The standard framework to benchmark read classifiers. Prokaryotic and viral complete genomes with OK taxonomy check status were downloaded from the NCBI RefSeq collection. Plasmids and short sequences were removed from prokaryotic genomes as a filtering step. For each species with at least two strains, 20% of its strains were randomly picked into a test set. All remaining genomes were put into a training set. Every tool built a custom reference database from the training set. Genomes in the test set were used to respectively generate three sets of random sequence fragments of 150 bp, 300 bp and normally distributed 1–10 kb lengths as simulated positive reads. On the other hand, Chromosome 1 of the RefSeq reference genome of the Arabidopsis thaliana plant was used to respectively generate three sets of random sequence fragments of corresponding lengths as simulated negative reads. All tools classified these test reads based on the uniform database using the same hardware, then their performances were compared.