Phylogeny Estimation Given Sequence Length Heterogeneity

Abstract

Phylogeny estimation is a major step in many biological studies, and has many well known challenges. With the dropping cost of sequencing technologies, biologists now have increasingly large datasets available for use in phylogeny estimation. Here we address the challenge of estimating a tree given large datasets with a combination of full-length sequences and fragmentary sequences, which can arise due to a variety of reasons, including sample collection, sequencing technologies, and analytical pipelines. We compare two basic approaches: (1) computing an alignment on the full dataset and then computing a maximum likelihood tree on the alignment, or (2) constructing an alignment and tree on the full length sequences and then using phylogenetic placement to add the remaining sequences (which will generally be fragmentary) into the tree. We explore these two approaches on a range of simulated datasets, each with 1000 sequences and varying in rates of evolution, and two biological datasets. Our study shows some striking performance differences between methods, especially when there is substantial sequence length heterogeneity and high rates of evolution. We find in particular that using UPP to align sequences and RAxML to compute a tree on the alignment provides the best accuracy, substantially outperforming trees computed using phylogenetic placement methods. We also find that FastTree has poor accuracy on alignments containing fragmentary sequences. Overall, our study provides insights into the literature comparing different methods and pipelines for phylogenetic estimation, and suggests directions for future method development.

MAGUS+eHMMs: improved multiple sequence alignment accuracy for fragmentary sequences

“…Percent gapped is the percentage of the alignment matrix occupied by dashes. 1000M1 has replicate 16 removed due to being identified as an outlier in Smirnov and Warnow (2020b ). Statistics regarding fragments are computed after introducing fragments to the datasets, while the rest are computed prior to fragmentation.…”

“…Each condition has 20 replicates and each replicate contains 1000 sequences with ∼ 1000 nucleotides. We explore fragmentary versions of the 1000M1 and 1000M2 conditions (also studied by Smirnov and Warnow, 2020b ), which have high and moderately high rates of evolution, respectively, and medium indel lengths. The RNASim ( Mirarab et al , 2015 ) datasets evolve under a complex evolutionary process that reflects selective pressures needed to conserve rRNA structure; hence, this simulation condition is more complex than the standard GTR+indel simulations (such as for the ROSE datasets).…”

Accuracy in near-perfect virus phylogenies

“…Let N F P be the number of splits onT but not T (false positives), and let N F N be the number of splits on T but notT (false negatives). When both trees are binary, N F P = N F N (Berry and Gascuel, 1996;Smirnov and Warnow, 2021); otherwise they can contribute differentially to error. The Robinson-Foulds (RF) distance (Robinson and Foulds, 1981), d RF = N F P + N F N , combines both errors in one measure of overall accuracy.…”

“…The Robinson-Foulds (RF) distance (Robinson and Foulds, 1981), d RF = N F P + N F N , combines both errors in one measure of overall accuracy. Here we distinguish between these errors explicitly by defining false positive and negative rates (Smirnov and Warnow, 2021):…”

TAPER: Pinpointing errors in multiple sequence alignments despite varying rates of evolution

“…Most commonly used methods further assume orthology, and errors in orthology detection are common (Laurin-Lemay et al, 2012;Salichos & Rokas, 2011). Alignment errors are also ubiquitous and can impact tree accuracy (Fletcher & Yang, 2010;Liu et al, 2009;Ogdenw & Rosenberg, 2006;Smirnov & Warnow, 2020;Wang et al, 2011). The prevalence of these errors in phylogenomic datasets has been appreciated (Hosner et al, 2016;Laurin-Lemay et al, 2012;Philippe et al, 2017;Sayyari et al, 2017;Springer & Gatesy, 2016, and several phylogenomics studies have now been criticized (Gatesy & Springer, 2014;Jeffroy et al, 2006;Salichos & Rokas, 2013;Shen et al, 2017;Springer & Gatesy, 2016.…”