Overview
| Name | Claude protein-design study (general model + public tools) |
|---|---|
| Full name | Study of a general language model acting as a design agent over public protein-design and co-folding tools |
| Ambiguous names (not auto-linked) | Claude |
| Family | Not a family entry. The subject is a study configuration, not a model family; the same partners' separate 2025-10 study is listed on its own |
| Organization | Anthropic, Adaptyv Bio, Twist Bioscience |
| Organization type | Mixed |
| Country | United States, Switzerland |
| Model class | Agent (general model orchestrating public design tools) |
| Domain | Proteins (also: Text, LLMs & agents, Molecule generation) |
| First release | Basis: Developer research post published (a partner case study followed the next day) source |
| Other milestones | |
| Latest update | · no later event Basis: Partner laboratory case-study report published The latest sourced event is the partner laboratory's case-study report; a joint competition ran 2026-09-28 to 10-31 with results due 2026-12-15 and nothing citable yet source The most recent sourced event, not a claim that nothing changed since. |
| Status | Product only, no public artifact · No public model artefact: the general model is closed and what is public is prompts and datasets |
| Openness | Code: Closed Weights: Closed The general model is closed and no code was released; prompts, computational models of the designed complexes and all in vitro and computational data are public as a dataset |
| Benchmarks reported | 1 counted (retrospective 1) |
| Wet-lab validation | 2 recorded campaigns |
| Unverified fields | † training_data.a_public, training_data.b_private_licensed, background.funding, other_evaluations[0] |
| As of | 2026-11-16 |
What it does
Not a new generative network: a general language model acts as an agent, orchestrating public backbone-generation, sequence-design and co-folding filtering tools per target, after which two outside organisations express the designs and test binding. Unlike BindCraft or BoltzGen the contribution is orchestration and selection; the tools called include RFdiffusion2/3, BoltzGen, Proteina-Complexa, Protenix and ESMFold2, though the tool list appears only in second-hand summaries.
Benchmarks reported
Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 1 of 1 pairings are counted.
| Benchmark | Relation | Eval set / protocol | Source type | Note and evidence |
|---|---|---|---|---|
| 28 Adaptyv competitions / BenchBB | Retrospective dataset use | — | Third-party | Organizer-run anonymous external test on 16 prior-competition targets; not a public competition · evidence |
Other evaluations (outside the 44)
- Per-target and per-configuration hit rates (Own task set) Third-party: Both parties report hit rates per target and per design configuration with wide variation; some per-target figures appear only in second-hand coverage
- The same partners' separate 2025-10 binder study (Case study) Self-reported: An earlier study applied the same approach to a different target class and sent designs for testing; it is a separate study and its figures are not merged here
Developers report: a general model can autonomously orchestrate public tools through the whole design pipeline, yielding binding designs for most targets at a high expression rate.
Independent evidence: No independent evidence found. The wet lab was executed and compiled by an interested partner that also chose the targets; no disinterested third-party evaluation was found
Training data
Group in Figure 7: Public data + wet lab for validation only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Not disclosed | The general model's training data is not public; each public tool it calls has its own training data source |
| b. Private or licensed data | Not disclosed | |
| c. Own wet-lab data used for training | No | No dedicated training in this study; wet-lab results were not fed back source |
| d. Wet lab used only for validation | Yes | Two outside organisations handled expression and binding assays, for validation only 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: 1,440 designs on 16 targets, of which 15 targets and 1,320 designs have usable data (one target's 120 designs excluded for aggregation and uncertain data); expression 1,235 of 1,296 tested (95.3%)
Attributing 354 to a single vendor is inaccurate: it is the pooled two-vendor final call, and per-configuration rates must be listed separately rather than merged
Executor: Executed by a partner laboratory and a gene-synthesis vendor; the organiser is an interested party, so this is not an independent evaluation
Denominator basis: Best design compared with the winning binder of each of six past competitions
The published "would have won five of six competitions" overstates the source: it is a best-binder affinity comparison, not a ranking simulation, and one competition's best binder was not beaten
Executor: Executed and compiled by the partner laboratory
Background
- Organization: The design side is a US public-benefit AI company founded in 2021 and based in San Francisco (the work is credited to its science team, with no individuals listed); the wet-lab side is a European protein measurement service company and a US gene-synthesis company, with another vendor supplying a high-throughput surface plasmon resonance platform.
- Funding: The design side announced a $30B Series G in 2026-02 at a $380B post-money valuation, co-led by several investors (second-hand reporting).
- Products and partnerships: The products are the company's general model family and its science programme; with the wet-lab partner it also ran a joint competition 2026-09-28 to 10-31, with results due 2026-12-15.
- Related models: BindCraft, BoltzGen, RFdiffusion2 / RFdiffusion3, Proteina-Complexa
Reading notes
- The comparison is against historical competition data rather than a matched control, and the interested organiser chose the targets.
- The pooled hit rate swings widely with target difficulty, including a zero-hit target, so it is not a stable rate.
- Only binding was measured; there are no functional, specificity or developability conclusions.
- The tool list exists only in second-hand summaries, so the gain cannot be split between the orchestration and the tools called.
Caveats
- The final hit call pools two vendors; one vendor's raw calls differ.
- The "would have won five of six" framing is wrong; correctly, the best designs out-bound five of the winners by affinity.
- The entry is retrospective dataset use and does not count toward reported benchmarks.
- The joint competition's results are not out and cannot be cited.
Sources
- Developer research post (company blog)
- Partner laboratory case study (company blog)
- Public dataset (prompts, computational models, in vitro data) (official site)
- Second-hand summary of the tool list and per-target figures (second-hand) second-hand
- Assay platform vendor coverage (press)
Corrections
2 corrections
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| wetlab_validation | 354 binders measured by a single vendor; one hit rate of 26.8% | 354/1,320 is the final call combining two vendors (336 for a single vendor); expression 1,235/1,296 = 95.3%; per-configuration 22.6%–35.1% listed separately | source | 2026-11-16 |
| wetlab_validation | Could have won 5 of 6 competitions | The best design's affinity beats 5 of the 6 winners, and does not exceed the best binder of one of them | 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.