VAXIN8: an artificial intelligence reverse vaccinology tool to predict potential vaccine candidates
1
Discipline of Public Health Medicine, University of KwaZulu-Natal, Durban, South Africa
Popul. Med. 2026;8(Supplement Supplement 1):A817
ABSTRACT
INTRODUCTION:
traditional vaccine development takes 10 to 15 years and requires cultivating pathogens in laboratories. Reverse vaccinology (RV) [1] is a computational approach that offers a faster alternative by identifying potential vaccine candidates directly from pathogen genomes. Artificial intelligence (AI) approaches such as protein language models (PLMs) pre-trained on millions of sequences offer new opportunities for antigen prediction in RV. However, optimal PLM application strategies remain unclear. This study compares fine-tuning (training the model further) versus frozen embeddings (using what the model already learned) for viral protective antigen (VPAg) prediction and benchmarks against traditional methods.
METHODS:
two PLMs were evaluated: ProteinBERT [2] and Evolutionary Scale Modeling 2 (ESM-2) [3]. For each, we compared fine-tuning versus extracting frozen embeddings combined with machine learning classifiers (logistic regression, support vector machine, k-nearest neighbors, random forest, XGBoost). Models were trained using the VPAgs-Dataset4ML dataset [4] containing 420 balanced viral sequences and externally validated on 2062 independent sequences (342 published antigens and 1720 non-antigens). Performance was assessed using area under the receiver operating characteristic curve (AUC) and compared against existing RV tools: VaxiJen [5] and Antigenic [6].
RESULTS:
frozen embeddings consistently outperformed fine-tuning. ESM-2 frozen embeddings with support vector machine achieved the highest AUC of 0.846 (VAXIN8), compared to 0.826 for ProteinBERT frozen with logistic regression. Fine-tuned models achieved lower scores: ProteinBERT fine-tuned achieved 0.816 and ESM-2 fine-tuned achieved 0.760. Traditional methods showed substantially lower performance: Antigenic (0.640) and VaxiJen (0.570).
CONCLUSIONS:
pre-trained PLMs encode immunological signals enabling accurate VPAg prediction without task-specific fine-tuning. As the first AI-powered RV tool developed in Africa, VAXIN8 enables vaccine researchers to input viral protein sequences and obtain antigen probability scores, helping prioritize candidates for experimental validation. These findings could advance vaccine candidate identification against viral pathogens, supporting public health preparedness and epidemic response.
traditional vaccine development takes 10 to 15 years and requires cultivating pathogens in laboratories. Reverse vaccinology (RV) [1] is a computational approach that offers a faster alternative by identifying potential vaccine candidates directly from pathogen genomes. Artificial intelligence (AI) approaches such as protein language models (PLMs) pre-trained on millions of sequences offer new opportunities for antigen prediction in RV. However, optimal PLM application strategies remain unclear. This study compares fine-tuning (training the model further) versus frozen embeddings (using what the model already learned) for viral protective antigen (VPAg) prediction and benchmarks against traditional methods.
METHODS:
two PLMs were evaluated: ProteinBERT [2] and Evolutionary Scale Modeling 2 (ESM-2) [3]. For each, we compared fine-tuning versus extracting frozen embeddings combined with machine learning classifiers (logistic regression, support vector machine, k-nearest neighbors, random forest, XGBoost). Models were trained using the VPAgs-Dataset4ML dataset [4] containing 420 balanced viral sequences and externally validated on 2062 independent sequences (342 published antigens and 1720 non-antigens). Performance was assessed using area under the receiver operating characteristic curve (AUC) and compared against existing RV tools: VaxiJen [5] and Antigenic [6].
RESULTS:
frozen embeddings consistently outperformed fine-tuning. ESM-2 frozen embeddings with support vector machine achieved the highest AUC of 0.846 (VAXIN8), compared to 0.826 for ProteinBERT frozen with logistic regression. Fine-tuned models achieved lower scores: ProteinBERT fine-tuned achieved 0.816 and ESM-2 fine-tuned achieved 0.760. Traditional methods showed substantially lower performance: Antigenic (0.640) and VaxiJen (0.570).
CONCLUSIONS:
pre-trained PLMs encode immunological signals enabling accurate VPAg prediction without task-specific fine-tuning. As the first AI-powered RV tool developed in Africa, VAXIN8 enables vaccine researchers to input viral protein sequences and obtain antigen probability scores, helping prioritize candidates for experimental validation. These findings could advance vaccine candidate identification against viral pathogens, supporting public health preparedness and epidemic response.
| eISSN: | 2654-1459 |
We process personal data collected when visiting the website. The function of obtaining information about users and their behavior is carried out by voluntarily entered information in forms and saving cookies in end devices. Data, including cookies, are used to provide services, improve the user experience and to analyze the traffic in accordance with the Privacy policy. Data are also collected and processed by Google Analytics tool (more).
You can change cookies settings in your browser. Restricted use of cookies in the browser configuration may affect some functionalities of the website.
You can change cookies settings in your browser. Restricted use of cookies in the browser configuration may affect some functionalities of the website.
