Development of Novel Multi-Task Prediction Methods for Large Scale Genomic Data

我的研究重点是开发用于多任务预测问题的新的统计和计算方法和软件，以及它们在基因组数据分析中的应用。背景：各种高通量实验技术正在产生大量的基因组数据，如下一代测序。大多数基因组数据的共同特征是高维的(基因表达数据高达数万，全基因组关联研究高达数百万)和高度相关性。由于这些数据通常具有小样本和中等样本大小，从数百到数千不等，因此信号稀疏，这给寻找与疾病相关的重要基因组预测因子带来了挑战。传统的统计方法主要是针对低维数据设计的。而流行的深度学习需要大样本来训练网络才能有更好的预测性能。因此，对于这样的数据，迫切需要基于统计学习的新方法。特别是，在开发多任务学习方法方面存在差距，这种方法同时预测多个研究结果，并从一个大的池中选择重要的基因组预测因子。目的：我最近在修改著名的堆积算法的基础上开发了一种新的算法，并证明在中等样本量的基因组数据中，它的性能优于神经网络和其他流行的统计学习方法。我建议将这项工作扩展到处理其他常见类型的结果，并计划将建议的方法纳入我之前开发的公开可用的R包，称为MTPS。该软件将成为处理具有高维关联数据的多任务问题的统一和综合的软件平台。我将在公平、多样性和包容性的指导下，通过培训高素质人才(HQP)来实施这项建议。影响：我的软件MTPS自2020年2月发布以来，已经被下载了4000多次。通过引入建议的方法，我们将为其添加新的功能，以解决更广泛的问题。我的新方法和软件将成为未来最受欢迎的多任务预测工具之一。我将通过应用所提出的方法来解决卫生科学和社会科学中的现实世界问题，从而在各个学科领域做出重大贡献。例如，我将与不列颠哥伦比亚大学医学院的世界领先研究人员合作，将这些方法应用于各种肺部疾病的生物标记物发现。已识别的生物标志物可用于开发诊断测试芯片或作为药物开发的靶标。我将与经济学家合作，将这些方法应用于美国股市的高频交易数据，并建立一个预测多只股票价格的模型。我相信，这样的应用将产生实质性的影响，并为加拿大自然科学和工程学的发展做出贡献。

My research focuses on developing novel statistical and computational methodologies and software for multi-task prediction problems and their application in the analysis of genomic data. Background: Tremendous genomic data are being generated by various high-throughput experimental technologies, such as Next Generation Sequencing. Common features of most genomic data are high dimensional (up to tens of thousands for gene expression data or up to millions for Genome-Wide Association Studies) and high correlation. Since those data usually have small and moderate sample sizes, ranging from hundreds to thousands, signals are sparse, creating challenges for finding important disease-associated genomic predictors. Traditional statistical methods are mainly designed for low dimensional data. And the popular deep learning requires a large sample to train the network to have better prediction performance. Therefore, for such data, novel methods based on statistical learning are in high demand. Particularly, there are gaps in developing multi-task learning methods, which simultaneously predict multiple research outcomes and select important genomic predictors from a large pool. Objectives: I recently developed a novel algorithm based on revising the well-known stacking algorithm and established that it outperforms the neural network and other popular statistical learning methods in genomic data for moderate sample sizes. I propose extending this work to handle other common types of outcomes and plan to incorporate the proposed methods into my previously developed publicly available R package, called MTPS. The software will become a unified and comprehensive software platform dealing with multi-task problems with high-dimensional correlated data. I will conduct the proposal through training highly qualified personnel (HQP) under the guidelines of equity, diversity, and inclusion. Impact: My software, MTPS, has been downloaded 4000+ times since released in February 2020. By incorporating the proposed methods, we will add new features to it to solve a broader range of problems. My novel methods and software will become one of the most popular tools for multi-task prediction in the future. I will make significant contributions to various subject areas by applying the proposed methods to solve real-world problems in health science and social science. For example, in collaboration with world-leading researchers at the University of British Columbia's medical school, I will apply the methods in biomarker discovery for various lung diseases. The identified biomarkers can be used to develop diagnosis test chips or as targets for drug development. In collaboration with economists, I will apply the methods to the US stock market's high-frequency trading data and build a model to predict multiple stocks' prices. I believe such applications will lead to substantial impacts and contribute to the development of Canadian Natural Sciences and Engineering.