A new phylogenetic protocol: Dealing with model misspecification and confirmation bias in molecular phylogenetics

Lars S Jermiin; Renee A Catullo; Barbara R Holland

Although the current phylogenetic protocol has many shortcomings, it also has many good attributes, including that it is easy to apply and implement as a pipeline. But to mitigate its limitations, it will be necessary to redesign the protocol to accommodate well-established, but largely-ignored, procedures as well as new feedback loops. Figure 2 shows a proposal for new phylogenetic protocol. It shares many features found in the current protocol (e.g., the first four steps). However, the fifth step (assess phylogenetic assumptions) will be novel to many researchers. As all phylogenetic methods are based on assumptions, it is sensible to validate these assumptions at this point in the protocol. Since many phylogenetic methods assume that the data (e.g., different genes) have evolved over the same tree, and that the chosen data partitions have evolved independently under the same time-reversible Markovian conditions, it is wise to survey the sub-MSA for evidence that the sequences actually have evolved under these conditions. If the data violate these phylogenetic assumptions, then it will be wise to avoid these phylogenetic methods and to employ other such methods. Alternatively, it may be worth following the relevant feedback loops in Figure 2 -perhaps something led to a biased sub-MSA? The relevance and benefits of this step are illustrated using a case study (Box 1), which focuses on determining whether a data set is consistent with the phylogenetic assumption of evolution under time-reversible conditions. Assessments of other phylogenetic assumptions require other types of tests and Next follows the choice of phylogenetic method, but now this choice is made on the basis of the previous step, rather than on cultural or computational reasons. If the sequences have evolved on a single tree under time-reversible Markovian conditions, there is a large set of phylogenetic methods to choose from (66-85). On the other hand, if these data have evolved under more complex Markovian conditions, the number of suitable phylogenetic methods is, frustratingly, rather limited (5,64,146-171), and most of these methods are aimed at finding the optimal model of sequence evolution for a given tree rather than finding the optimal set of trees. Users of phylogenetic methods therefore are sometimes confronted by a dilemma: