Measuring protein dynamics using smFRET
Don C. Lamb
In a recent, international, multi-laboratory blind study, we investigated both how accurately the FRET efficiency could be determined and how reliably the presence of dynamics could be detected in proteins. This studied focus on spFRET measurements of freely diffusing molecules in solution. The results revealed highly reproducible spFRET efficiency histograms and mostly consistent determination of protein dynamics. For establishing the presence of dynamics, most groups used either the burst variance analysis or plots of the donor-lifetime versus the intensity-based FRET efficiency. We also investigated what scale of motion can be detected using this approach.
For investigating dynamic processes that are slower than the diffusion time in a confocal setup, sample immobilization is normally used. The dynamics of individual molecules can then be directly observed. In addition, by including an additional color, it is also possible to elucidate the presence of correlated motions within molecules. However, this comes at a cost in complexity and involves time-consuming evaluation of the individual traces. To reduce the effort involved in data analysis, we have developed Deep-Learning assisted, single-molecule imaging analysis (DeepLasi). DeepLasi builds upon the machine learning approaches developed for two-color FRET. It contains an ensemble of pre-trained deep neural networks (DNNs) which are designed for rapid classification, segmentation, autocorrection and extraction of state dynamics for multi-color smFRET experiments. Accepting the raw intensity traces as input, DeepLasi will categorize the traces, determine corrected FRET histograms from the data and calculate transition density plots (TDPs) on a timescale of approx. 30-100 ms per trace. We benchmarked DeepLasi using ground truth simulations and a L-shaped DNA origami structure with a tunable flexible pointer. DeepLasi accurately recovers the ground truth parameters and is in high accordance with manual evaluation of the experimental data including the commonly used Hidden Markov Model (HMM) approach for kinetic evaluation.