Deep learning-based classification of complex intracellular calcium concentration patterns

Intracellular calcium ion (Ca2+) exhibits diverse dynamical behaviors, including complex oscillations such as bursting and chaos. The distinction of diverse dynamical patterns in time series data holds significant biochemical and biophysical implications, as these patterns are intimately related to physiological cellular states associated with health and disease. In this work, we introduce a deep learning framework based on a large kernel convolutional neural network (LKCNN) for the classification of dynamical states of intracellular Ca2+ dynamics. We use a chemical Langevin equation to generate synthetic data for training the LKCNN, simulating intracellular Ca2+ concentration patterns mimicking real experimental traces. We show that the LKCNN framework reliably classifies diverse intracellular Ca2+ dynamical regimes, achieving near 90% accuracy across all dynamical states. To this end, we propose an optimized kernel size for the LKCNN classifier, which captures both short-lived fluctuations and long-range correlations. While steady states, bursting, and simple oscillations are robustly distinguished with near-perfect accuracy, the performance slightly degrades for chaotic and multiple periodicity states under realistic levels of intrinsic fluctuations, reflecting genuine overlap in their temporal signatures. We validate our classifier with experimental Ca2+ concentration data, showing strong agreement with manual labeling and confirming generalizability beyond synthetic datasets. These results establish LKCNNs as flexible and scalable tools for systematic classification of cellular dynamics, with broad potential applications to other oscillatory biological processes.