Fragmentary gene sequences negatively impact gene tree and species tree reconstruction
收藏资源简介：
Analysis Pipeline for Insect Data Set We used the amino acid sequence data provided by Misof et al. (2014) as “supplementary 7.” Filtering Strategy Before identifying fragmentary sequences, we first remove extremely gappy sites, defined as those with >90% gaps. We then remove species that have <20% (1/5), 25% (1/4), 33% (1/3), 50% (1/2), 66% (2/3), 75% (3/4), or 80% (4/5) amino-acids (i.e., characters other than gaps). We use a tool called seqtools, implemented as part of the PASTA. Next, we reestimate gene trees. In order to track the occupancy and bootstrap support, we use in-house scripts, available online https://github.com/esayyari/discoVista. Gene Trees and Species Trees Gene trees are estimated using FastTree2 (Price et al. 2010) using its default amino acid substitution model JTT (Jones et al. 1992) or RAxML (Stamatakis 2014) with the automatic amino acid model selection. We use RAxML (Stamatakis 2014), version 8.2.9 with ten runs of inference using different starting trees. We used RAxML’s automatic model selection approach. When several species have identical sequences for a gene, we keep only one of them (i.e., remove redundant ones) in our RAxML runs and add the removed species back to the final inferred gene tree as a polytomy. For performing gene tree bootstrapping using FastTree, we first generate bootstrap sequences using RAxML and then run FastTree on those to estimate the bootstrapped gene trees. We then draw those bootstrap gene trees on ML gene tree branches using the Newick utility (Junier and Zdobnov 2010). For RAxML gene trees, we use the rapid bootstrapping option on reduced sequences (after removing identical sequences). After gene tree estimations, we add back the identical species and draw these bootstrap gene trees on the best ML gene trees (RAxML) following the same procedure using the Newick utility. We use ASTRAL-II to estimate the species trees summarizing gene trees with at least four taxa left after filtering. Simulation Procedure We use one model condition of a previously simulated data set from Mirarab and Warnow (2015) ASTRAL-II paper with 100 ingroup taxa and one outgroup. For each of the 50 replicates in this data set, Simphy (Mallo et al. 2016) was used to simulate a species tree according to the Yule model, and then 1,000 gene trees were simulated using the MSC model which captures ILS. The data set has moderate levels of ILS; the average distance between true gene trees and true species trees is 0.33. We subsampled genes to create three different data sets with 50, 200, or 1,000 genes. DNA sequences of varying lengths were simulated down the gene trees using Indelible (Fletcher and Yang 2009) with GTR parameters and stationary distributions estimated from published biological data sets, as detailed by Mirarab and Warnow (2015). Note that simulated sequences did not include any indels and thus were already aligned. Mirarab and Warnow (2015) suggested removing two replicates that include almost no phylogenetic signal, and we use the same strategy, leaving us with 48 replicates. This creates our unfiltered base data set. We add fragmentation to our complete simulated data set using a procedure that seeks to emulate patterns of fragmentation in the insect biological data set. 1) For each replicate, we order species in the biological data set and the simulated data set with respect to the tip-to-root distances. 2) We randomly select 100 of the biological species and map them to the simulated species with the same position in the order. The main outgroup (Ixodes scapularis) in the biological and simulated data sets always map to each other. 3) For each replicate in the simulated data set, we randomly sample (with replacement) 1,000 genes in the insect data sets that have at least 101 species, including the main outgroup. 4) For each species in each simulated gene, we compute the portion of gap sites in the corresponding gene alignment for the corresponding species in the biological data and remove the same portion of sites in the simulated data set at random positions. When a species is missing from a gene in the biological data set, we use the same species from another randomly chosen gene. Filtering Fragments and Gene Trees and Species Trees We first remove sites with >90% gaps, removing between 0.0% and 2.0% of the total number of characters in all sequences. We then remove from each gene any species that has less than a certain fraction (e.g., 10–80%) of the full gene. For example, at 10%, we remove only sequences that have 90% or more gaps. For each threshold, we then estimate gene trees using both RAxML (Stamatakis 2014) version 8.2.9 with two starting trees and FastTree (Price et al. 2010) version 2.1.9 Double precision using the GTR + Γ model of sequence evolution (Tavaré 1986). We infer the species tree using ASTRAL-II (Mirarab and Warnow 2015) version 4.11.1. We build species trees using all 1,000 genes or using randomly chosen subsets of 200 or 50 genes.
昆虫数据集分析流程 我们采用Misof等（2014）发表的作为“补充材料7”的氨基酸序列数据。 过滤策略 在识别碎片化序列前，我们首先移除极度缺失位点，即缺失率超过90%的位点。随后移除氨基酸（即非缺失字符）占比低于20%（1/5）、25%（1/4）、33%（1/3）、50%（1/2）、66%（2/3）、75%（3/4）或80%（4/5）的物种。我们使用集成于PASTA中的seqtools工具完成上述操作。随后，我们重新估计基因树。为追踪位点占据率与自展支持率，我们使用公开在线的内部脚本discoVista，其托管地址为https://github.com/esayyari/discoVista。 基因树与物种树 基因树的估计采用FastTree2（Price等，2010），使用其默认的JTT氨基酸替换模型（Jones等，1992）；或采用RAxML（Stamatakis，2014）并启用自动氨基酸模型选择功能。本次研究使用RAxML 8.2.9版本，采用10次不同起始树的推断，并启用其自动模型选择方法。当多个物种在某一基因上的序列完全一致时，我们在RAxML推断过程中仅保留其中一条序列（即移除冗余序列），并将移除的物种作为多歧支添加至最终推断的基因树中。 针对使用FastTree的基因树自展分析，我们首先通过RAxML生成自展序列，随后使用FastTree对这些序列进行推断以得到自展基因树，再通过Newick工具（Newick utility，Junier与Zdobnov，2010）将自展基因树绘制于最大似然（ML）基因树的分支上。针对RAxML基因树，我们对去重后的序列使用快速自展选项。完成基因树推断后，我们将重复物种重新添加回数据集，并通过相同的Newick工具流程，将自展基因树绘制于最优ML基因树（RAxML推断所得）之上。我们使用ASTRAL-II来估计总结基因树的物种树，该分析仅保留过滤后至少剩余4个类群的基因树。 模拟流程 我们采用Mirarab与Warnow（2015）ASTRAL-II论文中已发表的模拟数据集的一组模型条件，该数据集包含100个内群类群与1个外群。该数据集包含50次重复，每次重复均通过Simphy（Mallo等，2016）依据Yule模型模拟物种树，随后采用MSC模型（可涵盖不完全谱系分选，ILS）模拟1000个基因树。该数据集的ILS水平中等，真实基因树与真实物种树间的平均距离为0.33。我们对基因进行抽样子集，构建得到3组不同的数据集，分别包含50、200或1000个基因。依据基因树，使用Indelible（Fletcher与Yang，2009）模拟不同长度的DNA序列，其GTR参数与平稳分布均由已发表的生物数据集估计得到，具体细节参见Mirarab与Warnow（2015）。需注意，模拟序列未包含任何插入缺失（indel），因此已完成序列比对。 Mirarab与Warnow（2015）建议移除两个几乎不含系统发育信号的重复，我们沿用该策略，最终得到48次重复，以此构建未过滤的基础数据集。我们通过特定流程为完整的模拟数据集添加碎片化特征，以模拟昆虫生物数据集的碎片化模式：1） 针对每次重复，我们根据端到根距离对生物数据集与模拟数据集中的物种进行排序；2） 我们随机选取生物数据集中的100个物种，并将其映射至排序位置一致的模拟物种；生物数据集与模拟数据集中的主要外群（Ixodes scapularis，硬蜱）始终相互映射；3） 针对模拟数据集的每次重复，我们从昆虫数据集中随机有放回地抽样1000个基因，这些基因需至少包含101个物种（包括主要外群）；4） 针对模拟基因中的每个物种，我们计算其在对应生物数据集基因比对中的缺失位点占比，并在模拟数据集的对应基因中随机位置移除相同比例的位点。当某一物种在生物数据集的某一基因中缺失时，我们从另一随机选取的基因中选取该物种进行替代。 片段过滤与基因树、物种树推断 我们首先移除缺失率超过90%的位点，该操作会移除所有序列总字符数的0.0%至2.0%。随后，我们针对每个基因移除那些在完整基因中占比低于特定阈值（例如10%至80%）的物种。例如，当阈值为10%时，我们仅移除缺失率达90%及以上的序列。针对每个阈值，我们均采用RAxML（Stamatakis，2014）8.2.9版本（使用2个起始树）与FastTree（Price等，2010）2.1.9双精度版本进行基因树推断，其中FastTree采用GTR+Γ序列进化模型（Tavaré，1986）。我们采用ASTRAL-II（Mirarab与Warnow，2015）4.11.1版本推断物种树，构建物种树时可采用全部1000个基因，或随机选取的200、50个基因的子集。
