Towards the Building of a Comprehensive Searchable Biological Experiment Database

描述（由申请人提供）： 生物医学文献的快速增长和学科生物医学研究的扩展，预示着高通量基因组科学和技术，已经淹没了科学家谁试图吸收他们的研究所需的信息。标题/摘要词搜索的广泛采用，例如非常可取的国家医学图书馆的PubMed系统，提供了自1879年医学索引起源以来生物科学家查找相关出版物的方式的第一个重大进步（Hunter和Cohen 2006）。为生物科学家开发有效的信息检索系统的重要性已经导致了世界范围内信息系统的发展（例如，Arrowsmith（Smalheiser和Swanson 1998）、BioText（赫斯特2003）、GeneWays（Friedman等2001; Rzhetsky等2004）、iHOP（Hoffmann和瓦伦西亚2005）和BioMedQA（Lee等2006 a）以及注释数据库（例如，SWISSPROT，OMIM（Hamosh等人，2005）和BIND（Alfarano等人，2005））。 然而，大多数信息系统仅以文本信息为目标，而不能提供对诸如图像（例如，数字）。在生物医学文献中，数字通常比任何其他文献都更能代表发现的“证据”。全文生物学文章几乎总是包含作为生物医学文献关键内容的图形/图像。我们对《美国国家科学院院刊》（PNAS）上的生物学文章进行了检查，发现平均每篇文章出现5.2张图像（Yu和Lee，2006 a）。生物学家需要访问图像数据来验证研究事实，并制定或测试新的研究假设。据评估，文献中经常报告的文本陈述是嘈杂的（即，包含“虚假事实”）（Krauthammer et al. 2002）。捕捉作为实验“证据”的图像来支持文本“事实”将有利于生物科学信息系统，数据库和生物科学家。 不幸的是，如果没有自动系统来组织这些图像，这些丰富的信息几乎无法访问。我们建议开发高级自然语言处理（NLP）工具来语义组织图像。我们假设与图像相关的文本在语义上包含图像内容，并且可以开发自然语言处理技术来准确地将文本与其图像相关联。此外，我们假设，图像可以通过标准生物本体指定的类别进行语义组织，并且自然语言处理方法可以准确地将本体类别分配给图像。 我们的具体目标是： 目标1：开发和评估NLP技术，用于识别与全文文章中的图像相对应的文本语句。我们将为两种类型的关联开发不同的方法。我们将首先提出基于规则和统计的方法来识别出现在全文文章中的相关文本。然后，我们将开发混合方法，将摘要中的句子与文章正文中的图像联系起来。 目标2：开发和评估NLP技术，用于将实验结果自动分类为类别（例如，免疫印迹、PCR验证等）。 我们将探索使用基于字典，基于规则，图像分类和机器学习的方法来实现这一目标。 目标3：开发和评估用于自动分配基因本体类别到实验的NLP技术，这将根据生物学特性（例如，催化活性）。我们将开发统计和机器学习方法来实现这一目标。 我们发现，在全文生物学文章中出现的大多数图像都是数字图像（Yu和Lee 2006 a），因此我们在本提案中只关注数字图像。具体目标1的交付成果将是一个有效的用户界面BioEx，生物科学家可以直接从摘要中的句子访问图像。BioEx有望改进传统的每篇文章单文档格式，自1665年第一篇科学文章出现以来，这种格式一直主导着生物科学出版物（Gross 2002）。Specific Aim 2和3的可交付成果将是开源算法和工具，可将图像准确映射到基因本体和在线协议指定的类别。这些算法和工具将增强生物科学信息检索、信息提取、摘要和问题回答。

DESCRIPTION (provided by applicant): The rapid growth of the biomedical literature and the expansion in disciplinary biomedical research, heralded by high-throughput genome sciences and technologies, have overwhelmed scientists who attempt to assimilate information necessary for their research. The widespread adoption of title/abstract word searches, such as highly desirable the National Library of Medicine's PubMed system, has provided the first major advance in the way bioscientists find relevant publications since the origin of Index Medicus in 1879 (Hunter and Cohen 2006). The importance of developing valid information retrieval systems for bioscientists has led to the development of information systems worldwide (e.g., Arrowsmith (Smalheiser and Swanson 1998), BioText (Hearst 2003), GeneWays (Friedman et al. 2001; Rzhetsky et al. 2004), iHOP (Hoffmann and Valencia 2005), and BioMedQA (Lee et al. 2006a), and annotated databases (e.g., SWISSPROT, OMIM (Hamosh et al. 2005) and BIND (Alfarano et al. 2005)). However, most of information systems target only text information and fail to provide access to other important data such as images (e.g., figures). More than any other documentation, figures usually represent the "evidence" of discovery in the biomedical literature. Full-text biological articles nearly always incorporate figures/images that are the crucial content of the biomedical literature. Our examination of biological articles in the Proceedings of the National Academy of Sciences (PNAS) revealed the occurrence of 5.2 images per article on average (Yu and Lee 2006a). Biologists need to access image data to validate research facts and to formulate or to test novel research hypotheses. It has been evaluated that textual statements reported in literature frequently are noisy (i.e., containing "false facts") (Krauthammer et al. 2002). Capturing images that are experimental "evidence" to support the textual "fact" will benefit bioscience information systems, databases, and bioscientists. Unfortunately, this wealth of information remains virtually inaccessible without automatic systems to organize these images. We propose the development of advanced natural language processing (NLP) tools to semantically organize images. We hypothesize that text that associated with images semantically entails the image content and natural language processing techniques can be developed to accurately associate the text to their images. Furthermore, we hypothesize that images can be semantically organized by categories specified by standard biological ontology, and that natural language processing approaches can accurately assign the ontological categories to images. Our specific aims are: Aim 1: To develop and evaluate NLP techniques for identifying textual statements that correspond to images in full-text articles. We will develop different approaches for two types of the associations. We will first propose rule-based and statistical approaches to identify the associated text that appears in the full-text articles. We will then develop hybrid approaches to link sentences in abstracts to images in the body of the articles. Aim 2: To develop and evaluate NLP techniques for automatic classification of experimental results into categories (e.g., Western-Blot, PCR verification, etc) specified in the experimental protocol Protocol-Online. We will explore the use of dictionary-based, rule-based, image classification, and machine-learning approaches for accomplishing this aim. Aim 3: To develop and evaluate NLP techniques for automatic assignment of Gene Ontology categories to experiments, which will provide a knowledge-based organization of experiments according to biological properties (e.g., catalytic activity). We will develop statistical and machine-learning approaches for accomplishing this aim. We found that most of the images that appear in full-text biological articles are figure images (Yu and Lee 2006a) and we therefore focus on figure images only in this proposal. The deliverable of Specific Aim 1 will be an effective user-interface BioEx from which bioscientists can access images directly from sentences in the abstracts. BioEx has the promise of improvement over the traditional single-document-per-article format that has dominated bioscience publications since the first scientific article appeared in 1665 (Gross 2002). The deliverables of Specific Aim 2 and 3 will be open-source algorithms and tools that accurately map images to categories specified by the Gene Ontology and the Protocol Online. Those algorithms and tools will enhance bioscience information retrieval, information extraction, summarization, and question answering.