Overview
| Name | AlphaProteo · aliases: AlphaProteo |
|---|---|
| Full name | Sources use AlphaProteo only; no expanded name |
| Ambiguous names (not auto-linked) | AlphaFold |
| Family | Not a family entry. One technical-report version only; no family members |
| Organization | Google DeepMind, Francis Crick Institute |
| Organization type | Big-tech lab |
| Country | United Kingdom |
| Model class | Design (de novo binder) |
| Domain | Proteins (also: Molecule generation) |
| First release | † partly verified Basis: Developer blog post (the technical report on arXiv, 2024-09-12, is listed separately) Note: The blog date comes from the existing note and search results; the page was not fetched this round source |
| Other milestones |
|
| Latest update | † partly verified · no later event Basis: Only the release event is sourced; no new version, weights or successor found in 2025-2026 No later sourced event found (as of 2026-11-16); whether the arXiv v1 report (2024-09-12) counts as an update is a policy question and is unverified. Note: Must never be phrased as 'no updates since' source The most recent sourced event, not a claim that nothing changed since. |
| Status | Maintained only · Closed; only the 2024-09 release event is sourced |
| Openness | Code: Closed Weights: Closed Closed: no code and no weights; still no release plan as of 2026-04 (second-hand) |
| Benchmarks reported | 0 counted · none of the 44 |
| Wet-lab validation | 3 recorded campaigns |
| Unverified fields | † latest_update, first_release, wetlab_validation[1], training_data.b_private_licensed, training_data.c_own_wetlab_for_training |
| As of | 2026-11-16 |
What it does
Given a target structure and optional hotspot residues it generates and filters candidate binder sequences, aimed at one design round plus a small wet-lab test. It was among the first big-lab design models to publish wet-lab data, and Latent-X/X2, BoltzGen and BindCraft all use it as a comparator, but it is closed and cannot be reproduced externally.
Benchmarks reported
None of the 44 benchmarks. This model is on the zero-pairing list.
How it evaluates itself: Seven-target binder wet-lab campaign; Success-rate comparison; Follow-on functional tests
Other evaluations (outside the 44)
- Seven-target binder wet-lab campaign (Wet-lab validation) Self-reported: One round of medium-throughput screening with self-reported better affinity and success than prior methods; VEGF-A succeeded and TNFa failed
- Success-rate comparison (Comparison against baselines only) Self-reported: Per-target success compared with the best prior methods; the BHRF1 value is internally inconsistent and unverified
- Follow-on functional tests (Case study) partner: The release blog describes VEGF-A and SARS-CoV-2 RBD binders tested in partner-lab assays
Developers report: Developers report that designs against seven targets, including the difficult VEGF-A, bound with 3- to 300-fold better affinity and higher success than the best prior methods after a single screening round, without iterative optimisation (self-reported).
Independent evidence: No independent evidence found. Closed and not externally runnable; later competitors only cite its numbers as comparison
Training data
Group in Figure 7: Public data + wet lab for validation only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | Distilled PDB and predicted-structure data; only the source databases are named, with no size, cutoff or filtering rules source |
| b. Private or licensed data | Not disclosed | The report does not mention private or licensed data source |
| c. Own wet-lab data used for training | Not disclosed | Whether own experimental data entered training is not disclosed source |
| d. Wet lab used only for validation | Yes | Yes: the developer's wet lab ran binding tests on eight targets, with partner functional assays 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: Eight targets attempted (BHRF1, SARS-CoV-2 RBD, IL-7Ra, PD-L1, TrkA, IL-17A, VEGF-A and TNFa) in one round of medium-throughput screening
Self-reported with no independent replication; TNFa must be described as the failed eighth target and never merged into the seven successes
Denominator basis: The body text and a figure caption of the same developer source give different per-target rates, so no figures are given (unverified)
The body text is the authoritative reading, but the source is internally inconsistent, so this record marks it partial and quotes no value
Denominator basis: Qualitative, no counts: some RBD binders blocked infection in cells
Partner-lab executed; not an independent replication
Background
- Organization: Google DeepMind is Alphabet's AI research unit (two teams merged in 2023), based in London and funded internally rather than externally; its protein design team previously built the structure-prediction series. A UK research institute contributed the wet-lab work.
- Funding: Internally funded by Alphabet; no external funding rounds
- Products and partnerships: Not commercialised; the developer says it is also exploring applications with a sister drug-design company.
- Related models: AlphaFold 3, AlphaMissense, Latent-X / Latent-X2, BindCraft
Reading notes
- The eight targets were self-chosen and the figures are within-target hit rates, not comparable to open-competition rates.
- The TNFa failure shows it is no universal solution; the seven successes must be read together with that eighth failure.
- Being closed means no third party can re-test it, so every figure stays self-reported.
- Only source databases are disclosed, with no scale or cutoff, so overlap with the evaluation targets cannot be checked.
Caveats
- Zero pairings: it reports no score on any site benchmark.
- No entry or submission record in the Adaptyv competition / BenchBB; the model is closed and cannot be run externally.
- The BHRF1 per-target rate is internally inconsistent within one source and is treated as unverified.
- Training data is disclosed only at database level and must not be written up as 'public data only'.
Sources
- AlphaProteo technical report (arXiv 2409.08022) (preprint) · †
- Developer release blog (company blog) · †
Corrections
3 corrections
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| wetlab_validation | TNFa was presented as one of the seven successful targets | TNFa is the failed eighth target and is listed apart from the seven successes | source | 2026-11-16 |
| wetlab_validation | The BHRF1 per-target rate was published as a single value | Body text and figure caption disagree, so it is marked partial and no figure is given | source | 2026-11-16 |
| latest_update | The latest update was phrased as there being no updates since release | Changed to 'no later sourced event found (as of 2026-11-16)' with event kind none-since-release | 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.