IglooALM: Structural Conditioning for Antibody Loops

• The paper introduces IglooALM, which extends IglooLM by conditioning loop tokens with per-residue multimodal tokens to enhance antibody loop redesign.
• It integrates backbone dihedral embeddings with masked language modeling finetuned from IgBert, enabling controlled sampling of diverse loop sequences.
• Empirical results show that IglooALM achieves higher sequence diversity and structural consistency compared to inverse folding models, despite some limitations in affinity prediction tasks.

Searching arXiv for the specified paper to ground the article in the cited source. IglooALM is a transformer-based protein LLM for structurally conditioned antibody loop sequence generation, introduced within the Igloo framework in "Tokenizing Loops of Antibodies" (Fang et al., 10 Sep 2025). It extends IglooLM by conditioning not only on a loop-level token $t$ summarizing loop conformation, but also on per-residue multimodal tokens $x_i$ derived from loop sequence and backbone dihedral angles. The model is designed for sampling antibody loops whose sequences are diverse while remaining consistent with a target loop backbone. Within the broader framework, IglooALM operationalizes multimodal loop tokenization for generative design, using backbone angles from Ibex-predicted structures and masked language modeling finetuning on paired OAS antibody sequences.

1. Definition and position within the Igloo framework

IglooALM is the generative counterpart to IglooLM in the Igloo framework (Fang et al., 10 Sep 2025). The framework centers on antibody and TCR loop regions, specifically the complementarity-determining regions H1, H2, H3, H4 and L1, L2, L3, L4, where CDR4 denotes the DE loop adjoining CDR1 and CDR2. Igloo itself is a multimodal tokenizer operating at the loop level rather than the residue level, encoding both sequence and backbone dihedral angles. It learns a continuous loop token $t$, a quantized token $\hat{t}$ via a learned codebook, and multimodal residue embeddings $x_i$ for loop residues.

The motivation for IglooALM follows three limitations identified for canonical clustering of antibody loops: limited coverage, structure-only categorization, and lack of direct integration into modern protein LLMs. The paper states that 20.3% of all loops have no canonical assignment, rising to 76.3% for H3 loops. Igloo addresses this by assigning tokens to all loops while retaining the ability to recover canonical loop conformations. IglooALM then uses these tokens for conditional sequence generation, with the aim of redesigning loops in a manner consistent with target structural constraints.

A plausible implication is that IglooALM treats loop conformation as the primary modular unit for conditioning, rather than relying exclusively on residue-local structure or on sequence-only contextualization. This places it between sequence-only masked LMs and inverse folding systems: it does not co-generate structure, but it uses structural tokens as conditioning variables during sequence generation.

2. Architectural formulation and input representation

IglooALM is finetuned from IgBert, a BERT-style encoder with 420M parameters (Fang et al., 10 Sep 2025). As in IglooLM, the base architecture remains a masked LLM; the distinctive feature is the insertion of Igloo-derived structural tokens into the LM token stream. Whereas IglooLM inserts only the loop token $t$ at the start of each CDR and an <end> token at the loop end, IglooALM inserts both $t$ and the multimodal residue tokens $x_i$ for all residues in the loop. The paper states that IglooALM omits <end> and instead provides $x_i$ per residue inside the loop.

The Igloo tokenizer constructs per-residue multimodal embeddings from sequence and dihedral angles. For each residue, each dihedral $\theta \in \{\phi,\psi,\omega\}$ is mapped to unit-circle coordinates, yielding

$x_i$0

These are linearly projected as

$x_i$1

with $x_i$2 and $x_i$3, so that $x_i$4. Sequence is embedded as $x_i$5 via 20 learnable amino-acid embeddings, and multimodal residue embeddings are formed by

$x_i$6

A learnable classification token $x_i$7 is prepended to the loop sequence and processed by a Transformer with ESM-2-style layers to produce the loop embedding and residue embeddings. Quantized tokens $x_i$8 are obtained by vector quantization, with the best configuration using a codebook size $x_i$9. Before insertion into IgBert, the tokens are linearly projected to the LM hidden dimension. The paper does not introduce custom positional encodings beyond the standard learned position embeddings of the base model.

This architecture yields two conditioning scales simultaneously. The loop token $t$0 provides a compressed representation of global loop conformation, while the residue tokens $t$1 inject local angle-informed context at each loop position. The paper’s framing suggests that this finer-grained conditioning is the defining distinction between IglooALM and IglooLM.

3. Training objectives, data, and finetuning protocol

IglooALM is finetuned with the standard masked language modeling objective over amino acids (Fang et al., 10 Sep 2025). During finetuning, 15% of residues are masked, following the 80/10/10 replacement rule: 80% replaced by [MASK], 10% changed to a random amino acid, and 10% left unchanged. The MLM loss over masked positions $t$2 is given as

$t$3

No additional structural consistency or auxiliary losses are used for IglooALM; structural conditioning is supplied entirely through the inserted Igloo tokens.

The tokens themselves are produced by the pretrained Igloo tokenizer. Igloo pretraining uses multiple losses: masked reconstruction of dihedral angles, masked reconstruction of amino acids, BCE-based contrastive learning over loop backbones, vector-quantized codebook learning with commitment, and unit-circle regularization. For dihedral reconstruction, the residue-level loss is

$t$4

The amino-acid reconstruction loss is standard cross-entropy,

$t$5

Contrastive learning over loop embeddings uses BCE rather than InfoNCE:

$t$6

Codebook learning is defined as

$t$7

and the overall Igloo loss is

$t$8

IglooALM and IglooLM finetuning use OAS paired antibodies with Ibex-predicted structures. The paper reports 2,447,258 antibodies with heavy and light chains folded, clustering of heavy and light chains separately with MMseqs2 at 90% sequence identity, 875,767 heavy clusters and 247,156 light clusters, and validation/test sampling of 20,000 heavy-chain clusters and 10,000 light-chain clusters. The resulting total training set comprises 4,598,332 antibody chains. Finetuning is performed for 53k steps over 5 epochs, taking 3 days on 4× NVIDIA H100. Igloo pretraining is performed on 1× NVIDIA H100 for 100 epochs in two stages, with best hyperparameters $t$9, $\hat{t}$0, $\hat{t}$1, $\hat{t}$2, 4 transformer layers, commitment weight $\hat{t}$3, maximum loop length 36, and batch size 64.

4. Structural conditioning and generation procedure

IglooALM performs conditional generation of loop amino-acid sequences given loop backbone angles and masked loop sequence (Fang et al., 10 Sep 2025). The conditioning variables are the loop token $\hat{t}$4, summarizing loop conformation, and the per-residue multimodal tokens $\hat{t}$5, which incorporate the backbone dihedrals at each loop position. The model then predicts amino-acid logits for masked positions using the MLM head and samples sequences at temperatures

$\hat{t}$6

The paper is explicit that IglooALM does not co-generate structure. Structural consistency is instead expected to emerge from the angle-based conditioning provided by $\hat{t}$7 and the global conformation token $\hat{t}$8. After sampling, Ibex is used to fold the sampled sequence, and the resulting structure is compared with the original loop to assess self-consistency. This design differs from inverse folding baselines such as AbMPNN and AntiFold, which operate in the inverse folding regime.

The structural distance function that underlies Igloo training and retrieval is based on backbone dihedral angles rather than RMSD. For loops $\hat{t}$9 and $x_i$0, the dihedral-angle distance is defined as

$x_i$1

where $x_i$2 is the number of aligned residues and $x_i$3 denotes the alignment, with dynamic time warping used when lengths differ after stem-region Kabsch alignment. The paper states that dihedral-angle distance is prioritized over RMSD because it respects chirality and side-chain orientation constraints, while RMSD is reserved for self-consistency evaluation in the generative setting.

A plausible implication is that IglooALM’s conditioning is biased toward preserving backbone-compatible conformational classes rather than reconstructing a full atomistic structural state. This is consistent with the model’s loop-centric abstraction and with the use of post hoc folding for evaluation.

5. Empirical behavior in loop sampling

IglooALM is evaluated on loop sampling by masking loop sequences, conditioning on Igloo tokens, sampling candidate sequences, folding them with Ibex, and comparing the resulting loops to the original structures (Fang et al., 10 Sep 2025). The evaluation samples 50 structures per loop category from the Igloo test set in SAbDab for H1, H2, H3, L1, L2, and L3. For each loop, 10 sequences are sampled at each of the seven temperatures, giving 50 structures × 10 sequences × 6 loop types × 7 temperatures = 21,000 sequences per model.

The reported evaluation metrics are sequence diversity, measured by sequence identity or edit distance to the original loop, and structural consistency, measured by loop self-consistency RMSD (scRMSD) between the folded sampled loop and the original loop, aligned on loop residues. No TM-score or pLDDT is reported for sampling. Fig. 4a stratifies scRMSD by sequence identity bins, and the text summarizes the outcome as higher structural consistency at comparable or greater sequence diversity than baselines.

The comparison includes AbMPNN, AntiFold, the base IgBert model, and IglooLM. The paper states that IglooALM samples antibody loops that are “diverse in sequence and more consistent in structure” than state-of-the-art antibody inverse folding models. In the highlighted SARS-CoV-2-neutralizing antibody example with PDB 7TCQ, IglooALM samples H3 loops at $x_i$4 with average edit distance 6.6 for a length-9 loop and achieves average loop RMSD 0.79 Å. The abstract and Section 1 further report average sequence identity $x_i$5 and $x_i$6 Å RMSD for redesigned H3 loops of 7TCQ.

These results situate IglooALM as a structurally conditioned generator rather than a mere masked sequence denoiser. The paper’s comparison to inverse folding baselines suggests that multimodal loop tokenization can preserve target loop geometry even when producing sequences substantially different from the original. This suggests that the combination of $x_i$7 and $x_i$8 carries enough information to constrain sampling to a backbone-compatible manifold.

6. Relationship to Igloo retrieval, IglooLM, and baseline behavior

The performance of IglooALM is best interpreted within the broader empirical profile of the Igloo framework (Fang et al., 10 Sep 2025). Igloo’s loop-level latent space is trained so that loops with similar dihedral backbones are close together, and its tokens support both retrieval and downstream modeling. In paratope retrieval, using cosine similarity among same-type and same-length loops, Igloo achieves strong performance across loop types. For H3 under the criterion $x_i$9, Igloo attains Precision@20 of 0.402, outperforming Amino Aseed at 0.379 by 5.9% and ESM-2 (3B) at 0.237 by 69.8%. Under RMSD $t$0 Å, Igloo reaches 0.278 P@20 for H3. The framework also shows high loop-type purity $t$1 and loop-length purity $t$2, with strong recovery of canonical clusters, including heavy CDR1 0.894, heavy CDR2 0.900, heavy CDR3 0.754, and heavy CDR4 0.983.

IglooLM, which uses only the loop token $t$3, is evaluated on AbBiBench for binding-affinity prediction of heavy-chain variants using ridge regression on sequence-level embeddings with 10-fold nested CV and Spearman $t$4 as the metric. It outperforms the base IgBert model on 8 out of 10 antibody–antigen pairs and performs comparably to larger sequence and multimodal protein LLMs, including examples such as 4fqi_h3 with $t$5 for IglooLM versus 0.973 for ESM-2 (3B) and 0.970 for IgBert, and 3gbn_h1 with 0.948 for IglooLM versus 0.953 for ESM-2 (3B) and 0.947 for IgBert.

By contrast, the paper states that IglooALM underperforms IglooLM on AbBiBench. The authors interpret this as being consistent with the difficulty of accurately capturing subtle residue-level structural differences from predicted backbones in DMS-like data. This contrast is important: IglooALM’s added residue-level conditioning is beneficial for controllable generation, but not uniformly beneficial across tasks. A plausible implication is that the granularity of structural tokenization interacts strongly with task type. For loop redesign, the per-residue structural tokens are informative; for small-effect variant prediction, they may inject noise when based on imperfect structural predictions.

7. Limitations, ablations, and practical use

Several limitations and robustness results delimit the scope of IglooALM (Fang et al., 10 Sep 2025). First, the model relies on backbone angles from Ibex-predicted structures. If those predicted structures are inaccurate, especially for subtle sequence perturbations, the residue-level tokens $t$6 can degrade performance, as observed in the AbBiBench comparison with IglooLM. Second, although Igloo assigns tokens to all loops and recovers canonical clusters with high purity, extremely rare or outlier conformations, particularly in H3, remain difficult; the paper notes that H3 exhibits the widest spread in the latent UMAP manifold. Third, the sampling results are in silico, and further experimental validation is required to determine whether redesigned loops preserve binding. Finally, structure prediction with Ibex is a computational dependency for generating $t$7, which may be costly at library scale.

The ablation studies attribute much of Igloo’s performance to its contrastive structural pretraining. Removing the contrastive loss reduces retrieval P@20 substantially, including declines of 11.8% for L3 and 20.0% for H3 relative to the full model. Dihedral angles are more important than sequence alone for retrieval, particularly for H3, where angles-only yields an 85.2% improvement over sequence-only at P@20, while adding sequence on top of angles further improves H3 by 12.7%. Training with multi-length positive pairs using dynamic time warping and a tolerance of one residue yields slight overall degradation, suggesting that most clusters remain length-specific and that same-length positive pairing is more reliable.

The paper provides a practical workflow for using IglooALM. Loop regions are identified with ANARCI/AHo numbering, structure is predicted with Ibex to obtain $t$8, $t$9, and $t$0 for loop residues, the Igloo tokenizer computes $t$1, $t$2, $t$3, and then produces the loop token $t$4 and residue tokens $t$5. These are inserted into the IgBert token stream, with the loop amino acids masked for resampling. Sequences are sampled at the desired temperature, folded with Ibex, and filtered by structural criteria such as RMSD $t$6 Å or by dihedral-distance thresholds and sequence-diversity constraints. The released code is at https://github.com/prescient-design/igloo, and the base model is Exscientia/IgBert on HuggingFace.

In this framework, IglooALM occupies a specific methodological niche: a loop-conditioned masked LLM that leverages multimodal loop tokenization to generate antibody CDR sequences compatible with target backbones. Its principal contribution lies not in replacing inverse folding or sequence-only modeling universally, but in demonstrating that loop-level and residue-level structural tokens can be integrated into a protein LM to produce antibody loops that are both sequence-diverse and structurally consistent.