Computational Statistical Methods for Population Genomics

现在扫描一个种群中多个个体的整个基因组是可能的，而且相对便宜。所得数据可用于推断种群历史的各个方面，包括诸如种群增长和迁移率、重组率、选择系数以及混合水平等参数的值。贝叶斯统计范式为这种推断提供了一个很好的框架，因为它允许从指定模型下的数据中最大限度地提取信息，并且因为背景信息可以通过先验分布合并。尽管贝叶斯范式在原理上很简单，但在实践中几乎不可能精确应用，因为大型数据集和复杂模型意味着计算时间非常大。在过去的几年里，出现了一些令人兴奋的发展，以额外的近似为代价，将模型复杂性和可分析数据集大小的界限向后推（例如，参见Hey J， Machado CA, NATURE REVIEWS GENETICS, 4 (7): 535-543 JUL 2003）。在有充足数据的情况下，这种近似通常值得在更现实的模型中进行推断。两个这样的进步是：(a)似然的计算可以用模拟步骤代替，其中在给定当前参数设置的模型下模拟数据，如果模拟数据接近观测数据，这些步骤是可以接受的。(b)不计算全部的可能性，而是计算或近似一个类似的函数，但全部的数据用一个汇总统计的向量代替。基于这种方法的计算贝叶斯方法被称为ABC，即近似贝叶斯计算。申请人对这两方面的进步都做出了重大贡献，现在他们提议系统地研究如何使它们更有效地工作，并开发用户友好的计算机软件，使它们更广泛地提供给种群基因组学、保护遗传学和相关领域的研究工作者。这些任务将由帝国理工学院的博士后研究助理进行。与此同时，雷丁大学的一名博士生将致力于将帝国理工学院开发的新方法应用于人口基因组学的具体问题。结果将是，对于许多比以前可行的更复杂的情况，例如整个动物物种历史的详细方面，至少可以做出近似的推断。其他研究人员也会有这种新方法有用性的明确例子。我们将开发的方法是非常通用的，可以应用于任何使用复杂模型和大量数据的科学领域。虽然我们的项目侧重于人口基因组学，这似乎是最有成效的应用领域，但流行病学中的疾病传播模型是另一个可能从我们将开发的方法中受益的领域的例子。

It is now possible, and relatively cheap, to scan the entire genomes of multiple individuals within a population. The resulting data can be used to infer aspects of the history of a population, including the values of parameters such as population growth and migration rates, recombination rates, and selection coefficients, as well as levels of admixture. The Bayesian statistical paradigm offers a good framework for such inferences, because it allows maximal extraction of information from data under the specified model, and because background information can be incorporated via the prior distribution. Although straightforward in principle, exact application of the Bayesian paradigm is virtually impossible in practice in this setting because the large datasets and complex models mean that computation times are prohibitively large.In the past few years a number of exciting developments have arisen that push back the boundaries of the model complexity and dataset size that can be analysed, at the cost of an extra approximation (see for example Hey J, Machado CA, NATURE REVIEWS GENETICS, 4 (7): 535-543 JUL 2003). In the presence of ample data, this approximation is often worthwhile to achieve inferences in more realistic models than would otherwise be possible. Two such advances are:(a) Computation of the likelihood may be replaced by a simulation step in which data are simulated under the model given the current parameter settings, and these are accepted if the simulated data are close to the observed data.(b) Instead of the full likelihood, an analogous function is calculated or approximated but with the full data replaced by a vector of summary statistics. Computational Bayesian methods based on this approach have come to be known as ABC, Approximate Bayesian Computation.The applicants have contributed substantially to both these advances, and now propose to investigate systematically ways to make them work more efficiently, and to develop user-friendly computer software to make them more widely available to research workers in population genomics, conservation genetics, and related fields. These tasks will be pursued by a post-doctoral research associate at Imperial College. At the same time, a PhD student at Reading will work on applications of the new methods developed at Imperial to specific problems in population genomics. The result will be that at least approximate inferences will be possible for many more complex situations than was previously feasible, for example detailed aspects of the history of entire animal species. Other researchers will also have explicit examples of the usefulness of this new methodology.The methods we will be developing are very general, and can be applied in any area of science that uses complex models and large amounts of data. Although our project focusses on population genomics, which seems the most fruitful area for application, disease transmission models in epidemiology is an example of another field that is likely to benefit from the methods that we will develop.