IgDesign

What is IgDesign?

IgDesign is an inverse-folding model from Absci for designing antibody complementarity-determining region (CDR) sequences from an antibody-antigen complex structure. It conditions sequence generation on the antibody backbone, framework sequence, light chain, and antigen context.

IgDesign can redesign heavy-chain CDRs (HCDR1–3), light-chain CDRs (LCDR1–3), or a selected combination. ProteinIQ runs the reviewed IgDesign 1.0.0 source revision on a cloud GPU and preserves its native prediction table.

Input requirements

Submit exactly one PDB structure or RCSB PDB ID. The structure must contain:

• distinct heavy, light, and antigen chains;
• protein ATOM records for all three selected chain IDs; and
• matching SEQRES records for all three selected chains.

IgDesign’s parser compares the coordinate-derived sequence with SEQRES. Structures without matching records are rejected before the model runs.

Settings

Chain configuration

Each chain ID must be one alphanumeric PDB chain character. The defaults are H for heavy, L for light, and A for antigen. ProteinIQ retains its reviewed 5NGV-style heavy-CDR preset. The original no-argument IgDesign command uses the separate 1N8Z example preset, which selects all six CDRs with heavy B, light A, and antigen C.

CDR regions

Select one or more heavy- or light-chain CDRs. ProteinIQ uses IgDesign’s own PDB parser and ANARCI integration to annotate each antibody chain in the IMGT scheme, then passes the resulting sequence indices to IgDesign. Positions are therefore derived from the submitted structure rather than assumed from a fixed example.

The numbered CDR decoding order shows the order sent to IgDesign. Use Move up and Move down to change it. Newly selected CDRs are appended to the list; removing and selecting a CDR again moves it to the end. This order affects sequence generation.

Under Advanced CDR positions, you can provide an ordered JSON list for each selected CDR, such as [25,24]. Positions start at 0 in the source-cropped variable domain of the selected heavy or light chain. They are not PDB residue numbers or IMGT numbering labels. The list is passed to IgDesign in the specified order; IgDesign applies the light-chain offset itself. Leave a field blank or use [] to keep the automatic full IMGT interval. Indices outside the parsed variable domain are rejected.

Sampling

IgDesign’s native output count is:

decoding orders × ProteinMPNN samples × language-model samples

All three settings default to 10, producing 1,000 rows. Reducing any factor reduces both output count and compute. The ProteinMPNN and language-model temperatures independently control diversity at their corresponding sampling stages; both default to 0.5.

The random seed defaults to 0, matching the reviewed source configuration. Independent CDR loss calculation is enabled by default.

Output

The Results tab displays every column from IgDesign’s native CSV in source order. Selected CDR columns contain generated sequences. When independent loss is enabled, the source adds CDR-specific columns such as ce_loss_independent_hcdr1; each cell contains the mean cross entropy across that selected CDR for one candidate. ProteinIQ uses a native batch size of 1, so each CDR score is a single value, not a per-residue loss vector.

The Files tab includes:

• the native prediction CSV;
• IgDesign’s resolved configuration pickle;
• JSON provenance with source revisions, checkpoint identities, effective settings, and derived IMGT regions; and
• the prediction command log.

Cross-entropy loss is a model score for the generated sequence under the configured structural context. IgDesign does not define universal score bands or claim that a particular loss guarantees binding affinity, so compare candidates within the same run and validate selected designs experimentally.

Applications

• Redesign one or more CDR loops while preserving a known binding geometry.
• Explore heavy- and light-chain sequence alternatives around an antibody-antigen complex.
• Generate candidates for downstream structural analysis and experimental screening.

Limitations

• IgDesign requires an antibody-antigen complex structure; it does not design from an antigen sequence alone.
• The submitted backbone is fixed during sequence design.
• Source-native scores do not replace binding, specificity, developability, or expression experiments.

Reference

Shanehsazzadeh A, Alverio J, Kasun G, Levine S, et al. IgDesign: Integrated Generative Modeling for Antibody Design. bioRxiv (2023). doi:10.1101/2023.12.08.570889