Our results demonstrate that the optimal strategy for obtaining accurate classification of sequencing reads from mixed-species samples requires three steps: first, alignment to a mixed reference genome, then, the selection of the optimal alignment for each read to partition reads into species-specific FASTQ files, followed by re-alignment to the appropriate genome. Aligning mixed samples using a pooled reference index resulted in errors, some of which underestimated counts unequally across genes. Surprisingly, an algorithm designed to focus on differences within a reference sequence (Kallisto) was less effective at distinguishing reads by species. An algorithm using indexed genomes (HISAT2) performed better, with better overall accuracy and reduced error, particularly when coupled with a method to separate input FASTQ files using alignment information. With pooled mouse and human RNAseq reads with various proportions, both byAS and byPrim methods offered over 95% accuracies.
