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

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.

The hippocampus is critical for learning and memory in animals and humans. Early 29 surgical lesions of the hippocampus in human patients, designed to alleviate intractable 30 epilepsy, resulted in severe memory loss and an inability to form new declarative or 31 episodic memories 1,2 . Hippocampal atrophy is also associated with diseases related to 32 memory loss and cognitive decline including dementia and Alzheimer's disease 3-7 . Many 33 mechanistic studies have highlighted the importance of the hippocampus for spatial, 34 contextual, and associative learning in a variety of animal models 8,9 .  The hippocampus is also required for context-dependent extinction learning 11 . 47 Extinction learning is traditionally considered new learning that overrides a previously 48 learned relationship. In the example of trace conditioning, the subject learns that the 49 previously established CS is no longer predictive of a subsequent US. Extinction learning 50 after trace conditioning can be tested by the presentation of the CS without the associated 51 3 US, and monitoring the strength or presence of a conditioned response. As new learning 52 occurs, subjects will suppress their conditioned response to the previously predictive tone 53 or light. Extinction learning has also been shown to be NMDA receptor dependent 28  Syn-GCaMP6f and implanted with a custom window that allowed optical access to dorsal 99 CA1 ( Figure 1C). 4-6 weeks after surgery, mice were habituated and then trained on a 100 classic trace eye-blink conditioning paradigm followed by an extinction training session 101 ( Figure 1B). The paradigm consisted of a 9500Hz, 350ms tone as a conditioned stimulus 102 (CS), followed by a 250ms trace interval, followed by a 100ms gentle puff of air to one 103 eye that served as the unconditioned stimulus (US) ( Figure 1D). Eye behavior was 104 monitored with a USB 3.0 Camera ( Figure 1A,Ei). Animals were trained for 60-80 CS-US 105 trials over 5-9 days, until they reached conditioned response criterion (conditioned 106 response on 65% of trials). After reliable conditioned response to CS presentations was 107 established, on the final day of imaging animals were given a block of 20-40 CS-US trials 108 (last training session), followed by a block of CS-only extinction trials, where the CS was 109 not followed by the US (extinction session, Figure 1B). In this paradigm, we could perform 110 calcium imaging of the same neurons during both learning conditions, allowing us to track 111 how activity of each neuron changes during extinction acquisition.

112
Behavioral response was quantified by segmenting the eye videos and averaging 113 each frame to calculate a temporal trace of eyelid movement (Figure 1Ei-iii). A movement 114 threshold was calculated for each eye trace, fit to a uniform distribution equal to the 115 average eye size. This thresholding method provided a consistent evaluation of 116 behavioral response for each session across mice: eye closure with an amplitude above 117 the threshold between the tone onset and puff onset (tone-puff window) was classified as 118 a conditioned response (Figure 1Ei-Eiv). Using this method, we were able to track the 119 strength of the response to the CS, as well as the strong, persistent eye closure in 120 response to the aversive US on each trial (Figure 1Eiv). This method also allows for  Benjamini-Hochberg procedure) (Figure 2A  averaged across all trials, and the entire population was sorted by average response 145 7 intensity during the time period between the tone onset and puff onset (tone-puff window).

146
During the last training session, after multiple days of conditioning, substantially more 147 neurons (11.3 ± 6.2%: mean ± s.d) exhibited an increased calcium response between the 148 tone and puff, compared to the first day of training (5.6 ± 3.7%: mean ± s.d., p=0.0463,   Figure 2B). When neuronal responses were averaged together and analyzed as 162 described above, we found that 8.3 ± 5.4% (mean ± s.d.) of cells were responsive to CS 163 during extinction session, which is significantly higher than the percentage expected by 164 chance (bootstrapped estimation, extinction: N=1000, p=0.011, one-tailed bootstrap, 165 alpha=0.05, Supp. Figure 1B). 166 We compared the responses of individual cells to the CS during conditioning trials   versus 11.3 ± 6.2%: mean ± s.d.: p=0.016, one-tailed paired t-test, Figure 2D).

181
The population of cells that responded to the CS (increased fluorescence during 182 the tone-puff window) during the last training session were termed conditioned (CO) cells.

183
In contrast, extinction (EX) cells were determined as those that responded to the CS 184 during extinction. These populations were largely discrete, as only 21.8 ± 15.7% (mean ± 185 s.d.) of CO cells were also EX cells, and similarly only 32.8 ± 22.3% (mean ± s.d.) of EX 186 cells were also CS cells. These results suggest that during extinction learning, neurons 187 are recruited to encode tone presentations rapidly and emerge in less than 40 trials.  showed responses to the CS on multiple conditioning or extinction trials (example CO 195 cells: Figure 3A; example EX cells: Figure 3B). However, most individual CA1 neurons    We generated network maps for both correct and incorrect trial co-occurrence 281 matrices to again investigate differences in activity patterns (Figure 5Bi). The correct trials 282 map included 66.6 ± 22.8% (mean ± s.d.) of the total edges present during the last training 283 session and the incorrect trials map included 38.7 ± 21.6% (mean ± s.d.) of these total 284 edges (p=0.1355, two-tailed paired t-test, Figure 5C). The connectivity density and 285 average degree of the incorrect trials network map were slightly smaller than those for 286 13 the correct trials map, but not significantly (p=0.2258 & 0.2849 for density and degree 287 respectively, two-tailed independent t-test, Supp. Figure 4). However, the edges were 288 again almost completely distinct (Figure 5Bii-iii). In fact, only 5.3 ± 8.0%: (mean ± s.d.) of 289 total edges appear on both the correct and incorrect trial maps (p=0.0001 & 0.006 for 290 correct and incorrect trials versus shared edges respectively, two-tailed paired t-test, 291 Figure 5C). These findings again highlight the diverse co-activity among neuron pairs in 292 the same behavioral task, but across different behavioral responses, while maintaining 293 the overall CA1 network activation.  Additionally, it can be used to break down and compare trials by specific behaviors (i.e., 364 correct versus incorrect) or other variables that may change across trials over time.

365
Finally, the co-occurrence matrix allows us to consider connectivity maps of entire neuron 366 populations, an intuitive way to visualize and investigate the patterns of neural activation.

367
Overall, the co-occurrence matrix is a useful technique for monitoring the evolution of 368 population responses over time from high-dimensional calcium imaging datasets.

369
Using the co-occurrence matrices, we found that CA1 neurons' connectivity 370 changes drastically between conditioned learning and extinction training, but also 371 between trials with the correct or incorrect behavioral response during conditioned  indicate that on each trial only a subset of the appropriate sub-population is activated to 382 encode the relevant features of that trial, but also that these subsets work together to 383 create a larger network that represents learning across an entire session, which may be 384 critical to the encoding or retrieval of learning and memory in CA1.

402
The pipette was lowered over 3 min and allowed to remain in place for 3 min before 403 infusion began. The rate of the infusion was 100 nL/min. At the conclusion of the infusion, 404 the pipette remained in place for 10 min before slowly being withdrawn over 2-3 minutes.     Objective 10 X 0.28). Images yielded a field of view 1.343 mm by 1.343 mm (1024x1024 453 pixels) and were acquired at a 20 Hz sampling rate and stored offline for analysis.  For sessions where ROIs were matched to one another, spatial ROI maps were 508 co-registered using frame-wise cross-correlation. ROIs were then matched using a 509 greedy method that required the centroid of cells to be within 50 pixels of one another 510 and had to have at least 50% of their pixels overlap, as was published previously 58 . Cells 511 that did not meet both of those criteria were removed from the matched dataset for 512 comparison. (100 pixels) and subtracting the area for the ROI from that circle. The pixels in this local 518 background were averaged together spatially to measure a temporal background trace.

519
Background traces were subtracted from each cell's measured trace to remove local 520 fluctuations from scattering in wide-field imaging. The baseline calcium level was 521 calculated for each cell by fitting a normal distribution to the lowest 50 percentile of the 522 data and using the mean of this distribution as the baseline calcium level. This baseline 523 was subtracted from each locally corrected trace, and data was scaled by 5% of the 524 maximum range of the full calcium trace.

526
For trial-averaged analysis, all trials of the last training session were included.

527
Fluorescence for the 12 data points (600 ms) within the interval between tone onset and 528 puff onset (tone-puff window) was compared to the 12 data points prior to the tone. As within that spatial distribution were calculated and a percentage of either CO or EX cells 559 was determined from the cell identities from the cells within that radius. For bootstrapping, 560 the same number of CO or EX cells that was segmented for each recording session were 561 randomly selected and the same calculation within a 100 micron radius was calculated.

25
The measured percentages were then compared to the bootstrapped values for statistical 563 confidence.

565
Individual trial co-occurrence matrices were created for each pair of cells across 566 every trial. For consistency in the co-occurrence analysis, the last 20 trials of the last 567 training session were included. For each tone-puff window, the mean value of the 600 ms 568 (12 data points) between the tone onset and puff onset was compared to the 600 ms 569 before the tone. If this value was greater than 1 on a normalized trace, corresponding to 570 5% the maximum peak value of a trace, then the neuron was labelled as responding. The 571 result was a binary vector of 0s and 1s of length N, where N is the number of cells 572 recorded in the population. The outer product of this vector was taken with itself for the 573 whole population to yield an NxN co-occurrence matrix. This matrix is 1 if both the ith 574 and jth cells were activated between the tone-puff, and 0 otherwise.

575
Once a co-occurrence network was generated for each trial, they could be 576 combined for further analyses by summing on certain trials of interest. For this analysis, 577 co-occurrence matrices were summed across either the last training session, the 578 extinction session, "correct" trials of the last training session, or "incorrect" trials of the last 579 training session. Once a trial combination map was created, spectral biclustering was 580 performed for a 3x3 cluster pattern using the Python machine learning package scikit-581 learn 41,59 .

583
Anatomical spatial information was combined with the co-occurrence matrix to 584 generate network maps using the Python library NetworkX. The centroid of each ROI was 585 26 used as the position of the corresponding node, which represent the cells of the imaging 586 session. Since the co-occurrence matrix is symmetric, the lower triangular matrix is used 587 to generate the edges of the network. Co-activity between two cells is represented as an 588 edge between the corresponding nodes. For example, the ith cell and jth cell would be 589 connected by an edge if Ai,j is non-zero, where A is the NxN co-occurrence matrix. The where is total edges in correct network and is total edges in incorrect network.

600
Python package NetworkX was used to calculate network density, which is defined