Detecting Evolution of Amino-Acid Fitness in Vertebrate Genomes

数百种脊椎动物的基因组测序现在已经接近完成，还有近10,000种将在未来几年内确定（“Genome10k”）。这项工作的主要动机是增加比较分析的能力，以阐明基因和基因组功能是如何进化的（以及它是如何进化的）。尽管在脊椎动物生物多样性的主要谱系中进行密集的分类抽样有望大大提高对进化过程中“重要”的遗传变化进行统计推断的能力，但仍存在大量的计算和分析障碍。该项目的主要目标是利用计算和建模创新来可靠地检测多达数百至数千种脊椎动物基因组中蛋白质编码区多个位置氨基酸适应度的协调变化（“适应度转移”）。这种转变是通过方向选择改变蛋白质功能需求的预期结果，因此可能意味着功能分化或适应。为了描述可靠检测的限制，将对不同实验设计（物种数量、分化水平等）下的统计信息进行详细计算，以确定比较数据如何很好地区分适应性变化与其他现象（例如，种群规模的减少）。快速马尔可夫链蒙特卡罗方法推断适应度转移在大型比较数据集将开发和评估。**科学方法**我们最近开发了几种通用方法，用于快速，大型系统基因组数据集的贝叶斯分析，这可以帮助消除计算瓶颈，并在某些情况下将数据分析时间从几个月减少到几分钟。在这个项目中，我们将整合这些技术，以及未发表的改进，以及利用在科学计算中日益重要的廉价多核协处理器的大规模并行性的算法。我们将使用这些方法来实现选择约束和种群大小的离散时空异质性模型，并将检查它们在大量脊椎动物单拷贝基因上的表现。在整个过程中，计算的可伸缩性和时间复杂性的降低（即使以明显温和的近似为代价）将被优先考虑，以最大化大型数据集的效用。我们最近开始开发的渐近功率分析方法（未发表）将被详细阐述并用于表征实验设计对功率的影响并评估推断的极限。随着脊椎动物基因组数量的不断增加，对包括功能分化和适应在内的基本进化过程的遗传基础的了解正在出现巨大的机会。检测功能差异的统计方法是最广泛使用的从基因组数据进行功能推断的方法之一。尽管人们普遍认为这种方法过于简化，而且在一些重要的方面存在缺陷，但由于计算方便，它们是可以容忍的。通过关注快速和可扩展推理的算法，希望分子进化的最新进展可以“放大”，以实现比较基因组学中更有原则的方法。通过实验设计的定量框架的发展，将出现更合理的方法来选择需要哪些物种和多少物种来解决一个特定的问题。因此，这里开发的工具将促进大规模比较基因组研究的合理设计和执行。

The genomes of hundreds of vertebrate species have now been sequenced to near-completion, and nearly 10,000 more are slated to be determined over the next several years ("Genome10k"). A major motivation for this work is to increase the power of comparative analysis to illuminate how gene and genome function evolves (and how it has evolved). Although dense taxonomic sampling across the major lineages of vertebrate biodiversity is expected to substantially increase the ability to make statistical inferences about the genetic changes that "mattered" during evolution, substantial computational and analytic barriers to progress exist.**Objectives**The primary objective of this project is to exploit computational and modeling innovations to reliably detect coordinated changes in amino-acid fitness ("fitness shifts") across multiple positions in the protein-coding regions of up to hundreds to thousands of vertebrate genomes. Such shifts are an expected outcome of changes in the functional requirements of a protein by directional selection, and thus may imply functional divergence or adaptation. To characterize the limits of reliable detection, detailed calculations of statistical information under alternative experimental designs (numbers of species, divergence levels, etc.) will be performed to determine how well comparative data can distinguish fitness shifts from other phenomena (e.g., reductions in population size). Fast Markov Chain Monte Carlo methods of inferring fitness shifts in large comparative datasets will be developed and evaluated. * *Scientific approach **We recently developed several general approaches for rapid, Bayesian analysis of large phylogenomic datasets, which can help eliminate computational bottlenecks and in some cases reduce data analysis times from months to minutes. In this project, we will integrate these techniques, along with unpublished improvements, with algorithms that exploit the massive parallelism of inexpensive, many-core coprocessors of emerging importance in scientific computing. We will use these approaches to implement models of discrete spatial and temporal heterogeneity in selective constraints and population size, and will examine their performance on a large set of vertebrate single-copy genes. Throughout, scalability of computations and reductions in time-complexity (even at the expense of demonstrably mild approximations) will be prioritized to maximize utility in large datasets. Asymptotic power analysis methods that we have recently begun developing (unpublished) will be elaborated and used to characterize the impact of experimental design on power and to evaluate the limits of inference.**Expected significance**With increasingly large numbers of vertebrate genomes now available, tremendous opportunities are emerging to advance knowledge of the genetic basis for fundamental evolutionary processes including functional divergence and adaptation. Statistical methods for detecting functional divergence are among the most widely used ways that functional inferences are made from genomic data. Although it is widely appreciated that such approaches are oversimplified and flawed in important ways, they are tolerated because of their computational convenience. By focusing here on algorithms for fast and scalable inference, it is hoped that recent progress in molecular evolution can be "scaled up" to enable more principled approaches in comparative genomics. Through the development of a quantitative framework for experimental design, more reasoned methods for selecting which and how many species are needed to address a particular question will emerge. Thus, the tools developed here will facilitate advances in both the rational design and execution of large-scale comparative genomic studies.