计划执行与检查节点
executing-plans
按既定实施计划完成工作,保持过程检查和反馈。
target-preparation【效率与协作】Validate a BindCraft target PDB and construct a target JSON with explicit chains, hotspots, binder length range, naming, and output-path checks.
独立 GitHub 快照 · 自动分类 · 以原作者文档为准
Use this sub-skill before launching a design campaign. It turns a target PDB and design intent into a checked, portable target-settings JSON; it does not run AF2, ProteinMPNN, PyRosetta, or a design trajectory.
Keep the target PDB trimmed to the smallest biologically justified region. Preserve the target chains and standard amino-acid residues needed for the intended interface; do not assume that trimming is harmless for a target-dependent result.
Copy the seven-key template and replace every path with a path valid on the machine that will launch BindCraft. Do not make the checked-in example's absolute paths a runtime dependency.
Set chains to the target chain IDs. Use explicit chain-qualified hotspots
for a multi-chain target when residue numbering could be ambiguous.
Set target_hotspot_residues to null for AF2-selected site discovery, or
to the documented residue/range syntax for a deliberate site. Validate
every selected chain and hotspot against the PDB when the PDB is available.
Run the bundled read-only validator:
python skills/disco/bindcraft/sub-skills/target-preparation/scripts/validate_target.py \
--target-json ./settings_target/my_target.json \
--pdb ./inputs/my_target.pdb
Omitting --pdb still validates JSON and reports declared paths, but cannot
prove that chains or residue numbers exist. The validator never creates,
rewrites, downloads, or imports from the BindCraft checkout.
Hand the validated JSON to the design route. For GPU/model, algorithm, filter, and launch decisions, see design-pipeline. For post-run artifacts and metrics, see results-analysis.
Schema and syntax details are in data-formats; input and filesystem failure recovery is in troubleshooting.
The JSON/PDB checks are CPU-safe inspection only. BindCraft's actual AF2/MPNN binder-design loop requires the prepared CUDA environment, model weights, and other documented prerequisites; a successful target validation is not evidence that a design can run or that a target will yield binders. PyRosetta licensing, AF2-weight availability, GPU memory, and target-dependent behavior remain separate launch gates.