Running predictions with ColabFold

ColabFold accelerates AlphaFold2 by replacing its slow genetic-database search with MMseqs2, which builds the multiple sequence alignment in seconds against the ColabFold public server. The prediction network still loads the original AlphaFold2 weights, so accuracy is comparable while runtime drops dramatically. The localcolabfold distribution installs a command-line entry point, colabfold_batch, that runs the whole pipeline outside a notebook.

Preparing the FASTA input

>ubiquitin
MQIFVKTLTGKTITLEVEPSDTIENVKAKIQDKEGIPPDQQRLIFAGKQLEDGRTLSDYNIQKESTLHLVLRLRGG

Running colabfold_batch

colabfold_batch \
  --num-recycle 3 \
  --num-models 5 \
  --model-type alphafold2_ptm \
  --msa-mode mmseqs2_uniref_env \
  --templates \
  --amber \
  --num-relax 1 \
  --use-gpu-relax \
  ubiquitin.fasta \
  ubiquitin_out/

The positional arguments are the input (here a single-record FASTA) and the output directory. --num-recycle sets how many times the model refines its own output; three is the default and usually enough, but disordered or large targets can benefit from more. --num-models and --model-type choose how many of the five AlphaFold2 networks to run and which weight set, where alphafold2_ptm adds the pTM head that produces the ptm score, while --templates pulls PDB templates and --amber with --num-relax runs an OpenMM/Amber minimization that fixes clashes in the top-ranked model.

Inspecting the output bundle

ls ubiquitin_out/*.pdb

jq '{mean_plddt: (.plddt | add / length | . * 100 | round / 100), ptm: .ptm, max_pae: .max_pae}' \
  ubiquitin_out/ubiquitin_scores_rank_001_alphafold2_ptm_model_3_seed_000.json

ubiquitin_out/ubiquitin_relaxed_rank_001_alphafold2_ptm_model_3_seed_000.pdb
ubiquitin_out/ubiquitin_unrelaxed_rank_001_alphafold2_ptm_model_3_seed_000.pdb
ubiquitin_out/ubiquitin_unrelaxed_rank_002_alphafold2_ptm_model_4_seed_000.pdb
ubiquitin_out/ubiquitin_unrelaxed_rank_003_alphafold2_ptm_model_1_seed_000.pdb
ubiquitin_out/ubiquitin_unrelaxed_rank_004_alphafold2_ptm_model_2_seed_000.pdb
ubiquitin_out/ubiquitin_unrelaxed_rank_005_alphafold2_ptm_model_5_seed_000.pdb

{
  "mean_plddt": 96.31,
  "ptm": 0.87,
  "max_pae": 8.42
}

Models are ranked by mean pLDDT, so rank_001 is the best and its relaxed PDB is the one to open first; with --num-relax 1 only that top model gets a relaxed structure. Each corresponding scores JSON stores the per-residue pLDDT array, the predicted aligned error matrix, its max_pae, and the ptm value, and the run also writes coverage, PAE, and pLDDT plots as PNGs. Low pLDDT stretches and high off-diagonal PAE flag the parts of the fold the model is least sure about, which is far more informative than the single global number.