Contamination of widely used databases compromises rRNA based taxonomic assignment of metagenomes and metatranscriptomes

Taxonomic annotation of metagenomic and metatranscriptomic datasets frequently relies on aligning small subunit (SSU) rRNA reads to annotated reference databases. However, widely used SSU databases, including SILVA, GTDB, Eukaryome, and those distributed with SortMeRNA, are contaminated with large subunit (LSU) rRNA sequences. Low levels of LSU contamination can lead to large numbers of LSU reads being incorrectly identified as SSU and have serious consequences for downstream taxonomic assignments. The problem can be avoided by rigorously removing LSU sequences from databases used for taxonomic annotation, which we have done for KSGP 4.0. Alternatively, initial SSU read selection can be carried out with a carefully curated small SSU database that is free of LSU contamination. We illustrate these two approaches in combination using a metatranscriptomic dataset from an estuarine sediment. Without database cleaning, LSU-derived reads can make up half of supposed SSU sequences and are assigned to a small number of apparently dominant but artefactual taxa. Database cleaning removes this problem and we provide a script, total_rnaseq, to annotate total RNASeq or metagenomic data using this approach. The KSGP 4.0 database provides substantially improved annotation of Archaea compared to the SILVA database and moderate and small improvements for eukaryotes and bacteria respectively.