Bayesian and machine-learning-based analysis of high-volume super-resolution microscopy data for molecular-level cell phenotyping

点击翻译按钮获取中文摘要

The use of Single Molecule Localisation Microscopy (SMLM) is booming within the biological research community. By exploiting the temporal separation of fluorescent molecules, proteins can be localised to nano-scale spatial precision allowing researchers to explore the structure and dynamics of cellular processes at unprecedented levels of detail. However, analysis of such data is still restricted to be largely qualitative and based around visual inspection of reconstructed images. Given the fantastic recent progress of Bayesian and machine-learning statistical methodology, there is the opportunity to develop a fully quantitative alternative approach. In fact, we propose that we can learn a full generative model for molecular-level cellular architecture. This approach will allow us to answer previously impossible phenotypical biological questions relating to the organisation and co-organisation of sub-cellular structures, for example: 1. "Is the clustering of a membrane protein caused by the underlying actin cytoskeleton?" (the hypothesis testing problem)2. "Does a drug change the relationship between a membrane protein and the cytoskeleton" (the classification problem)3. "Is the emergence of a cancerous cell in a normal population detectable from changes in its nanoscale architecture" (the anomaly detection problem)We postulate that complex architectural arrangements can be described by the composition and coordination of simpler morphological elements. At their most basic, these might be simple clusters or aggregates of molecules. More complicated elements could include a complex, branched fibrous structure. We have been at the forefront of developing cluster analysis methodology for SMLM. We will approach the (much) larger problem of describing cellular architecture by first building a full cluster and fibre analysis toolkit for 2 and 3D SMLM data. Next, we will acquire high-volume multi-colour SMLM data and develop Bayesian machine-learning methodology to learn the probabilistic composition and coordination of simpler morphological elements (e.g. as identified through our toolkit) to give a full generative model for cellular architecture. Once achieved, we will further develop a statistical and machine-learning layer of inference methodology to address phenotypical questions such as those given above. The main biological application we will examine during this project is the architecture of the T cell immunological synapse which is key for regulating T cell activation and hence the immune response. This opens the exciting opportunity to collaborate with clinicians and drug developers to test how therapeutic agents affect molecular-level cellular architecture.