BALTIC · Machine learning based analytics for bacteria cell cycle characterization using super resolution microscopy

Machine learning based analytics for bacteria cell cycle characterization using super resolution microscopy

Super resolution microscopy provides an unprecedented insight into the molecular organization in cells. Whilst retaining molecular specificity and multiplexing, it generates images with near molecular resolution in the case of single molecule localisation microscopy (SMLM), typically 10-30nm. This gain in resolution however comes at a cost: complex sample preparation and, for the most part, loss of dynamic information (i.e., fixed cells). This research project aimed at tackling some of these limitations for the study of bacteria cell cycle (BCC), a fundamental dynamic process with clinical applications (e.g., bacterial infection, drug targeting). SMLM enables to visualize and investigate a mixed population of fixed cells (both cell wall and specific protein markers) at various cell stages at the nanoscale. We aimed at retrieving from these static images the otherwise lost BCC. Existing approaches for such "pseudo time" analysis, also called ordering, often rely heavily on pre-designed models and/or multiple protein markers. The significant drawback of such approaches is that they are heavily user-biased, cell specific and using most of the available resources the microscope has to offer. As a result, they remain constrained for the most part to cell specific studies, focusing primarily, or only, on characterizing the cell cycle, rather than providing a general framework for more complex and scalable studies. What I proposed instead is to rely on generative models, a subclass of unsupervised machine leaning tools, to infer hidden information, including the BCC, directly from the images. Generative models, including variational auto encoders (VAE) such as implemented for this work, are designed to learn how to best replicate the inputted data, in this case images of segmented bacteria. VAE have the particularity to do so by compressing the inputted images to a very low dimensional space or latent space. In this latent space, only the essential information is retained, which includes the BCC along which the collected images/cells get ordered in an automated fashion. It requires no a priori knowledge and, as such, uniquely allows for simultaneous multi cell types studies. Such data driven approach had a significant impact on our understanding of complex data sets such as genomics, as it provides a unique high information content representation of the inputted data. We foresee that it will play a similar role: enabling super resolution microscopy to become a routinely used tool to address fundamental and complex biological issues.

Текст от CORDIS, на английски · Данни: CORDIS, © Европейски съюз