Algorithmic Optimization and Application for Viral Analysis in Metagenomics

Abstract

Viruses are the most abundant biological entities in the biosphere, playing pivotal roles in global ecosystems as obligate intracellular parasites. Beyond their impact on human health, livestock, and crops, viruses drive global biogeochemical cycles by infecting diverse hosts ranging from prokaryotes to eukaryotes. While the advent of metagenomics has exponentially expanded the known virosphere, a vast “viral dark matter” remains. As of this writing, the ICTV has officially recorded only 16,313 species, far fewer than the estimated 10³¹ viral particles on Earth and millions of viral sequences identified. Given the profound ecological significance and immense diversity of viruses, it is imperative to characterize their composition and function within ecosystems. As viruses are ubiquitous, a typical analytical workflow begins with identifying viruses from a wide range of environmental and host-associated samples, followed by host prediction and taxonomic classification. The thesis presents integrated contributions to these three domains.

Some intrinsic viral properties complicated their study. First, viruses exhibit mutation rates over ten times higher than cellular organisms, driven by strong selective pressure from host immune systems to enable rapid adaptation. Coupled with their small genome sizes, these factors lead to extraordinary genetic diversity and heterogeneity. Second, although metagenomic sequencing accelerates virus discovery, it predominantly generates fragmentary contigs rather than complete genomes, limiting signals for downstream analyses. Third, known viruses merely represent the tip of the iceberg. Incomplete reference databases and evolving taxonomic frameworks demand models with strong generalizability and adaptability. Despite multiple available tools for viral analysis, these challenges collectively define the complexity of the field.

We first focused on the identification and host prediction of RNA viruses, followed by the taxonomic classification of all viruses. Among viruses, RNA viruses exhibit even higher mutation rates due to the lack of proofreading mechanisms during replication. Their low abundance in metagenomic data requires highly specific detection tools, while their high genetic diversity challenges alignment-based methods. To accurately identify RNA viruses from metagenomic samples, we developed VirBot, a tool based on protein families and adaptive score cutoffs. To enhance sensitivity, we constructed a comprehensive RNA viral protein family database comprising 1,384 families. We built profile hidden Markov models (pHMMs) for these families to identify conserved sites and retrieve remote homologs. By scanning millions of non viral sequences, we established adaptive score cutoffs for each pHMM, significantly reducing false positive rates. Benchmarked against state of the art tools on simulated and real metagenomic datasets, VirBot demonstrated superior specificity and sensitivity in RNA virus identification. When applied to a novel RNA phage dataset, VirBot achieved an 8.5% higher recall than the next best tool. Notably, VirBot successfully identified nearly all viral sequences from medical samples and provided correct taxonomic labels, highlighting its potential for medical applications.

Determining hosts of RNA viruses is of particular importance, as they are causative agents of some infectious diseases and frequently exhibit host-jumping events, raising public health and ecological concerns. High throughput sequencing has accelerated virus discovery but complicated host identification for metagenome assembled viruses. Even in host associated samples, assigning the sample origin as the host is not always correct, not to mention the environmental samples. Unlike DNA viruses (mostly bacteriophages), which may share homologs with hosts through horizontal gene transfer, RNA viruses primarily infect eukaryotes and generally do not share homologs with their hosts. Their high mutation rates and vast diversity further obscure host inference. Existing methods are often restricted to specific RNA virus types or particular hosts such as humans. In this work, we developed RNAVirHost, a machine learning-based tool that predicts RNA virus hosts solely from viral genomes. We formulated the problem as a hierarchical multi class classification problem, training separate models per virus order to mitigate extreme class imbalance. After comprehensive evaluation of commonly used features, methods, and architectures across multiple datasets, RNAVirHost, combining sequence similarity and genomic traits, achieved an optimal balance between accuracy and generalizability. In leave one genus out scenarios, it improved accuracy by 12.1% over alignment based methods. When tested on newly identified viruses, RNAVirHost attained over 70% accuracy.

Classifying viruses into taxonomic units is essential for understanding their roles and functions. However, viral mosaic genomes increase classification complexity. Genetic recombination and horizontal gene transfer with hosts blur evolutionary boundaries, resulting in heterogeneous evolutionary distances between virus genera. Additionally, the rapid growth of novel viruses and the evolving taxonomic framework make this an open set problem, requiring frequent updates and adaptability. Current tools struggle to perform well at the genus level. To address practical needs for computational taxonomic classification, we developed ViraTaxM, an alignment based tool designed to resolve genus level ambiguity in viral taxonomy. To tackle taxonomic uncertainty caused by homologous regions between viruses from different genera, it introduced two alignment based uncertainty measures When tested on newly identified viruses, RNAVirHost attained over 70% accuracy. Across multiple benchmark datasets, ViraTaxM outperformed popular classifiers while maintaining an optimal balance between classification rate, precision, and efficiency. In open class detection scenarios, it reduced misclassification by 10%. For viruses from novel genera, the integrated clustering module enabled rapid yet reliable genus level analysis.

In summary, our study explored key challenges in viral metagenomics and developed computational tools to address them. In collaboration with biologists, we translated biological insight into practical solutions. We developed three tools: VirBot for RNA virus identification, RNAVirHost for RNA virus host prediction, and ViraTaxM for genus-level virus taxonomic classification. Through comprehensive evaluation under diverse scenarios, our tools demonstrated robust performance and contributed to a deeper understanding of viral diversity and ecology.

Date of Award	21 Aug 2026
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Yanni SUN (Supervisor)