protein-mpnn
Use ProteinMPNN to design protein sequences from backbones, prepare constraint JSONL inputs, interpret scores/probability outputs, retrain models, and use custom checkpoints in a local ProteinMPNN checkout.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- protein-mpnn
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Use ProteinMPNN to design protein sequences from backbones, prepare constraint JSONL inputs, interpret scores/probability outputs, retrain models, and use custom checkpoints in a local ProteinMPNN checkout.
- disable-model-invocation
- true
02 资源文件 3
- 参考文档references/repo-provenance.md2.9 KB
- 参考文档references/repo-routing-metadata.json329 B
- 参考文档references/troubleshooting.md2.4 KB
03 SKILL.md 正文
ProteinMPNN
Use this skill when a user is working in or around a ProteinMPNN checkout and needs help with sequence design, scoring, constraint preparation, output interpretation, retraining, or custom checkpoints.
What This Skill Covers
- Running
protein_mpnn_run.pyfor single-PDB and parsed-JSONL design workflows. - Preparing helper JSONL inputs for designed chains, fixed positions, tied positions, bias, omit masks, and PSSM-guided design.
- Choosing vanilla, soluble, CA-only, or custom checkpoint weights.
- Interpreting generated FASTA headers, score/probability NPZ files, and output folders.
- Validating training data layout, adapting training commands, and routing custom checkpoints back into inference.
Minimum Runtime Assumptions
ProteinMPNN is a script-style repository, not a packaged library. Future agents should assume the user has a ProteinMPNN checkout containing protein_mpnn_run.py, protein_mpnn_utils.py, helper_scripts/, training/, and any model weights they intend to use. The Python environment needs at least PyTorch and NumPy for inference; training and CIF preparation may need additional scientific dependencies.
Do not copy model checkpoints into this skill. Point ProteinMPNN commands at the user's own checkout weights or training outputs.
Route By Task
- Run design, scoring, or probability output: use
sub-skills/inference-design/forprotein_mpnn_run.pycommands, model-family choices, direct-PDB versus JSONL mode, score-only mode, probability flags, output interpretation, and runtime troubleshooting. - Prepare constraints or JSONL inputs: use
sub-skills/constraint-inputs/forhelper_scripts/commands, parsed-PDB JSONL, chain assignment, fixed/design-only positions, tied positions, homooligomers, amino-acid bias, omit masks, PSSM dictionaries, and schema validation. - Train or use custom checkpoints: use
sub-skills/training-custom-models/for training data layout, debug training, SLURM adaptation, checkpoint naming, resume behavior, and using--path_to_model_weightswith inference.
Common First Questions
- Is the user starting from one PDB file or a folder/batch of PDBs?
- Are constraints needed before inference, such as fixed chains, tied residues, PSSM, or amino-acid bias?
- Is the desired task sequence generation, score-only evaluation, conditional/unconditional probabilities, or training?
- Which model family is intended: vanilla full-backbone, soluble full-backbone, CA-only, or custom checkpoint?
- Is the run expected to be a safe dry run, CPU-compatible smoke check, or a full GPU/HPC job?
Safe Starter Checks
Use these checks before a full design or training run:
python - <<'PY'
import numpy, torch
print('numpy', numpy.__version__)
print('torch', torch.__version__, 'cuda', torch.cuda.is_available())
PY
python protein_mpnn_run.py --help
For constraint files, use the validator bundled in sub-skills/constraint-inputs/scripts/validate_constraint_jsonl.py. For custom training data, use sub-skills/training-custom-models/scripts/check_training_layout.py.
Shared References
references/repo-provenance.mdrecords the repository evidence baseline used to create this skill.references/troubleshooting.mdcovers cross-cutting install/import, checkout layout, dependency, and routing issues.
Boundaries
- This skill is not a substitute for the original model code or model weights.
- Do not run long training jobs, GPU-heavy inference, network downloads, or notebook workflows unless the user explicitly asks and the environment is appropriate.
- Do not depend on source repo examples or docs as runtime skill content; use the bundled references and scripts here, then adapt commands to the user's checkout.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).