DART: Deep learning for the Analysis and Reconstruction of Transcriptional dynamics from live-cell imaging data

Transcriptional bursting, characterized by stochastic switching between promoter states, underlies cell-to-cell variability in gene expression. Accurately inferring promoter activity from live-cell imaging data remains challenging because the fluorescence signal at any given point is influenced by the history of promoter states. Here, we present DART (Deep learning for the Analysis and Reconstruction of Transcriptional dynamics), a deep learning framework that infers promoter onand off-states from fluorescence intensity traces, enabling the estimation of activation and inactivation rates and the selection of the most appropriate promoter-switching model. DART utilizes a neural network architecture that combines convolutional neural networks and long short-term memory layers to binarize fluorescence traces. Using extensive synthetic datasets spanning a wide range of transcriptional bursting levels, we demonstrate that DART outperforms current binarization methods, including conventional and augmented hidden Markov models, in both accuracy and robustness. Furthermore, a reanalysis of published experimental data using DART reveals a strong linear coupling between activation and inactivation rates, contradicting previous claims of independence. Our approach provides a powerful and generalizable tool for quantitative analysis of transcriptional kinetics from live-cell imaging data.