Interpretable graphical models for large multi-modal COPD data (R01HL159805)

大型多模态 COPD 数据的可解释图形模型 摘要 当今精准医学时代最重要的任务之一是了解其机制和作用 影响临床结果发展的因素。另一项重要任务是开发可解释的、 结果的预测模型。在过去的几年里，许多机器学习方法主导了以下任务： 预测建模，包括深度学习、随机森林等。他们受到前所未有的推动 某些研究领域产生的数据量。然而，这些方法的可解释性 并不简单，当仅使用中小型训练数据集时，其准确性会降低 可用。此外，他们的预测模型并没有揭示其他人之间复杂的相互作用网络。 数据集中的变量，这对于充分理解疾病机制至关重要。此外，大多数此类方法 不太适合在同一数据集中容纳混合数据类型（例如连续、离散）。 概率图模型（PGM）为经典机器学习方法提供了一种有前景的替代方案， 因为它们灵活且多才多艺。他们可以识别变量之间的直接（因果）关系， 指出疾病机制，并根据不同的数据建立预测模型，即使使用 较小的训练数据集。它们已被用于分类、生物标志物选择、可修饰的识别 结果的风险因素，或疾病网络扰动的机制研究。前几年 我们将 PGM 理论框架扩展到混合连续变量和离散变量的分析，或者 没有不可测量的混杂因素；我们现在可以评估并将先验信息合并到混合数据中 图学习。我们成功地将这些方法应用于各种临床重要问题，包括 未确定的肺结节的恶性预测、微生物组的鉴定和其他影响因素 肺炎，SNP生物标志物的选择用于癌症患者的联合治疗。 我们的目标是开发新颖的可解释方法来分析任何类型的数据，并用它们来解决 慢性阻塞性肺病（一种重要的慢性肺部疾病）的临床相关问题。方法评估将在 合成数据和真实数据，包括基因组、遗传、成像和临床慢性阻塞性肺病数据的并行数据集。我们的 中心目标是利用不同方式的患者数据来确定疾病进展机制的因素。 可交付成果将是 (1) 用于对任何类型数据进行综合分析的新 PGM 方法； (2)全新的、完整的 可以合并到其他管道中的记录软件包（R、Python）； （3）新的门户网站 向不懂计算机的慢性阻塞性肺病研究人员传播我们的方法； (4)发病机制的结果 慢性阻塞性肺疾病（COPD）的预测特征。这个跨学科团队项目 由于我们方法的普遍性，预计将产生超出上述可交付成果的积极影响 使它们适合研究任何疾病；它们可以轻松集成到个性化医疗中 当高通量数据收集将成为所有医院的常规诊断程序时的策略。

INTERPRETABLE GRAPHICAL MODELS FOR LARGE MULTI-MODAL COPD DATA ABSTRACT One of the most important tasks in today’s era of precision medicine is to understand the mechanisms and the factors affecting the development of clinical outcomes. Another important task is to develop interpretable, predictive models for outcomes. In the last years, many machine learning methods have dominated the task of predictive modeling, including deep learning, random forests and others. They are fueled by the unprecedent volume of data that have been generated in some research areas. However, the interpretability of these methods is not straight forward and their accuracy decreases when only small to medium size training datasets are available. Furthermore, their predictive models do not uncover the complex web of interactions between other variables in the dataset, which is essential for fully understanding disease mechanisms. Also, most such methods are not well suited to accommodate mixed data types (e.g., continuous, discrete) in the same dataset. Probabilistic graphical models (PGMs) offer a promising alternative to classical machine learning methods, because they are flexible and versatile. They can identify both the direct (causal) relations between variables, pointing to disease mechanisms, and build predictive models over diverse data, with good results even with smaller training datasets. They have been used for classification, biomarker selection, identification of modifiable risk factors of an outcome, or for mechanistic studies of perturbations of disease networks. In the previous years we extended the PGM theoretical framework to the analysis of mixed continuous and discrete variables, with or without unmeasured confounders; and we can now evaluate and incorporate prior information in mixed data graph learning. We successfully applied those methods to diverse clinically important problems, including malignancy prediction of undetermined lung nodules, identification of microbiome and other factors affecting pneumonia, selection of SNP biomarkers for combination treatment of cancer patients. Our objective is to develop novel interpretable methods for analysis of any-type data and use them to address clinically relevant questions in COPD, an important chronic lung disease. Method evaluation will be done on synthetic and real data, including parallel datasets with genomic, genetic, imaging and clinical COPD data. Our central aim is to identify factors of disease mechanisms of progression using different modalities of patient data. The deliverables will be (1) new PGM approaches for integrative analysis of any-type data; (2) a new, fully documented software package (in R, Python) that can be incorporated in other pipelines; (3) a new web portal to disseminate our methodologies to non-computer-savvy COPD researchers; (4) results on the pathogenesis and predictive features of chronic obstructive pulmonary disease (COPD). This cross-disciplinary team project is expected to have a positive impact beyond the above deliverables, since the generality of our approaches makes them suitable for studying any disease; and they can be easily integrated into personalized medicine strategies when high-throughput data collection will become a routine diagnostic procedure in all hospitals.