Thorsten Wohland

Thorsten Wohland1,2, Wai Hoh Tang1,3, Shao Ren Sim1, Daniel Ying Kia Aik2, Ashwin Venkata Subba Nelanuthala1, Adrian Roellin3

1- Department of Biological Sciences and NUS Centre for Bio-Imaging Sciences, National University of Singapore. 2- Department of Chemistry, National University of Singapore. 3- Department of Statistics and Data Science, National University of Singapore

Imaging Fluorescence Correlation Spectroscopy (Imaging FCS) has developed into a useful tool to observe 2D and 3D biomolecular dynamics and interactions in live systems. However, it faces several challenges that need to be addressed to make it an easy-to-use tool. First, the sheer amount of data recorded by array detectors, requires new efficient approaches to data analysis. Second, new illumination geometries require new solutions for data fitting as often analytic solutions cannot be found for the theoretical correlation functions that are required for data fitting. Third, measurement times are on the order of minutes, precluding the reliable detection of changes that occur below that limit.

Here, we address these issues by employing GPUs to accelerate real-time evaluations and deep learning to improve Imaging FCS time resolution and data fitting. GPUs accelerate data evaluation between 10-40 times depending on the evaluation modality if more than ~100 pixels are evaluated. Below that limit of 100 pixels, the overhead of transferring data to the GPUs makes them less efficient than the inbuilt CPUs. In deep learning we developed two convolutional neural networks (CNNs), FCSNet and ImFCSNet, which use either correlation functions or raw intensity traces as input, respectively. Both networks are trained on simulated, synthetic data, as it is difficult to obtain ground truth data for FCS over wide parameter ranges. These approaches allow us to a) use arbitrary experimental geometries, as analytic fit models are no longer required, b) reduce the amount of data required by about one order of magnitude to achieve the same precision as conventional nonlinear least-squares data fitting, and c) reduce the evaluation time by about two orders of magnitude. Finally, we will present several applications of these computational approaches in cells and on data recorded on total internal reflection and light sheet microscopes.