IgGM

What is IgGM?

IgGM (Immunoglobulin Generative Model) is a generative foundation model for designing antibodies and nanobodies against target antigens. Developed by Tencent AI4S and published at ICLR 2025, it can generate sequence proposals and, for some tasks, structural outputs as well.

The model excels at designing complementarity-determining regions (CDRs), the hypervariable loops responsible for antigen recognition. On benchmark tests, IgGM achieved a 36% sequence recovery rate for CDR-H3 (the most flexible and challenging region), representing a 22% improvement over prior state-of-the-art methods.

IgGM supports both conventional antibodies (heavy + light chain) and single-domain nanobodies. It won a top-three prize in the AIntibody competition, an experimentally validated antibody design challenge.

How does IgGM work?

IgGM combines three components:

1. Pre-trained language model (ESM-PPI): Extracts evolutionary and structural features from protein sequences. ESM-PPI extends ESM2 with improved handling of inter-chain relationships in multi-chain complexes.
2. Feature learning module: Processes 3D coordinate information through structure encoding and inter-chain interaction modules, integrating spatial relationships between antibody and antigen.
3. Diffusion-based prediction module: Uses SO(3) diffusion to model rotational degrees of freedom in protein structure. The system samples initial translations from a Gaussian distribution and rotations from the standard SO(3) group, then iteratively refines both sequence and structure.

Training proceeds in two phases. First, a diffusion model learns structure prediction while preserving original sequence information. Then a consistency model is distilled from the diffusion model, enabling fast generation from arbitrary noise levels. This two-phase approach proved essential for model performance.

How to use IgGM online

ProteinIQ provides GPU-accelerated IgGM without local installation or the complexity of managing PyTorch Geometric dependencies.

Inputs

Input	Description
Heavy Chain (VH)	Antibody heavy chain variable region sequence. Mark positions to design with X characters.
Light Chain (VL)	Light chain variable region (optional for nanobodies).
Antigen Structure	Required. PDB containing the target antigen chain. Upload directly or fetch by PDB ID.
Antigen Sequence	Optional one-record FASTA. For multi-chain PDBs, its identifier must be the antigen chain ID.
Original Sequence	Required for affinity maturation. Supply one FASTA record for a nanobody or matching H/L records for a two-chain antibody.

The X character indicates positions for the model to redesign. For CDR design, mask the CDR residues while providing the framework sequence.

Design tasks

Task	Description
CDR Design	Design CDR loops to bind the target epitope. Provide framework sequence with X at CDR positions. For a related CDR-focused model, compare with Proteo-R1.
Inverse Design	Generate a sequence compatible with an existing antibody structure.
Framework Design	Redesign framework regions while preserving CDRs. Useful for humanization.
Affinity Maturation	Optimize an existing antibody for improved binding. Provide original sequence for comparison.

Settings

Setting	Description
Number of designs	Samples to generate (1–100, default 1). More samples provide diversity but increase runtime.
Epitope residues	Ordered, 1-based positions in the selected antigen sequence (e.g., 45,46,47,50-55). If omitted, IgGM uses its native automatic behavior.
Apply PyRosetta relaxation	Source-supported energy minimization for CDR and framework design PDB outputs. Increases runtime.
Sampling steps	Number of IgGM sampling steps (default 10).
Chunk size	Chunk size for long-chain inference (default 64).
Calculate epitope only	Return IgGM’s native interface positions without generating designs. Requires an antibody-antigen complex PDB with matching H/L chains.
Max antigen size	Maximum antigen length retained before IgGM’s native cropping step (default 2000).

Outputs

IgGM can produce:

• Designed sequences in FASTA format for each sample
• Predicted structures as PDB files when the selected task/run emits structure outputs
• Calculated epitope positions as CSV and text for epitope-only runs
• Run provenance and logs, including the source revision, model checksums, effective settings, command arguments, stdout, and stderr

Sequence-only and epitope-only runs are shown in task-appropriate result tabs instead of the 3D viewer. All emitted files remain downloadable as a ZIP archive.

Epitope specification

The epitope—the antigen region the antibody should bind—can guide design:

1. Provide comma-separated residue numbers (e.g., 417,449,453,455-456). Numbers correspond to 1-based positions in the selected antigen sequence, not PDB residue IDs.
2. Leave the setting empty to let IgGM run its own automatic cal_ppi behavior. If that source calculation cannot establish an interface, IgGM falls back to model-determined placement.
3. Enable Calculate epitope only to return IgGM’s native interface positions for an antibody-antigen complex without generating designs.

ProteinIQ does not replace IgGM’s automatic behavior with a separate distance, surface-area, or residue-count heuristic.

Limitations

• Antigen structure required: Unlike some tools that accept sequence-only input, IgGM needs a 3D antigen structure for epitope-guided design.
• Memory constraints: Large antigens may need truncation (the max_antigen_size parameter, 2000 residues by default in ProteinIQ unless overridden).
• Binding affinity not predicted: IgGM generates plausible sequences but does not estimate binding strength. Experimental validation remains necessary.
• Rigid docking: The antigen structure is treated as fixed during design.
• No deterministic seed control: The released design.py does not include a seed option.

Reproducibility

ProteinIQ runs IgGM source commit 06abc563b3fc8c7ea020543add16b69b6f8a1c8d and verifies the published Zenodo checkpoint files by MD5 during image creation. Each completed job includes an iggm_provenance.json file and separate command, stdout, and stderr logs.