Multi-omic single-cell, electronic health record, and biomedical knowledge graph data integration using interpretable deep learning approaches

随着电子健康记录（EHR）和单细胞多组测序等基因组技术的成熟，我们正处于使用一种类型的数据（例如单细胞基因组）来理解另一种类型的数据（反之亦然）的融合边缘。单细胞基因组和异质电子病历数据的多模式数据集成将导致对人类复杂表型的新的机制见解。拟议的国际合作的目标是建立一种机器学习（ML）方法来整合电子病历和基因组数据。我们的方法将通过利用来自大型系统（如通用医学语言系统（UMLS）、基因本体（GO）、字符串数据库等）的现有综合管理的生物医学知识图谱，解决多个研究数据集成中常见的数据缺失问题，字符串数据库由已知的网络间和网络内连接组成，用于表型和基因顶点。通过使用深度学习方法，我们将从知识图中为每个基因或每种表型计算一个数字向量，然后可用于模拟实际的EHR和基因表达数据。我们将通过它们在基因和表型上的有限组潜在主题分布的主题混合隶属关系来表示这些数值向量。在同一主题下，对高概率的顶级基因和表型进行检查，将揭示控制表型合并症的基因调控网络模块。此外，当整合来自多个来源的数据时，我们将考虑多模态数据分布和混淆效应。此次合作将结合康奈尔大学的王飞博士和麦吉尔大学的李跃博士的专业知识，解决这个高度跨学科的研究项目，弥合机器学习、基因组学和健康信息学等学科之间的差距。我们将建立一个智能数字平台，使研究人员和健康相关从业人员不仅可以探索合理的表型和基因组联系，还可以整合他们自己的数据。该项目的成功将加速加拿大使用ML方法的综合分析框架的范式转变。

As electronic health record (EHR) and genomic technologies such as single-cell multi-omic sequencing becoming mature, we are on the verge of convergence in terms of using one type of data (e.g., single-cell genomic) to understand the other and vice versa. Multi-modal data integration of the single-cell genomic and heterogeneous EHR data will lead to new mechanistic insights into the human complex phenotypes. The objective of the proposed international collaboration is to build a machine learning (ML) method to integrate EHR and genomic data. Our approach will address missing data problem that are common in data integration from multiple studies by leveraging existing comprehensively curated biomedical knowledge graph from large systems like the Universal Medical Language System (UMLS), Gene Ontology (GO), String database, which consists of known inter-network and intra-network connections for the phenotype and gene vertices. By using a deep learning approach, we will compute for each gene or each phenotype a numerical vector from the knowledge graph, which can then be used to model the actual EHR and gene expression data. We will represent these numerical vectors by their topic mixture memberships for a finite set of latent topic distributions over the genes and phenotypes. Examining top genes and phenotypes with high probabilities under the same topic will reveal gene regulatory network modules that govern the phenotypic comorbidities. Furthermore, we will account for multi-modal data distribution and confounding effects when integrating data from multiple sources. The collaboration will combine domain expertise from Dr. Fei Wang from Cornell University and Dr. Yue Li from McGill University in tackling the highly interdisciplinary research project, bridging gaps across disciplines spanning across ML, genomics, and health informatics. We will build an intelligent digital platform that enables researchers and health-related practitioners to not only explore plausible phenotypic and genomic connections but also integrate their own data. The success of the project will accelerate the paradigm shift of integrative analytical frameworks using ML approaches in Canada.