VI. TR&D4 - Abstract The overall goal of this project is to develop tools for making maximal use of the information in biological images to enable the construction of predictive, multiscale models of structure and dynamics at the subcellular and cellular level. The tools will be especially useful for studies of how cell organization is created and maintained and how that organization differs from cell type to cell type and during disease. While existing software primarily provides descriptions of images, the focus of this project is on construction of generative models of cell organization. Generative models are learned from a collection of images and are capable of producing new images that are statistically equivalent to the images used for training. These models have distinct advantages over discriminative or descriptive approaches. They attempt to make use of all information in images, rather than to just extract selected descriptors or features. Further, while features are not useful for comparing and communicating results between different laboratories due to their dependence upon the specifics of image acquisition, generative models capture the underlying reality that gave rise to images and can therefore be compared across different microscopes and laboratories. They are also combinable and reusable, in that models can be linked together to make predictions about new relationships, and models for organelle shape and distribution learned for one cell type can be provisionally extended to new cell types. Work during the prior funding led to the development of extensive generative model capabilities that were incorporated into the open source CellOrganizer system. We propose to build upon this work to build new capabilities for constructing models that consider the extensive interrelationships between organelles and structures in cells, and for modeling the dynamics of proteins and organelles. In conjunction with TR&D3, we will also develop new methods for using images to constrain estimation of the affinities between components of a biological system. Lastly, we will develop new approaches for constructing models from both electron and fluorescence microscope images. The proposed work makes use of best available methods in machine learning and computer vision, including advanced inference methods and convolutional neural nets (so called ?deep learning? methods). The work builds on the extensive progress that has been made under what was Aim 1 of TR&D3 in the prior funding period, which resulted in eleven publications that acknowledged P41 support.
Showing the most recent 10 out of 180 publications
