III: Small: Deep Learning for Gene Expression Pattern Image Analysis

生物图像信息学是计算生物学中的一个新兴前沿，因为迫切需要从人工检查图像转向计算分析，以加速科学发现。传统方法通常使用浅层机器学习模型，在该模型中计算手工制作的图像表示并在模型构建中使用。这些方法在很大程度上依赖于数据和问题的先验知识来计算适当的图像表示。在深度学习方法在图像相关领域取得成功的推动下，本项目的目标是开发用于生物图像自动表征学习的高级深度学习模型。该项目还促进了新课程和实验室基础设施的开发，以吸引研究生、本科生和高中生，重点是那些来自代表性不足群体的学生。具体地说，该项目专注于分析果蝇和老鼠的时空基因表达模式图像。关键挑战在于如何捕捉生物学问题的内在结构，以及如何在人工标记的小型生物学数据集上进行有效的模型训练。这个项目开发了多实例、多任务、分层和正规化的深度学习模型，以纳入生物学问题的结构。多实例模型和多任务模型分别捕捉输入和输出之间的复杂关系。层次化和正则化模型对问题结构进行了显式编码，并使结果具有可解释性。此外，还开发了转移和无监督学习方法，以便在小标签数据集上进行有效的模型训练。这些是通过跨多个域集成已标记和未标记的数据集来实现的。总而言之，该项目有望产生一套先进的深度学习方法，用于高效和有效地分析生物图像。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Biological image informatics is an emerging frontier in computational biology, as there is an urgent need to move beyond the manual inspection of images to computational analysis for accelerating scientific discoveries. Conventional methods commonly employ shallow machine learning models in which handcrafted image representations are computed and used in model construction. These approaches heavily rely on prior knowledge on the data and problems to compute appropriate image representations. Motivated by the recent success of deep learning methods in image-related domains, the objective of this project is to develop advanced deep learning models for automated representation learning from biological images. This project also facilitates the development of new courses and laboratory infrastructure for attracting graduate, undergraduate, and high school students, with an emphasis on those from underrepresented groups.Specifically, this project focuses on the analysis of spatiotemporal gene expression pattern images in fruit fly and mouse. The key challenges lie in how to capture the intrinsic structures of biological problems and how to enable effective model training on small, manually labeled biological data sets. This project develops multi-instance, multi-task, hierarchical, and regularized deep learning models for incorporating the structures of biological problems. The multi-instance and multi-task models capture the complex relationships among inputs and outputs, respectively. The hierarchical and regularized models explicitly encode problem structures and make the results interpretable. In addition, transfer and unsupervised learning methods are developed to enable effective model training on small labeled data sets. These are achieved by integrating both labeled and unlabeled data sets across multiple domains. Altogether, this project is expected to result in a set of advanced deep learning methods for the efficient and effective analysis of biological images.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.