Remote MSA generation

Use the ColabFold server without local search binaries or sequence databases. Install the ml extra, then call:

from atomworks.ml.preprocessing.msa.colabfold_server import make_msas_colabfold_server

make_msas_colabfold_server(
    ["MKTAYIAKQRQISFVKSHFSRQLEERLGLIEVQ"], "output_msas/", use_templates=True
)

Sequences are sent to the public ColabFold service by default. Set host_url to use another compatible server. Results use AtomWorks’ hash-sharded MSA layout. The atomworks msa generate CLI supports local MMseqs2 and HHblits; use this Python API for remote generation.

Paired queries and templates

Use use_pairing=True to submit the sequences of one complex as a paired query. Templates are available only for unpaired queries. To generate both unpaired MSAs and paired complex MSAs, use make_msas_colabfold_server_batch() from the same module.

Find template hit tables with find_template_alignments() from atomworks.ml.preprocessing.msa.finding. To retrieve their structures, use fetch_template_structures_from_m8_file(m8_path, output_dir) from atomworks.ml.preprocessing.msa.template_structures. It uses the PDB mirror where available and downloads missing entries from RCSB. Returned files are raw structures: chain mapping, release-date filtering and template featurization remain the caller’s responsibility.