Overview
| Name | PoET-2 · aliases: PoET-2 |
|---|---|
| Full name | The paper uses the model name only, with no expanded form |
| Ambiguous names (not auto-linked) | PoET |
| Family | Not a family entry. The earlier PoET is a separate model; this record covers PoET-2 only |
| Organization | OpenProtein.AI |
| Organization type | Platform company |
| Country | Singapore |
| Model class | Protein language model (in-context learning) |
| Domain | Proteins (also: Structure & co-folding) |
| First release | · preprint Basis: Preprint v1 submission source |
| Other milestones | |
| Latest update | · paper revision Basis: Preprint v2 revision Paper revised to v2 (v1 was 2025-08-05); no new weights or product tier found source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · Paper revised within 12 months and the platform is still sold |
| Openness | Code: Partly open Weights: Unclear The abstract says weights are public in a repository, but neither repository nor terms were verified this round; it is also sold as a commercial platform (unverified) |
| Benchmarks reported | 1 counted (self 1) |
| Wet-lab validation | None recorded |
| Unverified fields | † open_status.code, open_status.weights, open_status.weights_license, open_status.code_license, background.funding, other_evaluations[0] |
| As of | 2026-11-16 |
What it does
Adds multimodal inputs such as structure on top of its predecessor and uses a set of homologous sequences as context to score or generate new sequences, pitched as a small model needing little data. It shares the retrieval-augmented route with Profluent-E1, which uses its predecessor as a comparison, and resembles VenusREM but is built around a platform product.
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 |
|---|---|---|---|---|
| 02 ProteinGym | Developer report, official protocol | zero-shot (substitutions, indels, higher-order, clinical) plus supervised settings | Self-reported | Paper table claims the official protocol; not found on the official table · evidence |
Other evaluations (outside the 44)
- Vendor low-data claim (Comparison against baselines only) Self-reported: Marketing material claims about 30x less experimental data with no public protocol, a vendor figure (unverified)
- Ensemble with another model (Own task set) Self-reported: The paper also reports an ensemble setting with a retrieval-augmented model, a developer-built control
Developers report: Developers report that multimodal retrieval augmentation and family in-context learning let a 182M model lead zero-shot tasks, especially indels and higher-order mutants, and claim about 30x less experimental data (a vendor figure, not an independent evaluation).
Independent evidence: No independent evidence found. Self-reported results are not confirmed on the official board or by third parties
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | 62 million homolog sets from a public sequence database, with public predicted structures for structure conditioning rather than experimental ones source |
| b. Private or licensed data | No | No private or licensed training data reported |
| c. Own wet-lab data used for training | No | |
| d. Wet lab used only for validation | No |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: A platform company headquartered in a Singapore biomedical park and listed in a commercial database as founded in 2021. It offers a no-code protein design platform, free for universities and licensed to pharma, with no public headcount.
- Funding: A commercial database lists two funds as investors with an early venture round logged in 2024-01 and no amount disclosed (second-hand); the company was also selected for a government research program on next-generation protein foundation models.
- Products and partnerships: The earlier PoET was a 2023 conference paper and PoET-2 followed in 2025. A large pharma company has used the platform since 2025 and announced an expansion into antibody discovery on 2026-03-31, with terms undisclosed. Competitors include Profluent-E1.
- Related models: Profluent-E1, VenusREM, ProSST
Reading notes
- The self-reported results are not separately confirmed on the official board and there is no public submission record.
- Scores on the same benchmark are not directly comparable between retrieval-augmented and sequence-only models.
- Reaching the top tier at 182M parameters shows parameters are not the only factor, but that is not a cross-family conclusion.
- The vendor's "30x less data" is marketing, not an independent evaluation.
Caveats
- Openness unverified: the abstract claims public weights, but neither repository nor licence terms were checked this round.
- No wet-lab validation; all evaluation is computational.
- A commercial platform and open weights are claimed side by side and must be cited separately.
Sources
- PoET-2 preprint (preprint)
- Preprint v1 full text with the self-reported table (preprint)
- Company publications page (official site)
- University news story on the platform (press)
- Commercial database funding entry (second-hand) (second-hand) second-hand
- Company announcement: selected for a government program (company blog)
- ProteinGym official repository (leaderboard)
Corrections
2 corrections
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| latest_update.event_kind | Recorded as a new version | Changed to a paper revision: 2026-02-26 is preprint v2 and the description itself says there are no new weights | source | 2026-11-16 |
| latest_update.description | Said only "v2 revision, no new weights or product tier" without tying the dates together | Adds the v1 and v2 dates alongside the unchanged "no new weights" statement | 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.