Computational modelling of the relationships between miscanthus genotype environment and phenotype

IBERS已经并正在生成大量基于基因型、表型和环境的数据，以及来自能源作物芒草的相关元数据。被任命为该项目的学生将在这些数据集的基础上，创建有关数据类型之间预测关系的新科学知识。这些知识将被编码到计算模型中。为了开发模型，学生将学习和应用最先进的数据挖掘和机器学习方法来识别和表征重要数量性状（QTL）的基因。这将有助于更好地了解芒草生物学，并通过定义芒草的理想型来指导芒草的育种。驯化一种新作物作为生物能源原料，需要广泛的技术应用，包括遗传学、生物化学、生理学、化学工程、生物信息学和系统生物学建模。在过去的五年里，IBERS建立了一个芒草研究平台，包括248种不同基因型的重复性状试验。对这一种群及其杂交后代进行了四年的密集表型分析，产生了超过10万个数据集。这些已被纳入一个强大的自定义编码集成芒草信息平台。示例描述包括：生长环境、形态、结构、开花时间、衰老、产量、春季出苗、倍性、遗传群、木质素、纤维素、半纤维素和细胞壁酚浓度。如果我们要最大限度地发挥数据的价值，从而加快作物改良的步伐，那么整合这些信息并在预测模型中加以利用是非常重要的。该学生将结合数据挖掘和机器学习来分析数据集，以确定重要的关联，并形成将芒草表型、基因型和环境联系起来的预测规则。这将使用最近开发的数据挖掘和机器学习技术，这些技术非常适合数据结构，以创建关联模型。这将与IBERS和Ceres的科学家密切互动，他们将提供关于模式和关联规则的反馈。这个过程将是一个互动的过程。通过数据挖掘识别的“有趣”模式（例如QTL）将提供对芒草生物学的深入了解，并将作为新的描述性属性输入机器学习系统，以改进建模。基因型/环境/表型数据的分析在技术上具有挑战性，因为有大量的描述符，其中许多是最好的关系表示。这意味着标准的统计方法可能不是最优的。Miscanthus数据适合于关系描述的典型方面有：时间关系；田野与植物间的空间关系；基因型间的家系关系等。许多统计遗传学，以及诸如空间统计学等领域，都致力于为关系问题开发特殊的解决方案，这些解决方案现在可以用原则的方式来解决。数据挖掘和机器学习方法已被用于利用人类基因型和环境数据开发预测表型模型；但据我们所知，在植物遗传学上还没有应用。为了评估和比较这些方法与标准统计遗传方法的性能，我们将使用标准的重新抽样方法，包括交叉验证和自举。这个提议的项目非常适合博士研究项目，因为它结合了后备元素（统计遗传学）和开放式元素（关系学习）。该提案有很强的培训元素，因为学生将获得遗传学、生物能源、数据挖掘和机器学习方面的经验，这些技能在未来的就业市场上很可能是高需求的

IBERS has generated, and is generating, large quantities of genotype, phenotype and environment based data, along with associated metadata from the energy crop Miscanthus. The student appointed to this project would build on these data sets to create new scientific knowledge about the predictive relationships between data types. This knowledge will be encoded in computational models. To develop the models the student will learn and apply state-of-the-art data-mining and machine-learning methods to identify and characterise genes underlying important quantitative traits (QTL). These will enable the better understanding of Miscanthus biology, and through the defining of ideotypes guide Miscanthus breeding. BACKGROUND The domestication of a new crop as a bioenergy feedstock, requires the application of a wide range of techniques including genetics, biochemistry, physiology, chemical engineering, bioinformatics and system biology modelling. Over the last five years at IBERS a Miscanthus research platform, including a replicated trait trial of 248 diverse genotypes, has been established. Four years of intensive phenotyping of this population, and their progeny from crosses, have resulted in the generation of more than 100,000 datasets. These have been incorporated into a powerful custom coded integrated Miscanthus informatics platform. Example descriptors include: growth environment, morphology, architecture, flowering time, senescence, yield, spring emergence, ploidy, genetic group, and lignin, cellulose, hemicellulose, and cell wall phenolic concentrations. The consolidation of this information and utilisation in predictive models is very important if we are to maximise the value of the data and therefore increase the pace of crop improvement. APPROACH The student will use a combination of data-mining and machine-learning to analyse datasets to identify important associations and to form predictive rules that relate Miscanthus phenotype, genotype and environment. This will use recently developed data-mining and machine-learning techniques, that are well suited to the structure of the data, to create association models. This will be done in close interaction with scientists at IBERS and Ceres who will provide feedback on the patterns and association rules. The process will be an interactive one. The 'interesting' patterns identified by data-mining (e.g. QTL) will provide insight into Miscanthus biology, and will be fed into the machine learning systems as new descriptive attributes to improve the modelling. The analysis of genotype/environment/phenotype data is technically challenging because there are a large number of descriptors, and many of these are best represented relationally. This means that standard statistical methods are probably sub-optimal. Typical aspects of Miscanthus data that lend themselves to relational description are: temporal relations; spatial relations across fields and plants; pedigree relationships between genotypes, etc. Much of statistical genetics, and fields such as spatial statistics, are devoted to developing ad hoc solutions to relational problems that can now be solved in principled ways. Data-mining and machine-learning methods have been used to develop predictive phenotype models using human genotype and environment data; but to the best of our knowledge there have been no application in plant genetics. To evaluate and compare the performance of the approaches with standard statistical genetic methods, we will use standard re-sampling approaches, including cross-validation and bootstrapping. TRAINING This proposed project is well suited to a PhD research project as it combines fallback elements (statistical genetics) along with open-ended elements (relational learning). There are strong training elements of the proposal as the student will gain experience in genetics, bio-energy, data-mining, and machine-learning, all skills likely to be in high-demand in the future job market