---

name: alphafold3-pytorch description: "Use AlphaFold 3 PyTorch for protein and biomolecular structure-prediction workflows, heterogeneous molecule inputs, PDB/mmCIF/MSA/template preparation, bounded model inference, training configuration, and local CLI or Gradio operation." disable-model-invocation: true metadata: disco-role: operating license: MIT

AlphaFold 3 PyTorch

Use this skill when a task names alphafold3-pytorch, AlphaFold 3 in PyTorch, protein/complex structure prediction, biomolecular diffusion, PDB/mmCIF inputs, MSA/template features, or the package's Alphafold3/Alphafold3Input APIs. This is operating guidance for the public package, not a claim that a checkpoint, training dataset, or production-scale result is available.

Start safely

1. Install the public distribution in an isolated environment: python -m pip install alphafold3-pytorch.
2. Confirm the package and its scientific dependencies import: python -c "import torch, alphafold3_pytorch as af3; print(af3.Alphafold3)".
3. Run the read-only environment probe at scripts/check_environment.py before using CUDA, optional encoders, MSA accelerators, or a CLI checkpoint.
4. Read references/package-overview.md for verified API boundaries and choose exactly one focused route below.
5. Read references/troubleshooting.md when an import, data dependency, shape, device, checkpoint, or output failure is involved.

The package has heavy runtime dependencies and production defaults. Begin with small synthetic inputs and reduced model dimensions; do not start full training, download PDB/AFDB data, or launch the interactive app as an exploratory smoke test.

Route by task

• Model construction, forward/loss versus sampling, confidence, ranking, diffusion, checkpoint loading, or memory: use model-inference.
• Proteins, RNA/DNA, ligands, ions, atom features, batching, serialization, missing atoms, or output structure conversion: use input-representation.
• PDB/mmCIF parsing, MSA/templates, cropping, weighted sampling, or dataset curation: use data-pipeline.
• Trainer, DataLoader, YAML/Pydantic configs, conductor phases, EMA, checkpoints, Fabric, or bounded training preflight: use training-configuration.
• Console commands, checkpoint-to-mmCIF planning, local Gradio UI, entity validation, or app cache/precision behavior: use cli-serving.

When a request crosses routes, start here, then follow the owning sub-skill's explicit sibling links. Keep data preparation separate from model execution so large downloads and expensive inference are not accidentally triggered.

Public contract

The primary public objects are Alphafold3, Alphafold3Input, AtomInput, BatchedAtomInput, PDBInput, PDBDataset, Trainer, and the Pydantic/YAML configuration factories. The two console entry points are alphafold3_pytorch and alphafold3_pytorch_app. Exact signatures, defaults, feature dimensions, return modes, and CLI flags live in the nearest references, not in this router.

Scope limits

• A real checkpoint is required for meaningful inference; this skill does not provide weights or validate biological quality by a tiny synthetic forward.
• CPU checks validate API/data correctness. CUDA is an optional stronger runtime path and must be probed explicitly; CPU evidence is not CUDA evidence.
• PDB/AFDB/CCD acquisition, filtering, clustering, Nim compilation, full training, and interactive server launch are intentionally bounded or reference-only. Follow the stop conditions in the focused route.
• Before refreshing this skill for a changed checkout, read references/repo-provenance.md and compare its commit, package version, dirty state, and evidence paths.