大规模机器学习驱动的细胞内蛋白模式变化定量分析研究

A critical issue in biological field focuses on automatically analyzing variations of the spatial distribution of proteins based on microscopy images. It can be used for detecting drug effects of drug screening, better understanding protein function and realizing disease diagnose. However, current methods are unable to accurately analyze the complicated patterns of protein distribution and capture miner but important local pattern variation. To this end, the large-scale machine learning methods are employed to analyze big dataset of microscopy images, explore both of local features and complicated interaction features for hierarchically representing protein distribution. It is expected to accurately and quantitatively capture hardly observed variations of protein distribution in a machine-vision way. In detail, our work contains three parts as follows. To start with, the idea of deep convolutional network is used to automatically learn multi-level denser features for segmenting single microtubules, resulting in single cell mask. We then recognize puncta, as well as obtaining local invariant features of puncta, which result in learning their types using a method of unsupervised classification. Finally, the multi-scale and accurate protein distribution is built based on the learning of punctum object types and the discovery of cell-level complex association features, which detailedly describes protein distribution and quantitatively evaluate the distribution variation. Our work is helpful to solve problems in cell biology and clinical research.

基于显微镜图像自动分析细胞内蛋白质分布模式及其变化是生物学领域研究的关键问题,有助于分析药物筛选中的药物效应、理解蛋白功能及相互作用、诊断疾病等。但目前的方法难以准确地分析蛋白在亚细胞的复杂分布模式及捕捉蛋白分布微小却重要的局部模式变化。为此,本项目拟利用大规模的机器学习方法分析海量荧光显微镜图像数据集，提取细胞内蛋白的局部特征及与其它组成成分的相互作用、复杂联系特征,分层次、多角度地表征蛋白分布,实现以数值的方式定量描述蛋白的分布模式及捕捉难以观察的模式变化。内容包括:1)基于深度卷积网络,自动学习多层次的稠密特征,分割图像中微管实例,提供单个细胞掩码;2)识别单个细胞内的蛋白点状对象，提取局部不变特征，利用无监督聚类方法学习对象类型；3)基于蛋白的局部特征学习和复杂联系特征发现,精准地模型化蛋白分布,定量评估模式变化。研究成果将为细胞生物学、临床医学等研究提供有效支持。

基于细胞内蛋白分布变化判断药物效应、蛋白功能、疾病诊断是生物学领域中药物筛选、亚细胞定位和组织病理学分类等研究的一个关键问题。其难点之一是依据显微镜图像如何自动分析细胞内蛋白质的分布模式及其相互作用。鉴于此，本项目基于海量荧光图像集，以机器视觉精准地捕捉蛋白分布细微却重要的变化及不同类型细胞间蛋白分布的相互作用及其差异性，具体内容包括：1) 集成MS模块和RBA模块的FUNet模型被提出用于图像分割。此深度学习模型能够捕捉图像微小但重要的语义细节，准确地将一幅图像划分为逻辑上有意义的区域，为蛋白分布建模提供单个细胞掩码； 2) 高维空间中建立基于低维流形的非参数密度模型，模拟蛋白在细胞内的分布。此模型有助于精细地、定量地评估细胞内蛋白间的相互作用及差异性；在本项目理论与应用相结合，从理论上建立不同的模型，定量地、客观地表征细胞内蛋白间的相互作用及其差异性，并应用于细胞生物学和临床医学研究。