CAREER: Knowledge-driven Analytics, Model Uncertainty, and Experiment Design

高通量、高内容数据获取技术的进步推动了现代科学研究从传统的还原论转向以数据驱动为重点的复杂系统研究的建构论。然而，大数据分析仍然存在重大问题，特别是关于数据驱动的研究结果的可重复性。为了使分析方法从传统的还原论科学研究转向借助丰富数据的建构论研究，提出了一个用于高维网络系统的系统预测和干预的综合贝叶斯框架，这将提供将前所未有的大量异质数据转化为可重复的科学知识的方法，以累积的方式，以客观的新概念，的不确定度量化以及在此基础上的优化实验设计。所提出的知识驱动分析具有强大的潜力，可以将可用的各种大规模数据转化为可复制的知识，以推动生命和材料科学研究。如果成功，它最终可以导致更有效的实验设计的计算工具，以最大限度地利用现有的大数据，并加快有效的疾病治疗和新材料发现的过程。该提案的跨学科性质有望通过研究和教育促进工程，生命科学和材料科学之间的思想交流。这一建议的更广泛的影响包括将拟议的研究与教育计划相结合：（1）加强与生命科学和材料科学合作者的互动。除了公开该项目的所有工具和研究成果外，还将向合作者分发所开发的算法和工具，包括必要的技术帮助，以便进行更有效的合作;（2）发展工程、数学、生命和材料科学的新课程，并将其纳入工程课程，为各级学生提供教育和研究培训，这将有助于新一代的研究人员建立跨学科研究的广泛和坚实的基础，并准备他们所需的技能，以应对现实世界的挑战。这些课程将在整个校园提供，以吸引对科学和工程研究感兴趣的女生和少数民族学生;（三）让本科生和研究生参与这一跨学科领域的研究，通过与REU项目和其他奖学金项目的合作，努力增加科学和工程领域代表性不足的群体的参与，以增加德州农工&大学（TAMU）该提案的科学重点是解决高维网络系统预测和不确定模型干预中存在的数学和计算问题。本论文将解决以下问题：（1）建立一个基于网络的贝叶斯框架和方法，用于异构数据集的系统分析;（2）定义一种新的基于客观的不确定性量化方法，用于评估模型的不确定性和数据的重要性，以使累积分析能够提高系统的理解和优化未来的实验设计;（3）基于不确定性量化，针对不同的作战目标，推导出最优的贝叶斯试验设计方案，从而最大限度地利用现有数据，实现未来试验设计和系统干预的有效性;以及（4）应用所开发的方法学来理解和治疗特定疾病，例如癌症和1型糖尿病;以及与生命和材料科学的合作者合作，以有效的方式设计新材料发现的实验，将知识转化为实际应用。该项目将为将现有数据转化为对生命、疾病和人造系统的系统理解奠定基础，以获得更深入的见解，并将其引导到有益于人类社会的理想系统行为。

The advancement of high-throughput high-content data acquisition techniques has pushed modern scientific research from traditional reductionism to constructionism for studying complex systems with a strong data-driven focus. However, there are still significant issues in big-data analytics, especially, regarding the reproducibility of data-driven research findings. To enable analytic methods from traditional reductionist scientific research to constructionist research with the help of rich data, an integrative Bayesian framework is proposed for systems prediction and intervention in high-dimensional network-based systems, which will provide ways to translate the unprecedented amount of heterogeneous data into reproducible scientific knowledge in a cumulative manner with novel concepts of objective-based uncertainty quantification and optimal experiment design based on that. The proposed knowledge-driven analytics has strong potential of transforming available diverse large-scale data for reproducible knowledge to drive life and materials science research. If successful, it can eventually lead to computational tools for more efficient experiment design to maximize the use of existing big data and speed up the process for effective disease therapeutics and new materials discovery. The interdisciplinary nature of this proposal promises to foster cross-fertilization of ideas between engineering, life science, and materials science through research and education. The broader impact of this proposal involves the integration of the proposed research with an educational plan: (1) to strengthen interactions with the collaborators in life and materials sciences. In addition to making all the tools and research outcomes from this project publicly available, the developed algorithms and tools, including necessary technical help, will be distributed to the collaborators for more effective collaboration; (2) to develop new courses interfacing engineering, mathematics, life and materials sciences and incorporate them into the engineering curriculum for education and research training of students at all levels, which will help the new generation of researchers to establish a broad and solid foundation for interdisciplinary research and prepare them with required skills to address real-world challenges. The courses will be available across campus to attract female and minority students who are interested in research in science and engineering; (3) to involve both undergraduate and graduate students in the research in this interdisciplinary field with the efforts to increasing the participation of underrepresented groups in science and engineering through the collaborations with the REU programs and other scholarship programs developed to increase diversity at Texas A&M University (TAMU). The scientific focus of this proposal is solving mathematical and computational problems that exist in high-dimensional network-based systems prediction and intervention with uncertain models. The following open problems will be addressed: (1) to develop a network-based Bayesian framework and methods for systematic analysis of heterogeneous data sets; (2) to define novel objective-based uncertainty quantification for assessing model uncertainty and data significance to enable cumulative analytics to improve systems understanding and optimize future experiment design; (3) to derive optimal Bayesian experiment design based on uncertainty quantification for different operational objectives, which can lead to the maximal use of existing data and effective future experiment design and systems intervention; and (4) to apply the developed methodologies for understanding and treating specific diseases, for example, cancer and type 1 diabetes; as well as designing experiments for new materials discovery in an efficient way, in collaboration with the collaborators in life and materials sciences, which will translate the knowledge to practical applications. This project will lay out the foundation for translating existing data into systems understanding of life, disease, and man-made systems to gain deeper insights and direct them to desirable systems behavior to benefit human society.