Advanced Stochastic Computation for Inference from Tree, Graph and Network Models

由于最近的实验进展，人类和其他生物正在获得大量的生物数据。这些数据带来的推断问题远远超出了标准统计工具的能力。在取得实验进展的同时，必须开发适当的生物统计学和生物信息学工具，这些工具将有效地利用复杂的数据，以提高我们对塑造基因组组织进化的遗传力的理解。分子进化和比较基因组学不再是收集数据是进步的主要障碍的领域。生物学中提出的许多概率模型都试图捕捉进化机制，反映数据生成机制。虽然这些模式仍然有局限性，但近年来已经有了很大的改进。我们认为，进展主要受到缺乏足够的计算工具的限制，无法从现有数据中提取信息并对复杂模型进行推理。概率模型在生物学中应用的一个主要瓶颈是，它们的校准计算昂贵，在许多情况下不可能使用现代技术。因此，研究人员往往倾向于使用简单的汇总统计数据来描述潜在的生物过程；这种方法显然不能令人满意。例如，拓扑汇总统计数据捕获了二元交互网络的基本特征，但受到不同类型的偏差的影响如此之大，以至于在得出结论时必须谨慎。当今系统生物学的一大挑战在于在概率建模的严格框架内开发统计和生物信息学工具，以便更好和更全面地了解细胞功能。在过去的几十年里，人们对分子数据的基于模型的推理进行了大量的研究，同时也出现了大量的研究，以开发高效的计算方法来促进分子数据的推理。一般而言，分子生物学中的统计推断主要有三种方法：(1)用于似然估计的重要性抽样(IS)；(2)马尔可夫链蒙特卡罗(MCMC)方法；(3)近似贝叶斯计算(ABC)。在这项建议中，我们集中在高级IS(或更广泛地说，顺序蒙特卡罗(SMC))和MCMC方法的组合，专注于遗传学和生物信息学中的马尔可夫模型。特别是，在这些技术的基础上，我们的目标是开发一个理论上合理的、在计算上可行的近似推理的通用框架，并且仍然能够准确地反映潜在随机模型的复杂性。我们将专注于的主要生命科学应用是：系谱树、蛋白质网络、系统发育树。

As a result of recent experimental advances, large amounts of biological data are becoming available for humans and other organisms. Such data pose inference problems well beyond the capabilities of standard statistical tools. Experimental advances must be accompanied by the development of suitable biostatistical and bioinformatic tools that will make efficient use of the complex data to improve our understanding of the genetic forces shaping the evolution of genome organization. Molecular evolution and comparative genomics are no longer fields where collecting data is the main obstacle to progress. Many probabilistic models proposed in biology try to capture the evolutionary mechanisms and reflect the data generating mechanism. Whilst these models still have limitations, there have been substantial improvements in recent years. We contend that progress is mainly limited by a lack of adequate computational tools for extracting information from existing data and performing inference for complex models. A major bottleneck in the application of probabilistic models to biology is that their calibration is computationally expensive and in many instances not possible using modern techniques. Thus, researchers often prefer to use simple summary statistics to characterize the underlying biological process; this approach is obviously unsatisfactory. E.g., topological summary statistics capture basic characteristics of binary interaction networks but are affected by different types of bias so great, that caution must be taken when drawing conclusions. A big challenge for systems biology nowadays consists in developing statistical and bioinformatics tools within the rigorous framework of probabilistic modelling that will allow for a better and more comprehensive understanding of cellular functions. In the last few decades a wealth of research has been performed on model-based inference for molecular data accompanied by an explosion of research in developing computationally efficient methods to facilitate it. Broadly speaking, there are three main approaches to statistical inference in molecular biology: (i) importance sampling (IS) for likelihood evaluation, (ii) Markov chain Monte Carlo (MCMC) methods (iii) Approximate Bayesian Computation (ABC) . In this proposal we concentrate on a combination of advanced IS (or more generally Sequential Monte Carlo (SMC)) and MCMC methods focussed upon Markov models in genetics and bioinformatics. In particular, building upon these techniques, we aim to develop a general framework for approximate inference which is theoretically sound, computationally feasable and still be able to accurately reflect the complexity of the underlying stochastic model. The main life science application on which we will concentrate are: genealogical trees, protein networks, phylogenetic trees.