Comprehensively Deciphering Plasmids: Enhancing Sequence Analysis with Textual Insights

Abstract

A plasmid is a small, circular DNA molecule mainly found in host bacteria. It plays a crucial role in carrying genes that confer beneficial traits to its host. For instance, the plasmid pKpQIL, derived from the bacterium Klebsiella pneumoniae ST 258, encodes the carbapenemase enzyme responsible for conferring resistance to carbapenem antibiotics. As a result, ST 258 can survive and even multiply in the presence of these antibiotics. Another key feature of plasmids is their ability to transfer between different bacteria through cell contact. For example, scientists have demonstrated that plasmid F, originally found in Escherichia coli, can be transferred to other bacteria within the Enterobacteriaceae family. In summary, these two characteristics of plasmids facilitate the rapid spread of traits, particularly antibiotic resistance and virulence, within pathogenic bacterial populations. This poses significant threats to public health and presents challenges for developing new treatment strategies, thereby underscoring the importance of plasmid research.

With advancements in metagenomic sequencing techniques and computational plasmid identification tools, researchers have successfully curated large-scale plasmid DNA data. For example, over 1.5 million plasmid sequences have been collected in the IMG/PR and PIPdb databases. However, most of the newly identified plasmids are not well-characterized in terms of their properties and functionalities. Therefore, in this thesis, I aim to address this gap by developing computational methods to characterize plasmid properties from various facets, including host range, protein function, ecosystem association, and risk assessment. In particular, we enhanced the comprehensive plasmid characterization tasks by integrating methodologies that span across the bioinformatics and natural language processing (NLP) fields. Specifically, as DNA and proteins possess specific structures and encode rich genetic information, there are significant commonalities between biological sequences and human language. This similarity inspires us to conceptualize plasmids as biological languages and to employ Transformer-based language models to learn their features.

The plasmid host range refers to the spectrum of bacteria in which a particular plasmid can be replicated, maintained, and transferred. Accurate prediction of the host range will help identify plasmids' role in the evolution and adaptation of bacteria, particularly human pathogens. Nevertheless, predicting the plasmid host range faces three challenges: the dissimilar sequence composition and copy numbers between plasmids and their hosts, the high diversity among plasmids, and the low functional annotation rate of plasmid-encoded proteins. The performance of current alignment-based host prediction tools is primarily limited by low recall for novel plasmids and high computational costs. On the other hand, machine learning-based tools demonstrate limited capability at lower taxonomic ranks. Here, we present HOTSPOT, which leverages a Transformer model to learn plasmid languages defined by protein clusters and replication- and mobility-related protein families for accurate host prediction from phylum to species level. A Monte Carlo Dropout-based early-stopping search is also designed for plasmids with a broad host range. Rigorous evaluation of HOTSPOT on multiple datasets, including both simulated and real metagenomes, demonstrated that it outperforms other popular methods. Additionally, our implemented top-down tree search algorithm enables us to retrieve higher taxonomic level results for plasmids associated with novel organisms and those that have multiple hosts.

Plasmid-specific proteins can be classified into two categories: core proteins and accessory proteins. Accurately annotating plasmid-encoded proteins is a crucial step for improving host prediction and facilitating the identification of plasmid-carried traits, which can further help pinpoint sources of certain bacterial infections. However, high-quality functional annotations for plasmid-encoded proteins remain limited. For example, only approximately 29.34% of the 678,197 non-redundant proteins encoded on 47,871 RefSeq complete plasmids are annotated with the most widely adopted Gene Ontology (GO) knowledgebase. Although many deep learning-based GO term prediction tools have been developed, none incorporate plasmid characteristics to optimize the GO annotation. Furthermore, two challenges are encountered: the high diversity of functions and the presence of unidentified signatures among the substantial number of unannotated proteins. In this study, we developed PlasGO, a three-layer architecture aimed at improving GO prediction by learning the modularization patterns of plasmids, where proteins with similar functions tend to be located close to each other. PlasGO's core architecture is a BERT model to capture the global context between proteins within a plasmid. PlasGO significantly increased the functional annotation rate by assigning high-confidence GO terms to over 95\% of previously unannotated proteins. Our protein annotation results with high predicted confidence scores can also correct existing erroneous annotations in public databases.

The rapidly expanding plasmid DNA data offers significant potential for comprehensively characterizing plasmids from multiple facets. However, there is a need for a versatile computational framework to address label discrepancies that arise between different databases, as they focus on various perspectives. One solution is to convert the various annotations into textual form and align the multi-modal features learned from both the textual annotations and plasmid sequences. Building on this concept, we introduce PlasRAG, a plasmid analysis tool designed for two main purposes: plasmid property characterization and plasmid DNA retrieval. When querying with plasmid DNA, PlasRAG predicts associated properties and relevant literature. We will also provide detailed predictions along with their corresponding confidence scores to give users insights into reliability. By integrating the predicted information through the retrieval-augmented generation (RAG) framework, we can leverage a powerful large language model (LLM) to provide accurate answers to user questions. Moreover, PlasRAG can retrieve eligible plasmids based on user-selected properties as queries. Quantitative ablation studies demonstrated that our DNA-text alignment model design is reasonable and achieves reliable performance. Additionally, the PlasRAG analysis of a novel human gut plasmid dataset sourced from multiple nations concluded that plasmids with a broader host range and those encoding antibiotic resistance genes (ARGs) tend to be more widely distributed. For all three developed tools, including HOTSPOT, PlasGO, and PlasRAG, we provide the source codes and training scripts, allowing for the updating of the software over time.

Date of Award	7 Aug 2025
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Yanni SUN (Supervisor)