Statistical and Machine Learning Methods to Address Biomedical Challenges for Integrating Multi-view Data

摘要 阿尔茨海默病(AD)是一种复杂而多样的疾病，影响着580万65岁或以上的成年人 在美国，老年AD是痴呆症最常见的原因，并呈现出显著的和不断增长的经济 和社会负担。我们从认知正常(CN)中诊断和分类AD的能力，或区分 患有阿尔茨海默病、早期轻度认知障碍[EMCI]或晚期轻度认知障碍(LMCI)的人是必不可少的 用于预防、诊断和治疗阿尔茨海默病。因为患有MCI的人有很高的机会转换 到AD，有效地区分转换为AD的人(MCI-C)和不转换的人(MCI-C) NC)对AD的早期诊断具有重要意义。AD的异质性进一步推动了分类的尝试 不同的AD亚群，以更好地了解潜在的生理学。有证据表明，使用数据 跨多种模式(例如，遗传学、成像、代谢组学)有可能更好地对AD亚组进行分类 而不是使用单一的医疗模式。然而，大多数使用多模式数据的AD研究都专注于成像 数据或仅限成像和遗传学数据。 本研究的目的是创新性地应用最新的机器学习和深度学习 我们已经开发了整合遗传学、成像、代谢组学、类脂组学和表型数据的方法-来自 NIAGADS-更好地了解AD的病因。我们的具体目标是：目标1(A)确定新的分子 可能区分AD病例和认知正常(CN)的特征和途径，MCI转换器[MCI-C]来自 MCI非转换者[MCI-NC]，使用多模式数据；(B)开发多基因风险评分和其他分子 风险评分，以确定患AD风险较高的个人，以帮助临床决策；目标2： (A)确定阿尔茨海默病进展和族裔群体中的分子变化[探索性研究]和(B)确定 AD的同质亚群，以亚群特有的分子和途径为特征。虽然ML和 DL已成功应用于AD研究，但其潜力尚未得到充分利用。 所提出的研究是可行的，有前景的，对AD的研究具有潜在的意义。我们希望能确定 一)分子签名和赋予AD风险或预防AD的途径二)风险较高的个人 用于开发AD和iii)AD分子亚群和亚群特有的分子生物标志物和途径。 最终，我们的发现有可能通过加深我们对AD的理解来为AD研究做出贡献 机制，完善个性化护理，并增强我们确定疾病治疗目标的能力。

Abstract Alzheimer’s disease (AD) is a complex and heterogeneous condition that affects 5.8 million adults 65 years or older in the U.S. AD is the most common cause of dementia and presents a substantial and increasing economic and social burden. Our ability to diagnose and classify AD from cognitive normals (CN), or discriminate among individuals with AD, early mild cognitive impairment [EMCI], or late mild cognitive impairment (LMCI), is essential for the prevention, diagnosis, and treatment of AD. Since individuals with MCI have a high chance of converting to AD, effectively discriminating between those who convert to AD (MCI-C) from those who do not convert (MCI- NC) is important for early diagnosis of AD. The heterogeneity of AD has further motivated attempts to classify distinct subgroups of AD to better inform the underlying physiology. There is evidence to suggest that using data across multiple modalities (e.g. genetics, imaging, metabolomics) has potential to classify AD subgroups better than using single modality. However, most AD studies that have used multimodal data have focused on imaging data or imaging and genetics data only. The purpose of this study is to innovatively apply state-of-the-art Machine learning (ML) and Deep Learning (DL) methods we have developed to integrate genetics, imaging, metabolomics, lipidomics, and phenotypic data– from NIAGADS– to better understand the etiology of AD. Our specific goals are: Aim 1 (a) Identify novel molecular signatures and pathways likely differentiating AD cases from cognitively normal (CN), MCI converters [MCI-C] from MCI non-converters[MCI-NC], using multimodal data; (b) Develop polygenic risk scores (PRS) and other molecular risk scores to identify individuals at a higher risk for developing AD to aid in clinical decision making; Aim 2: (a) Characterize molecular changes in AD progression and in ethnic groups [exploration study] and (b) Identify homogeneous subgroups of AD characterized by subgroup-specific molecules and pathways. Although ML and DL have been successfully used in AD research, their potential have not been fully harnessed. The proposed research is feasible, promising and potentially significant to AD research. We expect to identify i) molecular signatures and pathways conferring risk for, or protection against, AD ii) individuals at a higher risk for developing AD and iii) AD molecular subgroups and subgroup-specific molecular biomarkers and pathways. Ultimately, our findings have the potential to contribute to AD research by furthering our understanding of AD mechanisms, refining personalized care, and enhancing our ability to identify targets for disease treatment.