1. Neuroscience
Download icon

Distinct neuronal populations contribute to trace conditioning and extinction learning in the hippocampal CA1

  1. Rebecca A Mount
  2. Sudiksha Sridhar
  3. Kyle R Hansen
  4. Ali I Mohammed
  5. Moona Abdulkerim
  6. Robb Kessel
  7. Bobak Nazer
  8. Howard J Gritton  Is a corresponding author
  9. Xue Han  Is a corresponding author
  1. Department of Biomedical Engineering, Boston University, United States
  2. Department of Electrical and Computer Engineering, Boston University, United States
Research Article
  • Cited 0
  • Views 1,422
  • Annotations
Cite this article as: eLife 2021;10:e56491 doi: 10.7554/eLife.56491

Abstract

Trace conditioning and extinction learning depend on the hippocampus, but it remains unclear how neural activity in the hippocampus is modulated during these two different behavioral processes. To explore this question, we performed calcium imaging from a large number of individual CA1 neurons during both trace eye-blink conditioning and subsequent extinction learning in mice. Our findings reveal that distinct populations of CA1 cells contribute to trace conditioned learning versus extinction learning, as learning emerges. Furthermore, we examined network connectivity by calculating co-activity between CA1 neuron pairs and found that CA1 network connectivity patterns also differ between conditioning and extinction, even though the overall connectivity density remains constant. Together, our results demonstrate that distinct populations of hippocampal CA1 neurons, forming different sub-networks with unique connectivity patterns, encode different aspects of learning.

Introduction

The hippocampus is critical for learning and memory in animals and humans. Early surgical lesions of the hippocampus in human patients, designed to alleviate intractable epilepsy, resulted in severe memory loss and an inability to form new declarative or episodic memories (Scoville and Milner, 1957; Scoville, 1954). Hippocampal atrophy is also associated with diseases related to memory loss and cognitive decline, including dementia and Alzheimer’s disease (Fein et al., 2000; Apostolova et al., 2010; Chow et al., 2012; Henneman et al., 2009; Camicioli et al., 2003). Many mechanistic studies have highlighted the importance of the hippocampus for spatial, contextual, and associative learning in a variety of animal models (Wirth et al., 2009; Jarrard, 1993).

Various experimental paradigms have been devised to probe hippocampal-dependent forms of learning and memory. One such well-established paradigm is trace eye-blink conditioning, which requires an intact hippocampus (McEchron et al., 1998; Moyer et al., 2015; Tseng et al., 2004). In this experimental design, subjects are presented with a conditioned stimulus (CS), such as a tone or light, which reliably predicts an unconditioned stimulus (US), such as a puff of air or electrical shock delivered to the subject’s eyelid. In trace conditioning, the CS and US are separated temporally by a quiescent memory trace interval. Over time, subjects learn to associate the CS with the US, generating a behavioral conditioned response to the CS (Gruart and Delgado-García, 2007; Li et al., 2008; McLaughlin et al., 2002; Kishimoto et al., 2001; Kronforst-Collins and Disterhoft, 1998; Takehara-Nishiuchi and McNaughton, 2008). Trace conditioning acquisition is thought to depend on signaling at both nicotinic and muscarinic acetylcholine receptors (Brown et al., 2010; Disterhoft et al., 1999; Fontán-Lozano et al., 2005; Raybuck and Gould, 2010; Woodruff-Pak, 2003; Woodruff-Pak et al., 2007; Flesher et al., 2011; Woodruff-Pak et al., 1994) and involves NMDA receptor-dependent plasticity (Sakamoto et al., 2005).

The hippocampus is also required for context-dependent extinction learning (Moyer et al., 2015). Traditionally, extinction learning is considered new learning that overrides a previously learned relationship. In the example of trace conditioning, the subject learns that the previously established CS is no longer predictive of a subsequent US. Extinction learning after trace conditioning can be tested by presenting the CS without the associated US and monitoring the strength or presence of a conditioned response. As new learning occurs, subjects suppress their learned response to the previously predictive CS. Extinction learning has also been shown to be NMDA receptor-dependent (Dillon et al., 2008) and requires the involvement of hippocampal inhibitory neurons (Lissek et al., 2017) and adult neurogenesis (Catlow et al., 2013).

While the hippocampus is known to be important in both conditioning and extinction learning, it is unclear how individual hippocampal neurons participate in these two types of learning and how neurons interact as learning emerges. Immediate early gene and synaptic tagging experiments revealed that in both the CA1 and dentate gyrus distinct populations of neurons were activated in fear conditioning and context-dependent fear extinction (Tronson et al., 2009; Lacagnina et al., 2019). While these studies suggest that distinct learning processes are encoded by different subsets of the neuron population, these experiments relied on quantification at later time points, after learning occurs. Such findings, while informative, cannot differentiate neuronal population changes that occur during learning from changes that occur as a result of plasticity in the minutes to hours after learning. To address this question, Zhang et al. recently identified a population of CA1 neurons that emerges to encode contextual fear conditioning using calcium imaging, but they did not examine whether different neuron populations were actively recruited during extinction training (Zhang et al., 2019). In order to better understand the mechanisms of evolving interactions between these two types of learning, we performed calcium imaging to measure the ongoing neuronal activity of individual CA1 neurons in mice during both trace eye-blink conditioning and subsequent extinction learning.

Calcium imaging allows us to measure several hundreds of neurons simultaneously with single-cell resolution across many trials, and in the same brain area over multiple days of learning (Mohammed et al., 2016; Hansen et al., 2018). In our experimental design, once robust conditioning was achieved, mice underwent a final conditioning session immediately followed by extinction training, enabling us to track the same neuron population during both conditions to reveal how extinction learning alters single cell and population encoding in the hippocampus. Hippocampal-dependent trace conditioning is well-suited to calcium imaging because generally both learning and the associated CA1 neuronal responses evolve gradually, unlike fear conditioning, in which learning can occur as rapidly as a single trial.

Our results indicate that different individual CA1 neurons showed CS-related responses during either trace conditioning or extinction learning, suggesting two functionally distinct sub-populations of cells within hippocampal CA1. To further understand how the CA1 network reflects learning as it is occurring, we analyzed co-activity between CA1 neuron pairs on a trial-by-trial basis. We found that distinct pairs of neurons are activated during trace conditioning versus extinction learning, highlighting differential network activity during these two learning processes.

Results

Conditioned responding increases across trace conditioning sessions in a classical eye-blink task and decreases during extinction learning

Trace conditioning experiments were performed in head-fixed mice (n = 9 mice) that were positioned under a custom wide-field microscope equipped with a scientific (sCMOS) camera, as previously described (Mohammed et al., 2016Figure 1A). Calcium activity in CA1 neurons was monitored via GCaMP6f fluorescence, which allows recording from hundreds of neurons simultaneously (Mohammed et al., 2016; Chen et al., 2013; Gritton et al., 2019). Prior to imaging, mice were injected with AAV-Synapsin-GCaMP6f and implanted with a custom window that allowed optical access to dorsal CA1 (Figure 1C). 4–6 weeks after surgery, mice were habituated and trained on a classical trace eye-blink conditioning paradigm followed by an extinction training session (Figure 1B). The conditioning paradigm consisted of a 9500 Hz, 350 ms tone as a CS, followed by a 250 ms trace interval, which was then followed by a 100 ms gentle puff of air to one eye that served as the US (Figure 1D). Eye behavior was monitored with a USB 3.0 Camera (Figure 1A, Ei). Animals were trained for 60–80 CS-US trials over 5–9 days until they reached conditioned response criterion (anticipatory eyelid movement on 65% of trials, described below). After reliable conditioned response to CS presentations was established, on the final day of imaging, animals were given a block of 20–40 CS-US trials, followed by a block of 40 CS-only trials, where the CS was not followed by the US (Figure 1B). With this behavioral design, we could perform calcium imaging of the same neurons during both tasks, allowing us to track how the activity of each neuron changes during extinction acquisition. To compare across different learning conditions, the final 20 conditioning trials were analyzed in the first imaging day (first training session) and the last imaging day (last training session). Because extinction learning emerges over time, we analyzed both the first 20 and final 20 CS-only extinction trials (extinction session trials 1–20 and trials 21–40, respectively) (Figure 1B).

Figure 1 with 1 supplement see all
Experimental design and quantification of animal behavior.

(A) Imaging and behavioral setup. The imaging setup consisted of a microscope with a sCMOS camera, standard wide-field fluorescence optics, and a 10× long working distance objective to image a head-fixed mouse. For the behavioral paradigm, a speaker was positioned near the mouse and a cannula for directing an air puff was placed in front of one eye. Eye responses were monitored using a USB 3.0 Camera. (B) Experimental timeline. Each animal was injected with AAV-Syn-GCaMP6f and allowed 1–2 weeks for virus expression before surgical window implantation above CA1. The first training session was 4–6 weeks after surgery, and animals were trained and imaged for 5–9 days. (C) Full field of view and selected extracted traces. Maximum-minus-minimum projection image for one motion corrected video to show example field of view of several hundred cells. Inset: several selected cells and their corresponding normalized fluorescence trace recordings. (D) Within trial design. Trials consisted of a 350 ms tone as the conditioned stimulus (CS), followed by a 250 ms trace interval with no sound, followed by a 100 ms puff of air to the eye as the unconditioned stimulus (US). (E) Video eye monitoring and segmentation. (Ei) Raw eye frames aligned to the CS, trace interval, and US windows shown above. (Eii) Eye frames after segmentation. (Eiii) Extracted eye trace and conditioned response (CR) threshold for one trial. (Eiv) Eye trace for all 40 trials of a first training session from one example mouse. Red indicates eye opening, while blue indicates eye closure. (Ev) Extracted eye trace averaged over all trials shown in Eiv.

Behavioral responses were quantified by first segmenting the eye videos to calculate eye area in each frame (Figure 1Ei, ii). An eye closure was identified if the eye area dropped below the threshold determined based on the distribution of eye area throughout the entire session. Eye closure between the tone onset, and the puff onset (tone-puff window), was classified as a conditioned response (Figure 1Ei–iv). Using this method, we were able to track the strength of the conditioned response to the CS, as well as the strong persistent eye closure in response to the aversive US on each trial (Figure 1Eiv). This method also allows for consistent calculation of conditioned responses within each training session (Figure 1Eiv, v) and across days (Figure 2A).

We first estimated basal spontaneous eye closure occurrence by examining the ‘non-stimulus periods’ of each trial, defined as the time periods that are more than 2 s before a tone onset or more than 8 s after a tone onset. Spontaneous eye closures occurred in about 20% of the 600 ms windows analyzed during the non-stimulus periods and remained stable across sessions (Figure 2—figure supplement 1). For all animals, conditioned eye closure occurrence during the first session was significantly higher than the spontaneous eye closure occurrence (p=0.0003, n = 9 mice, Wilcoxon rank-sum test, Figure 2—figure supplement 1A). While most animals (n = 7 mice) showed gradual learning acquisition over the first training session with an eye closure occurrence of 30–70% during the tone-puff window, two animals exhibited rapid learning and correctly responded to over 75% of first session trials. After reaching conditioned response criterion over several days of training, all animals maintained a conditioned response occurrence of 55–90% during the last training session, again significantly higher than the spontaneous eye closure occurrence (p=0.0003, n = 9 mice, Wilcoxon rank-sum test, Figure 2—figure supplement 1B).

In order to examine how CA1 neural activity changes during learning, we verified conditioned learning by comparing behavioral responses between the first and last days of training. To most accurately capture the effects of learning that occurs after many days of training, we excluded the two rapid learners that achieved over 75% conditioned responding during the first session from analysis involving comparisons between the first and last training sessions. The remaining seven animals showed significantly more conditioned responding during the last training session (68.9 ± 11.9%) compared to the first training session (45.0 ± 16.8%, p=0.025, n = 7 mice, Wilcoxon rank-sum test, Figure 2A, Bi). Next, we analyzed performance during the extinction session and found that conditioned responding trended downward throughout the extinction session (Figure 2Bii). By the end of the extinction session, most animals exhibited significantly reduced conditioned responding compared to the last training session. Across all animals, the conditioned response dropped to 47.8 ± 21.1% during trials 21–40 of the extinction session, significantly lower than the conditioned response rate during the last training session (p=0.034, n = 9 mice, Wilcoxon rank-sum test, Figure 2A, Bii).

Calcium dynamics in CA1 neurons track behavioral responses during trace conditioning

To evaluate how CA1 neural activity is modulated between the first and last sessions of conditioning training, we imaged a large number of individual CA1 neurons during trace conditioning (3241 and 2332 total neurons recorded in the first and last/extinction sessions respectively, Supplementary file 1). When extinction was introduced, we imaged the same neurons during trace conditioning and extinction learning, enabling us to investigate whether conditioning and extinction recruit unique cell populations or repurpose the same population. In order to assess the activity of individual cells, calcium fluorescence videos were first motion corrected, and then a projection image was generated across each video for cell segmentation using a semi-automated software. Fluorescence traces for each cell were extracted by averaging the fluorescence intensity across all pixels within the cell and normalized for each neuron across each imaging session. To visualize the population response, neuronal responses for each cell were averaged across all trials, and the entire population was sorted by average response intensity during the time period between the tone onset and puff onset (tone-puff window, Figure 2C).

To quantify changes in neural responses, individual calcium events were identified in each normalized fluorescence trace (Figure 2—figure supplement 2A, B). Consistent with previous observations of sparse calcium events (Gritton et al., 2019), we found that the basal calcium event rate was 1.7 ± 0.4 events/min (0.03 ± 0.006 events/s) during the non-stimulus periods (n = 297,336 events, across n = 9 animals from n = 27 sessions). We then binarized calcium traces by assigning ones to the rising phases of all calcium events and zeros to the rest of the trace. To estimate the basal spontaneous calcium event amplitude distribution for each trace, we randomly shuffled the non-stimulus periods of the binarized calcium trace 1000 times for each cell in each imaging session. In each shuffle, we randomly selected 20 non-overlapping segments from the non-stimulus periods of each cell in the imaging session, calculated the mean calcium event amplitude from these 20 windows, and plotted this mean calcium event amplitude to form the basal distribution for that cell. To determine if a cell is responsive to the tone, we then calculated the mean calcium event amplitude within the 1 s windows following the 20 tone onsets. If a cell’s mean tone-relevant calcium event amplitude was greater than the 95th percentile of the basal distribution for that cell, that cell was identified as a ‘responsive cell’ .

After multiple days of conditioning, 14.7% of neurons met the criteria as responsive to the tone during the last training session (286 responsive neurons of total 1946 recorded neurons in seven mice, Supplementary file 2), significantly greater than the responsive population during the first training session (12.2%, 288 responsive neurons of total 2367 recorded neurons from seven mice, p=0.017, Fisher’s exact test, Figure 2Ci, Di, Ei, Supplementary file 2, Figure 2—figure supplement 3A; Figure 2—figure supplement 3B). In addition to the increase in the fraction of responsive cells after conditioning training, calcium response latency was significantly shorter during the last training session (337.6 ± 202.5 ms) as compared to the first training session (436.9 ± 202.9 ms, p=0.001, one-way ANOVA post-hoc Tukey test, n = 795 neurons, Figure 2—figure supplement 4). Shorter calcium response latency indicates that with repeated training CA1 neurons shift their responsivity closer to the CS tone presentations, further confirming more robust CA1 responses after trace conditioning. Consistent with previous studies (Zhang et al., 2019), our findings demonstrate that CA1 neurons begin to encode the CS on the first day of training, and they more reliably encode the CS over several days of continued conditioning training.

Figure 2 with 5 supplements see all
Conditioned responses (CRs) and neuronal calcium responses increase during conditioning and decrease during extinction.

(A) Extracted eye traces across days. Red indicates an increase in eye area, while blue indicates a reduction in eye area. (Ai) Eye trace for 10 tone-only trials and the first 10 trials of the first training session from the same example mouse in Figure 1E. (Aii) Eye trace for the last training session for the same example mouse. (Aiii) Eye trace for the extinction session for the same example mouse. (B) Quantification of CR. (i) Percentage of trials with a CR during the first session (orange) and last training session (red), *p=0.025, n = 7 mice for first vs. last session. (ii) Percentage of trials with a CR during the last session (red), trials 1–20 of the extinction session (blue), and trials 21–40 of the extinction session (blue). p=0.37, n = 9 mice for last session vs. trials 1–20 of extinction session, *p=0.034, n = 9 mice for last session vs. trials 21–40 of extinction session, Wilcoxon rank-sum test. (C) Trial-averaged calcium recordings for the first session and late training session. (Ci) Trial-averaged recordings sorted by average fluorescence between the tone and the puff for the first training session from an example mouse. (Cii) Trial-averaged recordings (plotted as in Ci) for the first trial block of the late training session for the same mouse, sorted by average fluorescence between the tone and the puff for the first trial block of the late training session. (Ciii) Trial-averaged recordings (plotted as in Ci) for the second trial block of the late training session, with cell sorting maintained from the first trial block of the session to identify spatially matched cells. (Civ) The same data as shown in Ciii, but resorted according to the fluorescence between the tone and the puff for the second trial block of the late training session. (D) Trial-averaged calcium recordings for the last training session and extinction session. (Di) Trial-averaged recordings (plotted as in C) for the last training session from the same mouse, sorted by average fluorescence between the tone and the puff for the last training session. (Dii) Trial-averaged recordings (plotted as in Di) of the extinction session, with cell sorting maintained from the last training session to identify spatially matched cells. (Diii) The same data as shown in Dii, but resorted according to the fluorescence between the tone and the puff for the extinction session. (E) Quantification of responsive cell properties. (Ei) Percentage of total cells that are identified as tone-responsive for the first session (orange), last training session (red), and extinction session (blue), **p=0.017, n = 7 mice for first vs. last training session, p=0.92, n = 7 mice for first vs. extinction session, ***p=0.0005, n = 9 mice for last training vs. extinction, Fisher’s exact test. (Eii) Percentage of cells that are present within both responsive cell populations of the first half and last half of the late training session (green) and both responsive cell populations of the last training session and extinction session (purple), ***p=1.88e-6, n = 9 mice, Fisher’s exact test. For all bar plots, error bars are ± s.d.

Extinction learning rapidly recruits new CA1 neurons

As described above, the last imaging day included both the last conditioning session and the extinction session, thus allowing us to track the same cells throughout both sessions. Behavioral analysis revealed that conditioned responding was significantly reduced during the extinction session as compared to response rate during the last training session (Figure 2B). To visualize the responses of individual cells to the CS during conditioning trials and extinction trials, we plotted the neuronal responses of the entire population during each session (Figure 2D). We discovered that many neurons with strong responses to the tone during conditioning do not respond to the tone during the extinction session (Figure 2Di, ii, Eii). When we re-sorted the neuronal responses based on their responses during the extinction session, we found a largely separate population of neurons that were responsive during extinction uniquely (Figure 2Diii). To visualize the temporal dynamics of response changes across the population, we plotted the average responses of individual CA1 neurons every 10 trials (Figure 2—figure supplement 5Ai, ii). Over the 20 trials of the last training session, the responsive CA1 population remained relatively constant between the two blocks of 10 trials. However, CA1 neuronal responses changed with extinction training, and a separate responsive CA1 population emerged within the first 20 extinction trials (Figure 2—figure supplement 5Aiii–vi, B). To capture these rapid population changes, we used the first 20 CS-only extinction trials as the ‘extinction session’ for all analyses of neuronal dynamics.

We found that 11.4% of neurons were responsive to the tone during the first 20 trials of the extinction session (266 responsive neurons from total 2332 recorded cells in nine mice, Supplementary file 3), significantly less than the 14.9% responsive neurons during the last training session and similar to the 12.2% responsive neurons during the first session (p=0.0005, n = 9 mice for last session vs. extinction session and p=0.92, n = 7 mice for first session vs. extinction session, Fisher’s exact test, Figure 2Ei, Figure 2—figure supplement 3, Supplementary file 3). The reduction in responsive neurons during the extinction session is likely a reflection of the network remodeling that is occurring during extinction learning, which would not have been present in the late training session as conditioned learning was well-established during that session. However, the existence of this responsive population during the extinction session demonstrates that neurons in CA1 show new encoding of the CS during extinction learning, after it is no longer paired with the US. To compare the identities of neurons during trace conditioning and extinction learning, we classified conditioned (CO) cells as those that responded to the CS during the last training session and extinction (EX) cells as those that responded to the CS during extinction. As behavioral performance and neural activities could vary throughout each training session, we estimated the chance of detecting discrete populations over the course of an imaging session by analyzing the imaging session prior to the last training session (late training session, Figure 1B). We analyzed two blocks of 20 CS-US trials (trials 1–20 and trials 21–40) of the late training session. To quantify the proportion of common responsive cells between sessions or trial blocks, we calculated the Jaccard index, defined as the number of cells responsive in both sessions divided by the total number of cells responsive in either session. The Jaccard index for the common responsive cells between the two trial blocks of the late training session was 20.5%, significantly greater than the Jaccard index observed between the last training and extinction sessions (10.7%, p=1.88e-6, n = 9 mice, Fisher’s exact test, Figure 2Eii, Supplementary file 5). Together, these results suggest that during extinction learning, in less than 20 trials, a largely unique population of neurons is recruited to encode tone presentations.

Temporally and spatially distributed populations of neurons encode either trace conditioning or extinction learning

Because calcium events are sparse, we next considered the heterogeneity of individual cell activation during conditioning and extinction learning. We analyzed the number of trials that CO and EX cells exhibited calcium event onsets in the 1 s following tone onset. While some cells showed calcium responses to the CS on a large number of conditioning or extinction trials (example CO cells: Figure 3A; example EX cells: Figure 3B), most CO and EX cells respond to a small number of trials. CO cells exhibited calcium events on 13.3 ± 2.3% of trials during the last training session, and EX cells exhibited calcium events on 7.0 ± 1.0% of trials during the extinction session. Because most neurons individually contributed to encoding only a small fraction of trials, trial encoding may thus depend on the contributions of a large population of sparsely active neurons.

Heterogeneous neuronal responses to the tone during conditioning and extinction learning.

(A) Responses across trials for three neurons that show reliable responding during the last training session, but not during the extinction session, termed CO cells. Outer columns are individual trials shown in gray, and the average of five trials shown in red. The pink box corresponds to the tone, and the orange box corresponds to the puff. Heat maps in the center show each trial for a 3 s time window surrounding the tone and puff presentations. (B) Three neurons that exhibit reliable responding during the extinction session, but not during the last training session, termed EX cells. (C, D) Spatial maps of all neurons from a representative animal during last conditioning training session (C) and extinction session (D), with CO cells in red, EX cells in blue, and all other cells in gray.

Comparing the spatial distributions of cells indicates that both CO and EX cells are not significantly clustered near one another and are heterogeneously distributed throughout the hippocampus (Figure 3C, D). We found that 3.8 ± 2.9% of cells within a 100 µm radius of CO cells were other CO cells, which was not significantly different from that expected by random shuffling of cell identity (shuffled = 2.8%, N = 1000, p=0.17, two-tailed bootstrap, α = 0.05). Similarly, 2.3 ± 2.6% of cells within a 100 µm radius of EX cells were other EX cells, which was also not significantly different from that expected by random shuffling (shuffled = 1.8%, N = 1000, p=0.24, two-tailed bootstrap, α = 0.05). These analyses reveal that individual CO cells and EX cells responded on a sparse subset of trials, and that CO and EX cells were heterogeneously distributed within CA1 with little spatial organization.

Co-occurrence analysis reveals differential connectivity between sub-populations of neurons during trace conditioning versus extinction learning

Because CO and EX cells responded on only a sparse subset of trials, but behavioral performance was robust across sessions, we hypothesized that population responses may be critical to the role of the CA1 network in learning and memory. While pairwise correlation can give reliable measures of network activation over many trials, or over longer periods of time, the limited number of imaging data points during the short (600 ms, 12 frames) tone-puff window of this study made pairwise correlation unsuitable for tracking neuronal calcium responses. Therefore, we quantified network responses based on co-activity of cell pairs on a trial-by-trial basis and summarized co-activity across all neuron pairs in a ‘co-occurrence matrix.’ For each trial, if a neuron exhibited a calcium event onset in the 1 s following tone onset, that neuron was assigned a 1 (Figure 4Ai). All other neurons (those without a calcium event onset following tone onset on that trial) were assigned a 0. Taking the outer product of this response vector yielded a co-occurrence matrix of all cell interactions in the population for a single trial (Figure 4Aii). These single-trial co-occurrence matrices were then combined over trials and clustered using spectral biclustering to visualize neurons that were co-modulated on those trials (Kluger et al., 2003Figure 4Aii). Clusters of co-active neuron pairs in the last training session appeared largely not co-active during the extinction session (Figure 4Bi, ii). However, re-clustering the extinction session matrix revealed new clusters of co-active neuron pairs on extinction trials (Figure 4Biii), consistent with our finding of a new population of cells encoding the tone during extinction learning.

Figure 4 with 2 supplements see all
Co-occurrence network analysis during last training session vs. extinction session.

(A) Schematic of method for constructing single-trial co-occurrence matrices. (Ai) A sub-population of cells for one trial that highlights how the response pattern was determined. If a cell exhibited a calcium event (denoted by vertical black line at event onset) during the 1 s window following tone onset, it was assigned a 1. (Aii) The outer product was taken of the vector of responses across the population with itself to generate a single-trial co-occurrence matrix. This is a binary matrix where if the ith and jth cells both exhibit a calcium event during the 1 s window following tone onset there is a 1, but a 0 otherwise. These individual trials can be combined as specific trials of interest and clustered with spectral biclustering to identify neurons with the highest degree of co-activity across those trials. (B) Representative co-occurrence matrices showing clusters of co-active neuron pairs in the last training session (Bi) and the extinction session, with sorting maintained from the last training session matrix (Bii), and re-clustered results based on the matrix during extinction session (Biii). (C) Connectivity maps created from co-occurrence matrices for the last training session (Ci, top), extinction session (Ci, bottom), and overlay (Cii). Edges from the last training session are shown in red, edges from the extinction session are shown in blue, and edges present during both sessions are shown in black. Insets: zoom-ins of four nodes. (Di) Quantification network edges present during the last training session (red), extinction session (blue), or both (gray). t = 0.98, p=0.36 for last vs. extinction sessions, ***t = 7.74, p=5.5e-5 for last session vs. shared and t = 5.73, p=0.0004 for extinction session vs. shared, n = 9 mice, two-tailed paired t-test. (Dii) Percentage of edges that are present in both the first half and last half of the late training session networks (green) vs. the last training session and extinction session networks (purple), ***p=2.48e-8, n = 9 mice, Fisher’s exact test. For all bar plots, error bars are ± s.d.

To quantify network connectivity, we anatomically mapped co-activity as edges between cells that were responsive during the last training session (CO cells) or the extinction session (EX cells, Figure 4Ci). We found no significant difference in the number of edges present in the last training session versus extinction session (60.0 ± 19.7% vs. 46.4 ± 19.6% of the total edges from the last training and extinction sessions combined, t = 0.98, p=0.36, n = 9 mice, two-tailed paired t-test, Figure 4Di). Additionally, the connectivity density (the observed number of edges divided by the total number of possible edges formed between neurons) and degree (two times the average number of edges per neuron) of the two maps were not different from one another (t = 1.78, p=0.11 and t = 1.82, p=0.11 for density and degree, respectively, n = 9 mice, two-tailed paired t-test, Figure 4—figure supplement 1).

However, when we overlaid the last session and extinction session maps, we noticed that edges during the last training session appeared largely distinct from edges present during the extinction session (Figure 4C, Di). To estimate whether the difference in edge identity between the last session and extinction session is above the chance of detecting discrete networks throughout a training day, we performed the same co-occurrence network analysis on two separate trial blocks (trials 1–20 and trials 21–40) from the late training session. As expected, co-activity amongst many neuron pairs during the first trial block of the late training session was maintained during the second trial block of the session (Figure 4—figure supplement 2A). To quantify changes in network edge identity, we calculated the Jaccard index as the common edges between the two sessions divided by the total edges in both sessions. We found a Jaccard index of 8.2% common edges between the network maps for the last training and extinction sessions, significantly smaller than the 10.4% common edges between the networks for the two trial blocks of the late training session (600 shared edges of 7306 total edges in last training/extinction sessions vs. 2159 shared edges of 20676 total edges in late training session, p=2.48e-08, n = 9 mice, Fisher’s exact test, Figure 4C, Dii, Figure 4—figure supplement 2B, Supplementary file 6). As individual responsive cells respond on a small number of trials, the common edge Jaccard index between co-active neuron pairs is naturally lower than the common cell Jaccard index described in Figure 2. However, the significant decrease in common edge Jaccard index for the last training and extinction sessions in comparison to that of the two trial blocks of the late training session is consistent with the observation of distinct responsive cells during extinction learning. Furthermore, the distinct activation of different cell pairs suggests unique network patterns of activity during each learning condition, while the overall involvement of the CA1 network (connectivity density) remained constant during the conditioning and extinction training sessions.

Discussion

In this study, we provide the first detailed, real-time evidence that largely distinct populations of neurons within the hippocampal CA1 region respond to a trace conditioned stimulus during either conditioned learning or extinction learning. Previous studies have looked at calcium activity associated with trace conditioning acquisition, but have not investigated whether a separate population of neurons contributes to extinction learning (Zhang et al., 2019; Mohammed et al., 2016; Modi et al., 2014). It has been previously reported that two functionally distinct neuron populations are activated by fear conditioning and extinction in the amygdala (Herry et al., 2008). Additionally, a subsequent study looking at the CA1 region of the hippocampus revealed distinct changes in gene phosphorylation states in largely non-overlapping neural populations in either fear conditioning or extinction (Tronson et al., 2009). However, in this study, cFos and pERK immuno-activities were used as markers of conditioning and extinction learning and measured hours to days after the respective training. While these results provided the first important insight into the potential for distinct population encoding, the indirect nature of the activity markers and the time course for immuno-quantification does not allow for the distinction between rapid or gradual evolution of conditioning and extinction neuron populations. A more recent synaptic tagging study found populations of cells in the dentate gyrus that are re-activated during either fear conditioning recall or extinction retrieval, presumably representing separate populations (Lacagnina et al., 2019). However, these populations were identified using the immediate early gene Arc in separate cohorts of mice, which does not allow for real-time monitoring or direct comparisons between the populations.

In order to better understand the dynamic relationship between conditioning and extinction learning in the hippocampus, and to further investigate whether distinct populations encode these learning events, we used calcium imaging to monitor the activity of individual cells throughout conditioning and extinction learning paradigms. We applied trace conditioning because (1) it lends the advantage of a singular defined stimulus to which neural activity can be easily aligned and measured between the two different paradigms, (2) it allows imaging of the same neurons during both conditioning and extinction in the same imaging session, and (3) learning during trace conditioning evolves over multiple trials, unlike fear-based paradigms where learning often occurs over very few trials. Interestingly, we found that the rate of conditioning was highly animal-dependent, evolving gradually in most animals, but rapidly in a small subset of mice that showed substantial responding to the CS within 40 conditioning trials. Although acquisition of trace eye-blink conditioning can occur over dozens of trials in a single session (Modi et al., 2014), most of our animals showed gradual behavioral improvement, reflected in the increase in conditioned response rate after multiple days of training and the increase in the proportion of neurons that responded to the tone from the first to the last training session. Overall, our results support the idea that robust conditioned learning gradually evolves over days as CA1 neurons are recruited to encode the stimulus and reflect previous electrophysiology studies in rabbits and rats where the time course of learning evolves through many CS-US pairings (Berger et al., 1983; Green and Arenos, 2007).

In our study, extinction learning evolved rapidly, and a new population of neurons that responded to the now-extinguished tone emerged within just 20 extinction trials. Previous work has implicated the prefrontal cortex and septal cholinergic inputs to be critical to the process of extinction, and these pathways may play a pivotal role in the rapidity of extinction neuron emergence (Acquas et al., 1996; Chang et al., 2010; Euston et al., 2012). More work will need to be done on this front to determine whether reducing or silencing these inputs could delay or block the emergence of extinction-selective neuron populations. In addition, it is possible that extinction learning may occur more rapidly because meaningful associations between conditioning cues and outcomes are already established. Studies probing location encoding of familiar places, where a spatial memory schema already exists, suggest that re-encoding of spatial memory occurs more rapidly when a spatial map is already established and the new information is updating that context (Tse et al., 2007; McKenzie et al., 2013). Accordingly, our observation of rapid emergence of extinction learning could be due to the presence of a contextual representation of the task stimuli prior to extinction training. Since extinction neuron activity can emerge rapidly, the mechanisms of interaction between conditioned neuron and extinction neuron activity may be an important future direction. Such work could benefit the treatment of anxiety-based disorders such as post-traumatic stress disorder, which is characterized by over-generalization of fearful stimuli to neutral contexts and impairments in the development of extinction learning (Orr et al., 2012; Maeng and Milad, 2017; Burriss et al., 2007; Wessa and Flor, 2007; Garfinkel et al., 2014; Milad et al., 2009).

Calcium imaging is a powerful tool to understand how large populations of neurons function at the network level. However, when investigating dynamic or rapid network changes, as in extinction learning, it can be difficult to decode the information present in the population using traditional analysis techniques. For example, pairwise correlation measured over our brief tone-puff window (600 ms, 12 data points) is unreliable on a trial-by-trial basis. Traditional single-trial analytic approaches usually cannot find meaningful correlations with such limited data. Thus, our development of a co-occurrence-based network approach provides a robust way to observe the trial-by-trial evolution of the population responses and a means to assess contributions of certain cells that might be otherwise overlooked, or overstated, in trial-averaged data. Additionally, co-activity allows us to consider functional connectivity maps of entire neuron populations, an intuitive way to visualize and investigate the patterns of neural activation. Overall, co-occurrence matrix analysis is a useful technique for monitoring the evolution of population responses over time from high-dimensional calcium imaging datasets.

Using co-occurrence matrices, we found that CA1 neurons’ connectivity patterns change drastically between conditioned trial performance and extinction learning. While some cells may participate in both learning conditions, unique pairs of neurons are differentially activated during the different types of learning, indicating a role of network activation in CA1. However, the network connectivity density and degree remain constant across these different behavioral paradigms, highlighting the constant overall involvement of the CA1 network during both learning tasks. Overall, our results provide important insight into how the hippocampus may represent and encode behaviorally relevant stimuli.

Materials and methods

Animal surgery and recovery

Request a detailed protocol

All animal procedures were approved by the Boston University Institutional Animal Care and Use Committee. A total of nine female C57BL/6 mice, 8–12 weeks old at the start of the experiments, were used in these studies (Taconic; Hudson, NY). To estimate sample size, power analysis was based on effect size differences found in our previous trace conditioning calcium results recorded in the hippocampus (Mohammed et al., 2016). Power analysis was performed using G*Power 3.1.9.6 (http://www.gpower.hhu.de), applying a one-tailed Wilcoxon signed-rank test utilizing α = 0.05 and a power (1-β) value of 0.80, and yielded a sample size of 9. Following arrival from the vendor, mice were group housed and allowed to habituate to the vivarium for 2+ weeks prior to surgery. Animals first underwent viral injection surgery targeting the hippocampus under stereotaxic conditions (AP: −2.0 mm, ML: +1.5 mm, DV: −1.6 mm). Mice were injected with 250 nL of AAV9-Synapsin-GCaMP6f.WPRE.SV40 virus obtained from the University of Pennsylvania Vector Core (titer ~6e12 GC/mL). All injections were made via pulled glass pipettes (diameter: 1.2 mm) pulled to a sharp point and then broken at the tip to a final inner diameter of ~20 μm. Virus was delivered via slow pressure ejection (10–15 psi, 15–20 ms pulses delivered at 0.5 Hz). The pipette was lowered over 3 min and allowed to remain in place for 3 min before infusion began. The rate of the infusion was 100 nL/min. At the conclusion of the infusion, the pipette remained in place for 10 min before slowly being withdrawn over 2–3 min. Upon complete recovery (7+ days after virus injection), mice underwent a second procedure for the implantation of a sterilized custom imaging cannula (OD: 3.17 mm, ID: 2.36 mm, height, 2 mm diameter), fitted with a circular coverslip (size 0; OD: 3 mm) adhered using a UV-curable optical adhesive (Norland Products). To access dorsal CA1, the cortical tissue overlying the hippocampus was carefully aspirated away to expose the corpus callosum. The white matter was then thinned until the underlying tissue could be visualized through the surgical microscope. The window was then placed and centered above the hippocampus. During the same surgery, a custom aluminum headplate was attached to the skull, anterior to the imaging cannula.

Animal training and trace conditioning behavioral paradigm

Request a detailed protocol

Animals were allowed at least 2 weeks to recover from window surgeries, followed by an additional 2–4 weeks of handling and habituation to being head-fixed underneath the microscope (Figure 1B). Each animal received at least three habituation sessions prior to the first imaging day. Habituation was performed in the dark with the imaging LED illuminated to the same intensity as it would be for recording sessions.

Following habituation, mice were trained on a trace eye-blink conditioning task similar to what has been described previously (Mohammed et al., 2016). Each trial consisted of a 350 ms long 9500 Hz tone (CS) at 78–84 dB followed by a 250 ms trace interval, followed by a 100 ms puff to the eye (US) at 4.2–6 psi (Figure 1Bi). The ambient noise level ranged between 55 and 60 dB. Inter-trial intervals for each presentation were pseudo-randomized within a recording session with an inter-trial interval of 35 ± 5 s. Animals were first presented with 20 tone-only trials. Animals were then presented with either 60 tone-puff trials per day for 8 days or 80 tone-puff trials per day for 4 days. The final recording session consisted of 20 or 40 tone-puff conditioning trials, followed by 40 extinction trials, where the puff was removed but the tone remained. For all analyses, we analyzed the final 20 trials of the first imaging day (first training session) and the final 20 CS-US conditioning trials on the last imaging day (last training session). For behavior analysis, we analyzed the first 20 and final 20 CS-only extinction trials on the last imaging day, and for analysis of neuronal dynamics we analyzed the first 20 CS-only trials on the last imaging day (extinction session).

Behavioral stimuli were generated using a custom MATLAB script that delivered TTL pulses for the tone and puff via an I/O interface (USB-6259; National Instruments, Austin, TX). Behavioral TTL pulses and imaging frame timing were digitized and recorded (Digidata 1440A; Axon CNS Molecular Devices, San Jose, CA, or RZ5D Bioamp Processor; Tucker Davis Technologies, Alachua, FL) to align behavioral data and imaging frames. Mouse eye behavior was captured using a Flea3 USB 3.0 camera (FL3-U3-13S2C-CS; Richmond, BC, Canada) and the Point Grey FlyCapture2 software. The mouse’s face and eye were illuminated with an infrared lamp positioned approximately 0.05–0.5 m away from the mouse.

Wide-field imaging

Request a detailed protocol

A custom wide-field microscope was used to record neuronal calcium responses during animal learning and behavior as previously described (Mohammed et al., 2016). Briefly, the animal was head-fixed below the microscope on an articulating base (SL20 Articulating Base Ball Stage; Thorlabs Inc, Morganville, NJ) via a custom-machined attachment for the headplate. The animal was gently wrapped by an elastic self-adherent wrap to reduce movement during recording. The microscope consisted of a scientific CMOS (sCMOS) camera (ORCA-Flash4.0 LT Digital CMOS camera C11440-42U; Hamamatsu, Boston, MA), standard optics for imaging GCaMP6f, and a ×10 magnification objective (Leica N Plan 10 × 0.25 PH1 or Mitutoyo Plan Apo Long WD Objective 10 × 0.28). Images yielded a field of view 1.343 mm by 1.343 mm (1024 × 1024 pixels) and were acquired at a 20 Hz sampling rate and stored offline for analysis.

Data analysis

Request a detailed protocol

All custom software is available on GitHub (github.com/HanLabBU/Distinct-populations-hippocampus) (Sridhar and Hansen, 2021a; copy archived at swh:1:rev:960e4b4d92e42697649b9b9a684ecf9c4cbb79f6Sridhar and Hansen, 2021b). 

Eye-blink segmentation and analysis

Request a detailed protocol

Each raw eye video was segmented in ImageJ (Fiji [Schindelin et al., 2012]) using the MorphoLibJ plugin (Legland et al., 2016) to generate a binary mask video. Each frame of this binary video was summed and normalized by the average eye area to generate a trace corresponding to the percentage of eye closure over time. First, image stacks were loaded as grayscale images, Gaussian filtered with a radius of 2, and thresholded to include only the eye range. Videos were converted to binary, holes were filled, and the Particle Analyzer feature was used to exclude all regions on the edges of the videos above the thresholded value. The MorphoLibJ plugin (Legland et al., 2016) was used to label connected components with a connectivity of 26. A custom Jython script (StepIntegers.py) was used to determine the connected components that existed across all image frames, which were merged into one connected component. Lastly, to capture any additional smaller connected components, which were commonly created at or around the time of eye closures, another custom Jython script (FindModalValues.py) was used to capture these remaining components, which were then merged into the final connected component. All other connected components not a part of this singular merged component were dropped from the binary mask stack, which was saved for eye-closure trace generation.

Eye traces over time were generated by summing the binary pixels corresponding to the segmented eye in each video frame and dividing by the average eye area across the whole video. A conditioned response was classified by a threshold of 2% eye area deviation below the standard eye area. The threshold was calculated by fitting a line to the central 95th percentile of the full eye trace. This threshold is equivalent to when the residuals deviated by 2% from a uniform distribution fit to the eye trace that was equal to the average eye area. Eye traces for six mice were manually inspected by an independent observer to confirm that eyelid movements (‘blinks’) identified using the 2% threshold matched blinks selected by visual inspection. Each time the eye trace showed a decrease below this threshold between the tone onset and puff onset, that trial was classified as a conditioned response trial.

As stated previously, the final 20 CS-US trials of the first imaging day were chosen for analysis. These trials were chosen to avoid potential startle response to novel stimuli (especially the puff) in the mice, and because calcium response latency was shorter during these trials as compared to the first 20 CS-US trials of the first imaging day. Latency of calcium response was determined as the average onset time of the first calcium event after tone onset across trials for each responsive cell. Average response latency in the first 20 trials of the first imaging day was 470.3 ± 214.1 ms, and by the last 20 trials of the same imaging day, average latency dropped to 436.9 ± 202.9 ms (Figure 2—figure supplement 4), suggesting a reduction in response to the eye puff after the first 20 trials of the first imaging day. Shorter response latency is consistent with the idea that the neuronal response is shifting from being puff-responsive to being tone-responsive.

The mice with the highest behavioral performance (=>75% correct response rate) during the first training session were excluded from all analyses comparing first session to last training session or first session to extinction session (n = 2 mice). Their rapid learning of the task prevents the capture of neuronal changes during learning. All mice were included for all analyses comparing last training session to extinction session and all analyses of the late training session.

Spontaneous eye closure occurrence calculation

Request a detailed protocol

Data was split into trial periods (2 s before tone onset to 8 s after tone onset, 201 imaging frames) and non-stimulus periods (all other frames). For spontaneous eye closure occurrence calculation, a sliding window of 12 consecutive points of the eye trace (600 ms, equivalent to the tone-puff window) during the non-stimulus periods was considered at a time. Each window was analyzed as described above; a decrease below the eye area threshold within the window was classified as a spontaneous eye closure.

Movement correction

Request a detailed protocol

Motion correction of videos was done using ptmc, an open-source, parallel Python version (github.com/HanLabBU/ptmc) of an image stabilization process published previously (Mohammed et al., 2016; Hansen, 2017a). Briefly, each frame was motion corrected by median filtering each image to remove noise, homomorphic filtering the image for edge detection, and comparing the frame with a reference image to determine how many pixels to shift that specific frame. This process was run in parallel by first motion correcting the first multi-page tiff stack (2047 frames) to an average projection image of the noisy, non-corrected video. This corrected video stack was used to generate a new reference image that was sent out in parallel with every frame across the whole video, including the first tiff stack used to generate the reference.

Neuronal trace extraction

Request a detailed protocol

After motion correction, regions of interest (ROIs) corresponding to cells were selected using a semi-automated custom MATLAB software called SemiSeg (github.com/HanLabBU/SemiSeg) (Hansen, 2017b). A projection image (maximum fluorescence minus minimum fluorescence) was calculated across the whole video stack for selecting ROIs. This static frame was loaded into SemiSeg, and the full boundary of the ROIs was selected by a user. This selected sub-region of the image was automatically thresholded to determine the pixels within that region that correspond to a cell. After all cells were selected from the projection image, pixels from each ROI were averaged together spatially to calculate a temporal trace for each neuron.

Co-registration of neurons

Request a detailed protocol

Due to a technical memory limitation on the recording computer, for some mice there was a temporal gap of about 20 min between the last training session and the extinction session, in which neither the mouse nor microscope were moved (n = 3 mice). These recording sessions necessitated co-registration of cells between the sessions, unlike the continuous recording sessions where the camera continuously recorded between the last training session and extinction session (n = 6 mice). For sessions where ROIs were matched to one another, spatial ROI maps were co-registered using frame-wise cross-correlation. ROIs were then matched using a greedy method that required the centroid of cells to be within 50 pixels of one another and had to have at least 50% of their pixels overlap, as published previously (Shen et al., 2018). Cells that did not meet both of those criteria were removed from the matched dataset for comparison.

Fluorescence trace normalization

Request a detailed protocol

A local background trace was calculated for each neuron by finding the centroid for each ROI and measuring a circle approximately 10 cell widths in radius (100 pixels) and subtracting the area for the ROI from that circle. The pixels in this local background were averaged together spatially to measure a temporal background trace. Corresponding background traces were subtracted from each cell’s measured trace to remove local fluctuations from scattering in wide-field imaging. After local background subtraction, each neuron’s fluorescence trace was normalized. Due to the length of recording sessions, traditional normalization techniques (such as mean baseline subtraction and standard deviation threshold selection) disproportionally affect cells that are more active. Therefore, the baseline calcium level was calculated for each cell by fitting a normal distribution to the lowest 50 percentile of the data (assuming a Gaussian noise fluctuation surrounding a true baseline calcium level) and using the mean of this distribution as the baseline calcium level (Figure 1—figure supplement 1A). This baseline was subtracted from each trace (Figure 1—figure supplement 1B, C).

The heat maps in Figure 2 and Figure 2—figure supplement 5 were generated by averaging the calcium imaging data across trials in a session for an individual mouse. Figure 2—figure supplement 4 was generated in a similar manner for all the cells from all mice. The average fluorescence of the 12 data points (600 ms, equivalent to the length of the tone-puff window) before the tone was subtracted from the averaged trace. Trial-averaged recordings were sorted by the average fluorescence during the tone-puff window for the self-sorted heat maps. To compare common responsive cells between two trial blocks, sorting was maintained between relevant heat maps in some figures.

Calcium event detection

Request a detailed protocol

Onsets of calcium events can be detected using Fourier analysis, where event onset coincides with increasing low-frequency power. First, the spectrogram from each trace was calculated (MATLAB chronux, mtspecgramc with tapers = [2 3] and window = [1 0.05]), and the power below 2 Hz was averaged. To detect any significant increases in power, the change in the power at each time point (powerdiff) was calculated, and then the outliers (three median absolute deviations away from the median power) in powerdiff (MATLAB function isoutlier) were identified (Shemesh et al., 2020).

For outliers that occurred at consecutive time points, the first outlier with positive powerdiff represented the first point of a calcium event onset. Any identified calcium event with amplitude (the signal difference between the peak and the event onset) less than seven standard deviations of the trace in the 10 s time window prior to calcium event onset was excluded. Event detection was manually inspected in about 75 traces from multiple mice to ensure validity. Baseline calcium event rate was calculated using the non-stimulus periods (defined in the section ‘Spontaneous eye closure occurrence calculation’).

Determination and analysis of responsive cells

Request a detailed protocol

Each calcium trace was binarized such that the durations of all calcium event rising phases were assigned ones and rest of the trace was assigned zeros. We calculated the basal spontaneous calcium event amplitude distribution for each cell by shuffling the non-stimulus periods for each imaging session 1000 times. Specifically, for each shuffle, 20 non-overlapping segments of 1 s (20 frames) were randomly selected from the non-stimulus periods. The number of ones in each segment was then averaged across all 20 segments to calculate the basal calcium event amplitude for that shuffle. This procedure was repeated 1000 times to generate the baseline calcium event amplitude distribution for each cell. To determine if a cell is responsive to the CS, we calculated a cell’s CS-relevant calcium event amplitude within the 1 s windows following the 20 tone onsets by averaging the number of ones in those 20 windows. If a cell’s CS-relevant calcium event amplitude was greater than the 95th percentile of the baseline distribution for that cell, that cell was identified as a ‘responsive cell.’ 1 s was chosen as the response window to allow for the slow kinetics of calcium signaling (Chen et al., 2013).

For analysis of common cells between trial blocks of the late training session, 40 consecutive trials of the imaging session were split into two blocks (trials 1–20 and trials 21–40). For the mice that recieved 80 trials per day (n=3 mice), trials 21-60 of the late training session were analyzed. For mice that recieved 60 trials per day (n=6 mice), trials 1-40 of the late training session were analyzed. These chosen trials correspond to the same trial numbers that were used from the last training/extinction imaging day for each individual mouse. Common percentage between two sets of responsive cells (S1 and S2) was calculated as a percentage of total responsive cells:

Commoncell%= |S1 S2||S1 S2|×100
=Cs(RCs1+RCs2-Cs) × 100

where Cs is shared (common) cells, RCs1 is responsive cells from set 1, and RCs2 is responsive cells from set 2. This calculation is similar to Jaccard index, which is used to assess the similarity between two finite sample sets.

Spatial cell identity bootstrapping

Request a detailed protocol

Bootstraps of shuffled cell identity distributions were calculated for comparison against the observed distribution of cell identities. A 100 µm radius (76 pixels at 1.312 µm/pixel) was drawn around each cell. The number of segmented cells that existed within that spatial distribution was calculated and a percentage of either CO or EX cells was determined from the cells within that radius. For bootstrapping, the same number of CO or EX cells that was segmented for each session was randomly selected from the total population and the same calculation within a 100 µm radius was calculated. The measured percentages were then compared to the bootstrapped values for statistical confidence.

Co-occurrence network creation

Request a detailed protocol

Individual trial co-occurrence matrices were created for each pair of cells across every trial. Trials 1–20 (first trial block) and trials 21–40 (second trial block) were used from the late training session. For each trial, cells that exhibited a calcium event in the 1 s following tone onset were assigned a 1, and all other cells were assigned a 0. This analysis results in a binary vector of 0s and 1s of length N, where N is the number of cells recorded in the population, for each trial. The outer product of this vector was taken with itself to yield an N × N co-occurrence matrix. This matrix is 1 if both the ith and jth cells exhibited a calcium event in the 1 s after tone onset, and 0 otherwise.

Once a co-occurrence matrix was generated for each trial, matrices could be combined for further analyses by summing certain trials of interest. For this analysis, co-occurrence matrices were summed across either the last training session, the extinction session, the first trial block, or the second trial block of the late training session. Once a trial combination matrix was created, spectral biclustering was performed for a 3 × 3 cluster pattern using the Python machine learning package scikit-learn (Kluger et al., 2003; Pedregosa, 2011). Spectral biclustering is an unsupervised clustering method initially developed for clustering microarray gene expression data (Kluger et al., 2003). It seeks to find checkerboard patterns within data matrices by simultaneously clustering two different features of interest (in this case, cell identities that are co-active) using spectral clustering, which determines a submatrix of the original data matrix with similar properties.

Network map generation

Request a detailed protocol

Anatomical spatial information was combined with the co-occurrence matrix to generate network maps using the Python library NetworkX. The centroid of each ROI was used as the position of the corresponding node, which represent the cells of the imaging session. Since the co-occurrence matrix is symmetric, the lower triangular matrix was used to generate the edges of the network. Co-activity between two cells (any value greater than zero between those two cells in the co-occurrence matrix) was represented as an edge between the corresponding nodes. For example, the ith cell and jth cell would be connected by an edge if Ai,j is non-zero, where A is the N × N co-occurrence matrix. To analyze the most robust ensembles, in the last training and extinction session maps, edges were plotted between all cells that were deemed ‘responsive’ in either the last training session or the extinction session (CO and EX cells). Similarly, in the late training session maps, edges were plotted between all cells that were deemed ‘responsive’ in either of the trial blocks of the late training session.

Quantification of network properties

Request a detailed protocol

The percentage of edges in each individual network is calculated as

Ei1Ei1+Ei2-Es×100

where Ei1 is the number of edges in the individual network of interest, Ei2 is the number of edges in the other network, and Es is the number of shared edges between the two networks.

The percentage of shared edges is calculated as

Ei1+Ei2-EtEt×100

where Et is the total edges in the entire network.

Python package NetworkX was used to calculate network density, which is defined as

2mn(n-1)

and degree is defined as

2mn

where m is the number of edges, and n is the number of nodes.

Data availability

Custom software is available at https://github.com/HanLabBU/Distinct-populations-hippocampus (copy archived at https://archive.softwareheritage.org/swh:1:rev:960e4b4d92e42697649b9b9a684ecf9c4cbb79f6).

References

    1. Pedregosa F
    (2011)
    Scikit-learn: machine learning in Python
    Journal of Machine Learning Research : JMLR 12:2825–2830.

Decision letter

  1. Joshua Johansen
    Reviewing Editor; RIKEN Center for Brain Science, Japan
  2. Laura L Colgin
    Senior Editor; University of Texas at Austin, United States
  3. Joshua Johansen
    Reviewer; RIKEN Center for Brain Science, Japan
  4. Kaori Takehara-Nishiuchi
    Reviewer; University of Toronto, Canada

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Acceptance summary:

The hippocampus is important for associative learning and memory underlying adaptive behavior. These memories can be altered by new experiences, but it is not clear if or how hippocampal neuronal ensembles dynamically encode these changes. This paper addresses this question using in-vivo calcium imaging to record the activity of hippocampal neurons during recall and subsequent extinction of an associative behavioral memory. Strikingly, the transition from stable memory expression to extinction engages a new population of neurons and reorganizes functional connectivity within neuronal assemblies in the hippocampus.

Decision letter after peer review:

Thank you for submitting your article "Distinct Neuronal Populations Contribute to Trace Conditioning and Extinction Learning in the Hippocampal CA1" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, including Joshua Johansen as the Reviewing Editor and Reviewer #1, and the evaluation has been overseen by Laura Colgin as the Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Kaori Takehara-Nishiuchi (Reviewer #2).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

Summary:

Hansen et al. describe experiments designed to characterize how neuronal responses in the dorsal hippocampus are modulated by trace conditioning and subsequent extinction using single cell calcium imaging in mice. The authors build on previous findings that both of these behavioral tasks depend on the integrity of the dorsal hippocampus. They report that distinct CA1 cells and ensembles are recruited during conditioning and extinction, which is consistent with expectations from previous literature (eg, Lacagnina et al., 2019, Nat Neuro; Tronson et al. J Neurosci 2009; Mehrab et al. eLife 2014; Zhang et al. PLOS 2019). However, this is the first instance where distinct learning and extinction cells have been characterized and where identification of different neuronal ensembles have been measured in concert with learning in CA1 during trace eyeblink conditioning. This represents an advance over these previous studies.

While the results are potentially interesting, critically important controls are missing, making it difficult to draw firm conclusions from the results about whether distinct populations and networks of cells are present during eyeblink training and extinction. Furthermore, there are a number of experimental and analytical issues which need to be addressed. We therefore would like to invite the authors to respond to these concerns before a decision can be made on whether this work could be published in eLife.

Essential revisions:

1. The authors claim that the CS excited two largely disjoint groups of neurons during 40 CS-US paired trials (trace conditioning) and subsequent 40 CS-only trials (extinction). However, it is possible these changes are simply dependent on time (or non-associative effects), and not learning per se. To further strengthen this argument, they must demonstrate that comparable groups of neurons respond to the CS if CS-US contingency was kept the same throughout the entire session. One possibility could be that the authors use the data collected on the day before last training/extinction and the final day and examine how CS-evoked activity differed between the first 40 CS-US pairings and the last 40 CS-US pairings. If their argument is correct, the percentage of neurons with similar CS-evoked responses (Figure 2Dii) and the co-activity patterns (Figure 4E) should be comparable between the first and last 40 CS-US trials. And, these measures in the session with the contingency change should be greater than those in the control session without the contingency change. Alternatively, the authors could break up the last training session into two epochs, and run the exact same single cell and co-occurrence/network map analysis comparing these two epochs. If learning is the driving factor behind the emergence of distinct ensembles over time, the percentage of overlapping cells and percentage of cells with shared edges should be larger when comparing epochs that both occurred during the last session vs comparing either of these to extinction. A final possibility would be to run another control group in which the final extinction session is replaced with further overtraining to examine whether the extinction induced changes are produced by extinction or some other non-associative process. In this case cell and ensemble (co-active cells) could be examined across the last session of training and the subsequent overtraining. The prediction would be that there would not be as much change in the cell/ensemble representation with overtraining compared with extinction. Something like this is required to make any claims regarding differences in cell populations during training and extinction.

2. On a similar note, the comparison between correct and incorrect trials needs controls. We would recommend that the authors run the same analysis (Figure 5) after shuffling the trial labels (i.e., correct or incorrect trials) assigned to each co-activity pattern. This analysis allows for testing whether the percentage of the shared edge (Figure 5c) between correct and incorrect trials is significantly lower than the chance level.

3. The method of cell registration across sessions is poorly described. As the quality of cross-registration is crucial for some aspects of this study (within session cross registration is less of a worry), evidence for correct registration is key. The stated method of registering cells within 50 pixels (~50um) of one another and having at least 50% of pixel overlap seems like it has the potential for many false positives, particularly with 1-photon imaging where the authors are effectively sampling cells across tens of μm in the z-dimension. How was registration validated? Also, it is unclear as to whether registration was performed across the last session and extinction session.

4. There is considerable variability among mice for the number of cells examined, yet each mouse's contribution is not weighted in the analysis to reflect this. Can the authors show that the results of their analyses are not sensitive to differing numbers of cells across animals? Otherwise, can they provide alternative analyses that are not subject to this influence (e.g., pool all neurons from all mice for a session type and run a fisher's exact test (or equivalent) to test for significance when assessing changes in the proportion of neural responses across session types (e.g., first vs last training session)). Alternatively using subsampling procedures using numbers from the mouse with the least number of cells could address this issue.

5. Similarly, there is considerable variability among mice for behavioral performance (Figure 2B). Is there a correlation between overall task performance and neural responses? All individual data points should be color coded for mouse identity so the reader can track how behavioral performance corresponds to neural data.

6. Although the absolute value of CR% varies depending on the criteria used to detect CRs, CR% of ~40% in the first session seems very high. Please include the frequency of spontaneous eyeblinks in Figure 2B for comparison.

7. For the data shown in Figure 2, a more detailed neuronal analysis of the response properties of CA1 cells is necessary. One important point is that extinction is not a static process, while, in theory, the last training session occurring before it is. One might imagine that during early extinction the network would be more similar to the last training session and that over the course of extinction the extinction defined cells would emerge. As it stands, the extinction session is lumped into one category. The authors should analyze this in more detail by determining the change in response properties over the course of extinction and comparing different portions of the extinction dataset to the last training data. Furthermore, it says that the data in Figure 2C comes from one example animal. The authors should show a heat plot for all cells from all animals to give a more comprehensive view. In addition, in Figure 2C most of the responsive cells have a high activity level prior to stimulus onset. This is confusing as it says in the methods that the cells were classified as having a larger response during the stimulus period compared with the pre-stimulus period. Furthermore, it is unclear why only cells with higher baseline/pre-stimulus activity would be the 'responsive' cells. Additionally, it would be helpful to show a population averaged peri-event time histogram of the responses in the different trial-periods shown in Figure 2C to give a better idea of the dynamics. Finally, for Figure 2D, the authors should say what the data is normalized to.

8. Is there a reason why the trial sample size is not matched across comparisons? For example, incorrect trials constitute, on average, only ~1/4 the amount of correct trials. Given the large variability in trial-to-trial neural activity, the number of edges identified via co-occurence analysis would likely increase with trial number, which could be driving the false impression of greater connectivity (albeit short of significance) during correct vs incorrect trials. Since for most animals the number of correct trials will be >2x the number of incorrect trials, the authors could also take subsets of correct trials and compare these against each other to test for consistency across correct trials.

9. There are several instances of 1-tailed tests throughout the manuscript (e.g., when comparing the proportion of tone-trace active cells during training vs extinction). A rationale for these 1-tailed tests should be provided or, ideally, 2-tailed tests should be used.

10. The authors should address why the specific value of 2% deviation from average eye size was chosen as the threshold for a conditioned response? Is this an arbitrary value? Empirically based?

11. Regarding the co-occurence network map, the authors should address whether the threshold for edge creation between two nodes just a single trial where the 2 nodes were coactive. If they increase this threshold so only more robust ensembles are counted, does the proportion of shared edges change appreciably?

12. The method of fluorescence trace normalization in which they fit a normal distribution to the lowest 50th percentile of the data, used the mean of this distribution as the baseline , subtracted this baseline from each trace and then scaled data by 5% of the maximum range of the full calcium trace seems somewhat convoluted and suggests that they cannot isolate single cells/ROIs. This should be explained and/or corrected.

13. The method of classifying a cell as responsive during the tone-puff window (described in lines 526-538) is not ideal. The classification threshold of a larger calcium response during the stimulus compared with the baseline period of 0.15 (several 5% or one 30% increase) seems somewhat arbitrary. It would be better to use a statistical comparison.

References:

Lacagnina et al. Distinct hippocampal engrams control extinction and relapse of fear memory. Nature Neuroscience 22, 2019

Mehrab et al. CA1 cell activity sequences emerge after reorganization of network correlation structure during associative learning. eLife 2014

Tronson et al. Segregated Populations of Hippocampal Principal CA1 Neurons Mediating Conditioning and Extinction of Contextual Fear. J. Neurosci. 29, 2009

Zhang et al. Dynamics of a hippocampal neuronal ensemble encoding trace fear memory revealed by in vivo Ca2+ imaging. PLOS One 2019

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

Thank you for submitting your article "Distinct Neuronal Populations Contribute to Trace Conditioning and Extinction Learning in the Hippocampal CA1" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, including Joshua Johansen as the Reviewing Editor and Reviewer #1, and the evaluation has been overseen by Laura Colgin as the Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Kaori Takehara-Nishiuchi (Reviewer #2).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

We would like to draw your attention to changes in our revision policy that we have made in response to COVID-19 (https://elifesciences.org/articles/57162). Specifically, when editors judge that a submitted work as a whole belongs in eLife but that some conclusions require a modest amount of additional new data, as they do with your paper, we are asking that the manuscript be revised to either limit claims to those supported by data in hand, or to explicitly state that the relevant conclusions require additional supporting data.

Our expectation is that the authors will eventually carry out the additional experiments and report on how they affect the relevant conclusions either in a preprint on bioRxiv or medRxiv, or if appropriate, as a Research Advance in eLife, either of which would be linked to the original paper.

Summary:

The authors have added new data and analyses which provide stronger support for the claims made in the paper. However, there are some remaining concerns which should be addressed before a final decision can be made.

Essential revisions:

1. The greatest concern with the initial submission was, that the authors show that distinct populations are engaged during late training vs extinction periods, but it is possible these changes are simply dependent on time (or non-associative effects, and not learning per se). The reviewers suggested two potential solutions to this: "One possibility could be that the authors use the data collected on the day before last training/extinction and the final day and examine how CS-evoked activity differed between the first 40 CS-US pairings and the last 40 CS-US pairings… Alternatively, the authors could break up the last training session into two epochs, and run the exact same co-occurrence/network map analysis comparing these two epochs." This additional analysis of splitting up the prior days session was run. However, the analysis of this is unclear and it could still benefit from modification to clarify and/or optimally address the concern it was in response to. Specifically, in Figure 2, the authors compare neural activity in the 'First Half' and 'Second Half' of Late Training, but it is unclear what trials these represent. Relatedly, when comparing Last Training to Extinction it is not clear what trials were used. This is confusing because they use either 60 or 80 trials on the Late Training day, 20 or 40 trials for the Last Training session and 40 Extinction trials. The ideal analysis would be to use 60 trials for late training, 20 last training trials and 40 extinction trials so that the late training (20 early, 40 late) comparison could be matched to the last day (20 last training, 40 extinction). The authors need to explain how this was done and clearly state this in the text/figure legend. If possible, they should also try to avoid gaps in trials in, for example, the Late Training or Last Training session to match trial numbers as this could present a confound.

2. Related to point 1, the authors state that 3 mice had an approx. 20 minute wait between the end of the CS-US session and the CS-only extinction session, where the animals remained headfixed. It introduces an additional 20-minute gap that is not present in the late session control data. Should these animals be excluded from analysis? Alternatively, they could analyze the data both ways and if there aren't major differences then combine them.

3. The authors have modified their criteria for task-responsive cells, but it still leaves much to be desired. The criteria of being active on 10%+ of trials was selected as a threshold because it is 3x the average background firing rate of ALL cells. However, cells with a higher baseline firing rate would be more likely to be labeled as responsive, just by virtue of their increased firing probability (and not due to any specific responsivity to the CS or trace). Why not tailor the cutoff on a cell by cell basis? Eg, run statistical analysis comparing CS/trace activity to background activity on a cell-by-cell basis? Or compare a cell's CS/trace activity to a shuffled distribution of that cell's background activity?

4. In the initial version of the manuscript there was a significant behavioral extinction detected, but in the current version there is no significant behavioral extinction (Figure 2B in previous and current versions and explained in lines 285-293). This change should be explained to the reviewers. Also, this is a bit of a problem for the interpretation of their data which relies on trial summed neural responses. They may be able to resolve this by statistically comparing the last training session CR to the later extinction session (similar to the first half to last half extinction analysis reported in lines 211-217). Furthermore, they could/should also show the trial-by-trial CRs during the entire extinction session (single trial or 2/3 trial bins). This may provide support for their contention that "…learning is dynamic during the extinction session" and give the reader more insight into the behavioral changes occurring during extinction.

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

Thank you for submitting your article "Distinct Neuronal Populations Contribute to Trace Conditioning and Extinction Learning in the Hippocampal CA1" for consideration by eLife. Your article has been reviewed by 2 peer reviewers, including Joshua Johansen as the Reviewing Editor and Reviewer #1, and the evaluation has been overseen by Laura Colgin as the Senior Editor.

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

Summary:

The authors have addressed most of the reviewer concerns, but there is one outstanding reviewer comment which we advise you to consider and incorporate into the final version of your manuscript (see below).

Essential revisions:

We thank the authors for the clarification regarding point number 1 of the essential revisions. The only remaining suggestion is that the authors incorporate the gap between the late session trial blocks used for the control Jaccard index comparison or discuss this issue in the conclusion of the paper. That is, instead of comparing trials 1-20 and 21-40, compare trials 1-20 and 61-80 (or 41-60 in animals that did not undergo 80 trials). The idea is for this control analysis to better reflect the gap between last and extinction session blocks (which appears to be either 40 trials, or presumably ~ 20 mins + 20 trials for the animals that were subject to the technical memory limitation), and thus more precisely control for any potential changes due to elapsed time or task engagement.

Reviewer #1:

The authors have addressed my previous concerns. I support publication.

Reviewer #3:

The authors have addressed the majority of previous concerns. The only remaining concern is the analysis of the late session, where the authors should incorporate the gap between the late session trial blocks used for the control Jaccard index comparison. That is, instead of comparing trials 1-20 and 21-40, compare trials 1-20 and 61-80 (or 41-60 in animals that did not undergo 80 trials). The idea is for this control analysis to better reflect the gap between last and extinction session blocks (which appears to be either 40 trials, or presumably ~ 20 mins + 20 trials for the animals that were subject to the technical memory limitation), and thus more precisely control for any potential changes due to elapsed time or task engagement.

https://doi.org/10.7554/eLife.56491.sa1

Author response

Essential revisions:

1. The authors claim that the CS excited two largely disjoint groups of neurons during 40 CS-US paired trials (trace conditioning) and subsequent 40 CS-only trials (extinction). However, it is possible these changes are simply dependent on time (or non-associative effects), and not learning per se. To further strengthen this argument, they must demonstrate that comparable groups of neurons respond to the CS if CS-US contingency was kept the same throughout the entire session. One possibility could be that the authors use the data collected on the day before last training/extinction and the final day and examine how CS-evoked activity differed between the first 40 CS-US pairings and the last 40 CS-US pairings. If their argument is correct, the percentage of neurons with similar CS-evoked responses (Figure 2Dii) and the co-activity patterns (Figure 4E) should be comparable between the first and last 40 CS-US trials. And, these measures in the session with the contingency change should be greater than those in the control session without the contingency change. Alternatively, the authors could break up the last training session into two epochs, and run the exact same single cell and co-occurrence/network map analysis comparing these two epochs. If learning is the driving factor behind the emergence of distinct ensembles over time, the percentage of overlapping cells and percentage of cells with shared edges should be larger when comparing epochs that both occurred during the last session vs comparing either of these to extinction. A final possibility would be to run another control group in which the final extinction session is replaced with further overtraining to examine whether the extinction induced changes are produced by extinction or some other non-associative process. In this case cell and ensemble (co-active cells) could be examined across the last session of training and the subsequent overtraining. The prediction would be that there would not be as much change in the cell/ensemble representation with overtraining compared with extinction. Something like this is required to make any claims regarding differences in cell populations during training and extinction.

We thank the reviewers for pointing out this important control. We agree with the recommendation of the reviewer, and as suggested, we analyzed the data collected on the day before the last training/extinction session (referred to as “late training day” here and in the manuscript) in the same way that we analyzed the last training session and extinction session. Briefly, we split the late training day into two blocks of 20 trials (“first half” and “second half”). We identified responsive cells in the same way that we identified populations for the last training and extinction sessions, and determined the percentage of cells that were common between the first half and second half responsive populations. Common cell percentage is significantly higher between the two halves of the late training session than between the last training session and extinction session, confirming that there are distinct populations that underlie conditioning and extinction. These results are summarized in Figure 2 and the Results section entitled “Extinction learning rapidly recruits new CA1 neurons.”

Additionally, we generated co-activity matrices for the two blocks of the late training session, and compared the edges shared between the networks built from these two matrices. We found that shared edges are significantly higher between the two halves of the late training session than between the last training session and extinction session, confirming that there are distinct sub-networks that underlie conditioning and extinction. These results are summarized in Figure 4, Figure 4—figure supplement 2, and the Results section entitled “Co-occurrence analysis reveals differential connectivity between sub-populations of neurons during trace conditioning versus extinction learning.”

2. On a similar note, the comparison between correct and incorrect trials needs controls. We would recommend that the authors run the same analysis (Figure 5) after shuffling the trial labels (i.e., correct or incorrect trials) assigned to each co-activity pattern. This analysis allows for testing whether the percentage of the shared edge (Figure 5c) between correct and incorrect trials is significantly lower than the chance level.

Unfortunately, calcium event rates are sparse, and the animals’ high percentages of correct responses means that all animals have 9 or fewer incorrect trials (one animal has only 3). As such, we do not have sufficient data to compare these conditions statistically. Thus, we have removed analysis between correct and incorrect trials, and Figure 5, from the revised manuscript. We now focus on differences between populations involved in the conditioning and extinction sessions.

3. The method of cell registration across sessions is poorly described. As the quality of cross-registration is crucial for some aspects of this study (within session cross registration is less of a worry), evidence for correct registration is key. The stated method of registering cells within 50 pixels (~50um) of one another and having at least 50% of pixel overlap seems like it has the potential for many false positives, particularly with 1-photon imaging where the authors are effectively sampling cells across tens of μm in the z-dimension. How was registration validated? Also, it is unclear as to whether registration was performed across the last session and extinction session.

Co-registration was necessary for three of the nine mice (mice #1-3); these mice received 40 trials as the last training session and 40 trials as the extinction session, for a total of 80 trials. To ensure a large field of view and high speed imaging, images collected from the sCMOS camera were stored in computer random access memory (RAM) in real time, and then transferred to hard drive storage after image collection. Due to computer memory limitations, imaging needed to be paused between the 40 conditioning trials and 40 extinction trials in order to transfer the recorded video from computer memory to hard drive storage. This transfer takes ~20 minutes. During this data transfer time, neither the mouse nor the microscope were touched or moved. Once the data transfer was complete, we immediately resumed imaging to begin the extinction session. During data processing, these two videos were motion corrected and segmented for neuronal ROIs separately. Co-registration was then performed automatically as described, and validated by visual inspection of visible static landmarks (blood vessel patterns, specific cells with distinctive shapes, etc.). We are confident that co-registration did not introduce significant errors because the mice and microscope did not move, most neurons remained in both videos, and results are similar to mice with continuous recordings. Nonetheless, to avoid co-registration in the other six mice, we subsequently performed imaging for a total of 60 trials per imaging session (20 last session trials and 40 extinction session trials) to avoid the computer memory limitation. This video was continuous over all 60 trials, and thus no co-registration was needed between the last training and extinction sessions (n=6 mice, mice #4-9). We have now added these details to the Methods section “Neuronal Trace Extraction.”

4. There is considerable variability among mice for the number of cells examined, yet each mouse's contribution is not weighted in the analysis to reflect this. Can the authors show that the results of their analyses are not sensitive to differing numbers of cells across animals? Otherwise, can they provide alternative analyses that are not subject to this influence (e.g., pool all neurons from all mice for a session type and run a fisher's exact test (or equivalent) to test for significance when assessing changes in the proportion of neural responses across session types (e.g., first vs last training session)). Alternatively using subsampling procedures using numbers from the mouse with the least number of cells could address this issue.

We thank the reviewer for suggesting a Fisher’s exact test for analyzing the proportion of neurons that respond across sessions, which is a more appropriate statistical test. We have included this statistical test in Figures 2Ei, 2Eii, and 4Eii. The Fisher’s exact test contingency tables for these figures are included as supplementary files 2-6.

5. Similarly, there is considerable variability among mice for behavioral performance (Figure 2B). Is there a correlation between overall task performance and neural responses? All individual data points should be color coded for mouse identity so the reader can track how behavioral performance corresponds to neural data.

We have edited all bar plots to include unique data markers for each mouse as the reviewer suggested. The marker assignments are summarized in supplementary file 1. Unfortunately, because of the small differences in performance for the majority of mice, we were not able to test correlation strength between neural responses and behavioral performance.

6. Although the absolute value of CR% varies depending on the criteria used to detect CRs, CR% of ~40% in the first session seems very high. Please include the frequency of spontaneous eyeblinks in Figure 2B for comparison.

We have added analysis of spontaneous blink occurrence to the manuscript as the reviewer suggested. A sliding 600ms window of the non-stimulus periods was analyzed for a blink in the same way that each tone-puff window was analyzed. Spontaneous blink occurrence is consistent at about 20% during the first, last, and extinction sessions. We believe that the 40% response rate during the first session reflects learning within the training session. We also noticed that 2 of the 9 mice exhibited very rapid learning, and correctly responded to more than 75% of trials in the first session. As such, they were excluded from analyses that compare the first session and the last training session. Spontaneous blink occurrence is summarized in the new Figure 2—figure supplement 1 and described in the new Methods section entitled “Spontaneous Eye Closure Occurrence Calculation.” Exclusion of the 2 rapid learners is described in the corresponding Results section (“Conditioned responding increases across trace conditioning sessions in a classical eye blink task and decreases during extinction learning”), and also detailed in the Methods section entitled “Eye-Blink Segmentation and Analysis.”

7. For the data shown in Figure 2, a more detailed neuronal analysis of the response properties of CA1 cells is necessary. One important point is that extinction is not a static process, while, in theory, the last training session occurring before it is. One might imagine that during early extinction the network would be more similar to the last training session and that over the course of extinction the extinction defined cells would emerge. As it stands, the extinction session is lumped into one category. The authors should analyze this in more detail by determining the change in response properties over the course of extinction and comparing different portions of the extinction dataset to the last training data.

We agree with the recommendation of the reviewer, and we have included new heat maps (from the same animal shown in Figure 2) that show the evolution of the population response from last training session to extinction session (new Figure 2—figure supplement 5). These heat maps were generated with 10 trials each. All heat maps in Figure 2—figure supplement 5A are sorted by the average fluorescence during the tone-puff window of the first 10 trials of the last training session. The responsive population is mostly maintained during the last training session, and many of these cells continue to respond during the first 10 trials of extinction. However, over the course of extinction learning, some cells stop responding. The emergence of the extinction population can be seen in Figure 2—figure supplement 5B, as all the heat maps are sorted by the average fluorescence during the tone-puff window of the last 10 trials (trials 31-40) of the extinction session. Based on these results, we have revised our analysis to use the last 20 trials of the extinction session for responsive cell analyses, to most accurately capture the population that has emerged to respond during extinction only.

Furthermore, it says that the data in Figure 2C comes from one example animal. The authors should show a heat plot for all cells from all animals to give a more comprehensive view.

We have included heat maps that include all cells from all animals for the first, last, and extinction sessions, in new Figure 2—figure supplement 3.

In addition, in Figure 2C most of the responsive cells have a high activity level prior to stimulus onset. This is confusing as it says in the methods that the cells were classified as having a larger response during the stimulus period compared with the pre-stimulus period. Furthermore, it is unclear why only cells with higher baseline/pre-stimulus activity would be the 'responsive' cells. Additionally, it would be helpful to show a population averaged peri-event time histogram of the responses in the different trial-periods shown in Figure 2C to give a better idea of the dynamics. Finally, for Figure 2D, the authors should say what the data is normalized to.

The original Figure 2C was sorted by average fluorescence during the tone-puff window to allow for easy visualization of the population responses. Thus, cells that have a naturally high baseline fluorescence (simply due to cell-to-cell variability) will be sorted to the bottom of those plots, even if they are not deemed a “responsive cell.” We agree that cells with high baseline fluorescence is distracting and misleading, so for visualization purposes only, all heat maps shown in the main figures and figure supplements are now “baseline subtracted.” Each cell’s average fluorescence value from the 600ms prior to tone onset is subtracted from that cell’s trace (from 1 second before to 3 seconds after tone onset). Discussion of this change is included in the Methods section entitled “Fluorescence Trace Normalization.” New Figure 2Ei (originally Figure 2D) is normalized to cell number per mouse. The data is shown as a percentage of total cells from each mouse in the relevant session.

8. Is there a reason why the trial sample size is not matched across comparisons? For example, incorrect trials constitute, on average, only ~1/4 the amount of correct trials. Given the large variability in trial-to-trial neural activity, the number of edges identified via co-occurence analysis would likely increase with trial number, which could be driving the false impression of greater connectivity (albeit short of significance) during correct vs incorrect trials. Since for most animals the number of correct trials will be >2x the number of incorrect trials, the authors could also take subsets of correct trials and compare these against each other to test for consistency across correct trials.

The figure comparing correct and incorrect trials (formerly Figure 5) has been removed (see response to comment 2 above).

9. There are several instances of 1-tailed tests throughout the manuscript (e.g., when comparing the proportion of tone-trace active cells during training vs extinction). A rationale for these 1-tailed tests should be provided or, ideally, 2-tailed tests should be used.

We thank the reviewer for pointing this out. As suggested, we have changed all relevant statistical tests to 2-tailed tests.

10. The authors should address why the specific value of 2% deviation from average eye size was chosen as the threshold for a conditioned response? Is this an arbitrary value? Empirically based?

The 2% deviation threshold was selected after plotting the eye blink traces. Blinks (conditioned responses) were both manually identified and automatically identified by threshold crossing, and manually compared. 2% was the value determined to match what would have been manually labelled as blinks by an independent observer. Discussion of the selection of this threshold has been added to the Methods section “Eye-Blink Segmentation and Analysis.”

11. Regarding the co-occurence network map, the authors should address whether the threshold for edge creation between two nodes just a single trial where the 2 nodes were coactive. If they increase this threshold so only more robust ensembles are counted, does the proportion of shared edges change appreciably?

We have changed the network maps shown in Figure 4 to include more robust ensembles by plotting edges only between cells that were identified as responsive cells during either the last training session or the extinction session. Similarly, the newly-included network maps for the late training session (Figure 4—figure supplement 2) also show more robust ensembles, as edges were only plotted between cells that were deemed responsive during the first half or second half of the late training session. All of these cells were active on at least 10% of trials. The number of common edges between the last training and extinction session maps is significantly lower than the number of common edges between the first half and second half late training session maps, as summarized in Figure 4Eii.

12. The method of fluorescence trace normalization in which they fit a normal distribution to the lowest 50th percentile of the data, used the mean of this distribution as the baseline, subtracted this baseline from each trace and then scaled data by 5% of the maximum range of the full calcium trace seems somewhat convoluted and suggests that they cannot isolate single cells/ROIs. This should be explained and/or corrected.

We apologize for any confusion our original description created. Normalization analysis is performed on each identified cell after ROI segmentation. It is unrelated to single cell/ROI segmentation, which is performed using the maximum-minus-minimum projection image of the whole video. We have expanded discussion of normalization in the Methods section entitled “Fluorescence Trace Normalization” and we have added Figure 1—figure supplement 1 to further explain this technique. Briefly, the mean of the lowest 50th percentile of the calcium trace is used as an estimate of baseline fluorescence level, which excludes data points corresponding to calcium spikes that have fluorescence levels in the upper 50th percentile. The lowest 50th percentile of the calcium trace for one neuron follows a parametric Gaussian distribution (validated by quantile-quantile plots of data vs. a normal distribution, Figure 1—figure supplement 1A), and thus allows us to estimate true baseline of the fluorescence in the absence of calcium spikes. Since raw GCaMP6 fluorescence varies among simultaneously recorded neurons, we used this baseline fluorescence level to normalize the fluorescence of each neuron to 100%, so that we can average fluorescence traces across neurons in the same session to visualize population changes on the same scale.

In our original manuscript, we scaled calcium data to 5% of the maximum range of each trace in order to more easily represent statistically significant fluorescence values in Figure 2. With this scaling factor, the scale of fluorescence ranged from 0-20. However, we agree that this scaling is confusing and unnecessary. We have thus removed the scaling step from data processing and edited all normalized fluorescence scale bars to reflect this change. Normalized fluorescence is now presented as a percentage of each neuron’s maximum fluorescence.

13. The method of classifying a cell as responsive during the tone-puff window (described in lines 526-538) is not ideal. The classification threshold of a larger calcium response during the stimulus compared with the baseline period of 0.15 (several 5% or one 30% increase) seems somewhat arbitrary. It would be better to use a statistical comparison.

We agree with the reviewer, and have revised our responsive cell classification technique. Cells are now deemed to be activated on a given trial if they have 1 calcium onset event in the 1-second window following tone onset. If a cell is activated on ≥10% of trials, it is classified as a responsive cell. The 1-second window for responsivity was chosen to account for the slow GCaMP fluorescence dynamics. 10% of trials was chosen as the relevant threshold because this level of activation is equivalent to 0.1 events/sec, which is more than three times higher than the basal spontaneous event rate (0.03 events/sec during the non-stimulus periods) for all mice. We have added two new Methods sections entitled “Event Detection” and “Determination of Responsive Cells,” and we have included new Figure 2—figure supplement 2 to visualize event detection. We also included a brief description of this classification technique in the corresponding Results section (“Calcium dynamics in CA1 neurons track behavioral responses during trace conditioning”).

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

Essential revisions:

1. The greatest concern with the initial submission was, that the authors show that distinct populations are engaged during late training vs extinction periods, but it is possible these changes are simply dependent on time (or non-associative effects, and not learning per se). The reviewers suggested two potential solutions to this: "One possibility could be that the authors use the data collected on the day before last training/extinction and the final day and examine how CS-evoked activity differed between the first 40 CS-US pairings and the last 40 CS-US pairings… Alternatively, the authors could break up the last training session into two epochs, and run the exact same co-occurrence/network map analysis comparing these two epochs." This additional analysis of splitting up the prior days session was run. However, the analysis of this is unclear and it could still benefit from modification to clarify and/or optimally address the concern it was in response to. Specifically, in Figure 2, the authors compare neural activity in the 'First Half' and 'Second Half' of Late Training, but it is unclear what trials these represent. Relatedly, when comparing Last Training to Extinction it is not clear what trials were used. This is confusing because they use either 60 or 80 trials on the Late Training day, 20 or 40 trials for the Last Training session and 40 Extinction trials. The ideal analysis would be to use 60 trials for late training, 20 last training trials and 40 extinction trials so that the late training (20 early, 40 late) comparison could be matched to the last day (20 last training, 40 extinction). The authors need to explain how this was done and clearly state this in the text/figure legend. If possible, they should also try to avoid gaps in trials in, for example, the Late Training or Last Training session to match trial numbers as this could present a confound.

We apologize for any confusion that our original descriptions created. We have now revised Figure 1B to include the late training session. To standardize analysis across all sessions and all animals, we examined the final 20 trials of each session for the first training session, last training session, and extinction session. We have added discussion of these trial blocks to the Results section related to Figure 1, and the trial blocks are also noted in the Methods sections entitled “Eye-Blink Segmentation and Analysis.” For the first session, the final 20 trials were used to avoid potential startle responses to novel stimuli in the mice (described in the Methods section “Eye-Blink Segmentation and Analysis”). For the extinction session, the final 20 trials were used to capture the most robust extinction learning, as learning emerges during the session (as described in the Results section entitled “Conditioned responding increases across trace conditioning sessions in a classical eye blink task and decreases during extinction learning”). For the last training session, we used the final 20 trials to minimize any temporal gaps. Finally, for the late training session, two consecutive blocks of 20 trials were used for analysis (trials 1-20 and trials 21-40 of the session). We have changed all wording in the paper from “first half” and “last half” of the late training session to “first trial block” and “second trial block” to more accurately reflect the breakdown into 20 trial groupings. Also, we now clearly state this trial breakdown in the Results sections entitled “Extinction learning rapidly recruits new CA1 neurons” and “Co-occurrence analysis reveals differential connectivity between sub-populations of neurons during trace conditioning versus extinction learning.” as well as the Methods sections entitled “Determination and Analysis of Responsive Cells” and “Co-Occurrence Network Creation.”

As mentioned above, we analyzed the final 20 trials of the extinction session (extinction trials 21-40), instead of the first 20 extinction trials (extinction trials 1-20), because extinction learning occurs at different rates across animals. Thus, we felt that analyzing the final 20 extinction trials provides the most consistent measure of the effects of extinction learning on neural activity across all animals (also see response to comment 4 below). This decision was based largely on analysis of the first 20 extinction trials, which occur immediately after conditioning. We found that the common cell Jaccard index for the last training session and the first 20 trials of the extinction session (13.2%) is higher than the Jaccard index for the last training session and the final 20 trials of the extinction session (11.1%). However, both of these indices are lower than the Jaccard index for the late training session (19.9%). These analyses indicate that during the first 20 trials of the extinction session, different cells begin to be recruited in response to the tone, and this trend becomes stronger as extinction continues. Thus, we feel that analyzing the final 20 extinction trials is more appropriate in describing the neuronal effect of extinction learning, while still streamlining statistical comparisons among various trial blocks as we have performed in the manuscript.

The above analysis also indicates that gaps between analyzed trial blocks do not drastically affect our results, as the effects seen between the last training session and extinction trials are the same whether the first 20 extinction trials or final 20 extinction trials are used for comparison.

2. Related to point 1, the authors state that 3 mice had an approx. 20 minute wait between the end of the CS-US session and the CS-only extinction session, where the animals remained headfixed. It introduces an additional 20-minute gap that is not present in the late session control data. Should these animals be excluded from analysis? Alternatively, they could analyze the data both ways and if there aren't major differences then combine them.

We compared the behavioral performance of the three animals with the gap to the animals without the gap, as suggested by the reviewer. These animals’ behavioral responses closely follow the overall trend that also includes animals without the temporal gap, therefore, we grouped all animals together for analysis. We have highlighted the markers for the three animals with a temporal gap in green in Author response image 1, which is taken from Figure 2B.

Author response image 1

3. The authors have modified their criteria for task-responsive cells, but it still leaves much to be desired. The criteria of being active on 10%+ of trials was selected as a threshold because it is 3x the average background firing rate of ALL cells. However, cells with a higher baseline firing rate would be more likely to be labeled as responsive, just by virtue of their increased firing probability (and not due to any specific responsivity to the CS or trace). Why not tailor the cutoff on a cell by cell basis? Eg, run statistical analysis comparing CS/trace activity to background activity on a cell-by-cell basis? Or compare a cell's CS/trace activity to a shuffled distribution of that cell's background activity?

We thank the reviewer for suggesting the inclusion of a more robust criterion for identifying responsive cells, and we have changed our method of identifying these cells as suggested by the reviewer. The results influenced by this new method are plotted in Figure 2E, as in the previous version of the manuscript. The details of the new criterion are described in both the Results section entitled “Calcium dynamics in CA1 neurons track behavioral responses during trace conditioning” and the Methods section entitled “Determination and Analysis of Responsive Cells.” Specifically, we first binarized each calcium trace by assigning 1s to the rising phases of all calcium events and 0s to the rest of the trace. We then randomly selected 20 1-second windows of the binarized calcium event traces from the non-stimulus periods of each imaging session (“non-stimulus periods” include the time periods that are more than 2 seconds before tone onset or more than 8 seconds after tone onset). From these randomly selected 20 windows, we calculated a basal spontaneous calcium event amplitude by averaging the number of 1s across the 20 windows. We repeated this procedure 1000 times for each trace to build a basal spontaneous calcium event amplitude distribution, which normalizes each cell by its own spontaneous event rate. To determine if a cell is responsive to the tone (CS), we calculated the calcium event amplitude within 1-second windows following the 20 tone onsets. If a cell’s CS-relevant calcium event amplitude was greater than the 95th percentile of the basal distribution for that cell, that cell was identified as a “responsive cell.”

4. In the initial version of the manuscript there was a significant behavioral extinction detected, but in the current version there is no significant behavioral extinction (Figure 2B in previous and current versions and explained in lines 285-293). This change should be explained to the reviewers. Also, this is a bit of a problem for the interpretation of their data which relies on trial summed neural responses. They may be able to resolve this by statistically comparing the last training session CR to the later extinction session (similar to the first half to last half extinction analysis reported in lines 211-217). Furthermore, they could/should also show the trial-by-trial CRs during the entire extinction session (single trial or 2/3 trial bins). This may provide support for their contention that "…learning is dynamic during the extinction session" and give the reader more insight into the behavioral changes occurring during extinction.

In the initial submission of the manuscript, behavioral extinction was analyzed using 40 trials. In the previous revision and current revision, for standardization purposes as detailed in response to point 1, behavioral extinction was analyzed using only the last 20 trials of the extinction session; thus, some statistical power is lost between the two versions. Additionally, in the original version of the manuscript, the statistical test used was a one-tailed t-test. At the suggestion of the reviewers, we changed this test to two-tailed in the previous revision, which resulted a borderline p value of 0.052. Upon further evaluation of these behavioral data, however, we noted that the behavioral data was not normally distributed. As such, we have now implemented the non-parametric Wilcoxon rank-sum test, instead of the two-tailed t-test, in the current revision. Using the non-parametric version of the test, we found significantly reduced conditioning responding during the extinction session as compared to the last training session (p=0.034). We note that during extinction, most animals showed a reduction in conditioned response (CR) rate, and thus some level of extinction occurred in most animals. While two animals showed a small increase in CR rate, suggesting delayed extinction learning, we chose to include all animals in this analysis since changes in network activity may precede changes seen in behavioral performance.

To show the progression of behavioral extinction through the session as requested by the reviewer, we have included a plot of the average CR rate within a sliding window of 10 trials from all 40 trials of the extinction session, averaged across all 9 animals (Author response image 2). Error shading is shown as standard error to the mean. While behavioral performance is slightly variable, CR rate is, on average, lower during the final 20 trials of the extinction session as compared to the first 20 trials. Because of the variations between animal learning acquisition and trial-by-trial responding, it is difficult to compare performance using a small number of trials from discrete windows. Thus we kept Figure 2B as an average of 20 trials, where the window of performance from the population is the most consistent.

Author response image 2

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

Essential revisions:

We thank the authors for the clarification regarding point number 1 of the essential revisions. The only remaining suggestion is that the authors incorporate the gap between the late session trial blocks used for the control Jaccard index comparison or discuss this issue in the conclusion of the paper. That is, instead of comparing trials 1-20 and 21-40, compare trials 1-20 and 61-80 (or 41-60 in animals that did not undergo 80 trials). The idea is for this control analysis to better reflect the gap between last and extinction session blocks (which appears to be either 40 trials, or presumably ~ 20 mins + 20 trials for the animals that were subject to the technical memory limitation), and thus more precisely control for any potential changes due to elapsed time or task engagement.

We agree with the reviewer that for each mouse, the trial blocks, and the gaps between those blocks, should be identical between the last/extinction sessions and late training session to “precisely control for any potential changes due to elapsed time or task engagement”. We have now addressed this by removing gaps in our trial block comparisons, and we compare the exact same trial blocks during the last training/extinction sessions and late training session. In detail, we now analyze 40 consecutive trials. On the last imaging day, we analyzed the 20 CS-US conditioning trials immediately preceding the extinction trials (“last training session”), and the first 20 CS-only extinction trials immediately after the last training session (“extinction session”). There is no gap between the last training and extinction sessions. Similarly, on the late training day, we analyzed 40 consecutive trials, using the same corresponding trial blocks for each individual mouse analyzed. We have now added these details to the Methods section and we have updated corresponding results from this new analysis. With this new analysis, we found that the differences in neuronal dynamics between the last training and extinction sessions are now more pronounced and significant. These results provide even stronger support of our claim of distinct populations of cells involved in these two learning processes. We would like to thank the reviewer for this excellent suggestion, which has further improved our manuscript.

https://doi.org/10.7554/eLife.56491.sa2

Article and author information

Author details

  1. Rebecca A Mount

    Department of Biomedical Engineering, Boston University, Boston, United States
    Contribution
    Conducted the animal experiments, helped with data analysis, wrote the manuscript
    Contributed equally with
    Sudiksha Sridhar and Kyle R Hansen
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-8962-1641
  2. Sudiksha Sridhar

    Department of Biomedical Engineering, Boston University, Boston, United States
    Contribution
    Performed data analysis, wrote the manuscript
    Contributed equally with
    Rebecca A Mount and Kyle R Hansen
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-1749-4302
  3. Kyle R Hansen

    Department of Biomedical Engineering, Boston University, Boston, United States
    Contribution
    Performed data analysis, wrote the manuscript
    Contributed equally with
    Rebecca A Mount and Sudiksha Sridhar
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-2782-7289
  4. Ali I Mohammed

    Department of Biomedical Engineering, Boston University, Boston, United States
    Contribution
    Conducted the animal experiments
    Competing interests
    No competing interests declared
  5. Moona Abdulkerim

    Department of Biomedical Engineering, Boston University, Boston, United States
    Contribution
    Helped with the animal experiment
    Competing interests
    No competing interests declared
  6. Robb Kessel

    Department of Biomedical Engineering, Boston University, Boston, United States
    Contribution
    Helped with the animal experiment
    Competing interests
    No competing interests declared
  7. Bobak Nazer

    Department of Electrical and Computer Engineering, Boston University, Boston, United States
    Contribution
    Helped with data analysis, edited the manuscript
    Competing interests
    No competing interests declared
  8. Howard J Gritton

    Department of Biomedical Engineering, Boston University, Boston, United States
    Contribution
    Conducted the animal experiment, wrote the manuscript.
    For correspondence
    hgritton@bu.edu
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-3194-3258
  9. Xue Han

    Department of Biomedical Engineering, Boston University, Boston, United States
    Contribution
    Supervised the study, wrote the manuscript
    For correspondence
    xuehan@bu.edu
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-3896-4609

Funding

National Science Foundation (CBET-1848029)

  • Xue Han
  • Bobak Nazer

National Institutes of Health (1R01MH122971-01A1)

  • Xue Han

National Institutes of Health (1F31MH123008-01A1)

  • Rebecca A Mount

National Academy of Engineering

  • Xue Han

National Science Foundation (DGE-1247312)

  • Kyle R Hansen

National Institutes of Health (F31 NS 105420)

  • Kyle R Hansen

National Institutes of Health (1R21MH109941-01)

  • Xue Han

National Science Foundation (CCF-1955981)

  • Xue Han
  • Bobak Nazer

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

We thank the members of Han lab for technical support.

Ethics

Animal experimentation: All animal procedures were approved by the Boston University Institutional Animal Care and Use Committee (protocol #201800680), and all experiments were performed in accordance with the relevant guidelines and regulations.

Senior Editor

  1. Laura L Colgin, University of Texas at Austin, United States

Reviewing Editor

  1. Joshua Johansen, RIKEN Center for Brain Science, Japan

Reviewers

  1. Joshua Johansen, RIKEN Center for Brain Science, Japan
  2. Kaori Takehara-Nishiuchi, University of Toronto, Canada

Publication history

  1. Received: February 28, 2020
  2. Accepted: April 9, 2021
  3. Accepted Manuscript published: April 12, 2021 (version 1)
  4. Version of Record published: April 23, 2021 (version 2)

Copyright

© 2021, Mount et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 1,422
    Page views
  • 205
    Downloads
  • 0
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

Further reading

    1. Neuroscience
    Kristen Lee, Chris Q Doe
    Research Advance Updated

    Individual neurons can undergo drastic structural changes, known as neuronal remodeling or structural plasticity. One example of this is in response to hormones, such as during puberty in mammals or metamorphosis in insects. However, in each of these examples, it remains unclear whether the remodeled neuron resumes prior patterns of connectivity, and if so, whether the persistent circuits drive similar behaviors. Here, we utilize a well-characterized neural circuit in the Drosophila larva: the moonwalker descending neuron (MDN) circuit. We previously showed that larval MDN induces backward crawling, and synapses onto the Pair1 interneuron to inhibit forward crawling (Carreira-Rosario et al., 2018). MDN is remodeled during metamorphosis and regulates backward walking in the adult fly. We investigated whether Pair1 is remodeled during metamorphosis and functions within the MDN circuit during adulthood. We assayed morphology and molecular markers to demonstrate that Pair1 is remodeled during metamorphosis and persists in the adult fly. MDN-Pair1 connectivity is lost during early pupal stages, when both neurons are severely pruned back, but connectivity is re-established at mid-pupal stages and persist into the adult. In the adult, optogenetic activation of Pair1 resulted in arrest of forward locomotion, similar to what is observed in larvae. Thus, the MDN-Pair1 neurons are an interneuronal circuit – a pair of synaptically connected interneurons – that is re-established during metamorphosis, yet generates similar locomotor behavior at both larval and adult stages.

    1. Neuroscience
    Philipp Schlegel et al.
    Research Article Updated

    The hemibrain connectome provides large-scale connectivity and morphology information for the majority of the central brain of Drosophila melanogaster. Using this data set, we provide a complete description of the Drosophila olfactory system, covering all first, second and lateral horn-associated third-order neurons. We develop a generally applicable strategy to extract information flow and layered organisation from connectome graphs, mapping olfactory input to descending interneurons. This identifies a range of motifs including highly lateralised circuits in the antennal lobe and patterns of convergence downstream of the mushroom body and lateral horn. Leveraging a second data set we provide a first quantitative assessment of inter- versus intra-individual stereotypy. Comparing neurons across two brains (three hemispheres) reveals striking similarity in neuronal morphology across brains. Connectivity correlates with morphology and neurons of the same morphological type show similar connection variability within the same brain as across two brains.