Computational Drug Repurposing for AD/ADRD with Integrative Analysis of Real World Data and Biomedical Knowledge

摘要 阿尔茨海默病（AD）和AD相关痴呆（AD/ADRD）是美国第六大死亡原因， 是一种与衰老相关的神经退行性疾病，具有复杂的致病机制， 预计到2021年将有620万美国人。AD/ADRD的发病机制和病理生理 复杂，在寻找有效的新的治疗或预防策略方面造成困难， 近十年来的投资。另一方面，大型临床研究网络（CRN）的激增， 使用真实世界数据（RWD），例如电子健康记录（EHR）、索赔和账单数据等， 提供了独特的机会，以产生现实世界的证据（RWE），将有直接的转化影响， AD/ADRD。过去，RWD（如EHR）在AD/ADRD药物再利用方面的用途有限，主要用于 仅用于验证和评价AD/ADRD分子水平预测产生的假设 重新利用代理人，部分原因是一些关键的方法差距：（1）缺乏整合， 现有丰富的AD/ADRD生物学和病理生理学知识可用于假设生成，（2）缺乏 经过验证的可计算表型（CP）和自然语言处理（NLP）算法和工具， 准确定义研究人群，提取关键相关患者特征和有意义的结局 (e.g., MMSE评分以确定严重程度），（3）缺乏对疾病异质性的考虑（即， AD/ADRD亚型），以及（4）缺乏对RWD固有偏倚的认识和应用 因果推理原则这个项目的目标是开发一个全面的机器学习的基础上， 因果推理框架，用于生成高通量和高质量的药物再利用假设， AD/ADRD通过整合异构信息源。这个项目有三个目标。目标1 开发可计算表型，以提取与AD/ADRD相关的关键患者特征和结局 来自RWD的药物再利用研究。目标2旨在发展一个以学习为基础的因果推理框架 用于从RWD生成药物再利用假设，用于生成 来自生物医学知识库（BKB）的药物再利用假设;以及互信息增强 结合来自RWD和BKB的信息，以进一步提高生成的 假设目标3旨在通过不同的数据来源和方法验证生成的假设。 该项目将利用来自两个大型临床研究网络（CRN）的患者数据， 国家以患者为中心的临床研究网络（PCORnet）-覆盖约1500万佛罗里达人和约1100万佛罗里达人。 百万纽约人开发的算法和软件将开源并广泛传播 在CRN和AD/ADRD研究社区内。

ABSTRACT Alzheimer’s disease (AD) and AD-related dementia (AD/ADRD) is the 6th leading cause of death in the United States (US) – is an aging-related neurodegenerative disease with complex pathogenic mechanism affecting an estimated 6.2 million Americans in 2021. Both the pathogenic mechanism and pathophysiology of AD/ADRD are complex, creating difficulties in finding effective new treatment or prevention strategies, despite significant investments in the last decade. On the other hand, the proliferation of large clinical research networks (CRNs) with real-world data (RWD), such as electronic health records (EHRs), claims, and billing data among others, offer unique opportunities to generate real-world evidence (RWE) that will have direct translational impacts on AD/ADRD. In the past, RWD such as EHRs have limited use for AD/ADRD drug repurposing and primarily used only for validating and evaluating the hypotheses generated by molecular level predictions of AD/ADRD repurposing agents, partially due to a number of key methodological gaps: (1) the lack of integration with existing rich biological and pathophysiological knowledge of AD/ADRD for hypothesis generation, (2) the lack of validated computable phenotyping (CP) and natural language processing (NLP) algorithms and tools that can accurately define the study populations, extract key relevant patient characteristics and meaningful outcomes (e.g., MMSE scores to determine severity), (3) the lack of consideration on the heterogeneity of the disease (i.e., AD/ADRD subtypes), and (4) the lack of recognition of the inherent biases in RWD and the need of applying causal inference principles. The goal of this project is to develop a comprehensive machine learning based causal inference framework for generating high-throughput and high-quality drug repurposing hypotheses for AD/ADRD by integrating heterogeneous information sources. There are three aims in this project. Aim 1 aims at developing computable phenotypes to extract key patient characteristics and outcomes relevant to AD/ADRD drug repurposing studies from RWD. Aim 2 aims at developing a learning-based causal inference framework for generating drug repurposing hypotheses from RWD, a deep knowledge embedding framework for generating drug repurposing hypotheses from biomedical knowledge bases (BKB); and a mutual information enhancement framework that combines the information from both RWD and BKB to further improve the quality of the generated hypotheses. Aim 3 aims at validating the generated hypotheses with diverse data sources and approaches. The project will leverage the patient data from two large clinical research networks (CRNs) contributing to the national Patient-Centered Clinical Research Network (PCORnet) – covering ~15 million Floridians and ~11 million New Yorkers. The developed algorithms and software will be open sourced and widely disseminated within the CRNs and the AD/ADRD research communities.