Bayesian methods for complex, high-dimensional functional data in cancer research

 描述（由申请人提供）：复杂的高维数据（如多平台基因组学和成像数据）可用于发现生物标志物，从而深入了解癌症病因、自然史、预后和治疗反应预测。然而，现有的分析方法还不够，因为大多数方法要么忽略数据中的重要结构，要么将分析限制为不使用数据中所有信息的简单摘要。这项研究将开发一套通用的灵活、自动化、新颖的贝叶斯方法，用于对复杂的高维功能数据进行回归分析，以使用解释其复杂结构的模型来发现生物标志物，产生可调整多重测试的推理，并可扩展到高维设置。虽然普遍适用，但这些方法将在我们的合作者进行的两项研究的背景下开发，以发现膀胱癌自然史中的早期基因组和表观遗传事件以及与戒烟成功相关并预测戒烟成功的神经影像生物标志物。具体目标 1：将多平台基因组数据建模为函数，我们将在由我们的 co-I Czerniak 开发的新型膀胱癌模型生成的网格上开发空间相关基因组数据的功能响应回归方法。我们将应用这些方法来识别膀胱癌的基因组和表观遗传变化，并确定在疾病自然史中首次观察到的时间，揭示可能是疾病驱动因素的早期畸变。我们将开发推理策略来执行基因组水平测试，然后在调整多样性的同时聚合基因组区域。具体目标 2：我们将根据 Cinciripini 和 Versace 共同进行的一项随机戒烟试验开发事件相关电位 (ERP) 的功能回归方法，以测试不同的情绪刺激是否会引起不同的神经反应，确定这些影响在戒烟尝试成功或不成功的个体之间是否存在差异，并评估 ERP 是否是成功的独立预测因素。我们的方法将通过空间函数过程或张量基函数灵活地捕获电极间相关性，并捕获电极内相关性 使用基函数建模，并通过策略来确定哪种基最适合 ERP。具体目标 3：我们将为戒烟试验中的 fMRI 数据开发功能回归方法，首先在受试者水平上识别不同视觉刺激差异激活的大脑区域，然后引入一种策略将我们的方法扩展到群体水平分析，以表征人群水平的神经系统差异，将其与戒烟成功联系起来，并评估其相对于 ERP 和标准人口统计学、心理测量和遗传预测因素的预测能力。我们的纵向相关体积数据模型将通过基函数建模捕获体积内相关性，引入一种新颖的混合基函数建模策略，该策略以考虑已知解剖结构、空间邻近性和由功能连接引起的遥远相关性的方式捕获脑内相关性。具体目标 4：我们将把这些新方法集成到一套通用的贝叶斯方法中，用于复杂、高维函数的空间和纵向相关功能响应回归、区分和推理，以及可广泛应用的免费、自动化、可扩展的软件。

 DESCRIPTION (provided by applicant): Complex, high-dimensional data like multi-platform genomics and imaging data can be used to discover biomarkers providing insight into cancer etiology, natural history, prognosis, and prediction of response to therapy. Existing analytical methods are not adequate, however, as most either ignore important structure in the data or limit analysis to simple summaries that do not use all of the information in the data. This research will develop a general suite of flexible, automated, novel Bayesian methods for performing regression analyses on complex, high dimensional functional data to discover biomarkers using models that account for their intricate structure, yield inference that adjusts fo multiple testing, and are scalable to high-dimensional settings. While generally applicable, these methods will be developed in the context of two studies conducted by our collaborators to discover early genomic and epigenetic events in the natural history of bladder cancer and neuroimaging biomarkers associated with and predicting smoking cessation success. Specific Aim 1: Modeling multi-platform genomic data as functions, we will develop methods for functional response regression for spatially correlated genomics data on a lattice generated by a novel bladder cancer model developed by our co-I Czerniak. We will apply these methods to identify genomic and epigenetic changes in bladder cancer and determine when first observed in the disease's natural history, revealing early aberrations that are potential disease drivers. We will develop inferential strategies to perform genome-level tests and then ag genomic regions while adjusting for multiplicity. Specific Aim 2: We will develop functional regression approaches for event-related potentials (ERPs) from a randomized smoking cessation trial conducted by our co-Is Cinciripini and Versace to test whether different emotional stimuli evoke differential neurological response, determine whether these effects vary between individuals successful or unsuccessful in their smoking cessation attempt, and assess whether ERPs are independent predictors of success. Our methods will flexibly capture inter-electrode correlation via spatial functional processes or tensor basis functions, and capture intra-electrode correlation using basis function modeling, with strategies to determine which basis is best for ERPs. Specific Aim 3: We will develop functional regression approaches for fMRI data from our smoking cessation trial, first at the subject level to identify brain regions differentially activaed by different visual stimuli, and then introducing a strategy to scale our approach up to group-level analyses to characterize population-level neurological differences, relate them to cessation success, and assess their predictive ability relative to ERP and standard demographic, psychometric, and genetic predictors. Our models for longitudinally correlated volumetric data will capture intra-volume correlation through basis functional modeling, introducing a novel hybrid basis function modeling strategy that captures within-brain correlation in a manner that accounts for known anatomy, spatial proximity, and distant correlations induced by functional connectivity. Specific Aim 4: We will integrate these new methods into a general suite of Bayesian methods for spatially and longitudinally correlated functional response regression, discrimination, and inference for complex, high-dimensional functions along with freely available, automated, scalable software that can be broadly applied.