Overview
| Name | Origin-1 · aliases: Origin-1, AbsciGen, AbsciBind |
|---|---|
| Full name | The name is a platform name; the preprint gives no expanded form |
| Ambiguous names (not auto-linked) | Origin |
| Family | Not a family entry. A single platform whose components are described inside one record |
| Organization | Absci |
| Organization type | AI-native biotech |
| Country | United States |
| Model class | Design (all-atom diffusion, sequence design, scoring) |
| Domain | Antibodies (also: Structure & co-folding) |
| First release | · preprint Basis: Preprint v1 posted, with a developer launch page the same day Note: Materials giving 2026-03 refer to a results-disclosure date, not first release source |
| Other milestones | |
| Latest update | · paper revision Basis: Preprint v2 The preprint was revised to v2; afterwards the public repository saw only maintenance edits to data tables and the readme (latest 2026-11-13), and no journal version was found source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · Preprint revision and repository maintenance within 12 months |
| Openness | Code: Partly open Weights: Unclear Data and evaluation are in a public repository; neither the paper nor the website states whether weights are released or under what terms |
| Benchmarks reported | 0 counted · none of the 44 |
| Wet-lab validation | 2 recorded campaigns |
| Unverified fields | † open_status.weights, training_data.c_own_wetlab_for_training, wetlab_validation[1] |
| As of | 2026-11-16 |
What it does
Conditions on an unbound antigen structure, an antibody framework and a user-specified epitope to generate the complex by all-atom diffusion, designs paired heavy and light chain sequences for six CDRs, then screens with co-folding scoring before the wet lab. Against RFantibody it is all-atom, allows antigen flexibility and explicitly targets zero-prior epitopes; against BoltzGen its antibody evidence lies in full-length antibodies rather than nanobodies.
Benchmarks reported
None of the 44 benchmarks. This model is on the zero-pairing list.
How it evaluates itself: In silico comparison vs RFantibody; AbsciBind discrimination; Zero-prior epitope design, 10 targets; IL36RA affinity maturation
Other evaluations (outside the 44)
- In-silico comparison with an open antibody design model (Comparison against baselines only) Self-reported: Compares the share of designs passing the developer's own score on four zero-prior targets; not an experimental hit-rate comparison
- Own assessment of the scoring module (Own task set) Self-reported: Separates binders from non-binders across eight antibody-antigen systems against several co-folding models, with some comparator values taken from other papers
Developers report: Developers report de novo full-length antibodies against targets with no known complex structure, with four of ten targets yielding developable binders from fewer than 100 designs each.
Independent evidence: No independent evidence found. No third-party assessment or replication was found; earlier critiques of this developer concern a separate 2023 piece of work, not this model
Training data
Group in Figure 7: Public data + wet lab for validation only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | An in-house antibody structure set from 10,045 public structure-database entries, covering all entries of a dedicated antibody database, with homologous complexes of the zero-prior targets excluded from every split source |
| b. Private or licensed data | No | source |
| c. Own wet-lab data used for training | Not disclosed | The methods section describes only public structure data, while the website also mentions years of proprietary data generation and closed-loop experiments without scale or composition source |
| d. Wet lab used only for validation | Yes | Binding, specificity, aggregation, structural and functional assays across ten targets 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: Ten targets with fewer than 100 designs each; hits counted per target as developable antibodies passing several orthogonal assays
Two of the targets also have cryo-EM structures matching the designs at 3.0 A and 3.1 A; the developers note few hits per target, few designs and stricter acceptance criteria
Executor: Run by the developer; no independent experimental arm
Denominator basis: An AI-guided single-mutation library whose size and number tested are not given on the launch page
The developers report a functional antagonist after maturation with a tens-fold affinity gain; self-reported
Executor: Run by the developer
Background
- Organization: The developer is a US-listed AI-native drug discovery company headquartered in Washington state with sites in New York, Switzerland and Serbia, and 140 employees at end-2025 per its annual report.
- Funding: A listed company; in 2025-01 a semiconductor company subscribed $20M of its shares per the annual report.
- Products and partnerships: The annual report lists collaborations with several research institutions and companies; the clinical pipeline centres on one antibody candidate and no trial application tied to this model is disclosed.
- Related models: RFantibody, BoltzGen, Chai-1
Reading notes
- It is relatively strong public evidence for zero-prior epitopes with full-length antibodies and structural confirmation, but hits per target and design counts are small.
- The gap against the open model uses the developer's own score on only four targets and is not an experimental hit rate.
- No antibody benchmark on this site lists it, so generalisation can only be judged by third parties reproducing it on zero-prior targets.
- Whether weights are public is unstated, which limits external reproduction.
Caveats
- All results are self-reported, with no peer-reviewed version and no independent experimental arm.
- The exact affinity-maturation value is absent from the launch page and must be sourced from the paper before citation.
- The proprietary data and closed-loop experiments mentioned on the website come without scale, so training data composition cannot be judged from them.
- Under this round's relation rules it has no countable pairing with any benchmark on this site.
Sources
- Preprint (developer PDF) (preprint) · 2026-01-14
- Preprint version record (registry)
- Developer launch page (official page)
- Data and evaluation repository (repository)
Corrections
2 corrections
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| wetlab_validation | Affinity maturation gives exact potency values | The release page gives only approximate values; exact values must be taken from the paper, so none are shown for now | source | 2026-11-16 |
| training_data.c_own_wetlab_for_training | Own experimental data used in training | The paper's methods section mentions only public structure data and the website claim has no scale, so it is treated as not disclosed | 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.