AI for Bio · Benchmark 43

Ginkgo GDPa1

Uniform-assay developability data on 246 clinical antibodies plus a one-off blind test on 80 private held-out antibodies.

bench-43Type A+BEvidence L4Antibodies◐ Low-frequency Board中文

Overview

NameGinkgo GDPa1 · aliases: GDPa1, Ginkgo GDPa1, AbDev competition
Full nameGDPa1 antibody developability dataset and the AbDev competition
Ambiguous names (not auto-linked)AbDev
FamilyNot a family entry. One dataset plus a one-off competition; no family members.
DomainAntibodies (also: Proteins)
InclusionSupplementary entry
TypeA B · Public dataset plus a private held-out blind test
Evidence levelL4 · Blind test with hidden truth, but the held-out antibodies are existing clinical antibodies, not prospectively synthesised
PublisherGinkgo Datapoints, Ginkgo Bioworks, Hugging Face (Platform company)
First release† partly verified
Basis: Official announcement of the public dataset and the competition (2025-09-08); the 2025-06-13 dataset-object creation date is a gated object and is not used as a first release
Note: Day precision still rests on the announcement page and stays flagged. source
Other milestones
  • First public evidence: † · Dataset object creation date (gated); a note only, never displayed as a release source
  • Journal online: · Results paper online in a journal (mAbs 18(1)) source
  • Official announcement: † · Competition and public dataset announcement source
Latest update† partly verified · data
Basis: Dataset card changelog and the platform's last-modified time
Dataset card updated to v1.4; before that: results paper 2026-02-22, competition closed 2025-11-18.
Note: The dataset is gated and the changelog unreadable, so the v1.4 label is carried from an existing record. source
The most recent sourced event, not a claim that nothing changed since.
Next expectedThe organizers have announced no date for a next round.
Status◐ Low-frequency · Low frequency: the competition closed (2025-09-08 to 2025-11-18); the public leaderboard still runs and the dataset is at v1.4.
Scale246 public clinical antibodies with 10 assays; 80 private held-out antibodies; 5 properties predicted; 113 teams.
  • Public antibodies: 246 (dataset card)
  • Assays: 10 (dataset card)
  • Private held-out antibodies: 80 (results paper abstract)
  • Properties predicted: 5 (competition 2025)
  • Teams: 113 (results paper abstract)
Wet-lab refreshFresh experiments each round · The 80 held-out antibodies were measured specifically for the competition, once only; not a standing round.
Own wet lab needed to evaluate a new modelNo, run it as a download
Models reported0 counted
Unverified fields† first_release, latest_update, release_milestones.first_public_evidence, scale.items[2], scale.items[4], ranking.entries[2], ranking.entries[3]
As of2026-11-16

What it measures

  • Task: Predict five developability properties from sequence, scored on 80 private held-out antibodies.
  • Truth: Experimental measurements from one laboratory under uniform assays: hydrophobicity, polyreactivity, self-association, thermostability and titer.
  • Primary metric: Spearman correlation per property
  • Secondary metrics: Per-property prize standings
  • Version traps: Distinguish GDPa1 cross-validation from the competition's private held-out protocol; the paper shows the two differ systematically.

Current ranking

Board

MetricSpearman correlation per property on the private held-out set
Eval setAbDev competition 2025, 80 private held-out antibodies, 5 properties
Snapshot
Boardhttps://huggingface.co/spaces/ginkgo-datapoints/abdev-leaderboard
Retrieved2026-11-16

Caveat. The competition closed 2025-11-18 and standings come from the results paper; the leaderboard Space errored on 2026-11-16 and was not loaded; no site model appears among the winners.

Shown in the overview table:

Anonymous
0.708
Spearman correlation per property on the private held-out set · AbDev competition 2025, 80 private held-out antibodies, 5 properties
Board ·

RankModelScore (verbatim)Metric · eval set / protocolSource type · dateSource and note
1Anonymous0.708Spearman correlation per property on the private held-out set
AbDev competition 2025, 80 private held-out antibodies, 5 properties
Board
source Not a site model; top-1 on hydrophobicity (HIC).
1microcrisprtm20.392Spearman correlation per property on the private held-out set
AbDev competition 2025, 80 private held-out antibodies, 5 properties
Board
source Not a site model; top-1 on thermostability (Tm2).
1PR_CHO_NOV0.356Spearman correlation per property on the private held-out set
AbDev competition 2025, 80 private held-out antibodies, 5 properties
Board
source Not a site model; top-1 on polyreactivity (abstract 0.356, Results 0.337; swapped with self-association between the two).
1AbDevelop0.337Spearman correlation per property on the private held-out set
AbDev competition 2025, 80 private held-out antibodies, 5 properties
Board
source Not a site model; top-1 on self-association (abstract 0.337, Results 0.356).
1meta20.310Spearman correlation per property on the private held-out set
AbDev competition 2025, 80 private held-out antibodies, 5 properties
Board
source Not a site model; top-1 on titer.

Models reported

Grouped by relation type; each line shows its source type. Counted relations add up to 0. A pairing whose evidence rests only on this survey's own page is shown as unverified.

None. No model among the 112 has a recorded pairing with this entry.

Caveats

  1. The key negative result of the blind test: cross-validation correlations are systematically higher than held-out correlations.
  2. Only 246 training and 80 test antibodies, all clinical-stage, so extrapolation to early-stage or designed antibodies is untested.
  3. A one-off competition: the held-out set is exposed once it closes, so later work can only compare cross-validation scores.
  4. Winning methods differ by property; no single architecture wins across the board.
  5. Not a prospective synthesis test; the held-out antibodies are existing clinical antibodies.

How to use it

Request GDPa1 from the data platform (terms must be accepted), use the recommended stratified cross-validation split, and report Spearman per property for all five. The private held-out set was consumed by the competition and cannot be submitted to again; a real blind test needs new antibodies measured under the same assays, or joining the related consortium. State whether a number is cross-validation or the competition held-out set.

Pairs well with: AbBiBench · FLAb / FLAb2 · AIntibody · Polaris / ASAP / OpenADMET

Sources

  1. Competition and dataset announcement (official site) ·
  2. Competition results paper (mAbs 18(1)) (journal) ·
  3. GDPa1 dataset card (gated) (official site) ·
  4. Public leaderboard (leaderboard)

Corrections

5 corrections
FieldWasNowBasisFound
first_release2025-06-13 (dataset on HF); competition announced 2025-09-082025-09-08source2026-11-16
first_release_basisHF blog/competition launch with public GDPa1 data (HF repo object created 2025-06-13, likely private/gated earlier)official announcement of the public GDPa1 dataset and AbDev competition, 2025-09-08. HF dataset object created 2025-06-13 (creation date, gated, not displayed as release).source2026-11-16
first_release2025-09-082025-06-13 (dataset on HF); competition announced 2025-09-08source2026-11-16
scale246 public antibodies, 10 assays; 80 private holdout; 113 teams (presented as fully verified)same figures, but 246/10 are verified on the dataset card while 80 and 113 come from the results paper abstractsource2026-11-16
ranking.entriespolyreactivity 0.337 and self-association 0.356polyreactivity 0.356 and self-association 0.337 per the abstract; the Results section swaps the twosource2026-11-16

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