Communication Dans Un Congrès Année : 2025

Improved strain-level metagenome assembly for modern long reads

Résumé

Advances in long-read sequencing technologies from PacBio and Oxford Nanopore Technologies (ONT) have paved the way for the development of methods that significantly improve the quality of metagenome assemblies, particularly with the modern PacBio HiFi and ONT R10.4 technologies. Nevertheless, standard metagenome assemblers, while providing high-quality species-level reconstructions, are often unable to distinguish and accurately reconstruct genomes of highly similar bacterial strains (i.e., with >99% sequence identity). Moreover, several long-readbased methods have been recently proposed to address the challenge of reconstructing bacterial strain genomes (or haplotypes) from metagenomic data.

In this work we show that existing metagenome assembly tools either produce highly fragmented assemblies or generate longer contigs at the expense of numerous strain-level misassemblies. We also present Strainberry2, an improved version of Strainberry, specifically designed for PacBio HiFi and ONT R10.4 sequencing data, aimed at providing a better balance between contiguity and strain-level accuracy. As is common for strain-separation methods, Strainberry2 takes as input a strain-oblivious metagenome assembly (i.e., a species-level assembly) together with a set of accurate long reads. It produces an augmented assembly that recovers strain diversity - absent in the input assembly - by generating strain-specific sequences. Strainberry2 distinguishes itself from competing methods by introducing an upstream dereplication step and a novel strain-phasing algorithm. The dereplication step is aimed at reducing ambiguous read mappings, which are critical for reliably detecting single-nucleotide and structural variants. The strain-phasing algorithm uses an iterative approach that leverages an implicit variant graph constructed from co-occurring single-nucleotide variants in the reads. At each iteration, candidate bacterial haplotypes are defined, while retaining only those supported by a sufficient number of reads.

We ran Strainberry2 and state-of-the-art standard and strain-aware metagenome assemblers on 64 simulated datasets, covering different sequencing technologies, varying numbers of strains, multiple bacterial species, and either uniform or staggered strain abundances. We further ran these tools with real sequencing data on the ZymoBIOMICS Gut Microbiome Standard mock community, which includes five highly similar strains of Escherichia coli and for which both PacBio HiFi and ONT R10.4 reads are publicly available. The resulting assemblies were evaluated with MetaQUAST. To overcome the limitations of MetaQUAST in assessing assemblies against highly similar reference genomes, we complemented the evaluation with a k-mer-based approach. On the 64 simulated datasets, Strainberry2 achieved, on average, comparable or significantly better results across all qualitative metrics (e.g., 2-7X fewer misassemblies according to MetaQUAST and at least 10X fewer k-mer-based strain-switch errors). On the real-data mock community, Strainberry2 showed the same trend, particularly with ONT R10.4 data. With PacBio HiFi reads, Strainberry2 delivered results on par with the state-of-the-art hifiasm-meta, providing slightly higher accuracy across qualitative metrics. While hifiasm-meta reached greater contiguity (4X higher NGA50), this came at the cost of 20% more duplicated sequences, highlighting the more balanced trade-off achieved by Strainberry2.

Fichier principal
Vignette du fichier
seqBIM-2025-abstract-vicedomini.pdf (265.3 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-05424057 , version 1 (18-12-2025)

Licence

Identifiants

  • HAL Id : hal-05424057 , version 1

Citer

Riccardo Vicedomini, Rayan Chikhi. Improved strain-level metagenome assembly for modern long reads. SeqBIM 2025 - Journées sur les Séquences en Bioinformatique, Informatique et Mathématiques, Nov 2025, Nantes, France. ⟨hal-05424057⟩
132 Consultations
103 Téléchargements

Partager

  • More