Variable Selection and Prediction for High-Dimensional Genetic Data with Complex Structures

精准医疗的挑战是为每个人提供合适的治疗或建议。大量资源被用于生成基因数据，希望它将提供量身定制的决策。由于基因分型的成本现在已经低于几种常规临床测试的成本，至少有七个大型医疗保健系统已经投资于对其大部分人口进行全基因组基因分型，在这些人口中，电子健康记录数据是可用的。这些数据被用于开发多基因风险评分（PRS），它可以根据遗传数据预测复杂的疾病，因此有可能通过精确医学改善临床护理，这将是加拿大人口非常感兴趣的。为了使这些投资的生产力最大化，需要提高预测准确性的分析工具。在临床决策的背景下，也有必要了解哪些变量驱动这些预测。事实上，实质性专家不愿意使用机器学习文献中所谓的黑箱算法，因为缺乏可解释性和透明度。很难知道算法是如何做出可能产生严重道德后果的决定的。另一方面，虽然统计文献中开发的许多模型是可解释的，并提供了参数估计的不确定性度量，但它们无法扩展到今天生成的大量数据。统计学家不仅要发展理论上合理的方法，而且要考虑计算算法、数据格式和软件等实际问题，这一点变得越来越重要。串联考虑每个组件是朝着在实践中使用更合适的方法迈出的一步。为此，本提案的目标主要围绕三个主题：1)发展新的高维线性混合模型的理论和计算算法，用于相关或搜索数据的变量选择和预测；2)提出关键暴露与高维数据集的交互模型（如基因-环境交互）；3)根据生存时间端点的高维数据开发预测工具。我们的方法将在用户友好的软件中实现，并仔细考虑数据格式和存储，以促进数据分析师更广泛地采用更复杂的模型。这个项目的成果将有助于我成为一名具有复杂结构的高维数据的变量选择和预测模型专业知识的新研究员，并使我在国内和国际上具有竞争力。

The challenge of precision medicine is to appropriately fit treatments or recommendations to each individual. Large amounts of resources are being used to generate genetic data with the hope that it will provide tailored decision making. Since genotyping costs have now dropped below those of several routine clinical tests, at least seven large health care systems have invested in genome-wide genotyping of a large proportion of their population, within whom electronic health record data are available. This data is being used to develop polygenic risk scores (PRS), which can predict complex diseases on the basis of genetic data, and thus have the potential to improve clinical care via precision medicine which would be of great interest to the Canadian population. Analytic tools increasing prediction accuracy are needed to maximize the productivity of these investments. In the context of clinical decision making, there is also a need to understand which variables are driving these predictions. Indeed, there is a reluctance among substantive experts to use so-called black-box algorithms from the machine learning literature because there is a lack of interpretability and transparency. It is difficult to know how the algorithm is making its decisions which can have serious ethical consequences. On the other hand, while many of the models developed in the statistical literature are interpretable and provide measures of uncertainty around their parameter estimates, they are not scalable to the massive amounts of data being generated today. It is becoming increasingly important for statisticians to not only develop theoretically justified methods, but also consider practical issues such as computational algorithms, data format and software. Considering each component in tandem is a step towards more appropriate methods being used in practice. To this end, the goal of this proposal is focused around three Themes: 1) to develop the theory and computational algorithms for new high-dimensional linear mixed models for variable selection and prediction in correlated or groped data; 2) propose interaction models between a key exposure and a high-dimensional dataset (e.g. gene-environment interactions); and 3) develop prediction tools from high-dimensional data for survival time endpoints. Our methods will be implemented in user friendly software, with careful considerations of data format and storage, in order to promote wider uptake of more complex models by data analysts. The results from this project will help to establish me as a new researcher with expertise in variable selection and prediction models for high-dimensional data with complex structures and make me competitive nationally and internationally.