Improving the Resolution of Single-Molecule Localization Microscopy by Leveraging Spatiotemporal Information

Abstract

Single-molecule localization microscopy (SMLM) is a cutting-edge super-resolution technique that enhances microscopy image resolution by exploiting the temporal separation of fluorophore localizations across multiple frames. In SMLM, several images are captured from a region of interest, with only a subset of fluorophores activated in each frame. This selective activation of fluorophores ensures well-separated observations, which significantly improves spatial resolution. The final high- resolution image is then constructed by integrating the localization data across all frames. However, the inherent stochastic blinking behaviour of fluorophores introduces significant challenges, as individual fluorophores can blink multiple times, appearing at slightly different positions due to measurement errors. These repeated observations are often misinterpreted as separate fluorophores, leading to overcounting artifacts that can compromise both the accuracy of molecule localization and the quality of quantitative analyses. Various methods have been proposed to address these blinking artifacts, such as applying fixed spatial thresholds or clustering observations within spatial dimensions. However, these approaches often fail to account for the temporal dynamics of fluorophore blinking, which can provide critical insights, particularly in densely populated regions. In this work, we present a robust algorithm designed to mitigate blinking artifacts and accurately localize true molecular positions, thereby enabling high-fidelity image reconstructions. Our approach effectively integrates both spatial and temporal information, with a particular focus on capturing long-term temporal dependencies inherent in fluorophore blinking behaviour. By leveraging the temporal dimension, our method provides more precise reconstructions and improves localization accuracy, especially in regions with overlapping or closely spaced molecules. We have trained Long Short-Term Memory (LSTM) models on simulated SMLM data generated from short experimental datasets, demonstrating the effectiveness of our approach in these limited time frames. Moving forward, we aim to extend this methodology to real-world, long-duration experiments. To achieve this, we plan to utilize Large Language Models (LLMs), which are particularly well-suited for capturing long-range temporal dependencies, thus further improving the resolution and accuracy of SMLM image reconstructions in more complex experimental scenarios.