Abstract
Single-molecule fluorescence resonance energy transfer (smFRET) has emerged as a pivotal technique for probing biomolecular dynamics over time at nanometer scales. Quantitative analyses of smFRET time traces remain challenging due to confounding factors such as low signal-to-noise ratios, photophysical effects such as bleaching and blinking, and the complexity of modeling the underlying biomolecular states and kinetics. The dynamic distance information shaping the smFRET trace powerfully uncovers even transient conformational changes in single biomolecules both at or far from equilibrium, relying on trace idealization to identify specific interconverting states. Conventional trace idealization methods based on hidden Markov models (HMMs) require substantial a priori knowledge of the system under study, manual intervention, and assumptions about the number of states and transition probabilities. Here, we present a deep learning framework using long short-term memory (LSTM) to automate the trace idealization, termed Kin-SiM. Our approach employs neural networks pretrained on simulated data to learn high-order correlations in the multidimensional FRET trajectories. Without user input of Markovian assumptions, the trained LSTM networks directly idealize the FRET traces to extract the number of underlying biomolecular states, their interstate dynamics, and associated kinetic parameters. On benchmark smFRET data sets, Kin-SiM achieves a performance similar to conventional HMM-based methods but with less hands-on time and lower risk of bias. We further systematically evaluate the key training factors that affect network performance to define the correct hyperparameter tuning for applying deep neural networks to smFRET data analyses.