An in-depth evaluation of metagenomic classifiers for soil microbiomes

Recent endeavours in metagenomics, exemplified by projects such as the human microbiome project and TARA Oceans, have illuminated the complexities of microbial biomes. A robust bioinformatic pipeline and meticulous evaluation of their methodology have contributed to the success of these projects. The soil environment, however, with its unique challenges, requires a specialized methodological exploration to maximize microbial insights. A notable limitation in soil microbiome studies is the dearth of soil-specific reference databases available to classifiers that emulate the complexity of soil communities. There is also a lack of in-vitro mock communities derived from soil strains that can be assessed for taxonomic classification accuracy. In this study, we generated a custom in-silico mock community containing microbial genomes commonly observed in the soil microbiome. Using this mock community, we simulated shotgun sequencing data to evaluate the performance of three leading metagenomic classifiers: Kraken2 (supplemented with Bracken, using a custom database derived from GTDB-TK genomes along with its own default database), Kaiju, and MetaPhlAn, utilizing their respective default databases for a robust analysis. Our results highlight the importance of optimizing taxonomic classification parameters, database selection, as well as analysing trimmed reads and contigs. Our study showed that classifiers tailored to the specific taxa present in our samples led to fewer errors compared to broader databases including microbial eukaryotes, protozoa, or human genomes, highlighting the effectiveness of targeted taxonomic classification. Notably, an optimal classifier performance was achieved when applying a relative abundance threshold of 0.001