Efficient Bayesian phylogenomic dating with new models of trait evolution and rich diversities of living and fossil species

随着物种的分化，它们在基因组中以几乎恒定的速度积累核苷酸替换。因此，替换可以作为推断物种分化的计时器。通过结合化石记录中的信息，可以将推断的物种形成时间校准到地质时间。这种被称为分子钟测年的方法在进化生物学中有着广泛的应用，例如研究病毒大流行的传播时间、动植物的古代多样化速度、物种进化与过去的气候或灭绝事件的关系、人类进化，或农业和动物驯化的起源。事实上，进化时间表比没有时间信息的树提供了关于物种历史的更丰富的信息，从而允许关于进化时间尺度的假设的制定和测试。目前，贝叶斯方法是分子时钟测年中最先进的方法，因为它们允许对进化过程进行灵活的建模，并在分析中整合化石的不确定性。贝叶斯时钟测年的进展包括谱系间速率变化的随机模型(所谓的松弛时钟模型)，现存和灭绝分类群特征进化的模型，以及“软界”和灵活的化石校准密度的发展。虽然这些进展使得贝叶斯方法对时钟日期测定很有吸引力，但贝叶斯计算依赖于MCMC抽样，这需要计算昂贵的随机模拟，排除了贝叶斯方法用于大规模数据集的分析。这是不幸的，因为大规模的分子数据集现在是司空见惯的：几个高通量基因组测序项目现在已经宣布或正在进行中，我们预计会有数千个物种的基因组规模数据(例如10K动物基因组和英国的66K真核基因组项目)。在基因组数据泛滥的同时，基于比较解剖学计算革命的形态数据集的数量也出现了爆炸性的增长--X射线断层扫描和摄影测量的广泛应用导致了巨大的特征数据数据库：MorPhoBank和Phenome10K现在存储了超过7,000个物种的64,200多个表面扫描。现在迫切需要能够利用这些新生成的数据集的计算工具。例如，用目前的方法，推断一个66K物种的时间表将需要至少55年的计算时间(从我们之前的一些分析推断)。显然，分析方法的效率没有跟上可用的数据量，越来越需要解决进化生物学中的大规模问题。在这个项目中，我们将克服物种分歧贝叶斯时钟测年中的两个主要挑战：(I)MCMC算法在分析大数据集时的混合和计算限制，以及(Ii)当前进化特征模型在时间表推断方面的限制。我们将利用MCMC算法设计的新思想来设计新的MCMC算法来提高混合效率，并通过代码改进和并行化来提高计算效率。我们将结合先进的特征模型来推断现存物种和化石物种的时间表。特别是，我们将采用特征模型来分析大型基因组特征数据集，如RNA-seq表达数据。新开发的算法将在我们的MCMCtree软件中实现，并应用于几个具有密集采样的现存物种和化石物种的大规模经验数据集。数据分析将为方法开发提供重要的动机，并通过解决进化生物学中的基本问题来展示我们的新软件。我们的建议涉及BBSRC的战略优先事项，即“数据驱动生物学”和“生物科学的系统方法”。

As species diverge, they accumulate nucleotide substitutions in their genomes at a rate approximately constant in time. Thus, substitutions serve as timepieces to infer species divergences. By incorporating information from the fossil record, the inferred speciation timings can be calibrated to geological time. This method, known as molecular-clock dating, has broad applications in evolutionary biology, such as studying the timing of spread of viral pandemics, ancient rates of diversification in animals and plants, the relationship of species evolution with past climate or extinction events, human evolution, or the origin of agriculture and animal domestication. Indeed, evolutionary timetrees provide much richer information about species histories than trees without temporal information, thus allowing the formulation and testing of hypotheses on evolutionary timescales.Currently, Bayesian methods are the-state-of-the-art in molecular-clock dating as they allow flexible modelling of evolutionary processes and integration of fossil uncertainties in the analysis. Progresses in Bayesian clock-dating include stochastic models of rate variation among lineages (so-called relaxed clock models), modelling of trait evolution in extant and extinct taxa, and development of "soft-bounds" and flexible fossil calibration densities. While these advances have made the Bayesian method attractive for clock-dating, Bayesian computation relies on MCMC sampling which requires computationally expensive stochastic simulation, precluding the Bayesian method for analysis of large-scale datasets. This is unfortunate since large scale molecular datasets are now commonplace: several high-throughput genome sequencing projects have now been announced or are in progress and we expect a flood of genome-scale data for several thousand species (e.g. the 10K animal genomes and the UK's 66K eukaryotic genomes projects). This deluge of genome data has been accompanied by an explosive increase in the number of morphological datasets based on a computational revolution in comparative anatomy - the widespread deployment of X-Ray Tomography and photogrammetry resulting in vast databases of trait data: MorphoBank and Phenome10K now store over 64,200 surface scans for over 7,000 species. Computational tools capable of exploiting these newly generated datasets are now urgently required. For example, with current methods, inference of a 66K-species timetree would require at least 55 years of computing time (extrapolating from some of our previous analyses). Evidently, the efficiency of analytic methods has not kept apace with the volume of data available and increasingly required to tackle large scale questions in evolutionary biology. In this project we will overcome two major challenges in Bayesian clock dating of species divergences: (i) the mixing and computational limitations of MCMC algorithms in analyses of large datasets, and (ii) the limitations of current trait models of evolution in timetree inference. We will design novel MCMC algorithms to improve the mixing efficiency making use of new ideas about MCMC algorithm design and improve the computational efficiency through code improvement and parallelization. We will incorporate advanced trait models to infer timetrees of extant and fossil species. In particular, we will adapt trait models to analyse large genomic trait datasets such as RNA-seq expression data. The newly developed algorithms will be implemented in our MCMCtree software, and applied to several large-scale empirical datasets with densely sampled extant and fossil species. The data analyses will provide important motivations for method development and serve to showcase our new software by addressing fundamental questions in evolutionary biology. Our proposal addresses the BBSRC's strategic priorities of "data driven biology" and "system approaches to the biosciences".