Fast Network Inference Methods for Connectome Analysis

描述(由申请人提供)：该R21项目的目标是能够从神经成像数据中进行生物物理上真实的、全脑的、“宏观尺度”的网络分析，数据尺度目前是不可能的。绘制人脑的连通性图是神经科学的重大开放问题之一。了解“连接体”很可能是解开大脑功能及其与正常行为和精神疾病的联系的关键一步。目前用于从神经成像数据中识别大脑活动网络的计算方法范围从描述性简单但计算高效，到描述性详细和生物物理动机，但计算昂贵。虽然后一种模型很流行，但目前计算成本如此之高，以至于不可能将它们用于全脑、数据驱动的网络推理--这种建模是重建人类连接体所必需的。在这项工作中，我们将开发和验证新一代网络推理算法，这些算法将使大型神经成像数据集的全脑、数据驱动、生物物理详细的网络建模成为可能。我们的方法侧重于评估或“评分”单个网络模型的计算成本--这是用于网络推理的组合优化算法核心的关键一步。我们用从数据中学习的快速近似值来代替精确的评分函数，从而在整体优化中产生数量级的加速。在这项工作中，我们将为两类广泛使用的神经成像数据网络模型开发、验证、表征和分发该技术的版本：动态贝叶斯网络(DBN)和动态因果模型(DCMS)。我们通过三个具体的目标来实现这一点：具体目标1：将我们之前在一般数据分析问题中的快速网络识别的工作扩展到用于神经成像分析的非线性可测量活动网络(传感器空间DBN)。具体目标2：为生物物理基础上的潜在变量网络模型，如动态因果模型(DCMS)开发适当的代理函数。具体目标3：与精确评分搜索相比，测试基于代理的搜索识别DBN和DCM的速度更快、规模更大，而不会显著降低模型质量的假设。这项工作为未来研究大脑功能和功能障碍的网络基础铺平了道路，包括探索精神分裂症、精神病或慢性药物成瘾等精神疾病的有效连接基础。

DESCRIPTION (provided by applicant): The goal of this R21 project is to enable biophysically realistic, whole-brain, "macro-scale" network anal- ysis from neuroimaging data, at data scales that are not currently possible. Mapping the connectivity of the human brain is one of the great open problems of neuroscience. Understanding the "connectome" is likely to be a key step in unraveling the function of the brain and its connection to both normal behavior and mental illness. Current computational methods for identifying the brain activity networks from neuroimag- ing data span a spectrum from descriptively simple, but computationally efficient, to descriptively detailed and biophysically motivated, but computationally expensive. While popular, the latter models are currently so computationally expensive that it is infeasible to use them for whole-brain, data-driven network inference - the kind of modeling that is necessary to reconstructing the human connectome. In this work, we will develop and validate a new generation of network inference algorithms that will enable whole-brain, data- driven, biophysically-detailed network modeling of large neuroimaging data sets. Our approach focuses on the computational expense of evaluating, or "scoring," individual network models - a key step in the heart of the combinatorial optimization algorithms used for network inference. We replace the exact scoring function with a fast approximation learned from data, yielding orders of magnitude speedup in the optimization as a whole. In this work, we will develop, validate, characterize, and distribute versions of this technique for two classes of widely-used network models for neuroimaging data: dynamic Bayesian networks (DBNs) and dynamic causal models (DCMs). We accomplish this through three specific aims: Specific Aim 1: Extend our prior work on fast network identification in general data analysis problems to nonlinear measurable activity networks (sensor space DBNs) for neuroimaging analysis. Specific Aim 2: Develop appropriate proxy functions for biophysically-grounded, latent variable network models, such as dynamic causal models (DCMs). Specific Aim 3: Test the hypothesis that proxy-based search identifies DBNs and DCMs faster, and at larger scale, with no significant loss of model quality, compared to exact scoring search. This work paves the way for future studies of the network underpinnings of brain function and dysfunction, including probing the effective connectivity substrates of mental illnesses such as schizophrenia, psychopa- thy, or chronic drug addiction.