Precise calcium-to-spike inference using biophysical generative models

The intramolecular dynamics of fluorescent calcium indicators distort the relationship between calcium signals and action potentials ("spikes"), hampering efficient spike inference from calcium imaging. To address this problem, we characterized the calcium response kinetics of three widely used indicators, GCaMP6f, jGCaMP7f, and jGCaMP8f, using in vitro stopped-flow measurements and brain slice recordings. We identify previously unreported kinetic features, including use-dependent slowing of fluorescence decay, that introduce systematic errors in linear model-based inference methods. Using these observations, we developed a multistate model of GCaMP and used it to create biophysically-inspired Bayesian Sequential Monte Carlo and machine learning inference models trained on synthetic datasets. These methods outperform existing methods on spike timing accuracy and correlation benchmarks derived from diverse cell types. Our results show that using synthetic data derived from our biophysical model yields a decoder that outperforms even those trained on extensive experimental data. By separating indicator characterization from inference, our framework, Calcium Spike Processing using Integrated Kinetic Estimation and Simulation (C-SPIKES), provides a generalizable strategy applicable to existing and future calcium indicators.