Abstract
The intramolecular dynamics of fluorescent indicators of neural activity can distort the accurate estimate of action potential ("spike") times. In order to develop a more accurate spike inference algorithm we characterized the kinetic responses to calcium of three popular indicator proteins, GCaMP6f, jGCaMP7f, and jGCaMP8f, using in vitro stopped-flow and brain slice recordings. jGCaMP8f showed a use-dependent slowing of fluorescence responses that caused existing inference methods to generate numerous false positives. From these data we developed a multistate model of GCaMP and used it to create Bayesian Sequential Monte Carlo (Biophys(SMC)) and machine learning (Biophys(ML)) inference methods that reduced false positives substantially. This biophysical method dramatically improved spike time accuracy, detecting individual spikes with a median uncertainty of 4 milliseconds, a performance level that reached the theoretical limit and is twice as accurate as any previous method. Our framework thus highlights advantages of physical model-based approaches over model-free algorithms.