Probabilities of Unranked and Ranked Anomaly Zones under Birth-Death Models

How to Validate a Bayesian Evolutionary Model

Biology has become a highly mathematical discipline in which probabilistic models play a central role. As a result, research in the biological sciences is now dependent on computational tools capable of carrying out complex analyses. These tools must be validated before they can be used, but what is understood as validation varies widely among methodological contributions. This may be a consequence of the still embryonic stage of the literature on statistical software validation for computational biology. Our manuscript aims to advance this literature. Here, we describe, illustrate, and introduce new good practices for assessing the correctness of a model implementation with an emphasis on Bayesian methods. We also introduce a suite of functionalities for automating validation protocols. It is our hope that the guidelines presented here help sharpen the focus of discussions on (as well as elevate) expected standards of statistical software for biology.

Theoretical and practical considerations when using retroelement insertions to estimate species trees in the anomaly zone

A potential shortcoming of concatenation methods for species tree estimation is their failure to account for incomplete lineage sorting. Coalescent methods address this problem but make various assumptions that, if violated, can result in worse performance than concatenation. Given the challenges of analyzing DNA sequences with both concatenation and coalescent methods, retroelement insertions (RIs) have emerged as powerful phylogenomic markers for species tree estimation. Here, we show that two recently proposed quartet-based methods, SDPquartets and ASTRAL_BP, are statistically consistent estimators of the unrooted species tree topology under the coalescent when RIs follow a neutral infinite-sites model of mutation and the expected number of new RIs per generation is constant across the species tree. The accuracy of these (and other) methods for inferring species trees from RIs has yet to be assessed on simulated data sets, where the true species tree topology is known. Therefore, we evaluated eight methods given RIs simulated from four model species trees, all of which have short branches and at least three of which are in the anomaly zone. In our simulation study, ASTRAL_BP and SDPquartets always recovered the correct species tree topology when given a sufficiently large number of RIs, as predicted. A distance-based method (ASTRID_BP) and Dollo parsimony also performed well in recovering the species tree topology. In contrast, unordered, polymorphism, and Camin-Sokal parsimony typically fail to recover the correct species tree topology in anomaly zone situations with more than four ingroup taxa. Of the methods studied, only ASTRAL_BP automatically estimates internal branch lengths (in coalescent units) and support values (i.e. local posterior probabilities). We examined the accuracy of branch length estimation, finding that estimated lengths were accurate for short branches but upwardly biased otherwise. This led us to derive the maximum likelihood (branch length) estimate for when RIs are given as input instead of binary gene trees; this corrected formula produced accurate estimates of branch lengths in our simulation study, provided that a sufficiently large number of RIs were given as input. Lastly, we evaluated the impact of data quantity on species tree estimation by repeating the above experiments with input sizes varying from 100 to 100 000 parsimony-informative RIs. We found that, when given just 1 000 parsimony-informative RIs as input, ASTRAL_BP successfully reconstructed major clades (i.e clades separated by branches > 0.3 CUs) with high support and identified rapid radiations (i.e. shorter connected branches), although not their precise branching order. The local posterior probability was effective for controlling false positive branches in these scenarios.

Heuristics for unrooted, unranked, and ranked anomaly zones under birth-death models

Species trees that can generate a nonmatching gene tree topology that is more probable than the topology matching the species tree are said to be in an anomaly zone. We introduce some heuristic approaches to infer whether species trees are in anomaly zones when it is difficult or impossible to compute the entire distribution of gene tree topologies. Here, probabilities of unrooted, unranked, and ranked gene tree topologies under the multispecies coalescent are used. A ranked tree can be viewed as an unranked tree with a temporal ordering of its internal nodes. Overall, considering probabilities of unrooted or unranked gene tree topologies within one nearest neighbor interchange from the species tree topology is a reasonable heuristic to infer the existence of anomalous unrooted or unranked gene trees, respectively. We investigated a test proposed by Linkem et al. (2016) which classifies a species tree as being in an unranked anomaly zone if there is a subset of four taxa in an unranked anomaly zone. We find this test to have high true positive rates, but it can also have high false positive rates. For ranked trees, because at least one of the most probable ranked gene tree topologies must have the same unranked topology as the species tree, we propose to use only those ranked gene trees that have topologies that match the unranked species tree topology. We find that the probability that the species tree is in unrooted and unranked anomaly zones tends to increase with the speciation rate, and the probability of all three types of anomaly zones increases rapidly with the number of taxa. We find that probabilities that species trees are in an anomaly zone can be quite high for moderately high speciation rates.

PRANC: ML species tree estimation from the ranked gene trees under coalescence

SUMMARY

PRANC computes the Probabilities of RANked gene tree topologies under the multispecies coalescent. A ranked gene tree is a gene tree accounting for the temporal ordering of internal nodes. PRANC can also estimate the maximum likelihood (ML) species tree from a sample of ranked or unranked gene tree topologies. It estimates the ML tree with estimated branch lengths in coalescent units.

AVAILABILITY AND IMPLEMENTATION

PRANC is written in C++ and freely available at github.com/anastasiiakim/PRANC.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.