AI for Bio · Model

Proteina-Complexa

All-atom binder design model combining generative pretraining with test-time compute

ProteinsBig-tech labUnited StatesActive中文

Overview

NameProteina-Complexa · aliases: Proteina-Complexa
Full nameThe paper uses the model name only; no expanded name
Ambiguous names (not auto-linked)Proteina
FamilyNot a family entry. The predecessors Proteina / La-Proteina and the downstream enzyme-design model have their own records; no family entry
OrganizationNVIDIA (GenAIR)
Organization typeBig-tech lab
CountryUnited States
Model classDesign (all-atom binder generation + test-time optimisation)
DomainProteins (also: Molecule generation, Small molecules)
First release
Basis: Developer research page publication (conference oral paper)
Note: Earliest sourced public evidence; the method paper's arXiv v1 is 2026-03-30 source
Other milestones
  • First public evidence: · Research page publication source
  • Preprint v1: · Method paper arXiv v1 source
  • Code release: † · GitHub 1.0.0 commit source
  • Weights release: † · Checkpoints on the model catalogue (binder / ligand / enzyme-motif sets) source
  • Official announcement: · Research page source
Latest update · paper revision
Basis: Wet-lab preprint posted (separate from the method paper)
The wet-lab campaign preprint went up; earlier distinct events are the method paper v1 (2026-03-30), code and checkpoints (2026-03-16) and the research page (2026-01-26). Commits through 2026-11-11 are install and environment changes
Note: Previously recorded as a code release; kind corrected to preprint posted source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Preprint and code activity within 12 months
OpennessCode: Open · Apache-2.0 announced on the code page; repository README points to LICENSE only
Weights: Research use only · NVIDIA Open Model License (model-catalogue page also carries a proprietary header; wording inconsistent)
Training code and configs are open and checkpoints come from the company catalogue; licence wording differs between pages and the synthetic dataset is released as dimer ID lists only
Benchmarks reported1 counted (modified protocol 1)
Wet-lab validation2 recorded campaigns
Unverified fields† open_status.code_license, open_status.weights_license, release_milestones.code_release, release_milestones.weights_release
As of2026-11-16

What it does

Unifies generative modelling with hallucination-style sequence optimisation, emitting binder sequence and structure directly without a separate inverse-folding step and spending extra compute at inference to raise hit rates; staged training (monomers, synthetic dimers, PDB multimers, protein-ligand pairs) with extensions to small-molecule targets and enzyme motif design. Like BoltzGen it is open generative binder design; compared with RFdiffusion2/3 it leans on test-time compute and synthetic data scale.

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.

BenchmarkRelationEval set / protocolSource typeNote and evidence
28 Adaptyv competitions / BenchBBDeveloper report, modified protocol—Self-reportedAdded by R7/fc_zero as a missing candidate and confirmed in the review. · evidence

Other evaluations (outside the 44)

  • Enzyme motif design task set (41 tasks) (Public benchmark not among the 44) Self-reported: The paper reports 41 atomic-motif enzyme tasks against an earlier design model; that task set is outside the 44
  • Own all-to-all wet-lab panel (Wet-lab validation) Self-reported: The 127-target panel is a developer-built readout, not a public benchmark
  • Partner campaign (programme-wide 133-target framing) (Case study) Self-reported: The project page cites over a million candidates across 133 targets; the 127-target panel is a subset
  • Compute-matched in-silico success comparison (Comparison against baselines only) Self-reported: The abstract claims higher in-silico success than prior generative and hallucination methods at matched compute

Developers report: pretraining on synthetic dimer data plus test-time compute gives higher in-silico success than prior methods, with on-target hits for 91% of 127 targets in a pooled screen.

Independent evidence: No independent evidence found. Wet-lab work was partner-executed and developer-compiled; no independent evaluation found

Training data

Group in Figure 7: Public data + wet lab for validation only.

ClassValueDetail
a. Public databasesYesThe predecessor's 46M structure-sequence pairs, PDB multimers, PLINDER protein-ligand pairs, plus a synthetic dimer dataset source
b. Private or licensed dataNot disclosed
c. Own wet-lab data used for trainingNoNo claim of feeding wet-lab results back into training source
d. Wet lab used only for validationYesLarge partner wet-lab campaign used 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.

All-to-all multiplexed phage display (wet-lab preprint (posted 2026-09-18), not the method paper)
Targets 127 · Designs tested 1015870 · Hits 115 · Success on-target 115/127 (91%); target-specific 92/127 (72%); poly-specific 78/127 (61%) · Best affinity —

Denominator basis: 1,015,870 designed sequences screened against 127 targets in one multiplexed campaign
Specific and poly-specific sets are not mutually exclusive; 133 is the programme-wide total that contains this 127-target panel and must never be added to it
Executor: Partner-lab executed; not independent replication

Self-reported · · verified · source
Compute-matched comparison (computational and experimental mix) (method paper)
Targets 83 · Designs tested 32469 · Hits 702 · Success 1.87% · Best affinity —

Denominator basis: 83 targets, 32,469 codesigned sequences; 702 on-target hits of which 607 specific
Comparator numbers in the same table are the developers' own runs, not independent re-runs
Executor: Developer-reported

Self-reported · · verified · source

Background

  • Organization: A generative-AI research group inside a listed GPU and AI platform company (Santa Clara, California); its life-science model line sits under the company's open digital-biology platform and is funded from corporate R&D with no separate financing. External co-author institutions include two research institutes in the UK and Canada and a Korean national university.
  • Funding: Not disclosed
  • Products and partnerships: Wet-lab partners include a multiplexed protein measurement company, a CRO, a large pharma company and four university labs, mostly validation collaborations with undisclosed terms; a downstream enzyme-design model with an academic institute adds explicit backbone coordinates and cuts sampling steps from about 400 to 60.
  • Related models: La-Proteina, RFdiffusion2 / RFdiffusion3, BoltzGen

Reading notes

  1. 127 and 133 are different things: 127 is the pooled panel, 133 the programme-wide total that contains it; cite hits as 115/127.
  2. Wet-lab figures must cite the wet-lab preprint, not the method paper; they are separate events.
  3. A hit means at least one binding design per target, not a per-design rate, and the result depends on test-time compute, so it is not directly comparable with single-shot models.
  4. Adaptyv participation is self-reported and unconfirmed by the organiser; MotifBench is not mentioned in the paper and PLINDER appears only as training data.

Caveats

  1. The only site pairing is self-reported participation the organiser has not confirmed.
  2. The large wet-lab campaign is validation, not training; it is not a data flywheel.
  3. Licence wording differs between the code page and the model catalogue page.
  4. The benefit of the synthetic-data route rests on developer reports.

Sources

  1. Method paper, arXiv 2603.27950 (preprint)
  2. Project page (with wet-lab preprint) (official site)
  3. Research publication page (official site)
  4. Code repository (repository)
  5. Model catalogue checkpoint page (official site)

Corrections

4 corrections
FieldWasNowBasisFound
wetlab_validation92 target-specific (no denominator), multispecific omitted115/127, 92/127, 78/127, citing the wet-lab preprintsource2026-11-16
latest_update2026-03-30, code release2026-09-18, wet-lab preprint onlinesource2026-11-16
benchmarks_reportedZero pairingsAdaptyv modified-protocol self-report; the model leaves the zero-pairing listsource2026-11-16
benchmarks_reportedCandidate pairings MotifBench, PLINDERNeither is a pairing: the paper does not mention the first, and the second is used only as training datasource2026-11-16

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.