Colabfold

ColabFold

[!note] Vault audit 2026-07-24 — USE-11 Use this for local/batch AlphaFold2-style monomer or multimer folding from FASTA; for the hosted OpenFold NIMs use openfold2-nim (monomer) / openfold3-nim (complex), for local Boltz-2 use struct-predictor, and to just retrieve existing/AlphaFold DB structures use structural-biology. Local AF2 predict vs hosted NIM vs retrieve is the routing axis.

Use this skill for local or batch AlphaFold2-style protein structure prediction with ColabFold. It is especially useful when users need a familiar reference workflow, multimer predictions, or many FASTA sequences processed consistently.

Routing

• Use struct-predictor when the task is about the vault's Boltz-oriented structure workflow.
• Use this skill when the user asks for AlphaFold2, ColabFold, colabfold_batch, pLDDT, PAE, MSA generation, or protein complex prediction.
• Use diffdock after structure prediction when the next step is ligand docking.
• Use molecular-dynamics when the next step is relaxation, simulation, or trajectory analysis.

Batch Pattern

Input FASTA:

>protein_a
MSEQUENCE...

Run:

colabfold_batch input.fasta colabfold_out

For multimers, encode chains with : in the sequence or use the format supported by the installed ColabFold version.

Interpretation

• pLDDT: local confidence. Treat low-confidence loops/tails cautiously.
• PAE: domain-domain and chain-chain placement confidence.
• Ranking score: useful for triage, not a substitute for biological validation.
• Multimer interfaces require special scrutiny; inspect PAE and interface contacts.

Practical Checks

• Remove signal peptides, tags, or low-complexity tails only when biologically justified.
• Use the same FASTA identifiers across prediction, docking, and MD outputs.
• Keep all JSON/PAE outputs, not just the top PDB.
• For publications, report ColabFold version, model preset, database/MSA mode, and whether templates were used.