Rajesh Poddar

Addressing how neural circuits underlie behavior is routinely done by measuring electrical activity from single neurons in experimental sessions. While such recordings yield snapshots of neural dynamics during specified tasks, they are ill-suited for tracking single-unit activity over longer timescales relevant for most developmental and learning processes, or for capturing neural dynamics across different behavioral states. Here we describe an automated platform for continuous long-term recordings of neural activity and behavior in freely moving rodents. An unsupervised algorithm identifies and tracks the activity of single units over weeks of recording, dramatically simplifying the analysis of large datasets. Months-long recordings from motor cortex and striatum made and analyzed with our system revealed remarkable stability in basic neuronal properties, such as firing rates and inter-spike interval distributions. Interneuronal correlations and the representation of different movements and behaviors were similarly stable. This establishes the feasibility of high-throughput long-term extracellular recordings in behaving animals.

Addressing the neural mechanisms underlying complex learned behaviors requires training animals in well-controlled tasks, an often time-consuming and labor-intensive process that can severely limit the feasibility of such studies. To overcome this constraint, we developed a fully computer-controlled general purpose system for high-throughput training of rodents. By standardizing and automating the implementation of predefined training protocols within the animal's home-cage our system dramatically reduces the efforts involved in animal training while also removing human errors and biases from the process. We deployed this system to train rats in a variety of sensorimotor tasks, achieving learning rates comparable to existing, but more laborious, methods. By incrementally and systematically increasing the difficulty of the task over weeks of training, rats were able to master motor tasks that, in complexity and structure, resemble ones used in primate studies of motor sequence learning. By enabling fully automated training of rodents in a home-cage setting this low-cost and modular system increases the utility of rodents for studying the neural underpinnings of a variety of complex behaviors.

Summary Addressing how neural circuits underlie behavior is routinely done by measuring electrical activity from single neurons during experimental sessions. While such recordings yield snapshots of neural dynamics during specified tasks, they are ill-suited for tracking single-unit activity over longer timescales relevant for most developmental and learning processes, or for capturing neural dynamics across different behavioral states. Here we describe an automated platform for continuous long-term recordings of neural activity and behavior in freely moving animals. An unsupervised algorithm identifies and tracks the activity of single units over weeks of recording, dramatically simplifying the analysis of large datasets. Months-long recordings from motor cortex and striatum made and analyzed with our system revealed remarkable stability in basic neuronal properties, such as firing rates and inter-spike interval distributions. Interneuronal correlations and the representation of different movements and behaviors were similarly stable. This establishes the feasibility of high-throughput long-term extracellular recordings in behaving animals. Highlights We record neural activity and behavior in rodents continuously (24/7) over months An automated spike-sorting method isolates and tracks single units over many weeks Neural dynamics and motor representations are highly stable over long timescales Neurons cluster into functional groups based on their activity in different states eTOC Blurb Dhawale et al. describe experimental infrastructure for recording neural activity and behavior continuously over months in freely moving rodents. A fully automated spike-sorting algorithm allows single units to be tracked over weeks of recording. Recordings from motor cortex and striatum revealed a remarkable long-term stability in both single unit activity and network dynamics.

Article Figures and data Abstract Introduction Results Discussion Materials and methods Data availability References Decision letter Author response Article and author information Metrics Abstract Addressing how neural circuits underlie behavior is routinely done by measuring electrical activity from single neurons in experimental sessions. While such recordings yield snapshots of neural dynamics during specified tasks, they are ill-suited for tracking single-unit activity over longer timescales relevant for most developmental and learning processes, or for capturing neural dynamics across different behavioral states. Here we describe an automated platform for continuous long-term recordings of neural activity and behavior in freely moving rodents. An unsupervised algorithm identifies and tracks the activity of single units over weeks of recording, dramatically simplifying the analysis of large datasets. Months-long recordings from motor cortex and striatum made and analyzed with our system revealed remarkable stability in basic neuronal properties, such as firing rates and inter-spike interval distributions. Interneuronal correlations and the representation of different movements and behaviors were similarly stable. This establishes the feasibility of high-throughput long-term extracellular recordings in behaving animals. https://doi.org/10.7554/eLife.27702.001 Introduction The goal of systems neuroscience is to understand how neural activity generates behavior. A common approach is to record from neuronal populations in targeted brain areas during experimental sessions while subjects perform designated tasks. Such intermittent recordings provide brief 'snapshots' of task-related neural dynamics (Georgopoulos et al., 1986; Hanks et al., 2015; Murakami et al., 2014), but fail to address how neural activity is modulated outside of task context and across a wide range of active and inactive behavioral states, (for exceptions see [Ambrose et al., 2016; Evarts, 1964; Gulati et al., 2014; Hengen et al., 2016; Hirase et al., 2001; Lin et al., 2006; Mizuseki and Buzsáki, 2013; Santhanam et al., 2007; Wilson and McNaughton, 1994]). Furthermore, intermittent recordings are ill-suited for reliably tracking the same neurons over time (Dickey et al., 2009; Emondi et al., 2004; Fraser and Schwartz, 2012; McMahon et al., 2014a; Santhanam et al., 2007; Tolias et al., 2007), making it difficult to discern how neural activity and task representations are shaped by developmental and learning processes that evolve over longer timescales (Ganguly et al., 2011; Jog et al., 1999; Lütcke et al., 2013; Marder and Goaillard, 2006; Peters et al., 2014; Singer et al., 2013). Addressing such fundamental questions would be greatly helped by recording neural activity and behavior continuously over days and weeks in freely moving animals. Such longitudinal recordings would allow the activity of single neurons to be followed over more trials, experimental conditions, and behavioral states, thus increasing the power with which inferences about neural function can be made (Lütcke et al., 2013; McMahon et al., 2014a). Recording continuously outside of task context would also reveal how task-related neural dynamics and behavior are affected by long-term performance history (Bair et al., 2001; Chaisanguanthum et al., 2014; Morcos and Harvey, 2016), changes in internal state (Arieli et al., 1996), spontaneous expression of innate behaviors (Aldridge and Berridge, 1998), and replay of task-related activity in different behavioral contexts (Carr et al., 2011; Foster and Wilson, 2006; Gulati et al., 2014; Wilson and McNaughton, 1994). While in vivo calcium imaging allows the same neuronal population to be recorded intermittently over long durations (Huber et al., 2012; Liberti et al., 2016; Peters et al., 2014; Rose et al., 2016; Ziv et al., 2013), photobleaching, phototoxicity, and cytotoxicity (Grienberger and Konnerth, 2012; Looger and Griesbeck, 2012), as well as the requirements for head-restraint in many versions of such experiments (Dombeck et al., 2007; Huber et al., 2012; Peters et al., 2014), make the method unsuitable for continuous long-term recordings. Calcium imaging also has relatively poor temporal resolution (Grienberger and Konnerth, 2012; Vogelstein et al., 2009), limiting its ability to resolve precise spike patterns (Vogelstein et al., 2010; Yaksi and Friedrich, 2006) (but see Gong et al., 2015 for an alternative high-speed voltage sensor). In contrast, extracellular recordings using electrode arrays can measure the activity of many single neurons simultaneously with sub-millisecond resolution (Buzsáki, 2004). Despite the unique benefits of continuous (24/7) long-term electrical recordings, they are not routinely performed. A major reason is the inherently laborious and difficult process of reliably and efficiently tracking the activity of single units from such longitudinal datasets (Einevoll et al., 2012), wherein firing rates of individual neurons can vary over many orders of magnitude (Hromádka et al., 2008; Mizuseki and Buzsáki, 2013) and spike waveforms change over time (Dickey et al., 2009; Emondi et al., 2004; Fraser and Schwartz, 2012; Okun et al., 2016; Santhanam et al., 2007; Tolias et al., 2007). To address this, we designed and deployed a modular and low-cost recording system that enables fully automated long-term continuous extracellular recordings from large numbers of neurons in freely behaving rodents engaged in natural behaviors and prescribed tasks. To efficiently parse the large streams of neural data, we developed an unsupervised spike-sorting algorithm that automatically processes the raw data from electrode array recordings, and clusters spiking events into putative single units, tracking their activity over long timescales. We validated our algorithm on ground-truth datasets and found that it surpassed the performance of spike-sorting methods that assume stationarity of spike waveforms. We used this integrated system to record from motor cortex and striatum continuously over several months, and address an ongoing debate (Clopath et al., 2017; Lütcke et al., 2013) about whether the brain is stable (Ganguly and Carmena, 2009; Greenberg and Wilson, 2004; McMahon et al., 2014b; Peters et al., 2014; Rose et al., 2016) or not (Carmena et al., 2005; Huber et al., 2012; Liberti et al., 2016; Mankin et al., 2012; Morcos and Harvey, 2016; Rokni et al., 2007; Ziv et al., 2013) over long timescales, an issue that has been previously addressed using intermittent calcium imaging (Huber et al., 2012; Liberti et al., 2016; Morcos and Harvey, 2016; Peters et al., 2014; Rose et al., 2016; Ziv et al., 2013) and extracellular recordings (Carmena et al., 2005; Ganguly and Carmena, 2009; Greenberg and Wilson, 2004; Mankin et al., 2012; McMahon et al., 2014b; Rokni et al., 2007). Our continuous, long-term recordings revealed a remarkable stability in basic neuronal properties of isolated single units, such as firing rates and inter-spike interval distributions. Interneuronal correlations and movement tuning across a range of behaviors were similarly stable over several weeks. Results Infrastructure for automated long-term neural recordings in behaving animals We developed experimental infrastructure for continuous long-term extracellular recordings in behaving rodents (Figure 1A, Figure 1—figure supplement 1). Our starting point was ARTS, a fully Automated Rodent Training System we previously developed (Poddar et al., 2013). In ARTS, the animal's home-cage doubles as the experimental chamber, making it a suitable platform for continuous long-term recordings. Figure 1 with 1 supplement see all Download asset Open asset Long-term continuous neural and behavioral recordings in behaving rodents pose challenges for traditional methods of spike-sorting. (A) Adapting our automated rodent training system (ARTS) for long-term electrophysiology. Rats engage in natural behaviors and prescribed motor tasks in their home-cages, while neural data is continuously acquired from implanted electrodes. The tethering cable connects the head-stage to a commutator mounted on a carriage that moves along a low-friction linear slide. The commutator-carriage is counterweighted to eliminate slack in the tethering cable. Behavior is continuously monitored and recorded using a camera and a 3-axis accelerometer. (B) Example of a recording segment showing high-resolution behavioral and neural data simultaneously acquired from a head-mounted 3-axis accelerometer (left) and a tetrode (right) implanted in the motor cortex, respectively. (Inset) A 2 ms zoom-in of the tetrode recording segment. (C) Drift in spike waveforms over time make it difficult to identify the same units across discontinuous recording sessions. Peak-to-peak spike amplitudes (top) and spike waveforms (bottom) for four distinct units on the same tetrode for hour-long excerpts, at 24 hr intervals, from a representative long-term continuous recording in the rat motor cortex 4 months after electrode implantation. Different units are indicated by distinct colors. We tracked units over days using a novel spike-sorting algorithm we developed to cluster continuously recorded neural data (see Figure 2). (D) Continuous extracellular recordings pose challenges for spike-sorting methods assuming stationarity in spike shapes. (Top) Peak-to-peak spike amplitudes of two continuously recorded units (same as in C) accumulated over 1 hr (left), 25 hr (middle) and 49 hr (right). (Bottom) Drift in spike waveforms can lead to inappropriate splitting (middle, right panels) of single-units and/or merging (right panel) of distinct units, even though these two units are separable in the hour-long 'sessions' shown in C. https://doi.org/10.7554/eLife.27702.002 To ensure that animals remain reliably and comfortably connected to the recording apparatus over months-long experiments, we designed a variation on the standard tethering system that allows experimental subjects to move around freely while preventing them from reaching for (and chewing) the signal cable (Figure 1A; see Materials and methods for details). Our solution connects the implanted recording device via a cable to a passive commutator (Sutton and Miller, 1963) attached to a carriage that rides on a low-friction linear slide. The carriage is counterweighted by a pulley, resulting in a small constant upwards force (<10 g) on the cable that keeps it taut and out of the animal's reach without unduly affecting its movements. The recording extension can easily be added to our custom home-cages, allowing animals that have been trained, prescreened, and deemed suitable for recordings to be implanted with electrode drives, and placed back into their familiar training environment (i.e. home-cage) for recordings. Extracellular signals recorded in behaving animals from 16 implanted tetrodes at 30 kHz sampling rate are amplified and digitized on a custom-designed head-stage (Figure 1B and Figure 1—figure supplement 1, Materials and methods). To characterize the behavior of animals throughout the recordings, the head-stage features a 3-axis accelerometer that measures head movements at high temporal resolution (Venkatraman et al., 2010) (Figure 1A–B). We also record continuous video of the rats' behavior with a wide-angle camera (30 fps) above the cage (Materials and methods). The large volumes of behavioral and neural data (~0.5 TB/day/rat) are streamed to custom-built high-capacity servers (Figure 1—figure supplement 1). A fast automated spike tracker (FAST) for long-term neural recordings Extracting single-unit spiking activity from raw data collected over weeks and months of continuous extracellular recordings presents a significant challenge for which there is currently no adequate solution. Parsing such large datasets must necessarily rely on automated spike-sorting methods. These face three major challenges (Rey et al., 2015a). First, they must reliably capture the activity of simultaneously recorded neurons whose firing rates can vary over many orders of magnitude (Hromádka et al., 2008; Mizuseki and Buzsáki, 2013). Second, they have to contend with spike shapes from recorded units changing significantly over time (Figure 1C) (Dickey et al., 2009; Emondi et al., 2004; Fraser and Schwartz, 2012). This can lead to sorting errors such as incorrect splitting of a single unit's spikes into multiple clusters, or incorrect merging of spikes from multiple units in the same cluster (Figure 1D). Third, automated methods must be capable of processing very large numbers of spikes in a reliable and efficient manner (in our experience, >1010 spikes per rat over a time span of 3 months for 64 channel recordings). Here we present an unsupervised spike-sorting algorithm that meets these challenges and tracks single units over months-long timescales. Our Fast Automated Spike Tracker (FAST) (Poddar et al., 2017) comprises two main steps (Figure 2). First, to compress the datasets and normalize for large variations in firing rates between units, it applies 'local clustering' to create de-noised representations of spiking events in the data (Figure 2A–C). In a second step, FAST chains together de-noised spikes belonging to the same putative single unit over time using an integer linear programming algorithm (Figure 2D) (Vazquez-Reina et al., 2011). FAST is an efficient and modular data processing pipeline that, depending on overall spike rates, runs two to three times faster on our custom built storage servers (Figure 1—figure supplement 1) than the rate at which the data (64 electrodes) is acquired. This implies that FAST could, in principle, also be used for online sorting, although we are currently running it offline. Figure 2 with 4 supplements see all Download asset Open asset Overview of fast automated spike tracker (FAST), an unsupervised algorithm for spike-sorting continuous, long-term recordings. (A–C) Step 1 of FAST. (A) Cartoon plot of spike waveform feature (such as amplitude) over time. Spikes are grouped into consecutive blocks of 1000 (indicated by gray dashed lines). (B) Superparamagnetic clustering is performed on each 1000 spike-block to compress and de-noise the raw spike dataset. Centroids (indicated by purple circles) of high-density clusters comprising more than 15 spikes are calculated. These correspond to units with higher firing rates. (C) Leftover spikes in low-density clusters (indicated by brown dots), corresponding to units with lower firing rates, are pooled into 1000 spike blocks and subject to the local clustering step. This process is repeated for a total of 4 iterations in order to account for spikes from units across a wide range of firing rates. Cluster centroids representing averaged spike waveforms from each round of local clustering are carried forward to Step 2 of FAST. (D–E) Step 2 of FAST. (D) Centroids from all rounds of local clustering in Step 1 are pooled into blocks of size 1000 (dashed grey lines) and local superparamagnetic clustering is performed on each block. (E) The resulting clusters are linked across time using a segmentation fusion algorithm to yield cluster-chains corresponding to single units. (F) Step 3 of FAST. In the final step, the output of the automated sorting (top) is visually inspected and similar chains merged across time to yield the final single-unit clusters (bottom). https://doi.org/10.7554/eLife.27702.004 To parse and compress the raw data, FAST first identifies and extracts spike events ('snippets') by bandpass filtering and thresholding each electrode channel (Materials and methods, Figure 2—figure supplement 1). Four or more rounds of clustering are then performed on blocks of 1000 consecutive spike 'snippets' by means of an automated superparamagnetic clustering routine, a step we call 'local clustering' (Blatt et al., 1996; Quiroga et al., 2004) (Materials and methods, Figure 2B and Figure 2—figure supplement 2A–D). Spikes in a block that belong to the same cluster are replaced by their centroid, a step that effectively de-noises and compresses the data by representing groups of similar spikes with a single waveform. The number of spikes per block was empirically determined to balance the trade-off between computation time and accuracy of superparamagnetic clustering (see Materials and methods). The goal of this step is not to reliably find all spike waveforms associated with a single unit, but to be reasonably certain that the waveforms being averaged over are similar enough to be from the same single unit. Due to large differences in firing rates between units, the initial blocks of 1000 spikes will be dominated by high firing rate units. Spikes from more sparsely firing cells that do not contribute at least 15 spikes to a cluster in a given block are carried forward to the next round of local clustering, where previously assigned spikes have been removed (Materials and methods, Figure 2C, Figure 2—figure supplement 2A–D). Applying this method of pooling and local clustering sequentially four times generates a de-noised dataset that accounts for large differences in the firing rates of simultaneously recorded units (Figure 2C, Materials and methods). The second step of the FAST algorithm is inspired by an automated method ('segmentation fusion') that links similar elements over cross-sections of longitudinal datasets in a globally optimal manner (Materials and methods, Figure 2D–E, Figure 2—figure supplement 2E–F). Segmentation fusion has been used to trace processes of individual neurons across stacks of two-dimensional serial electron-microscope images (Kasthuri et al., 2015; Vazquez-Reina et al., 2011). We adapted this method to link similar de-noised spikes across consecutive blocks into 'chains' containing the spikes of putative single units over longer timescales (Figure 2F). This algorithm allows us to automatically track the same units over days and weeks of recording. In a final post-processing and verification step, we use a semi-automated method (Dhawale et al., 2017a) to link 'chains' belonging to the same putative single unit together, and perform visual inspection of each unit (Figure 2F–G and Figure 2—figure supplement 3). A detailed description of the various steps involved in the automated spike-sorting can be found in Materials and methods. Below, we describe how we validated the spike-sorting performance of FAST using ground-truth datasets. Validation of FAST on ground-truth datasets To measure the spike-sorting capabilities of FAST, we used a publicly available dataset (Harris et al., 2000; Henze et al., 2009) comprising paired intracellular and extracellular (tetrode) recordings in the anesthetized rat hippocampus. The intracellular recordings can be used to determine the true spike-times for any unit also captured simultaneously on the tetrodes (Figure 3A). To benchmark the performance of FAST, we compared its error rate to that of the best ellipsoidal error rate (BEER), a measure which represents the optimal performance achievable by standard clustering methods (Harris et al., 2000) (see Materials and methods). We found that FAST performed as well as BEER on these relatively short (4–6 min long) recordings (Figure 3B and Figure 3—figure supplement 1A, n = 5 recordings). Figure 3 with 2 supplements see all Download asset Open asset Validation of spike-sorting performance of FAST using ground-truth datasets. (A) Example traces of simultaneous extracellular tetrode (black) and intracellular electrode (red) recordings from the anesthetized rat hippocampus (Harris et al., 2000; Henze et al., 2009). The intracellular trace can be used to identify ground-truth spike times of a single unit recorded on the tetrode. (B) Spike-sorting error rate of FAST on the paired extracellular-intracellular recording datasets (n = 5) from (Henze et al., 2009), in comparison to the best ellipsoidal error rate (BEER), a measure of the optimal performance of standard spike-sorting algorithms (see Materials and methods, Harris et al., 2000). The error rate was calculated by dividing the number of misclassified spikes (sum of false positives and false negatives) by the number of ground-truth spikes for each unit. (C) A representative synthetic tetrode recording dataset in which we model realistic fluctuations in spike amplitudes over 256 hr (see Materials and methods). The four plots show simulated spike amplitudes on the four channels of a tetrode. Colored dots indicate spikes from eight distinct units while gray dots represent multi-unit background activity. For visual clarity, we have plotted the amplitude of every 100th spike in the dataset. (D) Spike-sorting error rates of FAST (blue bars) applied to the synthetic tetrode datasets (n = 48 units from six tetrodes), in comparison to the BEER measure (red bars) over different durations of the simulated recordings. Error-bars represent standard error of the mean. * indicates p<0.05 after applying the Šidák correction for multiple comparisons. we to ability to and track units over time in their spike waveforms. Due to the of experimental ground-truth datasets recorded continuously over days to timescales, we synthetic tetrode recordings (Dhawale et al., et al., 2009) from eight distinct units whose spike amplitudes over 256 hr a process et al., 2016) (Figure Figure 3—figure supplement and Materials and methods). To make these synthetic recordings as realistic as we several the of the of spike and of using our long-term recordings in the rodent motor cortex and striatum (see Figure 3—figure supplement Materials and methods, and the long-term recordings from striatum and motor We then compared the sorting performance of FAST to BEER on from 1 to the 256 hr of the synthetic recordings (n = 48 units from simulated Our was for the performance of BEER on with accumulated time in the recording as the of between different units in spike waveform feature and that is also we (Figure In contrast, the performance of FAST, which was to BEER for short recording significantly it the recordings days (Figure FAST was also found to perform than BEER over recording durations longer than 64 hr for single electrode recording that are used in (Figure 3—figure supplement The large differences in the sorting error rates between the experimental (Figure and synthetic datasets (Figure can be to relatively small (Harris et al., 2000) lower = n = 5 of units in intracellular and extracellular recordings in comparison to our in vivo extracellular recordings on which the simulated datasets are = n = units, Figure Figure 4 with 1 supplement see all Download asset Open asset units isolated and tracked from continuous months-long recordings made in striatum and motor cortex of behaving (A) The of applying FAST to a representative continuous tetrode recording in the rodent motor cortex point represents spike amplitudes on channel 2 of a with different representing distinct units. For visual we have plotted the amplitude of de-noised spike centroids (see Figure not raw spike (bottom) show spike waveforms of six units on the four channels of the tetrode at 24 and hr into the recording. (B) times for all units recorded in the and by (C) of units recorded in (left) and (right) from two over a of recording times are indicated by and are by they were first in the recording. and indicate times at which the electrode array was into the brain by the Open indicate times at which the population of recorded units of simultaneously recorded units in the and as a function of time in the two recordings. (D) of simultaneously recorded units in the and as a function of time in the recordings shown in C. (E) of cluster for all units recorded in and Cluster was by the and of inter-spike 2 ms (bottom). the for each of these (F) of for all units recorded in and validated spike-sorting performance on experimental and synthetic we next describe how our algorithm data acquired from continuous long-term tetrode recordings, but we that it can be adapted to efficiently of electrode array or single electrode recordings whether continuous or Continuous long-term recordings from striatum and motor cortex To the of our experimental platform and analysis pipeline for long-term neural recordings, we implanted tetrode 64 into striatum (n = or motor cortex (n = of (Materials and methods). We recorded and behavioral data continuously (Figure with brief for more than 3 We that our recordings not of with the or recording but we deemed the experiments to have their We used our automated spike-sorting method (FAST) to cluster spike waveforms into putative single units, a total of units from motor cortex and units from striatum (Figure we track single units over several days days for motor cortex and days for with significant of units tracked continuously for more than a and in motor cortex and and even a in motor cortex and in Figure of stable recordings were by of the electrode or spontaneous to the movement of the (Figure we recorded simultaneously from 15 units in motor cortex and units in striatum standard shown in Figure The of single unit was by measures of cluster that cluster (Harris et al., et al., and of a et al., 2011; (Figure the cluster of the units over their recording we found that of motor (n = and of units (n = our et al., 2009; et al., 2005; et al., 2010) and of 2 ms of motor units (n = and of units (n = these for at least of recording. The of units isolated from the and was found to be and (Figure to track populations of recorded units over time have on units across discontinuous recording sessions on such as spike waveform (Dickey et al., 2009; Emondi et al., 2004; Fraser and Schwartz, 2012; Ganguly and Carmena, 2009; Greenberg and Wilson, 2004; and 2007; McMahon et al., 2014a; Okun et al., 2016; and Tolias et al., and activity measures firing rates and inter-spike interval (Dickey et al., 2009; Fraser and Schwartz, 2012). that spike waveforms of single-units can significant changes even a (Figure a major with discontinuous tracking methods is the in their we were to reliably track units continuously over we used the single-units as data with which to benchmark the performance of discontinuous tracking methods (Figure supplement 1). In our data we found discontinuous tracking over to be (Figure supplement