Abstract
The combination of image analysis and superresolution microscopy methods allows for unprecedented insight into the organization of macromolecular assemblies in cells. Advances in deep learning (DL)-based object recognition enable the automated processing of large amounts of data, resulting in high accuracy through averaging. However, while the analysis of highly symmetric structures of constant size allows for a resolution approaching the dimensions of structural biology, DL-based image recognition may introduce bias. This prohibits the development of readouts for processes that involve significant changes in size or shape of amorphous macromolecular complexes. Here we address this problem by using changes of septin ring structures in single molecule localization-based superresolution microscopy data as a paradigm. We identify potential sources of bias resulting from different training approaches by rigorous testing of trained models using real or simulated data covering a wide range of possible results. In a quantitative comparison of our models, we find that a trade-off exists between measurement accuracy and the range of recognized phenotypes. Using our thus verified models, we find that septin ring size can be explained by the number of subunits they are assembled from alone. Furthermore, we provide a new experimental system for the investigation of septin polymerization.