list[string] \| null	no		Atom selections for token-level templates (e.g. ['A', 'B/*/1-10']).
`ground_truth_conformer_selection`	list[string] \| null	no		Atom selections fixed to their ground-truth conformers (e.g. ['C', 'D']).
`cyclic_chains`	list[string] \| null	no		Chain identifiers to model as cyclic (e.g. cyclic peptides).
`early_stopping_plddt_threshold`	number	no	≥0.0; ≤1.0; default `0.5`	Early-stopping threshold as a fraction in [0, 1] (note: output pLDDT scores use a 0–100 scale). Sampling stops early when the pLDDT fraction falls below this value.
`one_model_per_file`	boolean	no	default `False`	Whether to write each predicted model to its own file.
`annotate_b_factor_with_plddt`	boolean	no	default `False`	Whether to store per-residue pLDDT in the B-factor column.
`include_pae`	boolean	no	default `False`	Whether to include the predicted aligned error matrix in the response.
`include_plddt`	boolean	no	default `True`	Whether to include per-residue pLDDT scores in the response.

RF3FoldRequestInput

Field	Type	Required	Constraints	Description
`name`	string	yes		Human-readable label for this prediction job; used as the output directory name.
`components`	list[RF3Component]	yes	items 1–	Biomolecular components that make up the complex to predict.
`bonds`	list[list] \| null	no		Custom covalent bonds as pairs of atom specifications.

Raw JSON Schema

{
  "$defs": {
    "RF3AlignmentDatabase": {
      "description": "MSA alignment database options.",
      "enum": [
        "mgnify",
        "small_bfd",
        "uniref90"
      ],
      "title": "RF3AlignmentDatabase",
      "type": "string"
    },
    "RF3Component": {
      "additionalProperties": false,
      "description": "A biomolecular component for structure prediction.",
      "properties": {
        "name": {
          "description": "Human-readable label for this component; used as an identifier within the prediction job.",
          "title": "Name",
          "type": "string"
        },
        "type": {
          "$ref": "#/$defs/RF3EntityType",
          "description": "Component entity type: protein, DNA, RNA, or ligand."
        },
        "sequence": {
          "anyOf": [
            {
              "maxLength": 2048,
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "A protein, DNA, or RNA sequence in single-letter codes (maximum 2048 characters).",
          "title": "Sequence"
        },
        "smiles": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Ligand structure as a SMILES string.",
          "title": "Smiles"
        },
        "ccd_code": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Chemical Component Dictionary code for a ligand.",
          "title": "Ccd Code"
        },
        "structure_path": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Container-local path to a template structure file (CIF/PDB/SDF). Not usable by external callers; prefer structure_cif instead.",
          "title": "Structure Path"
        },
        "structure_cif": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Template structure in mmCIF format (inline text). Use this to supply a structure template.",
          "title": "Structure Cif"
        },
        "chain_id": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Chain identifier to operate on (e.g. \"A\").",
          "title": "Chain Id"
        },
        "msa_path": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Container-local path to an MSA file (.a3m). Not usable by external callers; prefer msa_content or alignment instead.",
          "title": "Msa Path"
        },
        "msa_content": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Multiple-sequence alignment for the query sequence, in A3M format.",
          "title": "Msa Content"
        },
        "alignment": {
          "anyOf": [
            {
              "additionalProperties": {
                "type": "string"
              },
              "propertyNames": {
                "$ref": "#/$defs/RF3AlignmentDatabase"
              },
              "type": "object"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "MSA alignments keyed by sequence database.",
          "title": "Alignment"
        }
      },
      "required": [
        "name",
        "type"
      ],
      "title": "RF3Component",
      "type": "object"
    },
    "RF3EntityType": {
      "description": "Entity types for RF3 input.",
      "enum": [
        "protein",
        "DNA",
        "RNA",
        "ligand"
      ],
      "title": "RF3EntityType",
      "type": "string"
    },
    "RF3FoldParams": {
      "additionalProperties": false,
      "description": "Parameters for RF3 structure prediction.",
      "properties": {
        "n_recycles": {
          "default": 10,
          "description": "Number of trunk recycling iterations.",
          "maximum": 20,
          "minimum": 2,
          "title": "N Recycles",
          "type": "integer"
        },
        "num_steps": {
          "default": 200,
          "description": "Number of diffusion sampling steps.",
          "maximum": 500,
          "minimum": 50,
          "title": "Num Steps",
          "type": "integer"
        },
        "diffusion_batch_size": {
          "default": 5,
          "description": "Number of diffusion samples to generate in parallel.",
          "maximum": 10,
          "minimum": 1,
          "title": "Diffusion Batch Size",
          "type": "integer"
        },
        "seed": {
          "anyOf": [
            {
              "type": "integer"
            },
            {
              "type": "null"
            }
          ],
          "default": 42,
          "description": "Random seed for reproducible sampling.",
          "title": "Seed"
        },
        "template_selection": {
          "anyOf": [
            {
              "items": {
                "type": "string"
              },
              "type": "array"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Atom selections for token-level templates (e.g. ['A', 'B/*/1-10']).",
          "title": "Template Selection"
        },
        "ground_truth_conformer_selection": {
          "anyOf": [
            {
              "items": {
                "type": "string"
              },
              "type": "array"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Atom selections fixed to their ground-truth conformers (e.g. ['C', 'D']).",
          "title": "Ground Truth Conformer Selection"
        },
        "cyclic_chains": {
          "anyOf": [
            {
              "items": {
                "type": "string"
              },
              "type": "array"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Chain identifiers to model as cyclic (e.g. cyclic peptides).",
          "title": "Cyclic Chains"
        },
        "early_stopping_plddt_threshold": {
          "default": 0.5,
          "description": "Early-stopping threshold as a fraction in [0, 1] (note: output pLDDT scores use a 0\u2013100 scale). Sampling stops early when the pLDDT fraction falls below this value.",
          "maximum": 1.0,
          "minimum": 0.0,
          "title": "Early Stopping Plddt Threshold",
          "type": "number"
        },
        "one_model_per_file": {
          "default": false,
          "description": "Whether to write each predicted model to its own file.",
          "title": "One Model Per File",
          "type": "boolean"
        },
        "annotate_b_factor_with_plddt": {
          "default": false,
          "description": "Whether to store per-residue pLDDT in the B-factor column.",
          "title": "Annotate B Factor With Plddt",
          "type": "boolean"
        },
        "include_pae": {
          "default": false,
          "description": "Whether to include the predicted aligned error matrix in the response.",
          "title": "Include Pae",
          "type": "boolean"
        },
        "include_plddt": {
          "default": true,
          "description": "Whether to include per-residue pLDDT scores in the response.",
          "title": "Include Plddt",
          "type": "boolean"
        }
      },
      "title": "RF3FoldParams",
      "type": "object"
    },
    "RF3FoldRequestInput": {
      "additionalProperties": false,
      "description": "Input specification for a structure prediction task.",
      "properties": {
        "name": {
          "description": "Human-readable label for this prediction job; used as the output directory name.",
          "title": "Name",
          "type": "string"
        },
        "components": {
          "description": "Biomolecular components that make up the complex to predict.",
          "items": {
            "$ref": "#/$defs/RF3Component"
          },
          "minItems": 1,
          "title": "Components",
          "type": "array"
        },
        "bonds": {
          "anyOf": [
            {
              "items": {
                "maxItems": 2,
                "minItems": 2,
                "prefixItems": [
                  {
                    "type": "string"
                  },
                  {
                    "type": "string"
                  }
                ],
                "type": "array"
              },
              "type": "array"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Custom covalent bonds as pairs of atom specifications.",
          "title": "Bonds"
        }
      },
      "required": [
        "name",
        "components"
      ],
      "title": "RF3FoldRequestInput",
      "type": "object"
    }
  },
  "additionalProperties": false,
  "description": "Request for RF3 structure prediction.",
  "properties": {
    "params": {
      "$ref": "#/$defs/RF3FoldParams",
      "description": "Optional parameters controlling this action (defaults are used when omitted)."
    },
    "items": {
      "description": "Batch of inputs to process in a single request. Up to 1 input per request.",
      "items": {
        "$ref": "#/$defs/RF3FoldRequestInput"
      },
      "maxItems": 1,
      "minItems": 1,
      "title": "Items",
      "type": "array"
    }
  },
  "required": [
    "items"
  ],
  "title": "RF3FoldRequest",
  "type": "object"
}

Response — RF3FoldResponse

Field	Type	Required	Constraints	Description
`results`	list[list[RF3FoldResponseResult]]	yes		Per-input results, returned in the same order as the request items.

Nested types

RF3ConfidenceScores

Field	Type	Required	Description
`ptm`	number \| null	no	Predicted TM-score (pTM) for the overall structure (0–1).
`iptm`	number \| null	no	Interface predicted TM-score (ipTM) for multi-chain complexes (0–1).
`ranking_score`	number \| null	no	Composite score used to rank diffusion samples.
`has_clash`	boolean \| null	no	Whether the predicted structure contains steric clashes.
`plddt`	list[number] \| null	no	Per-residue pLDDT confidence score (0–100; higher is more confident).
`pae`	list[list[number]] \| null	no	Predicted aligned error (PAE) matrix, in Ångströms.

RF3FoldResponseResult

Field	Type	Required	Constraints	Description
`structure_cif`	string	yes		Predicted structure in mmCIF format.
`confidence`	RF3ConfidenceScores	yes		Confidence scores for the prediction (pTM, ipTM, pLDDT, PAE, ranking).
`early_stopped`	boolean	no	default `False`	Whether this sample was stopped early on low confidence.
`sample_idx`	integer	yes		Index of this sample within the diffusion batch.

Raw JSON Schema

{
  "$defs": {
    "RF3ConfidenceScores": {
      "description": "Confidence metrics for a prediction.",
      "properties": {
        "ptm": {
          "anyOf": [
            {
              "type": "number"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Predicted TM-score (pTM) for the overall structure (0\u20131).",
          "title": "Ptm"
        },
        "iptm": {
          "anyOf": [
            {
              "type": "number"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Interface predicted TM-score (ipTM) for multi-chain complexes (0\u20131).",
          "title": "Iptm"
        },
        "ranking_score": {
          "anyOf": [
            {
              "type": "number"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Composite score used to rank diffusion samples.",
          "title": "Ranking Score"
        },
        "has_clash": {
          "anyOf": [
            {
              "type": "boolean"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Whether the predicted structure contains steric clashes.",
          "title": "Has Clash"
        },
        "plddt": {
          "anyOf": [
            {
              "items": {
                "type": "number"
              },
              "type": "array"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Per-residue pLDDT confidence score (0\u2013100; higher is more confident).",
          "title": "Plddt"
        },
        "pae": {
          "anyOf": [
            {
              "items": {
                "items": {
                  "type": "number"
                },
                "type": "array"
              },
              "type": "array"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Predicted aligned error (PAE) matrix, in \u00c5ngstr\u00f6ms.",
          "title": "Pae"
        }
      },
      "title": "RF3ConfidenceScores",
      "type": "object"
    },
    "RF3FoldResponseResult": {
      "description": "Single prediction output from RF3.",
      "properties": {
        "structure_cif": {
          "description": "Predicted structure in mmCIF format.",
          "title": "Structure Cif",
          "type": "string"
        },
        "confidence": {
          "$ref": "#/$defs/RF3ConfidenceScores",
          "description": "Confidence scores for the prediction (pTM, ipTM, pLDDT, PAE, ranking)."
        },
        "early_stopped": {
          "default": false,
          "description": "Whether this sample was stopped early on low confidence.",
          "title": "Early Stopped",
          "type": "boolean"
        },
        "sample_idx": {
          "description": "Index of this sample within the diffusion batch.",
          "title": "Sample Idx",
          "type": "integer"
        }
      },
      "required": [
        "structure_cif",
        "confidence",
        "sample_idx"
      ],
      "title": "RF3FoldResponseResult",
      "type": "object"
    }
  },
  "description": "Response from RF3 structure prediction.",
  "properties": {
    "results": {
      "description": "Per-input results, returned in the same order as the request items.",
      "items": {
        "items": {
          "$ref": "#/$defs/RF3FoldResponseResult"
        },
        "type": "array"
      },
      "title": "Results",
      "type": "array"
    }
  },
  "required": [
    "results"
  ],
  "title": "RF3FoldResponse",
  "type": "object"
}

Usage¶

One-line summary: All-atom biomolecular structure prediction network using transformer + diffusion architecture, capable of predicting protein, DNA, RNA, and ligand complexes with confidence scoring.

Overview¶

RosettaFold3 is an all-atom biomolecular structure prediction network competitive with leading open-source models. It improves on tasks such as prediction of chiral ligands and fixed-backbone or fixed-conformer docking.

This implementation wraps the RosettaFold3 model from the RosettaCommons/foundry repository into a Modal app following the BioLM patterns. RF3 can:

Predict protein structures with or without MSAs
Predict multi-component complexes (protein-protein, protein-DNA, protein-RNA, protein-ligand)
Handle non-canonical amino acids and covalent modifications
Template portions of structures while folding others
Predict chiral small molecules accurately
Early-stop low-confidence predictions to save compute

Architecture¶

Property	Value
Architecture	Transformer + Diffusion
Training objective	All-atom structure prediction
Input modalities	Sequence, Structure, SMILES, MSA
Input molecules	Protein, DNA, RNA, Ligand, Complex
Output	Predicted 3D structure (mmCIF format)
Max sequence length	2048 tokens

Model Files¶

__init__.py - Package initialization
app.py - Main Modal app with RF3 inference engine
schema.py - Pydantic request/response schemas
config.py - Model family configuration
download.py - Model weight download logic
fixture.py - Test fixture generator
test.py - Model tests

Available Checkpoints¶

The "latest" checkpoint is used by default (most recent release with bugfixes). The upstream IPD server also hosts preprint and benchmark checkpoint variants, but only the latest checkpoint is exposed through this API.

Capabilities & Limitations¶

CAN be used for: - Protein structure prediction (single chain or multi-chain) - Protein-protein, protein-DNA, protein-RNA, and protein-ligand complex prediction - Templated folding (fixing portions of a structure) - Chiral small molecule prediction - MSA-enhanced structure prediction - Multi-sample diffusion batching with confidence ranking

CANNOT be used for: - Sequences longer than 2048 residues - Real-time interactive folding (diffusion sampling is computationally expensive) - Training or fine-tuning (inference only)

Other considerations: - MSAs significantly improve prediction quality for proteins - Early stopping can save 10-20x compute for low-confidence predictions - Output structures are returned in mmCIF format (may be gzipped) - Multiple samples from the diffusion batch are ranked by confidence

Usage Examples¶

Simple Protein Folding¶

from models.rf3.schema import (
    RF3FoldRequest,
    RF3FoldRequestInput,
    RF3FoldParams,
    RF3Component,
    RF3EntityType,
)

request = RF3FoldRequest(
    params=RF3FoldParams(
        n_recycles=10,
        num_steps=200,
        diffusion_batch_size=5,
    ),
    items=[
        RF3FoldRequestInput(
            name="simple_fold",
            components=[
                RF3Component(
                    name="protein",
                    type=RF3EntityType.PROTEIN,
                    sequence="MKLLISC...",
                    chain_id="A",
                )
            ],
        )
    ],
)

Protein with MSA¶

from models.rf3.schema import (
    RF3FoldRequest,
    RF3FoldRequestInput,
    RF3FoldParams,
    RF3Component,
    RF3EntityType,
)

request = RF3FoldRequest(
    params=RF3FoldParams(
        n_recycles=10,
        include_plddt=True,
    ),
    items=[
        RF3FoldRequestInput(
            name="with_msa",
            components=[
                RF3Component(
                    name="protein",
                    type=RF3EntityType.PROTEIN,
                    sequence="MKLLIS...",
                    chain_id="A",
                    msa_content=">seq1\nMKLLIS...\n>seq2\nMKLLVS...",
                )
            ],
        )
    ],
)

Protein-Ligand Complex¶

from models.rf3.schema import (
    RF3FoldRequest,
    RF3FoldRequestInput,
    RF3FoldParams,
    RF3Component,
    RF3EntityType,
)

request = RF3FoldRequest(
    params=RF3FoldParams(
        diffusion_batch_size=5,
        ground_truth_conformer_selection=["B"],
    ),
    items=[
        RF3FoldRequestInput(
            name="protein_ligand",
            components=[
                RF3Component(
                    name="protein",
                    type=RF3EntityType.PROTEIN,
                    sequence="MKLLIS...",
                    chain_id="A",
                ),
                RF3Component(
                    name="ligand",
                    type=RF3EntityType.LIGAND,
                    smiles="CC(=O)OC1=CC=CC=C1C(=O)O",  # Aspirin
                    chain_id="B",
                ),
            ],
        )
    ],
)

Templated Folding¶

from models.rf3.schema import (
    RF3FoldRequest,
    RF3FoldRequestInput,
    RF3FoldParams,
    RF3Component,
    RF3EntityType,
)

request = RF3FoldRequest(
    params=RF3FoldParams(
        template_selection=["A/*/1-50"],
        n_recycles=10,
    ),
    items=[
        RF3FoldRequestInput(
            name="templated_fold",
            components=[
                RF3Component(
                    name="protein",
                    type=RF3EntityType.PROTEIN,
                    sequence="MKLLIS...",
                    structure_cif="...",
                    chain_id="A",
                )
            ],
        )
    ],
)

Architecture & training¶

Architecture¶

Model Type & Innovation¶

RosettaFold3 (RF3) is an all-atom biomolecular structure prediction network from the Baker Lab (RosettaCommons). It combines a Transformer-based trunk with a diffusion-based structure generation module, enabling prediction of protein structures, multi-component complexes (protein-protein, protein-DNA, protein-RNA, protein-ligand), and non-canonical modifications.

Key innovations include: - Diffusion-based structure sampling that generates multiple diverse conformations - Support for small molecule ligands via SMILES/CCD input alongside macromolecular sequences - Template conditioning for partial structure information - Early stopping based on pLDDT confidence to save compute on low-quality predictions - Cuequivariance operations for efficient equivariant neural network computation on GPU

RF3 processes inputs through a trunk network with recycling (iterative refinement), then samples 3D coordinates via a learned diffusion process that denoises random coordinates into predicted structures.

Parameters & Layers¶

Component	Details
Architecture	Transformer trunk + diffusion structure module
Trunk recycling	2--20 recycles (default: 10)
Diffusion steps	50--500 (default: 200)
Diffusion batch size	1--10 (default: 5) output structures
Input modalities	Sequence, SMILES, MSA (A3M), structure (CIF/PDB/SDF)
Output format	mmCIF with confidence scores
Checkpoint	rf3_foundry_01_24_latest.ckpt (default)
Framework	PyTorch + Lightning + AtomWorks

Training Data¶

Property	Details
Training structures	PDB (protein, nucleic acid, ligand complexes)
Sequence databases	UniRef90, MGnify, Small BFD (for MSA generation)
Ligand representations	CCD (Chemical Component Dictionary), SMILES
Training cutoff	09/21 for benchmark checkpoint; 01/24 for latest

Loss Function & Objective¶

RF3 is trained with a combination of: - Diffusion loss for structure denoising (noise-conditioned score matching) - Auxiliary losses including pTM (predicted TM-score), pLDDT (predicted LDDT), and PAE (Predicted Aligned Error) - Clash penalty for steric violations

Tokenization / Input Processing¶

Protein input: Amino acid sequences with optional MSA (A3M format)
DNA/RNA input: Nucleotide sequences
Ligand input: SMILES strings or CCD codes
Structure templates: mmCIF, PDB, or SDF format (inline or file path)
MSA provision: Direct A3M content, file path, or database alignment dictionary
Component specification: Each input entity is a named component with type, sequence/structure, and optional chain ID
Bonds: Custom covalent bonds can be specified between atoms
Internal format: Inputs are converted to a JSON specification consumed by the RF3InferenceEngine

Performance & Benchmarks¶

Published Benchmarks¶

From Corley, Mathis et al., bioRxiv (2025):

RF3 is competitive with leading open-source biomolecular structure prediction models. It improves on several tasks including:

Task	RF3 Performance	Notes
Protein monomer folding	Competitive with AF2/AF3	Single-chain proteins
Protein-protein complexes	Competitive	Multi-chain predictions
Protein-ligand docking	Improved chiral ligand handling	Better than RF2 for chirality
Fixed-backbone docking	Improved	Template-conditioned prediction
Cyclic peptides	Supported	Via `cyclic_chains` parameter

BioLM Verification Results¶

Test Case	Input	Tolerance	Status
Simple protein fold	95-residue single chain	rel_tol 1e-2, RMSD < 5A	PASS
Protein-ligand complex	Protein + ibuprofen SMILES	rel_tol 1e-2, RMSD < 5A	PASS
Multi-chain complex	Two protein chains	rel_tol 1e-2, RMSD < 5A	PASS
MSA-enhanced prediction	Protein + A3M MSA	rel_tol 1e-2, RMSD < 5A	PASS

Comparison to Alternatives¶

Model	Strength	When to prefer
RF3	All-atom; ligands; templates; diffusion sampling	Multi-modal complexes; ligand binding; template-guided
AlphaFold2/3	Well-validated; highest accuracy on many benchmarks	Maximum single-chain accuracy
Boltz	Open-source AF3-like	Similar use cases to RF3
Chai-1	Multi-modal structure prediction	Alternative to RF3/Boltz
ESMFold	Very fast; no MSA needed	Rapid screening; single chains

Strengths & Limitations¶

Pros¶

Supports diverse biomolecular inputs: proteins, DNA, RNA, ligands, and their complexes
Diffusion-based sampling generates multiple diverse conformations per prediction
Template conditioning allows incorporating partial structural knowledge
Early stopping on low-confidence predictions saves compute
Provides comprehensive confidence metrics (pTM, ipTM, pLDDT, PAE, ranking score)
Supports cyclic peptides and non-canonical modifications
MSA-enhanced predictions for higher accuracy when alignments are available

Cons¶

Requires A100 40GB GPU -- significant resource cost
Longer runtime than single-pass methods (diffusion requires many steps)
Large model checkpoint (~several GB)
Batch size limited to 1 input item per request
Maximum sequence length of 2048 residues
Requires foundry package with specialized dependencies (cuequivariance)

Known Failure Modes¶

Very large complexes may exceed GPU memory
Low-confidence predictions trigger early stopping (by design)
Ligand predictions require valid SMILES or CCD codes
MSA quality significantly affects prediction accuracy for proteins

Implementation Details¶

Inference Pipeline¶

Request
  |-- 1. Validate input components (sequences, SMILES, structures)
  |-- 2. Convert to RF3 input specification JSON
  |-- 3. Write MSA files if alignment provided
  |-- 4. Create RF3InferenceEngine with parameters
  |-- 5. Run diffusion-based inference
  |     |-- Trunk recycling (n_recycles iterations)
  |     |-- Diffusion sampling (num_steps denoising steps)
  |     |-- Generate diffusion_batch_size structures
  |-- 6. Check for early stopping (pLDDT threshold)
  |-- 7. Read output CIF files (may be gzipped)
  |-- 8. Read confidence JSON files
  |-- 9. Optionally read PAE and pLDDT arrays
  |-- 10. Return structures with confidence scores

Memory & Compute Profile¶

Resource	Value
GPU	A100 40GB
Memory	64 GB RAM
CPU	8.0 cores
Batch size	1
Max sequence length	2048
Max output structures	10 (diffusion_batch_size)

Determinism & Reproducibility¶

Setting	Value
Torch manual seed	Yes (42 at model load)
CUDA manual seed	Yes (42, if available)
User-specified seed	Supported via `seed` parameter (default: 42)
Diffusion sampling	Stochastic; seed controls randomness

Different seeds produce different structural samples from the diffusion process. This is by design -- multiple samples enable exploration of conformational space.

Caching Behavior¶

Response caching is available as an optional, off-by-default gateway feature (BIOLM_CACHE_ENABLED) -- see the gateway docs; it is not handled by the model container. Especially useful given the high compute cost of each prediction.

Versions & Changelog¶

Version	Date	Changes
v1	2025	Initial implementation using foundry commit 6866d61

Available Checkpoints¶

Checkpoint	Filename	Use Case
Latest (default)	rf3_foundry_01_24_latest.ckpt	Production use with bugfixes
Preprint	rf3_foundry_01_24_preprint.ckpt	Reproducing paper results
Benchmark	rf3_foundry_09_21_preprint.ckpt	Fair benchmarking (09/21 PDB cutoff)

Biology¶

Molecule Coverage¶

Primary Molecule Type(s)¶

RosettaFold3 is a general-purpose all-atom biomolecular structure prediction model covering:

Proteins: Single chains and multi-chain complexes, with or without MSA
DNA: Single-stranded and double-stranded DNA
RNA: Single-stranded RNA
Small molecule ligands: Via SMILES strings or Chemical Component Dictionary (CCD) codes
Complexes: Any combination of the above (protein-protein, protein-DNA, protein-RNA, protein-ligand, etc.)

The model also supports cyclic peptides, non-canonical amino acids (via CCD codes), and covalent modifications. Template structures can be provided to condition predictions on partial experimental data.

Cross-Applicability¶

Molecule Type	Applicability	Evidence	Caveats
Protein monomers	High	Primary training target	MSA improves accuracy
Protein-protein complexes	High	Multi-chain support	Large complexes may exceed memory
Antibodies	Moderate	Handles as general proteins	Specialized models (ImmuneFold) may be more accurate for CDR loops
Protein-ligand complexes	High	SMILES/CCD input for ligands	Improved chiral ligand handling over RF2
Protein-DNA complexes	High	DNA sequence input supported	Less validated than protein-only
Protein-RNA complexes	High	RNA sequence input supported	Less validated than protein-only
Cyclic peptides	Moderate	Via `cyclic_chains` parameter	Specialized feature
Small molecules alone	Low	Designed for biomolecular context	Use RDKit/molecular docking for small molecules alone

Biological Problems Addressed¶

Protein Structure Prediction¶

Biological context: Knowing the 3D structure of a protein is fundamental to understanding its function, designing drugs that target it, and engineering proteins with desired properties. Experimental methods (X-ray crystallography, cryo-EM, NMR) are slow and expensive, while computational prediction from sequence enables rapid structural characterization.

How RF3 helps: RF3 predicts full-atom protein structures from amino acid sequences, optionally enhanced with MSA information. The diffusion-based approach generates multiple diverse conformations, allowing exploration of conformational space rather than producing a single point estimate. Confidence metrics (pTM, pLDDT) enable quality assessment.

Output interpretation: Each output is an mmCIF structure file with associated confidence scores. ptm indicates global fold confidence (>0.7 is generally reliable). plddt per-residue scores identify confident vs uncertain regions. Multiple samples from diffusion allow selection of the best structure via ranking_score.

Multi-Component Complex Prediction¶

Biological context: Most biological processes involve interactions between multiple molecules -- protein-protein signaling, transcription factor-DNA binding, ribosome-mRNA interactions, enzyme-substrate complexes. Understanding these interactions at atomic resolution is critical for drug design and mechanistic biology.

How RF3 helps: RF3 natively handles multi-component inputs: users specify each molecular entity (protein chain, DNA strand, RNA strand, or ligand) as a named component, and RF3 predicts the full complex structure including inter-molecular interfaces. The iptm (interface pTM) score specifically evaluates the quality of predicted interfaces.

Protein-Ligand Docking¶

Biological context: Drug discovery requires predicting how small molecules bind to protein targets. Accurate binding pose prediction enables virtual screening, lead optimization, and understanding of drug resistance mechanisms.

How RF3 helps: Ligands can be specified via SMILES strings or CCD codes as input components alongside protein sequences. RF3 jointly predicts protein structure and ligand binding pose. Template conditioning via ground_truth_conformer_selection can fix the ligand conformation while folding the protein around it.

Template-Guided Structure Prediction¶

Biological context: In many scenarios, partial structural information is available -- an experimental structure of a homolog, a known domain fold, or crystallographic data for part of a complex. Incorporating this information improves prediction accuracy.

How RF3 helps: The template_selection parameter allows users to specify which parts of the input should be treated as structural templates. RF3 conditions its prediction on these templates while freely predicting the remaining structure. This is valuable for: - Homology-guided folding - Loop prediction with known framework - Complex prediction with known subunit structures

Applied Use Cases¶

Published Applications¶

RF3 is based on the foundry framework from RosettaCommons. Published and anticipated use cases include:

Structure-based drug design: Predicting protein-ligand complex structures for virtual screening
Antibody-antigen complex modeling: Predicting binding interfaces for therapeutic antibody development
Protein engineering: Understanding structure-function relationships for directed evolution
Nucleic acid-protein interactions: Predicting transcription factor-DNA binding for gene regulation studies

Anticipated Use Cases¶

Cryo-EM model building: Using RF3 predictions as starting models for cryo-EM refinement
Allosteric mechanism discovery: Sampling multiple conformations via diffusion to identify allosteric states
Multi-drug resistance prediction: Modeling how mutations affect drug binding

Predecessor Models¶

RosettaFold2 (Baek et al., 2023): The previous generation from the Baker Lab, which RF3 improves upon with diffusion-based sampling and better ligand handling.
AlphaFold2 (Jumper et al., 2021): The landmark structure prediction model that RF-series models build upon and compete with.

Complementary Models¶

ESM2: Provides protein language model embeddings that can be used independently of structure prediction for sequence fitness assessment.
Boltz: Alternative open-source structure prediction that can serve as a cross-validation for RF3 predictions.
ProperMAB: Can use RF3-predicted antibody structures for extracting developability features (though ProperMAB uses ABodyBuilder2 internally).

Alternative Models¶

Alternative	Advantage over RF3	Disadvantage vs RF3
AlphaFold2/3	Higher accuracy on some single-chain benchmarks	Closed-source (AF3); no diffusion sampling (AF2)
Boltz	Similar AF3-like architecture; open-source	May not handle templates/cyclic peptides as well
Chai-1	Multi-modal; commercially supported	Different licensing terms
ESMFold	Much faster; no MSA needed	Single-chain only; lower accuracy on multi-chain

Biological Background¶

Protein Folding Problem¶

The protein folding problem -- predicting 3D structure from amino acid sequence -- was considered one of biology's grand challenges for over 50 years. AlphaFold2's breakthrough in 2020 demonstrated that deep learning could achieve experimental-level accuracy. RF3 builds on this foundation with additional capabilities for multi-molecular complexes and diffusion-based sampling.

Diffusion Models for Structure Prediction¶

RF3 uses a diffusion-based approach for structure generation. In the forward process, atomic coordinates are progressively corrupted with noise. The model learns to reverse this process -- given noisy coordinates, it predicts the denoising step that recovers the true structure. This approach has several advantages:

Multiple samples: Each denoising trajectory produces a different structure, enabling exploration of conformational space
Confidence-aware: The model can identify low-confidence predictions early and stop
Diverse outputs: Different random seeds produce structurally diverse predictions that can be ranked by confidence

Multiple Sequence Alignments (MSAs)¶

MSAs are collections of evolutionarily related sequences aligned to a query. Co-evolutionary patterns in MSAs encode 3D contact information: residue pairs that co-evolve tend to be spatially close in the folded structure. Providing MSAs significantly improves prediction accuracy, especially for proteins with many homologs. RF3 supports MSAs in A3M format from databases including UniRef90, MGnify, and Small BFD.

Confidence Metrics¶

RF3 provides several confidence measures: - pTM (predicted TM-score): Global structural similarity metric, 0--1. Values >0.7 generally indicate reliable folds. - ipTM (interface pTM): Evaluates the quality of inter-chain interfaces in multi-chain predictions. - pLDDT (predicted LDDT): Per-residue confidence, 0--100. Values >90 indicate very high confidence. - PAE (Predicted Aligned Error): Per-residue-pair distance error matrix, indicating relative positioning confidence. - Ranking score: Composite score: 0.8 * ipTM + 0.2 * pTM - 100 * has_clash.

Sources & license¶

License: BSD-3-Clause (text)

Papers

Accelerating Biomolecular Modeling with AtomWorks and RF3 — bioRxiv preprint, 2025 · DOI

Source repositories

github: https://github.com/RosettaCommons/foundry

Cite

@article{corley2025accelerating,
  title={Accelerating biomolecular modeling with atomworks and rf3},
  author={Corley, Nathaniel and Mathis, Simon and ...},
  journal={bioRxiv},
  year={2025}
}

RosettaFold3¶

At a glance¶

API & schema¶

fold¶

Usage¶

Overview¶

Architecture¶

Model Files¶

Available Checkpoints¶

Capabilities & Limitations¶

Usage Examples¶

Simple Protein Folding¶

Protein with MSA¶

Protein-Ligand Complex¶

Templated Folding¶

Architecture & training¶

Architecture¶

Model Type & Innovation¶

Parameters & Layers¶

Training Data¶

Loss Function & Objective¶

Tokenization / Input Processing¶

Performance & Benchmarks¶

Published Benchmarks¶

BioLM Verification Results¶

Comparison to Alternatives¶

Strengths & Limitations¶

Pros¶

Cons¶

Known Failure Modes¶

Implementation Details¶

Inference Pipeline¶

Memory & Compute Profile¶

Determinism & Reproducibility¶

Caching Behavior¶

Versions & Changelog¶

Available Checkpoints¶

Biology¶

Molecule Coverage¶

Primary Molecule Type(s)¶

Cross-Applicability¶

Biological Problems Addressed¶

Protein Structure Prediction¶

Multi-Component Complex Prediction¶

Protein-Ligand Docking¶

Template-Guided Structure Prediction¶

Applied Use Cases¶

Published Applications¶

Anticipated Use Cases¶

Related Models¶

Predecessor Models¶

Complementary Models¶

Alternative Models¶

Biological Background¶

Protein Folding Problem¶

Diffusion Models for Structure Prediction¶

Multiple Sequence Alignments (MSAs)¶

Confidence Metrics¶

Sources & license¶

`fold`¶