ICLR 2025

VenusVaccine

Immunogenicity prediction with dual attention enables vaccine target selection — multimodal sequence, structure, and physicochemical encoding for protective versus non-protective antigen classification.

Song Li*, Yang Tan*, Song Ke, Liang Hong, Bingxin Zhou
(* equal contribution)

Paper Code
VenusVaccine overview figure
VenusVaccine classifies protective antigens via dual attention over sequence, structure, and physicochemical multimodal inputs.

What the model does

VenusVaccine predicts whether an antigen is protective or non-protective, supporting vaccine target selection from protein antigens. Inputs combine amino-acid sequence, structure-derived tokens, and physicochemical descriptors in a single multimodal encoder with dual attention.

Released checkpoints cover three antigen settings — Bacteria, Virus, and Tumor — so the same inference path can be pointed at the matching pathogen type.

Multimodal encoding

Beyond the amino-acid sequence, VenusVaccine consumes structure and property streams that are extracted into a shared JSON feature format:

  • Foldseek secondary-structure sequences for local backbone geometry.
  • ESM3 structure-sequence encoding of 3D conformation.
  • E-descriptors (5-D) and Z-descriptors (3-D) for physicochemical context.

Pretrained protein language models provide sequence backbone features through lightweight adapters, with support for ESM, Bert, and Ankh families among others.

Inference workflow

  1. Obtain a PDB (experimental or predicted, e.g. ESMFold / AlphaFold).
  2. Convert structures with pdb2json.py to assemble sequence, Foldseek, ESM3, and E/Z features.
  3. Run infer.py -i input.json -t Bacteria|Virus|Tumor against the matching checkpoint.

Outputs include a binary protective-antigen label and a probability score, suitable for ranking candidate vaccine targets before wet-lab follow-up.

Citation

If you use VenusVaccine, please cite:

@inproceedings{li2025immunogenicity,
title={Immunogenicity Prediction with Dual Attention Enables Vaccine Target Selection},
author={Song Li and Yang Tan and Song Ke and Liang Hong and Bingxin Zhou},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=hWmwL9gizZ}
}