Indexing, Mining and Modeling Spatio-Temporal Patterns of Gene Expressions

卡内基梅隆大学获得了一项资助，用于开发新的机器学习和数据挖掘方法，以发现高等真核生物复杂生物环境中基因表达的时空模式。重点将是挖掘多细胞系统的原位杂交（ISH）图像，并捕获果蝇胚胎中基因表达的细胞水平组织学背景。对于果蝇胚胎ISH图像的特征提取功能、距离度量、文本/图像数据融合方法和时空模型的设计是目前尚不完善的问题。主要的新颖之处是：(a)基于最先进的图像处理技术，如数学形态学、各种滤波器、小波、图论和概率分割等，更显著的特征提取；(b)图像/文本融合和跨模态查询的新型基于图的方法和概率模型；(c)新的潜在空间模型捕捉更高层次的“语义相似性”，而不是在可变形态背景下功能或行为相似基因的直接特征相似性；(d)卡尔曼滤波和非线性动态模型，以模拟和预测基因表达的时空演变。此外，还将建立一个ISH图像库，其中包含来自公共资源的果蝇基因组中所有已知基因的表达。结构化数据库可以通过“图像示例”或关键字（都是自动派生的）进行搜索。基因表达模式的原位杂交图像的系统分析将引起高度的兴趣。需要强大而复杂的计算机算法来分析这些数据。所提出的系统将满足这些需求，并将提供对胚胎发育以及哪些基因/蛋白质影响什么的更好理解。该项目是实现最终长期目标的必要步骤，即了解胚胎发育的分子机制，并揭示参与这一过程的基因调控网络。它还为遗传学和发育生物学教育提供了一个新的基于图像的平台。

Carnegie Mellon University is awarded a grant to develop novel machine learning and data mining methods to find spatio-temporal patterns of gene expressions in complex biological contexts of higher eukaryotic organisms. The focus will be on mining in situ hybridization (ISH) images of multi-cell systems and on capturing the cell-level histological context of gene expression in Drosophila embryos. The challenge is to design a good feature extraction function, distance measure, text/image data fusion methods, and spatio-temporal models, which to our knowledge remain very underdeveloped, for embryonic ISH images in Drosophila. The main novelties are: (a) more salient feature extraction based on state-of-the-art image processing techniques such as mathematical morphology, a variety of filters, wavelets, graph-theoretic and probabilistic segmentations, etc.; (b) novel graph-based methods and probabilistic models for image/text fusion and cross-modal querying; (c) novel latent- space models capturing higher-level "semantic similarity" rather than direct feature similarity of functionally or behaviorally similar genes in variable morphology contexts, (d) Kalman filters and non-linear dynamic models, to model and predict spatio-temporal evolutions of gene expression. Moreover a repository of ISH images will be created with the expressions of all known genes in the Drosophila genome from public sources. The structured database can be searched 'by image example' or by keyword (all automatically derived). Systematic profiling of in situ hybridization images of gene expression patterns will attract high interest. Powerful and sophisticated computer algorithms will be needed to analyze these data. The proposed system will meet these needs and will provide a better understanding of embryo development as well as which genes/proteins affect what. This project is a necessary step towards the ultimate, long term goal, of understanding the molecular mechanism of embryo development, and the unraveling of the gene regulation network involved in this process. It also offers a new image-based platform for genetics and developmental biology education.