Deep learning-enhanced single-molecule spectrum imaging

Abstract

Fluorescence is widely used in biological imaging and biosensing. Rich information can be revealed from the fluorescence spectrum of fluorescent molecules, such as pH, viscosity and polarity of the molecule’s environment, and distance between two FRET molecules. However, constructing the fluorescence spectrum of a single fluorescent molecule typically requires a significant number of photons, which can suffer from photobleaching and, therefore, limit its potential applications. Here, we propose a deep learning-enhanced single-molecule spectrum imaging method (SpecGAN) for improving the single-molecule spectrum imaging efficiency. In SpecGAN, the photon flux required to extract a single-molecule fluorescence spectrum can be reduced by 100 times, which enables two orders of magnitude higher temporal resolution compared to the conventional single-molecule spectrometer. The concept of SpecGAN was validated through numerical simulation and single Nile Red molecule spectrum imaging on support lipid bilayers (SLBs). With SpecGAN, the super-resolution spectrum image of the COS-7 membrane can be reconstructed with merely 12 000 frames of single-molecule localization images, which is almost half of the previously reported frame count for spectrally resolved super-resolution imaging. The low photon flux requirement and high temporal resolution of SpecGAN make it a promising tool for investigating the molecular spectrum dynamics related to biological functions or biomolecule interactions.