Germinal

Germinal is a generative pipeline for designing antibodies that bind a chosen epitope entirely de novo, rather than discovering binders by screening large immune or synthetic libraries. Given a target antigen and a set of epitope residues, it produces complementarity-determining regions (CDRs) grafted onto a fixed antibody framework, biasing the binder toward the specified surface while preserving the constant scaffold that governs expression and developability. The method was introduced by a team from Stanford University and the Arc Institute (Mille-Fragoso, Wang, Driscoll, Dai, Widatalla, Zhang, Hie, and Gao) in a bioRxiv preprint first posted in September 2025, with version 3 released in April 2026.

The central problem Germinal addresses is the experimental cost of antibody discovery. Traditional campaigns and even many computational approaches require testing thousands of candidates to find functional binders. Germinal instead aims for high enough in silico success rates that only dozens of designs need to be expressed and assayed, making de novo antibody generation tractable in a standard wet lab without high-throughput infrastructure.

It sits alongside structure-guided binder design tools such as RFdiffusion and AlphaProteo, but is specialized for the immunoglobulin fold: it couples a structure predictor with an antibody-specific protein language model so that the designed CDRs are both structurally plausible against the epitope and consistent with natural antibody sequence statistics.

#Key Features

• Epitope-targeted design: Users specify exact target residues, and the pipeline biases binders toward that epitope while freely redesigning the CDRs and holding the antibody framework fixed.
• Antibody language model in the loop: A pretrained antibody protein language model (AbLang by default, with IgLM as an alternative) keeps designed CDRs within realistic antibody sequence space, improving expression and developability.
• Structure-based hallucination: The hallucination stage builds on ColabDesign and AlphaFold-Multimer backpropagation, with AlphaFold3-class cofolding (AF3, Chai-1, or the open-source Protenix) for evaluation.
• Low-n experimental validation: Functional binders were recovered after testing only 43-101 designs per antigen, a regime compatible with low-cost, manual wet-lab workflows.
• Nanobody and scFv formats: VHH (nanobody) design is fully validated; single-chain variable fragment (scFv) support is included but preliminary.

#Technical Details

Germinal runs a three-stage loop: (1) hallucination of antibody sequences with ColabDesign, biased toward the chosen epitope; (2) selective CDR redesign with AbMPNN, an antibody-adapted ProteinMPNN; and (3) cofolding of the antibody-antigen complex with a structure prediction model, followed by structural filtering. The antibody language model scores and constrains CDR sequences during the process. Across epitopes from four diverse antigens (PD-L1, IL3, IL20, and BHRF1), the authors tested only 43-101 de novo nanobodies per target and reported experimental success rates of roughly 4-22%, recovering 2-11 binders per target with sub-micromolar dissociation constants by BLI, the strongest reaching nanomolar affinity (around 140 nM). The code is released under Apache 2.0 and depends on JAX/GPU, PyRosetta, and large downloaded model parameters; some components (IgLM, PyRosetta) carry non-commercial academic licenses.

#Applications

Germinal is aimed at researchers and antibody engineers who need binders against a defined epitope, such as a functional site, a conserved region, or an interface that conventional immunization or panning struggles to target. Because it requires testing only dozens of designs, it lowers the barrier for academic labs and small teams to pursue de novo nanobodies for research reagents, diagnostics, and early-stage therapeutic discovery without large screening operations.

#Impact

By demonstrating functional, epitope-specific de novo antibodies at low experimental throughput, Germinal pushes generative antibody design closer to practical adoption and provides an open, reproducible reference pipeline that integrates antibody language models with modern structure predictors. As a preprint released with code and full computational and experimental protocols, its long-term influence will depend on independent reproduction across more antigens and formats; the demonstrated affinities approach but do not yet reach mature therapeutic potency, and scFv design remains experimental. An independent, fully open-source reimplementation named OpenGerminal, released by Bing Han and Sheng Li at the University of Virginia, reproduces the three-stage architecture while removing the non-commercial PyRosetta dependency in favor of an open stack (OpenMM, FreeSASA, FASPR, and sc-rs) and substituting the AbLang1 antibody language model for IgLM. Benchmarked on the PD-L1 and IL-3 targets, it reported higher cofolding pass rates (33.7% versus 18.6% for PD-L1 and 24.6% versus 8.0% for IL-3) at roughly 1.5 times the runtime, providing the first systematic comparison of the two antibody language models within this pipeline. The Germinal method is covered by a provisional patent filed by Stanford University and the Arc Institute, which downstream commercial users must account for.