Supporting Knowledge Discovery in Life Sciences

大量的公开数据是人工智能在生命科学中发挥关键作用的绝佳机会。自动化方法已被证明在支持生命科学研究方面是有效的，但挖掘复杂和非结构化数据仍然是一个主要挑战。在这种背景下，我的研究计划的目标是通过简化现有知识的获取，并支持其探索，为生命科学的知识发现做出贡献。我建议通过创建联合检索和挖掘文本和非文本数据的算法来实现这一目标。寻找现有知识的生命科学家面临着严峻的挑战，比如在文档中发现实体，检索与特定主题相关的文档和数据，或者根据他们对实验的贡献来分析数据。因此，在接下来的五年里，我的研究将集中在两个目标上：1。研究从自然语言（NL）查询中联合检索各种类型文档的新模型和算法。文档的检索是生命科学的关键步骤，因为检索的结果可以用作各种任务的输入，例如管理、分类或生物网络建模。在理解NL查询和检索异构类型的文档方面存在双重挑战。目的是研究分析自然语言查询的最佳方法，以扩展它们的方向，从而触发文章、基因或蛋白质序列、相关数据库条目、实验数据等的检索。探索新的算法来发现文档中的生物实体，并将它们链接到相关的知识库。尽管在社交媒体和新闻中的实体发现和链接（EDL）方面已经做了很多工作，但生命科学仍然面临许多挑战。由于自动注释文档支持研究人员建立生物过程的计算模型，因此在生物实体发现和链接任务方面的进一步工作是必要的。EDL在基因组学中非常具有挑战性，因为生物实体通常是高度模糊的，并且通常很少有上下文可用于消除歧义。目的是研究解决EDL任务的通用方法如何适用于基因组学领域，以及如何将几个参考数据库一起使用以支持生物实体的链接和消歧。这个研究项目是跨学科的。在计算机科学领域，该程序结合了自然语言处理、信息检索、机器学习和大数据挖掘。我与基因组学研究人员的合作提供了一个涉及真实用户的具有挑战性的环境。参与人员将接受自然语言处理、应用机器学习和文本/数据挖掘方面的高级培训。发布的工作将是开源的，以便于社区重用，并转移到行业。

The massive amount of publicly available data is an amazing opportunity for artificial intelligence to play a key role in life sciences. Automatic approaches have proven to be effective in supporting life sciences research, yet mining complex and unstructured data is still a major challenge. In this context, the objective of my research program is to contribute to knowledge discovery in life sciences by easing access to existing knowledge, and supporting its exploration. I propose to reach this objective by creating algorithms to jointly retrieve and mine textual and non-textual data. Life scientists looking for existing knowledge face critical challenges such as discovering entities in documents, retrieving documents and data relevant to specific topics, or analyze data according to their contribution to experiments. Over the next five years, my research will hence focus on two objectives:O1. The investigation of new models and algorithms to jointly retrieve various types of documents from natural language (NL) queries. The retrieval of documents is a critical step for life sciences since the retrieved results can be used as input for a variety of tasks, such as curation, triage, or biological network modeling. There is a twofold challenge in understanding NL queries, and retrieving heterogeneous types of documents. The objective is to investigate the best way of analyzing NL queries to expand them in directions that trigger the retrieval of articles, gene or protein sequences, related database entries, experimental data, etc.O2. The exploration of new algorithms to discover bio-entities in documents, and link them to relevant knowledge bases. Though much work has been done toward entity discovery and linking (EDL) in social media and news, many challenges still remain in life sciences. As automatically annotated documents support researchers in building computational models of biological processes, further work on the bio-entity discovery and linking task is necessary.EDL is very challenging in genomics because bio-entities are often highly ambiguous, and little context is usually available for disambiguation. The objective is to investigate how generic approaches for solving the EDL task can be adapted to the genomics field, and how several reference databases can be used together to support linking and disambiguation of bio-entities.This research program is cross-disciplinary. In the computer science domain, the program combines natural language processing, information retrieval, machine learning, and big data mining. My collaboration with genomics researchers provides a challenging environment involving real users. Involved Highly Qualified Personal will get advanced training in natural language processing, applied machine learning, and text/data mining. The released work will be open-source in order to be easily reused by the community, and transferred to the industry.