Neural Network Informed Photon Filtering Reduces Artifacts in Fluorescence Correlation Spectroscopy Data

Fluorescence Correlation Spectroscopy (FCS) techniques are well-established tools to investigate molecular dynamics in confocal and super-resolution microscopy. In practice, users often need to handle a variety of sample or hardware-related artifacts, an example being peak artifacts created by bright, slow-moving clusters. Approaches to address peak artifacts exist, but measurements suffering from severe artifacts are typically non-analyzable. Here, we trained a 1-dimensional U-Net to automatically identify peak artifacts in fluorescence time-series and then analyzed the purified, non-artifactual fluctuations by time-series editing. We show that in samples with peak artifacts, the transit time and particle number distributions can be restored in simulations and validated the approach in two independent biological experiments. We propose that it is adaptable for other FCS artifacts, such as detector dropout, membrane movement, or photobleaching. In conclusion, this simulation-based, automated, open-source pipeline makes measurements analyzable which previously had to be discarded and extends every FCS user’s experimental toolbox.