Igloo

The complementarity-determining regions (CDRs) of an antibody are the flexible loops that do most of the work of antigen recognition, and since Chothia and Lesk's 1987 analysis the standard way to organize them has been to sort them into "canonical clusters" of recurring backbone conformations. That scheme has two limits: 20.3% of loops in the structural antibody database SAbDab match no canonical cluster at all, rising to 76.3% for the hypervariable H3 loop, and the clusters encode backbone geometry alone — no sequence, and no representation a protein language model can consume.

Igloo (ImmunoGlobulin LOOp Tokenizer) is a multimodal tokenizer that replaces canonical clusters with learned tokens. Developed at Prescient Design, Genentech's machine learning group, with Harvard University, and released as an arXiv preprint in September 2025, it encodes a loop's backbone dihedral angles together with its amino acid sequence into a single continuous embedding plus a discrete codebook token. Where existing multimodal protein language models such as SaProt, ProstT5, and ESM-3 tokenize structure one residue at a time, Igloo tokenizes at the level of an entire loop — the substructure that antibody engineers actually reason about.

The release covers three artifacts: the tokenizer itself plus two derived antibody language models. IglooLM inserts the loop token at the start of each CDR and targets representation quality; IglooALM adds per-residue multimodal tokens for structure-conditioned generation.

#Key Features

• Loop-level multimodal tokens: Dihedral angles are mapped to unit-circle coordinates and summed with learned amino acid embeddings, so each token reflects both conformation and chemistry rather than backbone geometry alone.
• Contrastive objective on dihedral distance: Equal-length loops below a dihedral distance of 0.1 are positive pairs and those above 0.47 negatives, pulling similar conformations together in latent space without any cluster labels.
• Complete coverage with a discrete codebook: A vector-quantized codebook of 8,192 entries assigns a token to every loop, including the H3 loops that canonical clustering leaves unlabeled, while still recovering known canonical conformations.
• Drop-in tokens for protein language models: A linear projection maps Igloo tokens into the hidden dimension of an existing antibody language model, so no architectural surgery is needed to make an established model structure-aware.
• Structure-consistent loop sampling: IglooALM conditions generation on the target backbone, producing sequence-diverse CDRs that fold back to the intended conformation.

#Technical Details

The tokenizer is a compact four-layer BERT-style transformer built on the ESM-2 transformer layer implementation, with an embedding dimension of 128, a maximum loop length of 36 residues, and a learnable classification token that becomes the loop representation. Four objectives combine — masked reconstruction of dihedral angles and of amino acid identities, the contrastive loss, and codebook learning — over 100 epochs on a single NVIDIA H100. Training data comprises 807,815 loops: 108,167 experimentally resolved CDR1–CDR4 loops from 18,303 SAbDab and STCRDab structures (14,341 antibodies, 3,095 nanobodies, 867 T-cell receptors) and 699,648 loops from paired Observed Antibody Space sequences folded with Ibex, Prescient Design's immune-protein structure predictor. IglooLM and IglooALM are fine-tuned from IgBert, a 420M parameter paired-antibody model, over 4,598,332 antibody chains.

On paratope retrieval at rank 20, Igloo leads on H3 loops with 0.402 precision by dihedral distance, 5.9% above the best structure tokenizer and 69.8% above ESM-2 (3B). Its codebook partition reaches 0.983 loop-type purity and 0.965 loop-length purity against reference clusters. On AbBiBench binding affinity regression, IglooLM beats its IgBert base on 8 of 10 antibody–antigen targets and ranks first or second on 7, matching ESM-2 (3B) with roughly seven times fewer parameters. Redesigning the CDR H3 of a SARS-CoV-2 neutralizing antibody (PDB 7TCQ), IglooALM samples loops at 0.27 mean sequence identity that fold to 0.79 Å RMSD of the original, improving on the antibody inverse folding models AbMPNN and AntiFold.

#Applications

Igloo is aimed at antibody discovery teams working at library scale. Tokens support retrieval of conformationally matched loops from structural databases, selection of library members sharing a seed binder's loop conformation, and downsampling of large libraries to maximize codebook coverage. IglooLM embeddings feed affinity-prediction regressors during lead optimization, and IglooALM redesigns CDRs while holding the backbone fixed — a complement to conformation-conditioned designers such as ConformAb. The recommended workflow requires ANARCI AHo-aligned sequences and antibody structures, experimental or Ibex-predicted; a lower-fidelity sequence-only path covers cases where folding millions of sequences is impractical.

#Impact

Igloo reframes structure tokenization around biologically meaningful substructures rather than individual residues, connecting four decades of canonical CDR classification to modern protein foundation models. Training, inference, and analysis code is released under Apache-2.0, though trained checkpoints are not distributed with the repository and no model card or data card accompanies the release. The work remains a preprint awaiting peer review, and its evaluation is entirely computational: the authors note that wet-lab validation is still needed to confirm that IglooALM-redesigned antibodies retain antigen binding.