An explainable deep learning approach for antimicrobial resistance proteins functional annotation

• Ano de defesa: 2025
• Autor(a) principal: Borelli, Tiago Cabral
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertacoes da USP; Universidade de São Paulo; Faculdade de Medicina de Ribeirão Preto
• Palavras-chave em Português: Anotação funcional; Aprendizado de máquina; Redes neurais convolucionais; Resistência antimicrobiana; Antimicrobial resistance; Convolutional neural networks; Functional annotation; Machine learning
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/17/17136/tde-17032026-164855/

Bacterial resistance to antimicrobials is one of the main global public health problems, according to the World Health Organization. It burdens public health systems more than AIDS and malaria combined and therefore needs to be addressed on multiple fronts: from the development of new antimicrobials to the creation of genomic surveillance strategies. However, the identification of proteins involved in antimicrobial resistance is limited due to the high similarity required between candidate proteins and those used as references for sequence alignment, resulting in a high number of false negatives. Therefore, there is a need to develop tools that do not rely on direct sequence alignment. This work presents a study on the current state of resistance protein databases as well as the development of DeepSEA, an alignment-independent tool for the functional annotation of resistance proteins using convolutional neural networks. The results show a high degree of redundancy both within and between databases. Annotation errors propagated during the concatenation of these databases were also found, particularly among multidrug class proteins. DeepSEA outperformed alignment-based \"state-of-the-art\" tools and showed performance comparable to more complex transformer-based models. Furthermore, it was possible to manipulate the weights of the convolutional layers and demonstrate that DeepSEA learned to recognize protein domains related to antimicrobial resistance.