Accurate Reconstruction of Microbial Strains from Metagenomic Sequencing Using Representative Reference Genomes

Exploring the genetic diversity of microbes within the environment through metagenomic sequencing first requires classifying these reads into taxonomic groups. Current methods compare these sequencing data with existing biased and limited reference databases. Several recent evaluation studies demonstrate that current methods either lack sufficient sensitivity for species-level assignments or suffer from false positives, overestimating the number of species in the metagenome. Both are especially problematic for the identification of low-abundance microbial species, e. g. detecting pathogens in ancient metagenomic samples. We present a new method, SPARSE, which improves taxonomic assignments of metagenomic reads. SPARSE balances existing biased reference databases by grouping reference genomes into similarity-based hierarchical clusters, implemented as an efficient incremental data structure. SPARSE assigns reads to these clusters using a probabilistic model, which specifically penalizes non-specific mappings of reads from unknown sources and hence reduces false-positive assignments. Our evaluation on simulated datasets from two recent evaluation studies demonstrated the improved precision of SPARSE in comparison to other methods for species-level classification. In a third simulation, our method successfully differentiated multiple co-existing Escherichia coli strains from the same sample. In real archaeological datasets, SPARSE identified ancient pathogens with[≤] 0.02% abundance, consistent with published findings that required additional sequencing data. In these datasets, other methods either missed targeted pathogens or reported non-existent ones. SPARSE and all evaluation scripts are available at https://github.com/zheminzhou/SPARSE.