Single-molecule fluorescence resonance energy transfer (FRET) technology represents a pioneering biophysical tool for investigating the dynamics and interactions of biomolecules at the nanoscale. Conventional long-term single-molecule FRET measurements were performed using total internal reflection fluorescence (TIRF) microscopy, which provides insights into molecular dynamics through the monitoring of individual fluorescent molecules and the changes in their fluorescence intensity. However, the vast quantity of data generated by this process represents a substantial challenge for classifying. Traditional manual or semi-automated methods are not only inefficient but also susceptible to introducing subjective bias and rely heavily on the experimenter's experience. 

In this context, the application of deep learning models has the potential to significantly reduce the incidence of human errors and enhance the speed and accuracy of data processing. However, models trained on limited datasets are unlikely to be applicable to all users, due to the existence of unique experimental conditions and designs. Furthermore, the application of single-molecule FRET in non-equilibrium system is becoming increasingly prevalent. However, our preliminary examination of current models on dynamically emerging fluorescence signals (defined as those whose fluorescent signals do not exist at the beginning of the acquisition process but emerge during the acquisition due to the interaction of diffusing labelled molecules with molecules immobilized on the surface) has revealed that they are not yet fully fit for purpose. It is therefore imperative to develop a universal method that can be applied in a variety of scenarios with minimal modification.

In light of the above, we put forth a Deep lEarning Based fRagmentatIon approach for Single-molecule fluorescence event identification (DEBRIS). DEBRIS represents a pioneering approach to accurately categorizing and analyzing four distinct classes of single-molecule fluorescence events using a single trained AI model. DEBRIS is capable of automatic classification of different types of traces by classifying short-fragment traces instead of long-duration ones, setting it apart from other methods.

This inspiration comes from the Tetris game and LEGO blocks, in which a limited number of building blocks can be assembled to create a multitude of structures. Similarly, the entirety of long-term single-molecule traces can be constituted by a finite set of short fragments, which are characterized by specific local features. DEBRIS model was trained based on simulated datasets comprising seven classes of two-channel short fragment intensity-time traces, designed to simulate local features commonly observed in single-molecule two-color fluorescence experiments (Fig. 1a). Subsequently, DEBRIS employs a sliding window to predict local features frame-by-frame, thereby converting intensity-time traces into pattern-time traces (Fig. 1b). Then, traces are classified into groups according to user-defined criteria, and single-molecule events are identified. The combination of local feature prediction in short fragments, the sliding window method, and user-defined criteria enables DEBRIS to adapt to two-channel fluorescence traces with arbitrary modes and arbitrary lengths, thereby providing a robust tool for the automated identification of single-molecule fluorescence events.

Figure 1. DEBRIS training set construction and general process.

To evaluate the classification performance of DEBRIS on two-color fluorescent trajectories, we employed five single-molecule two-color FRET datasets (Fig. 2). Upon utilizing the expert classified results as a benchmark, the FRET distributions obtained by DEBRIS exhibited a 99% similarity to those obtained manually. This result demonstrates that DEBRIS is comparable to manual classification in terms of accuracy, while offering significant advantages in terms of efficiency, objectivity and reproducibility.

Figure 2. DEBRIS classification performance evaluation.

We employed the DEBRIS to achieve the first accurate capture of fluorescence signals that emerge dynamically. Moreover, DEBRIS model, which was trained using two-color traces, can also be extended to be applicable to one-color single-molecule fluorescence traces. For a detailed account of the applications of DEBRIS in the aforementioned scenarios, please refer to our published paper.

By accurately identifying local features of single-molecule fluorescence traces and allowing flexible adjustment of the classification criteria according to the experimental design, DEBRIS achieves accurate identification of steady and dynamic single-molecule fluorescence signals under two-color and one-color experimental conditions without modifying the neural network structure. This innovative approach provides an efficient data analysis tool for the application of single-molecule fluorescence technology, thereby greatly expanding its potential in biophysical research.

Paper link:

https://doi.org/10.1038/s42003-024-07122-4

Links to relevant papers from our group:

https://doi.org/10.1093/nar/gkae604

https://doi.org/10.1038/s41586-024-07486-x

https://doi.org/10.1038/s41586-024-07295-2

https://doi.org/10.1038/s41467-023-37233-1

https://doi.org/10.1021/jacs.2c12635