Precise hyperacuity estimation of spike timing from calcium imaging

Abstract

25Two-photon imaging is a major recording technique in neuroscience, but it suffers 26 from several limitations, including a low sampling rate, the nonlinearity of calcium 27 responses, the slow dynamics of calcium dyes and a low signal-to-noise ratio, all of which 28 impose a severe limitation on the application of two-photon imaging in elucidating 29 neuronal dynamics with high temporal resolution. Here, we developed a hyperacuity 30 algorithm (HA_time) based on an approach combining a generative model and machine 31 learning to improve spike detection and the precision of spike time inference. First, 32 Bayesian inference estimates the calcium spike model by assuming the constancy of the 33 spike shape and size. A support vector machine employs this information and detects 34 spikes with higher temporal precision than the sampling rate. Compared with conventional 35 thresholding, HA_time improved the precision of spike time estimation up to 20-fold for 36 simulated calcium data. Furthermore, the benchmark analysis of experimental data from 37 different brain regions and simulation of a broader range of experimental conditions 38 showed that our algorithm was among the best in a class of hyperacuity algorithms. We 39 encourage experimenters to use the proposed algorithm to precisely estimate hyperacuity 40 spike times from two-photon imaging. 42Recently, two-photon imaging has been one of the major means of recording 43 multineuronal activities in neuroscience to obtain the precise morphology and location of 44 the target neurons because of its high spatial resolution 1-6 . However, its utility is still 45 constrained by its relatively low temporal resolution due to the mechanical scanning of 46 two-photon rays. The other problems are the nonlinearity, the slow dynamics and the low 47 signal-to-noise ratio (SNR) of the calcium (Ca) responses 7-10 . Many algorithms have been 48 proposed to reconstruct spike trains from Ca imaging data, including conventional 49 thresholding 11 , deconvolution 12-15 , template matching 16-20 , Bayes inference 21-23 and 50 machine learning 24,25 , to overcome these problems. Few of them, however, have 51 addressed the two challenging goals simultaneously: reliable spike detection and spike 52 time estimation with high temporal precision in the presence of the nonlinearity, slow 53 dynamics and low SNR of the Ca responses 26 . For the former goal, the spike dynamics 54 of the target neurons and/or kinematics of the Ca responses may vary dramatically across 55 brain regions and different Ca dyes. For the latter goal, a trade-off between the number 56 of recorded neurons and temporal resolution exists. The slow kinematics and the low 57 SNR of the currently available Ca dyes may also limit the temporal precision of the 58 information conveyed by the Ca responses. These factors impair reliable spike detection 59 as well as precise spike time estimation for high-frequency firing that is frequently 60 encountered in cortical cells 27-29 . 61 Here, we propose an approach combining a g...