Learning representations from heterogeneous data for digital health

分子生物标记物，包括基因、miRNAs、lncRNAs、CircRNAs等，不仅有助于理解人类复杂疾病的发病机制，而且有助于其诊断、预后、治疗和药物开发。新兴的高通量技术产生了丰富的异质数据，如序列数据、表达数据、语义数据和一些经过实验验证的关联(网络)数据。虽然现有的机器学习方法可以应用于从这些异质数据中进行各种预测任务，但仍然存在一些挑战和局限性，以及一些有待克服的障碍，包括缺乏对异质数据的集成、缺乏平衡的训练数据、缺乏模型解释或可理解性。在这项拟议的研究计划中，我的长期目标是开发新的人工智能模型和算法，以便有效和高效地从异质数据中学习生物标记物、药物和疾病的强大综合表示，同时将它们应用于各种预测任务，这些任务是精准医学数字健康的关键。为了实现我的长期目标，这项提案设计了三个具体目标如下。目标1：开发改进的机器学习模型，通过整合异质数据来学习生物标记物、药物和疾病的表示；目标2：开发深度神经网络模型，用于从异质数据中学习生物标记物、药物和疾病的表示。目标3：开发可理解的机器学习模型，该模型可以学习生物标志物、药物和疾病的表示，以获得可解释的预测结果，以便人类能够理解和解释结果。我们开发的模型和算法将被应用于几个预测任务，如基因-疾病关联、miRNA-疾病关联、CircRNA-疾病关联、lncRNA-疾病关联或DUG重新定位。这一拟议的研究计划将在人工智能和数字健康方面取得进展，最终将有助于处理、分析、理解和利用不同种类的健康数据。这些方法也可以被其他领域(自然语言处理、图像分析等)的研究人员用于指导他们的方法，并通过我们提出的知识翻译策略发现新的知识。通过这项拟议的研究计划，三名博士和几名硕士。培养本科生学习人工智能、数字健康、软件开发等领域的高级知识，以及数字健康等人工智能不可或缺的领域的问题解决技能。此外，对人工智能和数字健康的高需求，再加上我的合作研究方法，为我的学生提供了大量在这个令人兴奋的多学科培训环境中工作的机会。

Molecular biomarkers, including genes, miRNAs, LncRNAs, circRNAs, etc., not only help understand the mechanisms of complex human diseases, but also help their diagnosis, prognosis, treatment and drug development. The emerging high throughput technologies generate copious heterogeneous data such as sequence data, expression data, semantic data, and some experimentally verified association(network) data. Although existing machine learning methods can be applied for various prediction tasks from such heterogeneous data, there are some challenges and limitations, as well as some obstacles yet to overcome, including lack of integration of heterogeneous data, lack of balanced training data, and lack of model explanation or intelligibility. In this proposed research program my long-term goal is to develop the novel artificial intelligence models and algorithms for effectively and efficiently learning strong comprehensive representations of biomarkers, drugs and diseases from heterogeneous data while applying them to various prediction tasks, which are key to digital health for precision medicine. To achieve my long-term goal, three specific objectives are designed in this proposal as follows. Objective 1: Developing improved machine learning models for learning representations of biomarkers, drugs and diseases by integrating heterogeneous data; Objective 2: Developing deep neural network models for learning representations of biomarkers, drugs and diseases from heterogeneous data. Objective 3: Developing intelligible machine learning models that can learn the representations of biomarkers, drugs and diseases for interpretable prediction results, so that humans can understand and interpret the outcomes. Our developed models and algorithms will be applied to several prediction tasks such as gene-disease associations, miRNA-disease associations, circRNA-disease associations, LncRNA-disease associations or dug repositioning. This proposed research program will make advances in both artificial intelligence and digital health that will ultimately aid in processing, analyzing, understanding, and exploiting heterogeneous health data. The methods to be developed could also be used by researchers in other areas (natural language processing, image analytics, just to name a few) in guiding their methods and discovering new knowledge through our proposed knowledge translation strategies. Through this proposed research program, three Ph.D., and several M.Sc. and undergraduate students will be trained to gain advanced knowledge in artificial intelligence, digital health, and software development, as well as the problem solving skills in digital health and other fields in which artificial intelligence is indispensable. In addition, high demand in artificial intelligence and digital health combined with my collaborative research approach provides my students with numerous opportunities for working in this exciting multidisciplinary training environment.