IgGM2

Designing immune receptors that bind a chosen target requires reasoning jointly about three coupled quantities: the amino-acid sequence of the variable loops, the full-atom conformation those loops adopt, and the binding geometry between receptor and target. Conventional pipelines break this into separate stages — structure prediction, inverse folding, and sidechain packing — which lets errors compound and prevents the backbone from adapting to newly proposed sequences. IgGM2 replaces that modular chain with a single all-atom generative model that predicts receptor structure and co-designs the complementarity-determining regions (CDRs) in one pass.

IgGM2 is a unified diffusion framework covering antibodies, nanobodies, and T-cell receptors (TCRs). Conditioned on a fixed target — an antigen for antibodies and nanobodies, or a peptide–MHC (pMHC) complex for TCRs — it jointly generates CDR residue identities and complete atomic receptor structures, so framework geometry relaxes around the designed loops without a separate inverse-folding or sidechain- packing step. A single pretrained checkpoint handles complex-structure prediction, receptor monomer prediction, and framework-conditioned CDR co-design across all three receptor classes without per-target retraining.

The model was developed by the Tencent AI for Life Sciences Lab together with collaborators at Shanghai Jiao Tong University, Zhejiang University, and Zhongguancun Academy. It extends IgGM (ICLR 2025), the same group's antibody and nanobody generative model, into an all-atom system that adds TCR–pMHC modeling and full-atom co-design.

#Key Features

• All-atom co-design: Generates CDR sequences and complete atomic structures simultaneously, letting the receptor framework adapt to designed loops rather than packing them onto a fixed backbone.
• Structure-to-design strategy: First learns how receptors position themselves around fixed target structures, then transfers that target-conditioned structural prior to CDR design.
• Unified receptor scope: One model spans antibody–antigen, nanobody–antigen, and TCR–pMHC complexes, covering both the humoral and cellular arms of adaptive immunity.
• Target-conditioned without retraining: A single checkpoint supports complex and monomer structure prediction and framework-conditioned CDR co-design across tasks.

#Technical Details

IgGM2 performs receptor-atom diffusion conditioned on a fixed structural context, denoising the receptor's atoms toward a target-consistent conformation while emitting CDR residue identities. Evaluated as a binder-only structure predictor, the IgGM2-P variant reaches a 73.8% success rate on the FoldBench antibody/nanobody–antigen benchmark using a single 1×1 sample without candidate ranking, and it captures receptor–target spatial relationships more accurately than AlphaFold3. On the STCRDab-derived TCR–pMHC benchmark it delivers strong TCR–pMHC modeling. For sequence design, IgGM2 achieves competitive amino-acid recovery and improves Rosetta-based interface preference metrics, indicating more favorable generated binding interfaces.

#Applications

IgGM2 is aimed at researchers designing therapeutic and diagnostic binders. Antibody and nanobody engineers can generate epitope-focused CDRs against a fixed antigen and obtain the predicted complex structure in a single step, while immunotherapy groups can model and design TCRs against peptide–MHC targets — a regime that most antibody tools do not address. Because the same checkpoint serves structure prediction and co-design, it fits into affinity maturation, humanization, and de novo binder campaigns without swapping models between tasks.

#Impact

By folding structure prediction, inverse folding, and sidechain packing into one all-atom generative process, and by extending coverage from antibodies and nanobodies to TCR–pMHC complexes, IgGM2 broadens the scope of immune-receptor design models built on the openly released IgGM lineage, whose code and weights are distributed under the MIT license. As a preprint it awaits peer review, and dedicated IgGM2 code and weights have not yet been released, so reported results are in-silico benchmarks that remain to be confirmed by experimental characterization.