IglooLM: Protein Language Model for Antibody Loops

• IglooLM is a novel antibody protein language model that tokenizes complementarity-determining regions using multimodal tokens combining sequence identity with backbone dihedral geometry.
• It finetunes a 420M BERT-style IgBert architecture by inserting specialized loop tokens, thereby improving tasks like binding affinity prediction and structural retrieval.
• The model achieves notable performance gains, including up to 20% improvement in loop retrieval accuracy and high purity scores, while overcoming limitations of canonical clustering.

IglooLM is a protein LLM for antibodies that incorporates loop-level multimodal tokens learned by the ImmunoGlobulin LOOp Tokenizer, Igloo. In the underlying formulation, antibody complementarity-determining regions are treated as tokenizable structural units rather than solely residue-wise sequences. Igloo encodes both amino-acid identity and backbone dihedral geometry, and IglooLM uses the resulting loop tokens as special inputs inside a finetuned BERT-style antibody LLM derived from IgBert. Within the reported framework, this design is used primarily for binding affinity prediction, while the broader Igloo family also supports structural retrieval and loop generation through the related IglooALM model (Fang et al., 10 Sep 2025).

1. Conceptual position and relation to antibody loop modeling

Igloo denotes a multimodal antibody loop tokenizer that encodes sequence and backbone structure, represented by dihedral angles. The central modeling choice is to represent each antibody loop as a higher-level object with a learned embedding and a discrete token assignment, rather than reducing loop description either to sequence alone or to membership in a pre-existing canonical cluster. IglooLM is the downstream protein LLM that uses these learned loop tokens inside an antibody sequence model; specifically, it is built by finetuning IgBert and inserting Igloo loop tokens as special tokens into the model input (Fang et al., 10 Sep 2025).

This places IglooLM at the intersection of antibody-specific protein language modeling and structure-aware tokenization. The paper also defines IglooALM, described as a “dihedral Angle LLM,” which extends the same basic idea by using both the loop token and multimodal residue-level tokens. In the reported experiments, IglooLM is the simpler variant, using only the loop-level token $\mathbf{t}$, whereas IglooALM augments it with residue-level multimodal tokens $\mathbf{x}_i$.

A key contextual point is the relationship to canonical clustering. Canonical clusters have historically been used to categorize CDR structural diversity, but the reported motivation for Igloo is that canonical approaches have limited coverage and are not readily incorporated into protein foundation models. Igloo is presented as retaining the ability to recover canonical loop conformations while assigning tokens to all loops, including those outside canonical taxonomies. This addresses a stated limitation of canonical clustering, for which 20.3\% of all loops and 76.3\% of H3 loops lack a known canonical cluster (Fang et al., 10 Sep 2025).

2. Loop representation and multimodal encoding

In the Igloo formulation, an antibody loop is modeled as a sequence of residues

$\mathbf{a}=(a_1,\dots,a_n)$

together with backbone dihedral angles

$\bm{\phi}, \bm{\psi}, \bm{\omega}\in(-\pi,\pi]^n.$

Each angle is converted to a unit-circle representation,

$$(\cos\bm{\phi}, \sin\bm{\phi}, \cos\bm{\psi}, \sin\bm{\psi}, \cos\bm{\omega}, \sin\bm{\omega}) \in [-1,1]^{n\times 6},$$

and then projected with a linear layer to obtain

$\mathbf{D}\in\mathbb{R}^{n\times d}.$

In parallel, each amino acid is mapped to a learned residue embedding

$\mathbf{A}\in\mathbb{R}^{n\times d}.$

The multimodal residue representation is the sum

$$\mathbf{X} = \mathbf{D}+\mathbf{A}\in\mathbb{R}^{n\times d}.$$

A learnable classification token $\mathbf{t}$ is prepended to represent the entire loop (Fang et al., 10 Sep 2025).

This representation is technically significant because it combines local sequence identity with explicit backbone geometry at the residue level, while simultaneously producing a loop-level summary token. The resulting loop representation is therefore neither a one-hot sequence encoding nor a direct canonical label; it is a learned embedding that reflects sequence-plus-structure correspondence. A plausible implication is that the loop token can act as a conformation-aware abstraction layer between structural immunoinformatics and sequence-based antibody language modeling.

The loops considered are the standard CDRs H1, H2, H3, H4, L1, L2, L3, L4. The training corpus for Igloo comprises loops from SAbDab, STCRDab, and Ibex-predicted structures from paired OAS sequences. The reported preprocessing yields 18,303 structures from SAbDab and STCRDab, consisting of 14,341 antibodies, 3,095 nanobodies, and 867 TCRs, which in turn produce 108,167 experimentally resolved loop structures. These are supplemented by 699,648 predicted loop structures from paired OAS / Ibex, for a total of 807,815 training loops (Fang et al., 10 Sep 2025).

3. Training objectives, contrastive geometry, and discrete tokenization

Igloo is trained self-supervised with four objectives: masked reconstruction of dihedral angles, masked reconstruction of amino-acid identities, contrastive learning of backbone similarity, and codebook learning. The total loss is reported as

$$\mathcal{L} = \ell_{\text{dihedral recon.}} + \ell_{\text{AA}} + \ell_{\text{contrastive}} + \ell_{\text{codebook}} + \lambda\,\ell_{\text{dihedral reg.}}.$$

For masked residues, dihedral reconstruction predicts $\mathbf{x}_i$0 using a 2-layer MLP and applies MSE to target unit-circle coordinates. Amino-acid reconstruction uses a 2-layer MLP trained with cross-entropy (Fang et al., 10 Sep 2025).

The contrastive objective is the methodological centerpiece. Similarity between loops $\mathbf{x}_i$1 and $\mathbf{x}_i$2 is defined using a dihedral angle distance $\mathbf{x}_i$3 from North et al.:

$\mathbf{x}_i$4

where $\mathbf{x}_i$5 aligns residues either one-to-one for equal-length loops or via DTW when lengths differ. Loop pairs are labeled positive if they have the same length and $\mathbf{x}_i$6, negative if they have different length or $\mathbf{x}_i$7, and are otherwise ignored. The contrastive loss is

$\mathbf{x}_i$8

with

$\mathbf{x}_i$9

so that loops with similar backbone geometry are pulled together in latent space and dissimilar ones are pushed apart.

Igloo additionally learns discrete tokens through a VQ-style codebook loss,

$\mathbf{a}=(a_1,\dots,a_n)$0

where $\mathbf{a}=(a_1,\dots,a_n)$1 is the stop-gradient operator. This makes the framework both a continuous embedding model and a discrete tokenizer for efficient retrieval.

The reported ablations support the structural role of these design choices. Removing the dihedral contrastive loss reduces retrieval performance, with improvements attributed to the contrastive objective of 11.8\% for L3 and 20.0\% for H3 at rank 20. The dihedral modality is described as the most important for retrieval, especially for H3, where the ablation discussion reports an 85.2\% improvement attributable to the dihedral track; sequence contributes as well, with 12.7\% improvement for H3 retrieval when included. Full multimodal Igloo performs best overall, although sequence-only Igloo can still recover many canonical clusters. Igloo is also reported to recover canonical clusters with 0.983 loop-type purity and 0.965 loop-length purity, with per-CDR purity values of 0.894 for H1, 0.900 for H2, 0.754 for H3, 0.983 for H4, 0.880 for L1, 0.975 for L2, 0.831 for L3, and 0.930 for L4 (Fang et al., 10 Sep 2025).

4. Incorporation into IgBert and the IglooLM / IglooALM family

IglooLM is constructed by finetuning IgBert, described as a 420M parameter BERT-style antibody LLM trained on paired and unpaired OAS sequences. The reported procedure takes the publicly available pretrained weights and hyperparameters from IgBert and continues finetuning with Igloo tokens. The base model is therefore not frozen; it is further optimized in the presence of the new loop-level tokenization scheme (Fang et al., 10 Sep 2025).

At the input level, the Igloo loop token $\mathbf{a}=(a_1,\dots,a_n)$2 is inserted as a special token at the start of each CDR loop, with an <end> token placed at the end of the loop. This makes loop boundaries explicit in the sequence presented to the LLM and provides a learned summary token for each loop. IglooLM uses only this loop-level token, whereas IglooALM uses both $\mathbf{a}=(a_1,\dots,a_n)$3 and multimodal residue-level tokens $\mathbf{a}=(a_1,\dots,a_n)$4.

For language-model finetuning, the reported structural preprocessing folds 2,447,258 antibodies with Ibex and trains on 4,598,332 antibody chains. Heavy and light chains are clustered separately at 90\% sequence identity; validation and test sets use cluster representatives, while training uses all sequences in training clusters. For retrieval and evaluation against known structures, the structural database is SAbDab using train/validation CDRs as the reference database and held-out test loops as queries.

The distinction between IglooLM and IglooALM is not merely architectural. The paper emphasizes that IglooLM is more effective when the task depends on loop conformation summaries and when residue-level structure is noisy or difficult to predict, as in subtle heavy-chain variant affinity prediction. The reported comparison therefore functions as a methodological caution against assuming that finer-grained structural tokenization is always preferable.

5. Retrieval, affinity prediction, and comparative evaluation

The principal downstream benchmark for IglooLM is binding affinity prediction for heavy-chain variants from AbBiBench. In the reported setup, mutated heavy-chain variants for each antibody-antigen pair are embedded, IglooLM embeddings are used as features, a ridge regressor is trained with 10-fold nested cross-validation, and performance is measured by the Spearman correlation coefficient $\mathbf{a}=(a_1,\dots,a_n)$5 between predicted and true binding affinity (Fang et al., 10 Sep 2025).

IglooLM is evaluated on 10 targets and is reported to outperform its base model IgBert on 8 out of 10 targets. It is also reported to be first or second on 7 out of 10 targets and to perform on par with existing state-of-the-art sequence-based and multimodal protein LLMs while being comparable to models with $\mathbf{a}=(a_1,\dots,a_n)$6 more parameters.

Target	Spearman $\mathbf{a}=(a_1,\dots,a_n)$7
1mlc	0.616
1n8z	0.675
2fxg	0.713
3gbn_h1	0.948
3gbn_h9	0.962
4fqi_h1	0.921
4fqi_h3	0.971
aayl49	0.625
aayl49_ML	0.531
aayl51	0.579

The paper highlights several direct comparisons to IgBert: 4fqi_h1: 0.921 vs 0.898, 4fqi_h3: 0.971 vs 0.970, 2fxg: 0.713 vs 0.694, and aayl51: 0.579 vs 0.566; for 1n8z, the value 0.675 vs 0.682 is noted as slightly lower. It is also stated that IglooLM is more than 7× smaller than ESM-2 (3B) yet performs better on average across the 10 antibody-antigen targets.

The broader evaluation includes paratope retrieval, using Precision at rank 20 with correctness defined by either $\mathbf{a}=(a_1,\dots,a_n)$8 or $\mathbf{a}=(a_1,\dots,a_n)$9 Å. Under the dihedral criterion for H3 retrieval, the reported values are Igloo: 0.402 and Amino Aseed: 0.379, corresponding to a +5.9\% relative improvement. The paper further states that Igloo exceeds ESM-2 (3B) by 69.8\% on H3 retrieval under the same dihedral criterion. Baselines named in the comparative analysis include sequence-only PLMs (ESM C, ESM-2 (3B), AbLang2, IgBert), multimodal PLMs (SaProt, ProstT5), and structure tokenizers or inverse-folding models (MIF, ProteinMPNN, Foldseek 3Di, Amino Aseed) (Fang et al., 10 Sep 2025).

6. Interpretation, limitations, and research significance

Several technical conclusions are explicit in the reported analyses. First, Igloo avoids the coverage problem of canonical clusters by assigning tokens to all loops while still recovering known structural taxonomy. Second, loop-level multimodal tokens appear especially useful for tasks in which conformation summaries matter more than residue-wise structural detail. Third, the empirical gap between IglooLM and IglooALM on heavy-chain variant affinity prediction indicates that added structural granularity can degrade performance when the structural source is imperfect (Fang et al., 10 Sep 2025).

The limitations are also clearly delineated. The generative results of IglooALM are reported to require more comprehensive experimental validation. At inference, the loop structures used by IglooLM and IglooALM come from Ibex predicted structures, so structure prediction errors can propagate into downstream representations. H3 remains the most difficult loop class, being both the most diverse and the least well covered by canonical clustering; the lower canonical-cluster purity for H3, 0.754, is consistent with that difficulty. In addition, the positive and negative contrastive thresholds, $\bm{\phi}, \bm{\psi}, \bm{\omega}\in(-\pi,\pi]^n.$0 and $\bm{\phi}, \bm{\psi}, \bm{\omega}\in(-\pi,\pi]^n.$1, are modeling choices, even though the margin region is ignored to mitigate overfitting to a hard threshold.

The practical implications reported for the framework include structural retrieval, antibody language-model improvement, affinity prediction, and loop generation/design, with relevance to CDR design, lead optimization, affinity maturation, and antibody library annotation. A plausible implication is that loop-level multimodal tokenization may provide a more stable abstraction than residue-level structural tokenization for antibody-specific foundation models, particularly when the available structural input is predicted rather than experimentally resolved. Within the scope of the reported results, IglooLM therefore represents a model family in which antibody loops function as explicit learned units of sequence-structure organization rather than as latent byproducts of residue-only language modeling.