Overview
| Name | JAM / JAM-2 · aliases: JAM, JAM-2 |
|---|---|
| Full name | The technical report gives no expanded form |
| Ambiguous names (not auto-linked) | JAM |
| Family | Part of the JAM family |
| Organization | Nabla Bio |
| Organization type | AI-native biotech |
| Country | United States |
| Model class | Design (antibody / VHH) |
| Domain | Antibodies (also: Proteins) |
| First release | † partly verified · product update Basis: No later than that month: the technical report was covered by a third party; the primary page carries no date Note: The previous generation's preprint is from 2025-05; the company page's entry date format is ambiguous and is not used source |
| Other milestones | |
| Latest update | † unverified · product update Basis: Follow-on report on T-cell-engaging multispecific antibodies against peptide-MHC class I targets The follow-on report extends to peptide-MHC class I targets and includes a dual-variant antibody, with cryo-EM confirmation by an imaging service provider; the report itself carries no version date and the month comes from a third-party post, so it is treated as unverified source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · A follow-on report within 12 months, dated only to the month and unverified |
| Openness | Code: Closed Weights: Closed No public weights or code; the material is a company-hosted technical report, not peer reviewed |
| Benchmarks reported | 0 counted · none of the 44 |
| Wet-lab validation | 4 recorded campaigns |
| Unverified fields | † first_release, latest_update, training_data.a_public, wetlab_validation[3] |
| As of | 2026-11-16 |
What it does
Designs antibodies and VHHs de novo for multi-pass membrane targets such as GPCRs, ion channels and transporters; the newer generation extends output to drug-like VHH-Fc and full-length mAb formats, supports user-specified epitopes and steers at inference time from validated molecules. It sits with closed antibody design efforts such as Chai-2 and Latent-X/X2, opposite the academic open route of Germinal.
Benchmarks reported
None of the 44 benchmarks. This model is on the zero-pairing list.
How it evaluates itself: Target Set 1: 16 unseen targets; Head-to-head vs Chai-2 numbers; Epitope-directed set; GPCR native-cell binding; Experiment-guided steering
Other evaluations (outside the 44)
- Comparison with another closed model's published numbers (Comparison against baselines only) Self-reported: Compares published binding outcomes on overlapping targets rather than re-running the other model
- Own wet-lab evaluation on hard targets (Wet-lab validation) Self-reported: The target-set, epitope-directed and native-cell experiments are all developer-defined and map to no benchmark on this site
Developers report: Developers report fully computational design of drug-like VHH-Fc and full-length mAbs with binders on all 16 unseen targets, epitope-level control and activity against GPCRs in native cells.
Independent evidence: No independent evidence found. The material is a company-hosted technical report, not peer reviewed, and no third-party re-test was found
Training data
Group in Figure 7: Public data + wet lab for validation only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Not disclosed | The technical report does not state training data sources source |
| b. Private or licensed data | Not disclosed | source |
| c. Own wet-lab data used for training | Not disclosed | Pretraining is undisclosed; inference-time steering prompted by validated molecules is not training source |
| d. Wet lab used only for validation | Yes | All reported results come from newly run wet-lab screens 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: 16 targets, 45 designs per modality per target, a binary binding readout; hits counted per target
Mean per-target success is 39% and 18% for the two modalities; of the 16 targets, 2 reached picomolar and 7 single-digit nanomolar
Executor: Run by the developer; the report contains no independent experimental arm
Denominator basis: 20 user-specified epitopes per target and about 600 designs per epitope; the report gives no per-epitope denominator
The report states that on half the targets 30-70% of epitopes yielded binders; a range without per-target denominators
Executor: Run by the developer
Denominator basis: Two chemokine receptors tested in cells; the report gives no design-count denominator
The report also claims that test-time compute scaling raises success substantially; self-reported
Executor: Run by the developer
Denominator basis: Second-hand source only, no denominator
Generating like molecules prompted by one validated molecule is a case study, not a measured evaluation
Executor: Executor undisclosed
Background
- Organization: The developer is headquartered in Massachusetts as an AI-native antibody design company spun out of a university lab; headcount of about twenty-odd comes from public data aggregators.
- Funding: An $11M seed in 2021-12 led by two venture firms and a $26M Series A in 2024-05 led by one firm. A 2026 Series B listed by a data aggregator has no company announcement or second source and is not used.
- Products and partnerships: It collaborates with three large pharmaceutical companies; in 2025-10 it signed a second multi-year deal with one of them with potential value above a billion dollars including milestones, upfront undisclosed. Competitors include the developers of Chai-2 and Latent-X/X2.
- Related models: Chai-2, Latent-X / Latent-X2, Germinal
Reading notes
- All results are self-reported, not peer reviewed, and there is no independent experimental arm.
- "Binders on all 16 targets" is a per-target statement; per-design hit rates are much lower and the two must not be swapped.
- The target set overlaps another closed model's study to enable comparison, and the comparison is between published results.
- Pretraining data is entirely undisclosed, so it cannot be filed as public-data-only.
Caveats
- Neither the first nor the follow-on report has a primary-page date, so dates are month-level only.
- The agonist count is second-hand and that figure is no longer given on the site.
- A second-hand claim that the previous generation got one hit in twenty thousand designs could not be found in the report.
- It reports no benchmark on this site, so comparison with other models is only indirect, through each party's own wet-lab protocol.
Sources
- Technical report PDF (technical report)
- Third-party coverage (2025-11-20) (news) · 2025-11-20
- Imaging provider blog (relating to the follow-on report) (blog)
- Seed round coverage (news)
Corrections
2 corrections
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| wetlab_validation | Cited the number of agonist antibodies (300+) | The number is second-hand and unverified; removed | source | 2026-11-16 |
| latest_update | Recorded only up to the technical report | Added the 2026-06 follow-up report on peptide-MHC class I, marked unverified (month-level) | 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.