Overview
| Name | Germinal · aliases: Germinal |
|---|---|
| Full name | The paper uses a single word as the name with no expanded form |
| Ambiguous names (not auto-linked) | None |
| Family | Not a family entry. A single pipeline with no family of versions |
| Organization | Stanford University (chemical engineering), Arc Institute |
| Organization type | Academic lab |
| Country | United States |
| Model class | Design (antibody and nanobody CDR redesign) |
| Domain | Antibodies (also: Proteins) |
| First release | · paper revision Basis: Preprint v1 posted (version API) Note: The site recorded 2025-09-19, which is the date inside the DOI string source |
| Other milestones | |
| Latest update | † partly verified · code release Basis: Repository commits with no tagged release Repository commits add scFv fixes, dependencies on an antibody language model and a scoring tool, and co-folding restraint fixes; there is no tagged release and no journal version was found. Note: The commits mix fixes and dependencies, so whether this is a substantive update is arguable source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · Repository commits within 12 months but no tagged release |
| Openness | Code: Open Weights: Unclear Code is open; it trains no weights of its own and depends on an antibody language model and a structure-modelling package whose licences are described second-hand |
| Benchmarks reported | 0 counted · none of the 44 |
| Wet-lab validation | 2 recorded campaigns |
| Unverified fields | † wetlab_validation, release_milestones.code_release, latest_update, open_status.code_license, open_status.weights, background.funding |
| As of | 2026-11-16 |
What it does
Redesigns the complementarity-determining regions on a fixed framework, scores binding with a co-folding model and constrains antibody-likeness with an antibody language model, aiming at epitope-directed antibodies without animal immunisation. It follows the same backpropagation-through-a-structure-model route as BindCraft, differing in its antibody focus and the added language model, and serves as the open academic counterpart to closed commercial models such as JAM-2 and Latent-X2.
Benchmarks reported
None of the 44 benchmarks. This model is on the zero-pairing list.
How it evaluates itself: Four-target nanobody campaign; Epitope check; Internal ablations
Other evaluations (outside the 44)
- Four-target nanobody campaign (Wet-lab validation) Self-reported: Dozens of designs on each of four targets, reporter screen then binding measurement; values unverifiable this round
- Epitope check (Wet-lab validation) Self-reported: Alanine substitutions at predicted interface residues to observe the change in binding
- Internal ablations (Own task set) Self-reported: Compares co-folding guidance alone with added custom losses, and the antibody-language-model weight on developability and humanness measures
Developers report: an open pipeline that, without animal immunisation, yields nanomolar binders on all four tested targets from under about a hundred designs each, with humanness improved by antibody-language-model guidance.
Independent evidence: No independent evidence found. No third-party evaluation or independent reproduction was found, and it has no pairing among the 44 benchmarks here
Training data
Group in Figure 7: Public data + wet lab for validation only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | No dedicated training; it reuses off-the-shelf weights from a co-folding model and two antibody language models, built on public sequence and structure archives source |
| b. Private or licensed data | No | |
| c. Own wet-lab data used for training | No | |
| d. Wet lab used only for validation | Yes | Wet lab is for validation only, with testing described second-hand as involving a gene-synthesis company source |
Wet-lab validation
Every rate carries its numerator and denominator. Claims that could not be verified are shown as unverified, without numbers; different paper versions are separate entries.
Denominator basis: Four targets with dozens of designs each, first screened by a reporter assay then measured for binding; the exact numerator and denominator could not be verified this round
The site previously published per-target hit counts, a 4-22% success range and nanomolar affinities; the full text was inaccessible this round
Executor: The preprint places the experiments in academic labs, with a gene-synthesis company involved in testing per second-hand sources
Denominator basis: Alanine substitutions at predicted interface residues to see whether binding drops; no design count is given
The experiment speaks to where the designs bind, not to a success rate
Executor: The same academic labs as the main campaign
Background
- Organization: A collaboration between a chemical engineering lab at a US university and a non-profit research institute in California, with no company entity; the institute's news post describes it as student-led.
- Funding: Funding is listed in the preprint acknowledgements with no amounts given.
- Products and partnerships: The output is the open pipeline itself; the same institute also produces genome and single-cell models such as Evo 2 and State.
- Related models: BindCraft, JAM / JAM-2, Latent-X / Latent-X2, Evo 2
Reading notes
- Four targets with dozens of designs each is a small sample, and the work is an unreviewed preprint.
- The hit counts and success range previously published here could not be verified this round; the full text must be re-obtained before citing.
- The target list includes known easier cases, and generalisation to hard targets is unevidenced.
- Its value is as an open, inspectable counterpart to closed commercial antibody design models rather than scale.
Caveats
- All wet-lab figures were unverifiable this round, so the page gives no numbers.
- The first-release date is corrected away from the date inside the DOI string.
- The repository has no tagged release, so the latest update is a commit date.
- The pipeline trains no weights of its own and is bounded by its upstream models.
- It has no pairing among the 44 benchmarks covered here.
Sources
- Preprint 2025 (preprint)
- Preprint version API record (archive)
- Institute news post (official site)
- Code repository (repository)
Corrections
2 corrections
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| first_release | 2025-09-19 (date in the DOI string) | 2025-09-24 (preprint v1 posted, version API) | source | 2026-11-16 |
| wetlab_validation | Per-target hit counts, 4–22% success rates and 140–560 nM affinities | The full text was rate-limited this round; all marked unverified and the numbers withdrawn | source | 2026-11-16 |
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.