A C. elegans whole-brain digital twin

近几十年来，大脑研究取得了令人瞩目的进步。然而，神经回路的动态，它们对动物行为的规范，对环境或内部状态的适应，以及个体之间的可变性，仍然知之甚少。为了整合神经元功能、电路级计算和全脑协调，自由行为动物的全脑成像是必不可少的。虽然这项技术在大多数动物身上都令人望而生畏，但在毫米长的线虫--线虫身上，这种技术是可用的，并迅速成熟。尽管线虫相对简单，但线虫是一种行为自由的动物，它做出决定，学习，遗忘，适应不断变化的条件，并参与集体行为，以求生存，觅食和逃避捕食。像所有动物一样，它发育、睡眠和衰老，它的研究已经证明它是神经生物学、神经遗传学、学习的神经基础、可塑性和行为以及神经退化的强大模型系统。虽然许多线虫神经元的功能已经被广泛研究，但理解更大电路的动力学提出了新的挑战：全脑成像提供了神经元活动的基本观察，但不是神经元之间的相互作用。因此，我们认为，要在细胞、电路和全局大脑水平上获得完整的理解，需要机械性和解释性的模型。这样的模型必须考虑到由动物的连接体指定的神经回路中出现的全脑活动。为了实现这一目标，我们的总体目标是建立线虫大脑的第一个数字双胞胎。数字双胞胎是真实世界系统的软件表示，用作预测、解释或控制系统在不同条件下的反应的模型。虽然通常应用于工程资产，但方法和挑战(尤其是对内部工作和输出的有限可观察性的有限访问)表明了从数据进行全脑建模的重要共性。具体目标包括：人工智能：开发人工智能工具来训练数字双胞胎，基于线虫连接限制的全脑活动数据。使用&GT的全脑活动数据，应用、测试和扩展单个动物的全脑模型的优化方法；50只动物。使用学习高维时间序列的低维表示(即神经活动轨迹)的深层神经模型来增强全脑数据并引导我们的优化方法。统一我们的框架以获得代表具有相似神经元激活模式和行为编码的模型动物集群的解的家族。开发和应用新的人工智能工具来训练基于数据集种群的模型种群，使用概率和种群密度工具。数字双胞胎：在细胞分辨率上开发线虫大脑的生物基础的机械模型。在线虫的神经生物学、生物学和电路学中实现具有适当基础的神经元和电路模型例如，保守和可变的连接体、已知的突触极性、双边对称性等。根据数据测试和评估优化模型，并基于生物现实主义实施成功解决方案的后选择机制。将成功的模型应用于模拟，为验证实验和未来研究得出新的假设，重点是了解分布式编码及其灵活性、适应性和变异性。如果成功，数字双胞胎将改变我们对线虫大脑的理解，从而改变我们对其他动物神经系统的理解。这个项目，将把人工智能工具放在适当的位置，让我们更接近这个目标。这种新颖的人工智能，以及人工智能、模拟和复杂数据的集成，将有利于生命科学和工程科学中其他数字孪生兄弟的构建。

Brain research has witnessed remarkable advances in recent decades. And yet, the dynamics of neural circuits, their specification of an animal's behaviours, adaptation to context or internal state, and variability across individuals, remain poorly understood. To integrate neuronal function, circuit-level computation, and brain-wide coordination, whole-brain imaging in freely-behaving animals is essential. While daunting in most animals this technology is available and fast-maturing in the mm-long nematode, C. elegans.Despite its relative simplicity, C. elegans is a freely behaving animal that makes decisions, learns, forgets, adapts to ever-changing conditions, and engages in collective behaviour, in order to survive, forage for food and escape predation. Like all animals, it develops, sleeps and ages, and its study has proved it a powerful model system for neurobiology, neurogenetics, the neural basis of learning, plasticity and behaviour, and neurodegeneration.While the functions of many C. elegans neurons have been studied extensively, understanding the dynamics of larger circuits poses new challenges: whole-brain imaging provides essential observation of neuronal activity, but not the interactions between neurons. We therefore argue that to obtain an integrated understanding at cellular, circuit and global-brain levels requires mechanistic and explanatory models. Such models must account for brain-wide activity that emerges from the neural circuitry, as specified by an animal's connectome. To address this goal, our overall aim is to build the first digital twin of the C. elegans brain.A digital twin is a software representation of a real-world system, used as a model to predict, explain or control the system's response under different conditions. While commonly applied to engineering assets, the methodology, and the challenges (in particular, limited access to the internal working and limited observables of the outputs) suggest important commonalities with whole-brain modelling from data.Specific objectives include:AI: To develop AI tools to train a digital twin, based on whole-brain-activity data constrained by the C. elegans connectome.To apply, test and extend optimisation methods for whole-brain models of individual animals, using brain-wide activity data for >50 animals.To augment whole-brain-data and bootstrap our optimisation methods using deep neural models that learn low-dimensional representations of high-dimensional time-series (i.e. neural activity traces).To unify our framework in order to obtain families of solutions representing clusters of model animals with similar neuronal activation patterns and behavioural encoding.To develop and apply novel AI tools for training populations of models based on populations of datasets, using probabilistic and population density tools.Digital Twin: To develop biologically-grounded mechanistic models of the C. elegans brain, at cellular resolution.To implement neuronal and circuit models with appropriate grounding in C. elegans neurobiology, e.g. the conserved and variable connectome, known synaptic polarities, bilateral symmetry, etc.To test and evaluate optimised models against data and implement post-selection mechanisms for successful solutions, based on biological realism.To apply successful models in simulations to derive predictions for validation experiments and new hypotheses for future research, with focus on understanding distributed encoding and its flexibility, adaptability and variability.If successful, a digital twin will transform our understanding of the C. elegans brain, and hence, the nervous systems of other animals. This project, will put in place AI tools that bring us closer to this goal. The novel AI, and the integration of AI, simulations and complex data, will benefit the construction of other digital twins, across life and engineering sciences.