Enhanced quantitative accuracy in single molecule imaging: a fidelity-guided denoising approach

Enayati, Aref, Baird, Hannah M., Weyland, Mathias S., Petri-Fink, Alke, Lattuada, Marco, Jamieson, W. David, Füchslin, Rudolf Marcel and Castell, Oliver ORCID: https://orcid.org/0000-0002-6059-8062 2026. Enhanced quantitative accuracy in single molecule imaging: a fidelity-guided denoising approach. Biophysical Journal 10.1016/j.bpj.2026.10.023

Full text not available from this repository.

Abstract

Single-molecule total internal reflection fluorescence (TIRF) microscopy provides unparalleled insight into biological processes, but reaches fundamental limits due to the challenges of imaging individual fluorophores, where limited photon counts mean signals often border background noise. Practically, trade-offs exist between emission rate, time resolution, and observation lifetime. Experimentalists increasingly seek higher time resolution and longer observational times, demanding accurate quantification of fluorescence intensity and super-resolved positional information. Consequently, single-molecule detection must operate at low signal-to-noise ratios, often requiring significant human oversight to optimize. Applying machine learning to improve signal-to-noise would reduce these analytical bottlenecks, allowing researchers to work at lower photon counts to further push observational limits. However, while existing computational methods enhance traditional fluorescent microscopy, single-molecule analysis presents a unique challenge: the benefits of denoising often come at the unacceptable cost of distorting the vital positional and brightness measurements. We address this critical trade-off with an intelligent deep learning-based denoising framework guided by a fidelity constraint, preventing the algorithm from creating artifacts or distorting molecular features. Our training objective is informed by the calibrated Poisson-Gaussian noise model of the detector. This cleans images, easing spot detection while preserving absolute intensity and positional data. Additionally, we present a version in which a reinforcement learning agent sets the strength of this constraint automatically during training, removing the need for manual parameter exploration. On simulated data, our approach raises the single-molecule detection F1-Score from 0.10 to 0.63 under the most challenging noise conditions tested, while preserving molecular features. Finally, on experimental data of fluorescently-labelled, membrane-interacting peptides, our method reduces errors in measuring molecular position by a factor of two to five and a half, and changes the measured brightness four to thirteen times less, compared to leading denoising methods.

Item Type:	Article
Date Type:	Published Online
Status:	In Press
Schools:	Schools > Pharmacy
Publisher:	Cell Press
ISSN:	0006-3495
Date of Acceptance:	26 October 2026
Last Modified:	10 Nov 2026 14:03
URI:	https://orca.cardiff.ac.uk/id/eprint/190753

Actions (repository staff only)

Edit Item