UI はローカライズ済みです。スキルの紹介と説明文書は選択した言語へ AI が自動翻訳し、コード・コマンド・パス・専門用語は原文のままです。原作者の文書が優先されます。
スキルを探す
生産性・共同作業 · SKILL.md

input-preparation

input-preparation

Build, validate, convert, and troubleshoot AlphaFold 3 input JSON files before prediction runs. Use when creating fold inputs, validating protein/RNA/DNA/ligand entries, converting AlphaFold Server JSON, adding MSAs/templates/user CCD/bonds, or explaining schema versions.

Independent GitHub snapshot · Automated classification · Original documentation governs

スキルの説明

GitHub

name: input-preparation description: "Build, validate, convert, and troubleshoot AlphaFold 3 input JSON files before prediction runs. Use when creating fold inputs, validating protein/RNA/DNA/ligand entries, converting AlphaFold Server JSON, adding MSAs/templates/user CCD/bonds, or explaining schema versions." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

AlphaFold 3 Input Preparation

Use this sub-skill when the task is about constructing or checking AlphaFold 3 fold-input JSON, not about running inference or interpreting outputs.

Route here for

  • Creating alphafold3 dialect JSON files with name, modelSeeds, sequences, dialect, and version.
  • Validating protein, RNA, DNA, ligand, ion, MSA, template, userCCD, userCCDPath, and bondedAtomPairs fields.
  • Converting AlphaFold Server-style fold jobs into AlphaFold 3 inputs and explaining conversion limits.
  • Debugging Input.from_json(...) errors before invoking a prediction run.
  • Checking path-relative external files such as unpairedMsaPath, pairedMsaPath, mmcifPath, and userCCDPath.

Do not handle here

  • Runtime database, model, GPU, bucket, or inference flags; route to ../running-predictions/.
  • Confidence JSON, ranking scores, mmCIF output, or result interpretation; route to ../output-interpretation/.
  • Lower-level model runner, data pipeline, or Python API integration beyond input parsing; route to ../python-apis/.

Fast workflow

  1. Draft the input with dialect: "alphafold3" and version: 4 unless compatibility with an older saved input is required.
  2. Keep one prediction job per AlphaFold 3 JSON file. Use a top-level list only for AlphaFold Server JSON that will be converted.
  3. Ensure every sequence entity has an uppercase alphabetic id; use a list of IDs only for identical copies.
  4. Prefer explicit path fields for external content: unpairedMsaPath, pairedMsaPath, mmcifPath, and userCCDPath.
  5. Validate before running prediction:
python sub-skills/input-preparation/scripts/validate_fold_input.py fold_input.json

Key references

  • references/input-json.md covers schema recipes, entity fields, conversion constraints, relative paths, and validation rules.
  • references/troubleshooting.md maps common parser and preparation failures to fixes.
  • scripts/validate_fold_input.py provides a safe local parser check using the installed AlphaFold 3 package.
保存したスキル