Spyglass: a framework for reproducible and shareable neuroscience research
Scientific progress depends on reliable and reproducible results. Progress can be accelerated when data are shared and re-analyzed to address new questions. Current approaches to storing and analyzing neural data typically involve bespoke formats and software that make replication, as well as the subsequent reuse of data, difficult if not impossible. To address these challenges, we created Spyglass, an open-source software framework that enables reproducible analyses and sharing of data and both intermediate and final results within and across labs. Spyglass uses the Neurodata Without Borders (NWB) standard and includes pipelines for several core analyses in neuroscience, including spectral filtering, spike sorting, pose tracking, and neural decoding. It can be easily extended to apply both existing and new, custom pipelines to datasets from multiple sources. We demonstrate these features in the context of a cross-laboratory replication by applying advanced state space decoding algorithms to publicly available data. New users can try out Spyglass on a Jupyter Hub hosted by HHMI and 2i2c: https://spyglass.hhmi.2i2c.cloud/.
"Constant cycling" of hippocampal activity encoding mutually exclusive hypotheticals
Cognitive abilities such as planning, imagination, and decision-making depend on the brain’s ability to represent hypothetical, rather than actual, experience. Despite this critical insight, the neurobiology underlying hypothetical representation remained poorly understood. In Kay et al., (2020), we report the discovery that hippocampal neural activity encoding mutually exclusive hypotheticals – here, divergent future paths in a maze – can alternate between hypotheticals every ~125 ms (8 Hz); a pattern we termed “constant cycling.” This finding establishes that neural representation of hypotheticals can have remarkable temporal organization: a combination of rapid expression (<1 s) and regularity (constant operation over successive 8 Hz cycles). We further found that cycle-to-cycle switching is most strongly expressed in specific hippocampal sub-regions, indicating that neural representation of hypotheticals has discrete structural organization within the hippocampus. Lastly, we found that constant cycling generalizes across multiple representations: constant cycling occurs not only for representations of position but also for representations of heading direction, a classical correlate of hippocampal activity that was first established more than 25 years ago. This raises the intriguing possibility that the brain uses a common mechanism to generate other types of hypotheticals, not only those of location or direction. Our findings also raise the possibility that neural processes that evaluate and choose among different possible futures in fact operate at the sub-second timescale, in contrast to the seconds-long timescale traditionally studied in decision-making. We are actively pursuing this possibility in current work. We believe that the wider scientific impact of our results will be to shift the study of complex cognition to short timescales and to shift focus to the brain’s capacity to generate representations of possible (vs. actual) experience.
Dinstinct hippocampal-nucleus accumbens networks support distinct cognitive functions
The hippocampus is critical for storing and then retrieving memories for the events of daily life. This process engages many other structures outside the hippocampus, but whether different parts of the hippocampus act together or separately to engage these structures was not clear. Specifically, the dorsal-ventral axis of the rodent hippocampus is associated with profound differences in gene expression and anatomical connectivity, but how those differences might translate into different patterns of engagement with other structures was not known. In Sosa, Joo, and Frank (2020) we addressed this question in the context of hippocampal interactions with the nucleus accumbens (NAc). Previous work using artificial stimulation or inactivation separately implicated either dorsal hippocampal (dHPC) or ventral hippocampal (vHPC) input to the NAc in linking spatial information to reward. These disparate results made it impossible to know whether one or both pathways were actually engaged during as animals navigated to rewarded locations. We recorded simultaneously from dHPC, vHPC, and NAc and used hippocampal sharp-wave ripples (SWRs) (markers of memory reactivation) to identify times of inter-regional information processing. We found that many NAc neurons are modulated at the times of SWRs, but that dHPC SWRs and vHPC SWRs occur at different times and engage distinct networks of neurons in the NAc. Further, we showed that only the NAc network associated with the dHPC encodes information relevant to reward and the traversal of spatial paths in an appetitive task. The NAc neurons associated with the dHPC also showed similar firing across different paths, suggesting generalization across experiences. Our results suggest that dHPC and vHPC and their associated downstream NAc neuronal networks support distinct cognitive functions, with the dHPC primarily engaged in spatial-reward learning. The opposition we observe between these networks during SWRs may support the encoding and recall of distinct aspects of experience at different times.
High density, long-term recordings in multiple brain areas
Brain functions engage activity across distributed networks of neurons spanning multiple brain structures. A critical barrier to understanding these brain functions, which include memory formation and memory-guided decision making, was the lack of tools capable of measuring millisecond timescale patterns of neural activity across these networks. Specifically, the available electrophysiological technologies were limited in their ability to access large numbers of neurons across multiple brain regions and to do so over the extended time scales (weeks to months) that characterize memory processes. We therefore developed a new methodology that overcomes these limitations. Our system uses electrode arrays fabricated of polyimide, a flexible and biocompatible substrate. These arrays are integrated with a modular headstage that supports up to 1024 recording channels in freely behaving rats. With this system we were able to carry out simultaneous recordings from hundreds of single units (putative single neurons) across a set of a spatially distributed brain regions (medial prefrontal cortex, orbitofrontal cortex, and nucleus accumbens). Our technology also allows very long lasting, high-quality recordings (over 160 days), continuous recordings (24 hours/day, 7 days a week), and exceptional stability, allowing us to measure spiking activity from large numbers of individual units (247 of 1150 measured) for at least a week. This technology has enabled studies in my laboratory where, for the first time, we can measure the activity of many hundreds of neurons from multiple brain structures as animals learn and perform memory-guided behaviors across days, weeks, and months (Chung et al., 2018). We are also working to distribute this technology to the community. Finally, I note that this technology is complementary to the HHMI/Welcomme/IMEC Neuropixels arrays, in that it permits much closer spacing of recording devices and produces much more stable recordings in freely behaving animals.
A hippocampal-cortical loop of information flow during sleep sharp-wave ripple events
Hippocampal replay during sharp-wave ripple events (SWRs) is thought to drive memory consolidation in hippocampal and cortical circuits. This is most often discussed as a process whereby the hippocampus replays memories to drive neocortical activity patterns that would reinforce a distributed memory representation. Here we asked whether that description is accurate. Specifically, previous work had shown that changes in neocortical activity can precede SWR events, but whether and how these changes influence the content of replay was not clear. In Rothschild et. al. (2017), we showed that during sleep there is actually a rapid cortical–hippocampal–cortical loop of information flow around the times of SWRs. We recorded neural activity in auditory cortex (AC) and hippocampus of rats as they learned a sound-guided task and during sleep. We found that patterned activation in AC precedes and predicts the subsequent content of hippocampal activity during SWRs, while hippocampal patterns during SWRs predict subsequent AC activity. Delivering sounds during sleep biased AC activity patterns, and sound-biased AC patterns predicted subsequent hippocampal activity. These findings provide a potential mechanism to explain observations from other laboratories that stimuli presented during sleep can enhance subsequent task performance. These results also suggest that activation of specific cortical representations during sleep influences the identity of the memories that are consolidated into long-term stores. We hypothesize that coordinated reactivation across sensory cortical regions immediately preceding SWRs facilitates a flow of reactivated sensory information into the hippocampus. This incoming information biases hippocampal reactivation, which then broadcasts an integrated representation back to the reactivated cortical networks, linking the patterns of activity across multiple cortical areas to consolidate a coherent memory representation.
Coordinated hippocampal-prefrontal reactivation during awake sharp-wave ripple events
If, as we have hypothesized, awake SWRs could support memory processes like retrieval or consolidation, they should engage a broad network of areas that are involved in representing different aspects of a memory. In Jadhav, Rothschild et. al. (2016) we asked whether that was the case by carrying out dual site recordings from hippocampus and medial prefrontal cortex (mPFC) in animals learning spatial tasks. We found that a surprisingly large proportion of mPFC cells showed spiking modulation during SWRs (~35%). Unlike in hippocampal area CA1, however, SWR-related activity in PFC comprised both excitation and inhibition of distinct populations. Within individual SWRs, excitation activated PFC cells with representations related to the concurrently reactivated hippocampal representation, while inhibition suppressed PFC cells with unrelated representations. These findings demonstrate that awake SWRs mark times of strong coordination between hippocampus and PFC that reflects structured reactivation of representations related to ongoing experience. More broadly, the patterns of excitation and inhibition in the mPFC during SWRs suggested that upon the initiation of an SWR, the representation in PFC related to the current state can be suppressed and replaced with a representation of recent active behavior consistent with the representation reactivated in the hippocampus. In conjunction with our observation described above that a similar pattern of suppression is seen in CA2 cells, we hypothesize that this could be important maintaining the separation between current experience and retrieved memories.
A hippocampal network for spatial location during immobility and sleep
Previous studies of hippocampal place representations have focused largely on the spatially specific activity seen when animals are moving from place to place. Normal behavior involves both movement as well as periods of immobility, leading us to ask the question “how does an animal know where it is when it stops moving?” In Kay et. al. (2016), we examined patterns of neural activity across multiple hippocampal subregions and identified a population of cells in area CA2 that preferentially encoded current location during periods of immobility. Interestingly, these cells, unlike all other previously identified principle cells in the hippocampus, are inhibited during the sharp-wave ripple events where we see memory replay. As a result, these cells come on as animals slow down and approach a stopping point. Their activity then continues, but is transiently interrupted by SWRs, indicating that the representation of current location is suppressed during memory replay events. We also saw this same sort of suppression in the prefrontal cortex (see below). We also identified a low frequency (~1-4 Hz) local field potential event that is associated with the activity of these CA2 cells, and using that event as a probe, we found a set of cells in other hippocampal subregions that also maintain a representation of current location during immobility. Further, these representations of current location resurfaced during sleep, specifically in a sleep phase known as SIA. These findings indicate that the hippocampal place code is maintained across states, and suggest that there are rapid alternations between representations of current experiences and replay related to past experiences.
Spiking patterns during sharp-wave ripple events predict decision errors
In Singer et al., (2013), we showed that, during learning, the specific patterns of spiking activity seen during SWRs predicts whether an animal’s next decision will be correct or incorrect. We examined SWR-related activity in the center arm of our W-track task, where animals need to remember where they came from to determine where they should go next. We found that during learning, more intense memory reactivation preceding a given choice predicted that the next choice would be correct, while less intense reactivation predicted an error. We also found that there were generally multiple memory replay events preceding correct trials. Individual events most often represented either a path involving the upcoming correct or incorrect choice, and in most cases these two possibilities were represented with equal frequency. These findings provided the first link between SWR reactivation and specific upcoming decisions. Further, because these reactivation events tend to represent upcoming possibilities, our findings suggest a role for the hippocampus in planning future actions. In particular, our results suggest that, during learning, the hippocampus broadcasts sequences representing future options to the rest of the brain, allowing other areas to decide on the correct course of action. These findings motivate our current work on real-time decoding, as this technology will allow us to test our hypotheses about the role of individual replay events in learning and decision-making.
Gamma synchronization of CA3-CA1 in awake replay of sharp-wave ripple events
We have shown that sharp-wave ripple (SWR) events are important for awake memory processes, and that during these events hippocampal place cells are reactivated in precise temporal sequences reflecting prior experience. Our goal here was to understand the extent of this coordination and the network mechanisms that make this replay possible. We noted that while we as experimenters can apply an externally defined clock to “decode” the mnemonic content of these events, the hippocampus does not have access to this external clock. Given that SWRs are generated within the hippocampus, something must act to synchronize activity across the hippocampus during memory replay, but the nature of that mechanism was unknown. In Carr, Karlsson, and Frank (2012), we first demonstrated that memory replay events consist of precisely timed sequences of CA3 and CA1 neural activity that are coordinated within and across hemispheres. We then showed that, during these events, there was a transient increase in slow gamma (20-50Hz) power, coherence, and phase locking across the CA3 and CA1 networks of both hemispheres. This transient coupling was also seen in spiking: CA1 neurons became more strongly locked to CA3 gamma specifically during SWRs and fired ~7 ms later than CA3 within each gamma cycle, suggesting entrainment of CA1 by CA3 output. Further higher levels of gamma synchrony during awake SWRs predicted higher quality replay of past experiences even when the number of active cells and other variables were taken into account. We also found differences between SWRs that occur in awake and more sleep-like states that may help explain our previous observations of higher quality replay during waking. Overall, our results indicate that CA3–CA1 gamma synchronization is a central component of awake replay, and suggest that transient gamma synchronization serves as a clocking mechanism to enable consistent memory reactivation across the hippocampal network.
Awake sharp-wave ripple events are important for memory-guided decision making
Animals use past experience to guide decisions, an ability that requires storing memories for the events of daily life and retrieving those memories as needed. This storage and retrieval depends on the hippocampus and associated structures in the medial temporal lobe, but the specific patterns of neural activity that support these memory functions remain poorly understood. We know that, during exploration, individual neurons fire in specific regions of space known as place fields. In contrast, during periods of slow movement, immobility, and slow-wave sleep, groups of neurons are active during sharp-wave ripple (SWR) events. This activity frequently represents a rapid timescale replay of a past experience. Awake SWRs in particular can reactivate sets of place fields encoding forward and reverse paths associated with both current and past locations. This reactivation has been hypothesized to contribute to multiple functions including learning, retrieval, consolidation, and trajectory planning, but its specific role in learning remained unclear. In Jadhav et al. (2012) we selectively disrupted awake SWRs in rats learning in our W-track task. We observed a specific learning and performance deficit that persisted throughout training. This deficit was associated with awake SWR activity, as SWR interruption left place field activity and post-experience SWR reactivation intact. These results provide a link between awake SWRs and hippocampal memory processes, and suggest that awake replay of memory-related information during SWRs supports learning and memory-guided decision making. More broadly, we hypothesize that the memory replay events seen during behavior propagate out to many other brain regions and engage circuits involved with outcome evaluation, planning and decision-making. We are currently exploring how replay events contribute to those processes.
Replay of past experiences during awake sharp-wave ripple events
One prominent theory of memory formation posits that there are two stages of memory formation. First, synaptic plasticity in the hippocampus would support memory storage during an experience. Subsequently, reactivation of hippocampal memory traces during sleep would promote synaptic plasticity in distributed hippocampal-neocortical networks, leading to more permanent storage of the memory. The presence of hippocampal replay during sleep is well established, but given that we do not sleep after every experience and that the strength of this replay decays with time since the experience, it is not clear how this scheme could allow us to remember experience throughout the day. In Karlsson and Frank (2009) we asked whether we could see replay of past experiences while the animals was awake. We found that that the hippocampus continually reactivates recently experienced memories, even when the animal is located outside of the place where the memory was formed. This reactivation was stronger during awake behavior than during sleep-like states, suggesting that waking replay could be particularly important for memory retrieval and the long-term storage of memories in distributed neocortical networks.
Reward enhances replay of recent experience
We do not remember all our experiences. Instead, the most exciting or emotional memories stick with us while the more mundane seem to fade over time. In addition, we are able to recall the emotional context of these memories, implying that we can somehow bind specific experiences to their outcomes. While these phenomena are well established, the mechanisms that link experiences to their outcomes and promote long term storage are not well understood. In Singer and Frank (2009) we investigated the relationship between memory reactivation and reward. We found that, as compared to not receiving a reward, receiving a reward after traversing a path led to an eight-fold increase in the likelihood that cells would be reactivated. Thus, receiving a reward led to strong reactivation of the hippocampal representation of paths associated with the rewarded location. By reactivating these paths after the outcome of traversing the path was known, reward-driven reactivation is ideally suited to allow the animal to learn the association between a set of actions and their consequence. Furthermore, the greater strength of reactivation following reward may help explain why salient or emotionally relevant events are well remembered.