IgSeek

The standard recipe for structure-based protein design is to generate a backbone with a diffusion model and then run inverse folding to infer a sequence that folds into it. That recipe breaks down on antibody complementarity-determining regions. CDRs — and CDR-H3 above all — are hypervariable loops with little of the regular secondary structure that inverse-folding models rely on, and sequence inference on them frequently hallucinates residues that produce non-functional antibodies.

IgSeek, from MBZUAI with Tencent AI Lab, the Chinese University of Hong Kong, and Beijing Institute of Technology, replaces generation with retrieval. A multi-channel equivariant graph neural network (MEGNN) encodes CDR backbone geometry into a fixed-length vector, and every solved CDR loop in a natural antibody database is embedded once into a searchable index. To design a sequence for a new loop, IgSeek embeds the query, retrieves its nearest structural neighbors, aligns their sequences, and samples each position from the resulting residue distribution. The design is therefore assembled from structurally conserved motifs that nature has already validated, rather than sampled from a learned generative distribution.

The approach sits deliberately outside the ProteinMPNN lineage of inverse-folding models, and it complements rather than replaces backbone generators: the paper pairs IgSeek with backbones from RFdiffusion and dyMEAN to show that a better structure hypothesis yields a better retrieved sequence.

#Key Features

• Retrieval instead of generation: Sequences come from real solved CDRs with similar geometry, which keeps designs inside the space of loops that fold and function.
• Equivariant structural encoder: The MEGNN represents each residue by its four backbone atoms and updates node features and multi-channel coordinates across layers while preserving E(3) equivariance, so the embedding is pose-independent.
• One index, many queries: Because the database is embedded once, design reduces to a vector lookup, and no model retraining is needed for a new target.
• Optional geometric re-ranking: An IgSeek+Kabsch variant re-scores top retrievals by RMSD until ten hits under 1 Å are found, which raises accuracy across all six CDR types.
• Cross-receptor generalization: The same trained encoder is applied to T-cell receptor loops without any TCR-specific training.

#Technical Details

The CDR vector database is built from 24,479 solved CDR loops deposited in SAbDab before 1 January 2024. Evaluation uses 4,449 loops released between 3 January and 29 May 2024 as a held-out test set, 5,111 TCR loops from STCRDab, and 5,000 predicted CDR-H3 loops from OAS for timing. For a query, the ten nearest neighbors of matching length are retrieved, their sequences aligned, and a per-position residue distribution sampled. Measured by average amino acid recovery, IgSeek improves on ProteinMPNN, ESM-IF1, AbMPNN, and AntiFold by at least 2.9% on light-chain CDRs and is comparable to the best baselines on heavy-chain loops; IgSeek+Kabsch beats every baseline on all six CDR types, with its largest margin on the hypervariable CDR-H3. On STCRDab it leads by at least 30% on both chain A and chain B loops with no retraining. Retrieval is 2.6× faster than FoldSeek at equal accuracy, and end-to-end inference is at least 20× faster than the design baselines.

#Applications

IgSeek suits antibody and TCR engineering workflows that already produce candidate loop backbones — from a structure predictor, a diffusion generator, or a crystal structure — and need sequences that are likely to express and bind. Its speed makes it practical for high-throughput campaigns where thousands of backbone hypotheses must each be converted to sequence, and its retrieval outputs carry a natural provenance: every design can be traced to the solved structures it was assembled from, which is useful when triaging candidates for synthesis.

#Impact

IgSeek reframes CDR design as a database problem, and in doing so offers a concrete answer to inverse-folding hallucination on hypervariable loops: constrain the output to sequences that have been observed on similar geometry. The transfer to T-cell receptors without retraining is the strongest evidence that the learned structural embedding captures loop geometry rather than antibody-specific sequence statistics. The main limitations are inherent to retrieval — coverage is bounded by what SAbDab contains, so genuinely novel loop conformations have no close neighbor to draw from — and no code, weights, or prebuilt index have been publicly released. The preprint has not been peer reviewed.