Biomedical Data Translator Technical Feasibility Assessment and Architecture Design

在生物医学研究人员之间转换洞察力是一个根本挑战，他们研究 生物机制，和临床医生，谁诊断患者的症状，是那么多的联系 在生物学过程和疾病病理生理学之间，人们知之甚少。一个 全面的生物医学翻译器必须支持跨对象的推理链，因为 基因突变，分子效应，组织特异性表达模式， 细胞过程、器官表型、疾病状态、患者症状和药物 回应，这是一个超出任何一个组织范围的挑战。 幸运的是，这条链上的许多单独的环节都是通过实验得出的 各个数据类型之间的统计关系。高通量扰动筛 将化学和遗传扰动与基因表达等细胞表型联系起来 模式，细胞存活，或磷酸化的变化。遗传关联研究链接 突变到人类疾病或中间表型和生物标记物。电子医疗 记录(EMR)将疾病或人类表型与诊断或当前程序联系起来 术语(CPT)代码和临床试验将药物和候选药物对 疾病状态。 原则上，将这些链接合并到推理链中可以将结果在 它们内部的全套数据类型。在实践中，每个链接都是由专家维护的 特定领域的实验、语义术语和方法标准。而一把钥匙 全球生物医学翻译面临的挑战是在以下领域建立一致的标准 这些现有的数据类型，一个更重要的目标是开发一个原则性和健壮性强的 (A)建立生物系统模型和研究这些系统的实验方法的框架； (B)组织有关生物机制和疾病的知识；和(C)纳入 不同的数据集，作为了解潜在和未知自然状态的窗口。 我们建议实现一个生物医学翻译器作为一个概率图形模型，一个 人工智能(AI)研究的范例。就像独立的研究社区形成一样 翻译过程的弱耦合部分，图形模型允许从 弱耦合的“节点”。这些推论要求每个节点只发布概率 发行版，无需全局实体解析即可实现互操作性 标准，并受益于质量控制、容错和相关性范例 评估在人工智能研究中很常见。我们假设只有有限数量的API， 作为世界各地社区的概率计算实现，将产生一个 生物医学翻译器作为弱耦合知识源的一个新特性。 根据图形模型的基本属性，这样的翻译器可以概率地翻译 在它内部连接的任何数据类型中，允许相对复杂的查询概念。为 例如：在以患者为基础的EMR中，哪些组织的细胞过程受到影响？什么 基因突变使细胞对小分子治疗效果敏感？哪些小分子 模仿预防疾病的基因“大自然实验”？ 为了说明这些资源和我们的体系结构范例的价值，我们建议 实施支持以下查询的生物医学翻译器的示范项目 小分子、生物过程、基因和疾病。示范项目将 为应对关键的数据集成和组织挑战迈出宝贵的第一步 将使以前不可能进行的查询成为可能，例如识别扰乱 在疾病背景下，人类遗传学所涉及的相同的生物过程。以这个身份， 这样的翻译器可以现实地识别已知症状的现有药物(即，改变用途)， 但更广泛地说，它可以作为假说产生和 生物学发现，建议临床前小分子基于它们的 观察到的生物活性，或在细胞蛋白质功能之间提供迄今新的联系 和疾病病理生理学。

A fundamental challenge to translate insights between biomedical researchers, who study biological mechanisms, and clinicians, who diagnose patient symptoms, is that many links between biological processes and disease pathophysiology are poorly understood. A comprehensive Biomedical Translator must enable chains of inference across objects as diverse as genetic mutations, molecular effects, tissue-specific expression patterns, cellular processes, organ phenotypes, disease states, patient symptoms, and drug responses, a challenge beyond the scope of any one organization. Fortunately, many individual links in this chain have been made by experiments yielding statistical connections between individual data types. High-throughput perturbation screens link chemical and genetic perturbations to cellular phenotypes such as gene-expression patterns, cell survival, or changes in phosphorylation. Genetic association studies link mutations to human disease or intermediate phenotypes and biomarkers. Electronic medical records (EMR) link diseases or human phenotypes to diagnostic or current procedural terminology (CPT) codes, and clinical trials link the impact of drugs and drug candidates on disease states. In principle, incorporating these links into chains of inference could translate results between the full set of data types within them. In practice, each link is maintained by experts with domain-specific experiments, semantic terminology, and methodological standards. While a key challenge faced by a global Biomedical Translator is to establish consistent standards across these existing data types, a more important goal is to develop a principled and robust framework to (a) model biological systems and experimental approaches to investigate them; (b) organize knowledge about biological mechanism and disease; and (c) incorporate diverse datasets that serve as windows into the underlying and unknown state of nature. We propose to implement a Biomedical Translator as a probabilistic graphical model, a paradigm from artificial intelligence (AI) research. Just as separate research communities form weakly coupled parts of the translation process, graphical models allow global inferences from weakly coupled “nodes”. These inferences require each node to publish only probability distributions, enabling interoperability without necessarily having global entity-resolution standards, and benefit from paradigms for quality control, fault tolerance, and relevance assessment common in AI research. We hypothesize that a limited number of APIs, implemented as probability computations by communities around the world, would yield a Biomedical Translator as an emergent property of weakly coupled knowledge sources. From basic properties of graphical models, such a Translator could probabilistically translate among any data types connected within it, allowing for relatively complex query concepts. For example: What cellular processes in which tissues are impacted in a patient-based EMR? What genetic mutations sensitize cells to small-molecule treatment effects? Which small molecules mimic genetic “experiments of nature” that protect against disease? To illustrate the value of these resources and our architectural paradigm, we propose a demonstration project to implement a Biomedical Translator supporting queries between small molecules, biological processes, genes, and disease. The demonstration project will provide a valuable first step to confront key data-integration and organizational challenges and will enable previously impossible queries, such as identifying small molecules that perturb the same biological processes implicated by human genetics in a disease context. In this capacity, such Translator could realistically identify existing drugs for known symptoms (i.e., repurposing), but could more broadly serve as an engine for hypothesis generation and biological discovery, suggesting pre-clinical small molecules to develop based on their observed biological activity, or providing heretofore novel links between cellular protein function and disease pathophysiology.