COINSTAC: decentralized, scalable analysis of loosely coupled data

 描述(由申请人提供)： 脑成像界从目前正在进行的广泛的数据共享努力中受益匪浅5，10。然而，现有的策略存在着显著的差距，这些策略侧重于匿名的、特殊后的共享，或者1)完整的原始或预处理数据[在开放研究的情况下]，或者2)手动计算的摘要测量[例如，在封闭(或尚未共享)研究的情况下，例如海马体体积11]，这是我们建议解决的问题。目前共享数据的方法往往包括共享数据的调查员以及请求数据的个人的重大后勤障碍(例如，往往需要美国和国际机构签署多项数据共享协议和批准)。这种情况需要改变，以便科学界成为一个可以收集、管理、广泛共享和分析数据的场所，同时也开放了对目前无法获得的(许多)数据集的访问(见我们小组2对此的最新概述)。 现有的大量数据需要一种方法，这种方法可以以分布式的方式分析数据，同时将源数据的控制权留给单独的调查员；这促使采用动态的、分散的方式来处理大规模分析。我们正在提出一个对等系统，称为用于匿名计算的协作信息学和神经成像套件工具包(COINSTAC)。该系统将提供一个独立的、开放的、不附带任何条件的工具，对分布在不同地点的数据集进行分析。因此，可以避免实际汇总数据的步骤，同时可以保留大规模分析的优势。为了实现这一点，在目标1中，我们提出的统一数据接口将使共享和协作变得容易。还将纳入强大和新颖的质量保证和可复制工具。协作和数据共享将通过形成临时的(需要的和基于项目的)虚拟研究集群，对其各自的数据进行自动生成的本地计算，并在全球推理程序中汇总统计数据来实现。社区组织将提供源源不断的大型项目，这些项目可以形成和完成，而不需要创建新的僵化组织或面向项目的存储库。在目标2中，我们开发、评估和整合隐私保护算法，以确保使用的数据即使在多次重复使用时也无法重新识别。我们还将为几个关键的多变量算法家族(一般线性模型、矩阵分解[例如独立分量分析]、分类)开发先进的分布式和隐私保护方法，以估计内在网络并执行数据融合。最后，在目标3中，我们将通过对具有多种成像模式的国家和国际场所的药物滥用数据集的分布式分析，在概念验证研究中展示这一方法的实用性。

 DESCRIPTION (provided by applicant): The brain imaging community is greatly benefiting from extensive data sharing efforts currently underway5,10. However, there is a significant gap in existing strategies which focus on anonymized, post-hoc sharing of either 1) full raw or preprocessed data [in the case of open studies] or 2) manually computed summary measures [such as hippocampal volume11, in the case of closed (or not yet shared) studies] which we propose to address. Current approaches to data sharing often include significant logistical hurdles both for the investigator sharing the dat as well as for the individual requesting the data (e.g. often times multiple data sharing agreements and approvals are required from US and international institutions). This needs to change, so that the scientific community becomes a venue where data can be collected, managed, widely shared and analyzed while also opening up access to the (many) data sets which are not currently available (see recent overview on this from our group2). The large amount of existing data requires an approach that can analyze data in a distributed way while also leaving control of the source data with the individual investigator; this motivates dynamic, decentralized way of approaching large scale analyses. We are proposing a peer-to-peer system called the Collaborative Informatics and Neuroimaging Suite Toolkit for Anonymous Computation (COINSTAC). The system will provide an independent, open, no-strings-attached tool that performs analysis on datasets distributed across different locations. Thus, the step of actually aggregating data can be avoided, while the strength of large-scale analyses can be retained. To achieve this, in Aim 1, the uniform data interfaces that we propose will make it easy to share and cooperate. Robust and novel quality assurance and replicability tools will also be incorporated. Collaboration and data sharing will be done through forming temporary (need and project-based) virtual clusters of studies performing automatically generated local computation on their respective data and aggregating statistics in global inference procedures. The communal organization will provide a continuous stream of large scale projects that can be formed and completed without the need of creating new rigid organizations or project-oriented storage vaults. In Aim 2, we develop, evaluate, and incorporate privacy-preserving algorithms to ensure that the data used are not re-identifiable even with multiple re-uses. We also will develop advanced distributed and privacy preserving approaches for several key multivariate families of algorithms (general linear model, matrix factorization [e.g. independent component analysis], classification) to estimate intrinsic networks and perform data fusion. Finally, in Aim 3, we will demonstrate the utility of this approach in a proof of concept study through distributed analyses of substance abuse datasets across national and international venues with multiple imaging modalities.