Systems neuroscience aims to explain how complex adaptive behaviors can arise from the interactions of many individual neurons. As a result, population recordings—which capture the activity of multiple neurons simultaneously—have become the foundational method for progress in this domain. Extracellular electrophysiology and calcium-dependent two-photon optical physiology are by far the most prevalent population recording techniques, due to their single-neuron resolution, ease of use, and scalability. Recent advances have made it possible to record simultaneously from thousands of neurons with electrophysiology (Jun et al., 2017; Siegle et al., 2021; Stringer et al., 2019a) or tens of thousands of neurons with calcium imaging (Sofroniew et al., 2016; Stringer et al., 2019b; Weisenburger et al., 2019). While insights gained from both methods have been invaluable to the field, it is clear that neither technique provides a completely faithful picture of the underlying neural activity. In this study, our goal is to better understand the inherent biases of each recording modality, and specifically how to appropriately compare results obtained with one method to those obtained with the other.

Head-to-head comparisons of electrophysiology and imaging data are rare in the literature, but are critically important as the practical aspects of each method affect their suitability for different experimental questions. Since the expression of calcium indicators can be restricted to genetically defined cell types, imaging can easily target recordings to specific sub-populations (Madisen et al., 2015). Similarly, the use of retro- or anterograde viral transfections to drive indicator expression allows imaging to target sub-populations defined by their projection patterns (Glickfeld et al., 2013; Gradinaru et al., 2010). The ability to identify genetically or projection-defined cell populations in electrophysiology experiments is far more limited (Economo et al., 2018; Jia et al., 2019; Lima et al., 2009). Both techniques have been adapted for chronic recordings, but imaging offers the ability to reliably return to the same neurons over many days without the need to implant bulky hardware (Peters et al., 2014). Furthermore, because imaging captures structural, in addition to functional, data, individual neurons can be precisely registered to tissue volumes from electron microscopy (Bock et al., 2011; Lee et al., 2016), in vitro brain slices (Ko et al., 2011), and potentially other ex vivo techniques such as in situ RNA profiling (Chen et al., 2015). In contrast, the sources of extracellular spike waveforms are very difficult to localize with sufficient precision to enable direct cross-modal registration.

Inherent differences in the spatial sampling properties of electrophysiology and imaging are widely recognized, and influence what information can be gained from each method (Figure 1A). Multi-photon imaging typically yields data in a single plane tangential to the cortical surface, and is limited to depths of <1 mm due to a combination of light scattering and absorption in tissue. While multi-plane (Yang et al., 2016) and deep structure (Ouzounov et al., 2017) imaging are both areas of active research, imaging of most subcortical structures requires physical destruction of more superficial tissues (Dombeck et al., 2010; Feinberg and Meister, 2015; Skocek et al., 2018). Extracellular electrophysiology, on the other hand, utilizes microelectrodes embedded in the tissue, and thus dense recordings are easiest to perform along a straight line, normal to the cortical surface, in order to minimize per-channel tissue displacement. Linear probes provide simultaneous access to neurons in both cortex and subcortical structures, but make it difficult to sample many neurons from the same cortical layer.

Figure 1

Download asset Open asset

The temporal resolutions of these two methodologies also differ in critical ways (Figure 1B). Imaging is limited by the dwell time required to capture enough photons to distinguish physiological changes in fluorescence from noise (Svoboda and Yasuda, 2006), and the kinetics of calcium-dependent indicators additionally constrain the ability to temporally localize neural activity (Chen et al., 2013). While kilohertz-scale imaging has been achieved (Kazemipour et al., 2019; Zhang et al., 2019), most studies are based on data sampled at frame rates between 1 and 30 Hz. In contrast, extracellular electrophysiology requires sampling rates of 20 kHz or higher, in order to capture the action potential waveform shape that is essential for accurate spike sorting. High sampling rates allow extracellular electrophysiology to pin-point neural activity in time with sub-millisecond resolution, enabling analyses of fine-timescale synchronization across simultaneously recorded neural populations. The fact that electrophysiology can measure action potentials—what we believe to be the fundamental currency of neuronal communication and causation—bestows upon it a more basic ontological status than on calcium imaging, which captures an indirect measure of a neuron’s spike train.

To date, there has been no comprehensive attempt to characterize how the choice of recording modality affects the inferred functional properties of neurons in sensory cortex. Our limited understanding of how scientific conclusions may be skewed by the recording modality represents the weakest link in the chain of information integration across the techniques available to neurophysiologists today. To address this, we took advantage of two recently collected large-scale datasets that sampled neural activity in mouse visual cortex using either two-photon calcium imaging (de Vries et al., 2020) or dense extracellular electrophysiology (Siegle et al., 2021). These datasets were collected using standardized pipelines, such that the surgical methods, experimental steps, and physical geometry of the recording rigs were matched as closely as possible (Figure 1C,D). The overall similarity of these Allen Brain Observatory pipelines eliminates many of the potential confounding factors that arise when comparing results from imaging and electrophysiology experiments. We note that this is not an attempt at calibration against ground truth data, but rather an attempt to reconcile results across two uniquely comprehensive datasets collected under highly standardized conditions.

Our comparison focused on metrics that capture three fundamental features of neural responses to environmental stimuli: (1) responsiveness, (2) preference (i.e. the stimulus condition that maximizes the peak response), and (3) selectivity (i.e. sharpness of tuning). Responsiveness metrics characterize whether or not a particular stimulus type (e.g. drifting gratings) reproducibly elicits increased activity. For responsive neurons, preference metrics (e.g. preferred temporal frequency) determine which stimulus condition (out of a finite set) elicits the largest response, and serve as an indicator of a neuron’s functional specialization—for example, whether it responds preferentially to slow- or fast-moving stimuli. Lastly, selectivity metrics (e.g. orientation selectivity, lifetime sparseness) characterize a neuron’s ability to distinguish between particular exemplars within a stimulus class. All three of these features must be measured accurately in order to understand how stimuli are represented by individual neurons.

We find that preference metrics are largely invariant across modalities. However, in this dataset, electrophysiology suggests that neurons show a higher degree of responsiveness, while imaging suggests that responsive neurons show a higher degree of selectivity. In the absence of steps taken to mitigate these differences, the two modalities will yield mutually incompatible conclusions about basic neural response properties. These differences could be reduced by lowering the amplitude threshold for valid ΔF/F events, applying a spikes-to-calcium forward model to the electrophysiology data (Deneux et al., 2016), or sub-selection of neurons based either on event rate or by contamination level (the likelihood that signal from other neurons is misattributed to the neurons under consideration). This reconciliation reveals the respective biases of these two recording modalities, namely that extracellular electrophysiology predominantly captures the activity of highly active units while missing or merging low-firing-rate units, while calcium-indicator binding dynamics sparsify neural responses and supralinearly amplify spike bursts.

We compared the visual responses measured in the Allen Brain Observatory Visual Coding (‘imaging’) and Allen Brain Observatory Neuropixels (‘ephys’) datasets, publicly available through brain-map.org and the AllenSDK Python package. These datasets consist of recordings from neurons in six cortical visual areas (as well as subcortical areas in the Neuropixels dataset) in the awake, head-fixed mouse in response to a battery of passively viewed visual stimuli. For both datasets, the same drifting gratings, static gratings, natural scenes, and natural movie stimuli were shown (Figure 1E). These stimuli were presented in a single 3 hr recording session for the ephys dataset. For the imaging dataset, these stimuli were divided across three separate 1 hr imaging sessions from the same group of neurons. In both ephys and imaging experiments, mice were free to run on a rotating disc, the motion of which was continuously recorded.

The imaging dataset was collected using genetically encoded GCaMP6f (Chen et al., 2013) under the control of specific Cre driver lines. These Cre drivers limit the calcium indicator expression to specific neuronal populations, including different excitatory and inhibitory populations found in specific cortical layers (see de Vries et al., 2020 for details). The ephys dataset also made use of transgenic mice in addition to wild-type mice. These transgenic mice expressed either channelrhodopsin in specific inhibitory populations for identification using optotagging (see Siegle et al., 2021 for details), or GCaMP6f in specific excitatory or inhibitory populations (see Materials and methods). Unlike in the imaging dataset, however, these transgenic tools did not determine which neurons could be recorded.

We limited our comparative analysis to putative excitatory neurons from five cortical visual areas (V1, LM, AL, PM, and AM). In the case of the imaging data, we only included data from 10 excitatory Cre lines, while for ephys we limited our analysis to regular-spiking units by setting a threshold on the waveform duration (>0.4 ms). After this filtering step, we were left with 41,578 neurons from 170 mice in imaging, and 11,030 neurons from 52 mice in ephys. The total number of cells for each genotype, layer, and area is shown in Figure 1F.

Calculating response magnitudes for both modalities

In order to directly compare results from ephys and imaging, we first calculated the magnitude of each neuron’s response to individual trials, which were defined as the interval over which a stimulus was present on the screen. We computed a variety of metrics based on these response magnitudes, and compared the overall distributions of those metrics for all the neurons in each visual area. The methods for measuring these responses necessarily differ between modalities, as explained below.

For the ephys dataset, stimulus-evoked responses were computed using the spike times identified by Kilosort2 (Pachitariu et al., 2016a; Stringer et al., 2019a). Kilosort2 uses information in the extracellularly recorded voltage traces to find templates that fit the spike waveform shapes of all the units in the dataset, and assigns a template to each spike. The process of ‘spike sorting’—regardless of the underlying algorithm—does not perfectly recover the true underlying spike times, and has the potential to miss spikes (false negatives) or assign spikes (or noise waveforms) to the wrong unit (false positives). The magnitude of the response for a given trial was determined by counting the total number of spikes (including false positives and excluding false negatives) that occured during the stimulation interval. This spike-rate–based analysis is the de facto standard for analyzing electrophysiology data, but it washes out information about bursting or other within-trial dynamics. For example, a trial that includes a four-spike burst will have the same apparent magnitude as a trial with four isolated spikes (Figure 2A).

Figure 2 with 4 supplements see all

Download asset Open asset

Methods for determining response magnitudes for neurons in imaging datasets are less standardized, and deserve careful consideration. The most commonly used approach involves averaging the continuous, baseline-normalized fluorescence signal over the trial interval. This method relies on information that is closer to the raw data. However, it suffers the severe drawback that, due to the long decay time of calcium indicators, activity from one trial can contaminate the fluorescence trace during the next trial, especially when relatively short (<1 s) inter-stimulus intervals are used. To surmount this problem, one can attempt to determine the onset of abrupt changes in fluorescence and analyze these extracted ‘events,’ rather than the continuous trace. There are a variety of algorithms available for this purpose, including non-negative deconvolution (Vogelstein et al., 2010; Friedrich et al., 2017), approaches that model calcium binding kinetics (Deneux et al., 2016; Greenberg et al., 2018), and methods based on machine learning (Theis et al., 2016; Berens et al., 2018; Rupprecht et al., 2021). For our initial comparison, we extracted events using the same method we applied to our previous analysis of the large-scale imaging dataset (de Vries et al., 2020). This algorithm finds event times by reframing ℓ₀-regularized deconvolution as a change point detection problem that has a mathematically guaranteed, globally optimal ‘exact’ solution (hereafter, ‘exact ℓ₀’; Jewell and Witten, 2018; Jewell et al., 2018). The algorithm includes a sparsity constraint (λ) that is calibrated to each neuron’s overall noise level.

For the most part, the events that are detected from the 2P imaging data do not represent individual spikes, but rather are heavily biased towards indicating short bouts of high firing rate, for example bursting (Huang et al., 2021). There is, however, rich information contained in the amplitudes of these events, which have a non-linear—albeit on average monotonic—relationship with the underlying number of true spikes within a window. Therefore, in our population imaging dataset, we calculated the trial response magnitude by summing the amplitudes of events that occurred during the stimulation interval (Figure 2B). In example trials for the same hypothetical neuron recorded with both modalities (Figure 2A,B), the response magnitudes are equivalent from the perspective of electrophysiology. However, from the perspective of imaging, the trial that includes a spike burst (which results in a large influx of calcium) may have an order-of-magnitude larger response than a trial that only includes isolated spikes.

Impact of event detection on functional metrics

We sought to determine whether our approach to extracting events from the 2P data could explain between-modality differences in responsiveness and selectivity. Prior work has shown that scientific conclusions can depend on both the method of event extraction and the chosen parameters (Evans et al., 2019). However, the impact of different algorithms on functional metrics has yet to be assessed in a systematic way. To address this shortcoming, we first compared two event detection algorithms, exact ℓ₀ and unpenalized non-negative deconvolution (NND), another event extraction method that performs well on a ground truth dataset of simultaneous two-photon imaging and loose patch recordings from primary visual cortex (Huang et al., 2021). Care was taken to ensure that the characteristics of the ground truth imaging data matched those of our large-scale population recordings in terms of their imaging resolution, frame rate, and noise levels, which implicitly accounted for differences in laser power across experiments.

The correlation between the ground truth firing rate and the overall event amplitude within a given time bin is a common way of assessing event extraction performance (Theis et al., 2016; Berens et al., 2018; Rupprecht et al., 2021). Both algorithms performed equally well in terms of their ability to predict the instantaneous firing rate (for 100 ms bins, exact ℓ₀ r = 0.48 ± 0.23; NND r = 0.50 ± 0.24; p = 0.1, Wilcoxon signed-rank test). However, this metric does not capture all of the relevant features of the event time series. In particular, it ignores the rate of false positive events that are detected in the absence of a true underlying spike (Figure 3A). We found that the exact ℓ₀ method, which includes a built-in sparsity constraint, had a low rate of false positives (8 ± 2% in 100 ms bins, N = 32 ground truth recordings), whereas NND had a much higher rate (21 ± 4%; p = 8e-7, Wilcoxon signed-rank test). Small-amplitude false positive events have very little impact on the overall correlation between the ground truth spike rate and the extracted events, so parameter optimization does not typically penalize such events. However, we reasoned that the summation of many false positive events could have a noticeable impact on response magnitudes averaged over trials. Because these events always have a positive sign, they cannot be canceled out by low-amplitude negative deflections of similar magnitude, as would occur when analyzing ΔF/F directly.

Figure 3

Download asset Open asset

We tested the impact of applying a minimum-amplitude threshold to the event time series obtained via NND. If the cumulative event amplitude within a given time bin (100 ms) did not exceed a threshold (set to a multiple of each neuron’s estimated noise level), the events in that window were removed. As expected, this procedure resulted in almost no change in the correlation between event amplitude and the ground truth firing rate (Figure 3B). However, it had a noticeable impact on both the average response magnitude within a given time window (Figure 3C), as well as the false positive rate (Figure 3D).

Applying the same thresholding procedure to an example neuron from our population imaging dataset demonstrates how low-amplitude events can impact a cell’s apparent selectivity level. The prevalence of such events differs between the two compared approaches to event extraction, the exact ℓ₀ method used in Figure 2 and NND with no regularization (Figure 3E). When summing event amplitudes over many drifting gratings presentations, the difference in background rate has a big impact on the measured value of global orientation selectivity (gOSI), starting from 0.91 when using the exact ℓ₀ and dropping to 0.45 when using NND. However, these differences could be reconciled simply by setting a threshold to filter out low-amplitude events.

Extracting events from a population dataset (N = 3095 Slc17a7+ neurons from V1) using NND resulted in much lower measured overall selectivity levels, even lower than for electrophysiology (Figure 3F). Thresholding out events at multiples of each neuron’s noise level (σ) raised selectivity levels; a threshold between 3 and 4σ brought the selectivity distribution closest to the ephys distribution, while a threshold between 4 and 5σ resulted in selectivity that roughly matched that obtained with exact ℓ₀ (Figure 2G). The rate of low-amplitude events also affected responsiveness metrics (Figure 3G). Responsiveness was highest when all detected events were included, matching or slightly exceeding the levels measured with ephys. Again, applying an amplitude threshold between 4 and 5σ brought responsiveness to the level originally measured with exact ℓ₀. By imposing the same noise-based threshold on the minimum event size that we originally computed to determine optimal regularization strength (essentially performing a post-hoc regularization), we are able to reconcile the results obtained with NND with those obtained via exact ℓ₀.

This analysis demonstrates that altering the 2P event extraction methodology represents one possible avenue for reconciling results from imaging and ephys. However, as different parameters were needed to reconcile either selectivity or responsiveness, and the optimal parameters further depend on the presented stimulus class, this cannot be the whole story. Furthermore, relying only on 2P event extraction parameters to reconcile results across modalities implies that the ephys data is itself unbiased, and all we need to do is adjust our imaging analysis pipeline until our metrics match. Because we know this is not the case, we explored the potential impacts of additional factors on the discrepancies between our ephys and imaging datasets.

Controlling for transgene expression

Given that imaging (but not ephys) approaches fundamentally require the expression of exogenous proteins (e.g. Cre, tTA, and GCaMP6f in the case of our transgenic mice) in specific populations of neurons, we sought to determine whether such foreign transgenes, expressed at relatively high levels, could alter the underlying physiology of the neural population. All three proteins have been shown to have neurotoxic effects under certain conditions (Han et al., 2012; Schmidt-Supprian and Rajewsky, 2007; Steinmetz et al., 2017), and calcium indicators, which by design bind intracellular calcium, can additionally interfere with cellular signaling pathways. To examine whether the expression of these genes could explain the differences in functional properties inferred from imaging and ephys experiments, we performed electrophysiology in mice that expressed GCaMP6f under the control of specific Cre drivers. We collected data from mice with GCaMP6f expressed in dense excitatory lines (Cux2 and Slc17a7) or in sparse inhibitory lines (Vip and Sst), and compared the results to those obtained from wild-type mice (Figure 4A). On average, we recorded 45.9 ± 7.5 neurons per area in 17 wild-type mice, and 55.8 ± 15.6 neurons per area in 19 GCaMP6f transgenic mice (Figure 4B). The distribution of firing rates of recorded neurons in mice from all Cre lines was similar to the distribution for units in wild-type mice (Figure 4C).

Figure 4

Download asset Open asset

Because some GCaMP mouse lines have been known to exhibit aberrant seizure-like activity (Steinmetz et al., 2017), we wanted to check whether spike bursts were more prevalent in these mice. We detected bursting activity using the LogISI method, which identifies bursts in a spike train based on an adaptive inter-spike interval threshold (Pasquale et al., 2010). The dense excitatory Cre lines showed a slight increase in burst fraction (the fraction of all spikes that participate in bursts) compared to wild-type mice (Figure 4D). This minor increase in burstiness, however, was not associated with changes in responsiveness or selectivity metrics that could account for the baseline differences between the ephys and imaging datasets. The fraction of responsive neurons was not lower in the GCaMP6f mice, as it was for the imaging dataset—in fact, in some visual areas there was an increase in responsiveness in the GCaMP6f mice compared to wild-type (Figure 4E). In addition, the distribution of selectivities was largely unchanged between wild-type and GCaMP6f mice (Figure 4F). Thus, while there may be subtle differences in the underlying physiology of GCaMP6f mice, particularly in the dense excitatory lines, those differences cannot explain the large discrepancies in visual response metrics derived from ephys or imaging.

Forward-modeling synthetic imaging data from experimental ephys data

Given the substantial differences between the properties of extracellularly recorded spikes and events extracted from fluorescence traces (Figure 2A–C), and the potential impact of event extraction parameters on derived functional metrics (Figure 3), we hypothesized that transforming spike trains into simulated calcium events could reconcile some of the baseline differences in response metrics we have observed. The inverse transformation—converting fluorescence events into synthetic spike times—is highly under-specified, due to the reduced temporal resolution of calcium imaging (Figure 1B).

To implement the spikes-to-calcium transformation, we used MLSpike, a biophysically inspired forward model (Deneux et al., 2016). MLSpike explicitly considers the cooperative binding between GCaMP and calcium to generate synthetic ΔF/F fluorescence traces using the spike trains for each unit recorded with ephys as input. We extracted events from these traces using the same exact ℓ₀-regularized detection algorithm applied to our experimental imaging data, and used these events as inputs to our functional metrics calculations (Figure 5A).

Figure 5 with 2 supplements see all

Download asset Open asset

A subset of the free parameters in the MLSpike model (e.g. ΔF/F rise time, Hill parameter, saturation parameter, and normalized resting calcium concentration) were fit to simultaneously acquired loose patch and two-photon-imaging recordings from layer 2/3 of mouse visual cortex (Huang et al., 2021). Additionally, three parameters were calibrated on the fluorescence traces from the imaging dataset to capture the neuron-to-neuron variance of these parameters: the average amplitude of a fluorescence transient in response to a spike burst (A), the decay time of the fluorescence transients (τ), and the level of Gaussian noise in the signal (σ) (Figure 5B). For our initial characterization, we selected parameter values based on the mode of the overall distribution from the imaging dataset.

The primary consequence of the forward model was to ‘sparsify’ each neuron’s response by washing out single spikes while non-linearly boosting the amplitude of ‘bursty’ spike sequences with short inter-spike intervals. When responses were calculated on the ephys spike train, a trial containing a 4-spike burst within a 250 ms window would have the same magnitude as a trial with four isolated spikes across the 2 s trial. After the forward model, however, the burst would be transformed into an event with a magnitude many times greater than the events associated with isolated spikes, due to the nonlinear relationship between spike counts and the resulting calcium-dependent fluorescence. This effect can be seen in stimulus-locked raster plots for the same neuron before and after applying the forward model (Figure 5C).

What effects does this transformation have on neurons’ inferred functional properties? Applying the forward model plus event extraction to the ephys data did not systematically alter the fraction of responsive units in the dataset. While 8% of neurons switched from being responsive to drifting gratings to unresponsive, or vice versa, they did so in approximately equal numbers (Figure 5D). The forward model did not improve the match between the distributions of response reliabilities (our responsiveness metric) for any stimulus type (Figure 5E). The forward model similarly had a negligible impact on preference metrics; for example, only 14% of neurons changed their preferred temporal frequency after applying the forward model (Figure 5F), and the overall distribution of preferred temporal frequencies still matched that from the imaging experiments (Figure 5G). In contrast, nearly all neurons increased their selectivity after applying the forward model (Figure 5H). Overall, the distribution of lifetime sparseness to drifting gratings became more similar to—but still did not completely match—the imaging distribution across all areas (Figure 5I). The average Jensen–Shannon distance between the ephys and imaging distributions was 0.41 before applying the forward model, compared to 0.14 afterward (mean bootstrapped distance between the sub-samples of the imaging distribution = 0.064; p < 0.001 for all areas, since 1000 bootstrap samples never exceeded the true Jensen–Shannon distance; see Materials and methods for details). These results imply that the primary effects of the forward model—providing a supralinear boost to the ‘amplitude’ of spike bursts, and thresholding out single spike events—can account for baseline differences in selectivity, but not responsiveness, between ephys and imaging.

To assess whether the discrepancies between the imaging and ephys distributions of responsiveness and selectivity metrics could be further reduced by using a different set of forward model parameters, we brute-force sampled 1000 different parameter combinations for one ephys session, using 10 values each for amplitude, decay time, and noise level (Figure 5—figure supplement 1A), spanning the entire range of parameters calibrated on the experimental imaging data. The fraction of responsive neurons did not change as a function of forward model parameters, except for the lowest values of amplitude and noise level, where it decreased substantially (Figure 5—figure supplement 1B). This parameter combination (A ≤ 0.0015, sigma ≤ 0.03) was observed in less than 1% of actual neurons recorded with two-photon imaging, so it cannot account for differences in responsiveness between the two modalities. Both the difference between the median lifetime sparseness for imaging and ephys, as well as the Jensen–Shannon distance between the full ephys and imaging lifetime sparseness distributions, were near the global minimum for the parameter values we initially used (Figure 5—figure supplement 1C,D).

It is conceivable that the inability of the forward model to fully reconcile differences in responsiveness and selectivity was due to the fact that we applied the same parameters across all neurons of the ephys dataset, without considering their genetically defined cell type. To test for cell-type-specific differences in forward model parameters, we examined the distributions of amplitude, decay time, and noise level for individual excitatory Cre lines used in the imaging dataset. The distributions of parameter values across genotypes were largely overlapping, with the exception of increasing noise levels for some of the deeper populations (e.g. Tlx3-Cre in layer 5, and Ntsr1-Cre_GN220 in layer 6) and an abundance of low-amplitude neurons in the Fezf2-CreER population (Figure 5—figure supplement 2A). Given that higher noise levels and lower amplitudes did not improve the correspondence between the ephys and imaging metric distributions, we concluded that selecting parameter values for individual neurons based on their most likely cell type would not change our results. Furthermore, we saw no correlation between responsiveness or selectivity metrics in imaged neurons and their calibrated amplitude, decay time, or noise level (Figure 5—figure supplement 2B–E).

Effect of ephys selection bias

We next sought to determine whether electrophysiology’s well-known selection bias in favor of more active neurons could account for the differences between modalities. Whereas calcium imaging can detect the presence of all neurons in the field of view that express a fluorescent indicator, ephys cannot detect neurons unless they fire action potentials. This bias is exacerbated by the spike sorting process, which requires a sufficient number of spikes in order to generate an accurate template of each neuron’s waveform. Spike sorting algorithms can also mistakenly merge spikes from nearby neurons into a single ‘unit’ or allow background activity to contaminate a spike train, especially when spike waveforms generated by one neuron vary over time, for example due to the adaptation that occurs during a burst. These issues all result in an apparent activity level increase in ephys recordings. In addition, assuming a 50-μm ‘listening radius’ for the probes (radius of half-cylinder around the probe where the neurons’ spike amplitude is sufficiently above noise to trigger detection) (Buzsáki, 2004; Harris et al., 2016; Shoham et al., 2006), the average yield of 116 regular-spiking units/probe (prior to QC filtering) would imply a density of 42,000 neurons/mm³, much lower than the known density of ~90,000 neurons/mm³ for excitatory cells in mouse visual cortex (Erö et al., 2018).

If the ephys dataset is biased toward recording neurons with higher firing rates, it may be more appropriate to compare it with only the most active neurons in the imaging dataset. To test this, we systematically increased the event rate threshold for the imaged neurons, so the remaining neurons used for comparison were always in the upper quantile of mean event rate. Applying this filter increased the overall fraction of responsive neurons in the imaging dataset, such that the experimental imaging and synthetic imaging distributions had the highest similarity when between 7 and 39% of the most active imaged neurons were included (V1: 39%, LM: 34%, AL: 25%, PM: 7%, AM: 14%) (Figure 6A). This indicates that more active neurons tend to be more responsive to our visual stimuli, which could conceivably account for the discrepancy in overall responsiveness between the two modalities. However, applying this event rate threshold actually increased the differences between the selectivity distributions, as the most active imaged neurons were also more selective (Figure 6B). Thus, sub-selection of imaged neurons based on event rate was not sufficient to fully reconcile the differences between ephys and imaging. Performing the same analysis using sub-selection based on the coefficient of variation, an alternative measure of response reliability, yielded qualitatively similar results (Figure 6—figure supplement 1).

Figure 6 with 1 supplement see all

Download asset Open asset

If the ephys dataset includes spike trains that are contaminated with spurious spikes from one or more nearby neurons then it may help to compare our imaging results only to the least contaminated neurons from the ephys dataset. Our initial QC process excluded units with an inter-spike interval (ISI) violations score (Hill et al., 2011, see Materials and methods for definition) above 0.5, to remove highly contaminated units, but while the presence of refractory period violations implies contamination, the absence of such violations does not imply error-free clustering, so some contamination may remain. We systematically decreased our tolerance for ISI-violating ephys neurons, so the remaining neurons used for comparison were always in the lower quantile of contamination level. For the most restrictive thresholds, where there was zero detectable contamination in the original spike trains (ISI violations score = 0), the match between the synthetic imaging and experimental imaging selectivity and responsiveness distributions was maximized (Figure 6C,D). This indicates that, unsurprisingly, contamination by neighboring neurons (as measured by ISI violations score) reduces selectivity and increases responsiveness. Therefore, the inferred functional properties are most congruent across modalities when the ephys analysis includes a stringent threshold on the maximum allowable contamination level.

Results across stimulus types

The previous results have primarily focused on the drifting gratings stimulus, but we observe similar effects for all of the stimulus types shared between the imaging and ephys datasets. Figure 7 summarizes the impact of each transformation we performed, either before or after applying the forward model, for drifting gratings, static gratings, natural scenes, and natural movies.

Figure 7

Download asset Open asset

Across all stimulus types, the forward model had very little impact on responsiveness. Instead, sub-selecting the most active neurons from our imaging experiments using an event-rate filter rendered the shape of the distributions the most similar. For the stimulus types for which we could measure preference across small number of categories (temporal frequency of drifting gratings and spatial frequency of static gratings), no data transformations were able to improve the overall match between the ephys and imaging distributions, as they were already very similar in the baseline comparison. For selectivity metrics (lifetime sparseness), applying the forward model played the biggest role in improving cross-modal similarity, although there was a greater discrepancy between the resulting distributions for static gratings, natural scenes, and natural movies than there was for drifting gratings. Filtering ephys neurons based on ISI violations further reduced the Jensen–Shannon distance, but it still remained well above zero. This indicates that the transformations we employed could not fully reconcile observed differences in selectivity distributions between ephys and imaging.

Interpreting higher order analyses in light of our findings

The differences in responsiveness and selectivity metrics computed from the ephys and imaging datasets suggest that functional properties of neurons in the mouse visual cortex can appear to be dependent on the choice of recording modality. These effects extend beyond responsiveness and selectivity, as higher order analyses often build on these more fundamental metrics. Here, we focus on a representative example, a functional classification scheme based on responses to four stimulus types (drifting gratings, static gratings, natural scenes, and natural movies), which, in our previous work, revealed distinct response classes in the imaging dataset (de Vries et al., 2020). We aim to illustrate, using this recent example, that the effect of recording modality on responsiveness goes far beyond merely impacting how many neurons might be included in an analysis but may also propagate to conclusions we draw about the functional properties of neuronal (sub)populations. Our original study, based on imaging data alone, revealed that only ~10% of the neurons in the imaging dataset responded reliably to all four stimuli, while the largest class of neurons contained those that did not respond reliably to any of the stimuli. This classification result suggested that many neurons in the mouse visual cortex are not well described by classical models and may respond to more intricate visual or non-visual features. Here, we perform an analogous analysis on the ephys dataset to show how this classification is impacted by modality specific biases.

As in our original study, we performed unsupervised clustering on the 4 x N matrix of response reliabilities for each unit’s preferred condition for each stimulus class, where N represents the number of units (Figure 8A). The resulting clusters were assigned class labels based on whether their mean response reliability was above an initial threshold of 25% (to match the percentage we used in our previous analysis of the imaging dataset). After labeling each cluster, we calculated the fraction of units belonging to each functional class. We averaged over 100 different clustering initializations to obtain the average class membership of units across runs (Figure 8B).

Figure 8

Download asset Open asset

Running this analysis on the ephys dataset ‘naively’, seemed to reveal a very different landscape of functional classes. In stark contrast to the published imaging results, for ephys, the largest class (~40% of the units) contained units that responded reliably to all four classes of stimuli, while the class that did not respond reliably to any of the stimuli was empty. This is consistent with the observation that responsiveness is higher in the ephys dataset for each stimulus type (Figure 2F). To account for this bias, we systematically raised the responsiveness threshold used to group clusters into functional classes for the ephys dataset and found that the distribution of classes became more similar to the distribution of classes for the imaging dataset (Figure 8C). A threshold of 40% response reliability for the ephys dataset minimized the Jensen–Shannon distance between the distributions, rendering the class assignments in ephys remarkably similar to those for the imaging dataset (Figure 8E). The class labels for each neuron reflect the pattern of cross-stimulus response reliability, and as such provide an indication of its ‘meta-preference’ for different stimulus types. Thus, once we account for the generally higher levels of responsiveness seen in the ephys dataset, we observe similar meta-preferences and thus functional organization as for the imaging dataset. This example highlights how recording-modality-specific biases can affect higher-order conclusions about functional properties, and how a fundamental understanding of such biases can be leveraged to explain and resolve apparent contradictions.

In this study, we have compared response metrics derived from mouse visual cortex excitatory populations collected under highly standardized conditions, but using two different recording modalities. Overall, we observe similar stimulus preferences across the two datasets (e.g. preferred temporal frequencies within each visual area), but we see systematic differences in responsiveness and selectivity. Prior to any attempt at reconciliation, electrophysiological recordings showed a higher fraction of units with stimulus-driven activity, while calcium imaging showed higher selectivity (sharper tuning) among responsive neurons.

Our comparison of 2P event detection methods showed that the rate of small-amplitude events can influence inferred functional metrics. While the prevalence of false positives in ground truth data has been analyzed previously (Theis et al., 2016; Rupprecht et al., 2021), this metric is not typically used to optimize event detection parameters. Instead, correlation with ground truth rate is preferred (Berens et al., 2018; Pachitariu et al., 2018). However, this measure does not account for changes in background rate associated with the prevalence of small-amplitude events. These events always have a positive magnitude, are often triggered by noise, and can dramatically affect measured selectivity levels. In fact, most neurons have selectivity levels near zero when all events detected by non-negative deconvolution are included in the analysis (Figure 3F). Measured responsiveness, on the other hand, was highest when all events were included (Figure 3G). While this could indicate a greater sensitivity to true underlying spikes, it could also result from contamination by fluctuations in background fluorescence during the visual stimulation interval. Because of confounds such as these, we found it more informative to carry out our comparison on ephys data that has been transformed by the forward model and processed using the same event extraction method as the experimental imaging data.

Notably, the forward model boosted selectivity of the ephys data due to the sparsifying effect of calcium dynamics and the exact ℓ₀ event extraction step. As large, burst-dependent calcium transients can have amplitudes several orders of magnitude above the median amplitude (Figure 2B), this causes the response to the preferred stimulus condition to be weighted more heavily in selectivity calculations than non-preferred conditions. Similarly, isolated spikes during non-preferred conditions can be virtually indistinguishable from noise when viewed through the lens of the forward model. When the same trials are viewed through the lens of electrophysiology, however, spike counts increase more or less linearly, leading to the appearance of lower selectivity.

Unexpectedly, the forward model did not change responsiveness metrics. We initially hypothesized that the lower responsiveness in the imaging dataset was due to the fact that single-spike events are often not translated into detectable calcium transients; in ground truth recordings, around 75% of 100 ms bins with actual spikes do not result in any events detected by the exact ℓ₀ method with the default sparsity constraint (Huang et al., 2021). Instead, our observation of unchanged responsiveness following the forward model suggests that differences between modalities are more likely due to electrophysiological sampling bias—that is extracellular recordings missing small or low-firing rate neurons, or merging spike trains from nearby cells. In order to reconcile some of these differences, we could either apply a very strict threshold on ISI violations score to the ephys dataset, or remove between 61 and 93% of the least active neurons from the imaging dataset (Figure 6). This should serve as a cautionary tale for anyone estimating the fraction of neurons in an area that appear to increase their firing rate in response to environmental or behavioral events. Without careful controls, contamination from other neurons in the vicinity can make this fraction appear artificially high. While it may be possible to minimize these impurities with improved spike-sorting algorithms, there will always be neurons that even the best algorithms will not be able to distinguish in the face of background noise.

Differences in laminar sampling and running behavior between the two modalities had almost no effect in our comparisons. The transgenic expression of GCaMP6f in specific neural populations also did not impact the distributions of functional metrics. Finally, the initial parameters chosen for the forward model produced metric distributions that were close to the optimum match over the realistic parameter space. Therefore, we conclude that the primary contribution to differences in the considered ephys and imaging metrics comes from (1) the intrinsic nature of the spikes-to-calcium transfer function (2) the selection bias of extracellular electrophysiology recordings.

Even after accounting for these known factors to the best of our ability, the overall population from our imaging experiments still displayed higher selectivity than its ephys counterpart (Figure 5I, Figure 8B–D). What could account for these remaining differences? One possibility is that there may be residual undetected contamination in our ephys recordings. An ISI violations score of 0 does not guarantee that there is no contamination, just that we are not able to measure it using this metric. Sampling the tissue more densely (i.e. increasing the spatial resolution of spike waveforms) or improving spike sorting methods could reduce this issue. Another possibility is that ‘missed’ spikes—especially those at the end of a burst—could result in reduced amplitudes for the simulated calcium transients. In addition, if the in vivo spikes-to-calcium transfer function is non-stationary, there could be stimulus-dependent changes in calcium concentration that are not captured by a forward model that takes a spike train as its only input. Simultaneous cell-attached electrophysiology and two-photon imaging experiments have demonstrated the existence of ‘prolonged depolarization events’ (PDEs) in some neurons that result in very large increases in calcium concentration, but that are indistinguishable from similar burst events in extracellular recordings (Ledochowitsch et al., 2019).

One potential limitation of our approach is that we have only imaged the activity of transgenically expressed calcium indicators, rather than indicators expressed using a viral approach. In addition, the vast majority of our imaging data comes from mice expressing GCaMP6f, with only a small number of GCaMP6s neurons recorded. While we would ideally want to perform the same experiments with viral expression and GCaMP6s, this would require an expensive multi-year effort of similar magnitude to the one that produced our existing imaging dataset. Instead, we have chosen to simulate the effects of these alternative conditions. In our analysis of the forward model parameter sweep, functional metrics remain relatively constant for a wide range of amplitude and decay time parameters (Figure 5—figure supplement 1). The full range of this sweep includes decay times that are consistent with GCaMP6s, and event amplitudes that are consistent with viral expression.

The forward models currently available in the literature (Deneux et al., 2016; Wei et al., 2019) are of comparable power, in that their most complex instantiations allow for non-instantaneous fluorescence rise as well as for a non-linear relationship between calcium concentration and fluorescence. To the best of our knowledge, none of these forward models explicitly model nonstationarities in the spike-to-calcium transfer function. Moreover, all currently available models suffer from the drawback that fits to simultaneously recorded ground truth data yield significant variance in model parameters across neurons (Ledochowitsch et al., 2019; Wei et al., 2019). We strove to mitigate this shortcoming by showing that brute-force exploration of MLSpike model parameter space could not significantly improve the match between real and synthetic imaging data.

Another potential confound is elevated activity around an implanted probe, which was characterized in a recent study (Eles et al., 2018). By performing calcium imaging around a silicon probe that was recently introduced into the brain, the authors found increased intracellular calcium lasting for at least 30 min after implantation. If this is true in our Neuropixels recordings, it could at least partially account for the higher overall firing rates and responsiveness in electrophysiology compared to imaging. Careful simultaneous measurements will be required in order to account for this relative activity increase.

While results obtained from ephys and imaging are sometimes treated as if they were interchangeable from a scientific standpoint, in actuality they each provide related but fundamentally different perspectives on the underlying neural activity. Extracellular electrophysiology tells us—with sub-millisecond temporal resolution—about a neuron’s spiking output. Calcium imaging doesn't measure the outgoing action potentials directly, but rather the impact of input and output signals on a neuron’s internal state, in terms of increases in calcium concentration that drive various downstream pathways (West et al., 2001). While voltage-dependent fluorescent indicators may offer the best of both worlds, there are substantial technical hurdles to employing them on comparably large scales (Peterka et al., 2011). Thus, in order to correctly interpret existing and forthcoming datasets, we must account for the inherent biases of these two recording modalities.

A recent study comparing matched populations recorded with electrophysiology or imaging emphasized differences in the temporal profiles of spike trains and calcium-dependent fluorescence responses (Wei et al., 2019). The authors found that event-extraction algorithms that convert continuous ΔF/F traces to putative spike times could not recapitulate the temporal profiles measured with electrophysiology; on the other hand, a forward model that transformed spike times to synthetic ΔF/F traces could make their electrophysiology results appear more like those from the imaging experiments. Their conclusions were primarily based on metrics derived from the evolution of firing rates or ΔF/F ratios over the course of a behavioral trial. However, there are other types of functional metrics that are not explicitly dependent on temporal factors, such as responsiveness, which cannot be reconciled using a forward model alone.

It is worth emphasizing that we are not suggesting that researchers use the methods we describe here to attempt to make all of their imaging data more similar to electrophysiological data, or vice versa. Since no one single method is intrinsically superior to the other, doing so would merely introduce additional biases. Instead, we recommend that readers examine how sensitive their chosen functional metrics, and, by extension, their derived scientific conclusions are to (1) signal contamination during spike detection and sorting (for electrophysiology data), (2) application of a spike-to-calcium forward model (for electrophysiology data), (3) filtering by event rate (for imaging data), and (4) false positives introduced during event detection (for imaging data). If the chosen functional metrics are found to be largely insensitive to the above transformations, then results can be compared directly across studies that employ different recording modalities. Otherwise, the sensitivity analysis can be used as a means of establishing bounds on the magnitude and direction of expected modality-related discrepancies.

More work is needed to understand the detailed physiological underpinning of the modality-specific differences we have observed. One approach, currently underway at the Allen Institute and elsewhere, is to carry out recordings with extremely high-density silicon probes (Neuropixels Ultra), over 10x denser (in terms of the number of electrodes per unit of area) than the Neuropixels 1.0 probes used in this study. Such probes can capture each spike waveform with 100 or more electrodes, making it easier to disambiguate waveforms from nearby neurons, and making it less likely that neurons with small somata would evade detection or that their waveforms would be mistaken for those of other units. These experiments should make it easier to quantify the selection bias of extracellular electrophysiology, as well as the degree to which missed neurons and contaminated spike trains have influenced the results of the current study. In addition, experiments in which silicon probes are combined with two-photon imaging—either through interleaved sampling, spike-triggered image acquisition, or improved artifact removal techniques—could provide more direct ground-truth information about the relationship between extracellular electrophysiology and calcium imaging.

Overall, our comparison highlights the value of large-scale, standardized datasets. The fact that functional metrics are sensitive not only to experimental procedures but also to data processing steps and cell inclusion criteria (Mesa et al., 2021), makes it difficult to directly compare results across studies. Having access to ephys and imaging datasets collected under largely identical conditions allowed us to rule out a number of potential confounds, such as laminar sampling bias and inter-modality behavioral differences. And due to the technical difficulties of scaling simultaneous ephys/imaging experiments, these will, for the foreseeable future, continue to complement and validate large-scale unimodal datasets, rather than replace them.

Ultimately, the goal of this work is not to establish the superiority of any one recording modality in absolute terms, since their complementary strengths ensure they will each remain essential to scientific progress for many years to come. Instead, we want to establish guidelines for properly interpreting the massive amounts of data that have been or will be collected using either modality. From this study, we have learned that extracellular electrophysiology likely overestimates the fraction of neurons that elevate their activity in response to visual stimuli, in a manner that is consistent with the effects of selection bias and contamination. The apparent differences in selectivity underscore the fact that one must carefully consider the impact of data processing steps (such as event extraction from fluorescence time series), as well as what each modality is actually measuring. Selectivity metrics based on spike counts (the neuron’s outputs) will almost always be lower than selectivity metrics based on calcium concentrations (the neuron’s internal state). Even with this in mind, however, we cannot fully reproduce the observed levels of calcium-dependent selectivity using spike times alone—suggesting that a neuron’s internal state may contain stimulus-specific information that is not necessarily reflected in its outputs.

In summary, we have shown that reconciling results across modalities is not straightforward, due to biases that are introduced at the level of the data processing steps, the spatial characteristics of the recording hardware, and the physical signals being measured. We have attempted to account for these biases by (1) altering 2P event detection parameters, (2) applying sub-selection to account for built-in spatial biases, and (3) simulating calcium signals via a forward model. We have shown that functional metrics are sensitive to all of these biases, which makes it difficult to determine which ones are most impactful. For example, to what degree does the bias of ephys for higher-firing-rate units stem from the recording hardware versus the spike sorting procedure? It is possible that more spatially precise sampling, or modifications to the spike sorting algorithm, will reduce this bias in the future. Similarly, how much of the 2P imaging bias stems from the calcium indicator versus the event extraction algorithm? New indicators (such as GCaMP8) may ameliorate some of these problems, as could further optimization of the event extraction step. In the end, these two modalities provide two imperfect yet complementary lenses on the underlying neural activity, and their respective strengths and limitations must be understood when interpreting the recorded activities across experiments.

Code and intermediate data files required to generate all manuscript figures are available through GitHub (https://github.com/AllenInstitute/ophys_ephys_comparison_paper). NWB files for the imaging and ephys sessions (excluding the ephys experiments in GCaMP6 mice) are available via the AllenSDK (https://github.com/AllenInstitute/AllenSDK, copy archived at https://archive.softwareheritage.org/swh:1:rev:b29058f4c136f906123b3f93cd184edb9158b4e4), and also as an AmazonWeb Services Public Data Set (https://registry.opendata.aws/allen-brain-observatory/). NWB files for the ephys experiments in GCaMP6 mice are available through Zenodo (https://https://zenodo.org/record/5090747).

1. 1. Siegle JH
   2. Ledochowitsch P
   3. Jia X
   4. Millman DJ
   5. Ocker GK
   6. Caldejon S
   7. Casal L
   8. Cho A
   9. Denman DJ
   10. Durand S
   11. Groblewski PA
   12. Heller G
   13. Kato I
   14. Kivikas S
   15. Lecoq J
   16. Nayan C
   17. Ngo K
   18. Nicovich PR
   19. North K
   20. Ramirez TK
   21. Swapp J
   22. Waughman X
   23. Williford A
   24. Olsen SR
   25. Koch C
   26. Buice MA
   27. de Vries SE

   (2021) Zenodo

   Neuropixels recordings from GCaMP6f+ mice.

   https://doi.org/10.5281/zenodo.5090747

1. 1. Allen Institute for Brain Science

   (2016) Allen Brain Atlas

   ID observatory.brain-map.org/visualcoding/. Allen Brain Observatory Visual Coding.

   http://observatory.brain-map.org/visualcoding/
2. 1. Allen Institute for Brain Science

   (2019) Allen Brain Map

   ID portal.brain-map.org/explore/circuits/visual-coding-neuropixels. Allen Brain Observatory Neuropixels.

   https://portal.brain-map.org/explore/circuits/visual-coding-neuropixels

1. 1. Berens P
   2. Freeman J
   3. Deneux T
   4. Chenkov N
   5. McColgan T
   6. Speiser A
   7. Macke JH
   8. Turaga SC
   9. Mineault P
   10. Rupprecht P
   11. Gerhard S
   12. Friedrich RW
   13. Friedrich J
   14. Paninski L
   15. Pachitariu M
   16. Harris KD
   17. Bolte B
   18. Machado TA
   19. Ringach D
   20. Stone J
   21. Rogerson LE
   22. Sofroniew NJ
   23. Reimer J
   24. Froudarakis E
   25. Euler T
   26. Román Rosón M
   27. Theis L
   28. Tolias AS
   29. Bethge M

   (2018) Community-based benchmarking improves spike rate inference from two-photon calcium imaging data

   PLOS Computational Biology 14:e1006157.

   https://doi.org/10.1371/journal.pcbi.1006157

   • PubMed
   • Google Scholar
2. 1. Bock DD
   2. Lee WC
   3. Kerlin AM
   4. Andermann ML
   5. Hood G
   6. Wetzel AW
   7. Yurgenson S
   8. Soucy ER
   9. Kim HS
   10. Reid RC

   (2011) Network anatomy and in vivo physiology of visual cortical neurons

   Nature 471:177–182.

   https://doi.org/10.1038/nature09802

   • PubMed
   • Google Scholar
3. 1. Buzsáki G

   (2004) Large-scale recording of neuronal ensembles

   Nature Neuroscience 7:446–451.

   https://doi.org/10.1038/nn1233

   • PubMed
   • Google Scholar
4. 1. Chen TW
   2. Wardill TJ
   3. Sun Y
   4. Pulver SR
   5. Renninger SL
   6. Baohan A
   7. Schreiter ER
   8. Kerr RA
   9. Orger MB
   10. Jayaraman V
   11. Looger LL
   12. Svoboda K
   13. Kim DS

   (2013) Ultrasensitive fluorescent proteins for imaging neuronal activity

   Nature 499:295–300.

   https://doi.org/10.1038/nature12354

   • PubMed
   • Google Scholar
5. 1. Chen KH
   2. Boettiger AN
   3. Moffitt JR
   4. Wang S
   5. Zhuang X

   (2015) RNA imaging. spatially resolved, highly multiplexed RNA profiling in single cells

   Science 348:aaa6090.

   https://doi.org/10.1126/science.aaa6090

   • PubMed
   • Google Scholar
6. 1. Cotterill E
   2. Charlesworth P
   3. Thomas CW
   4. Paulsen O
   5. Eglen SJ

   (2016) A comparison of computational methods for detecting bursts in neuronal spike trains and their application to human stem cell-derived neuronal networks

   Journal of Neurophysiology 116:306–321.

   https://doi.org/10.1152/jn.00093.2016

   • PubMed
   • Google Scholar
7. 1. de Vries SEJ
   2. Lecoq JA
   3. Buice MA
   4. Groblewski PA
   5. Ocker GK
   6. Oliver M
   7. Feng D
   8. Cain N
   9. Ledochowitsch P
   10. Millman D
   11. Roll K
   12. Garrett M
   13. Keenan T
   14. Kuan L
   15. Mihalas S
   16. Olsen S
   17. Thompson C
   18. Wakeman W
   19. Waters J
   20. Williams D
   21. Barber C
   22. Berbesque N
   23. Blanchard B
   24. Bowles N
   25. Caldejon SD
   26. Casal L
   27. Cho A
   28. Cross S
   29. Dang C
   30. Dolbeare T
   31. Edwards M
   32. Galbraith J
   33. Gaudreault N
   34. Gilbert TL
   35. Griffin F
   36. Hargrave P
   37. Howard R
   38. Huang L
   39. Jewell S
   40. Keller N
   41. Knoblich U
   42. Larkin JD
   43. Larsen R
   44. Lau C
   45. Lee E
   46. Lee F
   47. Leon A
   48. Li L
   49. Long F
   50. Luviano J
   51. Mace K
   52. Nguyen T
   53. Perkins J
   54. Robertson M
   55. Seid S
   56. Shea-Brown E
   57. Shi J
   58. Sjoquist N
   59. Slaughterbeck C
   60. Sullivan D
   61. Valenza R
   62. White C
   63. Williford A
   64. Witten DM
   65. Zhuang J
   66. Zeng H
   67. Farrell C
   68. Ng L
   69. Bernard A
   70. Phillips JW
   71. Reid RC
   72. Koch C

   (2020) A large-scale standardized physiological survey reveals functional organization of the mouse visual cortex

   Nature Neuroscience 23:138–151.

   https://doi.org/10.1038/s41593-019-0550-9

   • PubMed
   • Google Scholar
8. 1. Deneux T
   2. Kaszas A
   3. Szalay G
   4. Katona G
   5. Lakner T
   6. Grinvald A
   7. Rózsa B
   8. Vanzetta I

   (2016) Accurate spike estimation from noisy calcium signals for ultrafast three-dimensional imaging of large neuronal populations in vivo

   Nature Communications 7:12190.

   https://doi.org/10.1038/ncomms12190

   • PubMed
   • Google Scholar
9. 1. Dombeck DA
   2. Harvey CD
   3. Tian L
   4. Looger LL
   5. Tank DW

   (2010) Functional imaging of hippocampal place cells at cellular resolution during virtual navigation

   Nature Neuroscience 13:1433–1440.

   https://doi.org/10.1038/nn.2648

   • PubMed
   • Google Scholar
10. 1. Economo MN
    2. Viswanathan S
    3. Tasic B
    4. Bas E
    5. Winnubst J
    6. Menon V
    7. Graybuck LT
    8. Nguyen TN
    9. Smith KA
    10. Yao Z
    11. Wang L
    12. Gerfen CR
    13. Chandrashekar J
    14. Zeng H
    15. Looger LL
    16. Svoboda K

    (2018) Distinct descending motor cortex pathways and their roles in movement

    Nature 563:79–84.

    https://doi.org/10.1038/s41586-018-0642-9

    • PubMed
    • Google Scholar
11. 1. Eles JR
    2. Vazquez AL
    3. Kozai TDY
    4. Cui XT

    (2018) In vivo imaging of neuronal calcium during electrode implantation: spatial and temporal mapping of damage and recovery

    Biomaterials 174:79–94.

    https://doi.org/10.1016/j.biomaterials.2018.04.043

    • PubMed
    • Google Scholar
12. 1. Erö C
    2. Gewaltig MO
    3. Keller D
    4. Markram H

    (2018) A cell atlas for the mouse brain

    Frontiers in Neuroinformatics 12:84.

    https://doi.org/10.3389/fninf.2018.00084

    • PubMed
    • Google Scholar
13. Preprint

    1. Evans MH
    2. Petersen RS
    3. Humphries MD

    (2019) On the use of calcium deconvolution algorithms in practical contexts

    bioRxiv.

    https://doi.org/10.1101/871137

    • Google Scholar
14. 1. Feinberg EH
    2. Meister M

    (2015) Orientation columns in the mouse superior colliculus

    Nature 519:229–232.

    https://doi.org/10.1038/nature14103

    • PubMed
    • Google Scholar
15. 1. Friedrich J
    2. Zhou P
    3. Paninski L

    (2017) Fast online deconvolution of calcium imaging data

    PLOS Computational Biology 13:e1005423.

    https://doi.org/10.1371/journal.pcbi.1005423

    • PubMed
    • Google Scholar
16. 1. Garrett ME
    2. Nauhaus I
    3. Marshel JH
    4. Callaway EM

    (2014) Topography and areal organization of mouse visual cortex

    Journal of Neuroscience 34:12587–12600.

    https://doi.org/10.1523/JNEUROSCI.1124-14.2014

    • PubMed
    • Google Scholar
17. 1. Glickfeld LL
    2. Andermann ML
    3. Bonin V
    4. Reid RC

    (2013) Cortico-cortical projections in mouse visual cortex are functionally target specific

    Nature Neuroscience 16:219–226.

    https://doi.org/10.1038/nn.3300

    • PubMed
    • Google Scholar
18. 1. Gradinaru V
    2. Zhang F
    3. Ramakrishnan C
    4. Mattis J
    5. Prakash R
    6. Diester I
    7. Goshen I
    8. Thompson KR
    9. Deisseroth K

    (2010) Molecular and cellular approaches for diversifying and extending optogenetics

    Cell 141:154–165.

    https://doi.org/10.1016/j.cell.2010.02.037

    • PubMed
    • Google Scholar
19. Preprint

    1. Greenberg DS
    2. Wallace DJ
    3. Voit K-M
    4. Wuertenberger S
    5. Czubayko U
    6. Monsees A
    7. Handa T
    8. Vogelstein JT
    9. Seifert R
    10. Groemping Y
    11. Kerr JND

    (2018) Accurate action potential inference from a calcium sensor protein through biophysical modeling

    bioRxiv.

    https://doi.org/10.1101/479055

    • Google Scholar
20. 1. Han HJ
    2. Allen CC
    3. Buchovecky CM
    4. Yetman MJ
    5. Born HA
    6. Marin MA
    7. Rodgers SP
    8. Song BJ
    9. Lu HC
    10. Justice MJ
    11. Probst FJ
    12. Jankowsky JL

    (2012) Strain background influences neurotoxicity and behavioral abnormalities in mice expressing the tetracycline transactivator

    Journal of Neuroscience 32:10574–10586.

    https://doi.org/10.1523/JNEUROSCI.0893-12.2012

    • PubMed
    • Google Scholar
21. 1. Harris KD
    2. Quiroga RQ
    3. Freeman J
    4. Smith SL

    (2016) Improving data quality in neuronal population recordings

    Nature Neuroscience 19:1165–1174.

    https://doi.org/10.1038/nn.4365

    • PubMed
    • Google Scholar
22. 1. Hill DN
    2. Mehta SB
    3. Kleinfeld D

    (2011) Quality metrics to accompany spike sorting of extracellular signals

    Journal of Neuroscience 31:8699–8705.

    https://doi.org/10.1523/JNEUROSCI.0971-11.2011

    • PubMed
    • Google Scholar
23. 1. Hoyer S
    2. Hamman JJ

    (2017) Xarray: n-d labeled arrays and datasets in Python

    Journal of Open Research Software 5:10.

    https://doi.org/10.5334/jors.148

    • Google Scholar
24. 1. Huang L
    2. Ledochowitsch P
    3. Knoblich U
    4. Lecoq J
    5. Murphy GJ
    6. Reid RC
    7. de Vries SE
    8. Koch C
    9. Zeng H
    10. Buice MA
    11. Waters J
    12. Li L

    (2021) Relationship between simultaneously recorded spiking activity and fluorescence signal in GCaMP6 transgenic mice

    eLife 10:e51675.

    https://doi.org/10.7554/eLife.51675

    • PubMed
    • Google Scholar
25. 1. Hunter JD

    (2007) Matplotlib: a 2D graphics environment

    Computing in Science & Engineering 9:90–95.

    https://doi.org/10.1109/MCSE.2007.55

    • Google Scholar
26. Preprint

    1. Jewell S
    2. Hocking TD
    3. Fearnhead P
    4. Witten D

    (2018) Fast nonconvex deconvolution of calcium imaging data

    arXiv.

    https://arxiv.org/pdf/1802.07380.pdf

    • Google Scholar
27. 1. Jewell S
    2. Witten D

    (2018) Exact spike train inference via ℓ₀optimization

    The Annals of Applied Statistics 12:2457–2482.

    https://doi.org/10.1214/18-AOAS1162

    • PubMed
    • Google Scholar
28. 1. Jia X
    2. Siegle JH
    3. Bennett C
    4. Gale SD
    5. Denman DJ
    6. Koch C
    7. Olsen SR

    (2019) High-density extracellular probes reveal dendritic backpropagation and facilitate neuron classification

    Journal of Neurophysiology 121:1831–1847.

    https://doi.org/10.1152/jn.00680.2018

    • PubMed
    • Google Scholar
29. 1. Jun JJ
    2. Steinmetz NA
    3. Siegle JH
    4. Denman DJ
    5. Bauza M
    6. Barbarits B
    7. Lee AK
    8. Anastassiou CA
    9. Andrei A
    10. Aydın Ç
    11. Barbic M
    12. Blanche TJ
    13. Bonin V
    14. Couto J
    15. Dutta B
    16. Gratiy SL
    17. Gutnisky DA
    18. Häusser M
    19. Karsh B
    20. Ledochowitsch P
    21. Lopez CM
    22. Mitelut C
    23. Musa S
    24. Okun M
    25. Pachitariu M
    26. Putzeys J
    27. Rich PD
    28. Rossant C
    29. Sun WL
    30. Svoboda K
    31. Carandini M
    32. Harris KD
    33. Koch C
    34. O'Keefe J
    35. Harris TD

    (2017) Fully integrated silicon probes for high-density recording of neural activity

    Nature 551:232–236.

    https://doi.org/10.1038/nature24636

    • PubMed
    • Google Scholar
30. 1. Kazemipour A
    2. Novak O
    3. Flickinger D
    4. Marvin JS
    5. Abdelfattah AS
    6. King J
    7. Borden PM
    8. Kim JJ
    9. Al-Abdullatif SH
    10. Deal PE
    11. Miller EW
    12. Schreiter ER
    13. Druckmann S
    14. Svoboda K
    15. Looger LL
    16. Podgorski K

    (2019) Kilohertz frame-rate two-photon tomography

    Nature Methods 16:778–786.

    https://doi.org/10.1038/s41592-019-0493-9

    • PubMed
    • Google Scholar
31. 1. Ko H
    2. Hofer SB
    3. Pichler B
    4. Buchanan KA
    5. Sjöström PJ
    6. Mrsic-Flogel TD

    (2011) Functional specificity of local synaptic connections in neocortical networks

    Nature 473:87–91.

    https://doi.org/10.1038/nature09880

    • PubMed
    • Google Scholar
32. Preprint

    1. Ledochowitsch P
    2. Huang L
    3. Knoblich U
    4. Oliver M
    5. Lecoq J
    6. Reid C
    7. Li L
    8. Zeng H
    9. Koch C
    10. Waters J

    (2019) On the correspondence of electrical and optical physiology in in vivo population-scale two-photon calcium imaging

    bioRxiv.

    https://doi.org/10.1101/800102

    • Google Scholar
33. 1. Lee WC
    2. Bonin V
    3. Reed M
    4. Graham BJ
    5. Hood G
    6. Glattfelder K
    7. Reid RC

    (2016) Anatomy and function of an excitatory network in the visual cortex

    Nature 532:370–374.

    https://doi.org/10.1038/nature17192

    • PubMed
    • Google Scholar
34. 1. Lima SQ
    2. Hromádka T
    3. Znamenskiy P
    4. Zador AM

    (2009) PINP: a new method of tagging neuronal populations for identification during in vivo electrophysiological recording

    PLOS ONE 4:e6099.

    https://doi.org/10.1371/journal.pone.0006099

    • PubMed
    • Google Scholar
35. 1. Lin J

    (1991) Divergence measures based on the shannon entropy

    IEEE Transactions on Information Theory 37:145–151.

    https://doi.org/10.1109/18.61115

    • Google Scholar
36. 1. Madisen L
    2. Garner AR
    3. Shimaoka D
    4. Chuong AS
    5. Klapoetke NC
    6. Li L
    7. van der Bourg A
    8. Niino Y
    9. Egolf L
    10. Monetti C
    11. Gu H
    12. Mills M
    13. Cheng A
    14. Tasic B
    15. Nguyen TN
    16. Sunkin SM
    17. Benucci A
    18. Nagy A
    19. Miyawaki A
    20. Helmchen F
    21. Empson RM
    22. Knöpfel T
    23. Boyden ES
    24. Reid RC
    25. Carandini M
    26. Zeng H

    (2015) Transgenic mice for intersectional targeting of neural sensors and effectors with high specificity and performance

    Neuron 85:942–958.

    https://doi.org/10.1016/j.neuron.2015.02.022

    • Google Scholar
37. Conference

    1. Martin D
    2. Fowlkes C
    3. Tal D
    4. Malik J

    (2001) A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics

    Proceedings of the Eighth IEEE International Conference on Computational Vision. pp. 416–423.

    https://doi.org/10.1109/ICCV.2001.937655

    • Google Scholar
38. Conference

    1. McKinney W

    (2010) Data structures for statistical computing in Python

    Proceedings of the 9th Python in Science Conference. pp. 51–56.

    https://doi.org/10.25080/MAJORA-92BF1922-00A

    • Google Scholar
39. 1. Mesa N
    2. Waters J
    3. de Vries SEJ

    (2021) The effect of inclusion criteria on the functional properties reported in mouse visual cortex

    Eneuro 8:ENEURO.0188-20.2021.

    https://doi.org/10.1523/ENEURO.0188-20.2021

    • PubMed
    • Google Scholar
40. Software

    1. Nicovich R

    (2021) AllenInstitute/AIBSOPT, version swh:1:rev:e38af7e25651fe7517dcf7ca3d38676e3c9c211e

    Software Heritage.

    https://archive.softwareheritage.org/swh:1:dir:8954e35a23b97f4ce14f9028e50f4bc5f073e44e;origin=https://github.com/AllenInstitute/AIBSOPT;visit=swh:1:snp:6b3d8870456826b0097613e88c8a513d46e6de1d;anchor=swh:1:rev:e38af7e25651fe7517dcf7ca3d38676e3c9c211e
41. 1. Niell CM
    2. Stryker MP

    (2010) Modulation of visual responses by behavioral state in mouse visual cortex

    Neuron 65:472–479.

    https://doi.org/10.1016/j.neuron.2010.01.033

    • PubMed
    • Google Scholar
42. 1. Olmos A
    2. Kingdom FA

    (2004) A biologically inspired algorithm for the recovery of shading and reflectance images

    Perception 33:1463–1473.

    https://doi.org/10.1068/p5321

    • PubMed
    • Google Scholar
43. 1. Ouzounov DG
    2. Wang T
    3. Wang M
    4. Feng DD
    5. Horton NG
    6. Cruz-Hernández JC
    7. Cheng YT
    8. Reimer J
    9. Tolias AS
    10. Nishimura N
    11. Xu C

    (2017) In vivo three-photon imaging of activity of GCaMP6-labeled neurons deep in intact mouse brain

    Nature Methods 14:388–390.

    https://doi.org/10.1038/nmeth.4183

    • PubMed
    • Google Scholar
44. Conference

    1. Pachitariu M
    2. Steinmetz NA
    3. Kadir SN
    4. Carandini M
    5. Harris KD

    (2016a)

    Fast and accurate spike sorting of high-channel count probes with KiloSort

    Advances in Neural Information Processing Systems.

    • Google Scholar
45. Preprint

    1. Pachitariu M
    2. Stringer C
    3. Dipoppa M
    4. Schröder S
    5. Rossi LF
    6. Dalgleish H
    7. Carandini M
    8. Harris KD

    (2016b) Suite2p: beyond 10,000 neurons with standard two-photon microscopy

    bioRxiv.

    https://doi.org/10.1101/061507

    • Google Scholar
46. 1. Pachitariu M
    2. Stringer C
    3. Harris KD

    (2018) Robustness of spike deconvolution for neuronal calcium imaging

    The Journal of Neuroscience 38:7976–7985.

    https://doi.org/10.1523/JNEUROSCI.3339-17.2018

    • PubMed
    • Google Scholar
47. Software

    1. Pachitariu M

    (2021) MouseLand/Kilosort, version swh:1:rev:db3a3353d9a374ea2f71674bbe443be21986c82c

    Software Heritage.

    https://archive.softwareheritage.org/swh:1:dir:f24b1f096a0f44b907337a5db613116e58f9ea47;origin=https://github.com/MouseLand/Kilosort;visit=swh:1:snp:a45f1825481650ddb3f287b8afeb71886738d762;anchor=swh:1:rev:db3a3353d9a374ea2f71674bbe443be21986c82c
48. 1. Pasquale V
    2. Martinoia S
    3. Chiappalone M

    (2010) A self-adapting approach for the detection of bursts and network bursts in neuronal cultures

    Journal of Computational Neuroscience 29:213–229.

    https://doi.org/10.1007/s10827-009-0175-1

    • PubMed
    • Google Scholar
49. Preprint

    1. Pedregosa F
    2. Varoquaux G
    3. Gramfort A
    4. Michel V
    5. Thirion B
    6. Grisel O
    7. Blondel M
    8. Müller A
    9. Nothman J
    10. Louppe G

    (2012) Scikit-learn: machine learning in Python

    arXiv.

    https://arxiv.org/abs/1201.0490

    • Google Scholar
50. 1. Peirce JW

    (2007) PsychoPy--psychophysics software in python

    Journal of Neuroscience Methods 162:8–13.

    https://doi.org/10.1016/j.jneumeth.2006.11.017

    • PubMed
    • Google Scholar
51. 1. Peterka DS
    2. Takahashi H
    3. Yuste R

    (2011) Imaging voltage in neurons

    Neuron 69:9–21.

    https://doi.org/10.1016/j.neuron.2010.12.010

    • PubMed
    • Google Scholar
52. 1. Peters AJ
    2. Chen SX
    3. Komiyama T

    (2014) Emergence of reproducible spatiotemporal activity during motor learning

    Nature 510:263–267.

    https://doi.org/10.1038/nature13235

    • PubMed
    • Google Scholar
53. 1. Renier N
    2. Wu Z
    3. Simon DJ
    4. Yang J
    5. Ariel P
    6. Tessier-Lavigne M

    (2014) iDISCO: a simple, rapid method to immunolabel large tissue samples for volume imaging

    Cell 159:896–910.

    https://doi.org/10.1016/j.cell.2014.10.010

    • PubMed
    • Google Scholar
54. Preprint

    1. Ruebel O
    2. Tritt A
    3. Dichter B
    4. Braun T
    5. Cain N
    6. Clack N
    7. Davidson TJ
    8. Dougherty M
    9. Fillion-Robin J-C
    10. Graddis N

    (2019) NWB:n 2.0: an accessible data standard for neurophysiology

    bioRxiv.

    https://doi.org/10.1101/523035

    • Google Scholar
55. Preprint

    1. Rupprecht P
    2. Carta S
    3. Hoffmann A
    4. Echizen M
    5. Blot A
    6. Kwan AC
    7. Dan Y
    8. Hofer SB
    9. Kitamura K
    10. Helmchen F
    11. Friedrich RW

    (2021) Database and deep learning toolbox for noise-optimized generalized spike inference from calcium imaging

    bioRxiv.

    https://doi.org/10.1101/2020.08.31.272450

    • Google Scholar
56. 1. Schmidt-Supprian M
    2. Rajewsky K

    (2007) Vagaries of conditional gene targeting

    Nature Immunology 8:665–668.

    https://doi.org/10.1038/ni0707-665

    • PubMed
    • Google Scholar
57. 1. Schwarz G

    (1978) Estimating the dimension of a model

    The Annals of Statistics 6:461–464.

    https://doi.org/10.1214/aos/1176344136

    • Google Scholar
58. 1. SciPy 1.0 Contributors
    2. Virtanen P
    3. Gommers R
    4. Oliphant TE
    5. Haberland M
    6. Reddy T
    7. Cournapeau D
    8. Burovski E
    9. Peterson P
    10. Weckesser W
    11. Bright J
    12. van der Walt SJ
    13. Brett M
    14. Wilson J
    15. Millman KJ
    16. Mayorov N
    17. Nelson ARJ
    18. Jones E
    19. Kern R
    20. Larson E
    21. Carey CJ
    22. Polat İ
    23. Feng Y
    24. Moore EW
    25. VanderPlas J
    26. Laxalde D
    27. Perktold J
    28. Cimrman R
    29. Henriksen I
    30. Quintero EA
    31. Harris CR
    32. Archibald AM
    33. Ribeiro AH
    34. Pedregosa F
    35. van Mulbregt P

    (2020) SciPy 1.0: fundamental algorithms for scientific computing in Python

    Nature Methods 17:261–272.

    https://doi.org/10.1038/s41592-019-0686-2

    • PubMed
    • Google Scholar
59. 1. Shoham S
    2. O’Connor DH
    3. Segev R

    (2006) How silent is the brain: is there a “dark matter” problem in neuroscience?

    Journal of Comparative Physiology A 192:777–784.

    https://doi.org/10.1007/s00359-006-0117-6

    • Google Scholar
60. 1. Siegle JH
    2. López AC
    3. Patel YA
    4. Abramov K
    5. Ohayon S
    6. Voigts J

    (2017) Open ephys: an open-source, plugin-based platform for multichannel electrophysiology

    Journal of Neural Engineering 14:045003.

    https://doi.org/10.1088/1741-2552/aa5eea

    • PubMed
    • Google Scholar
61. 1. Siegle JH
    2. Jia X
    3. Durand S
    4. Gale S
    5. Bennett C
    6. Graddis N
    7. Heller G
    8. Ramirez TK
    9. Choi H
    10. Luviano JA
    11. Groblewski PA
    12. Ahmed R
    13. Arkhipov A
    14. Bernard A
    15. Billeh YN
    16. Brown D
    17. Buice MA
    18. Cain N
    19. Caldejon S
    20. Casal L
    21. Cho A
    22. Chvilicek M
    23. Cox TC
    24. Dai K
    25. Denman DJ
    26. de Vries SEJ
    27. Dietzman R
    28. Esposito L
    29. Farrell C
    30. Feng D
    31. Galbraith J
    32. Garrett M
    33. Gelfand EC
    34. Hancock N
    35. Harris JA
    36. Howard R
    37. Hu B
    38. Hytnen R
    39. Iyer R
    40. Jessett E
    41. Johnson K
    42. Kato I
    43. Kiggins J
    44. Lambert S
    45. Lecoq J
    46. Ledochowitsch P
    47. Lee JH
    48. Leon A
    49. Li Y
    50. Liang E
    51. Long F
    52. Mace K
    53. Melchior J
    54. Millman D
    55. Mollenkopf T
    56. Nayan C
    57. Ng L
    58. Ngo K
    59. Nguyen T
    60. Nicovich PR
    61. North K
    62. Ocker GK
    63. Ollerenshaw D
    64. Oliver M
    65. Pachitariu M
    66. Perkins J
    67. Reding M
    68. Reid D
    69. Robertson M
    70. Ronellenfitch K
    71. Seid S
    72. Slaughterbeck C
    73. Stoecklin M
    74. Sullivan D
    75. Sutton B
    76. Swapp J
    77. Thompson C
    78. Turner K
    79. Wakeman W
    80. Whitesell JD
    81. Williams D
    82. Williford A
    83. Young R
    84. Zeng H
    85. Naylor S
    86. Phillips JW
    87. Reid RC
    88. Mihalas S
    89. Olsen SR
    90. Koch C

    (2021) Survey of spiking in the mouse visual system reveals functional hierarchy

    Nature 592:86–92.

    https://doi.org/10.1038/s41586-020-03171-x

    • PubMed
    • Google Scholar
62. Software

    1. Siegle J

    (2021) ecephys_spike_sorting, version swh:1:rev:995842e4ec67e9db1b7869d885b97317012337db

    Software Heritage.

    https://archive.softwareheritage.org/swh:1:dir:0a622ac41d878c751d74ba9deccc37691b7380a0;origin=https://github.com/AllenInstitute/ecephys_spike_sorting;visit=swh:1:snp:dd5be4e7aeea9666292804748f8ed8b6c3537c8d;anchor=swh:1:rev:995842e4ec67e9db1b7869d885b97317012337db
63. 1. Skocek O
    2. Nöbauer T
    3. Weilguny L
    4. Martínez Traub F
    5. Xia CN
    6. Molodtsov MI
    7. Grama A
    8. Yamagata M
    9. Aharoni D
    10. Cox DD
    11. Golshani P
    12. Vaziri A

    (2018) High-speed volumetric imaging of neuronal activity in freely moving rodents

    Nature Methods 15:429–432.

    https://doi.org/10.1038/s41592-018-0008-0

    • PubMed
    • Google Scholar
64. 1. Sofroniew NJ
    2. Flickinger D
    3. King J
    4. Svoboda K

    (2016) A large field of view two-photon mesoscope with subcellular resolution for in vivo imaging

    eLife 5:e14472.

    https://doi.org/10.7554/eLife.14472

    • PubMed
    • Google Scholar
65. 1. Steinmetz NA
    2. Buetfering C
    3. Lecoq J
    4. Lee CR
    5. Peters AJ
    6. Jacobs EAK
    7. Coen P
    8. Ollerenshaw DR
    9. Valley MT
    10. de Vries SEJ
    11. Garrett M
    12. Zhuang J
    13. Groblewski PA
    14. Manavi S
    15. Miles J
    16. White C
    17. Lee E
    18. Griffin F
    19. Larkin JD
    20. Roll K
    21. Cross S
    22. Nguyen TV
    23. Larsen R
    24. Pendergraft J
    25. Daigle T
    26. Tasic B
    27. Thompson CL
    28. Waters J
    29. Olsen S
    30. Margolis DJ
    31. Zeng H
    32. Hausser M
    33. Carandini M
    34. Harris KD

    (2017) Aberrant cortical activity in multiple GCaMP6-Expressing transgenic mouse lines

    Eneuro 4:ENEURO.0207-17.2017.

    https://doi.org/10.1523/ENEURO.0207-17.2017

    • PubMed
    • Google Scholar
66. 1. Stringer C
    2. Pachitariu M
    3. Steinmetz N
    4. Reddy CB
    5. Carandini M
    6. Harris KD

    (2019a) Spontaneous behaviors drive multidimensional, brainwide activity

    Science 364:eaav7893.

    https://doi.org/10.1126/science.aav7893

    • Google Scholar
67. 1. Stringer C
    2. Pachitariu M
    3. Steinmetz N
    4. Carandini M
    5. Harris KD

    (2019b) High-dimensional geometry of population responses in visual cortex

    Nature 571:361–365.

    https://doi.org/10.1038/s41586-019-1346-5

    • PubMed
    • Google Scholar
68. 1. Svoboda K
    2. Yasuda R

    (2006) Principles of two-photon excitation microscopy and its applications to neuroscience

    Neuron 50:823–839.

    https://doi.org/10.1016/j.neuron.2006.05.019

    • PubMed
    • Google Scholar
69. 1. Theis L
    2. Berens P
    3. Froudarakis E
    4. Reimer J
    5. Román Rosón M
    6. Baden T
    7. Euler T
    8. Tolias AS
    9. Bethge M

    (2016) Benchmarking spike rate inference in population calcium imaging

    Neuron 90:471–482.

    https://doi.org/10.1016/j.neuron.2016.04.014

    • PubMed
    • Google Scholar
70. 1. van der Walt S
    2. Colbert SC
    3. Varoquaux G

    (2011) The NumPy array: a structure for efficient numerical computation

    Computing in Science & Engineering 13:22–30.

    https://doi.org/10.1109/MCSE.2011.37

    • Google Scholar
71. 1. van Hateren JH
    2. van der Schaaf A

    (1998) Independent component filters of natural images compared with simple cells in primary visual cortex

    Proceedings of the Royal Society of London. Series B: Biological Sciences 265:359–366.

    https://doi.org/10.1098/rspb.1998.0303

    • Google Scholar
72. 1. Vinck M
    2. Batista-Brito R
    3. Knoblich U
    4. Cardin JA

    (2015) Arousal and locomotion make distinct contributions to cortical activity patterns and visual encoding

    Neuron 86:740–754.

    https://doi.org/10.1016/j.neuron.2015.03.028

    • PubMed
    • Google Scholar
73. 1. Vinje WE
    2. Gallant JL

    (2000) Sparse coding and decorrelation in primary visual cortex during natural vision

    Science 287:1273–1276.

    https://doi.org/10.1126/science.287.5456.1273

    • PubMed
    • Google Scholar
74. 1. Vogelstein JT
    2. Packer AM
    3. Machado TA
    4. Sippy T
    5. Babadi B
    6. Yuste R
    7. Paninski L

    (2010) Fast nonnegative deconvolution for spike train inference from population calcium imaging

    Journal of Neurophysiology 104:3691–3704.

    https://doi.org/10.1152/jn.01073.2009

    • PubMed
    • Google Scholar
75. 1. Wang Q
    2. Ding S-L
    3. Li Y
    4. Royall J
    5. Feng D
    6. Lesnar P
    7. Graddis N
    8. Naeemi M
    9. Facer B
    10. Ho A
    11. Dolbeare T
    12. Blanchard B
    13. Dee N
    14. Wakeman W
    15. Hirokawa KE
    16. Szafer A
    17. Sunkin SM
    18. Oh SW
    19. Bernard A
    20. Phillips JW
    21. Hawrylycz M
    22. Koch C
    23. Zeng H
    24. Harris JA
    25. Ng L

    (2020) The allen mouse brain common coordinate framework: a 3D reference atlas

    Cell 181:936–953.

    https://doi.org/10.1016/j.cell.2020.04.007

    • Google Scholar
76. 1. Wei Z
    2. Lin B-J
    3. Chen T-W
    4. Daie K
    5. Svoboda K
    6. Druckmann S

    (2019) A comparison of neuronal population dynamics measured with calcium imaging and electrophysiology

    Neuroscience 16:e1008198.

    https://doi.org/10.1371/journal.pcbi.1008198

    • Google Scholar
77. 1. Weisenburger S
    2. Tejera F
    3. Demas J
    4. Chen B
    5. Manley J
    6. Sparks FT
    7. Martínez Traub F
    8. Daigle T
    9. Zeng H
    10. Losonczy A
    11. Vaziri A

    (2019) Volumetric Ca²⁺ imaging in the mouse brain using hybrid multiplexed sculpted light microscopy

    Cell 177:1050–1066.

    https://doi.org/10.1016/j.cell.2019.03.011

    • PubMed
    • Google Scholar
78. Book

    1. Welles O

    (1958)

    Touch of Evil

    Universal - International.

    • Google Scholar
79. 1. West AE
    2. Chen WG
    3. Dalva MB
    4. Dolmetsch RE
    5. Kornhauser JM
    6. Shaywitz AJ
    7. Takasu MA
    8. Tao X
    9. Greenberg ME

    (2001) Calcium regulation of neuronal gene expression

    PNAS 98:11024–11031.

    https://doi.org/10.1073/pnas.191352298

    • Google Scholar
80. 1. Yang W
    2. Miller JE
    3. Carrillo-Reid L
    4. Pnevmatikakis E
    5. Paninski L
    6. Yuste R
    7. Peterka DS

    (2016) Simultaneous Multi-plane imaging of neural circuits

    Neuron 89:269–284.

    https://doi.org/10.1016/j.neuron.2015.12.012

    • PubMed
    • Google Scholar
81. 1. Zhang T
    2. Hernandez O
    3. Chrapkiewicz R
    4. Shai A
    5. Wagner MJ
    6. Zhang Y
    7. Wu CH
    8. Li JZ
    9. Inoue M
    10. Gong Y
    11. Ahanonu B
    12. Zeng H
    13. Bito H
    14. Schnitzer MJ

    (2019) Kilohertz two-photon brain imaging in awake mice

    Nature Methods 16:1119–1122.

    https://doi.org/10.1038/s41592-019-0597-2

    • PubMed
    • Google Scholar

Author details

1. Joshua H Siegle

   MindScope Program, Allen Institute, Seattle, United States

   Contribution

   Conceptualization, Formal analysis, Writing - original draft, Writing - review and editing

   Contributed equally with

   Peter Ledochowitsch

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-7736-4844

2. Peter Ledochowitsch

   MindScope Program, Allen Institute, Seattle, United States

   Contribution

   Conceptualization, Formal analysis, Writing - original draft, Writing - review and editing

   Contributed equally with

   Joshua H Siegle

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-0835-3444

3. Xiaoxuan Jia

   MindScope Program, Allen Institute, Seattle, United States

   Contribution

   Conceptualization, Writing - review and editing

   Competing interests

   No competing interests declared

4. Daniel J Millman

   MindScope Program, Allen Institute, Seattle, United States

   Contribution

   Conceptualization, Writing - review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-6255-6085

5. Gabriel K Ocker

   MindScope Program, Allen Institute, Seattle, United States

   Present address

   Boston University, Boston, United States

   Contribution

   Conceptualization, Writing - review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0001-9627-9576

6. Shiella Caldejon

   MindScope Program, Allen Institute, Seattle, United States

   Contribution

   Investigation

   Competing interests

   No competing interests declared

7. Linzy Casal

   MindScope Program, Allen Institute, Seattle, United States

   Contribution

   Project administration

   Competing interests

   No competing interests declared

8. Andy Cho

   MindScope Program, Allen Institute, Seattle, United States

   Contribution

   Investigation

   Competing interests

   No competing interests declared

9. Daniel J Denman

   Allen Institute for Brain Science, Allen Institute, Seattle, United States

   Present address

   University of Colorado Denver Anschutz Medical Campus, Aurora, United States

   Contribution

   Conceptualization, Writing - review and editing

   Competing interests

   No competing interests declared

10. Séverine Durand

    MindScope Program, Allen Institute, Seattle, United States

    Contribution

    Supervision, Investigation

    Competing interests

    No competing interests declared

11. Peter A Groblewski

    MindScope Program, Allen Institute, Seattle, United States

    Contribution

    Supervision

    Competing interests

    No competing interests declared

    "This ORCID iD identifies the author of this article:" 0000-0002-8415-1118

12. Gregg Heller

    MindScope Program, Allen Institute, Seattle, United States

    Contribution

    Investigation

    Competing interests

    No competing interests declared

13. India Kato

    MindScope Program, Allen Institute, Seattle, United States

    Contribution

    Investigation

    Competing interests

    No competing interests declared

14. Sara Kivikas

    MindScope Program, Allen Institute, Seattle, United States

    Contribution

    Investigation

    Competing interests

    No competing interests declared

15. Jérôme Lecoq

    MindScope Program, Allen Institute, Seattle, United States

    Contribution

    Supervision

    Competing interests

    No competing interests declared

16. Chelsea Nayan

    MindScope Program, Allen Institute, Seattle, United States

    Contribution

    Investigation

    Competing interests

    No competing interests declared

17. Kiet Ngo

    Allen Institute for Brain Science, Allen Institute, Seattle, United States

    Contribution

    Investigation

    Competing interests

    No competing interests declared

18. Philip R Nicovich

    Allen Institute for Brain Science, Allen Institute, Seattle, United States

    Present address

    Cajal Neuroscience, Seattle, United States

    Contribution

    Supervision

    Competing interests

    No competing interests declared

19. Kat North

    MindScope Program, Allen Institute, Seattle, United States

    Contribution

    Investigation

    Competing interests

    No competing interests declared

20. Tamina K Ramirez

    MindScope Program, Allen Institute, Seattle, United States

    Contribution

    Investigation

    Competing interests

    No competing interests declared

21. Jackie Swapp

    MindScope Program, Allen Institute, Seattle, United States

    Contribution

    Investigation

    Competing interests

    No competing interests declared

22. Xana Waughman

    MindScope Program, Allen Institute, Seattle, United States

    Contribution

    Investigation

    Competing interests

    No competing interests declared

23. Ali Williford

    MindScope Program, Allen Institute, Seattle, United States

    Contribution

    Supervision

    Competing interests

    No competing interests declared

24. Shawn R Olsen

    MindScope Program, Allen Institute, Seattle, United States

    Contribution

    Supervision, Writing - review and editing

    Competing interests

    No competing interests declared

    "This ORCID iD identifies the author of this article:" 0000-0002-9568-7057

25. Christof Koch

    MindScope Program, Allen Institute, Seattle, United States

    Contribution

    Conceptualization, Writing - review and editing

    Competing interests

    No competing interests declared

26. Michael A Buice

    MindScope Program, Allen Institute, Seattle, United States

    Contribution

    Conceptualization, Formal analysis, Supervision, Writing - review and editing

    Competing interests

    No competing interests declared

    "This ORCID iD identifies the author of this article:" 0000-0002-2196-1498

27. Saskia EJ de Vries

    MindScope Program, Allen Institute, Seattle, United States

    Contribution

    Conceptualization, Formal analysis, Supervision, Writing - original draft, Writing - review and editing

    For correspondence

    saskiad@alleninstitute.org

    Competing interests

    No competing interests declared

    "This ORCID iD identifies the author of this article:" 0000-0002-3704-3499

• 5,570

  views

• 544

  downloads

• 33

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.