Automated Stoichiometry Analysis of Single-Molecule Fluorescence Imaging Traces via Deep Learning

Abstract

The stoichiometry of protein complexes is precisely regulated in cells and is fundamental to protein function. Singe-molecule fluorescence imaging based photobleaching event counting is a new approach for protein stoichiometry determination under physiological conditions. Due to the interference of the high noise level and photoblinking events, accurately extracting real bleaching steps from single-molecule fluorescence traces is still a challenging task. Here, we develop a novel method of using convolutional and long-short-term memory deep learning neural network (CLDNN) for photobleaching event counting. We design the convolutional layers to accurately extract features of steplike photobleaching drops and long-short-term memory (LSTM) recurrent layers to distinguish between photobleaching and photoblinking events. Compared with traditional algorithms, CLDNN shows higher accuracy with at least 2 orders of magnitude improvement of efficiency, and it does not require user-specified parameters. We have verified our CLDNN method using experimental data from imaging of single dye-labeled molecules in vitro and epidermal growth factor receptors (EGFR) on cells. Our CLDNN method is expected to provide a new strategy to stoichiometry study and time series analysis in chemistry.