High frequency spike inference with particle Gibbs sampling

Abstract

AbstractFluorescent calcium indicators are an indispensable tools for monitoring the spiking activity of large neuronal populations in animal models. However, despite the plethora of algorithms developed over the last decades, accurate spike time inference methods for rates greater than 20 Hz are lacking. More importantly, little attention has been devoted to the quantification the statistical uncertainties in spike time estimation, which is essential for assigning a confidence in the inference for a particular recording. To address these challenges, we introduce an auto-regressive generative model that accounts for bursting neuronal activity and baseline fluorescence modulation, and it also applies recent sequential Monte Carlo approaches to obtain joint posterior distributions of static and dynamic model parameters. We show that our inference method is competitive with state-of-the-art algorithms by analysing the CASCADE benchmark datasets. We also show that spike time intervals as short as five milliseconds can be inferred from fluorescence transients recorded using a state-of-the-art genetically encoded indicator. Overall, our study describes a Bayesian inference method to detect neuronal spiking patterns and their uncertainty. The use of particle Gibbs samplers allows for unbiased estimates of all model parameters and it provides a statistical framework to test more specific models of calcium indicators.