Protein Structure Prediction — Recipes

This is a cookbook for the Protein Structure Prediction service. If you’re new to the service, start with the Tutorial — it walks you through a single-protein prediction end-to-end. Once you’re comfortable with the form, the recipes here cover the patterns that go beyond the defaults: multi-chain complexes, ligands, MSAs you supply or have computed, and reruns.

For the full reference of every form field see the Quick Reference Guide.

Choosing a recipe

Your situation

Recipe

Single protein, just want a structure

Tutorial — uses Auto → ESMFold

Multi-chain protein complex

Multi-chain complex with Boltz

Protein + a small-molecule cofactor (ATP, NAD, …)

Protein + ligand by CCD code

Protein + a novel small molecule

Protein + SMILES ligand

Need an MSA but don’t have one

Let BV-BRC compute the MSA

Have an MSA, want best-quality fold

Boltz / OpenFold / Chai with uploaded MSA

Already ran one — try a different engine

Rerun with a different tool

Recipe 1 — Multi-chain protein complex with Boltz

Goal: predict the structure of a complex (e.g. an antibody Fv with heavy + light chains, or a homo-tetramer).

Inputs: one FASTA file with multiple > records — each record becomes one chain. Soft limit: 26 chains and 10,000 total residues per job.

>heavy_chain
QVQLVQSGAEVKKPGASVKVSCKASGYTFT...
>light_chain
DIQMTQSPSSLSASVGDRVTITCRASQDIS...

Form settings:

Field

Value

Prediction Tool

Boltz (or Auto if you’ve uploaded an MSA)

Protein

the multi-chain FASTA

MSA Source

Use MSA Server or Service (recommended) or Precomputed MSA from Workspace if you already have one

Other inputs

leave empty

Job Name

something descriptive — e.g. antibody-fv-boltz

Why Boltz for complexes? Boltz, OpenFold 3, and Chai-1 are all AF3-class diffusion models with strong multimer performance. Boltz is the auto-selector’s first choice when an MSA is available; OpenFold and Chai are reasonable swaps if you want to compare.

Recipe 2 — Protein with a cofactor (CCD code)

Goal: fold a protein with a known cofactor in place (e.g. an ATP-binding domain with ATP modeled).

Inputs: the protein FASTA plus a CCD code in the Ligands input.

Form settings:

Field

Value

Prediction Tool

Boltz (or Auto)

Protein

your FASTA

Ligands → Notation

CCD codes

Ligands → input

ATP (one per line if you have more)

MSA Source

Use MSA Server or Service or Precomputed

Common CCD codes: ATP, ADP, GTP, NAD, FAD, HEM, MG, ZN, CA. Glycans use their CCD codes too (NAG, MAN, BMA).

Limitation: the form’s Notation selector carries one notation per submission. To mix CCD ligands and SMILES ligands in the same job, you need the CLI (see CLI Reference).

Recipe 3 — Protein with a SMILES ligand

Goal: fold a protein with a novel small molecule (no CCD code yet) in place — e.g. a drug candidate.

Inputs: the protein FASTA plus the molecule as a SMILES string.

Form settings:

Field

Value

Prediction Tool

Boltz

Protein

your FASTA

Ligands → Notation

SMILES strings

Ligands → input

the SMILES (e.g. CCO for ethanol, one per line for multiple)

MSA Source

Use MSA Server or Service

The form validates SMILES as you type; the first invalid line is reported inline. Standard SMILES syntax — branching parentheses must balance, atoms come from the typical organic set + halogens.

Tip: if you have the molecule’s 2D structure in another tool (e.g. RDKit, ChemDraw), export to canonical SMILES first.

Recipe 4 — Let BV-BRC compute the MSA with ColabFold

Goal: predict a structure that needs an MSA, but you don’t have one and don’t want to build it yourself.

Form settings:

Field

Value

Prediction Tool

Boltz, OpenFold, or Chai

Protein

your FASTA

MSA Source

Use MSA Server or Service

The service runs MMseqs2 against UniRef + ColabFoldDB and feeds the result to the selected engine. Expect 30 s – 3 min of extra wall time before folding starts.

When not to use this: if you already have an MSA from your own pipeline (JackHMMER, ColabFold local, etc.), upload it instead — your MSA is likely deeper / better filtered than the server’s defaults.

Recipe 5 — Uploaded MSA with Boltz, OpenFold, or Chai

Goal: highest-quality fold, with an MSA you’ve curated.

Inputs: a .a3m, .sto, or .pqt MSA file uploaded to your workspace.

Form settings:

Field

Value

Prediction Tool

Boltz (default), or OpenFold / Chai to compare

Protein

the query FASTA matching the MSA’s first sequence

MSA Source

Precomputed MSA from Workspace

MSA File

the uploaded .a3m / .sto / .pqt file

The service auto-converts A3M ↔ Parquet for the engine that needs it. Stockholm is converted to A3M-internally for Boltz / OpenFold.

Tip: a deeper MSA (more sequences) usually helps until ~256 effective sequences, after which gains flatten. ColabFold’s default depth of 5,000 raw → ~256 effective is a reasonable target.

Recipe 6 — Rerun an existing job with a different tool

Goal: you’ve run a fold and want to compare a different engine on the same inputs.

  1. Open the completed job in your workspace.

  2. Click the Rerun button in the Action Bar.

  3. The form re-opens pre-filled with the prior submission’s parameters.

  4. Change the Prediction Tool to the engine you want to compare against.

  5. Update the Job Name (the default will collide otherwise — crambin-esmfold → crambin-boltz, etc.).

  6. Submit.

This is the cleanest way to A/B-test engines without rebuilding the form from scratch.

What’s not on the form

These advanced knobs are not surfaced on the web form. Reach them via the CLI or JSON-RPC API:

  • num_samples — how many independent predictions to generate (default 5; useful for assessing ensemble disagreement)

  • num_recycles — recycling iterations (engine-specific defaults; AlphaFold defaults to 3)

  • seed — random seed for reproducibility (engine-specific)

  • output_format — restrict output to a subset (pdb, cif, report, confidence, …)

  • Per-engine flags — see each engine’s adapter in PredictStructureApp/adapters/

See the CLI Reference and API Reference for the full parameter surface.

See also