Background

With advancements in metagenomics, the complex ecosystems of microorganisms in the environment and within the human body are gradually being revealed. Bacteriophages, viruses that infect bacteria, are of particular interest as they regulate bacterial populations and are closely involved in human health. However, completely reconstructing phage genomes using genomic sequencing technologies faces several limitations. Phage genomes are small and contain many repetitive sequences, making them prone to fragmentation when sequencing data is assembled. Existing assembly tools have limitations in connecting these fragmented sequences into a complete genome map, which has hindered the elucidation of their actual roles in microbial ecosystems.

Key Findings

The research team, led by Dr. Ruihang Wang from the DeepOmics Lab at the City University of Hong Kong, developed 'PALACE' (Phage Assembly via Deep-learning And Conjugate graph-based framEwork), a hybrid analysis framework that combines homology information and deep learning models. PALACE adopts two complementary strategies to sensitively detect phage-derived sequences in metagenomic data. It uses homology-based searches to find similarities with known viral proteins and simultaneously employs a deep learning prediction model trained on the unique sequence patterns of novel phages. Based on this cross-validated viral signal data, the research team applied 'conjugate graph' theory to refine the assembly process. By mathematically reconstructing the relationships between fragmented sequences, they were able to restore long, connected phage genomes while reducing errors.

In performance verification tests using artificial data, PALACE achieved excellent results with an F1-score (a measure of genome reconstruction accuracy) between 0.92 and 1.00. This represents an improvement of 0.21 to 0.48 F1-score points compared to existing state-of-the-art viral identification and assembly tools. Furthermore, when the research team applied PALACE to actual human gut microbiome metagenomic data, they achieved a 56% improvement in the median genome completeness of the reconstructed phage genomes compared to competing methods. Based on this, the researchers were able to identify thousands of high-quality phage genomes.

Significance and Prospects

This research is considered a breakthrough in the field of viral metagenomics, where information is prone to fragmentation, by combining artificial intelligence (AI) and graph theory. With the ability to reconstruct high-quality phage genomes, it opens the way to more accurately track the biological mechanisms by which phages affect human diseases or microbial community structures. However, the research team notes that in diverse environmental samples or rare microbial communities, the efficiency of homology searches may be partially reduced due to a lack of reference databases. A future challenge is to further reduce database dependence by strengthening unsupervised learning techniques.