Overview
| Name | La-Proteina · aliases: La-Proteina |
|---|---|
| Full name | No separate expanded name |
| Ambiguous names (not auto-linked) | Proteina |
| Family | Not a family entry. Single-model entry, not a family |
| Organization | NVIDIA |
| Organization type | Big-tech lab |
| Country | United States |
| Model class | All-atom protein generation |
| Domain | Proteins (also: Molecule generation) |
| First release | · preprint Basis: arXiv 2507.09466 v1 source |
| Other milestones | |
| Latest update | · new version Basis: arXiv v2 posted arXiv v2 posted; the last code push is a small fix on 2025-09-23 source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · New paper version within 12 months |
| Openness | Code: Open · Apache-2.0 Weights: Open · NVIDIA Open Model License Code Apache-2.0; weights under the NVIDIA Open Model License |
| Benchmarks reported | 1 counted (board 1) |
| Wet-lab validation | None recorded |
| Unverified fields | None |
| As of | 2026-11-16 |
What it does
Generates all-atom proteins with flow matching in a partially latent space: an explicit backbone plus per-residue latents carrying sequence and side-chain detail; co-designable proteins reach 800 residues.
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 |
|---|---|---|---|---|
| 40 MotifBench | Official leaderboard or results | — | Board | The board lists this model (29.75, 2026-07); the Chinese model page says there is no score, which is wrong. · evidence |
Other evaluations (outside the 44)
None recorded.
Developers report: Source states no developer claim
Independent evidence: No independent evidence found. No independent evidence found
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | Two public structure-database subsets (~344.5k and ~46.94M structures); PDB not used |
| b. Private or licensed data | No | |
| c. Own wet-lab data used for training | No | |
| d. Wet lab used only for validation | No | No wet lab in the paper; a downstream model did validation |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: NVIDIA is a listed GPU and AI company; this work comes from its generative-AI research lab and is internally funded; accepted at ICLR 2026
- Funding: Not disclosed or not applicable
- Products and partnerships: It is the base for Proteina-Complexa, and Dyno Psi follows the same line
- Related models: Proteina-Complexa, Dyno Psi-1 / Psi-Phi, RFdiffusion2 / RFdiffusion3
Reading notes
- Its only pairing is an official board entry; the site model page previously said none, a published error
- The paper has no wet lab; wet-lab evidence comes from a downstream model
Caveats
- Training uses predicted-structure subsets only, not the experimental structure database
- Its metrics are computational self-consistency, not wet-lab feasibility
Sources
- Official page export.arxiv.org (company blog)
- GitHub repository NVIDIA-BioNeMo/la-proteina (repository)
- arXiv preprint 2507.09466 (preprint)
Corrections
1 correction
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| benchmarks_reported | The Chinese model page said no MotifBench entry was found | The official board lists it; relation is official participation | 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.