Predicting gene function from images of cells

Abstract

This dissertation shows that biologically meaningful predictions can be made by analyzing images of cells. In particular, groups of related genes and their biological functions can be predicted using images from large gene-knockdown experiments. Our analysis methods focus on measuring individual cells in images from large gene-knockdown screens, using these measurements to classify cells according to phenotype, and scoring each gene according to how reduction in its expression affects phenotypes. To enable this approach, we introduce methods for correcting biases in cell images, segmenting individual cells in images, modeling the distribution of cells showing a phenotype of interest within a screen, scoring gene knockdowns according to their effect on a phenotype, and using existing biological knowledge to predict the underlying biological meaning of a phenotype and, by extension, the function of the genes that most strongly affect that phenotype. We repeat this analysis for multiple phenotypes, extracting for each a set of genes related through that phenotype, along with predictions for the biology of each phenotype. We apply our methods to a large gene-knockdown screen in human cells, validating it on known phenotypes as well as identifying and characterizing several new cellular phenotypes that have not been previously studied.