BoltzGen
Diffusion pipeline for binder design; jobs mount your YAML at /inputs and reuse cached weights in /ref.
Paths and expectations
- Arguments are passed verbatim to the
boltzgenentry point. - Design YAML plus referenced structures live under
/inputs; keep relative paths consistent. - Results land in
/outputs(look forfinal_ranked_designs/,all_designs_metrics.csv,results_overview.pdf). --cache=/refkeeps downloads inside the shared reference volume to skip re-fetching weights.
Quickstart
Sanity-check with the upstream vanilla protein template:
run
/inputs/example/vanilla_protein.yaml
--output=/outputs/protein_anything_quickstart
--protocol=protein-anything
--num_designs=10
--budget=2
--cache=/ref
Example YAML to pair with it:
entities:
binder:
id: B
sequence: 90-130
targets:
- file: /inputs/targets/1g13prot.cif
chains: [A]
protocol: protein-anything
output:
path: /outputs/protein_anything_quickstart
Targeted peptide scan
Short run that stops after design plus inverse folding:
run
/inputs/peptide_site.yaml
--output=/outputs/site_scan
--protocol=peptide-anything
--num_designs=24
--budget=4
--steps=design,inverse_folding
--cache=/ref
Rerank existing designs
Filtering only; explore different budgets without regenerating designs:
run
/inputs/long_run.yaml
--output=/outputs/reranked_budget10
--steps=filtering
--budget=10
--alpha=0.2
--cache=/ref
Submit
Start jobs from New Job -> BoltzGen. Keep design YAML and assets under /inputs.