AI for Bio · Benchmark 36

AIntibody

Antibody-side prospective blind test: submitted designs get uniform affinity and developability measurement.

bench-36Type AEvidence L5Antibodies◐ Low-frequency Board中文

Overview

NameAIntibody · aliases: AIntibody, AIntibody challenge
Full nameThe organizers use only the name AIntibody; no expanded form exists.
Ambiguous names (not auto-linked)None
FamilyNot a family entry. Only one completed round exists, so there are no family members.
DomainAntibodies (also: Proteins)
InclusionSupplementary entry
TypeA · Prospective blind challenge scored per task.
Evidence levelL5 · Blind test with prospective wet-lab work under uniform protocols; no independent replication.
PublisherSpecifica (IQVIA) (Platform company)
First release
Basis: Challenge design paper online (Nature Biotechnology, 10.1038/s41587-024-02469-9) source
Other milestones
  • Journal online: · Results paper online (Nature Biotechnology) source
  • Official announcement: · Challenge design paper online source
Latest update · data
Basis: Results paper online (Nature Biotechnology)
Round-1 results paper online: 29 organizations submitted 511 sequences, measured for affinity and developability; vendor releases followed on 2026-08-27. source
The most recent sourced event, not a claim that nothing changed since.
Next expectedSecondary reports say a society will run the next round, but the organizers have announced no date or target.
Status◐ Low-frequency · First round finished; the next round's host appears only in secondary reports, with no date or target.
Scale29 organizations submitted 511 sequences across 3 tasks; fewer than 511 unique sequences were actually tested.
  • Organizations: 29 (round 1)
  • Submitted sequences: 511 (round 1)
  • Unique sequences tested: <511 (round 1)
  • Tasks: 3 (round 1)
  • Developability assays: 5 (round 1)
Wet-lab refreshFresh experiments each round · New measurements each round: affinity by both HT-SPR and KinExA, and a five-assay developability panel run by independent laboratories under uniform protocols.
Own wet lab needed to evaluate a new modelNo, run it as a download
Models reported1 counted (board 1, unconfirmed 1)
Unverified fields† ranking.entries[1], ranking.caveat, next_expected, models_reported[0]
As of2026-11-16

What it measures

  • Task: Three tasks: computational affinity maturation from sequencing output, ranking affinity within HCDR3 clusters, and de novo CDR design against a protein absent from the selection output.
  • Truth: Newly measured experimental results under uniform protocols: affinity plus five developability assays, against SARS-CoV-2 spike RBD.
  • Primary metric: Measured equilibrium dissociation constant KD.
  • Secondary metrics: Whether the five developability thresholds are passed.
  • Version traps: Placements cannot be pooled across the three tasks; headline numbers need the task and assay method, and participation counts differ between vendor releases and the paper.

Current ranking

Board

MetricMeasured KD (HT-SPR and KinExA) plus five developability screens; placement per task
Eval setAIntibody round 1, 3 challenges, 511 submitted sequences from 29 organizations, SARS-CoV-2 spike RBD
Snapshot
Boardhttps://doi.org/10.1038/s41587-026-03238-6
Retrieved2026-11-16

Caveat. The results paper is paywalled, so figures come from vendor releases and the site's earlier note; placements sit with organizations or internal models, vendor claims conflict, and task 3 values must not be mixed with task 1.

Shown in the overview table:

AuraIDE / AuraBind (Aureka Biotechnologies, internal model)
KD 94.7 pM (KinExA); best experimental antibody 113 pM
Measured KD (HT-SPR and KinExA) plus five developability screens; placement per task · AIntibody round 1, 3 challenges, 511 designs from 29 organizations, SARS-CoV-2 spike RBD; all designs made and measured in one lab
Board ·

RankModelScore (verbatim)Metric · eval set / protocolSource type · dateSource and note
1AuraIDE / AuraBind (Aureka Biotechnologies, internal model)KD 94.7 pM (KinExA); best experimental antibody 113 pMMeasured KD (HT-SPR and KinExA) plus five developability screens; placement per task
AIntibody round 1, 3 challenges, 511 designs from 29 organizations, SARS-CoV-2 spike RBD; all designs made and measured in one lab
Board
source Not a site model; first in task 1, with second and fifth also reported by the same organization; any link to open weights is a company statement only.
3IgGM (Tencent AI4S)No scoreMeasured KD (HT-SPR and KinExA) plus five developability screens; placement per task
AIntibody round 1, 3 challenges, 511 designs from 29 organizations, SARS-CoV-2 spike RBD; all designs made and measured in one lab
Self-reported
†
source Developer claims a top-three placement in a 2025-08 repository entry; no value, no task named, not confirmed in official materials.

Models reported

Grouped by relation type; each line shows its source type. Counted relations add up to 1. A pairing whose evidence rests only on this survey's own page is shown as unverified.

Official leaderboard or results (1)

  • IgGM / IgGM-FM Board · The team claims a top-three placement in the affinity-optimization task; no per-team standing is cited and official materials do not confirm it. low confidence †unverified · evidence: link

Unconfirmed (1, not counted)

  • OpenDDE Unconfirmed source Challenge 1 (computational affinity maturation) · The placement belongs to that company's internal model and cannot be credited to the open weights; the site matrix and model page disagree. Resolution: Ruled unconfirmed: the public weights cannot be matched to the competition entry. †disputed · evidence: link

Caveats

  1. The paper's conclusion is conservative: results do not transfer across tasks and most designs did not beat conventional screening.
  2. Except for one model, predicting high-affinity clones within HCDR3 clusters did worse than picking at random.
  3. Participation and submission counts differ between sources, and much of the coverage is vendor press releases.
  4. A single 2.9 pM winning claim appears only on social media, with no public confirmation that it passed developability screening.
  5. Placements sit with organizations or internal models, open weights have no attributable placement, and there is no independent replication.

How to use it

There is no open submission: either wait for the next round or reproduce the same pipeline. Screen internally on the AbBiBench and FLAb retrospective sets first, then keep a prospective arm. Report the three tasks separately without pooling placements, and give the assay method, KD and all five developability results with the number of sequences tested.

Pairs well with: Ginkgo GDPa1 · AbBiBench · FLAb / FLAb2 · FoldBench · Adaptyv competitions / BenchBB

Sources

  1. AIntibody results paper (Nature Biotechnology 2026) (journal) ·
  2. AIntibody challenge design paper (Nature Biotechnology 2024) (journal) ·
  3. Press coverage: results and next-round arrangements (press)
  4. Press coverage: vendor placement claims and the next round's host (press)
  5. Company press release (syndicated) (press)

Corrections

4 corrections
FieldWasNowBasisFound
scale511 designs from 29 organizations; same-lab KinExA plus five developability assays; 14 days / max 10 sequences per task511 submitted sequences from 29 organizations; affinity by HT-SPR and KinExA; five-assay developability panel run by independent laboratories under uniform protocols; fewer than 511 unique sequences were actually testedsource2026-11-16
scale14 days and at most 10 sequences per team per taskremovedsource2026-11-16
wetlab_notefresh KinExA affinity and five developability assays in one labHT-SPR and KinExA affinity, five-assay developability panel run by independent laboratories under uniform protocolssource2026-11-16
models_reportedleaderboard entry recorded for the open-weight modeldisputed-or-unconfirmed; the placement belongs to a company-internal modelsource2026-11-16

Part of the AI for Bio benchmarks survey · all 44 benchmarks · open in the explorer. Record as of 2026-11-16.