Deep Learning-Based Computational Tools for Microbial Protein Sequence Annotation

Abstract

Microbial protein annotation remains a fundamental challenge in understanding pathogen biology, host–microbe interactions, and the molecular basis of clinically important traits such as antimicrobial resistance and virulence. Traditional homology-based and manual feature engineering–driven approaches often struggle to generalize across diverse and rapidly evolving protein families, leaving a substantial fraction of microbial proteins poorly characterized. Recent advances in deep learning provide new opportunities to capture complex sequence and structural patterns beyond conventional techniques. This dissertation develops deep learning–based frameworks for microbial protein annotation across three biologically important and complementary contexts spanning bacteria, phages, and fungi. In Chapter 2, we introduce DeepMRP, a model for predicting bacterial metal resistance proteins, genetic elements that allow bacteria to withstand metal-based antimicrobials while also promoting the co-selection and persistence of antibiotic resistance across microbial communities. DeepMRP overcomes the limitations of fixed-threshold best-hit homology methods by leveraging bit score–based similarity distributions, enabling more robust identification of distantly related sequences. Chapter 3 focuses on phage structural proteins, which are essential for host recognition and phage–host interactions. We propose DeePSP-GIN, a structure-aware graph neural network that integrates predicted protein 3D structures with protein language model embeddings to capture both spatial and sequential dependencies, improving the identification and classification of phage structural proteins. In Chapter 4, we address fungal virulence factor proteins, key determinants of fungal pathogenicity and host invasion, by constructing a curated database and developing DeepFVFP, a protein language model–based framework with a convolutional neural network that captures biologically meaningful sequence patterns and enables interpretation of model predictions to uncover signals associated with virulence. Collectively, this work demonstrates that integrating sequence, structural, and learned representations through deep learning provides a powerful and flexible framework for microbial protein annotation. The approaches developed in this dissertation improve predictive performance and offer the potential to be extended to other functionally important protein classes, ultimately contributing to a deeper understanding of microbial systems and their roles in health and disease.