We present a novel identification tool called PhyloKey, based on the method of morphology-based, phylogenetic binning developed within the software package RAxML. This method takes a reference data set of species for which both molecular and morphological data are available, computes a molecular reference tree, maps the morphological characters on the tree, and computes weights based on their level of consistency versus homoplasy using maximum likelihood (ML) and maximum parsimony (MP). Additional units for which only morphological data are known are then binned onto the reference tree, calculating bootstrap support values for alternative placements. This approach is modified here to work as an identification tool which uses the same character coding approach as interactive keys. However, rather than identifying individual samples through a progressive filtering process when entering or selecting characters, query samples are binned in batch mode to all possible alternative species in the tree, with each placement receiving a bootstrap support adding to 100% for all alternative placements. In addition to the fact that, after scoring a character matrix, a large number of specimens can be identified at once in short time, all possible alternative identifications are immediately apparent and can be evaluated based on their bootstrap support values. We illustrate this approach using the basidiolichen genus Cora, which was recently shown to contain hundreds of species. We also demonstrate how the PhyloKey approach can aid the restudying of herbarium samples, adding further value to these collections and contributing with large quantitative data matrices to ‘non-molecular museomics’. Our analysis showed that PhyloKey identifies species correctly with as low as 50% of the characters sampled, depending on the nature of the reference tree and the character weighting scheme. Overall, a molecular reference tree worked best, but a randomized reference tree gave more consistent results, whereas a morphological reference tree performed less well. Surprisingly, even character weighting gave the best results, followed by parsimony weighting and then maximum likelihood weighting.
