protein-property-prediction

name	protein_property_prediction
description	Physicochemical properties, surface/SS features, and finetuned protein/residue function prediction. Use when the user asks for solubility, optimal temperature, activity/binding/conserved sites, RSA/SASA/secondary structure, or property tables from FASTA/PDB. Do NOT use for zero-shot mutation ranking (zero_shot_mutation_workflow) or custom model training (train tools / future finetune workflow).
license	Apache-2.0
metadata	{"version":"1.0","skill-author":"VenusFactory2"}

Protein Property & Function Prediction

Overview

Orchestrates VenusFactory predict tools: FASTA physicochemical features, PDB RSA/SASA/SS, and finetuned sequence-level / residue-level function heads (Ankh / ESM2 / ProtT5 adapters under ckpt/).

Project Tools (VenusFactory2)

Tool	Args	When
calculate_physchem_from_fasta	fasta_file, optional out_dir	Length, MW, pI, gravy, etc.
calculate_rsa_from_pdb	pdb_file, chain_id	Relative solvent accessibility
calculate_sasa_from_pdb	pdb_file	Absolute SASA
calculate_ss_from_pdb	pdb_file, chain_id	Secondary structure
predict_protein_function	fasta_file, task, model_name	Solubility, Optimal Temperature, …
predict_residue_function	fasta_file, task, model_name	Activity / Binding / Conserved Site / Motif
predict_structure_esmfold	sequence	Need PDB before RSA/SASA/SS

Workflow

Properties from sequence

1. calculate_physchem_from_fasta
2. Optional predict_protein_function with an explicit task matching platform dataset names.

Structure features

1. Obtain PDB (protein_structure_pipeline or user upload).
2. Run calculate_rsa_from_pdb / calculate_sasa_from_pdb / calculate_ss_from_pdb as needed.

Residue sites

1. predict_residue_function with task in {Activity Site, Binding Site, Conserved Site, Motif}.
2. Cross-check with InterPro (interpro_domain_annotation) before proposing edits to catalytic residues.

Common mistakes

• Inventing task names not present in platform constant.json / ckpt mapping.
• Running RSA tools on FASTA without a structure.
• Treating finetuned scores as wet-lab labels without reporting model/task.