MultiSAAI

Deciding whether a candidate antibody binds a given antigen is the rate-limiting step in early therapeutic discovery, and doing it experimentally is slow and expensive. Structure-based prediction is accurate when a complex is available but struggles with the conformational flexibility of the CDR loops that do most of the binding. Sequence-based prediction sidesteps the structure requirement, at the cost of throwing away the interface geometry that makes the problem tractable.

MultiSAAI is a sequence-based antibody-antigen interaction predictor that tries to put some of that geometry back without needing a solved structure. Developed by Guijun Zhang's group at Zhejiang University of Technology and posted as a preprint in June 2025, it takes three amino acid sequences — the antibody heavy chain variable region, the light chain variable region, and the antigen — and returns a binding probability.

Its distinguishing choice is to treat the heavy and light chains as playing different roles rather than concatenating them, and to combine four feature families at different scales: protein language model embeddings, physicochemical descriptors, geometric constraints, and residue substitutability. A multi-scale network then evaluates both global residue-pair compatibility across the whole sequences and local amino acid fitness at the predicted binding interface.

#Key Features

• Chain-aware sequence input: Heavy and light chain variable regions are modelled separately, reflecting their asymmetric contributions to antigen recognition, with an antibody-specific Roformer encoder for the antibody chains and ESM-2 for the antigen.
• Interface-focused features: CDR-H3 residues are identified and numbered with ANARCI, epitopes are predicted with an in-house tool, and the pair is encoded as a contact matrix capturing hydrogen bonds, ionic interactions, van der Waals contacts, and hydrophobic effects.
• Multi-scale interactive learner: A bilinear attention network combines global and local views so that pairwise compatibility and interface-level fitness inform a single score.
• Two trained variants: The generic model is trained on a curated SAbDab-derived dataset; MultiSAAI_SARS2 is trained on the same SARS-CoV-2 data used by A2binder for direct comparison.
• Hosted inference: A public web server accepts sequences directly, so screening does not require a local install or GPU.

#Technical Details

The generic training set was built from the July 2024 SAbDab release: 4,440 complete antibody-antigen complexes with paired heavy chain, light chain, and antigen sequences, reduced to 2,609 non-redundant pairs by CD-HIT at 98% identity. Antigens were clustered into 747 groups at 90% identity with MMseqs2, and two augmentation strategies — intra-group antibody pairing and CDR-H3-restricted antigen exchange — produced 3,790 positives, balanced 1:1 with negatives drawn from stringent inter-group pairing. All 7,580 samples were clustered phylogenetically with ClustalW into nine evolutionarily distinct groups, each split 4:1 into training and test to limit leakage. Language model layers are fine-tuned at a reduced learning rate of 2e-5 under binary cross-entropy loss.

MultiSAAI reaches an AUROC of 0.757 on the generic antibody-antigen benchmark and 0.946 on the SARS-CoV-2 dataset, outperforming A2binder and AbAgIntPre. Ablations trace the gains: adding physicochemical, geometric, and evolutionary features to the base ESM-2 and multi-scale CNN configuration helps, the bilinear attention network helps further, and swapping ESM-2 for the antibody Roformer on the antibody side yields a modest additional gain. Attention weights recover known interface residues in case studies, including antigen positions H133, K137, R223, and Y225.

#Applications

The intended use is high-throughput triage: given an antigen sequence for a newly emerged pathogen or a new target, rank a large antibody panel and prioritise a shortlist for experimental validation. The authors demonstrate this by scoring HER2 against every protein-antigen-targeting antibody in SAbDab. Because inference needs only sequences and is available through a web server, the model fits naturally into early discovery workflows where structures are not yet available.

#Impact

MultiSAAI is a useful example of how far explicit interface features can push a sequence-only interaction predictor, and its ablations make an unusually clear case for each component. Its caveats are real: an AUROC of 0.757 on the generic benchmark is modest and reflects how hard cross-antigen generalisation remains, the 0.946 SARS-CoV-2 figure comes from a much narrower distribution, and the epitope prediction tool the pipeline depends on had not been publicly released at the time of the preprint. The GitHub repository carries no licence file, which leaves reuse terms unresolved, and the work has not yet been peer reviewed.