RosettaFold3

Predict¶

Run the predict action for RosettaFold3.

from biolmai import BioLM
response = BioLM(
    entity="rf3",
    action="predict",
    params={'n_recycles': 3,
     'num_steps': 50,
     'diffusion_batch_size': 1,
     'seed': 42,
     'early_stopping_plddt_threshold': 0.5,
     'one_model_per_file': False,
     'annotate_b_factor_with_plddt': False,
     'include_pae': False,
     'include_plddt': True},
    items=[{'name': 'simple_protein_fold',
      'components': [{'name': 'protein_chain',
                      'type': 'protein',
                      'sequence': 'MKLLILAVVFTVFGTAVQAAYPYDDAAQLTEEQRKNEELRGQLQPTEGRPVEMDAPPSTGPIQRLANMLYTGIISFWMCSGNMLWVFFPVVLFWALVQY',
                      'chain_id': 'A'}]}]
)
print(response)

curl -X POST https://biolm.ai/api/v3/rf3/predict/ \
  -H "Authorization: Token YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "params": {
    "n_recycles": 3,
    "num_steps": 50,
    "diffusion_batch_size": 1,
    "seed": 42,
    "early_stopping_plddt_threshold": 0.5,
    "one_model_per_file": false,
    "annotate_b_factor_with_plddt": false,
    "include_pae": false,
    "include_plddt": true
  },
  "items": [
    {
      "name": "simple_protein_fold",
      "components": [
        {
          "name": "protein_chain",
          "type": "protein",
          "sequence": "MKLLILAVVFTVFGTAVQAAYPYDDAAQLTEEQRKNEELRGQLQPTEGRPVEMDAPPSTGPIQRLANMLYTGIISFWMCSGNMLWVFFPVVLFWALVQY",
          "chain_id": "A"
        }
      ]
    }
  ]
}'

import requests

url = "https://biolm.ai/api/v3/rf3/predict/"
headers = {
    "Authorization": "Token YOUR_API_KEY",
    "Content-Type": "application/json"
}
payload = {
      "params": {
        "n_recycles": 3,
        "num_steps": 50,
        "diffusion_batch_size": 1,
        "seed": 42,
        "early_stopping_plddt_threshold": 0.5,
        "one_model_per_file": false,
        "annotate_b_factor_with_plddt": false,
        "include_pae": false,
        "include_plddt": true
      },
      "items": [
        {
          "name": "simple_protein_fold",
          "components": [
            {
              "name": "protein_chain",
              "type": "protein",
              "sequence": "MKLLILAVVFTVFGTAVQAAYPYDDAAQLTEEQRKNEELRGQLQPTEGRPVEMDAPPSTGPIQRLANMLYTGIISFWMCSGNMLWVFFPVVLFWALVQY",
              "chain_id": "A"
            }
          ]
        }
      ]
    }

response = requests.post(url, headers=headers, json=payload)
print(response.json())

library(httr)

url <- "https://biolm.ai/api/v3/rf3/predict/"
headers <- c("Authorization" = "Token YOUR_API_KEY", "Content-Type" = "application/json")
body <- list(
  params = list(
    n_recycles = 3,
    num_steps = 50,
    diffusion_batch_size = 1,
    seed = 42,
    early_stopping_plddt_threshold = 0.5,
    one_model_per_file = FALSE,
    annotate_b_factor_with_plddt = FALSE,
    include_pae = FALSE,
    include_plddt = TRUE
  ),
  items = list(
    list(
      name = "simple_protein_fold",
      components = list(
        list(
          name = "protein_chain",
          type = "protein",
          sequence = "MKLLILAVVFTVFGTAVQAAYPYDDAAQLTEEQRKNEELRGQLQPTEGRPVEMDAPPSTGPIQRLANMLYTGIISFWMCSGNMLWVFFPVVLFWALVQY",
          chain_id = "A"
        )
      )
    )
  )
)

res <- POST(url, add_headers(.headers = headers), body = body, encode = "json")
print(content(res))

POST /api/v3/rf3/predict/¶

Predict endpoint for RosettaFold3.

Request Headers:

• Content-Type – application/json

• Authorization – Token YOUR_API_KEY

Request

• params (object, optional) — Configuration parameters:

  • n_recycles (int, range: 0-20, default: 10) — Number of trunk recycles

  • num_steps (int, range: 50-500, default: 200) — Number of diffusion sampling steps

  • diffusion_batch_size (int, range: 1-10, default: 5) — Number of output structures to generate

  • seed (int, optional, default: 42) — Random seed for reproducibility

  • template_selection (array of strings, optional) — Atom selections for token-level templates

  • ground_truth_conformer_selection (array of strings, optional) — Atom selections for ground truth conformers

  • cyclic_chains (array of strings, optional) — List of chain IDs to cyclize

  • early_stopping_plddt_threshold (float, range: 0.0-1.0, default: 0.5) — pLDDT threshold for early stopping

  • one_model_per_file (bool, default: False) — Save each model to a separate file

  • annotate_b_factor_with_plddt (bool, default: False) — Annotate B-factor column with pLDDT

  • include_pae (bool, default: False) — Include Predicted Aligned Error matrix in the output

  • include_plddt (bool, default: True) — Include per-residue pLDDT scores in the output

• items (array of objects, min: 1, max: 1) — Input specification for structure prediction:

  • name (string, required) — Name for this prediction task

  • components (array of objects, min: 1, required) — List of biomolecular components:

    • name (string, required) — Component name

    • type (string, required, enum: [“protein”, “DNA”, “RNA”, “ligand”]) — Entity type

    • sequence (string, optional) — Sequence string

    • smiles (string, optional) — SMILES string for small molecule

    • ccd_code (string, optional) — Chemical Component Dictionary code

    • structure_path (string, optional) — Path to structure file (CIF/PDB/SDF)

    • structure_cif (string, optional) — Structure in mmCIF format

    • chain_id (string, optional) — Chain identifier

    • msa_path (string, optional) — Path to MSA file (.a3m)

    • msa_content (string, optional) — MSA content in A3M format

    • alignment (object, optional) — MSA alignments by database:

      • mgnify (string, optional) — Alignment for mgnify database

      • small_bfd (string, optional) — Alignment for small_bfd database

      • uniref90 (string, optional) — Alignment for uniref90 database

  • bonds (array of arrays of strings, optional) — Custom bonds as pairs of atom specifications

Example request:

http Copy

POST /api/v3/rf3/predict/ HTTP/1.1
Host: biolm.ai
Authorization: Token YOUR_API_KEY
Content-Type: application/json

      {
  "params": {
    "n_recycles": 3,
    "num_steps": 50,
    "diffusion_batch_size": 1,
    "seed": 42,
    "early_stopping_plddt_threshold": 0.5,
    "one_model_per_file": false,
    "annotate_b_factor_with_plddt": false,
    "include_pae": false,
    "include_plddt": true
  },
  "items": [
    {
      "name": "simple_protein_fold",
      "components": [
        {
          "name": "protein_chain",
          "type": "protein",
          "sequence": "MKLLILAVVFTVFGTAVQAAYPYDDAAQLTEEQRKNEELRGQLQPTEGRPVEMDAPPSTGPIQRLANMLYTGIISFWMCSGNMLWVFFPVVLFWALVQY",
          "chain_id": "A"
        }
      ]
    }
  ]
}

Status Codes:

• 200 OK – Successful response

• 400 Bad Request – Invalid input

• 500 Internal Server Error – Internal server error

Response

• results (array of arrays of objects) — One result per input item, in the order requested:

  • structure_cif (string) — Predicted structure in mmCIF format, including all atoms, chains, and coordinates

  • confidence (object) — Confidence metrics for this prediction sample:

    • ptm (float, optional) — Predicted TM-score for overall fold accuracy

    • iptm (float, optional) — Interface predicted TM-score for inter-chain interfaces

    • ranking_score (float, optional) — Scalar ranking metric derived from confidence outputs

    • has_clash (boolean, optional) — Flag indicating presence of steric clashes in the predicted structure

    • plddt (array of floats, optional, range: 0.0-100.0, length: number of polymer residues) — Per-residue pLDDT scores

    • pae (2D array of floats, optional, shape: [N, N], N = number of polymer residues) — Predicted Aligned Error matrix

  • early_stopped (boolean) — Indicates whether sampling was terminated early based on early_stopping_plddt_threshold

  • sample_idx (int) — Index of this diffusion sample within the diffusion_batch_size for the corresponding input item

Example response:

http Copy

HTTP/1.1 200 OK
Content-Type: application/json

      {
  "results": [
    [
      {
        "structure_cif": "data_simple_protein_fold_model\n#\n_entry.id       simple_protein_fold_model\n_entry.author   unknown_user\n_entry.date     2026-01-06\n_entry.time     02:19:24\n#\nloop_\n_chem_comp_bond.pdbx_ordinal \n_chem_... (truncated for documentation)",
        "confidence": {
          "ptm": 0.37070754170417786,
          "iptm": 0.0,
          "ranking_score": 0.0741,
          "has_clash": false
        },
        "early_stopped": false,
        "sample_idx": 0
      }
    ]
  ]
}

Performance¶

• Predictive accuracy relative to other BioLM structure models: - Compared to AlphaFold2, RF3 is similar on standard single-chain proteins and protein–protein interfaces (with MSAs), but provides more reliable side-chain packing and ligand geometry, especially for protein–ligand complexes and mixed L/D peptides - RF3 consistently outperforms Boltz-2 on recent PDB benchmarks for protein–protein, protein–ligand, and protein–nucleic acid interfaces (higher median interface lDDT and better ligand poses), and typically matches or improves on Chai-1 for protein–ligand and protein–nucleic acid interfaces under sequence/structure-only settings - Versus ESMFold, RF3 achieves higher backbone and side-chain accuracy for monomers when MSAs are available and is substantially more accurate on any multi-component, nucleic-acid, or ligand-containing complex

• Chirality, ligand geometry, and complex interfaces: - RF3’s chirality-aware diffusion improves stereocenter correctness for ligands and D-residues (≈88% correct ligand chiral centers on held-out tests, comparable to AF3 and notably above Boltz-2 without guidance) and maintains this without inference-time constraints - On mixed L/D macrocyclic peptides outside all training cutoffs, RF3 attains ≈86% correct chiral centers and mean backbone RMSD ≈1.7 Å, better than AF3 and Boltz-1x under matched settings - In protein–ligand docking benchmarks, RF3 yields higher median interface lDDT than Boltz-2 and earlier RosettaFold variants; when provided rigid ligand conformers or distance constraints, interface lDDT typically increases by ~0.07–0.08 and ligand-only lDDT approaches 0.99, indicating strong adherence to user-specified geometry

• Coverage of biomolecular categories and comparison to antibody-focused models: - RF3 narrows the gap between open-source models and AF3 across protein–protein, protein–ligand, protein–DNA, protein–RNA, and RNA-only benchmarks, with training extended to 2024-01 further improving median interface lDDT in all categories - On a de-leaked antibody–antigen benchmark, RF3 delivers intermediate DockQ performance between AF3 and open-source AF3-like models (Boltz-2, Chai-1); it is generally preferred over Boltz-2 for heterogeneous protein interfaces, while antibody-specialized pipelines on BioLM (e.g., ABodyBuilder3-based stacks) remain more accurate for paratope-local tasks without ligands or nucleic acids

• Robustness, geometry quality, and practical model choice on BioLM: - AtomWorks-based atom-level preprocessing standardizes bond orders, charges, occupancies, and reference conformers, reducing unphysical geometries; lower PoseBusters energies of these conformers correlate with higher RF3 accuracy, particularly for underrepresented ligands - Relative to AlphaFold2 and ESMFold, RF3’s explicit all-atom representation leads to fewer geometric artifacts for ligands, ions, and covalent modifications, and its unified handling of proteins, nucleic acids, and small molecules often makes it the preferred predictor for all-atom, multi-entity assemblies, especially in design workflows where RF3 is used to validate and rank candidate complexes generated by separate design models

Applications¶

• Structure-guided protein engineering for therapeutic and industrial proteins using all-atom complex prediction across proteins, DNA/RNA, and ligands, enabling teams to prioritize mutations that maintain global fold and binding interfaces before wet-lab work; RF3’s diffusion-based sampling and interface pLDDT/ipTM scores let you explore multiple plausible conformations of a variant and rank designs for stability or partner specificity, while being mindful that very large assemblies may exceed GPU limits and that very low-homology targets can yield lower-confidence models

• Protein–ligand binding pose prediction to support structure-based lead optimization, by jointly folding the protein and docking small-molecule ligands from SMILES or CCD codes and leveraging RF3’s explicit chiral features and atom-level diffusion to better respect stereocenters and fixed ligand conformers; medicinal chemistry and protein engineering groups can use these complex models to rationalize SAR, propose pocket mutations to improve affinity or selectivity, and triage ligand series when crystallography or cryo-EM are not yet available, recognizing that RF3 is not a full replacement for exhaustive docking or free-energy methods for subtle affinity ranking

• Enzyme active-site redesign around small molecules or transition-state analogs by conditioning RF3 on a known ligand conformer or partial backbone template and letting the model fold the protein around the bound state, using its atom-level distance conditioning to enforce key catalytic geometries; industrial biotechnology and biocatalysis teams can rapidly screen active-site mutations for compatible binding modes and access new substrate scopes, with the caveat that RF3 does not by itself predict catalytic rates or full reaction energetics and should be combined with downstream QM/MM or kinetics modeling for final selection

• Protein–DNA and protein–RNA complex modeling to support design of synthetic regulators and sequence-specific binders, using RF3’s joint modeling of proteins with nucleic acids and its training on nucleic acid distillation datasets to predict interface geometries and base- or nucleotide-contact patterns; genome engineering and synthetic biology groups can apply these models to evaluate how mutations in DNA-binding domains or RNA-binding modules affect target recognition and off-target risk, while understanding that RF3 does not replace high-throughput binding assays for final specificity profiling

• Mixed L/D cyclic peptide and macrocycle structure prediction for therapeutic peptide programs, taking advantage of RF3’s explicit chirality representation and training on mixed-chirality macrocycles to generate backbone-resolved, all-atom structures that respect specified stereocenters; peptide drug discovery teams can use these models to rationalize permeability hypotheses, guide scaffold selection, and inform NMR or cryo-EM study design, noting that RF3 provides structural hypotheses rather than full ADME or permeability predictions and is not optimized for standalone small-molecule-only conformer generation

Limitations¶

• Maximum Input Size and Throughput: RF3 requests accept exactly one prediction item per call (items list length is 1) and use a fixed batch_size of 1. The total modeled sequence length across all polymer components (proteins, DNA, RNA) must not exceed max_sequence_len = 2048 residues/nucleotides. Very large complexes, many chains, or heavily liganded systems may still fail due to GPU memory limits even if under 2048 residues.

• Sampling Cost and Latency: RF3 uses a diffusion process controlled by num_steps (default 200, allowed 50–500) and generates multiple samples per request via diffusion_batch_size (default 5, max 10, bounded by max_num_samples = 10). Increasing either parameter improves sampling diversity but linearly increases runtime and GPU memory use; RF3 is not suitable for ultra–high-throughput screening compared to lighter models (e.g., ESMFold for single chains or coarse-grained docking tools).

• Scope of Supported Inputs: Each component type must be one of protein, DNA, RNA, or ligand (see RF3EntityType). Proteins and nucleic acids are provided via sequence (optionally with MSAs using msa_content, msa_path, or alignment with RF3AlignmentDatabase keys such as 'mgnify' or 'uniref90'). Ligands are provided via smiles or ccd_code. RF3 is not a general small-molecule or materials model: purely small-molecule-only systems, large non-biological assemblies, or docking campaigns without a biomolecular receptor should use specialized chemoinformatics or docking tools.

• MSA and Template Dependence: RF3 can run without MSAs or templates, but protein and complex accuracy is typically higher when high-quality MSAs (alignment, msa_content, or msa_path) and appropriate template_selection / ground_truth_conformer_selection are supplied. For low-homology targets with poor MSAs and no reliable templates, accuracy can degrade; in such settings, faster single-pass models without MSA (e.g., ESMFold) may be preferable for initial screening before selective RF3 refinement.

• Confidence, Early Stopping, and Failure Modes: RF3 may terminate denoising early if the running pLDDT estimate falls below early_stopping_plddt_threshold (default 0.5), setting early_stopped = true in outputs. Early-stopped structures, or predictions with low confidence.ptm / confidence.iptm or uniformly low confidence.plddt (if include_plddt is true), should be treated as unreliable. RF3 can also yield physically unusual or weakly ranked conformations for highly flexible regions, intrinsically disordered segments, or chemistries that are rare or absent in the training data.

• When RF3 Is Not the Best Choice: RF3 is an all-atom, GPU-intensive model aimed at detailed complex prediction and docking with full-atom coordinates. It is not optimal for: rapid folding of large numbers of single chains where backbone-only accuracy is sufficient (consider ESMFold or AlphaFold2-style monomer runs); antibody-specific modeling and developability (antibody-focused models generally perform better on CDR loops and paratopes); exhaustive virtual screening over large ligand libraries against a fixed receptor (consider docking engines or ML docking surrogates); or workflows that primarily need sequence-level embeddings or generative design (use protein language models or sequence generators, then optionally re-rank a small subset with RF3).

How We Use It¶

RosettaFold3 underpins BioLM’s structure-centric workflows as a single, all-atom engine for proteins, nucleic acids, and protein–ligand complexes. We use RF3 structures as the core structural layer for protein engineering campaigns (enzyme redesign, antibody maturation, mixed L/D macrocycles) to rank generative candidates, interpret multi-round mutational scans, and connect 3D context to manufacturability and developability assessments. RF3’s handling of ligands, cyclic peptides, and mixed biomolecular assemblies enables standardized structure prediction across discovery stages, while our data science and ML engineering patterns tie RF3 outputs to sequence embeddings, property predictors, and assay data for lab-in-the-loop optimization.

• RF3 structures are combined with BioLM generative models and ProteinMPNN/LigandMPNN-style sequence design to support closed-loop design–predict–test cycles for enzymes, binders, and small-molecule binding sites.

• For commercial programs such as antibody–antigen complexes and protein–ligand docking, RF3-based complex models enable higher-confidence hit triage, tighter experimental panels, and scalable in silico screening that reduces variants advanced to synthesis and wet-lab testing.

Related¶

• AlphaFold2 – High-accuracy protein structure predictor for single chains and simple complexes; useful as a baseline or cross-check alongside RF3’s all-atom, ligand-aware complex predictions.

• Chai-1 – Multi-modal complex structure prediction model; can be run in parallel with RF3 to compare interface geometries and confidence on difficult protein–protein and protein–ligand assemblies.

• Boltz-2 – AF3-like diffusion model for biomolecular complexes; complements RF3 for ensemble predictions and robustness checks on protein–protein, protein–nucleic acid, and protein–ligand interfaces.

• ESMFold – Fast single-chain protein structure predictor without MSAs; suitable for rapid screening or generating initial monomer models that can be refined or assembled into complexes with RF3.

References¶

• Corley, N., Mathis, S., Krishna, R., Bauer, M. S., Thompson, T. R., Ahern, W., Kazman, M. W., Brent, R. I., Didi, K., Kubaney, A., McHugh, L., Nagle, A., Favor, A., Kshirsagar, M., Sturmfels, P., Li, Y., Butcher, J., Qiang, B., Schaaf, L. L., Mitra, R., Campbell, K., Zhang, O., Weissman, R., Humphreys, I. R., Cong, Q., Funk, J., Sonthalia, S., Liò, P., Baker, D., & DiMaio, F. (2025). Accelerating biomolecular modeling with AtomWorks and RF3. bioRxiv preprint.