Abstract
The hippocampal spatial code’s relevance for downstream neuronal populations – particularly its major subcortical output the lateral septum (LS) - is still poorly understood. Here, using calcium imaging combined with unbiased analytical methods, we functionally characterized and compared the spatial tuning of LS GABAergic cells to those of dorsal CA3 and CA1 cells. We identified a significant number of LS cells that are modulated by place, speed, acceleration, and direction, as well as conjunctions of these properties, directly comparable to hippocampal CA1 and CA3 spatially modulated cells. Interestingly, Bayesian decoding of position based on LS spatial cells reflected the animal’s location as accurately as decoding using the activity of hippocampal pyramidal cells. A portion of LS cells showed stable spatial codes over the course of multiple days, potentially reflecting long-term episodic memory. The distributions of cells exhibiting these properties formed gradients along the anterior-posterior and dorsal-ventral axes of the LS, directly reflecting the topographical organization of hippocampal inputs to the LS. Finally, we show using trans-synaptic tracing that LS neurons receiving CA3 and CA1 excitatory input send projections to the hypothalamus and medial septum, regions that are not targeted directly by principal cells of the dorsal hippocampus. Together, our findings demonstrate that the LS accurately and robustly represents spatial, directional as well as self-motion information and is uniquely positioned to relay this information from the hippocampus to its downstream regions, thus occupying a key position within a distributed spatial memory network.
Introduction
The lateral septum (LS) is an anatomically complex structure uniquely positioned within a broader distributed spatial memory network. The LS consists almost exclusively of highly interconnected inhibitory GABAergic cells [1, 2] that are densely innervated by the principal cells of the hippocampus [3, 4], in contrast to the medial septum (MS), which receives inputs solely from the inhibitory neurons of the hippocampus [5]. Previous research has implicated the LS in a wide variety of behaviors, including spatial and working memory [6–10], regulation of feeding [11–14], anxiety [15–18], locomotion [19, 20], and social behaviors [21–24].
To date, functional characterizations of the LS have often focused on the spatial coding properties of these neurons, primarily due to the extensive excitatory input from positionally-tuned hippocampal pyramidal neurons, commonly referred to as ’place cells’ [4,24–26]. Yet, this body of work has not yet yielded a consensus. Several studies have reported place-like cells in the LS, but the number, information content, and stability vary widely from study to study [20,27–31], while others described a lack of canonical place cells altogether [4]. As previous studies often depended on variable criteria of what defines a place cell, estimates range from 5.3% [4] to 56.0% [20] of LS cells classified as spatially-modulated. Strikingly, even estimates by the same authors on data acquired from the same subjects engaging the same task ranged between 26.5% [32] to 56.0% [20]. When reported, LS place cells are typically described as of lesser quality than classic hippocampal place cells, with lower within-session stability and larger place fields [30, 31]. Adding to this complexity, the LS is a large structure that is cytoarchitecturally subdivided into dorsal, intermediate, and ventral subregions that spread across its anterior-posterior axis [5,26,33]. This suggests that, in addition to subjective criterias, the striking differences in LS spatial coding characteristics between studies may be in part a product of variations in recording locations within the region. In addition to positional information, LS neurons may also encode direction and self-motion information, including acceleration and velocity [20,30,31].
Beyond their spatial tuning properties, hippocampal place cells are also noted for the rapid reorganization of the place fields across time (on the order of days) and navigational contexts, a phenomenon termed remapping [34–38]. If the LS spatial code is directly inherited from its hippocampal inputs, one would expect comparable dynamics over a similar timescale. While the LS, similar to the hippocampus, has been implicated in behaviors that may require a stable representation of space over time, including context-reward [39, 40], contextual fear conditioning [16,41–44], spatial learning and memory [32,45,46] remapping dynamics in the LS remain largely understudied. Thus, whether and how the LS spatial code changes across time is crucial to clarifying its relationship to the upstream hippocampal spatial code, with important implications for its involvement in this broader spatial memory network and how it supports behaviors across longer timescales.
To resolve these outstanding questions, we functionally characterize hippocampal connectivity of LS, and asked how information rich LS GABAergic neurons are in relation to their main pyramidal inputs in dorsal CA3 and CA1. To this end, we used head-mounted miniaturized microscopes to record ∼2000 LS neurons across the anterior-posterior and dorsal-ventral axes of the LS, and compared their firing characteristics to pyramidal cells recorded in CA1 and CA3, as animals navigated linear track and open field environments. Results demonstrate that the LS accurately and robustly represents spatial, speed, and directional information in an anatomically-organized fashion. While the information of LS neurons is generally comparable to CA3 and CA1, our tracing studies suggest that LS information is sent to downstream brain regions such as the MS and hypothalamus that are not directly connected by the principal cells of the dorsal hippocampus.
Results
LS GABAergic neurons exhibit stable spatial activity during alternation on a linear track
Previous electrophysiological studies of the LS disagree on both the proportion of place encoding cells in this region and their spatial content. Cells have been previously described to be of lower quality than the classically described hippocampal place cells [20,30,32] or to be virtually absent [4].
First, to assess the distribution of excitatory pyramidal inputs from each hippocampal subregion to the GABAergic cells of the LS, we injected a monosynaptic retrograde rabies tracer in dorsal LS (Fig 1A, S1 Fig, n = 5 mice), utilizing the monosynaptically restricted, Cre-dependent retrograde tracing features of this technique. We observed strong labelling of pyramidal neurons of the hippocampus in areas CA1 and CA3 (Fig 1B, S1 Fig). To investigate how the spatial coding properties of the LS compare to these main dorsal hippocampal inputs, we expressed the calcium (Ca2+) indicator GCaMP6f in GABAergic cells of the LS or in the pyramidal cells of the dorsal CA1 (dCA1) and dorsal CA3 (dCA3) of mice (Fig 1C, S2 Fig, S3 Fig). We recorded Ca2+ activity from 1082 LS GABAergic neurons (n = 15 mice), 1251 dCA1 pyramidal cells (n = 5 mice) and 464 cells from dCA3 (n = 6 mice) from animals alternating on a linear track. We subsequently filtered and binarized the recorded calcium traces (S4A Fig) [47] and excluded periods of immobility (< 5cm.s-1). Using an information-theoretic framework [48–50], we computed the mutual information (MI) between calcium activity and spatial location for each cell and expressed the results in bits (S4B Fig). As we observed a significant correlation between probability of a cell to be active (P(A)) and MI (S4B Fig), we computed the MI value for 30 bootstrap samples (50% data sampling) and compared those to 30 circularly shuffled surrogates, which allowed us to compute significance of the MI value independently of P(A) (S4C-E Fig) [47]. We found that LS GABAergic activity was significantly spatially modulated (p < .01, Fig 1D, S4H Fig). Indeed, both mutual information and within-session stability was higher for spatially modulated LS cells compared to non-modulated cells (Mann-Whitney test, p < .0001, S4E-F Fig). Using this approach, a total of 43.90% of LS GABAergic cells qualified as spatially modulated (475/1082 cells), with LS cells similarly distributed along the linear track as place cells recorded from dorsal CA1 and dorsal CA3 (Fig 1E). In contrast to previous reports [4,20,32], we observed a significant difference in the proportion of spatially modulated cells across recording regions (LS: 43.57 ± 4.46 %; dCA1: 29.60 ± 2.99 %; dCA3: 27.01 ± 7.07 %; Kruskal-Wallis, H(3) = 6.212, p = .0448; Dunn’s multiple comparisons test: LS-dCA1: p = 0.1757; LS-dCA3: p = 0.1101; Fig 1F). We found that within-session stability of spatially modulated cells (expressed as a correlation where 0 means no stability while 1 means perfectly stable) was significantly different between LS (0.44 ± .0165) and dCA3 (0.73 ± 0.26; Kruskal-Wallis, H(3) = 80.48, p < .0001; Dunn’s multiple comparisons test, p < .0001), and between dCA1 and dCA3 (p < .0001; Fig 1G). Even though LS GABAergic cells are not fast spiking (S5 Fig, S6 Fig) [51], P(A) in LS (0.02005 ± 4.692 x 104) is higher than both dCA1 (0.0094 ± 2.544 x 104, Kruskal-Wallis, H(3) = 502.0, p < .0001; Dunn’s multiple comparisons test, p > .0001) and dCA3 (0.009225 ± 3.759 x 104, p < .0001, S6A,B). Therefore, we assessed the mutual information in bits carried in each binarized event, and observed that the LS (5.624 x 105 ± 1.151 x 106 bits/binarized event) carries significantly less information per binarized event than CA1 (1.430 x 104 ± 1.613 x 106 bits/binarized event; Kruskal-Wallis, H(3) = 1288, p < .0001; Dunn’s multiple comparisons test, p < .0001); and CA3 (1.130 x 104 ± 2.507 x 106; bits/binarized event; p < .0001, Fig 1H). To confirm that the observed differences in proportion of spatially modulated cells and within-session stability were not due to differences in P(A), we used an activity cut-off of P(A) > .001 and confirmed our previous results (S6D Fig).
In contrast to their omnidirectional firing properties in open field environments, hippocampal neurons rapidly develop strong directional selectivity on linear tracks [52–54]. To test whether the LS place code developed comparable directional selectivity, we ranked spatially modulated cells as a function of the difference between MI for left and right traversals (S7A-D Fig), and found LS cells tuned to the absolute location regardless of direction (7.67%, 83/1082 cells), while other cells were selective to the left- (15.52%, 168/1082 cells), or right-trajectories (22.08%, 239/1082 cells).
Previous studies have used decoding methods to predict behavioral variables from calcium imaging data recorded in the hippocampus [34,38,55], yielding insights into the amount of information encoded by a neuronal assembly. Here, we asked whether we could reliably estimate the mouse location solely from LS neuronal activity patterns (Fig 2). Using 30 bootstrap samples of 60 cells (Fig 2B,C), decoding location using LS neuronal activity significantly outperformed a decoder that was trained using shuffled data (paired t-test, t(6) = 13.56, p < .0001, Fig 2D), even when decreasing the number of neurons used (S8A,B Fig). Similarly, a decoder trained using CA1 or CA3 data also yielded significantly less error than one using shuffled surrogates (CA1: t(4) = 7.558, p = .0016; CA3: t(3) = 3.233, p = .0481, Fig 2D, S8C,D Fig). While using a small bootstrap sample size allowed a fair comparison between recording regions, it also induced higher error rates. To ensure the proper functioning of our decoder, we confirmed that the decoding error decreased significantly when increasing bootstrapped sample size (S8C,D Fig).
In order to compare decoding accuracy for LS, CA1 and CA3, a decoding score was computed (Fig 2E, see methods), and we found that LS significantly outperformed CA3 (one-way ANOVA, F(2,13) = 6.277, p = .0124; Holm-Sidak’s multiple comparisons test p < .05), but not CA1. Overall, spatially modulated cells recorded in LS were significantly more active than those recorded in the hippocampus (H(3) = 155.1, p < .0001, S5A,B Fig), which could be a contributing factor to the overall higher decoding accuracy using LS cells. To confirm that the differences observed in decoding error are not due to the inclusion of a large number of pseudo-silent cells in the dCA1 or dCA3 bootstrapped samples, we confirmed our findings using an activity cut-off of P(A) > .001 (S8E-F). Similarly, to confirm that the observed differences in decoding error are not due to the use of a temporal smoothing window, we replicated our results using a decoder that omitted temporal filtering (S8G-H).
LS GABAergic neurons exhibit stable spatial activity during free exploration in a 2D environment
In order to assess spatial coding during free exploration in a two-dimensional environment, we recorded Ca2+ activity of 1899 GABAergic neurons in 28 mice implanted in distinct subregions of the LS, while the animals were freely exploring a novel open field. Out of 1899 recorded cells, 37.80% (718/1899 cells, n = 28 mice) of LS cells displayed significant spatial information (p < .01; Fig 3A-B, S9 Fig), as compared to 31.33% (323/1031 cells, n = 6 mice) in CA1 and 25.84% (138/534 cells, n = 8 mice) CA3 cells. Despite the LS’ known involvement in feeding-related behaviors [11–14], we did not observe any increased number of cells with firing fields around food zones, nor did we observe overrepresentation of objects or walls (S10 Fig). To assess reward-modulation in the context of increased task demand, we assessed spatial firing of LS GABAergic neurons recorded while animals performed a delayed non-matching to place task in a T-maze, but did not observe any significant reward code compared to shuffled surrogates (S11 Fig).
With theta-rhythmic cells in the LS having higher firing rates than theta-rhythm independent cells [56], and the suggestion that theta-rhythmic cells could receive direct hippocampal inputs (though see [57, 58]), we asked whether LS cells with higher firing rates were also more likely to be spatially modulated. Indeed, spatially-modulated cells were significantly more active than non-modulated cells (Mann-Whitney test, U = 342219, p < .0001; S9G-H Fig) and displayed a higher bursting index, defined as the probability of a cell being active, given it was already in an active state, or P(At|At-1), than non spatially modulated cells (Mann-Whitney test, U = 368276, p < .0001; S9H Fig).
We compared the place field properties of cells recorded in LS to those of dorsal CA1 and CA3 (Fig 3). For each animal, we computed the portion of significantly positionally modulated cells per region, which was not significantly different across LS (36.71 ± 2.16 %), dCA1 (38.32 ± 9.36 %), and dCA3 (28.39 ± 6.73 %; one-way ANOVA, F(2,39) = 1.105, p = .3414; Fig 3B). We computed the information for each binarized event, and observed that the LS (1.020 x 104 ± 7.56 x 107 bits/binarized event) carries significantly less information per binarized event than CA1 (1.200 x 104 ± 1.067 x 106 bits/binarized event; Kruskal-Wallis, H(3) = 264.6, p < .0001; Dunn’s multiple comparisons test, p < .0001) and CA3 (1.199 x 104 ± 1.605 x 106 bits/binarized event; p < .0001). As expected on the basis of prior results [4,28,30,31], within-session stability of LS cells (0.31 ± 0.011) was significantly lower than that of spatially modulated cells recorded in CA1 (0.44 ± 0.017; Kruskal-Wallis, H(3) = 89.52, p < .0001; Dunn’s multiple comparisons test, p < .0001) and CA3 (0.56 ± 0.027; p < .0001; Fig 3D), and place fields were more dispersed (LS: 14.47 ± 0.13 cm; CA1: 12.00 ± 0.207 cm; CA3: 10.61 ± 0.35 cm; Kruskal-Wallis, H(3) = 155.7, p < .0001, Fig 3E). Similar to recordings in a 1D environment, we observed large differences in P(A) between spatially modulated cells recorded from LS and those from dorsal CA1 and CA3 (S12A Fig). Therefore, we also assessed for each region the proportion of spatially modulated cells, split-half stability and mean-dispersion using an activity cut-off of P(A) > .001, and confirmed our results (S12B Fig).
When decoding the location of the animals using bootstrap samples of 80 cells from each region (Fig 3F), decoding error was significantly lower in LS (20.16 ± 1.48 cm) compared to shuffled surrogates (26.03 ± 0.84 cm; two-way RM ANOVA, F(1,13) = 89.09, p < .0001 for main effect of shuffle), even when decreasing the number of neurons used to only 40 neurons (S13A,B). Omission of temporal filtering recapitulated these results (S13E Fig). The same held true for decoders trained using CA1 or CA3 data respectively (S13C,D). Strikingly, a decoder trained on data from LS does not perform significantly worse than a decoder trained on CA1 or CA3 data (F(2,13) = 2.186, p = 0.2384 for main effect of region; Fig 3H). Similarly to recordings on the linear track, we used an activity cut-off of P(A) > .001 to confirm that the differences observed in decoding error between regions are not due to the inclusion of a large number of pseudo-silent cells in the bootstrapped sample (S13F Fig).
A subset of LS cells has stable spatial representations up to 8 days, similar to CA1
To test whether the LS and the hippocampus exhibit comparable evolutions of their spatial code over time, we recorded from animals implanted in LS and dorsal CA1 in a novel open field over both short (3 days) and longer timescales (8 days, Fig 4, S14 Fig). We assessed the stability of the spatial map across sessions, and observed a subset of LS spatially modulated cells that were stable over time (Fig 4B). Correlating the tuning maps of aligned LS cell pairs over days lead to a significant increase in mean pairwise correlation value over days (two-way ANOVA, F(2, 808) = 3.259, p = 0.0389, main effect of time) with LS cells being statistically more stable than shuffled comparisons (F(2,808) = 277.0, p < .0001, main effect of shuffling, Fig 4C, S14C Fig). We observed the opposite pattern for pyramidal cells recorded from dorsal CA1, with a significant decrease in mean pairwise correlation value over days (two-way ANOVA, F(2,766) = 33.47, p < 0.0001 main effect of time, F(1,766) = 319.8, p < 0.0001 main effect of shuffle, Fig 4C). For cells recorded in the LS, we observed the strongest within-session stability on day 3 (Kruskal-Wallis, H(4) = 19.16, p = 0.0003; S14E Fig), as well as an increase in spatial information (Kruskal-Wallis, H(4) = 13.27, p = 0.0041; S14B Fig). There was no significant increase in the proportion of stable cell pairs over subsequent days (one-way ANOVA, F(3) = .400, p = .9537, S14D Fig), or the proportion of spatially modulated cells per animal for each days (one-way ANOVA, F(3,16) = 1.618, p = .2247, S14E Fig).
We next assessed the significant tuning map correlations on subsequent day-pairs for those cells that were significantly spatially modulated on day 1 and were subsequently found on all recording days (Fig 4D), as well as those that passed our criteria for spatial modulation on all days (Fig 4E). We then compared the place field correlation for different day pairs (F(5,363) = 2.625, p = .0239, interaction effect), and we found that on shorter timescales, tuning maps for LS cells (0.316 ± 0.0507) are significantly less correlated than CA1 (0.497 ± 0.0411; Fisher’s LSD day 1-2, p = 0.0153, Fig 4F). On the other hand, the correlation for day pairs 3-8 is higher for the LS (0.369 ± 0.0513) compared to dorsal CA1 (0.174 ± 0.0669; Fisher’s LSD, p = 0.0098; Fig 4F). Together, this suggests that a subset of LS spatial cells encodes spatial information over longer periods of time.
LS represents direction, velocity, and acceleration information
In CA1, place cells have been found to carry directional [59–61] and speed-related information [61–64]. Previous studies found that subsets of LS neurons show some degree of modulation by direction of travel [31] as well as velocity and acceleration [20] through correlative measures in spatial alternation or T-maze navigation tasks, where the relationship between cell activity patterns and movement-associated variables may be confounded with task-dependent variables. Here, we again employed an information-theoretic approach to compute the MI between calcium activity and each of these self-motion correlates in a subset of mice recorded during a 15 min free foraging task in the open field. We found that 28.13% of LS cells are significantly modulated by head-direction (346/1230 cells, n =19 mice, Fig 5A, S15 Fig), 18.70% by velocity (230/1230 cells, Fig 5C) and 24.63% by acceleration (303/1230 cells, Fig 5E). We assessed the stability of LS directional and self-motion tuning over short (3 days) and longer timescales (8 days, S16 Fig), and we observed a subset of LS directionally modulated cells that were stable over time (S16A Fig). Mean tuning vectors correlation of aligned LS cell pairs significantly increased over days (two-way ANOVA, F(2,774) = 3.682, p = 0.0256, main effect of time) with LS cells being statistically more stable than shuffled surrogates (F(1,774) = 54,42, p < .0001, main effect of shuffling, S16C Fig). For velocity or acceleration tuning, we did not observe such stable tuning over time (two-way ANOVA, velocity: F(1,506) = 2.609, p = .1069, acceleration: F(1,320) = .1986, p = .6561, main effect of shuffle, S16D,E).
We compared the proportion of LS cells significantly modulated by each of these variables with the proportion of those cell-types found in dCA1 and dCA3. The proportion of directionally modulated cells was significantly different across regions (LS: 27.89 ± 2.16, dCA1: 46.11 ± 4.98, dCA3: 28.92 ± 4.52; Kruskal-Wallis, H(3) = 6.586, p = .0371, Fig 5B) with a higher proportion of directionally modulated cells in CA1 as compared to LS (Dunn’s multiple comparisons test, p = .0340; Fig 5B). The proportion of velocity encoding cells was not significantly different between LS (20.43 ± 2.598), CA1 (24.38 ± 7.73), or CA3 (20.04 ± 3.62; Kruskal-Wallis, H(3) = 0.1016, p = 0.9504, Fig 5D). Similarly, no difference was found in the proportion of acceleration encoding cells between LS (25.11 ± 2.34), CA1 (24.14 ± 4.13), or CA3 (22.60 ± 2.98, Kruskal-Wallis, H(3) = 0.3337, p = 0.8463, Fig 5F). We found a proportion of cells that were significantly tuned to more than one of these variables in both LS as well as dCA1 and dCA3 (Fig 5G, S15A-D Fig, S17 Fig), with 22.03% of LS cells (271/1230 cells) encoding more than one modality (Fig 5H). Together, this conjunctive coding for location, directionality, and velocity indicates that LS cells fire in response to more complex navigational features, similarly to the hippocampus.
To understand the function of this downstream copy of the hippocampal spatial code, we assessed the downstream targets of the dCA1/dCA3 to LS projection. First, we used a Cre-dependent, anterograde tracing approach to confirm LS projections to the hypothalamus and ventral tegmental area in the mouse (S18 Fig), as it was initially described in the rat [3,5,25]. Next, we leveraged the anterograde transsynaptic properties of Cre-expressing AAV1 viral vector [65] and assessed the targets of the dorsal CA1-lateral septum (dCA1-LS) as well as dorsal CA3-LS (dCA3-LS) projections (Fig 6). For either dCA1 or dCA3 injections (Fig 6A, S19 Fig), we observed a direct pathway from the principal cells of the hippocampus, via the LS, leading to dense innervation of the medial septum (Fig 6B) and the hypothalamic regions (Fig 6C). We observed that dCA1-LS preferentially targets the LH, and dCA3-LS targets the hypothalamus more broadly as well as the nuclei of the medial zone. Whether the GABAergic cells of the lateral septum form functional synapses at the level of the medial septum/diagonal band complex is debated [5,26,66–68]. Thus, in order to determine whether the projections observed at the level of the MS with both our anterograde tracing approaches constituted passing fibers or synaptic connections, we used an AAV-Flex-Synaptag [69], a Cre-dependent viral construct to express mRuby in the cytosol and eGFP bound to synaptophysin, a protein primarily expressed in the presynaptic endings of cells participating in synaptic transmission (Fig 6D). For this, we used a CaMKIIα-Cre mouse line, expressing Cre only in CAMKII-positive cells, a marker that is abundant in the GABAergic cells of the LS but absent in the MS, thereby preventing any unspecific labelling. After expressing AAV-Flex-Synaptag in the intermediate LS, we observed significant synaptophysin-bound projections in the MS (Fig 6E), confirming the existence of a synaptic interface between the LS and the MS. Together, our work shows that LS receives direct inputs from dorsal CA1 and CA3 and in turn projects to regions that are not directly receiving inputs from the hippocampus itself (Fig 6F).
LS-cells coding for space, direction and self-motion are non-uniformly distributed
To map out the monosynaptic connections between hippocampal and LS neurons, we injected the anterograde transsynaptic Cre-expressing AAV1 viral vector [65], in either dCA1 or dCA3 (S19 Fig, S20A Fig). We subsequently injected a Cre-dependent eYFP expressing virus in the LS to visualize cells that receive direct hippocampal input (S20B Fig). Importantly, this approach allowed us to quantify the cells along the LS anterior-posterior and dorsal-ventral axis receiving direct hippocampal inputs, and assess their distribution density within the anterior regions of the LS (Bregma +1.0 to 0.0 anterior-posterior). Because we can only observe secondary labelled cell-bodies at locations where there was infusion of the secondary, Cre-dependent virus, we used a relatively large injection volume (total 800 nL, targeted at the intermediate LS). Both dorsal dCA1 and dCA3 injections resulted in a majority of labelled LS cell bodies in dorsal regions compared to intermediate and ventral LS (S20D-F Fig). Hence, LS neurons receiving direct inputs from dorsal CA1 and CA3 appear to be localized at more dorsal-posterior regions as compared to more ventral-anterior regions of LS.
Next, we examined whether spatially modulated cells recorded in LS had spatial properties that were dependent on the localization of their recording. To measure this, we classified the location of GRIN lens implants along the dorsal-ventral and anterior-posterior axis (Fig 7A). Post-mortem histological verification of GRIN lens implant combined with within-animal analysis of the location of recorded cells enabled us to approximate the location of each cell (Fig 7B). While we observed no differences in cell activity characteristics, such as bursting index or activity probability along the anterior-posterior axis, medial-lateral or dorsal-ventral axis (S21 Fig), we found a pronounced increase in the portion of stable spatially modulated cells (within-session stability > .5) along the anterior-posterior axis (linear regression, R2 = 0.1071, p = 0.0022, the proportion of stable cells per 0.2 mm bin per animal, n = 24 mice, Fig 7C). We observed a similar gradient along the dorsal-ventral axis, with a larger proportion of stable spatial cells found at the more dorsal regions of the LS, although this trend failed to reach significance (linear regression, R2= 0.1399, p = .0718, n = 24 mice, Fig 7C). Taken together, these data strongly suggest that regions of the LS that receive stronger innervation from the dorsal hippocampus have a larger proportion of cells that reliably encode space. This suggests that this information could be directly inherited from the hippocampus, which should be tested using targeted inactivation of the hippocampal pyramidal inputs to LS.
We also assessed the topographical distribution of direction and self-motion tuned cells within the LS. We observed a relatively strong dorsal-ventral gradient in the distribution of stable (within-session stability > .5) directionally modulated cells (linear regression, R2 = 0.3837, p = 0.0061, n = 17 mice, Fig 7C), with most directionally modulated cells located at the more dorsal regions of the LS, but we did not observe such a gradient along the anterior-posterior axis (linear regression, R2 = 0.0544, p = .0615, n = 24 mice, Fig 7C). Velocity modulated cells were primarily found at the ventral pole of the LS (linear regression, R2 = 0.1836, p = 0.0367, n = 24 mice), but no such gradient was found along the anterior-posterior axis (linear regression, R2 = .00348 p = .0874, n = 24 mice). Acceleration-modulated cells were found evenly distributed throughout the LS (DV: R2 = .0133, p = .2928; AP: R2 = .01332, p = .2929, n = 24 mice).
Discussion
While spatial coding has been extensively characterized in the hippocampal formation, how downstream regions integrate this information has only recently begun to receive attention. The LS has re-emerged as a critical region implicated in a space encoding network, although electrophysiological recordings of the LS in rats have led to disparate estimates of the quantity and quality of place cells in the region [4,20,28,32]: LS place cells recorded on a circular track were described to be almost absent [4], but were found to be abundant in reward-seeking tasks [20, 32]. Using a large calcium imaging dataset and unbiased information metric and decoding approaches, we found that 37.80% of GABAergic LS cells robustly code for space during free exploration in an open field and 43.90% of cells during linear track alternation. One of the major features of the linear track is the induction of directional place fields in LS, which may underlie some of the differences we observed in 1D versus 2D spatial coding characteristics.
In addition to spatial information, we found that the LS also reliably encodes velocity, acceleration, and directional information, suggesting that the LS encodes more complex navigational features than previously thought. Neuronal representations in the hippocampus have been found to change over time, with place code ensembles changing rapidly: an estimated ∼40% [34] to ∼75-85% [38] of CA1 neurons were found to remap over days. The stability of the LS spatial code has received little attention to date. In the current study, we found a subpopulation of cells in the posterior LS that display stable place fields over eight days. The functional role of such stability remains to be elucidated. One possibility is that this ensemble may mediate the stable encoding of contexts over longer durations of time, which could account for the critical importance of the LS for a number of spatial behaviors, including contextual fear conditioning [16,41–44], context-reward associations [39, 40] and radial-maze navigation tasks [45, 46].
Similar to the change in properties of hippocampal place cells along the dorsal-ventral axis [70, 71], previous work has hinted at a dorsal to ventral organization of the LS place code: electrophysiology studies recorded a slightly larger proportion of LS place cells in the more dorsal regions [30]. A complementary anatomical microstructure has been previously described for the rate-independent phase code in the LS, where the strength of the phase code increased as a function of recording depth along the dorsal-ventral axis [4]. Here we find that lateral septal spatially modulated cells are arranged along an anterior-posterior gradient similar to the gradient of hippocampal inputs into the LS – a finding that helps to reconcile the variation in previously reported estimates of spatial cells in the LS, which ranged between 5.3% of cells recorded along the dorsal-ventral axis [4] up to 56% recorded at the most dorsal level of the LS [20]. This pattern suggests that the inhibitory neurons of the LS may inherit a place code from the hippocampus, with subregions receiving dorsal hippocampal inputs being most similar to the classical hippocampal place cells. Interestingly, this is not the only case of GABAergic neurons inheriting some spatial properties from hippocampal pyramidal cells: interneurons within CA1 have been shown to exhibit significant spatial modulation and comparable information as pyramidal place cells, although with a higher firing rate and greater spatial dispersion [72]. It should be considered that, although the hippocampus seems a likely source of the positional information encoded by the LS, this information could also arise from incoming projections from any other region projecting to LS, such as the entorhinal cortex [5, 25] which could account for part of head direction and velocity tuning.
In addition to spatial modulation of LS cells, we observed a significant number of velocity and acceleration modulated cells in the LS. These cells have been previously described by others on the basis of correlative measures, although estimated proportions vary widely: for velocity tuning in reward-seeking tasks, estimates range from almost absent [4] to almost 60% of cells [20]. For acceleration, approximately 45% of LS cells were found to show some degree of correlation, with almost 30% of recorded cells correlated with both speed and acceleration [20]. Here, using information metrics and a free exploration task in which firing characteristics are not confounded with task parameters, we find that 18.70% of cells were modulated by velocity and 24.63% by acceleration. Cells reliably coding for both velocity and acceleration were relatively rare (2%). Velocity and acceleration cells were distributed dissimilarly in the LS, with velocity cells being more abundant towards the ventral pole, whereas acceleration modulated cells were found to be distributed equally. This raises the question whether velocity and acceleration information are inherited from different areas upstream of the LS. Interestingly, cells coding for direction were more abundant towards the dorsal portion of the LS, but distributed equally along the anterior-posterior axis. Previous studies have identified classical head direction cells in multiple, well-described circuits, including the postsubiculum [73, 74], anterodorsal thalamic nucleus [75], mammillary bodies [76], entorhinal cortex [77], retrosplenial cortex [78, 79] and parasubiculum [80, 81]. A likely source of directional information to the LS is the entorhinal cortex [82]. Our work supports the perspective that LS neurons combine a wide range of modalities and may form a complex representation of context over longer timescales than previously reported. In addition to space, velocity, and reward information, hippocampal pyramidal cells have been shown to encode variables such as time [83, 84], odor [85, 86] and sound frequency [87]. Whether this information is also relayed to the LS remains to be elucidated. Despite this interconnectedness with the LH and previous reports of LS place cells being skewed towards reward in a spatial navigation task [32], we did not observe any overrepresentation of LS spatially modulated cells around food zones during free exploration or around the reward zone in a T-maze non-match-to-place paradigm.
The question arises of what the function could be of having a region downstream of the hippocampus that expresses such seemingly similar coding characteristics. Previous work estimated the convergence of hippocampal efferents onto LS neurons to be 20 to 800 times more dense than to any of its cortical targets [4]. Understanding how individual LS neurons integrate the thousands of synaptic inputs they receive from these hippocampal pyramidal neurons will be critical to understanding how the hippocampal map is processed downstream. Through the process of synaptic integration, a target neuron can fire an action potential upon receiving sufficient temporally coincidental excitatory input on its dendrites. The activation of a LS neuron could thus require multiple hippocampal pyramidal cells to spike simultaneously or in close temporal proximity. Hippocampal pyramidal cells will fire concomitantly when their place fields are in overlapping regions or in close vicinity of another. Due to the high interconnectivity of LS GABAergic neurons [88, 89], the activation of one LS cell could subsequently lead to the inhibition of other neurons in the nucleus [24, 90], thereby reducing noise. Thus, through this process of coincidence detection and recurrent inhibition, the spatial map could converge to be represented by a much lower number of cells. As such, the LS could accurately convey hippocampal information to downstream regions using a much more distributed code, and thereby support more effective information transmission and associative learning [91, 92].
Finally, our tracing work shows LS cells receiving hippocampal inputs from principal cells in dorsal CA1 and CA3 in turn project directly to the LH and the MS. Although the existence of an LS-MS projection has long been debated [5,26,66,68,93], we observed both dense innervation at the level of the MS after transduction of dCA1-input and dCA3-input receiving LS cells, as well as synaptic densities between the LS and MS. This suggests that principal cells of the hippocampus send indirect projections to the MS through LS neurons. This circuit is complementary to the well-described projections of hippocampal GABAergic interneurons directly to the MS which may play a key role in theta rhythm generation [94, 95]. The MS plays a role in generating and propagating theta rhythms throughout the hippocampal formation [96–98], which in turn organizes hippocampal place cell activity [99], as well as a behavioral role in the initiation and velocity of locomotion [100, 101]. One hypothesis is that the information on locomotion-related information reported in specific MS neurons [101] may originate from the LS. Moreover, we observed a direct projection from CA1 and CA3, via the LS, to the hypothalamus. This pathway was previously shown by previously using non-specific monosynaptic anterograde tracing [3], with similar unilateral projection patterns for CA1 and bilaterally for CA3 as described here. Here we show for the first time that the same LS neurons that receive inputs from dorsal CA1 and CA3 project directly to the hypothalamus. The hypothalamus is a highly connected region known for its role in regulating feeding behaviors [102], arousal and motivation [103], and more recently learning and memory [104–107]. Additionally, the LH is crucial for the control of locomotion [108–111] and thought to mediate motivational and goal-directed processes underlying feeding [12,102,112,113].
Together, our findings show that GABAergic cells of the LS may provide the hypothalamus and MS with information about location, direction, and speed, and therefore constitute a core node within a distributed spatial memory network. Within this network, the LS may be necessary for the translation of spatial information to goal-directed or environmentally appropriate actions necessary for survival.
Methods
Ethics statement
All procedures were approved by the McGill University and Douglas Hospital Research Centre Animal Use and Care Committee and in accordance with the Canadian Council on Animal Care.
Animals
Naive mice (8-16 weeks old) were housed individually on a 12-hour light/dark cycle at 22°C and 40% humidity with food and water ad libitum. All experiments were carried out during the light portion of the light/dark cycle. Both male and female mice were used for this study.
Virus injections
Mice, 8-16 weeks old, were anesthetized with isoflurane (5% induction, 0.5-2% maintenance) and placed in a stereotaxic frame (David Kopf Instruments). Body temperature was maintained with a heating pad, and eyes were hydrated with gel (Optixcare). Carprofen (10 ml/kg) was administered subcutaneously. All viral injections were performed using a glass pipette connected to a Nanoject III (Drummond) injector at a flow rate of 1 nL/second. All stereotaxic coordinates are taken from Bregma (in mm). After surgery, animals were continuously monitored until recovery. For three days after surgery, mice were provided with a soft diet supplemented with Carprofen for pain management (MediGel CPF, ∼5mg/kg/day).
Virus-mediated expression of genetically encoded calcium indicator
Adeno-associated virus (AAV) of serotype 9 containing the genetically encoded calcium indicator (GECI) 6 fast (GCaMP6f) under a flex promoter was obtained from the University of Pennsylvania Vector Core (AAV2/9.Syn.Flex.GCaMP6f.WPRE.SV40, CS0641, Penn Vector Core via Addgene). Lateral septum viral injections were targeted at the following coordinates: Anterior LS, AP: 0.86; ML: 0.35; DV: -3.0; Intermediate LS, AP: 0.38; ML: 0.35; DV: -2.7, Posterior LS, AP: 0.10; ML: 0.35; DV: -2.5, with DV + or - .2 to adjust for more dorsal or ventral placement of injection. Lateral septum injections were done in VGAT-IRES-Cre transgenic mice (Slc32a1tm2(cre)Lowl/J, JAX stock #016962). For hippocampal principal cell imaging, C57Bl/6J mice (JAX stock #000664) were injected with a viral vector specifically targeting GCaMP6f expression to CaMKII-positive cells (AAV5.CamKII.GCaMP6f.WPRE.SV40, Penn Vector Core via Addgene, CS1024). GCaMP6f injections were targeted at the following coordinates: CA1, AP: -1.8; ML: 1.5; DV: -1.5; CA3, AP: -2.1; ML: 2.3; DV: -2.2. Three of CA3 animals were Grik4-Cre transgenic mice (C57BL/6-Tg(Grik4-cre)G32-4Stl/J, JAX stock #006474), restricting GCaMP6f expression to the pyramidal cells of CA3 specifically.
Retrograde tracing
Retrograde tracing studies were done using a Cre-dependent retrograde Rabies tracing approach. Briefly, 100 nL of helper virus (AAVdj.hSyn.Flex.TVA.P2A.eGFP.2A.oG, NeuroPhotonics Center, Laval, Canada; [114, 115]) was injected in the LSd (coordinates as above) of VGAT-IRES-Cre transgenic mice (n = 5 animals, Slc32a1tm2(cre)Lowl/J, JAX stock #016962). Cre specificity was tested by the NeuroPhotonics center. After 4 weeks of incubation time to allow for the complete expression of the helper virus, 200 nL of inactivated EnvA pseudotyped rabies virus coupled to mCherry was injected (RABV-EnvA- deltaG-mCherry, NeuroPhotonics Center) at the same coordinates as the helper virus injection, and brains were collected 7 days later.
Anterograde transsynaptic tracing
To distinguish synaptic targets from passing projection fibers of dCA1 and dCA3 within the LS, C57Bl/6J mice (JAX stock #000664) received a bilateral injection of transsynaptic Cre-inducing anterograde virus (AAV1-hSyn-Cre-WPRE-hGH, Penn Vector Core via Addgene, V24784) mixed with a Cre-dependent tdTom-expressing reporter virus to visualize injection sites (AAV2-Ef1a-flex-tdTomato, NeuroPhotonics Center, AAV602) in a 3:1 ratio, 80 nL total injection volume per hemisphere (n = 5 mice per group). Injections were targeted at the following coordinates: CA1: AP: -1.86; ML: 1.5; DV: -1.5; CA3: AP: -2.1; ML: 2.3; DV: - 2.2. Following 1 week of incubation time, animals received a bilateral injection of a Cre- dependent anterograde vector in LS expressing enhanced yellow fluorescent protein (400 nL, bilateral, AP: 0.5; ML: 0.35; DV: -2.6, AAVdj-DIO-Cheta-eYFP, NeuroPhotonics Center, AAV938). Brains were collected and processed three weeks post-injection.
Synaptag tracing
To verify presence of synaptic connections at the level of the LS-MS projection, AAVdj-hSyn-flex-mRuby2-syp-eGFP construct (lot AAV799, NeuroPhotonics Center; Oh et al., 2014) was injected in the LS of CaMKIIα-Cre mice (B6.Cg-Tg(Camk2a-cre)T29-1Stl/J, JAX stock #005359), with 100 nL total injection volume injected at 1 nL/second (unilateral, AP: 0.5; ML: 0.35; DV: -2.7). After 4 weeks of incubation time, animals were perfused, and brains were processed as described below.
GRIN lens implant and baseplate
Two to four weeks post GCaMP6f injection, a 0.5 mm diameter gradient refractive index (GRIN, Inscopix) lens was implanted above LS, dorsal CA1 or CA3, or a 1.8 mm GRIN (Edmund Optics) lens was implanted above dorsal CA1. For .5 mm LS lens implant, the lens was glued to an aluminium baseplate prior to the surgery, which provided two main advantages: 1) it allowed us to implant the baseplate/lens assembly in one surgery (instead of separate surgeries) and 2) during the surgery the miniscope (V3, miniscope.org) was used to image for fluorescence while lowering the lens. The mouse was anesthetized with isoflurane, and the skull was cleared. A ∼.6 mm diameter hole was drilled in the skull at the level of the injection site. An anchor screw was placed above the contralateral cerebellum to stabilize the implant. The baseplate and GRIN combination were attached to the miniscope and together secured to the stereotaxic frame. After making a leading track using a .5 mm diameter needle, the implant was lowered in place. While lowering the GRIN lens into the tissue, an increase in fluorescent signal indicates that the injection site has been reached. Both the GRIN lens and baseplate were permanently secured using C&B-Metabond (Patterson Dental). A plastic cap was used to protect the GRIN lens from scratching during the recovery period and in between recording sessions. The animal was given 4 - 6 weeks of recovery time before experiments started. LS implants were targeted at the following coordinates: Anterior LS, AP: 0.86; ML: 0.35; DV: -2.9; Intermediate LS, AP: 0.38; ML: 0.35; DV: -2.65; Posterior LS, AP: 0.10; ML: 0.35; DV: -2.45; with DV + or - .2 to adjust for dorsal, ventral or intermediate targets; CA3 implants were targeted at AP: -2.1; ML: 2.2; DV: -2.2.
For 1.8 mm GRIN lens implantation, a ∼2 mm diameter cranial window was prepared. Again, an anchor screw was used to secure the implant in place. After removing the dura, a portion of the cortex above the injection site was gently aspirated using a vacuum pump without applying pressure on the underlying hippocampal tissue. The 1.8 mm GRIN lens was lowered at the following coordinates: AP: -1.8; ML: 1.5; DV: -1.8. The GRIN lens was permanently attached to the skull using Metabond, and Kwik-Sil (World Precision Instruments) silicone adhesive was placed on the GRIN to protect it. On average 4 weeks post-implant, the silicone cap was removed and CA1 was imaged using a miniscope mounted with an aluminium base plate while the mouse was under light anesthesia (∼0.5 % isoflurane) to allow the visualization of cell activity. When a satisfying field of view was found (large number of neurons, visible landmarks), the baseplate was cemented above the GRIN lens and a protective cap was used to protect the GRIN from scratches. Imaging would start approximately 1 week after baseplating.
Miniscope recordings
The animals were habituated by being gently handled for ∼5 min per day for 3 days. Animals were then water or chow-scheduled (2h access per day). Miniscope recordings performed at 30 Hz for 15 minutes, with just one recording session per day to minimize photobleaching. For open field recordings, animals were freely foraging for randomly placed 10% sucrose or 10% sweetened condensed milk in water rewards.
Recording environments consisted of a 45 x 45 cm dark grey open field with visual cues, or a 49 x 49 cm white plexiglass open field, which was placed in different recording chambers. For linear track recordings, rewards (30 ul of 10% sucrose water) were placed at each end of the 100 cm linear track, or a 130 cm linear track, and the mouse had to consume one reward before getting the next one delivered. To ensure that all animals received the same amount of exposure, for analysis of spatial and self-motion modulation, only recordings were included from the first time the animal spent time in the open field.
For analysis of modulation by food zones or objects, animals were placed again in the, now familiar, open field on day 1, 3 and 5. On day 2 and 4, food zones or objects were added in opposing corners of the familiar open field without changing any of the existing visual cues. Food zones or objects were visible to the animal. Time spent in the region of interest was computed by taking all timestamps in which animals had their nose within less than 2 cm from the edge of the food zone or the object.
Non-match to place paradigm: Mice were water scheduled (2 h per day) and trained in an automatized T-maze (MazeEngineers) to a non-match to place task. Briefly, each trial was divided into two distinct phases: sample and test. In the initial sample phase, a randomly selected arm was blocked, which forced mice to explore the opposite arm where the animal received a 10% sucrose water reward. After consuming the reward, animals would return to the start box for the test phase, during which both arms could be explored. In the test phase, only the previously unexplored arm is baited, so that mice have to alternate locations between sample and test phases. Mice were subjected to 10 trials (sample + test) per day, for 10 consecutive days, and calcium activity in LS was recorded on all days. Data provided in S11 Fig are from Day 10 of training.
Miniscope and behavior video acquisition
Miniscopes (V3 and V4, miniscope.org) were assembled using open-source plans as described previously (miniscope.org) [116, 117]. Imaging data were acquired using a CMOS imaging sensor (Aptina, MT9V032) and multiplexed through a lightweight coaxial cable. Data was acquired using a data acquisition (DAQ) box connected via a USB host controller (Cypress, CYUSB3013). Animal behavior was recorded using a webcam mounted above the environment. Calcium and behavioral data were recorded using Miniscope custom acquisition software. The DAQ simultaneously acquired behavioral and cellular imaging streams at 30 Hz as uncompressed .avi files (1000 frames per file), and all recorded frames were timestamped in order to perform the subsequent alignment.
Histology
After completion of all experiments, animals were deeply anesthetized with ketamine/xylazine/acepromazide (100, 16, 3 mg/kg, respectively, intraperitoneal injection). Mice were then transcardially perfused with 4% paraformaldehyde in PBS (PFA). Brains were extracted and postfixed in PFA at 4 °C for a minimum of 48 hours. Brains were sectioned at 50 μm using a vibratome and cryoprotected in a solution of 30% ethylene glycol, 30% glycerol and 40% PBS until used.
Immunohistochemistry
Sections were washed 3 x 20 min in PGT (0.45% Gelatin and 0.25% Triton in 1x PBS) at room temperature. Next, primary antibodies (1:1000 goat anti-GFP from Novus Biologicals or 1:10 000 rabbit anti-RFP from VWR (Rockland) were incubated with PGT overnight at 4°C. Following 10, 20, and 40 min washes in PGT, sections were incubated with secondary antibodies (1:1000 donkey anti-goat coupled to A488 or 1:1000 donkey anti-rabbit coupled to A555, both from Molecular Probes) in PGT at room temperature for 1 hour. Sections were subsequently washed for 10 and 20 minutes in PGT and 30 minutes in PBS. Sections were then mounted on glass slides and permanently coverslipped with Fluoromount that contained DAPI.
In-vitro patch clamp electrophysiology
Mice were deeply anesthetized and acute brain slices were obtained following the protective recovery method [118]. Briefly, mice were transcardially perfused with N-methyl-D-glutamine (NMDG-based) solution containing (in mM): 93 NMDG, 93 HCl, 2.5 KCl, 1.2 NaH2PO4, 30 NaHCO3, 20 HEPES, 25 glucose, 5 sodium ascorbate, 2 thiourea, 3 sodium pyruvate, 10 MgSO4 and 0.5 CaCl2, (pH 7.4, oxygenated with carbogen). The brain was then quickly extracted and coronal slices (300 μm) were cut using a vibrating microtome (Leica-VT1000S). Slices were incubated for 10–12 min in 32°C NMDG solution before being transferred to a holding chamber filled with artificial cerebrospinal fluid (aCSF), kept at room temperature and containing (in mM): 124 NaCl, 24 NaHCO3, 2.5 KCl, 1.2 NaH2PO4, 5 HEPES, 12.5 glucose, 2 MgSO4, and 2 CaCl2, (pH 7.4, oxygenated with carbogen). For electrophysiology, slices were transferred to a submerged recording chamber perfused with aCSF (3 ml /min flow rate, 30 °C). Patch-clamp recordings were obtained from LS GABAergic (VGAT-positive) neurons expressing the fluorescent marker GFP and visualized using a 40x water immersion objective on an upright Olympus microscope. Recordings were performed and analysed using an Axon Multiclamp 700B amplifier and the Clampfit10 software (Molecular Devices). The intrapipette solution contained (in mM) 126 K-gluconate, 4 KCl, 10 HEPES, 4 Mg2ATP, 0.3 Na2GTP, and 10 PO-Creatine, adjusted to pH 7.25 with KOH (272 mosm). Pipette resistance was 4-6MΩ.
Calcium imaging analysis
Calcium imaging videos were analyzed using the Miniscope Analysis pipeline (https://github.com/etterguillaume/MiniscopeAnalysis) as described previously [47]. Non-rigid motion correction was applied using NoRMCorre [119], and videos were spatially downsampled (3x) before concatenation. Calcium traces were extracted using CNMFe [120] using the following parameters: gSig = 3 pixels (width of gaussian kernel), gSiz = 20 pixels (approximate neuron diameter), background_model = ’ring’, spatial_algorithm = ’hals’, min_corr = 0.8 (minimum pixel correlation threshold), min_PNR = 8 (minimum peak-to-noise ratio threshold). After extraction, cells and traces were visually inspected, and cells with low signal to noise ratio were removed. Raw calcium traces were filtered to remove high-frequency fluctuations and binarized (normalized amplitude > 2 sd and the first-order derivative > 0) [47].
Mouse position and head orientation was tracked using a DeepLabCut [121, 122] model trained on several mouse pose markers including head centroid (for position) and nose tip (used to compute head direction angle, see below). Velocity was extracted by dividing Δd/Δt where d is the distance and t is time, and subsequently smoothed using a gaussian filter with sigma = 33 ms to remove movement artefacts. Acceleration was computed by differentiating (Matlab function diff) velocity. Head direction was computed as the angle between the vertical axis and the line formed by the head centroid and nose tip. Location data was interpolated to calcium imaging sampling frequency using linear interpolation. Head directions were interpolated using the same method, but direction data was first unwrapped, interpolated and converted back into radians.
Activity and bursting index
Following the binarization of raw calcium traces, we compute the probability of a neuron to be active P(A) using the following formula: The bursting index was computed as the probability of a cell being active, given it was already in an active state, or P(Active, t | Active, t-1). The activity index is computed as the probability of a cell becoming active when it is in an inactive state, or P(Active, t | Inactive, t-1).
Spatial modulation
Before processing, input variables were binned (Position in open field: 17 x 17 bins of 3 x 3 cm; Position in linear track: 34 bins of 3 cm; velocity, min = 2.5 cm/s, max = 30 cm/s, 20 bins; acceleration: min = -2 cm/s2, max = +2 cm/s2, 20 bins; head direction: 40 bins of 9°). Velocity was smoothed over thirty frames to reduce discontinuities. For head direction and positional information, all frames were velocity was < 5cm/s were excluded. The probability of being active P(A) informs on the activity rate of a neuron. Next, we compute the probability of spending time in a given bin i (spatial, velocity, acceleration or head direction bin). Next, we compute the joint probability: And the probability that a cell is active given the animal is in a bin: After plotting the binarized activity in space, we perform n = 1000 random circular permutations. Circular permutations are used to remove the temporal relationship between neuronal activity and behavior, but it preserves the temporal structure of calcium transients and therefore leads to more conservative results (as opposed to complete randomization of every data point, which inflates the significance value of results). Because shuffled surrogates were not systematically normally distributed, we used a non-parametric approach where the p-value corresponds to the number of datapoints from the shuffled distribution that are greater than the actual data for each bin, divided by the number of permutations [123]. A threshold of p < .01 is used to determine the significant data points that make up a significant place field. In our case, for a cell to be considered spatially modulated, at least one of the spatial bins must contain activity that is significantly non-random. To compute within-session stability, the halfway point in the recording was determined by taking the timestamps where half of the total number of binarized events for the session was reached. Tuning maps were computed for both the first and second half of the session using the methodology described above. A Gaussian smoothing filter (σ = 3) was used to smooth the tuning maps before computing a correlation between the two. A within-session correlation value of >.5 was considered a stable field.
Calculating mutual information content
Mutual information is used to describe the amount of information about one variable (spatial location, head direction, velocity or acceleration) through the observation of neuronal activity, and was calculated using the following formula: Where M is the total number of possible behavioral states, P(Si∩Aj) is the joint probability of the animal being in bin i concurrently with activity level j. As we are using binarized activity, j can only be active or inactive. MI was calculated using 30 bootstrapped surrogates using 50 percent of the data (randomly chosen data points) with replacement, to allow for computation of mean MI and SEM. Next, the trace was circularly shuffled in order to calculate a shuffled MI and SEM. Cells were then sorted along with the magnitude of the difference between mean MI and mean shuffled MI, and a two-way ANOVA with a p < .01 threshold was used to determine whether the difference between the two numbers was significant in order for the cell to be deemed significantly spatially modulated. In case the assumptions for parametric ANOVA are not met (normal distribution, variance homogeneity), the significance value was computed by taking the sum of the shuffled MI values of greater magnitude than actual MI values, divided by the number of shuffles.
Bayesian decoding
A Bayesian decoder was used to evaluate how well LS neural activity estimated the animal’s location as compared to neural activity recorded from dorsal CA1 and CA3. We used decoding methods specifically adapted for calcium imaging data with binarized calcium transients [47]. Using only epochs with velocity > 5 cm/s, a training dataset was generated using 50% of the data. The remaining 50% of the data was used for testing. Decoding efficiency (as measured by a decoding error and a decoding agreement score as outlined below) was calculated using 30 bootstrapped surrogates using randomly chosen data points with replacement, to allow for computation of mean MI and SEM. In order to allow for fair comparison between recordings with different numbers of cells recorded, this approach was done using either 40, 60, 80 or 100 randomly chosen cells with replacement for each bootstrapped sample. The posterior probability density function, which is the probability that an animal is in a particular bin Si, given the neural activity A is calculated using the following equation: P(S|A) is the posterior probability distribution of states given neuronal activity. No prior assumption is made about the location of the mouse on the linear track, P(S) is kept uniform to attribute equal probability for each location. Every neuron is assumed to be independent of each other. To construct the tuning curves from multiple neurons, we use the following equation: With P(S|A) a vector of a posteriori behavioral states and N corresponding to the number of neurons used. For every neuron k the tuning curves are constructed, and corresponding posterior location probability can be derived.
To reconstruct the position of the mouse, we will consider the location associated with the maximum a posteriori: With the estimated state among all possible states S. We used a temporal filtering window in the open field of 1.5 seconds and in the linear track of 0.5 seconds to remove erratic jumps in the decoded position of the mouse. Decoding accuracy was measured as 1) decoding error and 2) decoding agreement. In one-dimensional space, decoding error was assessed as In two-dimensional space, decoding error was assessed as Decoding score was computed as Decoding agreement was defined as the portion of time where the exact location or head direction bin was successfully decoded:
Tracking cells across sessions
Neurons were tracked over multiple days using a probabilistic method as previously described [124]. Briefly, spatial footprints were aligned using non-rigid alignment (transformation smoothness = 2) to correct for brain tissue displacements or miniscope placement. After alignment, we considered candidate sets of cells to be the same neuron if their maximal distance was < 12 um, and the spatial correlation between sets of cells was > .65. Next, we visually identified candidate cell pairs across sessions and manually deleted those cell pairs that were erroneously selected (< 1% of cases). We assessed the stability of the spatial representation in LS using simple field correlation, which was assessed by correlating smoothed place fields of cell pairs. To generate the null hypothesis for place fields’ displacements between pairs of days, we used the place fields tuning maps but shuffled cell identities over days (30 shuffles).
Materials availability
This study did not generate new unique reagents.
Data and Code availability
All data used for main and supplemental figures are publicly available at [insert DOI from Gnode]. Code for cell extraction is available at https://github.com/etterguillaume/MiniscopeAnalysis. Code for behavior extraction is available at https://github.com/etterguillaume/PIMPN. Code for extraction of tuning curves and decoding were adapted from [47] and can be downloaded at the following address: https://github.com/etterguillaume/CaImDecoding. All other code is available at https://github.com/suzannevdveldt/spatialcodingLS.
Statistics
Statistics Statistical analyses were performed using Matlab (Mathworks) and GraphPad Prism version 6.00 (GraphPad Software, La Jolla, California USA). All data are presented as mean ± standard error of the mean (SEM) and statistical test details are described in the corresponding results. All t-tests are two-tailed. Normality distribution of each group was assessed using the Shapiro-Wilk normality test and parametric tests were used only when distributions were found normal, using Student’s t-tests, one-way ANOVA or two-way ANOVA. For non-normal distributions, non-parametric tests Mann-Whitney or Kruskal-Wallis tests were performed and described where applicable. Significant main effects or interactions were followed up with appropriate post hoc testing using Bonferroni corrections were applicable. p < 0.05 was considered statistically significant, or p < .01 and described where applicable *, p < 0.05; **, p < 0.01; ***, p < 0.001, ****, p < 0.0001.
Funding
This work was supported by funding from the Canadian Institutes for Health Research (CIHR), Fonds de la Recherche en Santé du Québec (FRSQ), and the Natural Sciences and Engineering Research Council of Canada (NSERC). SVDV is supported by a Vanier Canada Graduate Scholarship and the Richard H. Tomlinson Doctoral Fellowship. C-AM is supported by a FRQS fellowship.
Competing interests
The authors declare no competing interests.
Supplemental figures
Acknowledgments
We thank Fernanda Sosa for help with behavioral recording experiments, training of the animals, immunohistological experiments and quantifications. Alexandra T. Keinath for valuable comments on a previous version of the manuscript, and Ke Cui for help with colony maintenance and mice perfusions. We also thank Daniel Aharoni (UCLA) for guidance in using the UCLA miniscope and Bruno Rivard for help building the miniscopes. The present study used the services of the Molecular and Cellular Microscopy Platform at the DHRC.
Abbreviations
- LH
- lateral hypothalamus LS lateral septum
- MS
- medial septum
- VTA
- ventral tegmental area
References
- 1.↵
- 2.↵
- 3.↵
- 4.↵
- 5.↵
- 6.↵
- 7.
- 8.
- 9.
- 10.↵
- 11.↵
- 12.↵
- 13.
- 14.↵
- 15.↵
- 16.↵
- 17.
- 18.↵
- 19.↵
- 20.↵
- 21.↵
- 22.
- 23.
- 24.↵
- 25.↵
- 26.↵
- 27.↵
- 28.↵
- 29.
- 30.↵
- 31.↵
- 32.↵
- 33.↵
- 34.↵
- 35.
- 36.
- 37.
- 38.↵
- 39.↵
- 40.↵
- 41.↵
- 42.
- 43.
- 44.↵
- 45.↵
- 46.↵
- 47.↵
- 48.↵
- 49.
- 50.↵
- 51.↵
- 52.↵
- 53.
- 54.↵
- 55.↵
- 56.↵
- 57.↵
- 58.↵
- 59.↵
- 60.
- 61.↵
- 62.
- 63.
- 64.↵
- 65.↵
- 66.↵
- 67.
- 68.↵
- 69.↵
- 70.↵
- 71.↵
- 72.↵
- 73.↵
- 74.↵
- 75.↵
- 76.↵
- 77.↵
- 78.↵
- 79.↵
- 80.↵
- 81.↵
- 82.↵
- 83.↵
- 84.↵
- 85.↵
- 86.↵
- 87.↵
- 88.↵
- 89.↵
- 90.↵
- 91.↵
- 92.↵
- 93.↵
- 94.↵
- 95.↵
- 96.↵
- 97.
- 98.↵
- 99.↵
- 100.↵
- 101.↵
- 102.↵
- 103.↵
- 104.↵
- 105.
- 106.
- 107.↵
- 108.↵
- 109.
- 110.
- 111.↵
- 112.↵
- 113.↵
- 114.↵
- 115.↵
- 116.↵
- 117.↵
- 118.↵
- 119.↵
- 120.↵
- 121.↵
- 122.↵
- 123.↵
- 124.↵