AI for Bio · Model

Pinal

Natural-language-to-protein design: turns a function description into structure tokens, then a sequence

ProteinsAcademic labChinaActive中文

Overview

NamePinal · aliases: Pinal, Denovo-Pinal
Full nameNo separate expanded name
Ambiguous names (not auto-linked)None
FamilyNot a family entry. Single-model entry, not a family
OrganizationWestlake University
Organization typeAcademic lab
CountryChina
Model classText-to-protein generation
DomainProteins
First release · preprint
Basis: bioRxiv v1 preprint source
Other milestones
  • Preprint v1: · bioRxiv v1 source
  • Journal online: · Nature online publication source
Latest update · journal
Basis: Nature online publication
Nature online publication; later GitHub and Hugging Face updates are minor documentation changes source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Journal publication within 12 months
OpennessCode: Open · Apache-2.0
Weights: Open · Apache-2.0
Weights and code on GitHub and Hugging Face under Apache-2.0; the full training corpus is withheld
Benchmarks reported0 counted · none of the 44
Wet-lab validation4 recorded campaigns
Unverified fieldsNone
As of2026-11-16

What it does

Translates a natural-language function description into structure tokens and then designs a sequence; about 16B parameters trained on 1.7B protein-text pairs. Built on the same team's SaProt.

Benchmarks reported

None of the 44 benchmarks. This model is on the zero-pairing list.

How it evaluates itself: Computational evaluation; Fluorescent protein; Metabolic H protein; PET hydrolase and alcohol dehydrogenase

Other evaluations (outside the 44)

  • Computational evaluation (Own task set) Self-reported: Foldability, description-function alignment and sequence diversity versus baselines; no summary number

Developers report: Developers report natural-language descriptions can be turned into functional proteins, with activity seen in four tasks without iterative optimization

Independent evidence: No independent evidence found. No independent evidence found

Training data

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

ClassValueDetail
a. Public databasesYes1.7B synthetic protein-text pairs generated from public sequence, structure, domain and enzyme annotation resources
b. Private or licensed dataNo
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationYesWet lab validated four tasks only

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.

Fluorescent protein (bioRxiv v9)
Targets — · Designs tested 29 · Hits 1 · Success 1 fluorescent design · Best affinity —

Denominator basis: 50,000 sequences generated, 29 filtered candidates tested, 1 fluoresced green
Verified against the v9 full text, so the earlier unverifiable flag is lifted; developer-run

Self-reported · No sourced date · verified · source
Metabolic H protein (bioRxiv v9)
Targets — · Designs tested 20 · Hits 5 · Success 14/20 (70%) soluble; 5 active in a reconstituted CO2-fixation system · Best affinity glycine titer 11.1 mM, 1.7x the native protein

Denominator basis: 50,000 initial candidates filtered to ~1,000 and 20 synthesised: 14/20 soluble, 5 active in a reconstituted system
Verified against the v9 full text; developer-run

Self-reported · No sourced date · verified · source
PET hydrolase (bioRxiv v9)
Targets — · Designs tested 94 · Hits — · Success — · Best affinity best variant Tm 56.84 ± 0.29 C

Denominator basis: 200,000 generated, 2,416 passed filtering, 94 synthesised and screened; no hit rate reported
Previously missing counts are now filled in; developer-run

Self-reported · No sourced date · verified · source
Alcohol dehydrogenase (bioRxiv v9)
Targets — · Designs tested 8 · Hits 5 · Success 5 expressed and purified across three subfamilies · Best affinity —

Denominator basis: 2,500 initial candidates, 1,000 high-confidence, 8 selected, 5 expressed and purified across three subfamilies
Substrate conversion shown chromatographically against a no-enzyme control with no numeric activity; developer-run

Self-reported · No sourced date · verified · source

Background

  • Organization: Academic team at Westlake University with collaborators at several universities, institutes and one biotech company; not a company
  • Funding: Funded by Chinese national and provincial research grants
  • Products and partnerships: Built on the same team's SaProt; the full corpus is withheld for dual-use reasons
  • Related models: SaProt, ProGen3

Reading notes

  1. All four wet-lab tasks are developer-run at a scale of tens of designs
  2. All counts are verified in the v9 full text, so figures can be given

Caveats

  1. No results on any of the 44 benchmarks; on the zero-pairing list under this round's rules
  2. The full training corpus is withheld, so external reproduction is incomplete

Sources

  1. bioRxiv preprint (preprint)
  2. Crossref record 10.1038/s41586-026-11205-z (journal)
  3. GitHub repository westlake-repl/Denovo-Pinal (repository)
  4. Hugging Face page westlake-repl/Pinal (repository)
  5. Journal article DOI 10.1038/s41586-026-11205-z (journal)
  6. bioRxiv preprint (preprint)

Corrections

None recorded.

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