[Paper Review] Immunogenicity Prediction with Dual Attention Enables Vaccine Target Selection

Li Song, Yang Tan+3 more|arXiv (Cornell University)|Oct 3, 2024

vaccines and immunoinformatics approachesBiochemistry, Genetics and Molecular Biology3 citations

TL;DR

This paper introduces ProVaccine, a deep learning model with a dual attention mechanism that integrates protein sequence, structure, and physicochemical features to improve immunogenicity prediction. It achieves state-of-the-art performance on a new, large-scale dataset of 9,500 antigen sequences from bacteria, viruses, and tumors, demonstrating superior accuracy and generalizability over existing methods.

ABSTRACT

Immunogenicity prediction is a central topic in reverse vaccinology for finding candidate vaccines that can trigger protective immune responses. Existing approaches typically rely on highly compressed features and simple model architectures, leading to limited prediction accuracy and poor generalizability. To address these challenges, we introduce VenusVaccine, a novel deep learning solution with a dual attention mechanism that integrates pre-trained latent vector representations of protein sequences and structures. We also compile the most comprehensive immunogenicity dataset to date, encompassing over 7000 antigen sequences, structures, and immunogenicity labels from bacteria, virus, and tumor. Extensive experiments demonstrate that VenusVaccine outperforms existing methods across a wide range of evaluation metrics. Furthermore, we establish a post-hoc validation protocol to assess the practical significance of deep learning models in tackling vaccine design challenges. Our work provides an effective tool for vaccine design and sets valuable benchmarks for future research. The implementation is at https://github.com/songleee/VenusVaccine.

Motivation & Objective

• To address the limited accuracy and poor generalizability of existing immunogenicity prediction models in reverse vaccinology.
• To develop a deep learning framework that effectively integrates multimodal protein representations—sequences, structures, and physicochemical properties.
• To compile the most comprehensive immunogenicity dataset to date, covering antigens from bacteria, viruses, and tumors.
• To establish a practical, post-hoc validation protocol that balances biological relevance with computational feasibility, avoiding reliance on costly wet-lab experiments.
• To provide a benchmark for future research in vaccine target selection through a robust model and evaluation framework.

Proposed method

• ProVaccine employs a dual-attention mechanism to enable cross-scale communication between amino acid residues at both local and global levels.
• It encodes protein sequences and structures using pre-trained latent vector representations, incorporating atomic-level and peptide-level structural tokens.
• Hand-crafted physicochemical features based on Z-descriptors and E-descriptors are integrated into the model to enhance immunogenicity prediction.
• Attention pooling is applied to AA-level hidden representations to generate protein-level embeddings that capture key immunological signals.
• The model uses a multimodal fusion strategy to combine sequence, structural, and physicochemical representations into a unified representation for binary classification.
• A comprehensive post-hoc validation protocol is designed, using biologically relevant pathogens like *Helicobacter pylori* and SARS-CoV-2, to assess practical utility beyond standard metrics.

Experimental results

Research questions

• RQ1 Can a deep learning model with dual attention mechanisms improve immunogenicity prediction accuracy by integrating multimodal protein representations?
• RQ2 How does ProVaccine perform in cross-species immunogenicity prediction compared to existing methods?
• RQ3 To what extent can a large-scale, multi-pathogen immunogenicity dataset enhance model generalizability and training robustness?
• RQ4 Can a post-hoc validation protocol effectively assess the practical significance of immunogenicity predictions without requiring wet-lab experiments?
• RQ5 What is the contribution of hand-crafted physicochemical features and structural tokens to model performance in vaccine target selection?

Key findings

• ProVaccine significantly outperforms existing methods across multiple evaluation metrics, including AUC, AUPRC, and F1-score, on the newly compiled Immuno dataset of over 9,500 labeled antigens.
• The model demonstrates strong generalizability, particularly in cross-species immunogenicity prediction, due to its ability to capture complex relationships between protein features and immune response.
• The dual-attention mechanism enables effective communication between amino acid residues at different structural scales, enhancing the model’s capacity to detect biologically relevant patterns.
• Post-hoc validation on *Helicobacter pylori* and SARS-CoV-2 confirmed that ProVaccine’s predicted immunogenicity scores effectively prioritize known protective antigens, validating its practical utility.
• The integration of hand-crafted Z- and E-descriptor features significantly improves model performance, highlighting the value of domain-specific physicochemical priors in deep learning for vaccine design.
• The study establishes a new benchmark dataset (Immuno) and evaluation protocol that set a higher standard for future research in immunogenicity prediction.