ai4protein Protein Engineering Hypothesis
Overview
Instruction-only skill: structure the scientific argument, then point MLS at concrete hub tools. Inspired by scientific-agent-skills hypothesis-generation, adapted to VenusFactory protein tools.
VenusFactory execution
- Load this skill.
- Produce a short plan with: Observation → Hypothesis → Computational test → Wet-lab falsifier.
- Execute tests via other skills/tools (do not reimplement).
Recommended evidence chain
| Question | Skills / tools |
|---|---|
| What is known about the protein? | uniprot_database, interpro_domain_annotation, pubmed |
| Where is it expressed? | hpa_expression_context |
| Structure confidence? | protein_structure_pipeline / alphafold_database |
| Which mutations look beneficial? | zero_shot_mutation_workflow |
| Redesign backbone-constrained seq? | proteinmpnn_design_workflow |
| Kinetics / EC? | brenda_database |
| Partners / pathways? | string_database, kegg_database |
| Train a custom head? | venus_finetune_workflow |
Output template (for CB/MLS)
### Observation
...
### Hypothesis (falsifiable)
...
### Computational tests (ordered)
1. tool/skill … success = …
### Predicted outcome
...
### Wet-lab falsifier (suggested, not executed here)
...
### Risks / confounders
- Model score ≠ experimental activity
Rules
- Never present zero-shot / PLM ranks as measured ΔΔG or activity.
- Prefer protecting catalytic residues (
interpro+ literature) before design. - End computational rounds with at least one figure when numeric tables exist.
When NOT to use
- User only wants a single tool call with no scientific framing — call that tool's skill directly.