Overview
| Name | SeedProteo · aliases: SeedProteo |
|---|---|
| Full name | No separate expanded name |
| Ambiguous names (not auto-linked) | None |
| Family | Not a family entry. Single-model entry, not a family |
| Organization | ByteDance Seed |
| Organization type | Big-tech lab |
| Country | China |
| Model class | De novo binder design |
| Domain | Proteins |
| First release | · preprint Basis: arXiv 2512.24192 v1 submitted source |
| Other milestones |
|
| Latest update | · new version Basis: arXiv v2 adding in vitro results arXiv v2 adds wet-lab results; no code or weights found, access only via the web server source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · New paper version within 12 months |
| Openness | Code: Closed Weights: Closed Web server only; no weights or code listed |
| Benchmarks reported | 0 counted · none of the 44 |
| Wet-lab validation | 1 recorded campaign |
| Unverified fields | None |
| As of | 2026-11-16 |
What it does
Adapts the SeedFold folding architecture into a generator that co-designs sequence and structure for de novo target binders. Available only through a web server.
Benchmarks reported
None of the 44 benchmarks. This model is on the zero-pairing list.
How it evaluates itself: Unconditional monomer generation; In silico binder design; SPR binding of designs
Other evaluations (outside the 44)
- Unconditional monomer generation (Own task set) Self-reported: Lengths 100–1000 and three topologies, compared with several open generative models
- In silico binder design (Own task set) Self-reported: 10 targets with about 1,000 designs each, judged by its own folding model
Developers report: Developers report the highest in silico binder design success among open methods and a 70–80% wet-lab hit range on two targets
Independent evidence: No independent evidence found. Third-party work does not name it
Training data
Group in Figure 7: Public data + wet lab for validation only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | ~500k monomers from public structure databases, ~50k two-chain PDB complexes, plus ~100k synthetic domain-interaction pairs |
| b. Private or licensed data | No | |
| c. Own wet-lab data used for training | No | |
| d. Wet lab used only for validation | Yes | Wet lab used only for validation on two targets |
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: 10 designs each on two targets, all expressed; the abstract gives a 70–80% hit range, while per-target hit counts and exact best affinities sit in figures and are not cited
The paper states a CRO executed the assays; not an independent replication. Only ranges are given, not the four figure values
Background
- Organization: The developer is ByteDance's foundation-model research division, whose AI for Science work also includes the open-source Protenix; internally funded
- Funding: Not disclosed or not applicable
- Products and partnerships: Wet-lab assays were bought from a CRO; press reports say its AI drug discovery unit is being spun out, without naming this model
- Related models: SeedFold, Protenix, BoltzGen, La-Proteina
Reading notes
- The wet lab was CRO-executed and the discriminator is the same team's folding model; not independent validation
- Only two targets with 10 designs each: small scale, and per-target values cannot be cited
Caveats
- Closed, web-service only; cannot be reproduced externally
- The abstract's hit range and the figure values differ; the range is recorded
Sources
- arXiv preprint 2512.24192 (preprint)
- Official page seedfold.github.io (company blog)
Corrections
None recorded.
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.