Quantifying cellular structures at the nanoscale with graph neural networks and super resolution microscopy : Applying machine learning to improve super resolution microscopy localisations

Sammanfattning

Single-Molecule Localization Microscopy (SMLM) produces high-resolution point cloud data of protein distributions at the nanoscale. While some molecular structures are well-known, many remain uncharacterised, posing a challenge for classification and analysis. This thesis explores the use of auto-encoders to learn a latent representation of SMLM data that can distinguish between known structures and generalize to previously unseen ones. Unlike conventional classification methods, this approach does not assume a fixed set of categories. Instead, it focuses on capturing meaningful geometric features from spatial distributions. The resulting model offers a flexible and compact representation that enables both structure differentiation and reconstruction, contributing to a more robust understanding of nanoscale biological organization.