Unraveling the genetic basis of cellular behaviors with deep learning and imaging-based reverse genetics - Project Summary Imaging and genomics are becoming increasingly intertwined, as multiplexed RNA FISH and multiplexed immunohistochemistry now make it possible to perform “omic” measurements while preserving spatial information. These new technologies are allowing us to create a new, descriptive understanding of normal and diseased tissues. For cell culture models, they offer the promise of measuring multiple facets of cellular behavior – ranging from cell shape to gene expression – all in the same cell. This can be done by pairing dynamic live-cell imaging data with end-point spatial genomics measurements. Such measurements could even be performed in the setting of perturbations, creating a powerful tool for mapping biological networks. In this proposal, I seek to make these methods accessible to the life science community by using large-scale data annotation, deep learning, and cloud computing to solve several outstanding cellular image analysis problems facing the spatial genomics field. I also propose to develop a simple, scalable approach to performing perturbations in imaging-based experiments. The work proposed here is three-fold. First, we will develop deep learning methods for performing whole cell segmentation in tissues as well as segmentation and lineage construction in live-cell imaging movies. To ensure these models generalize across tissues, cell lines, and imaging platforms we will undertake a large-scale data annotation effort to create a standardized collection of images that have been annotated with single cell resolution. Second, we will also develop new deep learning methods for unsupervised learning of cellular behaviors. Third, we will create a new approach to imaging-based reverse genetic screens. In this approach, we will use CRISPR-Display to create multi-color spatial patterns in cell nuclei. This will allow us to link cells and perturbations in images while minimizing the number of collected images. Libraries with 100’s of thousands of perturbations would be interpretable with only 1-2 rounds of low-magnification 4 color imaging. Achieving these high-risk, high-reward goals will constitute a transformative advance as it will empower researchers studying living systems with imaging at the resolution of a single cell with both ease and scale. Once finished, this work will place the microscope back at the center of the biologist’s toolkit and enable images to become a universal datatype for biology.
