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.
Introduction
Single-molecule fluorescence resonance energy transfer (smFRET) is a cornerstone of modern biophysics, allowing researchers to observe molecular dynamics in real-time. However, experiments are fundamentally constrained by a "photon budget": fluorophores emit a finite number of photons before irreversible photobleaching. This creates a restrictive trade-off between temporal resolution and observation duration. High-speed imaging requires high laser power, which bleaches molecules rapidly, while low-power imaging results in a signal-to-noise ratio (SNR) too low for accurate analysis. Conventional mitigation strategies rely on oxygen scavenging systems (OSS), but these chemical additives can be toxic to live cells or interfere with biological dynamics. This work presents a computational solution to bypass these physical limitations.
Methods
The authors developed MUFFLE (Multi-frame dUal-channel Fusion network for single-molecule FLuorEscence), a supervised deep-learning framework specifically architected for smFRET data. MUFFLE leverages three types of redundancy:
- Spatial Redundancy: A U-Net-style convolutional neural network captures the features of point-like fluorescent spots.
- Temporal Redundancy: Bi-directional Convolutional Long Short-Term Memory (bi-ConvLSTM) modules analyze 15 neighboring frames to reconstruct the signal, assuming relatively mild molecular motion between frames.
- Cross-Channel Fusion: Crucially, MUFFLE exploits the physical anti-correlation between FRET donor and acceptor channels, using information from one to help denoise the other.
The model was trained on a dataset of over 800,000 frames of DNA Holliday Junctions (HJ). Training data was acquired using a custom "dual-view" microscope that simultaneously captured high-photon "bright" ground-truth images and low-photon "dark" versions (attenuated to 4%–20% brightness) of the same scene.
Results
MUFFLE demonstrates the ability to maintain accurate FRET efficiency distributions and kinetic dwell times even with a 10-fold reduction in photons per frame. By operating at lower laser powers and computationally recovering the signal, the authors achieved up to a 28-fold extension in total observation time for individual molecules.
The framework successfully resolved fast molecular dynamics at 3 ms per frame—a speed that typically causes near-instant bleaching. Furthermore, MUFFLE enabled reliable long-term FRET measurements in "naked" buffers without oxygen scavengers. Interestingly, these OSS-free experiments revealed that HJ dynamics are slightly slower when chemical photostabilizers are removed, highlighting the importance of minimally invasive imaging.
Discussion
MUFFLE shifts the burden of signal detection from hardware to software, effectively making fluorophores appear more stable. While the current implementation is optimized for immobilized molecules and EMCCD cameras, it provides a foundation for future applications in live-cell imaging where low-photon budgets and the avoidance of toxic additives are critical. Future work will focus on adapting the model for moving molecules and sCMOS camera architectures through transfer learning.
Summary updated