AGILE: A Cloud Approach to Automatic Gene Expression Pattern Recognition and Annotation Over Large-Scale Images

现代生物医学研究大量使用大型数据集。云计算正在成为一种具有成本效益的解决方案，它可以按需提供虚拟计算机和存储磁盘，以有效地存储和处理大量数据，而无需大量的前期成本。尽管取得了一些进展，但云计算在生物医学研究中的应用仍处于非常早期的阶段。关于如何最好地利用云来加速大规模生物医学应用，存在着各种各样的问题。特别是，生物医学应用程序是否可以直接迁移到云端而无需任何修改？如何开发基于云的生物医学应用？云中应用程序的性能和成本是什么？性能和成本是否可接受？我们是否有最佳的方法来将应用程序的性能和总体成本保持在云中可接受的范围内？该项目将通过解决上述问题，为真实的生物医学数据密集型任务开发一种云方法，用于大规模图像数据的有效基因表达模式识别和注释。选择这项任务主要是因为它在生物医学研究中的重要性。这种密集的数据分析任务在生物医学科学中越来越普遍。这个特殊的任务涉及小鼠胚胎的发育解剖学：通过使用解剖学注释来识别与胚胎中的发育和生理功能相关的基因相互作用和网络是非常有趣的。基因表达模式识别和注释代表用小鼠发育的解剖学术语标记胚胎图像。如果一个图像被标记了一个术语，这意味着相应的解剖成分显示了该基因的表达。目前，这项任务主要由领域专家手动完成。然而，由于大量数据的可用性，手动注释是昂贵且耗时的。此外，手动注释还可能产生由人类注释者引入的跨图像的标签的不一致性，因为它被证明是高度主观的。为了缓解手动注释的问题，我们采用了数据挖掘技术来自动识别胚胎图像中的解剖成分，并使用所提供的术语对图像进行注释。由于这项工作涉及使用非常大规模的图像，我们打算利用云计算来解决这项工作中的海量数据问题。预计这项项目的成功完成将为访问和利用云计算技术分析基于图像的大规模生物医学数据提供典型范例。该提案的一个重要且新颖的方面是，将解决限制云计算在生物医学应用中更广泛使用的主要问题。该工作的理论部分旨在提供（1）基于云的实用用户友好的生物医学数据挖掘工具，用于有效的基因表达模式识别和注释，以及（2）一套标准服务（例如图像处理算法、数据挖掘算法）和用于性能增强和成本降低的新颖的自动数据重用机制，其可以被重复使用并插入到类似的生物医学应用中。

Modern biomedical research makes significant use of large datasets. Cloud computing is emerging as a cost-effective solution by providing virtual computers and storage disks on demand to store and process massive data efficiently without large upfront costs.Despite some progress made, the use of cloud computing in the biomedical research is still at the very early stage. There exist various concerns on how to best utilise the cloud for accelerating large-scale biomedical applications. Especially, can a biomedical application be directly migrated to the cloud without requiring any modification? How to develop a cloud-based biomedical application? What are the performance and the cost of an application in the cloud? Are the performance and the cost acceptable? Do we have optimal methods to keep both performance and the overall cost of applications within the acceptable range in the cloud?This project will develop a cloud approach for a real biomedical data intensive task for effective gene expression pattern recognition and annotation over large-scale image data through addressing the concerns above. This task is chosen largely for its importance in the biomedical research. This type of intensive data-analysis task is increasingly common in the biomedical sciences. This particular task concerns developmental anatomy of mouse embryo: it is of great interest to identify gene interactions and networks that are associated with developmental and physiological functions in the embryo by using anatomical annotation. The gene expression pattern recognition and annotation represents labelling embryo images with anatomical terms for mouse development. If an image is tagged with a term, it means the corresponding anatomical component shows expression of that gene. Currently, this task is mainly taken manually by domain experts. However, with the availability of the vast amount of data, a manual annotation is expensive and time consuming. Additionally, the manual annotation may also produce the inconsistency of labels across images introduced by the human annotators as it proves to be highly subjective. To alleviate issues with the manual annotation, we have employed data mining techniques to automatically identify an anatomical component in the embryo image and annotate the image using the provided terms. As this task involves the use of very large-scale images, we intend to exploit cloud computing for this task to address the massive data problems.It is expected that the successful completion of this project will provide a typical exemplar for accessing and exploiting cloud computing technologies to analyse large-scale image-based biomedical data. An important, and novel, aspect of this proposal is that the major concerns that limit the more widespread use of cloud computing for biomedical applications will be addressed. The theoretical component of the work aims to provide (1) a practical user-friendly biomedical data-mining tool based on the cloud for effective gene expression pattern recognition and annotation and (2) a set of standard services (e.g. image processing algorithms, data mining algorithms) and a novel automatic data reuse mechanism for performance enhancement and cost reduction, which can be reused and plugged into the class of similar biomedical applications.