FENNEC: photon-level deep learning for classifying bursts in diffusion-based single-molecule FRET

Abstract

Single-molecule Forster resonance energy transfer (smFRET) reports on biomolecular conformational dynamics by measuring distance changes between donor and acceptor fluorophores. In diffusion-based smFRET, however, the detection of genuine conformational exchange is routinely confounded by photophysical artefacts, notably acceptor photobleaching and blinking, which produce similar burst-level signatures. Many existing methods for resolving conformational dynamics and dye photophysics rely on fitting kinetic models with a fixed number of states, which is typically unknown. Here we present FENNEC (Fluorescence Event Neural Network for Evaluating and Classifying bursts), a dilated convolutional neural network for diffusion-based smFRET data that simultaneously detects conformational dynamics, acceptor photobleaching, and acceptor blinking within individual bursts, directly from raw photon arrival times. FENNEC is trained entirely on simulated data, and requires no experimental data with assigned labels for training. Crucially, the dynamics classification is independent of the number of underlying states, and therefore provides an analysis and filtering method that complements established methods that extract the number of states and their kinetics. FENNEC can identify a high-confidence subset of static and dynamic bursts, while ambiguous bursts can be excluded or set aside for further analysis. Applied to a dynamic DNA hairpin, FENNEC recovers the expected population distributions. Together, these results provide proof of principle that a classifier trained on simulated photon-level data can identify conformational dynamics and photophysical artefacts in experimental smFRET data, and we invite further evaluation on a range of instruments and systems. FENNEC is freely available at https://github.com/jacrossley/FENNEC.