INTEGRATIVE GRAPHICAL MODELS FOR LARGE MULTI-MODAL BIOMEDICAL DATA

 描述(申请人提供)：新的高性价比的高通量基因组和成像技术彻底改变了临床诊断和研究领域，但它们也带来了一些新的和重大的挑战。新的数据集很大，而且经常是多模式的，即测量的变量有不同类型：连续(组学、功能磁共振测量)、二进制(SNPs、性别)、数字(年龄、药物剂量)、类别(家族病史、转移组织)、顺序(肿瘤分期、吸烟)。一个关键的分析方面是发现变量之间的直接(因果)联系。这有许多重要的原因：它可以用于分类、生物标记物选择、药物效应，或者用于疾病网络扰动的机制研究。过去曾使用图形模型，但它们没有针对(A)多模式数据和(B)大型数据集进行调整。这项应用的目标是开发新的方法，以确定因果或部分因果网络，可用于支持和增强准确的疾病预测和亚疾病分类，并帮助确定疾病分子机制的关键相互作用。我们将开发和测试基于混合变量部分因果图形(MVPCG)模型的新方法。将对合成数据集和真实数据集进行评估，包括包含基因组、遗传和表观遗传学数据、临床信息和时间序列诊断图像数据的并行数据集。我们的中心假设是，对不同形式的诊断患者数据进行综合的计算分析，可以识别临床和其他疾病相关特征之间的复杂关联和因果关系，从而帮助破译分子疾病机制。交付成果将是(1)用于整合和联合分析多模式生物医学和临床数据的新的图形方法；(2)新的、完全记录的、可无缝地并入其他算法的用于MATLAB和R的软件包；(3)新的完全支持的图形用户界面(GUI)，以进一步向非计算机熟练用户传播我们的方法；(4)关于转移性黑色素瘤患者的发病机制和预测特征的结果；以及(5)关于自闭症谱系受试者和神经典型患者的预测特征的结果。如果成功，这个跨学科团队项目将产生以上可交付成果之外的积极影响，因为我们方法的普遍性使其适用于任何疾病的研究，并使它们在未来海量高通量数据收集将成为所有医院的常规诊断程序时，很容易集成到个性化医疗策略中。

 DESCRIPTION (provided by applicant): The new cost-effective high-throughput genomic and imaging technologies have revolutionized the field of clinical diagnosis and research, but they have also created a number of new and significant challenges. The new datasets are large and frequently multi-modal, i.e. the measured variables are of different types: continuous (omics, fMRI measurements), binary (SNPs, gender), numerical (age, drug dosage), categorical (family history of disease, tissue of metastasis), ordinal (tumor stage, smoking). A key analysis aspect is to discover the direct (causal) associations between variables. This is important for many reasons: it can be used for classification, biomarker selection, drug effect, or for mechanistic studies of network perturbations in disease. Graphical models have been used in the past but they are not tuned for (a) multi-modal data and (b) large datasets. The objective of this application is to develop novel methodologies that will identify causal or partially causal networks, which can be used to support and enhance accurate disease prediction, and sub disease classification and help identify key interactions of the molecular mechanisms of diseases. We will develop and test new methodologies based on mixed variable partially causal graphical (MVPCG) models. Evaluation will be done on synthetic and real datasets, including parallel datasets with genomic, genetic and epigenetic data, clinical information and time series diagnostic image data. Our central hypothesis is that an integrative, computational analysis of different modalities of diagnostic patient data can identify complex associations and causal relations between clinical and other disease relevant features and thus help decipher the molecular disease mechanisms. The deliverables will be (1) new graphical approaches for integration and co-analysis of multi-modal biomedical and clinical data; (2) a new, fully documented software package for MatLab and R that can be seamlessly incorporated in other algorithms; (3) a new fully supported graphical user interface (GUI) to further disseminate our methodologies to non computer-savvy users; (4) results on the pathogenesis and predictive features of metastatic melanoma patients; and (5) results on predictive features of autistic spectrum subjects and neurotypicals. If successful, this cross-disciplinary team project will have a positive impact beyond the above deliverables, since the generality of our approaches makes them suitable for studying of any disease and makes them easily integratable into personalized medicine strategies in the future when massive high-throughput data collection will become a routine diagnostic procedure in all hospitals.