Title:	Flow matching for super resolution microscopy
DNr:	Berzelius-2025-45
Project Type:	LiU Berzelius
Principal Investigator:	Juliette Griffie <juliette.griffie@dbb.su.se>
Affiliation:	Stockholms universitet
Duration:	2025-02-12 – 2025-09-01
Classification:	10610
Keywords:

Abstract

Single Molecule Localization Microscopy (SMLM) is a powerful imaging technique that enables nanoscale resolution beyond the diffraction limit. However, despite its high spatial precision, there remains significant uncertainty in reconstructed structures due to noise, drift, and sample dynamics. Additionally, live imaging remains a challenge, as high photon yield and minimal phototoxicity must be balanced against temporal resolution. Addressing these limitations requires innovative computational methods that can refine localization precision, extract meaningful spatial relationships, and reconstruct continuous representations of molecular structures from discrete localizations. In this project, we propose to leverage Flow Matching, a machine learning framework inspired by generative modeling, to enhance both static and dynamic analyses of SMLM data. Flow Matching learns vector field dynamics by directly matching observed data distributions to a continuous flow, providing a data-driven approach to infer high-resolution structures and underlying molecular movements. Our approach builds upon SE(3)-equivariant Graph Neural Networks (GNNs), which respect the rotational, translational, and reflectional symmetries inherent in molecular organization. This enables efficient learning and prediction of spatial and temporal molecular distributions without the need for explicit trajectory linking. The first proof of concept will focus on improving resolution in SMLM data by reconstructing finer molecular details from sparse and noisy localizations. Beyond this, the project aims to explore different applications of Flow Matching and GNNs in SMLM, providing a versatile framework for analyzing nanoscale biological processes. High-performance computing resources are essential for this project due to the computational intensity of training SE(3)-equivariant GNNs on large-scale molecular datasets. The training and evaluation phases involve computing pairwise interactions across millions of localizations, requiring parallelized processing on GPU clusters. Access to a supercomputing facility will enable us to train deeper network architectures, perform extensive hyperparameter tuning, and test our model on diverse SMLM datasets. By obtaining supercomputer resources, we aim to develop a cutting-edge computational tool that integrates physics-informed deep learning with experimental biophysics. This work will provide a foundation for future studies in nanoscale dynamics modeling, benefiting researchers in both computational and experimental biology. The outcome of this project will be open-source, allowing broader adoption in the scientific community to enhance data analysis capabilities in SMLM and beyond.