AI for Bio · Model

VERID

Semantic digital twin of a lab workstation that loops agents between virtual dry runs, safety checks and real wet-lab experiments

Text, LLMs & agentsBig-tech labUnited StatesActive中文

Overview

NameVERID · aliases: VERID
Full nameThe name is an acronym recorded as the official pages write it, not expanded
Ambiguous names (not auto-linked)None
FamilyNot a family entry. A single infrastructure whose two companion applications are described in one record
OrganizationGoogle Research
Organization typeBig-tech lab
CountryUnited States
Model classAgent infrastructure / lab digital twin
DomainText, LLMs & agents (also: Proteins)
First release
Basis: Official research blog post source
Other milestones
  • First public evidence: · Official research blog source
  • Preprint v1: † · Main preprint; the two companion papers are dated 2026-10-17 and 2026-10-23 source
  • Official announcement: · Official research blog source
Latest update† partly verified · no later event
Basis: Companion papers posted; no sourced substantive event since
The main paper and one companion paper appeared the same day as part of the same release rather than a new version; nothing states whether code or weights are released source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · First published within 12 months, with companion papers
OpennessCode: Unclear
Weights: Unclear
Built on an open-source lab robotics library; neither the blog nor the paper abstracts state whether its own code or weights are released
Benchmarks reported0 counted (unconfirmed 6)
Wet-lab validation4 recorded campaigns
Unverified fields† open_status.code, open_status.weights, training_data.a_public, release_milestones.paper_v1
As of2026-11-16

What it does

Maintains a persistent semantic digital twin of a lab workstation, tracking robot geometry, well volumes, reagent identity and consumable stock, and offering introspection, time-compressed simulation, experiment provenance and a safe execution framework; two companion applications connect agents on top of it to real wet-lab work (three rounds of aptamer design and test, and binder design that trims the target to an interface window to speed design up). It divides work with Co-Scientist, which produces hypotheses while this addresses reliability once a hypothesis meets physical experiments; the companion work's structure and diffusion steps use tools such as AlphaFold 3.

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. 0 of 6 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
14 PubMedQAUnconfirmed (not counted)—Unconfirmed sourceNo report found in official materials †disputed · no source link
18 LAB-Bench / LABBench2Unconfirmed (not counted)—Unconfirmed sourceNo report found in official materials †disputed · no source link
23 BioASQUnconfirmed (not counted)—Unconfirmed sourceNo report by it was found in the official material. †disputed · no source link
29 BixBenchUnconfirmed (not counted)—Unconfirmed sourceNo report found in official materials †disputed · no source link
41 BioML-benchUnconfirmed (not counted)—Unconfirmed sourceNo report found; the only evidence is a site page, so it is shown as unconfirmed. †disputed · no source link
42 BioMysteryBenchUnconfirmed (not counted)—Unconfirmed sourceNo report found; the only evidence is a site page, so it is shown as unconfirmed. †disputed · no source link

Other evaluations (outside the 44)

  • Digital-twin protocol cases (Case study) Self-reported: Three dye-and-water liquid-handling cases: catching protocol errors in advance, recovering mid-run from a fault, and diagnosing subtle physical failures afterwards
  • Wet-lab readout of the closed loop (Wet-lab validation) Self-reported: Aptamer loops on two known targets and a binder experiment on one target, all designed and read out by the authors

Developers report: Developers report that a semantic digital twin makes agent-run wet labs safer and that closed-loop design raises hit rates round on round.

Independent evidence: No independent evidence found. No third-party evaluation or reproduction was found; the experiments were designed and read out by the authors

Training data

Group in Figure 7: Own wet-lab training data.

ClassValueDetail
a. Public databasesNot disclosedTraining data of the underlying language model and design tools is undisclosed; public structure-database entries serve as design targets source
b. Private or licensed dataNot disclosed source
c. Own wet-lab data used for trainingYesExperimental feedback of about 1,000 candidates per round and some 4,400 constructs in total tunes design rules online in the aptamer loop; this is not pretraining source
d. Wet lab used only for validationYesThe binder side's yeast display and sorting is validation 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.

Aptamer design loop (first target) (blog / aptamer paper)
Targets 1 · Designs tested — · Hits — · Success round 1 0%, round 2 26.0%, round 3 42.1% at 98.5% specificity · Best affinity —

Denominator basis: Bead-based selection with bead-corrected hit rates reported per round; the number tested per round is not given
The zero first round is attributed to fixed primer overhangs disrupting folding in nearly half the sequences; the intermediate round was previously missing on the site and all three rounds must be shown
Executor: Designed and read out by the developer; no blind third party

Self-reported · · verified · source
Aptamer design loop (second target) (blog / aptamer paper)
Targets 1 · Designs tested — · Hits — · Success round 2 specificity 75.0% to 19.1% with bead-corrected hit rate 10.2% to 3.7%; round 3 specificity 72.4% with on-target hit rate 33.5% · Best affinity —

Denominator basis: The same loop's second target, reported as specificity paired with hit rate rather than three plain hit rates
The second round's drop in specificity came from one structural motif dominating the designs, and the third round recovered through experimental feedback
Executor: Designed and read out by the developer

Self-reported · · verified · source
New binder experiment (single target) (companion paper)
Targets 1 · Designs tested 1250 · Hits — · Success — · Best affinity single-digit nanomolar (top designs)

Denominator basis: New experiment: a 5,000-member library with 1,250 designs per arm, read out by yeast display and sorting
This is the newly run experiment and must not be merged with the computational analysis on existing data below
Executor: Run by the developer

Self-reported · † · verified · source
Computational verification on existing data (not wet lab) (companion paper)
Targets 15 · Designs tested 3666 · Hits — · Success computational hit rate 5.06% to 9.33% at equal compute · Best affinity —

Denominator basis: The 15 targets and 3,666 designs are a trimmed analysis subset of an existing benchmark, not new experimental material
It must be presented as computational verification on existing data, and these figures must never be read as new wet-lab scale
Executor: A computational analysis with no experimental executor

Self-reported · † · verified · source

Background

  • Organization: The developer is a large technology company's internal research division with labs in several regions; the location of this project's wet-lab work is not public.
  • Funding: Internal company budget; no separate funding.
  • Products and partnerships: The blog lists no outside partners; related work from the same company includes Co-Scientist and AlphaProteo.
  • Related models: Co-Scientist, AlphaProteo, BindCraft

Reading notes

  1. This is method and infrastructure rather than a model that can be ranked, so the absence of board figures is expected.
  2. The improvement in hit rate comes with a clear error-correction narrative, but the sample is two known targets in a molecule class with extensive prior literature.
  3. The mid-loop drop on the second target shows the agent can be led astray by its own design rules and recovers only through experimental feedback.
  4. Unlike competition-style wet-lab evaluations, these experiments are designed and read out by the authors with no blind third party.

Caveats

  1. The binder side must separate two things: a new experiment on one target, and a computational analysis of 15 targets and 3,666 designs on existing data.
  2. The aptamer loop's intermediate round was previously omitted and all three rounds must be shown.
  3. The second target's figures pair specificity with hit rate and must not be read as three plain hit rates.
  4. It reports no benchmark on this site; the six related entries are treated as unconfirmed.

Sources

  1. Official research blog (2026-10-16) (blog) · 2026-10-16
  2. Main preprint (preprint)
  3. Companion paper: aptamer loop (preprint)
  4. Companion paper: binder design (preprint)
  5. Companion paper full text (binder design) (preprint)

Corrections

2 corrections
FieldWasNowBasisFound
wetlab_validationThe binder entry merged new experiments with computational analysis of existing dataSplit into two items: a new experiment on one target (5,000-member library, 1,250 designs per arm) and computational validation on existing data (15 targets, 3,666 designs)source2026-11-16
wetlab_validationThe aptamer closed loop recorded only the first and last roundsAdded the middle round, and wrote the second target's numbers as specificity paired with hit ratesource2026-11-16

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