iggm - SKILL.md Agent Skill

name: iggm description: > Antibody and nanobody CDR design using IgGM (generative model by TencentAI4S). Use this skill when: (1) Designing nanobody (VHH) CDR loops against a target, (2) Designing full antibody (heavy + light chain) CDRs, (3) Redesigning existing antibody CDRs, (4) Need antigen-conditioned antibody generation, (5) Generating diverse antibody candidates with specific epitope targeting.

For VHH mask-based design (scaffold preserved), use mber. For general protein binder design, use boltzgen or bindcraft. license: MIT category: design-tools tags: [antibody, nanobody, vhh, sequence-design, generative] source: https://github.com/hgbrian/biomodals

IgGM Antibody & Nanobody Design

IgGM is a generative model for functional antibody and nanobody design from TencentAI4S. It designs CDR loops conditioned on a target antigen structure.

Prerequisites

Requirement	Minimum	Recommended
GPU VRAM	16GB	24GB (A10G)

FASTA Format (Critical)

IgGM uses a specific FASTA format with chain identifiers:

Header	Meaning
`>H`	Heavy chain (or VHH for nanobody)
`>L`	Light chain (antibody only)
`>A`, `>B`, ...	Antigen chain ID in the PDB file

Use X to mark positions to design (CDR loops).

Nanobody (VHH) design

>H
QVQLVESGGGLVQPGGSLRLSCAASGFTFSXXXXXXXXXXXXXXXXXXXXXXTRV...CDR3...WGQGTLVTVSS
>A
(empty line — specifies chain A from antigen PDB)

Full antibody design

>H
VQLVESGGGLVQPGGSLRLSCAASXXXXXXXYMN...CDR2...WVRQAPGKGLEW...CDR3...
>L
DIQMTQSPSSLS...XXXXXXWYQQKPGKAPKLL...CDR2...KASSLES...CDR3...
>A
(specifies chain A from antigen PDB)

How to run

# Nanobody design against PDB target (chain A)
modal run modal_iggm.py \
  --input-fasta nanobody.fasta \
  --antigen antigen.pdb \
  --epitope "41,42,43" \
  --task design

# Antibody design
modal run modal_iggm.py \
  --input-fasta antibody.fasta \
  --antigen target.pdb \
  --task design \
  --num-designs 50

# Just CDR3 redesign (fix framework, design CDR3 only)
modal run modal_iggm.py \
  --input-fasta template.fasta \
  --antigen target.pdb \
  --task design

Key parameters

Parameter	Default	Description
`--input-fasta`	required	FASTA with X-marked positions
`--antigen`	required	Target PDB file
`--epitope`	None	Comma-separated residue numbers
`--task`	`design`	`design` or `optimize`
`--num-designs`	10	Number of sequences to generate

Output format

output/
├── design_0.fasta    # Designed sequences
├── design_0.pdb      # Co-folded complex
└── scores.json       # Confidence scores

IgGM vs mBER

Aspect	IgGM	mBER
Input	X-masked FASTA	Masked VHH sequence
Antigen conditioning	Yes (explicit epitope)	Yes (via AF-Multimer)
Output type	CDR loops	Full VHH sequence
Antibody support	Yes (H+L)	No (VHH only)
Best for	De novo CDR design	VHH refinement

Decision tree

Antibody/nanobody design?
├─ Full antibody (H+L chains) → IgGM ✓
├─ Nanobody (VHH) de novo → IgGM ✓
├─ VHH with known scaffold → mBER ✓
└─ General protein binder → boltzgen or bindcraft

Troubleshooting

Error	Cause	Fix
`KeyError: chain`	Antigen chain not found	Check PDB chain IDs match FASTA
`No X positions`	No design positions	Mark CDR positions with X
`CUDA out of memory`	Long antigen	Trim antigen to binding domain

Next: Validate with chai or protenix → protein-qc for filtering.