smDeepFLUOR: single-molecule deep learning fluorescence classification Authors: Jinseob Lee, Byungju Kim, Gayun Bu, Muhammad Tehseen, Vlad-Stefan Raducanu, Samir M. Hamdan, Jong-Bong Lee Published: 2026 ABSTRACT Fluorescence intensity variation has long served as a primary readout for monitoring biological events. However, single-fluorophore signals arising from distinct molecular events often exhibit similar intensity profiles, making further classification challenging using conventional methods. In this study, we introduce smDeepFLUOR, a deep learning–based framework that resolves seemingly homogeneous spatiotemporal fluorescence signals by uncovering latent features imperceptible to conventional analyses. By leveraging a three-dimensional convolutional neural network trained on image sequences captured over 7 × 7 × 10 voxel windows, smDeepFLUOR reliably distinguishes specific from nonspecific protein binding, even across different experimental days, with an accuracy of up to 97%. Remarkably, smDeepFLUOR also captures real-time DNA synthesis kinetics by identifying subtle changes in the spatial distance between the fluorophore and the 3’ end of nascent DNA, a feature undetectable by traditional methods. These classifications were achieved without incorporating explicit physical rules or engineered features, implying the presence of intrinsic, previously unrecognized differences in emission patterns. This approach significantly extends the analytical capabilities of single-molecule fluorescence imaging and opens new avenues for minimally labeled and label-free protein activities. Introduction Pixelated single-molecule fluorescence signals can offer unparalleled insights into the dynamics of individual biomolecules by analyzing their intensity profiles1. These signals can emerge from interactions between fluorophores and surrounding environments, leading to abrupt shifts in fluorescence signals via mechanisms such as protein-induced fluorescence enhancement or quenching (PIFE/PIFQ)2 and nucleic acid-induced enhancement (NAIFE)3,4 or quenching (NAIFQ)5. Although it remains challenging to generalize fluorescence responses across diverse environments, single-fluorophore signals not only reflect the presence of a biomolecule but also exhibit inherent variations depending on their local surroundings6. Under the premise that subtle environmental changes can induce distinct states in fluorophores, we investigated whether advanced deep-learning methods could classify these states, even when the average fluorescence intensity appears indistinguishable. Although fluorescence signals captured by an EMCCD camera exhibit Gaussian-like intensity profiles over only a few pixels, previous studies on single-molecule emission patterns have revealed that these intensity distributions contain more intricate information than intuitively expected7,8. Modern deep-neural networks9–11, particularly convolutional neural networks (CNNs)12, have shown robust capabilities in extracting essential features and achieving remarkable classification accuracy. Recent advancements in applying deep-learning to fluorescence imaging data have enhanced traditional methods, enabling the differentiation of features that were previously considered indistinguishable13–16. In this study, we developed an innovative approach called single-molecule deep-learning fluorescence classification (smDeepFLUOR), designed to classify distinct states of proteins or DNA labeled with a single fluorophore. This method uses ten frames of a two-dimensional image consisting of 7 × 7 pixels. The three-dimensional (3D) CNN architecture of smDeepFLUOR processes a 7 × 7 × 10 image of fluorescence intensity distribution, rather than relying on a single averaged intensity, thereby capturing comprehensive spatiotemporal data (Fig. 1 and Supplementary Fig. 1).Fig. 1Schematic of smDeepFLUOR.Fluorophore-conjugated biomolecules are imaged using TIRF microscopy. The smDeepFLUOR estimates the spatiotemporal fluorescence signals within a 7 × 7 × 10 voxel window using a 3D convolutional neural network. “px” denotes pixels. Schematic of smDeepFLUOR. Fluorophore-conjugated biomolecules are imaged using TIRF microscopy. The smDeepFLUOR estimates the spatiotemporal fluorescence signals within a 7 × 7 × 10 voxel window using a 3D convolutional neural network. “px” denotes pixels. We demonstrate that smDeepFLUOR can distinguish the specific binding of poly(A)-binding protein (PABP) to poly(A) RNA tails (93% accuracy) from its nonspecific binding to the surface of an imaging chamber (97% accuracy). The trained model reliably classifies the specific (85% accuracy) and nonspecific (90% accuracy) binding of PABP using data obtained from independent experiments performed on multiple days. Surprisingly, smDeepFLUOR can also detect subtle changes in emission patterns arising from the proximity of the 3′ end of nascent DNA to Cy3 labeled at the 5′ end, whose differences are indistinguishable using conventional methods. This smDeepFLUOR capability allows us to monitor the kinetics of DNA synthesis with a single fluorophore attached to DNA, even in the absence of discernible PIFE as DNA polymerase approaches the fluorophore. We show the capabilities of smDeepFLUOR in revealing previously unrecognized intrinsic information in single-molecule fluorescence signals, enabling a new class of fluorescence-based assays that operate beyond the sensitivity of conventional intensity-based methods. Results Classification of specific interactions by smDeepFLUOR We first investigated whether a fluorophore directly conjugated to proteins could sense changes in the local environment arising from specific or nonspecific protein interactions. To assess smDeepFLUOR to detect binding specificity, we prepared a human poly(A)-binding protein fused to a single mNeonGreen (mNG) fluorescent protein. The mNG-PABP specifically binds to the poly(A) region of a DNA/RNA partial duplex immobilized on a surface passivated with a mixture of polyethylene glycol (PEG) and PEG-Biotin (Fig. 2a). Specific-binding datasets were collected by cropping mNG-PABP signals colocalized with Cy5 signals at a DNA/RNA partial duplex, while mNG signals in the absence of the DNA/RNA partial duplex were collected for nonspecific mNG-PABP binding to the surface, imaged by total internal reflection fluorescence (TIRF) microscopy. Of the 40,000 image sets collected in the same sample, 80% were used to train smDeepFLUOR. Testing with the remaining 20% showed that smDeepFLUOR achieved an identification accuracy of approximately 97 and 93% for classifying nonspecific surface-bound and specific poly(A)-bound mNG-PABP, respectively (Fig. 2a). Notably, despite indistinguishable pixelated image sequences and average fluorescence intensities between the two classes, smDeepFLUOR successfully distinguished them, revealing the presence of subtle yet discriminative spatiotemporal features.Fig. 2Identification of specific interactions between poly(A) tail and PABP.a Fluorescence signals from mNG-PABP were used for training (green star in red box). The intensity of mNG-PABP did not significantly differ between specific and nonspecific binding (histogram) (n = 62,901 units for nonspecific binding; n = 51,134 for specific binding). The dataset was split into 80% training and 20% testing sets. The smDeepFLUOR trained on the full dataset achieved 97% accuracy for identifying nonspecific surface binding and 93% accuracy for identifying actual specific binding to the poly(A) tail. b The smDeepFLUOR achieved 90% accuracy for nonspecific binding and 85% accuracy for specific binding on unseen data from independent experiments conducted on different days. Data are presented as mean ± SD from Nexp = 2 independent experiments, where each independent experiment denotes a separately prepared imaging dataset acquired on a different day not used for model training. c Experimental design for classifying protein binding using mNG-PABP. Fluorescence signals from Cy5-labeled DNA/RNA partial duplex were used for training (red star in red box). Despite similar Cy5 intensities between protein-bound and unbound states (histogram) (n = 42,118 units for unbinding; n = 30,110 for binding), the model achieved 83% accuracy for the bound state and 88% for the unbound state. d Fluorescence signals from Cy5-DNA/RNA partial duplex were used for training. The model predicted the binding state with 82% accuracy and the unbinding state with 84% accuracy. Identification of specific interactions between poly(A) tail and PABP. a Fluorescence signals from mNG-PABP were used for training (green star in red box). The intensity of mNG-PABP did not significantly differ between specific and nonspecific binding (histogram) (n = 62,901 units for nonspecific binding; n = 51,134 for specific binding). The dataset was split into 80% training and 20% testing sets. The smDeepFLUOR trained on the full dataset achieved 97% accuracy for identifying nonspecific surface binding and 93% accuracy for identifying actual specific binding to the poly(A) tail. b The smDeepFLUOR achieved 90% accuracy for nonspecific binding and 85% accuracy for specific binding on unseen data from independent experiments conducted on different days. Data are presented as mean ± SD from Nexp = 2 independent experiments, where each independent experiment denotes a separately prepared imaging dataset acquired on a different day not used for model training. c Experimental design for classifying protein binding using mNG-PABP. Fluorescence signals from Cy5-labeled DNA/RNA partial duplex were used for training (red star in red box). Despite similar Cy5 intensities between protein-bound and unbound states (histogram) (n = 42,118 units for unbinding; n = 30,110 for binding), the model achieved 83% accuracy for the bound state and 88% for the unbound state. d Fluorescence signals from Cy5-DNA/RNA partial duplex were used for training. The model predicted the binding state with 82% accuracy and the unbinding state with 84% accuracy. Furthermore, to evaluate whether smDeepFLUOR could reliably analyze images from independent experiments that share the common feature of being the same protein but have entirely different background characteristics17, we tested its performance on datasets collected on different days. When trained on the dataset from Fig. 2b, smDeepFLUOR achieved an identification accuracy of 89.7 ± 4.7% (mean ± SD) for nonspecific binding to the surface and 84.6 ± 2.1% for specific binding to poly(A) in images obtained from these separate experiments (Fig. 2b). These results suggest that even when trained on data from a limited number of experiments, smDeepFLUOR demonstrated accuracy levels comparable to those achieved with the full dataset, highlighting its robustness. Notably, smDeepFLUOR can maintain consistent performance in independent experiments. To address the challenge of recognizing protein binding when changes in fluorescence intensity fall outside the detectable range of PIFE or PIFQ, we first trained smDeepFLUOR using mNG-PABP binding events with a single Cy5 dye attached to a DNA/RNA partial duplex (Fig. 2c). This time, we tested whether a fluorophore not attached to the protein could detect environmental changes induced by a nearby binding protein. For training, PABP-bound datasets were prepared by extracting Cy5 signals colocalized with mNG-PABP while Cy5 signals in the absence of mNG-PABP were used for PABP-unbound datasets (Fig. 2c). Remarkably, smDeepFLUOR achieved an identification accuracy of about 86%, with 83% accuracy for unbinding events and 88% for binding events to the template, for classifying protein binding events based on Cy5 signals on the template (Fig. 2c). These results suggest that smDeepFLUOR may enable the identification of specific protein binding to the template, even in regimes where PIFE and PIFQ are not applicable. Classification of real-time protein binding without the need for labeled proteins In our previous studies, we found that PABP could occupy over 90% of poly(A) templates in the presence of 10 nM PABP molecules, owing to its low dissociation constant (Kd) of about 1.0 nM18. This high proportion of unlabeled PABP-bound data enabled us to effectively train smDeepFLUOR using unlabeled PABP molecules19 (Supplementary Fig. 2). The resulting identification accuracy for unlabeled PABP binding reached 83% (Fig. 2d). When compared to smDeepFLUOR trained on mNG-PABP binding data based on colocalization with Cy5 (Fig. 2c), its identification accuracy matches well that of mNG-PABP (Supplementary Fig. 3). Taken together, this result suggests the potential for broad application of this approach in training with unlabeled proteins. Since labeling the protein of interest is often a significant challenge in performing fluorescence imaging, including new single-molecule experiments, we anticipate that smDeepFLUOR for analyzing unlabeled proteins will facilitate the development of innovative experimental concepts. To validate smDeepFLUOR to classify single protein binding events in real-time, we further analyzed individual trajectories by examining the colocalization of PABP and its template: mNG-PABP was introduced into an imaging chamber where Cy5-labeled DNA/RNA partial duplexes were immobilized, unbound proteins were washed away, and then both fluorescence channels were sequentially imaged (Fig. 3a). The model, trained on non-labeled PABP (Fig. 2d), accurately detected the timing of mNG-PABP appearance corresponding to binding to a DNA/RNA partial duplex, even when the Cy5 intensity remained unchanged. We showed that smDeepFLUOR successfully identified PABP binding states by recognizing the spatiotemporal overlap between mNG-PABP and Cy5-DNA/RNA partial duplex, although it had not been trained with fluorescently labeled PABP (Fig. 3c). These results demonstrate the model’s capacity to monitor real-time protein binding at the single-molecule level, without relying on significant fluorescence intensity fluctuations and labeled proteins for training.Fig. 3Real-time classification of PABP binding to poly(A) tail.a Schematic illustration of the colocalization experiment showing mNG-PABP binding to a Cy5-DNA/RNA partial duplex. b Representative intensity traces of a single mNG-PABP molecule (green) bound to a single Cy5-DNA/RNA duplex (gray), overlaid with model predictions of protein binding (red) and unbinding (blue) events. Using mNG-PABP as a true indicator of protein binding, we found that the moment of PABP binding could be accurately identified by predicting the Cy5 signal on the DNA/RNA partial duplex using the model. c Averaged prediction results of mNG-PABP traces bound to Cy5-labeled DNA/RNA partial duplexes (n = 124). Real-time classification of PABP binding to poly(A) tail. a Schematic illustration of the colocalization experiment showing mNG-PABP binding to a Cy5-DNA/RNA partial duplex. b Representative intensity traces of a single mNG-PABP molecule (green) bound to a single Cy5-DNA/RNA duplex (gray), overlaid with model predictions of protein binding (red) and unbinding (blue) events. Using mNG-PABP as a true indicator of protein binding, we found that the moment of PABP binding could be accurately identified by predicting the Cy5 signal on the DNA/RNA partial duplex using the model. c Averaged prediction results of mNG-PABP traces bound to Cy5-labeled DNA/RNA partial duplexes (n = 124). Classification of two distinct DNA constructs To assess whether smDeepFLUOR could classify DNA structural states based not only on protein-induced changes but also on environmental variations introduced by the DNA itself, we trained the model on two DNA constructs labeled with Cy3 at the 5′ end: a 32-nt (nucleotide) gap DNA (serving as a DNA synthesis template) and a nick DNA (a product of DNA synthesis) (Fig. 4a). This setup enabled us to investigate whether the structural transition from a gap to a nick influences the local fluorophore environment in the absence of additional binding proteins. We reasoned that this structural transition could induce sufficiently distinct local fluorophore environments to be detectable by smDeepFLUOR.Fig. 4Classification of two different DNA constructs, gap and nick DNA.a Schematics of Cy3-labeled 32-nt gap DNA and nick DNA. Fluorescence signals from Cy3 were used as input for model training (green star in red box). b In the absence of PCNA, RFC, and dNTPs, smDeepFLUOR achieved 90% accuracy in classifying gap DNA and 91% accuracy for nick DNA. c smDeepFLUOR achieved 78% accuracy for gap DNA and 80% accuracy for nick DNA on unseen data from independent experiments acquired on different days. Data are presented as mean ± SD from Nexp = 2 independent experiments. d DNA constructs for various gap lengths (Nnt = 17 nt, 15 nt, 12 nt, 9 nt, 6 nt, and 3 nt). e Prediction results for these test DNAs were obtained using the model trained on 32-nt gap DNA and nick DNA. The 17-nt, 15-nt, 12-nt, and 9-nt gap DNAs were predicted as nick DNA with accuracies of 17, 31, 20, and 19%, respectively. In contrast, the misclassification rate increased sharply for shorter gaps: the 6-nt gap and 3-nt gap DNAs were predicted as nick DNA with 84 and 87% accuracy, respectively. Classification of two different DNA constructs, gap and nick DNA. a Schematics of Cy3-labeled 32-nt gap DNA and nick DNA. Fluorescence signals from Cy3 were used as input for model training (green star in red box). b In the absence of PCNA, RFC, and dNTPs, smDeepFLUOR achieved 90% accuracy in classifying gap DNA and 91% accuracy for nick DNA. c smDeepFLUOR achieved 78% accuracy for gap DNA and 80% accuracy for nick DNA on unseen data from independent experiments acquired on different days. Data are presented as mean ± SD from Nexp = 2 independent experiments. d DNA constructs for various gap lengths (Nnt = 17 nt, 15 nt, 12 nt, 9 nt, 6 nt, and 3 nt). e Prediction results for these test DNAs were obtained using the model trained on 32-nt gap DNA and nick DNA. The 17-nt, 15-nt, 12-nt, and 9-nt gap DNAs were predicted as nick DNA with accuracies of 17, 31, 20, and 19%, respectively. In contrast, the misclassification rate increased sharply for shorter gaps: the 6-nt gap and 3-nt gap DNAs were predicted as nick DNA with 84 and 87% accuracy, respectively. Using Cy3 intensity profiles obtained via TIRF microscopy, smDeepFLUOR was trained on Cy3-labeled 32-nt gap DNA and nick DNA. It achieved 90% accuracy in identifying the gap DNA and 91% accuracy for the nick DNA, resulting in an overall classification accuracy of 90.5% in distinguishing between the two structural states (Fig. 4b). To evaluate whether the model could generalize to mixed populations, the model was tested on 1:1 and 9:1 mixtures of the two DNA constructs. The predicted distributions closely matched the expected ratios of 45:55 and 92:8, respectively (Supplementary Fig. 4). These results suggest that smDeepFLUOR can reliably detect structural transitions between 32-nt gap DNA and nick DNA with minimal labeling. Importantly, we applied the model, trained solely on 32-nt gap and nick DNA, to DNA molecules containing shorter gaps of 17, 15, 12, 9, 6, and 3 nt. The 17-, 15-, 12-, and 9-nt gap constructs were consistently classified as 32-nt gap DNA with accuracies ranging from 69 to 83% (Fig. 4e). In contrast, the 6-nt and 3-nt gap DNA molecules were predominantly predicted as nick DNA, achieving 84 and 87% accuracy, respectively (Fig. 4e). Interestingly, these findings align with previous reports of NAIFE occurring within a 10-nt range: short single-stranded segments can measurably alter the fluorophore’s local environment, even in the absence of proteins5. While the change in fluorescence intensity in our DNA structures may be too subtle to be detected by conventional methods, it is likely that smDeepFLUOR can recognize such variations through its sensitivity to spatiotemporal emission patterns. Classification of the DNA synthesis To assess smDeepFLUOR’s ability to detect the transition of gap DNA to nick DNA by DNA polymerase, we first trained the model using gap and nick DNA molecules in the presence of proliferating cell nuclear antigen (PCNA), replication factor C (RFC), and dNTPs, key components required for DNA synthesis, but in the absence of DNA polymerase. The classification accuracies under these conditions were slightly reduced, with 81% for gap DNA and 82% for nick DNA (Fig. 5a), suggesting that DNA synthesis components (PCNA, RFC, and dNTPs) other than DNA polymerase have minimal influence on smDeepFLUOR classification performance.Fig. 5DNA synthesis classification monitored with smDeepFLUOR.a In the presence of PCNA, RFC, and dNTPs, smDeepFLUOR achieved classification accuracies of 81% for gap DNA and 82% for nick DNA. b Schematic of the polymerase activity assay illustrating the gap-to-nick transition of DNA in the presence of DNA polymerases. c Predicted nick-to-gap DNA ratios under various conditions. The smDeepFLUOR predicted 62% nick formation by DNA polymerase δ at 5 min, and 37 and 67% by DNA polymerase η at 5 and 30 min, respectively. Without PCNA, predicted nick ratios were 20% for DNA polymerase δ at 5 min, and 40 and 56% for DNA polymerase η at 5 and 30 min, respectively. d Representative real-time trace of smDeepFLUOR predictions showing the measured gap-to-nick transition time (τ), including the association time of DNA polymerases. e, f The mean gap-to-nick transition time was 1.7 ± 0.3 s for DNA polymerase δ and 9.9 ± 2.0 s for DNA polymerase η, as shown in the resulting histograms. DNA synthesis classification monitored with smDeepFLUOR. a In the presence of PCNA, RFC, and dNTPs, smDeepFLUOR achieved classification accuracies of 81% for gap DNA and 82% for nick DNA. b Schematic of the polymerase activity assay illustrating the gap-to-nick transition of DNA in the presence of DNA polymerases. c Predicted nick-to-gap DNA ratios under various conditions. The smDeepFLUOR predicted 62% nick formation by DNA polymerase δ at 5 min, and 37 and 67% by DNA polymerase η at 5 and 30 min, respectively. Without PCNA, predicted nick ratios were 20% for DNA polymerase δ at 5 min, and 40 and 56% for DNA polymerase η at 5 and 30 min, respectively. d Representative real-time trace of smDeepFLUOR predictions showing the measured gap-to-nick transition time (τ), including the association time of DNA polymerases. e, f The mean gap-to-nick transition time was 1.7 ± 0.3 s for DNA polymerase δ and 9.9 ± 2.0 s for DNA polymerase η, as shown in the resulting histograms. We then applied smDeepFLUOR to analyze the activities of two distinct human DNA polymerases: polymerase δ20 and polymerase η21 (Fig. 5b). DNA polymerase δ exhibits a stronger PCNA-dependent association and a higher DNA synthesis rate than polymerase η, whose activity is relatively independent of PCNA22. Despite an identical average intensity of the probe Cy3 on DNA (Supplementary Fig. 5), smDeepFLUOR successfully detected their distinct biochemical behaviors in the classification of gap and nick DNA. In the presence of all components, including polymerase δ, the model predicted that 62% of gap DNA transitioned to nick DNA within 5 min after reaction initiation. In contrast, in the absence of PCNA, only 20% of gap DNA was classified as nick DNA (Fig. 5c). As expected, smDeepFLUOR also detected the gap-to-nick transition by DNA polymerase η, albeit with a sixfold slower rate than polymerase δ at 30 min reaction time (Fig. 5c). Moreover, the smDeepFLUOR classification indicated that polymerase η-mediated DNA synthesis occurred largely independent of PCNA, as the predicted nick ratios were nearly identical regardless of PCNA presence at the same reaction time (Fig. 5c). The results demonstrate that smDeepFLUOR accurately classifies gap and nick DNA, in strong agreement with previously reported biochemical data22–24. Furthermore, real-time prediction trajectories revealed that the nick transition time for gap (Fig. 5d; Supplementary Fig. 6) after 30 nM polymerase δ injection (τδ = 1.7 ± 0.3 s; Fig. 5e) was significantly faster than that of 30 nM polymerase η (τη = 9.9 ± 2.0 s; Fig. 5f). The nick transition time (τ) reflects both the association and synthesis phases, but it is primarily governed by the association step, as the synthesis rates, 100 s−1 for polymerase δ and 67 s−1 for polymerase η22, are rapid for a 32-nt synthesis. Our observation is consistent with previous reports22–24, indicating that the association rate of polymerase η to PCNA-loaded DNA is generally slower and weaker than that of DNA polymerase δ. Together, these results suggest that smDeepFLUOR can serve as a highly sensitive and quantitative tool for monitoring DNA synthesis or excision in a minimally labeled system containing only a single fluorophore on the template strand. Discussion Single-molecule fluorescence imaging offers deep insights into biomolecular interactions. However, its interpretability has long been constrained by the limited dimensionality of fluorescence signals, which are typically confined to fluorescence intensity or lifetime. In this study, we present smDeepFLUOR, a deep learning–based classification framework designed to extract high-content spatiotemporal features from the fluorescence emission dynamics of a single fluorophore, enabling discrimination beyond the capabilities of conventional intensity-based analyses. Remarkably, smDeepFLUOR maintains robust performance without the need for engineered features or multiplexed labeling, even when the mean fluorescence intensities between distinct molecular states were indistinguishable by conventional methods. This underscores the existence of subtle but learnable information embedded within the raw fluorescence dynamics. Recent deep learning approaches in single-molecule imaging have primarily focused on exploiting dynamic photophysical properties of fluorophores, such as blinking kinetics13, or on employing engineered and structured point spread functions (PSFs) that encode dipole orientation or wavefront distortions7,8,25. In contrast, our work demonstrates that meaningful classification of biomolecular states can be achieved from subtle, intrinsic variations in the PSF arising from changes in the local environment of a single fluorophore. To our knowledge, this is the first demonstration of PSF-mediated state classification without any explicit optical engineering or photophysical modulation. To elucidate the origin of this discriminatory signal captured by smDeepFLUOR, we conducted control experiments using 5′-Cy5-labeled double-stranded DNA (dsDNA) of varying total lengths, while maintaining the local fluorophore environment and DNA sequences identical (Supplementary Fig. 7). In these conditions, smDeepFLUOR failed to differentiate among the dsDNA molecules, which indicates that global structural properties or molecule-wide diffusion are unlikely to contribute to classification. Instead, experiments with variable gap lengths revealed that classification accuracy showed a sharp distinction between 9-nt and 6-nt gaps, suggesting that the fluorophore is responsive to localized environmental differences within a spatially restricted range (Fig. 4c). This is reminiscent of distance-dependent effects observed in mechanisms such as PIFE and NAIFE, which typically operate over distances of ~4 nm26 or ~7–9 nucleotides5. However, smDeepFLUOR achieved this discrimination without relying on abrupt intensity change, implying that it is likely to detect subtle, spatially confined variations in the emission profile rather than apparent intensity shifts. Consistent with this interpretation, autocorrelation analysis of fluorescence fluctuations revealed no significant differences in temporal fluctuation kinetics between gap and nick DNA (Supplementary Fig. 8), suggesting that smDeepFLUOR classification is not driven by simple differences in intensity dynamics. To elucidate the primary features exploited by smDeepFLUOR in distinguishing gap and nick DNA, we performed a series of complementary control experiments. When trained and evaluated on substrates containing random-sequence ssDNA gaps, smDeepFLUOR achieved a high classification accuracy of ~90% (Fig. 4). Although no discernible difference in fluorescence intensity was observed, this result is consistent with the previous reports of stacking-induced fluorescence enhancement of cyanine dyes, which addressed the intensity difference of Cy3 at gap and nick DNA4. Replacing the heterogeneous sequence gap with homogeneous poly(A) reduced the accuracy to ~79% (Supplementary Fig. 9), yet the performance remained well above chance. Because poly(A) ssDNA is known to adopt a distinct helical conformation compared with dsDNA27,28, this decrease suggests that conformational properties of the gap influence the local fluorophore environment sensed by the model. Nevertheless, the retained performance indicates that smDeepFLUOR remains sensitive to intrinsic structural differences between gap and nick DNA. Consistent with this interpretation, smDeepFLUOR was able to distinguish random sequence and poly(A) gap DNA constructs sharing an identical 32-nt gap length but differing only in sequence composition, with a classification accuracy of ~77% (Supplementary Fig. 10). Interestingly, smDeepFLUOR also detected sequence-dependent effects on the Cy3 attached to the 5′ end of dsDNA, despite the absence of a significant difference in fluorescence intensity (Supplementary Fig. 11). In previous studies29,30, Cy3 and Cy5 labeled at the terminus of ssDNA exhibited pronounced intensity changes depending on the identity of five consecutive nucleotides immediately adjacent to the dye. Taken together, these results suggest that smDeepFLUOR classification is dominated by DNA structural features in the immediate vicinity of the fluorophore, while retaining sensitivity to local sequence-dependent physicochemical environments that subtly modulate spatiotemporal fluorescence emission, consistent with short-range fluorescence sensitivity mechanisms such as PIFE and NAIFE. To further explore the mechanism underlying smDeepFLUOR classification, we examined how fluorophore photophysics and imaging conditions influence model performance (Supplementary Figs. 12–15). Gap–nick discrimination was robust for DNA labeled with the cyanine dye Cy3. In contrast, when identical labeling geometry and imaging conditions were maintained but Cy3 was replaced with the rhodamine-based dye Alexa Fluor 546 (AF546), a new network trained exclusively on the AF546 dataset exhibited a pronounced reduction in classification accuracy (~61%) and frequent misclassification (Supplementary Fig. 12). These results suggest that rhodamine-based dyes generate less discriminative spatiotemporal emission patterns when attached to DNA. Accordingly, smDeepFLUOR performance is strongly modulated by fluorophore photophysics, and dyes exhibiting pronounced environmental sensitivity, such as cyanine dyes and fluorescent proteins like mNeonGreen, are better suited for capturing the localized structural and dynamic signatures exploited by the model. Future extensions of smDeepFLUOR may include incorporating transfer learning31,32 or self-supervised training33 schemes to adapt the model to new molecular targets with limited annotation data. In particular, unsupervised classification of the spatiotemporal patterns of fluorescent molecule PSFs is of great significance. It has been observed that even slight changes in the local environment can induce detectable shifts in the emission profile of a single fluorophore, giving rise to distinct spatiotemporal PSFs. Therefore, if we can develop analysis methods that do not require training data, such as interpreting pixel-wise temporal fluctuations through autocorrelation or cross-correlation matrices34, it may become feasible to identify arbitrary and previously undefined molecular states. In addition to analytical approaches based on interpretable mathematical functions, the application of self-supervised training methods, such as masked autoencoders (MAE)35 in a label-free manner, could offer a powerful framework for uncovering latent molecular states previously inaccessible by conventional means. In conclusion, our findings establish smDeepFLUOR as a robust analytical framework capable of decoding hidden molecular information from seemingly homogeneous single-fluorophore signals. Its ability to operate with minimal labeling, generalize across independent experimental batches, and track single-molecule dynamics in real time makes smDeepFLUOR a valuable tool for probing subtle molecular interactions in diverse biological systems. Future work may integrate explainability techniques (e.g., saliency maps36 or attention mechanisms37) to further clarify which aspects of the fluorescence signal drive classification, potentially revealing new physical insights. Furthermore, extending smDeepFLUOR to broader classes of biomolecules, time scales, and imaging platforms may open new frontiers in real-time, minimally invasive single-molecule analysis. Methods Protein purification Human poly(A) binding protein C1 (PABP) purification Expression and purification procedures of human PABPs were performed using the Ni-NTA Fast Start Kit (QIAGEN) and followed by the kit protocol (Purification of 6xHis-tagged Proteins Under Native Conditions) with minor modification. PABP was expressed in NEBExpress® Iq Competent E. coli cells (NEB)38. Cells were cultured in LB broth at 37 °C to an OD₆₀₀ of 0.4, induced with 0.4 mM isopropyl β-D-1-thiogalactopyranoside (IPTG), and incubated for an additional 4 h at 37 °C. Cells were harvested by centrifugation (4000 × g, 10 min). All subsequent steps were performed on ice. Cells were resuspended in lysis buffer containing 1 mg/ml lysozyme and 250U benzonase nuclease solution and incubated for 30 min. The lysate was clarified by centrifugation (14,000 × g, 30 min), and the supernatant was loaded onto the Fast Start Column. The column was washed two times with 4 ml of native Wash Buffer and eluted using 1 ml Native Elution Buffer. Human mNG-poly(A) binding protein C1 (PABP) purification mNG-PABP was expressed and initially purified using the same Ni-NTA-based procedure described above for non-labeled PABP, followed by further purification using Anti-DYKDDDDK Affinity Resin (Sino Biological). The eluate was incubated with Anti-DYKDDDDK Affinity Resin overnight. To remove the nonspecific or unbound proteins, the resin was centrifuged (1000 × g , 5 min) and washed with phosphate-buffered Saline (PBS). Subsequently, 0.1 M elution buffer (100 mM Glycine, 10 mM NaCl, pH 3.0) was added to the resin and incubated with gentle shaking for 10 min at room temperature. Finally, the resin was centrifuged (5000 × g, 3 min) and the eluate was neutralized with 2 M Tris-buffer. The purified protein was analyzed by Coomassie-stained SDS-PAGE, and its concentration was determined using a NanoPhotometer. Human DNA polymerase η Human full-length DNA polymerase η was expressed in E. coli BL21 (DE3) cells using an N-terminal (His)₆-SUMO tag from the pE-SUMO-pro expression vector (LifeSensors, USA). Cells were cultured in 2YT medium at 24 °C to an OD₆₀₀ of 0.8, induced with 0.1 mM isopropyl β-D-1-thiogalactopyranoside (IPTG), and incubated for an additional 19 h at 16 °C. Cells were harvested by centrifugation and resuspended in lysis buffer [50 mM Tris-HCl (pH 8.0), 750 mM NaCl, 40 mM imidazole, 5 mM β-mercaptoethanol, 0.2% NP-40, 1 mM PMSF, 5% glycerol, and 1 EDTA-free protease inhibitor tablet per 50 mL (Roche, UK)]. All subsequent steps were performed at 4 °C. Cells were lysed enzymatically with 2 mg/mL lysozyme and mechanically by sonication. The lysate was clarified by centrifugation (22,040 × g, 60 min), and the supernatant was loaded onto a 5 mL HisTrap HP affinity column (Cytiva, USA) pre-equilibrated with buffer A [50 mM Tris-HCl (pH 7.5), 500 mM NaCl, 40 mM imidazole, 5 mM β-mercaptoethanol, and 5% glycerol]. The column was washed with 50 mL of buffer A and eluted using a linear gradient of buffer B [same as buffer A, but with 500 mM imidazole]. Polymerase η eluted at approximately 310 mM imidazole. Peak fractions were pooled and dialyzed overnight in dialysis buffer [50 mM Tris-HCl (pH 7.5), 500 mM NaCl, 5 mM β-mercaptoethanol, and 5% glycerol] in the presence of SUMO protease (LifeSensors) to remove the SUMO tag and yield native polymerase η. The dialyzed sample was re-applied to the HisTrap column, and the untagged polymerase η was collected from the flow-through. Polymerase η-containing fractions were pooled, concentrated, and further purified by size-exclusion chromatography using a HiLoad 16/600 Superdex 200 pg column (Cytiva, USA) equilibrated with storage buffer [50 mM Tris-HCl (pH 7.5), 300 mM NaCl, 10% glycerol, and 1 mM DTT]. Final protein fractions were concentrated, flash-frozen in liquid nitrogen, and stored at −80 °C. Human DNA polymerase δ purification DNA polymerase δ was expressed in Sf9 insect cells and purified as previously described20. Bacmid DNA containing all four subunits was transfected into Sf9 insect cells using FuGENE® HD transfection reagent (Promega, USA) according to the manufacturer’s instructions. The resulting supernatant was collected as the P1 viral stock and subsequently amplified to generate P2 and P3 viral stocks for large-scale protein expression. For polymerase δ expression, 4 L of Sf9 suspension culture at a density of 2 × 10⁶ cells/mL was infected with P3 viral stock and incubated for 72 h. Cells were harvested by centrifugation at 5500 × g for 10 min and resuspended in lysis buffer [50 mM Tris-HCl (pH 7.5), 500 mM NaCl, 40 mM imidazole, 5 mM β-mercaptoethanol, 0.2% NP-40, 1 mM PMSF, 5% glycerol, and EDTA-free protease inhibitor cocktail]. Cell lysis was performed by sonication, and cellular debris was removed by ultracentrifugation at 95,834 × g for 1 h at 4 °C. The clarified lysate was loaded onto a 5 mL HisTrap HP affinity column pre-equilibrated with buffer A [50 mM Tris-HCl (pH 7.5), 500 mM NaCl, 40 mM imidazole, 5 mM β-mercaptoethanol, and 5% glycerol]. The column was washed sequentially with buffer A containing 40 mM and 100 mM imidazole to remove non-specifically bound proteins. Subsequently, the salt concentration was reduced by washing with buffer B [50 mM Tris-HCl (pH 7.5), 50 mM NaCl, 40 mM imidazole, 5 mM β-mercaptoethanol, and 5% glycerol]. Bound proteins were eluted using a linear gradient of buffer C [50 mM Tris-HCl (pH 7.5), 50 mM NaCl, 600 mM imidazole, 5 mM β-mercaptoethanol, and 5% glycerol]. Fractions containing all polymerase δ subunits were pooled and further purified using a Mono Q 5/50 GL (Cytiva, USA) anion exchange column pre-equilibrated with buffer D [50 mM Tris-HCl (pH 7.5), 50 mM NaCl, 5 mM β-mercaptoethanol, and 5% glycerol]. The unbound proteins were removed by washing with 20 ml of buffer D containing 150 mM NaCl. Polymerase δ was eluted using a 20 mL linear gradient from 150 mM to 1 M NaCl. Fractions containing all polymerase δ subunits were combined, concentrated to 1 mL and subjected to size exclusion chromatography on a HiLoad 16/600 Superdex 200 pg column equilibrated with the storage buffer [50 mM Tris-HCl (pH 7.5), 200 mM NaCl, 1 mM DTT, and 5% glycerol]. Purified protein fractions were pooled, flash frozen, and stored at −80 °C. PCNA purification N-terminally His-tagged PCNA WT was expressed in E. coli and purified as previously described20. The PCNA expression plasmid was transformed into BL21 (DE3) E. coli cells and cultured in 2YT medium supplemented with ampicillin at 37 °C until an OD600 of 1.2 was reached. Protein expression was induced with 0.5 mM IPTG, and cultures were incubated for an additional 19 h at 16 °C. Cells were harvested by centrifugation at 5500 × g for 10 min and resuspended in lysis buffer [50 mM Tris-HCl (pH 7.5), 750 mM NaCl, 20 mM imidazole, 5 mM β-mercaptoethanol, 0.2% Nonidet P-40, 1 mM PMSF, 5% glycerol, and EDTA-free protease inhibitor cocktail]. Cells were lysed with 2 mg/mL lysozyme at 4 °C for 30 min followed by sonication. Cellular debris was removed by centrifugation at 22,040 × g for 20 min at 4 °C, and the clarified lysate was directly loaded onto a 5 mL HisTrap HP affinity column pre-equilibrated with buffer A [50 mM Tris-HCl (pH 7.5), 500 mM NaCl, 20 mM imidazole, 5 mM β-mercaptoethanol, and 5% glycerol]. The column was washed with buffer A and subsequently with buffer B [50 mM Tris-HCl (pH 7.5), 100 mM NaCl, 20 mM imidazole, 5 mM β-mercaptoethanol, and 5% glycerol] to reduce the salt concentration. Bound PCNA was eluted using a linear gradient of buffer C [50 mM Tris-HCl (pH 7.5), 100 mM NaCl, 500 mM imidazole, 5 mM β-mercaptoethanol, and 5% glycerol]. Eluted fractions were pooled and directly applied onto a HiTrap Q HP 5 mL (Cytiva, USA) anion exchange column pre-equilibrated with buffer D [50 mM Tris-HCl (pH 7.5), 100 mM NaCl, 5 mM β-mercaptoethanol, and 5% glycerol]. After washing with buffer D, protein was eluted using a linear gradient with buffer E [50 mM Tris-HCl (pH 7.5), 1 M NaCl, 5 mM β-mercaptoethanol, and 5% glycerol]. Peak fractions were pooled, concentrated to 1.5 mL and further purified by size exclusion chromatography using a HiLoad 16/600 Superdex 200 pg column equilibrated with the storage buffer [50 mM Tris-HCl (pH 7.5), 150 mM NaCl, 1 mM DTT, and 5% glycerol]. Purified protein fractions were pooled, flash frozen, and stored at −80 °C. RFC purification ΔN-RFC (amino acids: 554–1148 in RFC1) was expressed in E. coli and purified as previously described20. Two plasmids, p14A-ΔNRFC1/4/2 and pCDFK-RFC5/3, were co-transformed into BL21(DE3) E. coli cells, and transformants were selected on agar plates supplemented with kanamycin and ampicillin. Protein overexpression was carried out in 10 L of TB medium containing both antibiotics. Cultures were grown at 20 °C to an OD600 of 0.8, followed by induction with 0.1 mM IPTG, and incubation was continued for an additional 24 h at 15 °C. Cells were harvested by centrifugation and resuspended in lysis buffer [50 mM HEPES-NaOH (pH 7.5), 500 mM NaCl, 10 mM Imidazole, 10 mM β-mercaptoethanol, 5% glycerol, 1 mM PMSF, and EDTA-free protease inhibitor cocktail]. Cells were lysed with 2 mg/mL lysozyme at 4 °C for 30 min followed by sonication. Cellular debris was removed by centrifugation at 22,040 × g for 45 min at 4 °C, and the clarified lysate was directly loaded onto a HisTrap HP 5 mL affinity column pre-equilibrated with buffer A [50 mM HEPES-NaOH (pH 7.5), 500 mM NaCl, 10 mM Imidazole, 10 mM β-mercaptoethanol, and 5% glycerol]. The column was washed with buffer A and subsequently with buffer B [50 mM HEPES-NaOH (pH 7.5), 250 mM NaCl, 10 mM Imidazole, 10 mM β-mercaptoethanol, and 5% glycerol] to reduce the salt concentration. Bound ΔNRFC subunits were eluted using a linear gradient of buffer C [50 mM HEPES-NaOH (pH 7.5), 250 mM NaCl, 300 mM Imidazole, 10 mM β-mercaptoethanol, and 5% glycerol]. Fractions containing all ΔNRFC subunits were pooled and directly loaded onto a HiTrap Heparin HP 1 mL column (Cytiva, USA) pre-equilibrated with buffer D [50 mM HEPES-NaOH (pH 7.5), 250 mM NaCl,10 mM β-mercaptoethanol, and 5% glycerol]. After washing with buffer D, the protein complex was eluted using a linear gradient with buffer E [50 mM HEPES-NaOH (pH 7.5), 1 M NaCl,10 mM β-mercaptoethanol, and 5% glycerol]. Fractions containing all ΔNRFC subunits were combined, concentrated, and subjected to size exclusion chromatography using a HiLoad Superdex 16/600 200 pg column equilibrated with the storage buffer [50 mM HEPES-NaOH (pH 7.5), 300 mM NaCl, 1 mM DTT, and 10% glycerol]. Fractions containing correctly assembled ΔNRFC complex were pooled, concentrated, flash frozen, and stored at −80 °C. Quartz preparation The quartz slides were cleaned through sequential sonication following a previously reported protocol39. The cleaning steps included sonication in 10% Alconox solution for 20 min, followed by rinsing with deionized water (Milli-Q, Merck Millipore, Germany) for 5 min, acetone for 15 min, 1 M KOH for 20 min, and methanol for 10 min with water bath sonication. They were then flame-treated with a torch for 30 s. After cleaning, the slides were incubated in a 2% (v/v) solution of (3-aminopropyl) triethoxysilane (Sigma, USA) in methanol for 20 min. A mixture of polyethylene glycol succinimidyl valerate (PEG-SVA, MW 5000, Laysan Bio, USA) and Biotin-PEG-SVA (MW 5000, Laysan Bio, USA) at a mass ratio of 40:1 was prepared in 0.1 M sodium bicarbonate buffer (pH 8.3). This solution was applied to the quartz slide surface and incubated for 6 h. Finally, the functionalized slides were rinsed with deionized water and dried with nitrogen gas. Total internal reflection fluorescence (TIRF) microscopy All the single-molecule TIRF images in this study were acquired with the 1.6× magnifier in a prism-type TIRF microscope (Olympus IX-71, water-immersion 60× objective NA = 1.2), EMCCD (ImagEM C9100-13, Hamamatsu), MetaMorph 7.6 (Molecular Devices) imaging software, and a Dual View optical setup (DV2, Photometrics) if required. Images were acquired at a frame rate of 100 ms with a 512 × 512-pixel format, corresponding to a pixel size of 171 nm, and an EM gain of 255. Fluorophores were excited using the laser lines (Cobolt MLD 488 nm-300 mW for mNG, Cobolt DPL 532 nm-100 mW for Cy3, and CrystaLaser 638 nm-70 mW for Cy5 and AF647). The illumination intensity at the sample plane was approximately 280 W/cm². The fluorescence signal was filtered with each emission filter (Chroma, ET525/50 m; Semrock, 585/40; Semrock, FF01-676/37-25). Single-molecule experiment conditions PABP binding to poly(A) For PABP-binding classification, poly(A)60 containing 60 adenine nucleotides was synthesized by in vitro transcription from the pSK_An plasmid using the HiScribe™ T7 Quick High Yield RNA Synthesis Kit (New England Biolabs, USA). For nonspecific PABP binding classification, poly(A)25 DNA/RNA hybrid oligonucleotides and Cy5/biotin-labeled DNA oligonucleotides were purchased from Integrated DNA Technologies (IDT, USA). The sequence of poly(A)25 was 5′-ctcgaggtcgacggtatcgataagcttgaagac(A)25-3′, and that of the DNA oligonucleotide was 5′-/Cy5/gcttatcgataccgtcgacctcgag/biotin/-3′. All poly(A)n were annealed with the DNA oligonucleotide prior to experiments. After PEG-biotinylated flow chamber preparation, streptavidin (0.05 mg/ml, Sigma-Aldrich) in blocking buffer (20 mM Tri-HCl, pH 7.5, 2 mM EDTA, 50 mM NaCl, and 0.0025% Tween 20(v/v), all from Sigma) was introduced into the chamber and incubated for 5 min. Following free streptavidin removal by washing with blocking buffer, Cy5-conjugated RNA/DNA hybrid (10 pM) in blocking buffer was injected into the channel and incubated for 10 min before being washed with reaction buffer (25 mM Tris-HCl, pH 7.5, 100 mM l-glutamic acid potassium salt monohydrate, 5 mM MgCl2, 0.0025% [v/v] Tween 20, 0.01% Triton X-100, all from Sigma). Next, mNG-PABP (200 pM) or non-labeled PABP (10 nM) in reaction buffer was introduced and incubated for 5 min. To acquire images of nonspecific bound PABP, mNG-PABP (1 nM) was incubated in the chamber for 30 min in the absence of RNA/DNA hybrid. Finally, fluorescence images were acquired in the presence of an oxygen scavenging system [2 mM Trolox (Sigma, USA), 5 mM PCA (Sigma, USA), and 200 nM rPCO (Oriental Yeast, Japan)]. In the protein binding detection experiments, Cy5-conjugated RNA/DNA hybrids in blocking buffer were first introduced into the flow chamber, followed by the injection of an oxygen-scavenging buffer. Unbound Cy5 molecules were imaged using a red emission filter (Semrock, FF01-676/37-25). Subsequently, mNG-PABP (200 pM) was introduced together with the oxygen-scavenging buffer, and the sample was incubated for approximately 5 min. During incubation, imaging was performed using a blue emission filter (Chroma, ET525/50 m). Immediately thereafter, the same red emission filter was used again to acquire images of Cy5 in the bound state. Only Cy5 signals exhibiting clear colocalization were classified as true binding events and included in the analysis. DNA polymerases with gap and nick DNA DNA oligos were purchased from IDT (Supplementary Table 1). To immobilize DNA templates on the chamber surface, we followed the same protocol as the previous PABP binding experiments, with one modification: after streptavidin incubation, Cy3-conjugated gap DNA or nick DNA (1 pM) in blocking buffer was injected into the channel. Fluorescently labeled molecules were incubated in the chamber for 10 min before being washed with reaction buffer. For the polymerase activity experiments, reaction buffer containing 1 mM ATP, 10 μM ADP, 1 mM DTT, 100 μg/ml BSA, 100 μM each dNTP, 5 nM PCNA, 2.5 nM RFC, and either 30 nM polymerase δδ or 90 nM polymerase ηη was introduced into the chamber. Control experiments were performed in identical conditions, excluding specific interacting molecules as detailed in the main text. Fluorescence images were acquired using the same imaging buffer containing an oxygen scavenging system. Single particle cropping from TIRF microscopy images Single particle detection and cropping were performed using the ImageJ plugin mosaic40. Prior to particle detection, background intensity was adjusted using rolling ball background subtraction in ImageJ. This procedure corrects for fluorescence background arising from the sample but does not constitute an independent dark-frame calibration of the camera offset (detector dark counts). Particle detection was conducted with the following hyperparameters: radius = 3, cutoff = 0.001, and absolute threshold = 250, 300, 350, with smear particles below the threshold being excluded. For particle tracking, we applied displacement = 1 and 2.5 and link range = 1. Except for the determination of colocalization between particles, the displacement hyperparameter was set to 1, meaning that a particle was allowed to move a maximum of one pixel between consecutive frames. Particle tracking provided sub-pixel localization coordinates (x, y) for each fluorescence spot. For cube extraction, these coordinates were used to determine the corresponding central camera pixel, around which a 7 × 7 pixel region of interest was cropped from each frame. No sub-pixel image shifting or interpolation was applied; cubes were extracted directly from the original camera frames to preserve raw pixel intensities for the whole dataset. Since smDeepFLUOR predicts states based on a 10-frame sequence, traces shorter than 10 frames were discarded. Image standardization Standardization of data is crucial for increasing training performance by ensuring faster and more stable convergence during gradient descent, improving optimization, and enabling more effective weight initialization. It also boosts the performance of activation functions like ReLU, which we used in our model, enhances the effectiveness of batch normalization, and helps reduce overfitting by ensuring fair contribution from all features. Intensity from each pixel was standardized within a 7 × 7 × 10 cube (490 pixels). The equation used for the standardization is as follows:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{Z}}}}=\frac{{I}_{i}-\mu }{{{{\rm{\sigma }}}}},\mu=\frac{{\sum }_{i=1}^{490}{I}_{i}}{490},\sigma=\sqrt{\frac{1}{490}\mathop{\sum }_{i=1}^{490}\left({I}_{i}-\mu \right)}$$\end{document}Z=Ii−μσ,μ=∑i=1490Ii490,σ=1490∑i=1490Ii−μ Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${I}_{i}$$\end{document}Ii is the intensity of each pixel, and ZZ is the z-score of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${I}_{i}$$\end{document}Ii in each distribution with mean μμ and standard deviation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{\sigma }}}}$$\end{document}σ. No separate dark-frame-based offset calibration was performed. However, because normalization is performed independently within each cube, any constant detector offset within that cube is eliminated during standardization. Training and validation Dataset construction and temporal input design The complete data-processing workflow, from single-molecule imaging using TIRF microscopy to CNN input preparation, is summarized in Supplementary Fig. 1. To construct the full dataset, we extracted particle-centered image stacks of size 7 × 7 × N, where N denotes the number of frames available before photobleaching, from spots that were tracked for more than 10 frames for each state. The full dataset was divided into training and independent-validation experiments. The training data were further split into training and test subsets, while independent experimental datasets acquired on different days were reserved exclusively for independent validation. To determine an appropriate temporal depth for model input, we systematically evaluated how classification performance depends on the number of time frames included in each input unit. Using the gap and nick DNA dataset, we constructed input representations with increasing temporal depth (7 × 7 × N, where N = 1, 2, 5, 10, or 20) and assessed model performance under otherwise identical training conditions (Supplementary Fig. 16). smDeepFLUOR were trained independently for each temporal depth without adaptation of the N = 10 model. Classification accuracy increased monotonically with temporal depth, with a pronounced improvement up to N = 10 frames, beyond which performance largely saturated. Based on this analysis, we selected 7 × 7 × 10 as the default input size, which provides sufficient temporal context while avoiding excessive temporal averaging. Model architecture and training procedure We trained a shallow 3D CNN to classify single-channel PSF cubes of shape 7 × 7 × 10 × 1 (H × W × T × C). The network comprised three Conv3D blocks—(32 filters, 2 × 2 × 1), (16 filters, 3 × 2 × 1), and (32 filters, 2 × 2 × 10)—each followed by ReLU activation, batch normalization, and dropout. The first two blocks focused on spatial feature extraction (temporal kernel = 1), whereas the third block integrated temporal dynamics using a kernel spanning all 10 frames. Extracted features were flattened and passed to a dense layer (16 units, ReLU, L2 regularization = 0.05), followed by a two-class softmax classifier. The final class assignment was determined by selecting the class with the higher predicted probability, without additional post-processing or thresholding. During training, dropout was used as a regularization strategy to mitigate overfitting. For each classification task, dropout ratios were selected by evaluating validation performance across a predefined range. The optimal dropout rate varied across experimental conditions, reflecting differences in data heterogeneity and overfitting behavior, and was therefore determined on a per-task basis. Training was performed using the Adam optimizer (learning rate = 1 × 10⁻⁴) with a batch size of 32 for up to 50 epochs. The model was trained using the sparse categorical cross-entropy loss, defined as2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L=-\left(\frac{1}{N}\right){\sum }_{i}\log \left({p}_{i,{y}_{i}}\right)$$\end{document}L=−1N∑ilogpi,yiwhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${p}_{i,{y}_{i}}$$\end{document}pi,yi denotes the predicted probability assigned to the correct class \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${y}_{i}$$\end{document}yi for sample ii, and NN is the number of samples in the minibatch. In this formulation, class labels are represented as integer indices rather than one-hot encoded vectors. Validation loss was computed using the same loss function but evaluated on the held-out validation dataset that was not used for parameter updates during training. Regularization included batch normalization, dropout, and L2 regularization on the penultimate layer. We applied ReduceLROnPlateau on validation loss (factor = 0.5, patience = 3; minimum learning rate = 1 × 10⁻⁶), early stopping on validation loss (patience = 6, restoring best weights), and checkpointing of the best-performing model based on validation accuracy. Model training was performed using Google Colab Pro with an A100 GPU. Typical training time for a smDeepFLUOR model was approximately 2–15 min on an NVIDIA A100 GPU, depending on the dataset size. Scikit-learn41 and TensorFlow42 were used for data processing and model implementation. Validation strategy and prediction integration Training–test splitting was performed at the particle/trajectory level: particle-centered stacks were randomly divided into training (80%) and test (20%) sets. After this split, each particle stack was segmented into sequential 7 × 7 × 10 cubes for model training and evaluation. As a result, cubes originating from the same particle trajectory could not appear in both training and test sets, thereby preventing particle-level information leakage. During model development, we observed that the degree of overfitting depended on the specific experimental condition, including fluorophore type, binding behavior, and signal heterogeneity. For each experiment, all available data were first randomly split into training and test sets at the particle level. Overfitting was monitored by comparing training accuracy, validation accuracy, and prediction accuracy on independent experimental datasets acquired on different days. When the difference between validation accuracy and independent experiment was large, approximately 15–20%, indicating reduced generalization across experimental sessions, we applied random subsampling within the training set to limit the effective training set size. In practice, this situation was most often observed when the number of training cubes exceeded ~60,000. Random subsampling was therefore applied to reduce the training set to approximately 20,000–50,000 cubes, depending on the task. This subsampling was performed uniformly at random, preserved class balance, and was applied only to the training dataset after the particle-level split. The subsampling did not influence the test sets and independent experimental datasets. All validation and independent experiment evaluations were conducted exclusively on data not included in the subsampled training sets. Independent-experiment predictions were performed on datasets acquired on different days that were not used for training, separate from the 80:20 X_test/y_test split. In addition, for samples that undergo state transitions (for example, PABP unbinding to binding or gap-to-nick transitions), we used within-particle transitions acquired under identical imaging conditions across multiple days as an additional source of independent prediction data. For prediction results not shown in real time, final particle-level predictions were obtained by averaging predictions from 7 × 7 × 10 cubes extracted from 7 × 7 × N image stacks of a single particle. These particle-level predictions were used to construct all confusion matrices. In contrast, real-time prediction statistics were computed at the cube level prior to averaging and are summarized in Supplementary Fig. 17. For visualization of state transitions, real-time predictions were averaged over 10 successive 7 × 7 × 10 cubes to improve temporal robustness using a sliding window along the temporal dimension, resulting in adjacent cubes sharing overlapping frames. To clarify how cube-level predictions are integrated into particle-level classification, we explicitly compared per-cube and per-particle prediction strategies (Supplementary Fig. 18). In the per-particle approach, predictions from multiple cubes derived from the same trajectory were aggregated by majority voting, whereas in the per-cube analysis, each cube was classified independently. Comparable overall performance between per-particle and per-cube confusion matrices indicated consistent model behavior across both levels. Although classification accuracy generally increased with the number of cubes available per particle, robust performance was maintained even with a small number of cubes (as few as 1–10), indicating that the model does not rely on extensive temporal averaging. Artifact and misclassification control analyses To evaluate whether residual cube-level misclassifications could be attributed to trivial photometric differences, we compared fluorescence intensity and signal-to-noise ratio (SNR) between correctly and incorrectly classified cubes (Supplementary Fig. 19). Differences were minimal for both gap and nick DNA, with misclassified samples showing only slightly reduced intensity and ~9% reduced SNR. These results indicate that transient cube-level misclassifications are not driven by systematic differences in photon budget or noise level but instead reflect subtle local spatiotemporal ambiguities that are effectively suppressed by particle-level aggregation. We further examined whether cube-level misclassifications could arise from spatially heterogeneous imaging conditions by performing a grid-wise (4 × 4) analysis across the original 512 × 512-pixel fields of view (Supplementary Fig. 20). While prediction fractions varied across grids (e.g., ~0.01–0.21 in certain regions), indicating modest spatial heterogeneity, quantitative comparison of SNR and mean intensity between correctly and incorrectly classified cubes showed no consistent spatial pattern that could account for the observed differences. In high-accuracy regions, occasional misclassifications were associated with lower SNR, but in other regions, no clear separation in SNR or intensity was observed. Overall, these results suggest that cube-level misclassifications are not systematically driven by uneven illumination, detector artifacts, or other spatially structured imaging biases. To exclude contributions from optical aberrations or systematic distortions of the PSF, we quantitatively analyzed the experimentally measured PSF (Supplementary Fig. 21). The average PSF closely matched an ideal diffraction-limited Gaussian model, with no detectable asymmetry or aberration-associated features. Strehl ratio distributions showed substantial overlap between gap and nick DNA datasets, indicating no class-dependent differences in PSF quality. Together, these results support the conclusion that smDeepFLUOR classification errors arise from intrinsic local spatiotemporal ambiguities rather than from systematic experimental or optical artifacts. Lastly, to confirm that smDeepFLUOR specifically learns fluorescence-dependent features, background images adjacent to fluorescence spots were extracted and evaluated. When 7 × 7 × 10 background cubes were classified using the same smDeepFLUOR architecture, model performance was poor across experimental conditions, demonstrating that smDeepFLUOR detects intrinsic fluorescence-related signatures rather than background variations (Supplementary Fig. 22). Representative training loss curves for the smDeepFLUOR models across different classification tasks were provided in Supplementary Fig. 23, showing stable convergence of training and validation loss during optimization. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Transparent Peer Review file Supplementary Information Reporting Summary Transparent Peer Review file Source data Source data Source data