IgGM

Therapeutic antibody engineering is normally split across a chain of specialized tools: one predicts the antibody–antigen complex, another performs inverse folding on a fixed backbone, a third scores point mutations for affinity maturation, a fourth handles humanization. Most require an experimentally determined complex or an existing functional antibody as a starting point, which rules them out precisely when a designer has only an antigen and a target epitope in hand.

IgGM, from the Tencent AI for Life Sciences Lab and academic collaborators, collapses that pipeline into one conditional generative model. It jointly learns antigen–antibody binding interactions and the mapping between antibody sequence and structure, so changing the conditioning inputs — antigen structure, epitope residues, which sequence positions are masked — switches the model between complex structure prediction, inverse design, affinity maturation, framework optimization, humanization, and de novo design. It handles conventional paired-chain antibodies and single-domain nanobodies alike, and lets users specify CDR loop lengths directly.

The model first appeared at ICLR 2025 as a sequence–structure co-design method; the September 2025 preprint extends it into a foundation model for antibody design with wet-lab validation against PD-L1, Protein A, TNF-α, IL-33, and SARS-CoV-2 RBD variants. It is the predecessor of IgGM2, which adds all-atom modeling and T-cell receptors.

#Key Features

• One model, six design tasks: Complex structure prediction, inverse design, affinity maturation, framework engineering, humanization, and de novo design are all reached by conditioning the same generative process on different inputs.
• Epitope-conditioned de novo design: Given only an antigen structure and a chosen epitope, IgGM generates a complete variable domain — all six CDRs, or three for a nanobody — plus its docked structure, with no template complex required.
• Frequency-based candidate screening: Sequences are ranked by how often identical designs recur across independent sampling runs, using reproducibility as a proxy for likelihood under the learned distribution instead of a separate scorer.
• Framework-region control: A targeted fine-tune on roughly 2,000 SAbDab complexes where framework residues contact the antigen enables humanization and framework redesign while holding CDR conformation fixed.

#Technical Details

IgGM is a hybrid diffusion model. Sequences are generated by discrete diffusion over the 20 amino-acid categories, while structure is generated by continuous diffusion over Cα coordinates combined with SO(3) diffusion over residue orientation frames. Sequence features come from a frozen pre-trained sequence model — an extension of ESM-2 adapted to multi-chain complexes — refined by a denoising network of 16 Sgformer blocks feeding 8 prediction modules built on SE(3)-equivariant invariant point attention. Training used SAbDab structures released through 31 December 2022: 6,448 paired heavy/light antibody–antigen complexes and 1,907 single-chain complexes, CD-HIT clustered at 95% identity into 2,436 clusters, with a time-split 60-complex test set (SAb23H2) from the second half of 2023.

On SAb23H2 complex structure prediction, IgGM reaches 2.14 Å antibody RMSD and 0.89 lDDT versus 2.25 Å and 0.88 for dyMEAN, with a DockQ success rate of 0.467 against dyMEAN's 0.067; initializing from AlphaFold 3 structures raises that success rate to 0.667. For CDR H3 co-design it recovers 36% of native residues against 13.6% for the previous best, and on the IgDesign inverse-design benchmark it exceeds ProteinMPNN, IgMPNN, and IgDesign on CDRs H1, H3, and L2. In a de novo campaign against PD-L1, 10,000 candidates across 24 CDR-length combinations were filtered to 60 for synthesis, yielding seven binders with KD from 0.084 nM to 2.89 nM, six of which blocked PD-1. Two rounds of affinity maturation improved an anti-IL-33 antibody from 52.0 nM to 9.75 nM, and humanized variants of a murine anti-RBD antibody retained sub-nanomolar affinity while raising humanness from 0.657 to 0.822.

#Applications

IgGM targets antibody discovery groups that want one system spanning the campaign. Teams can model an antigen–antibody complex, mature a lead by iterative single-position redesign, humanize a murine antibody while preserving CDR geometry, engineer framework regions for Protein A purification, or generate binders de novo against a nominated epitope. Nanobody support and explicit CDR-length control make it usable for single-domain formats, and frequency-based ranking gives a concrete rule for choosing which of thousands of designs to express.

#Impact

By showing that one conditioned generative model can serve the whole antibody engineering workflow — and by backing each task with measured binding data rather than recovery statistics alone — IgGM helped push antibody design away from task-specific tools. Code is released under the MIT license and weights via Zenodo under CC-BY 4.0, and the model placed in the top three of the AIntibody in-silico design challenge. Limitations remain: the antigen and epitope are rigid inputs, so binding-induced conformational change is not modeled; only backbone atoms are generated, leaving side-chain packing to downstream relaxation; and complex antigens such as peptide–MHC are outside its scope. IgGM2 was built to address those gaps. The foundation-model preprint has not yet been peer reviewed.