IgGM
Antibody + nanobody design and structure prediction in one diffusion model. Predict an antibody-antigen complex, redesign CDRs or framework, mature affinity, or recover sequence from structure, epitope-guided, from the antigen you upload.
What it is for
Pick IgGM to design or humanize an antibody / nanobody against your antigen, or to predict the antibody-antigen complex, all in one model. For VHH backbones use RFantibody; for paired scFv CDRs use ESMFold2 design; to validate a designed binder's fold use Boltz-2.
IgGM (Wang et al., ICLR 2025). A generative diffusion foundation model for antibody and nanobody engineering. One model covers antibody-antigen complex structure prediction, CDR design, framework redesign / humanization, affinity maturation, and inverse (sequence-from-structure) design, all epitope-guided against the antigen you upload.
When it fits:
- You have an antibody or nanobody sequence and want to redesign its CDRs (or framework) against a specific antigen and epitope.
- You want to humanize a framework or mature affinity from a wild-type reference.
- You want a fast antibody-antigen complex structure prediction before committing to a wet-lab campaign.
Inputs
You will need:
- Antigen structure as PDB or mmCIF (the antigen sequence is read from this file — you do not type it).
- Antibody heavy chain sequence (>H); light chain (>L) optional (omit it for a nanobody / VHH). Mark positions to design with X.
- Optional: an epitope — click antigen residues on the structure and IgGM guides design toward them.
Each run uses a preset that sets the scale and scope:
- Antibody-antigen complex prediction
- Fold the antibody-antigen complex from the full heavy (and light) chain sequences against your antigen. Add an optional epitope to guide docking. Fastest mode.
- CDR design (H3 or all CDRs)
- Mask CDR positions with X in the antibody FASTA and IgGM redesigns them against your antigen. Add an epitope for epitope-guided design.
- Framework redesign / humanization
- Mask framework positions with X; IgGM redesigns them. Use for humanization or framework engineering.
- Affinity maturation
- Improve binding by exploring variants at the positions you mask with X, against a wild-type reference (same length, no masks). It designs one variant per masked position per sample, so a few samples over a short loop already gives a rich set.
- Inverse design (sequence from structure)
- Recover the antibody sequence given the complex backbone.
Parameters you set on the form:
- Antibody FASTA
- Paste the heavy chain as
>Hand, for a conventional antibody, the light chain as>L. Mark residues to design withX. Omit>Lfor a nanobody / VHH. Do not include an antigen record — it comes from the uploaded PDB. - Antigen PDB
- Upload the target as .pdb / .cif. The antigen sequence is extracted from the chain you select, so the structure is the single source of truth.
- Antigen chain
- The chain ID in the uploaded PDB that IgGM should treat as the antigen (e.g.
A). - Epitope
- Optional. Click residues on the antigen structure; IgGM guides design toward them. Positions are handled correctly regardless of the PDB's residue numbering.
- Mode
- Complex prediction folds the complex; CDR design / framework redesign redesign masked (X) positions; affinity maturation generates improved variants from a wild-type reference; inverse design recovers sequence from the backbone.
Typical runtime:
- complex_prediction
- ~2 min
- cdr_design
- ~3 min
- fr_design
- ~3 min
- affinity_maturation
- scales with samples x masked positions
- inverse_design
- ~2 min
How to read the results
Per design: the predicted antibody-antigen complex PDB, the designed sequence, and an epitope-contact count (how many of your chosen epitope residues the designed antibody engages). Sequence statistics and amino-acid distribution plots are attached as artifacts.
Where a tool reports them, the scores mean:
- ipTM
- Predicted confidence in the binder to target interface. Higher is better. Aim above roughly 0.7 on a tractable target.
- pLDDT
- Per-residue confidence in the predicted fold. Higher means the model is more sure of that part of the structure.
- i_pAE and pAE
- Predicted alignment error, at the interface (i_pAE) or across the whole structure (pAE). Lower is better.
References
Wang et al., ICLR 2025