界面已本地化;技能介绍与说明文档按所选语言由 AI 自动翻译,代码、命令、路径和专业术语保留原文,以原作者文档为准。
探索技能
效率与协作 · SKILL.md

input-representation

input-representation

【效率与协作】Build, validate, batch, serialize, and export AlphaFold 3 molecule and atom inputs without running model inference.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准

技能说明

GitHub

name: input-representation description: "Build, validate, batch, serialize, and export AlphaFold 3 molecule and atom inputs without running model inference." disable-model-invocation: true metadata: disco-role: operating license: MIT

Input representation

Use this sub-skill when a task starts from protein, RNA, DNA, ligand, or metal entities and needs a valid AlphaFold 3 input graph, atom-level features, collation, serialization, or structure-output conversion.

Route first

  1. Read the API contract for entity fields and the conversion chain.
  2. Read the feature and shape contract before supplying tensors, custom embeddings, masks, or windows.
  3. Use the workflows for direct construction, heterogeneous batches, round trips, and coordinate export.
  4. Use troubleshooting when validation fails.
  5. Run scripts/inspect_input.py --help before using the bundled, no-inference validator. It accepts a small JSON specification, emits deterministic summaries, and can validate a temporary serialization round trip.

The normal conversion is:

Alphafold3Input → molecule-length input → AtomInput → BatchedAtomInput → model-facing keyword arguments.

Prefer Alphafold3Input for sequence/SMILES/ion descriptions and AtomInput when atom-level features already exist. Keep MSA, template, PDB/mmCIF loading, cropping, and dataset preparation in data-pipeline. Keep forward, loss, sampling, checkpoint, and confidence behavior in model-inference. Keep trainer and dataloader configuration in training-configuration.

Operating guardrails

  • Validate entity alphabets, molecule counts, atom indices, feature dimensions, and padding before a model call.
  • Double-stranded sequences are represented by the supplied strand followed by its reverse complement; do not add the complement a second time.
  • A direct Alphafold3Input must contain at least one supported entity. Use explicit AtomInput data for precomputed or unusual atom-level cases.
  • The inspection helper never downloads data, trains, starts a server, or runs a model. It is a preflight tool, not a structure-quality check.
  • Treat output coordinates as a separate contract: model sampling is normally [batch, atoms, 3], while standalone alphafold3_input_to_biomolecule expects token-major [tokens, 47, 3] coordinates. Do not interchange them.
我的收藏