Hidden Markov model analysis of fluorescence blinking in fluorescently labeled DNA.

We investigate the transition processes between the emitting (ON) and non-emitting (OFF) states of fluorescent molecules using a machine-learning approach. In fluorescently labeled DNA, continuous fluorescence is observed under irradiation; however, the system occasionally transitions to a non-emitting state, often associated with a charge-separated configuration. The resulting fluorescence trajectories exhibit characteristic blinking behavior —alternating ON and OFF states— which is heavily obscured by various sources of noise, making reliable state classification challenging. Because such trajectories represent typical stochastic time-series data, advanced analytical techniques are required. In this study, we apply a hidden Markov model to extract hidden ON/OFF states from noisy fluorescence trajectories using the forward-filtered backward-blocking Gibbs sampling algorithm, and construct probability density functions of the ON- and OFF-state durations to characterize the blinking dynamics. From these distributions, the characteristic relaxation times are evaluated as 17.6 ms for the ON state and 7.8 ms for the OFF state. The relatively long OFF period indicates that the charge-separated state in the DNA-ATTO655 system is fairly stable, suggesting suppressed charge recombination. In addition, we discuss the characteristic timescale of the light absorption–emission process in the ON state in terms of the average photon count per time bin. These results provide new insights into the fluorescence dynamics of single DNA-fluorophore systems. Finally, we discuss the detailed conditions required for reliable time-series analysis in terms of the photon-count histogram shape and the time-bin width used in the trajectories.