AI for Bio · Model

Nesso-1

Open protein-ligand affinity model trading coarse-grained structure for inference speed

Small moleculesAI-native biotechUnited States, CanadaActive中文

Overview

NameNesso-1 · aliases: Nesso-1
Full nameThe technical report gives no expanded name
Ambiguous names (not auto-linked)Nesso
FamilyNot a family entry. A single version with no family members
OrganizationValence Labs, Recursion Pharmaceuticals
Organization typeAI-native biotech
CountryUnited States, Canada
Model classStructure-to-affinity (coarse-grained co-folding affinity)
DomainSmall molecules (also: Structure & co-folding)
First release† partly verified · new weights
Basis: The "Publish v1.0.0" commit on the weights host; a technical report PDF appeared the same month with no exact date found
Note: Repository creation is inadmissible as a first release; the weights-publication commit of the same day is used instead, and the report's exact date is still missing source
Other milestones
  • First public evidence: † · Weights publication commit source
  • Preprint v1: · Preprint posted date (registry record; the 08-01 in the DOI string is not used as a date) source
  • Weights release: † · Weights host publication commit source
Latest update · paper revision
Basis: Preprint posted (registry posted date)
The preprint went up on 2026-08-03; later activity on the weights host is model-card and asset edits only (config 2026-07-28, figures 2026-08-06)
Note: The kind previously read "no update since release", which contradicted the description; corrected to preprint posted source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Weights release and preprint within 12 months
OpennessCode: Unclear
Weights: Open · Apache-2.0 per model card
Weights are public on a model host with an Apache-2.0 model card; no separate code repository was found
Benchmarks reported4 counted (modified protocol 3, retrospective 1)
Wet-lab validationNone recorded
Unverified fields† first_release, release_milestones.first_public_evidence, release_milestones.weights_release, training_data.b_private_licensed, open_status.code, other_evaluations[3]
As of2026-11-16

What it does

Reuses the public training data and optimisation framework of Boltz-2 but drops its atom-level diffusion module, trading coarse-grained poses for inference speed in the same spirit as TerraBind; it outputs protein-ligand binding affinity and is positioned as a prefilter for large virtual screens.

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. 4 of 4 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
05 PoseBustersDeveloper report, modified protocol—Self-reportedA missing candidate added and confirmed in the review · evidence
30 PLINDER / PINDER / Runs N' PosesDeveloper report, modified protocol—Self-reportedA missing candidate added and confirmed in the review · evidence
31 FoldBenchDeveloper report, modified protocol—Self-reportedAdded by R7/fc_zero as a missing candidate and confirmed in the review. · evidence
01 CASPRetrospective dataset use—Third-partyAdded as a missing candidate in round 7 and confirmed in review. · evidence

Other evaluations (outside the 44)

  • Public affinity benchmarks (off-site) (Public benchmark not among the 44) Self-reported: Correlation reported on a viral-protease compound set and several free-energy comparison sets, with baselines largely copied from other papers
  • Virtual-screening activity benchmark (off-site) (Public benchmark not among the 44) Self-reported: Average precision and enrichment reported, with baselines taken from the comparison model's paper
  • 25 internal biochemical assays (Own task set) Self-reported: Compound-weighted correlation comparison; the report concedes limited gains on some assays
  • Inference speed comparison (Comparison against baselines only) Self-reported: Throughput compared with the baseline on three internal targets of ten compounds each on a single accelerator; a higher multiple circulates in company remarks (second-hand)

Developers report: an open coarse-grained co-folding affinity model more than an order of magnitude faster than the baseline, matching or beating it on their chosen public affinity and structure benchmarks, while cautioning that these are largely in-distribution with limited zero-shot generalisation.

Independent evidence: No independent evidence found. No third-party evaluation found; all figures come from the technical report and most baselines are copied

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesReuses the public training data and optimisation procedure of Boltz-2 source
b. Private or licensed dataNot disclosed
c. Own wet-lab data used for trainingNoNo own wet-lab training data source
d. Wet lab used only for validationNoInternal biochemical assays are held-out evaluation on existing data rather than new wet-lab work; the scale in internal projects is not disclosed source

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: The developer is the AI research unit of a listed US AI drug-discovery company that merged with a UK AI drug-design company in 2024 and has had a new chief executive since 2026-01-01 (the previous one moved to chair, then to strategic adviser from 2026-04).
  • Funding: Not disclosed
  • Products and partnerships: The company co-released the predecessor co-folding affinity model with a US university (2025-06), whose team later spun out as a public benefit corporation (see the BoltzGen record); this model is a slimmed, faster version of that predecessor and explicitly draws on TerraBind.
  • Related models: Boltz-2, TerraBind, BoltzGen, AQAffinity

Reading notes

  1. The headline is speed rather than an accuracy breakthrough, which suits prefiltering in virtual screens.
  2. All three structural benchmark entries are self-reported on a modified protocol: filtering admittedly differs from the cited paper and the poses are coarse-grained.
  3. The CASP entry is post-hoc reuse of a released affinity set, not participation, and does not count toward reported benchmarks.
  4. The report concedes limited generalisation and notes that on one benchmark a naive molecular-weight baseline is almost as good, so results should not be extrapolated to new targets.

Caveats

  1. All figures are self-reported with baselines mostly copied from other papers and no third-party recomputation.
  2. The first-release basis partly rests on repository creation and the report's exact date is missing; partly verified.
  3. The speed multiple has two framings (technical report and company remarks), the latter second-hand only.
  4. The scale of internal assay use is not disclosed.

Sources

  1. Technical report PDF (preprint)
  2. Weights and model card (package registry)
  3. Preprint registry record (archive)
  4. Weights host commit history (repository)

Corrections

3 corrections
FieldWasNowBasisFound
latest_update_kindNo later sourced event foundPreprint online (2026-08-03)source2026-11-16
benchmarks_reportedZero pairings (no pairing among the original 30)Three structure benchmarks are modified-protocol self-reports and one is retrospective dataset use; the model leaves the zero-pairing listsource2026-11-16
release_milestones.paper_v1Preprint 2026-08-01 (DOI string)2026-08-03 (posted date from the registration agency)source2026-11-16

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