Desenvolvimento de uma pipeline para vacinologia reversa baseada em arquitetura transformer

Abstract

Reverse vaccinology, combined with bioinformatics, enables the prediction of immunogenic proteins from genomic data, facilitating the development of biotechnological products. The in silico prediction of vaccine targets involves models based on natural language processing (NLP), applied to the structural and functional analysis of proteins, biomedical text mining, and synthetic data generation. This study presents ReVarcine, a tool designed to predict protein characteristics relevant to immunogenicity. Its process includes preprocessing sequences from the UniProt and SwissProt databases, removing incomplete or redundant entries, and using Keras-based tokenization for numerical representation. The sequences were divided into training and testing sets, distinguishing Gram-positive and Gram-negative bacteria. Eight deep learning models were developed to predict signal peptides, cleavage sites, subcellular localization, and secondary structures such as beta barrels and alpha helices. The architectures include embedding layers, transformers, LSTM, and dense layers, integrated into a modular pipeline. This design ensures scalability and adaptability for different organisms and analytical needs. Performance evaluation used metrics such as MCC and F1 score, comparing ReVarcine with established tools, including SignalP, PSORT, and PSIPRED. The results indicate that ReVarcine often outperforms traditional predictors in proteomic analysis, standing out for its versatility and accuracy. ReVarcine innovates by unifying multiple predictive tasks, reducing reliance on various specialized tools. This optimizes computational time and minimizes errors in data processing across platforms. Its integrated approach sets a new standard for immunoinformatics, making it a promising tool to meet the growing demands of reverse vaccinology.