Computer vision for the analysis of intra-cellular RNA localization patterns

Overview & Synopsis

RNAs do not necessarily have a random distribution in the cell. Specific localization patterns have been observed and such asymmetric distributions are of functional importance for the spatial regulation of gene expression. Current methods to study intra-cellular RNA localization use single molecule FISH (smFISH) protocols or its multiplexing variants. Each RNA molecule is targeted with fluorescent probes and can be imaged by fluorescence microscopy. These techniques enable single cell level analyses and can be scaled to thousands of transcripts. However, computational methods to quantify RNA localization suffer different flaws, such as the consistency of spot detection between experiments. A complete analysis pipeline typically requires the use of different frameworks and the development of in-house pieces of code, making smFISH analysis difficult to scale and reproduce. In this thesis, I present a new version of FISH-quant, a comprehensive computational framework to process smFISH images. It provides improved modules for the most common operations needed for smFISH analysis such as RNA detection, nucleus and cell segmentation as well as feature engineering, all in the same framework, designed to be flexible and scalable. In addition, I investigate the use of point cloud models directly trained from simulations to address the classification of RNA localization patterns, without the need to compute hand-crafted features. Lastly, I apply FISH-quant to experimental datasets in biological studies that include several dozens of transcripts, analyzed through 100,000 individual cells. For this, I design a classification pipeline to discriminate several generic RNA localization patterns. This allows us to discover translation factories, a new mechanism of spatial control of gene expression. Third, we focus on the centrosomal and protrusion patterns to provide a statistical description of several genes of interest. Altogether, these studies reveal regulation mechanisms more complex than expected with various spatial and temporal dynamics.