Antigen-specific Antibody Multi-modal Foundation Model for Functional Antibody Design

The paper introduces AAMFM, an Antigen-specific Antibody Multimodal Foundation Model for functional antibody design. It addresses limitations in existing protein language models (PLMs) and specialized antibody design models by integrating antibody sequence, structure, and antigen information into a unified representation.

The model is built upon ESM3, a multimodal protein language model, and is adapted to the antibody domain through multi-stage training. First, it undergoes domain adaptation pre-training on approximately 1.4 million paired antibody sequence-structure data from the Observed Antibody Space (OAS) database. Second, it is fine-tuned on high-resolution experimental antibody-antigen complex data from the Structural Antibody Database (SAbDab). A key innovation is the Antigen Geometric-Epitope-aware Adapter, a lightweight module (5.1M parameters, 0.36% of the 1.4B base model) that integrates antigen geometric features (via GearNet) and epitope interface information into the model's latent space.

The paper also introduces a novel preference optimization framework using Calibrated Direct Preference Optimization (Cal-DPO). Candidate antibody sequences are generated by the SFT model and scored using Protenix (a reproduction of AlphaFold3) for binding confidence (AF3 score) and AntiBERTy for sequence plausibility (pseudo-log-likelihood, PLL). These scores are weakly correlated (Spearman ρ = 0.13), providing orthogonal signals. Preference pairs are constructed using a dual-threshold rule: a sequence is preferred only if it exceeds another by at least 0.2 in AF3 score and 0.1 in PLL. This yields approximately 30k preference pairs used for Cal-DPO fine-tuning.

Experiments are conducted on two tasks: de novo antibody sequence-structure co-design (designing all six CDRs) and epitope-binding CDR-H3 co-design. Results show AAMFM-CalDPO achieves state-of-the-art performance on key metrics including PLL, AF3-score, pTM, and ipTM. For full antibody design, AAMFM-CalDPO achieves PLL of-0.87, AF3-score of 0.892, pTM of 0.908, and ipTM of 0.888, outperforming baselines like ESM3, Diffab, dyMEAN, and AbX. For CDR-H3 design, it achieves PLL of-1.18, AF3-score of 0.894, and ipTM of 0.890, also surpassing baselines.

Ablation studies confirm each component's contribution: removing SFT leads to less stable DPO optimization; removing the antigen adapter causes the most significant functional degradation; disabling epitope input degrades foldability and binding confidence; and Cal-DPO improves sequence plausibility, foldability, and antigen-binding confidence compared to SFT alone.

The main limitation is that designed antibodies have not yet been validated through in vitro experiments, though the authors are collaborating with experimental groups. Future work includes generalizing the framework beyond antibodies to other protein complexes. The code is available at https://github.com/XL-S224/AAMFM.

Improvements to AI systems:

1. Multimodal integration of sequence, structure, and binding context – Extend the AAMFM approach to other protein-protein interactions (e.g., enzyme-substrate, receptor-ligand) by replacing the antigen adapter with a generic binding partner adapter that fuses geometric features and interface residues from any target. This enables a single foundation model to design binders for arbitrary protein targets, not just antibodies.

2. Calibrated preference optimization with orthogonal scoring signals – Generalize the Cal-DPO framework to any generative design task where multiple, weakly correlated quality metrics exist (e.g., drug-likeness + synthesis accessibility, or protein stability + activity). The dual-threshold rule can be adapted to automatically construct preference pairs from any two scoring functions, improving sample efficiency and robustness of DPO fine-tuning.

3. Lightweight domain adaptation adapters – Use the 5.1M-parameter adapter concept to rapidly specialize large multimodal models (e.g., ESM3, or future LLMs) to new biological or chemical domains without full fine-tuning. This reduces compute by 99.6% and enables on-the-fly adaptation to new target families (e.g., nanobodies, peptides, or small-molecule binders) with minimal data.

4. Epitope-aware geometric conditioning – Incorporate explicit interface-aware geometric features (from GearNet) into any generative model for molecular complexes. This improves the model's ability to respect spatial constraints at binding interfaces, leading to higher foldability and binding confidence in generated candidates—applicable to designing inhibitors, protein-protein interaction blockers, or synthetic antigen receptors.

5. Orthogonal scoring for self-improvement loops – Adopt the weakly correlated scoring pair (AF3-score + PLL) as a general reward model for iterative self-training. This can be used in active learning pipelines where the model proposes candidates, scores them with orthogonal metrics, and selects high-confidence pairs for further fine-tuning—accelerating discovery cycles in drug design or protein engineering.

What the improved AI system can do:

• Design high-affinity binders for any protein target (not just antibodies) by inputting the target's structure and epitope residues, and generating full sequences and 3D structures of candidate binders with high binding confidence and foldability.

• Adapt a single large foundation model to new protein families in minutes using the lightweight adapter, without retraining the base model—enabling rapid response to emerging targets (e.g., viral variants, cancer neoantigens).

• Automatically generate and rank candidate designs using two independent quality signals, reducing false positives and improving the reliability of in silico screening before experimental validation.

• Iteratively improve its own design capability via the calibrated preference optimization loop, using its own generated candidates as training data—leading to continuous performance gains with each cycle.

• Provide explainable interface-aware designs by explicitly encoding epitope geometry, allowing researchers to see which residues are critical for binding and to rationally mutate designs for improved specificity or reduced off-target effects.