AI for Bio · Model

AQAffinity

Open protein-ligand affinity model adding an affinity head to an open co-folding backbone

Small moleculesPlatform companyUnited StatesActive中文

Overview

NameAQAffinity · aliases: AQAffinity
Full nameOfficial materials give no expanded name
Ambiguous names (not auto-linked)None
FamilyNot a family entry. A single version with no family members
OrganizationSandboxAQ
Organization typePlatform company
CountryUnited States
Model classCo-folding / affinity prediction (affinity head)
DomainSmall molecules (also: Structure & co-folding)
First release
Basis: Developer blog post date (the weights repository was created 2025-12-05, but repository creation is inadmissible as a first release) source
Other milestones
  • First public evidence: · Official blog post source
  • Code release: † · Training and inference code released with the blog post source
  • Weights release: † · Weights public (the host page is gated with automatic approval) source
  • Official announcement: · Official blog post source
Latest update† unverified · patch
Basis: Last modified time of the weights repository
The latest sourced event is the weights repository's last modification; the commit history is gated so the nature of the change cannot be verified, and no new weights or version were found
Note: Must not be recorded as maintenance only: for a weights repository that reading is unsupported and the field displays as unverified source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Weights and code released within 12 months
OpennessCode: Open · Apache-2.0
Weights: Open · Apache-2.0
Weights plus training and inference code are open, positioned as a licence-clean commercially usable affinity baseline; the weights host page requires click-through approval
Benchmarks reported1 counted (retrospective 1)
Wet-lab validationNone recorded
Unverified fields† latest_update, release_milestones.code_release, release_milestones.weights_release, background.funding
As of2026-11-16

What it does

Attaches a reimplemented affinity head to the open co-folding backbone OpenFold3 to predict protein-ligand affinity; it differs from Boltz-2 mainly in backbone and licence rather than modelling idea, while TerraBind in the same direction is closed and relies on internal data.

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
01 CASPRetrospective dataset use—Third-partyCASP16 L1000/L3000 post hoc, figure only, L3000 overlaps training †partly verified · evidence

Other evaluations (outside the 44)

  • Out-of-distribution assay comparison on a licensed database (Comparison against baselines only) Self-reported: The repository itself states it is clearly weaker than the model it reimplements on two out-of-distribution assays

Developers report: a licence-clean, commercially usable open structure-to-affinity baseline shipped with training code, rather than an accuracy breakthrough.

Independent evidence: No independent evidence found. No third-party evaluation found; all existing comparisons come from developer materials

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesThree public activity and compound databases source
b. Private or licensed dataNoA licensed commercial activity database was used only for evaluation, not training source
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNoNo wet-lab validation reported source

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: An AI and quantum technology company spun out in 2022-03 from a quantum team formed inside a US technology group in 2016; headquartered in California (about 55 staff at spin-out, with press releases of different years giving different cities).
  • Funding: A $500M Series D in 2022-03; a Series E in 2025-04 totalling about $450M with an additional $150M bringing in two large technology companies at a valuation of about $5.75B, roughly $950M raised in total (all from wire-service reporting with slightly varying framings).
  • Products and partnerships: Its drug-discovery unit launched 2023-06 with two large pharma companies named as customers; on 2025-06-18 it released a synthetic protein-ligand structure dataset with affinity labels (licence wording differs across pages); in 2023-11 it announced an AI-simulation collaboration with a chipmaker; it participates in an open co-folding consortium whose backbone carries this affinity head.
  • Related models: OpenFold3, Boltz-2, TerraBind, Nesso-1

Reading notes

  1. The value lies in the open licence and training code, not in leading accuracy.
  2. In the developers' own comparison it trails the model it reimplements, so it should not be treated as state of the art.
  3. One CASP subset closely resembles the training targets, so the figure cannot stand for generalisation, and it gives no values.
  4. The nature of the latest change is unverifiable (the commit history is gated), so do not record it as maintenance only.

Caveats

  1. The nature of the latest update is unverified; the repository commit history is not visible.
  2. No verified site benchmark value exists; the only entry is retrospective dataset use.
  3. No wet-lab validation.
  4. No third-party evaluation yet.

Sources

  1. Official release blog post (company blog)
  2. Weights and model card (package registry)
  3. Weights repository metadata API (repository)
  4. Dataset release press release (press)

Corrections

2 corrections
FieldWasNowBasisFound
latest_update_kindMaintenance only (code-maintenance-only)Changed to a patch-type event and marked unverified: the weights repository commit history requires approval, so the nature of the change has no basissource2026-11-16
first_releaseThe basis included the weights repository creation date (2025-12-05)2026-01-20, based on the official blog postsource2026-11-16

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