III: Small: Graph Generative Deep Learning for Protein Structure Prediction

III：小：用于蛋白质结构预测的图生成深度学习

• 批准号: 1907805
• 负责人: Liang Zhao
• 金额: $ 49.98万
• 依托单位: George Mason University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2019
• 资助国家: 美国
• 起止时间: 2019-08-01 至 2021-02-28
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1907805&HistoricalAwards=false
• 关键词: III; Small; Graph; Generative; Deep

Decades of scientific enquiry beyond molecular biology have demonstrated just how fundamental form is to function, whether in understanding phase transitions in statistical physics, predicting the evolution and dynamics of real networks in network science, or successfully steering an articulated robot arm to a target pose. A fundamental question in all these scientific domains is how to effectively explore the space of all possible forms of a dynamic system to uncover those that satisfy non-trivial constraints imposed by function. The most visible instantiation of this question in computational structural biology is de-novo protein structure prediction (PSP). PSP takes a structure-driven view of understanding molecular mechanisms in the cell and seeks to determine one or more biologically-active/native structures of a protein from knowledge of its chemical composition. Elucidating such structures is central to inferring the biological activities of a rapidly-growing number of protein-encoding gene sequences and thus advancing our understanding of the inner workings of a cell. While PSP has a natural formulation under stochastic optimization, current efforts are approaching a saturation point. This project proposes a radically-different, complementary approach. Inspired by recent momentum in generative deep learning, the project approaches de-novo PSP under the umbrella of generative, adversarial deep learning. The approach is firmly grounded in information integration and informatics, as it proposes generative models that learn in an adversarial setting to generate native-like tertiary protein structures. The project benefits researchers in machine learning, deep learning, and information integration with interests in graph generative models, molecule generation, and protein structure prediction. The project will result in open-source codes, online teaching modules and tutorials, publicly-available data and models, workshops, software demos, and will broaden the participation in computing of under-represented students.The activities in this project chart a new algorithmic path under the umbrella of information integration and informatics to address the current impasse in structure-function related problems in molecular biology. The focus is on the de-novo protein structure prediction problem. With experimental structure determination lagging behind the rapidly-growing number of protein-encoding gene sequences by high-throughput sequencing technologies, computational approaches have a central role in molecular biology research. Great progress has been made through stochastic optimization, but current approaches are experiencing diminishing returns, partly due to fundamental challenges concerning the resource-aware exploration-exploitation control in complex search spaces and inherently inaccurate scoring functions. This project puts forth a novel approach to structure prediction under the umbrella of generative, adversarial deep learning, leveraging recent advances and opportunities in graph generative learning, adversarial learning, and deep learning. Generative models learn in an adversarial setting to generate native-like tertiary protein structures. The proposed activities span multiple disciplines and promise to make general contributions in machine learning, deep learning, explainable AI, molecular modeling, and computational biology. The work will also benefit researchers and students interested in modeling complex, dynamic systems. The investigators will disseminate the proposed research via open-source codes in C++ and Python so as to reach diverse communities of researchers and students, online teaching modules and tutorials, trained models and data. They will actively educate involved communities through workshops, tutorials, and software demonstrations. This interdisciplinary project also creates excellent opportunities to broaden the participation in computing of under-represented students of all backgrounds.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

几十年来，在分子生物学之外的科学探索已经证明了基本形式是如何发挥作用的，无论是在理解统计物理学中的相变，预测网络科学中真实网络的进化和动态，还是成功地将铰接机器人手臂转向目标姿势。所有这些科学领域的一个基本问题是如何有效地探索一个动态系统的所有可能形式的空间，以发现那些满足函数施加的非平凡约束的形式。在计算结构生物学中，这个问题最明显的实例是de-novo蛋白质结构预测（PSP）。PSP采用结构驱动的观点来理解细胞中的分子机制，并试图从其化学组成的知识中确定蛋白质的一个或多个生物活性/天然结构。阐明这些结构对于推断数量迅速增长的蛋白质编码基因序列的生物活动至关重要，从而促进我们对细胞内部工作原理的理解。虽然PSP在随机优化下具有自然的公式，但目前的努力正在接近饱和点。该项目提出了一种完全不同的互补方法。受最近生成式深度学习发展势头的启发，该项目在生成式、对抗性深度学习的框架下进行了从头开始的PSP。该方法牢固地建立在信息集成和信息学的基础上，因为它提出了在对抗环境中学习以生成原生三级蛋白质结构的生成模型。该项目有利于机器学习、深度学习和信息集成方面的研究人员，并对图生成模型、分子生成和蛋白质结构预测感兴趣。该项目将产生开源代码、在线教学模块和教程、公开可用的数据和模型、研讨会、软件演示，并将扩大代表性不足的学生对计算机的参与。本项目的活动在信息集成和信息学的保护下绘制了一条新的算法路径，以解决当前分子生物学中结构-功能相关问题的僵局。重点研究了de-novo蛋白结构预测问题。由于实验结构的确定落后于高通量测序技术快速增长的蛋白质编码基因序列，计算方法在分子生物学研究中发挥了核心作用。随机优化已经取得了很大的进展，但目前的方法正在经历收益递减的过程，部分原因在于复杂搜索空间中资源感知勘探开发控制的基本挑战以及固有的不准确的评分函数。该项目提出了一种在生成式、对抗性深度学习的保护下进行结构预测的新方法，利用了图生成式学习、对抗性学习和深度学习方面的最新进展和机会。生成模型在对抗环境中学习生成原生的三级蛋白质结构。提议的活动跨越多个学科，并承诺在机器学习，深度学习，可解释的人工智能，分子建模和计算生物学方面做出普遍贡献。这项工作也将使对复杂动态系统建模感兴趣的研究人员和学生受益。研究人员将通过c++和Python的开源代码传播拟议的研究，以便接触到不同的研究人员和学生社区、在线教学模块和教程、训练过的模型和数据。他们将通过研讨会、教程和软件演示积极地教育相关社区。这个跨学科的项目也创造了极好的机会，扩大了所有背景的代表性不足的学生对计算机的参与。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

Dynamic Activation of Clients and Parameters for Federated Learning over Heterogeneous Graphs

• DOI: 10.1109/icde55515.2023.00126
• 发表时间: 2023-04
• 期刊: 2023 IEEE 39th International Conference on Data Engineering (ICDE)
• 作者: Zishan Gu;Ke Zhang;Liang Chen;Sun

Zero-Shot Cross-Lingual Machine Reading Comprehension via Inter-Sentence Dependency Graph

• DOI: 10.1609/aaai.v36i10.21407
• 发表时间: 2021-12
• 期刊: ArXiv
• 作者: Liyan Xu;Xuchao Zhang;Bo Zong;Yanchi Liu;Wei Cheng;Jingchao Ni;Haifeng Chen;Liang Zhao;Jinho D. Choi

Schematic Memory Persistence and Transience for Efficient and Robust Continual Learning

• DOI: 10.1016/j.neunet.2021.08.011
• 发表时间: 2021-05
• 期刊: Neural networks : the official journal of the International Neural Network Society
• 作者: Yuyang Gao;G. Ascoli;Liang Zhao

Designing a Direct Feedback Loop between Humans and Convolutional Neural Networks through Local Explanations

• DOI: 10.1145/3610187
• 发表时间: 2023-07
• 期刊: Proceedings of the ACM on Human-Computer Interaction
• 作者: Tong Sun;Yuyang Gao;Shubham Khaladkar;Sijia Liu;Liang Zhao;Younghoon Kim;S. Hong

Metagraph Aggregated Heterogeneous Graph Neural Network for Illicit Traded Product Identification in Underground Market

• DOI: 10.1109/icdm50108.2020.00022
• 发表时间: 2020-11
• 期刊: 2020 IEEE International Conference on Data Mining (ICDM)
• 作者: Yujie Fan;Yanfang Ye;Qian Peng;Jianfei Zhang;Yiming Zhang-;Xusheng Xiao;C. Shi;Qi Xiong;Fudong Sh