| name | alphafold2-structure-prediction |
| description | Run reproducible AlphaFold2 monomer or multimer inference with pinned code, parameters, sequence databases, template cutoff, and seeds. Use when a local AlphaFold2 installation or approved container/database setup is available; use AlphaFold DB lookup instead for an existing canonical model. |
| license | MIT |
AlphaFold2 Structure Prediction
Gate
Ask once before installing or pulling a pinned container, downloading model
parameters or sequence databases, querying remote MSA services, and using GPU
compute. State that a full official setup can require hundreds of GB of downloads
and several TB of disk. Prefer an existing validated installation. Never upload
private sequences without explicit approval. Follow $cx-compute-environment.
Workflow
- Resolve exact sequences, constructs, oligomeric state, chain stoichiometry,
modifications, and template cutoff. Use
$cx-alphafold-structure-analysis
instead when AlphaFold DB already contains the desired canonical monomer.
- Pin the AlphaFold repository/container digest, model parameters, database
snapshots, model preset, template date, MSA mode, recycle count, relaxation,
seeds, GPU, and numerical precision.
- Smoke-test the pinned environment on a tiny known input. Run monomer or
multimer inference into
artifacts/<run-id>/alphafold2/; retain FASTA, MSAs,
template hits, ranked structures, PAE, confidence JSON, timings, and logs.
- Report per-residue pLDDT, PAE, pTM/ipTM where available, chain/interface
uncertainty, disordered regions, template dependence, and model diversity.
- Record the complete run with
$science-provenance and review claims with
$science-review.
Boundaries
- Prediction confidence is not experimental validation or proof of interaction.
- Audit template/database overlap for benchmarks and keep apo/holo assumptions
explicit. Do not compare runs with different database snapshots as identical.
