Interpretable Machine Learning to Identify Alzheimer's Disease Therapeutic Targets

项目摘要 阿尔茨海默病(AD)是一个紧迫的国家和国际研究重点。淀粉样斑块和 神经纤维缠结是阿尔茨海默病的标志。它们的组成单元分别是淀粉样蛋白β(Aβ)和tau。 目前，我们缺乏对影响斑块和缠结形成的一系列基因的了解 对这些毒肽的保护性和病理学反应。 生物学家现在正在从人脑组织中收集基因表达数据以及Aβ和tau测量。这个 目前的方法试图找到一组最能预测结果(Aβ)的特征(这里是基因表达水平 或tau水平)。识别的特征，生物标记物，可以帮助确定斑块和缠结的分子基础。 不幸的是，假阳性生物标志物非常常见，复制成功率很低就是明证。 数据独立，临床应用成功率低(不到1%)。我们寻求从根本上改变这股潮流 通过解决当前方法的三个基本问题来实现生物标志物发现的范例， 理论上有充分基础的机器学习(ML)方法从数据中学习可解释的模型。 目标1.从公开的、高通量的大脑数据中学习可解释的特征表征。 高维、隐藏变量和复杂的要素关联会在 可预测性(即观察到的统计关联)和真正的生物相互作用。增加……的机会 识别真正的阳性生物标记物，我们需要新的特征选择标准来学习更好地解释 而不是简单地预测结果。为了做到这一点，我们提出的ML算法将识别 可能通过推断基因的功能来对结果(β或tau水平)做出有意义的解释 在导致阿尔茨海默病的细胞过程和来自许多现有大脑数据集的基因相互作用网络中。 目标2.使用一个统一的框架来解释模型预测，从而做出可解释的预测。由于 疾病异质性、复杂模型(例如，深度学习或集成模型)通常更准确地描述 基因和结果之间的关系比更简单的线性模型，但缺乏可解释性。我们会 开发一种新的ML框架，通过估计每个预测的重要性来解释复杂的模型预测 特定预测的特征，这将把对每个个体具有高度重要性的特征识别为个性化 根据这些重要性估计对主题进行标记和分类。 目的3.使用AD的强大蠕虫模型验证已识别的候选生物标志物。分析 没有进行干预实验的观测数据不能证明因果关系。在协作中 与我的同事Matt Kaeberlein一起，我们将利用AD强大的线虫模型来测试我们关于这一角色的假设 将某些基因作为疾病修饰物，并在此基础上开发一种新的方法来完善模型。 该项目的成功完成将导致Aβ和tau水平的分子基础以前未知， 潜在的治疗靶点，以及广泛适用于许多其他数据科学问题的通用ML技术。

Project Summary Alzheimer’s disease (AD) is an urgent national and international research priority. Amyloid plaques and neurofibrillary tangles are the hallmark of AD. Their building blocks are Amyloid-β (Aβ) and tau, respectively. At present, we lack an understanding of the set of genes that affect formation of plaques and tangles along with protective and pathological responses to these toxic peptides. Biologists are now gathering gene expression data and Aβ and tau measures from human brain tissues. The current approach attempts to find a set of features (here, gene expression levels) that best predict an outcome (Aβ or tau level). The identified features, biomarkers, can help determine the molecular basis for plaques and tangles. Unfortunately, false positive biomarkers are very common, as evidenced by low success rates of replication in independent data and low success reaching clinical practice (less than 1%). We seek to radically shift the current paradigm in biomarker discovery by resolving three fundamental problems with the current approach using novel, theoretically well-founded machine learning (ML) methods to learn interpretable models from data. Aim 1. Learn an interpretable feature representation from publicly available, high-throughput brain data. High-dimensionality, hidden variables, and complex feature correlations create a discrepancy between predictability (i.e., observed statistical associations) and true biological interactions. To increase the chance to identify true positive biomarkers, we need new feature selection criteria to learn a model that better explains rather than simply predicts the outcome. To do so, our proposed ML algorithms will identify the genes that are likely to give a meaningful explanation of the outcome (Aβ or tau level) by inferring both the functions of genes in the cellular processes contributing to AD and the gene interaction network from many existing brain datasets. Aim 2. Make interpretable predictions using a unified framework to explain model predictions. Due to disease heterogeneity, complex models (e.g., deep learning or ensemble models) often more accurately describe relationships between genes and an outcome than simpler, linear models, but lack interpretability. We will develop a novel ML framework that interprets complex model predictions by estimating the importance of each feature to a specific prediction, which will identify features of high importance for each individual as personalized markers and classify subjects based on these importance estimates. Aim 3. Validate the identified candidate biomarkers using powerful worm models of AD. Analyzing observational data without doing interventional experiments cannot prove causal relationships. In collaboration with co-I Matt Kaeberlein, we will utilize powerful nematode models of AD to test our hypotheses on the role of certain genes as disease modifiers, and develop a new way to refine the models based on this knowledge. Successful completion of this project will result in previously unknown molecular basis for Aβ and tau levels, potential therapeutic targets, and general ML techniques widely applicable to many other data science problems.