On the use of calcium deconvolution algorithms in practical contexts

Abstract

1Calcium imaging is a powerful tool for capturing the simultaneous activity of large 2 populations of neurons. Here we determine the extent to which our inferences of neu-3 ral population activity, correlations, and coding depend on our choice of whether and 4 how we deconvolve the calcium time-series into spike-driven events. To this end, we 5 use a range of deconvolution algorithms to create nine versions of the same calcium 6 imaging data obtained from barrel cortex during a pole-detection task. Seeking suit-7 able values for the deconvolution algorithms' parameters, we optimise them against 8 ground-truth data, and find those parameters both vary by up to two orders of mag-9 nitude between neurons and are sensitive to small changes in their values. Applied to 10 the barrel cortex data, we show that a substantial fraction of the processing methods 11 fail to recover simple features of population activity in barrel cortex already estab-12 lished by electrophysiological recordings. Raw calcium time-series contain an order of 13 magnitude more neurons tuned to features of the pole task; yet there is also qualitative 14 disagreement between deconvolution methods on which neurons are tuned to the task. 15 Finally, we show that raw and processed calcium time-series qualitatively disagree on 16 the structure of correlations within the population and the dimensionality of its joint 17 activity. Collectively, our results show that properties of neural activity, correlations, 18 and coding inferred from calcium imaging are highly sensitive to the choice of if and 19 how spike-evoked events are recovered. We suggest that quantitative results obtained 20 from population calcium-imaging be verified across multiple processed forms of the 21 calcium time-series. 22 28 Pnevmatikakis et al., 2016; Friedrich et al., 2017; Keemink et al., 2018; Giovannucci et al., 29 2019). As somatic calcium is proportional to the release of spikes, so we wish to use these 30 fluorescence time-series as a proxy for spiking activity in large, identified populations of 31 neurons. But raw calcium fluorescence is slow on the time-scale of spikes, nonlinearly 32 related to spiking, and contains noise from a range of sources. 33 These issues have inspired a wide range of deconvolution algorithms (Theis et al., 2016; 34 Berens et al., 2018; Stringer and Pachitariu, 2018), which attempt to turn raw somatic cal-35 1 cium into something more closely approximating spikes. Deconvolution algorithms them-36 selves range in complexity from simple deconvolution with a fixed kernel of the calcium 37 response (Yaksi and Friedrich, 2006), through detecting spike-evoked calcium events (Jew-38 ell and Witten, 2018; Pachitariu et al., 2016), to directly inferring spike times (Vogelstein 39 et al., 2010; Lütcke et al., 2013; Deneux et al., 2016). This continuum of options raises 40 the further question of the extent to which we should process the raw calcium signals. We 41 address here the question facing any systems neuroscientist using...

Information theoretic approaches to deciphering the neural code with functional fluorescence imaging

“…Overall, applying FOOPSI to fluorescence traces led to a poorer recovery of ground truth 798 information using SMGM compared to direct application of SMGM to the florescence traces ( � ). This 799 finding is consistent with recent research that has questioned the accuracy of deconvolution methods 800 (Evans et al, 2019). We also tested other metrics to measure mutual information directly from the 801 fluorescence time traces (KSG, Binned Estimator (uniform bins), and Binned Estimator (occupancy 802 binned)) and found these alternatives produced highly variable, saturating measurements of recovered 803 versus ground truth information.…”

“…When applied to spiking data, there is also 119 a change in units: rather than action potential counts, functional fluorescence traces are typically plotted 120 in units of florescence change with respect to the baseline (ΔF/F). One possible solution to these issues 121 would be to deconvolve calcium traces to recover APs; however, deconvolution is an active area of 122 research, and the accuracy of these methods has recently been questioned (Evans et al, 2019). Ideally, the 123 calcium traces could be used directly to measure spiking information, without the need for such an in 124 between, potentially error introducing, step.…”

Database and deep learning toolbox for noise-optimized, generalized spike inference from calcium imaging

“…Quantitative inference of spike rates is critical for the analysis of existing and future calcium imaging datasets 4,5,21 . Although ΔF/F can in theory only report on spike rate changes, we found that absolute spike rates can be reliably inferred when the baseline activity is sufficiently sparse, which was the case in all datasets examined here (Fig.…”

“…As a consequence, experimental conditions of novel datasets are often not well matched to those of available ground truth data. It is therefore not clear how an algorithm based on a specific ground truth dataset generalizes to other datasets, which causes problems for the inference of spike rates from calcium imaging data under most experimental conditions 12,13,20,21 .…”

Learning a tactile sequence induces selectivity to action decisions and outcomes in the mouse somatosensory cortex

“…We based our analyses of two-photon recordings on the fluorescence (ΔF/F0) time series rather than on activity reconstructed by deconvolution. The results of deconvolution methods can be highly sensitive to parameter choice, giving rise to variable conclusions as to neuronal response properties [109,110]. On the other hand, any analysis based on ΔF/F0 will overestimate neuronal correlations and give estimates on neuronal encoding of task variables that are likely to be a smoothed, filtered version of that occurring in reality [110,111].…”