OpenGerminal

July 2, 2026 · View on GitHub

A PyRosetta-free implementation of the Germinal antibody design pipeline (VHH nanobodies and scFv).

OpenGerminal replaces PyRosetta — the primary commercial dependency in Germinal — with fully open-source alternatives, enabling the pipeline to be used in academic and commercial settings without software licensing barriers. The Germinal team added AbLang1 as a non-default alternative language model option in commit 1e1c1a5 (PR #55), but IgLM remained the default and no comparative data were reported; OpenGerminal adopts AbLang1 as the sole language model for all VHH configurations, removing IgLM entirely.

⚠️ Patent notice: The Germinal methodology is subject to a provisional patent application filed by Stanford University and Arc Institute (Mille-Fragoso et al., bioRxiv 2025). A provisional patent does not confer enforceable rights, but a formal patent may be granted in the future. OpenGerminal removes software licensing barriers (PyRosetta, IgLM) only. Users intending commercial applications are solely responsible for assessing compliance with any patent rights that may arise from the Germinal methodology. This repository and its authors make no representations regarding the patent status of the underlying design method.

📄 Preprint: Han & Li (2026). OpenGerminal: an open-source implementation of the Germinal antibody design pipeline. bioRxiv. https://doi.org/10.64898/2026.06.25.734527

What Changed

Component	Original Germinal	OpenGerminal	License
Antibody LM (gradient)	IgLM v0.1.0	AbLang1-heavy v0.3.1 (ablang2 v0.2.1) (updated by Germinal team, commit 1e1c1a5; adopted in OpenGerminal as default)	Apache 2.0
Structure relaxation	PyRosetta FastRelax	OpenMM + FASPR	MIT / Apache 2.0
SASA calculation	PyRosetta	FreeSASA	LGPL
Shape complementarity	PyRosetta	sc-rs	MIT
Interface geometry	PyRosetta	Biopython	MIT
Gradient-based design	AF-Multimer	AF-Multimer	CC BY 4.0
Initial & final cofolding filter	Chai-1 v0.6.1 (1 model)	Chai-1 v0.6.1 (1 model)	Apache 2.0
Sequence redesign	AbMPNN	AbMPNN	MIT

Note: AbLang1 was added to Germinal as a non-default alternative language model by the Germinal team (commit 1e1c1a5, PR #55/#64/#68); IgLM remained the default and no comparative benchmarking was reported. OpenGerminal adopts AbLang1 as the sole language model, removing IgLM as a selectable option, and provides the first systematic benchmarking of the two models within this architecture.

Known approximations (inherited from FreeBindCraft): binder_score, interface_dG, and interface_hbonds use fixed pass-through values, as these metrics depend on the Rosetta force field and have no open-source equivalent. binder_score and interface_dG are recorded for reference only and are not applied as filter thresholds; interface_hbonds is used as a Stage 4 filter threshold, but the fixed value always passes it, effectively disabling this filter. binder_score is also used by the Stage 2 ensemble-selection logic, which is degraded as a result (all candidates appear identical, so the first is always selected rather than the energetically optimal one); practical impact is expected to be limited.

Pipeline Overview

OpenGerminal Pipeline

Requirements

Linux x86_64
NVIDIA GPU with ≥ 40 GB VRAM (A100 recommended)
Apptainer / Singularity
AF-Multimer parameters (see below)
Chai-1 weights (auto-downloaded on first run)

Installation

1. Download the container

# Option A: Download pre-built container (recommended)
# Download opengerminal_v1.0.0.sif (6.1 GB) from Zenodo:
# https://doi.org/10.5281/zenodo.20756013

# Option B: Build from source
git clone https://github.com/teaninja/OpenGerminal.git
cd OpenGerminal
docker build -f Dockerfile.free -t opengerminal:v1.0.0 .
docker save opengerminal:v1.0.0 | gzip > opengerminal_docker.tar.gz
# Convert to sif on HPC:
apptainer build --fakeroot opengerminal_v1.0.0.sif docker-archive://opengerminal_docker.tar.gz

2. Download AF-Multimer parameters

mkdir -p params && cd params
aria2c -q -x 16 https://storage.googleapis.com/alphafold/alphafold_params_2022-12-06.tar
tar -xf alphafold_params_2022-12-06.tar
rm alphafold_params_2022-12-06.tar

Alternatively, extract from an existing germinal.sif:

apptainer exec germinal.sif tar -cf params_from_sif.tar /workspace/params
tar -xf params_from_sif.tar --strip-components=2 -C params/

3. Prepare AbLang weights directory

AbLang weights are downloaded automatically on the first run. Provide a writable directory:

mkdir -p ablang_weights

Usage

Directory structure

/your/workdir/
  opengerminal_v1.0.0.sif
  params/                    ← AF-Multimer parameters
  pdbs/
    nb.pdb                   ← VHH framework template
    <antigen>.pdb             ← target antigen structure
  chai_cache/                ← Chai-1 weights (auto-downloaded)
  ablang_weights/            ← AbLang weights (auto-downloaded)
  results/                   ← output directory
  <jobname>/
    <target>.yaml            ← target configuration
    run.sh                   ← SLURM submission script

Target configuration (YAML)

target_name: pdl1
target_pdb_path: pdbs/pdl1.pdb
target_chain: A
binder_chain: B
target_hotspots: A37,A39,A41,A96,A98
hotspot_residue: W40
dimer: false
structure_model: chai

SLURM submission script

#!/bin/bash
#SBATCH -p gpu
#SBATCH --gres=gpu:a100:1
#SBATCH --nodes=1
#SBATCH -c 8
#SBATCH --mem=80G
#SBATCH --time=12:00:00
#SBATCH --account=<your_account>

cd /your/workdir
module load apptainer

apptainer exec --nv \
  --bind $PWD/params:/workspace/params \
  --bind $PWD/results:/workspace/results \
  --bind $PWD/pdbs:/workspace/pdbs \
  --bind $PWD/chai_cache:/opt/conda/envs/germinal/lib/python3.10/site-packages/downloads \
  --bind $PWD/ablang_weights:/opt/conda/envs/germinal/lib/python3.10/site-packages/ablang2/model-weights-ablang1-heavy \
  --bind $PWD/<jobname>/<target>.yaml:/workspace/configs/target/<target>.yaml \
  --pwd /workspace \
  opengerminal_v1.0.0.sif \
  python run_germinal.py target=<target> experiment_name=<exp> structure_model=chai

Benchmark

Tested on NVIDIA A100 80GB against original Germinal (commit 88d7f85).

Metric	Original Germinal	OpenGerminal v1.0.0
PD-L1 accepted designs (seeds)	3 seeds / 274 trajectories (1.1%)	9 seeds / 499 trajectories (1.8%)
PD-L1 cofolding entry rate	51 / 274 (18.6%)	168 / 499 (33.7%)
PD-L1 avg. time per trajectory (Stage 1)	180 sec (3.0 min)	258 sec (4.3 min)
IL-3 accepted designs (seeds)	0 / 376 trajectories (0%)	0 / 729 trajectories (0%)
IL-3 cofolding entry rate	30 / 376 (8.0%)	179 / 729 (24.6%)
IL-3 avg. time per trajectory (Stage 1)	156 sec (2.6 min)	252 sec (4.2 min)
PyRosetta required	✅ Yes	❌ No
IgLM required	✅ Yes	❌ No
Commercial use	⚠️ License required	✅ Free

For full methods, benchmark details, and analysis, see the accompanying paper (preprint coming soon).

Per-trajectory runtime (Stage 1, hallucination only) is slower for OpenGerminal on both targets: 43% for PD-L1 (180s vs 258s) and 62% for IL-3 (156s vs 252s). Stage 1 does not invoke OpenMM relaxation — relaxation only occurs in Stages 2 and 4 (cofolding filters). Since Stage 1 uses the same AF-Multimer model in both versions (verified by identical checksums of the core model code), we attribute the slowdown to the AbLang1 gradient steps replacing IgLM, not the relaxation backend.

"Accepted designs (seeds)" counts only designs whose originating trajectory has a complete log record, matching the rate reported in the manuscript (Section 3.3). For PD-L1, the original Germinal pipeline produced 7 accepted seeds (16 PDB structures) in total — of which 3 are traceable to a logged trajectory; all 9 OpenGerminal accepted seeds are log-traceable. Design-quality metrics for all accepted designs (independent of log traceability) are evaluated in the accompanying paper.

License

OpenGerminal code modifications are released under the Apache 2.0 License. See LICENSE for full text.

Software dependencies

The open-source components introduced in this work are distributed under the following licenses:

Component	License
OpenMM 8.5.1	MIT License
FASPR	Apache 2.0 License
FreeSASA	LGPL v2.1
sc-rs v1.0.0	MIT License
Biopython	Biopython License Agreement (BSD-like)
ablang2 v0.2.1	Apache 2.0 License

Patent disclaimer

The Germinal pipeline methodology is subject to a provisional patent application filed by Stanford University and Arc Institute. Provisional patents do not yet confer enforceable rights, but a formal patent may be granted in future. This notice is provided for transparency; it is not legal advice. Users are solely responsible for determining whether their intended use requires a patent license from the relevant rights holders.

This repository builds upon:

Germinal (Apache 2.0) — Mille-Fragoso et al., 2025
FreeBindCraft (Apache 2.0) — Ring, 2025
AbLang2 (Apache 2.0) — Oxford Protein Informatics Group
OpenMM (MIT)
FreeSASA (LGPL)

Citation

If you use OpenGerminal, please cite:

@article{opengerminal2026,
  title   = {OpenGerminal: an open-source implementation of the Germinal antibody design pipeline},
  author  = {Han, Bing and Li, Sheng},
  journal = {bioRxiv},
  year    = {2026},
  doi     = {10.64898/2026.06.25.734527},
  url     = {https://doi.org/10.64898/2026.06.25.734527}
}

Please also cite the original Germinal paper:

@article{germinal2025,
  title   = {Efficient generation of epitope-targeted de novo antibodies with Germinal},
  author  = {Mille-Fragoso, Santiago and others},
  journal = {bioRxiv},
  year    = {2025},
  doi     = {10.1101/2025.09.19.677421}
}

Affiliation

Department of Molecular Physiology and Biological Physics, University of Virginia School of Medicine
Kenworthy Lab

This work was conducted as part of research activities in the Kenworthy Lab.

Acknowledgments

OpenGerminal is built on the work of the Germinal team (Santiago Mille-Fragoso et al.) and draws on the PyRosetta-free approach pioneered by FreeBindCraft (Aaron Ring, Ariax Bio). We thank both teams for making their work open source.