Fig. 2: Diffraction-limited fluorescent spots hold binding-type information that can be utilized to more accurately differentiate between single binding events.

a Two types of DNA-binding domains used to collect diffraction-limited fluorescent spots. Blinking kinetics refers to the binary signals of binding on and off. (Revised from ref. ¹¹). b The correlation between domain labels and some common statistics (n = 25,530 independent binding events), as well as the classification accuracy accounted for by the missing information. The “YZ,” “XZ,” and “XY” statistics refer to those calculated from projections of the video onto the Y-Z, X-Z, and X-Y planes, with X as the width, Y as the height, and Z as the time axis. c The correlation between the final classification output of the proposed T2C CNN (FC2 feat1 and FC2 feat2) and some common statistics, as well as the classification accuracy accounted for by the additional information captured by the model and the remaining missing information. d Scatter plot of mean on-time versus mean off-time. Below, the Gaussian kernel density estimate of the probability distribution function (PDF) for the mean on-time is displayed. Based on multiple binding events, 6.33% of binding sites are misclassified. e A 2D uniform manifold approximation and projection (UMAP) plot of the output embeddings from the convolutional layers of the proposed T2C CNN. This model achieves a lower error rate (3.62%) in more fine-grained recognition of single binding events, significantly reducing the time required to determine the binding type for a binding site. Clearer visualizations with each domain plotted on top are provided in Supplementary Fig. 2. Source data are provided as a Source Data file.