Machine learning for graph-structured data: Understanding complex biological systems

许多复杂的生物和人造系统可以有效地描述为描述所涉及的元素及其相互关系的图(网络)。例子包括但不限于：人脑、蛋白质分子、药物和化合物、电网和人类交通网络。网络科学是为了补充图论(数学的一个分支)提供的抽象框架而创建的，这使得研究复杂的网络系统及其时间(随时间)的演化变得有用。然而，对这类系统的行为进行微观细节建模并不总是可能的，首先也是最重要的是，因为它们的行为和功能通常以一种非常重要的方式从相对简单的组件的相互作用中显现出来。最近提出的基于深度神经网络的机器学习方法(擅长解决历史上使用计算机难以解决的问题)允许基于图形结构的数据进行预测。这一有趣的方法受到了这样一个事实的挑战，即通常情况下，代表系统演变的观测之间存在着不平凡的时间依赖性。不幸的是，大多数预测行为变化的统计学方法都假设独立的输入没有时间顺序。此外，这类方法在学习过程中通常必须受到监督，而在许多旨在预测复杂系统行为异常或变化的应用中，很难获得可靠的监督。我的长期目标是开发无监督机器学习方法，通过分析具有时间相关性的图形结构数据来预测复杂网络系统中的行为变化并做出反应。在短期内，我的研究将由三个相互关联的目标组成。(1)首先，我将为图形结构的数据设计自回归和状态空间模型(支持基于先前看到的行为预测预期的未来行为)。有了这些知识，我的团队将处理两个具有重大社会意义的应用程序。第一个领域是计算神经科学。(2)我将与我的学生一起，设计新的数据驱动模型，通过分析脑电波(大脑活动扫描)形式的大脑记录，预测和应对耐药患者癫痫发作的发作。第二个应用是在结构生物学方面。(3)我们还将设计其他新颖的数据驱动模型，以处理代表蛋白质-蛋白质和蛋白质-配体复合体随时间演变的一系列图表(以帮助了解我们身体的重要功能)。这最终应该会导致设计出更稳定、更有效的药物。我的HQP参与了拟议的研究，将在先进的机器学习方法及其对图形结构数据的应用方面获得重要的专业知识，这些技能在学术界和行业都已成为战略。

Many complex biological and man-made systems can efficiently be described as graphs (networks) describing the elements involved and their mutual relations. Examples include, but are not limited to: the human brain, protein molecules, drugs and chemical compounds, power grids, and human transport networks. Network science has been created to complement the abstract framework offered by graph theory (a branch of mathematics), making it useful to study complex networked systems and their temporal (over time) evolution. However, modeling the behaviour of such systems down to microscopic details is not always possible, first and foremost because their behaviour and function usually emerges in a highly non-trivial way from the interactions of relatively simple components. Recently proposed machine learning methods based on deep neural networks (which are good at solving problems that have been difficult to solve historically using computers) allow predictions to be made based on graph-structured data. This intriguing approach is challenged by the fact that, usually, there exists a non-trivial temporal dependency between the observations representing the evolution of the system. Unfortunately, most statistical methods for predicting changes in behaviour assume independent inputs that have no temporal ordering. Moreover, such methods must typically be supervised in their learning and reliable supervision is difficult to obtain in many applications aiming at predicting anomalies or changes in the behaviour of complex systems. My long-term goal is to develop unsupervised machine learning methods for predicting and reacting to behaviour changes in complex networked systems by analyzing graph-structured data having temporal correlations. In the short-term, my research will consist of three (3) interrelated objectives. (1) First, I will design autoregressive and state-space models (that support forecasting expected future behaviour based on previously seen behaviour) for graph-structured data. With this knowledge, my team will tackle two applications of great societal importance. The first is in computational neuroscience. (2) Together with my students, I will design novel data-driven models for predicting and reacting to the onset of epileptic seizures in drug-resistant patients by analyzing brain recordings in the form of intracranial EEGs (brain activity scans). The second application is in structural biology. (3) We will also design other novel data-driven models to process a sequence of graphs representing the evolution of protein-protein and protein-ligand complexes over time (to help understand important functions in our bodies). This should ultimately lead to the design of more stable and effective drugs. My HQP involved in the proposed research will gain significant expertise in advanced machine learning methods and their application for graph-structured data, skills that have become strategic both in academia and industry.