Germinal

Germinal Antibody and Nanobody Design

Germinal is an open pipeline for epitope-targeted de novo antibody and nanobody design. It hallucinates CDRs on a fixed framework, designs sequences with AbMPNN, and cofolds with a structure predictor (it downloads AlphaFold-Multimer params). Runnable through biomodals.

The biomodals author notes Germinal is finicky and suggests BoltzGen for general binder design; treat Germinal as the antibody-format option, not a default.

Prerequisites

| Requirement | Value | |-------------|-------| | Runner | Modal (biomodals) | | GPU | H100 (default; GPU env var) | | Setup | See [Getting started](../../docs/getting-started.md) |

How to run

git clone https://github.com/hgbrian/biomodals && cd biomodals

uv run --with modal --with PyYAML modal run modal_germinal.py \
  --target-yaml target_example.yaml \
  --max-trajectories 1 \
  --max-passing-designs 1

Key parameters

| Parameter | Default | Description | |-----------|---------|-------------| | --target-yaml | required | Target config (targetname, targetpdbpath, targetchain, binderchain, targethotspots, length) | | --run-type | vhh | vhh (nanobody) or scfv | | --max-trajectories | 100 | Trajectories to run | | --max-passing-designs | 10 | Stop after this many passing designs | | --out-dir | ./out/germinal | Output directory |

Target YAML

target_name: PDL1
target_pdb_path: target.pdb
target_chain: A
binder_chain: B
target_hotspots: "45,67,89"
length: 120

Decision tree

Antibody-format binder?
│
├─ Nanobody / VHH → germinal (run-type vhh) or mber
├─ scFv → germinal (run-type scfv)
└─ Miniprotein (not antibody) → binder-design (boltzgen, bindcraft, mosaic)

For VHH nanobodies, biomodals also has modal_mber.py (mBER) and modal_iggm.py (IgGM) as alternatives.

Cost

Adaptyv's own tests of these models showed Germinal costing about $1.60 per accepted design, averaged across 7 targets.

Troubleshooting

| Issue | Cause | Fix | |-------|-------|-----| | Pipeline fails early | Missing PyYAML | Add --with PyYAML to the invocation | | No passing designs | Hard epitope or low budget | Raise --max-trajectories | | OOM | Large target | Use the default H100 or trim the target |

---

Next: Validate with boltz or chai, rank with ipsae, filter with protein-qc.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: adaptyvbio
• Source: adaptyvbio/protein-design-skills
• License: MIT

Install and usage instructions live in the source repository linked above.