Biomedical Data Translator Development of Autonomous Relay Agent: ARAX

该项目将通过一个多站点的翻译联盟继续开展协作工作， 团队（"团队X射线"）在俄勒冈州州立大学（PI斯蒂芬拉姆齐）和两个合作伙伴 宾夕法尼亚州立大学（PI Koslicki）和系统生物学研究所 (PI Eric多伊奇; Co-I Jared Roach）。X射线团队在翻译器的可行性方面非常有成效 评估阶段，团队为Translator带来关键的专业知识（请参阅参考资料）。 组件类型：我们建议创建、验证和集成一个自主中继代理 (ARA)叫做ARAX。ARAX将成为新的Translator架构中的中间件组件 这将大大超出原型推理工具（RTX）的能力， 我们在可行性评估阶段创建的。根据输入请求，ARAX的主 输出将是分等级的子图，并清楚地解释了分等级的依据。ARAX将利用代码 和算法，并将有一个明确的应用程序的重点领域，如下所述。 ARAX试图解决的主要问题：生物医学知识中的联系 图具有高度可变的（i）置信度（由于模糊的谓词和/或由于 知识类型的可靠性程度差异很大）和（ii）与 用户的查询。这种边缘显著性变异性会导致不正确且难以解释 这些结果共同构成了创建广泛有用的工具的重大问题， 计算机生物医学推理我们建议以明确的方式解决这个问题， 考虑到推理算法中这两种类型的边缘可变性， ARAX将提供给Translator的一系列生物医学查询类型。除了这些广泛的 能力，如项目计划中所述，我们将在ARAX中采用先进的算法 用于响应与疾病治疗相关的查询，包括（1）已知的药物重新定位 疾病，利用有关疾病发病机制的知识[1];和（2）治疗 根据症状和推定的致病遗传变异对罕见疾病提出建议。 项目实施计划：在我们的项目计划中，我们描述了五年时间轴， 在Translator中创建、验证和集成ARAX，从三个月的冲刺开始 到2020年3月中旬，ARAX的原型。该计划的主要组成部分包括：（1） 利用BioThings Explorer软件框架，使ARAX能够动态映射 化合物、蛋白质、途径、变体、表型和疾病之间的关系， Translator注册表中的知识源应用程序编程接口;（2）利用 COHD及相关Translator资源获取生物医学语义距离信息; (3)利用RTX生物医学知识的应用程序编程接口端点 图，KG 2;以及（4）实现利用出处的概率推理算法 信息和动态确定的边缘相关性分数来改进推理。我们将 系统地使用机器学习将排名分数与输出质量的衡量标准相结合。 我们团队的协作优势包括：（i）制定技术标准， Translator软件代理之间的通信（利用PI多伊奇的丰富经验 （二）开发知识图谱标准（利用PI Ramsey和PI Koslicki的专业知识）;以及（iii）导出用例小插曲， 翻译的潜力（利用Co-I Roach和PI Ramsey的专业知识）。发展 在此阶段，我们的团队将继续与其他团队和NIH利益相关者合作， 自适应、高带宽和团队边界无关的方式，如项目计划中所述。 构建拟议系统的关键挑战是（1）需要能够将“链接”在一起 工具之间的分析步骤，以及（2）需要合作制定标准， 使Translator组件能够相互作用;我们在项目计划中详细说明了这些问题。

This project would continue collaborative work within the Translator consortium by a multi-site team (“Team X-ray”) at Oregon State University (PI Stephen Ramsey) and at two partner institutions, Pennsylvania State University (PI Koslicki) and the Institute for Systems Biology (PI Eric Deutsch; Co-I Jared Roach). Team X-ray was highly productive in Translator's feasibility assessment phase and the team brings critical expertise to Translator (see Resources). Component type: We propose to create, validate, and integrate an autonomous relay agent (ARA) called ARAX . ARAX will be a middleware component in the new Translator architecture that will extend significantly beyond the capabilities of the prototype reasoning tool (RTX) that we created in the feasibility assessment phase. Depending on the input request, ARAX's main output will be ranked subgraphs with clearly explained ranking basis. ARAX will leverage code and algorithms from RTX and will have an explicit application focus area, as described below. Main problems that ARAX is trying to address: Connections within a biomedical knowledge graph have highly variable degrees of (i) confidence (due to ambiguous predicates and/or due to highly variable degrees of reliability of knowledge types) and (ii) potential relevance to the user's query. Such edge-significance variability leads to both incorrect and difficult-to-interpret results which together pose a significant problem for creating broadly useful tools for computer-based biomedical reasoning. We propose to address this problem by explicitly accounting for these two types of edge variability in the reasoning algorithms–spanning a broad range of biomedical query types–that ARAX will provide to Translator. In addition to these broad capabilities, as described in the Project Plan, we will incorporate advanced algorithms in ARAX for responding to queries relating to disease therapy, including (1) drug repositioning for known disease, leveraging knowledge about the disease’s pathogenesis [1] ; and (2) therapeutic recommendations for rare diseases based on symptoms and the putative causal genetic variant. Plan for implementation of the project: In our Project Plan we describe a five-year timeline for creating, validating, and integrating ARAX within Translator, beginning with a three-month sprint leading to a prototype of ARAX by mid-March 2020. Key components of the plan include: (1) leveraging the BioThings Explorer software framework to enable ARAX to dynamically map between compounds, proteins, pathways, variants, phenotypes, and diseases based on knowledge source application programming interfaces in the Translator registry; (2) leveraging COHD and related Translator resources to obtain biomedical semantic distance information; (3) leveraging an application programming interface endpoint for the RTX biomedical knowledge graph, KG2; and (4) implementing probabilistic reasoning algorithms leveraging provenance information and dynamically determined edge relevance scores to improve reasoning. We will systematically use machine-learning to align ranking scores with measures of output quality. Collaboration strengths of our team include (i) developing technical standards for communications between Translator software agents (leveraging PI Deutsch’s extensive past experience); (ii) developing knowledge graph standards (leveraging PI Ramsey’s and PI Koslicki’s expertise); and (iii) deriving use-case vignettes that speak to the transformative potential of Translator (leveraging Co-I Roach’s and PI Ramsey’s expertise). In the development phase, our team would continue to collaborate with other teams and with NIH stakeholders in an adaptive, high-bandwidth, and team-boundary-agnostic fashion, as detailed in the Project Plan. Key challenges to building the proposed system are (1) the need to be able to "chain" together analytical steps between tools and (2) the need for cooperative development of standards that enable Translator components to interact; we address them in detail in the Project Plan.