Research map / Label-efficient and robust learning / Recognition and long-tail

Immunogenicity Prediction with Dual Attention Enables Vaccine Target Selection

ICLR 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Dense features and foundation priors · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes VenusVaccine, a dual-attention model fusing pre-trained PLM sequence embeddings, two-scale structure tokens and E/Z physicochemical descriptors for immunogenicity prediction, and releases the ImmunoDB benchmark (~7.2k antigens).

Problem

Predicting whether protein or peptide antigens from bacteria, viruses, and tumors are immunogenic (protective) for reverse vaccinology; prior methods use heavily compressed features and simple models, giving limited accuracy and poor cross-species generalization.

Approach

Cross-attention between sequence PLM embeddings (keys/values) and two structural token scales (FoldSeek 20-dim finer tokens and ESM3 4096-dim coarser tokens as queries), with RoPE, then attention pooling over concatenated features plus handcrafted E- and Z-descriptors.

Key results

• On ImmunoDB-Virus, VenusVaccine-Ankh reaches 92.2% accuracy, 84.3 MCC and 0.85 KS, above the best baseline (XGBoost 89.1% ACC).
• On ImmunoDB-Bacteria, VenusVaccine-ProtBert achieves 84.5% accuracy and 65.9 MCC versus 83.3% ACC for VaxiJen3.0.
• On the H. pylori post-hoc test, VenusVaccine identifies 123 PVCs containing all 11 known protective antigens (100% recall) with fold-enrichment 15.11 vs 8.81 for the best tree baseline.

Measured effects reported in the paper

Surprising findings

• On the Tumor dataset, VaxiJen3.0 predicts all samples positive (100% recall, 39.0% accuracy), showing recall-only metrics can be misleading.
• AlphaFold2 structures gave better performance than ESMFold structures in the ablation, but the gap is small for some models (e.g., ProtBert on bacteria shows higher accuracy with AF2).

Limitations

• Negative labels for bacteria and virus rely on BLAST and VaxiJen filtering, which the authors note can bias toward predicting positives.
• Web-server baselines could not be retrained, possibly giving them unfair advantage or leakage, and VaxiJen3.0 was used for negative labeling.
• Post-hoc validation is limited to a few proteomes and does not include new wet-lab experiments.

Experimental setup

Datasets: ImmunoDB-Bacteria ImmunoDB-Virus ImmunoDB-Tumor UniProtKB NCBI IEDB Protegen

Benchmarks: ImmunoDB-Bacteria ImmunoDB-Virus ImmunoDB-Tumor H. pylori proteome post-hoc SARS-CoV-2 proteome post-hoc

Metrics: accuracy AUC-ROC precision recall F1 MCC top-30 precision (T30) cross entropy KS statistic fold-enrichment

Baselines: Random Forest Gradient Boosting XGBoost SGD Logistic Regression MLP SVM KNN VaxiJen2.0 VaxiJen3.0 Vaxi-DL VirusImmu

Builds on: ESM2 ESM3 FoldSeek RoPE VaxiJen

Closest papers by meaning

• AsEP: Benchmarking Deep Learning Methods for Antibody-specific Epitope Prediction (NeurIPS 2024 · similarity 0.83)
• Improving Antibody Humanness Prediction using Patent Data (ICML 2024 · similarity 0.82)
• A Label Disambiguation-Based Multimodal Massive Multiple Instance Learning Approach for Immune Repertoire Classification (AAAI 2024 · similarity 0.82)
• Cross-Gate MLP with Protein Complex Invariant Embedding Is a One-Shot Antibody Designer (AAAI 2024 · similarity 0.81)
• Pareto-Optimal Energy Alignment for Designing Nature-Like Antibodies (NeurIPS 2025 · similarity 0.81)
• Relation-Aware Equivariant Graph Networks for Epitope-Unknown Antibody Design and Specificity Optimization (AAAI 2025 · similarity 0.80)