Significant phylogenetic signal is not enough to trust phylogenetic predictions

Abstract

In a recent study, Cantwell-Jones et al. (2022) proposed a list of 1044 species as promising key sources of B vitamins based primarily on phylogenetic predictions. To identify candidate plants, they fitted lambda models of evolution to edible species with known values in each of six B vitamins (232 to 280 species) and used the estimated parameters to predict B-vitamin profiles of edible plants lacking nutritional data (6460 to 6508 species). The latter species were defined as potential sources of a B vitamin if the predicted vitamin content was [≥]15% towards recommended dietary allowances for active females between 31-50 years per 100 g of fresh edible plant material consumed. Unfortunately, the reliability of the predictions that informed the list of candidate species is questionable due to insufficient phylogenetic signal in the data (Pagels {lambda} between 0.171 and 0.665) and a high incidence of species with missing values (over 95% of all the species analyzed in the study). We found that of the 1044 species proposed as promising B-vitamin sources, 626 to 993 species showed accuracies that were indistinguishable from those obtained under a white noise model of evolution (i.e. random predictions conducted in absence of any phylogenetic structure) in at least one of the vitamins, which proves the weakness of the inference drawn from imputed information in the original study. We hope this commentary serves as a cautionary note for future phylogenetic imputation exercises to carefully assess whether the data meet the requirements for the predictions to be valuable, or at least more accurate than expected by chance.