Supervised multi-frame dual-channel denoising enables long-term single-molecule FRET under extremely low photon budget

Abstract

Camera-based single-molecule techniques have emerged as crucial tools in revolutionizing the understanding of biochemical and cellular processes due to their ability to capture dynamic processes with high precision, high-throughput capabilities, and methodological maturity. However, the stringent requirement in photon number per frame and the limited number of photons emitted by each fluorophore before photobleaching pose a challenge to achieving both high temporal resolution and long observation times. In this work, we introduce MUFFLE, a supervised deep-learning denoising method that enables single-molecule FRET with up to 10-fold reduction in photon requirement per frame. In practice, MUFFLE extends the total number of observation frames by a factor of 10 or more, greatly relieving the trade-off between temporal resolution and observation length and allowing for long-term measurements even without the need for oxygen scavenging systems and triplet state quenchers.

Fig. 1. Training and validation of MUFFLE.

a Optical diagram of training data acquisition. Based on a TIRF microscope, two dual-view modules were added to achieve four channels. b Schematic of training data. Light yellow (donor-bright) and light red (acceptor-bright) served as the ground-truth images that closely resembled the ones collected under our normal single-molecule FRET measurements, while dark yellow (donor-dark) and dark red (acceptor-dark) were their corresponding weak noisy attenuated images. c Composition of training data. Training images were collected under 24 different conditions, which were the combination of four attenuation levels, three exposure times, and two Mg²⁺ concentrations to achieve sufficient coverage of attenuation and dynamic ranges. d Network structure and training process. The network recovers clean signals by exploring spatial-temporal redundancy, as well as cross-channel redundancy. e Examples of ground-truth, attenuated, and reconstructed images. Images of HJ in 500 mM Mg²⁺ were collected under 10% attenuation level and 100 ms exposure time. Attenuated images were linearly enhanced by 10 folds for reference. Recon. stands for reconstruction. f Single-molecule fluorescence trajectories of the same molecule extracted from the ground truth and reconstructed images, respectively, as shown in (e). g FRET distributions of Cy3/Cy5 dual-labeled HJ in 1 mM Mg²⁺ (top panel) and 500 mM Mg²⁺ (bottom panel) determined from bright ground-truth and reconstructed images using different F_n values. Corresponding attenuated images were collected under 10% attenuation level and 33 ms exposure time. h Comparison of dynamics of HJ in 500 mM Mg²⁺ determined from bright ground-truth images and their corresponding reconstructed images using different F_n values. We used HaMMy software to calculate FRET states and their dwell times from the traces. For each set of traces under the same attenuation-exposure conditions, We performed frequency statistics on the dwell times obtained and calculated the mean dwell time by fitting with the 1-component exponential decay model. Data are presented as mean values ± standard error. The sample size is detailed in the Statistics and Reproducibility section in Methods. Each sub-figure contains dwell times of the high or low FRET states determined under 12 combinations of exposure times and attenuation levels. Source data are provided as a Source Data file.

Fig. 2. Application of the MUFFLE.

a–d Extended observation time under low laser power. a Typical single-molecule FRET trace of HJ at 1 mM Mg²⁺ with 33 ms exposure time under regular 31.8 mW or 3.6 mW laser powers. Comparison of FRET distributions (b) and dynamics (c) of HJ determined from MUFFLE-reconstructed movies under a series of power reductions to the ones determined from raw movies under regular laser power. Experiments were performed under a series of [Mg²⁺]. FRET states and their dwell times were calculated from traces using HaMMy software. For each specific laser power and [Mg²⁺] condition, frequency statistics of dwell times were performed, and the mean dwell time was calculated by fitting it to the 1-component exponential decay model. Data are presented as mean values ± standard error. The sample size is detailed in the Statistics and Reproducibility section in “Methods”. d Extended observation time achieved by MUFFLE. Durations were measured on HJ samples in 1 mM Mg²⁺ conditions, with four repeated measurements. Data are presented as mean values ± standard deviation. e–g Improvement of temporal resolution without affecting observation time. e 0.5s-long single-molecule traces from the movie acquired under a laser power of 31.8 mW and exposure time of 33 ms, 10 ms, and 3 ms. 10 mM Mg²⁺ was used. f Improved temporal resolution without sacrificing total observation time. Observation times were measured on HJ samples under 30 mM Mg²⁺ conditions, with 3 repeats for the 33 ms group, 6 repeats for the 10 ms group, and 8 repeats for the 3 ms group, ensuring comparable total frame counts across groups. Data are presented as mean values ± standard deviation. g FRET distributions of samples in (e). h–k Long-term single-molecule FRET measurements without the imaging buffer. h Single-molecule traces acquired from the movie without an oxygen-scavenging system and triplet state quenchers under regular 31.8 mw or 3.6 mW laser powers. i Prolonged observation time without imaging buffer achieved by MUFFLE. Observation times were measured on HJ samples under 500 mM Mg²⁺ conditions, with four repeated measurements. Data are presented as mean values ± standard deviation. Comparison of dynamics (j) and FRET distribution (k) of HJ determined from MUFFLE-reconstructed movies without the imaging buffer to the ones determined from raw movies under regular laser power with the imaging buffer. Dwell times were calculated using the same method as in (c). Data are presented as mean values ± standard error. The sample size is detailed in the Statistics and Reproducibility section in “Methods”. Source data are provided as a Source Data file.