Enhancing Vaxign-DL for Vaccine Candidate Prediction with added ESM-Generated Features

Area of Science:

• Computational Biology
• Vaccinology
• Machine Learning

Background:

• Existing machine learning models like Vaxign-DL predict bacterial protective antigens using 509 features.
• Deep learning techniques are crucial for advancing vaccine design.

Purpose of the Study:

• To evaluate the impact of protein folding features on vaccine antigen prediction.
• To improve the accuracy and performance of the Vaxign-DL model.

Main Methods:

• Utilized the ESM program to derive 1,280 protein folding features.
• Integrated ESM-derived features with existing Vaxign-DL features for antigen prediction.
• Conducted Leave-One-Pathogen-Out Validation (LOPOV) on 10 bacterial pathogens.

Main Results:

• ESM-derived features alone achieved prediction performance comparable to the original Vaxign-DL.
• Combining ESM and traditional features significantly improved prediction performance across seven metrics (specificity, sensitivity, AUROC).
• LOPOV study demonstrated significant improvement in predicting vaccine antigens using ESM features.

Conclusions:

• Protein folding features offer added value for vaccine antigen prediction.
• The enhanced Vaxign-DL model with ESM features shows improved predictive capabilities.
• This study pioneers the use of protein folding features in vaccine antigen discovery.