arxiv: 2607.20057
license: mit
tags:
- protein
- antibody
- antibody-design
- multimodal
- biological-sequence-modeling
- structure-aware
AAMFM: Antigen-specific Antibody Multi-modal Foundation Model
AAMFM is a multi-modal foundation model for antigen-specific functional antibody design. It jointly models antibody sequence and structure, conditioned on antigen geometry and epitope annotations.
Hierarchical training and inference pipeline of AAMFM.
Model description
AAMFM follows a hierarchical training strategy:
- Antibody-domain adaptation: continual pre-training on paired antibody sequence--structure data.
- Antigen-specific supervised fine-tuning: a lightweight adapter fuses antigen geometric features with epitope information.
- Preference alignment: calibrated DPO aligns the model with structure-based functional preferences and sequence plausibility.
At inference time, the model is conditioned on an antigen, epitope-site annotations, and antibody framework sequences to generate CDR sequences and a complete antibody structure.
Checkpoint
| File | Description |
|---|---|
| AAMFM.pt | AAMFM model checkpoint. |
Usage
Use this checkpoint with the official AAMFM code repository. The repository contains the training, generation, evaluation, and pseudo-log-likelihood utilities, together with instructions for the required external assets.
Generation requires a compatible base-model backend, its externally supplied weights, antigen geometric features, and antigen/epitope inputs. See the repository README for the complete environment and runtime setup.
Intended use and limitations
This checkpoint is intended for research in computational antibody design. Its outputs are computational candidates, not experimentally validated antibodies, and must not be used for clinical, diagnostic, or therapeutic decisions without appropriate experimental validation and expert review.
Citation
If you use AAMFM, please cite:
@article{shi2026aamfm,
title={Antigen-specific Antibody Multi-modal Foundation Model for Functional Antibody Design},
author={Shi, Xiaoliang and Wang, Zichen and Ma, Runze and Zhang, Zhongyue and Zheng, Shuangjia},
journal={arXiv preprint arXiv:2607.20057},
year={2026}
}
License
The AAMFM checkpoint and the accompanying source-code repository are released under the MIT License.