AI for Bio · Model

IgGM / IgGM-FM

Generative model for antibodies and nanobodies covering structure prediction, de novo design, maturation and humanization

AntibodiesBig-tech labChinaMaintained only中文

Overview

NameIgGM / IgGM-FM · aliases: IgGM, IgGM-FM
Full nameNo separate expanded name
Ambiguous names (not auto-linked)None
FamilyCovers the original IgGM and its foundation-model version IgGM-FM; the wet-lab results come from the IgGM-FM preprint
  • IgGM ·
  • IgGM-FM ·
OrganizationTencent AI for Life Sciences Lab
Organization typeBig-tech lab
CountryChina
Model classAntibody generative model
DomainAntibodies
First release · preprint
Basis: bioRxiv v1 preprint (repository creation date not used; earlier than the conference paper) source
Other milestones
  • Preprint v1: · bioRxiv v1 source
Latest update · new version
Basis: IgGM-FM preprint posted
The IgGM-FM foundation-model preprint was posted; later GitHub activity is small fixes (last 2026-05-28) source
The most recent sourced event, not a claim that nothing changed since.
StatusMaintained only · Only small fixes after 2025-09
OpennessCode: Open · MIT
Weights: Open · MIT
Code and weights on GitHub under the MIT licence
Benchmarks reported1 counted (board 1)
Wet-lab validation3 recorded campaigns
Unverified fields† wetlab_validation[1]
As of2026-11-16

What it does

One model covers complex structure prediction, de novo design, affinity maturation, inverse design and humanization for antibodies and nanobodies; the foundation-model version broadens the task range.

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
36 AIntibodyOfficial leaderboard or results—BoardThe team claims a top-three placement in the affinity-optimization task; no per-team standing is cited and official materials do not confirm it. †unverified · evidence

Other evaluations (outside the 44)

None recorded.

Developers report: Developers report one model handles antibody structure prediction, de novo design, maturation and humanization, with nM-level PD-L1 binders

Independent evidence: No independent evidence found. No independent evaluation found

Training data

Group in Figure 7: Public data + wet lab for validation only.

ClassValueDetail
a. Public databasesYesPublic antibody complexes released before 2022-12-31: 6,448 heavy-light and 1,907 single-chain complexes in 2,436 clusters
b. Private or licensed dataNo
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationYesAffinity measurements on several antigens used for validation

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.

PD-L1 de novo antibodies (bioRxiv v1)
Targets — · Designs tested 60 · Hits 7 · Success 7/60 binders · Best affinity KD 0.084–2.89 nM / KD 0.084–2.89 nM

Denominator basis: 10,000 candidates were generated per CDR-length combination (24 combinations), filtered by edit distance and structure/frequency, and 60 sequences were tested, with 7/60 binders
The 10,000 figure is per length combination, not the pool the 60 were drawn from; 6 of the 7 blocked ligand binding in ELISA; developer-run

Self-reported · No sourced date · verified · source
PD-L1 antibody inhibition (unverified) (bioRxiv v1)
Targets — · Designs tested — · Hits — · Success — · Best affinity —

Denominator basis: Unverified: the previously published lowest IC50 value cannot be found in the v1 full text, so no figure is given
The earlier published figure is withdrawn and treated as unverified

Self-reported · No sourced date · †unverified · source
Affinity maturation and humanization (bioRxiv v1)
Targets — · Designs tested 20 · Hits 5 · Success — · Best affinity KD from 52.02 nM to 9.753 nM (5.3x); humanized variants 0.139–0.486 nM

Denominator basis: 20 candidates selected by generation frequency, yielding 5 humanized antibodies
All figures verified; developer-run

Self-reported · No sourced date · verified · source

Background

  • Organization: An in-house lab in Shenzhen funded by the company, with university and Chinese Academy of Sciences collaborators
  • Funding: Not disclosed or not applicable
  • Products and partnerships: The sister system tFold shares authors and comparison targets
  • Related models: tFold, RFantibody, Origin-1

Reading notes

  1. IgGM's 7/60 and nM affinities must never be attributed to tFold; the two were previously mixed up on the site
  2. The AIntibody placement is only self-stated and not confirmed by organizer material

Caveats

  1. The old lowest-IC50 figure cannot be found in the full text and is withdrawn
  2. Generation counts are per length combination and must not be read as a screening pool

Sources

  1. bioRxiv preprint (preprint)
  2. bioRxiv preprint (preprint)
  3. GitHub repository TencentAI4S/IgGM (repository)
  4. Official page iclr.cc (company blog)
  5. bioRxiv preprint (preprint)

Corrections

2 corrections
FieldWasNowBasisFound
wetlab_validationThe 10,000 figure was the pool the 60 were drawn fromThe 10,000 figure is per CDR-length combination (24 combinations)source2026-11-16
wetlab_validationLowest IC50 5.29 nMThe value is not in the full text; treated as unverified with no figuresource2026-11-16

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