Leave-one-out validation: examples. Individual query genomes were fragmented (1'000 bp fragments) and then placed into reference trees from which the corresponding genomes (or entire clades) had been removed. The assignments are shown graphically (small circles). Note how the placements become increasingly scattered and imprecise upon removal of increasingly deep reference information. MLTreeMap is shown compared to two popular approaches (note that MEGAN, while the least accurate, applies to a much larger fraction of reads in a given sample and thus achieves the best coverage). Definitions of test success: *assignments are designated as correct when they are no more than two nodes away from the target position in the tree. **for MEGAN, assignments are designated as correct when they are mapping to the target phylum.
Stark et al. BMC Genomics 2010 11:461 doi:10.1186/1471-2164-11-461