Pattern separation of spiketrains in hippocampal neurons

Pattern separation is a process that minimizes overlap between patterns of neuronal activity representing similar experiences. Theoretical work suggests that the dentate gyrus (DG) performs this role for memory processing but a direct demonstration is lacking. One limitation is the difficulty to measure DG inputs and outputs simultaneously. To rigorously assess pattern separation by DG circuitry, we used mouse brain slices to stimulate DG afferents and simultaneously record DG granule cells (GCs) and interneurons. Output spiketrains of GCs are more dissimilar than their input spiketrains, demonstrating for the first time temporal pattern separation at the level of single neurons in the DG. Pattern separation is larger in GCs than in fast-spiking interneurons and hilar mossy cells, and is amplified in CA3 pyramidal cells. Analysis of the neural noise and computational modelling suggest that this form of pattern separation is not explained by simple randomness and arises from specific presynaptic dynamics.


38
How does the brain allow us to discriminate between similar events in our past? This question is 39 a central challenge in the neurobiology of memory and remains elusive. To prevent confusion 40 between memories that share similar features, the brain needs to store distinct activity patterns to when a partial cue common to several patterns is presented, to the reactivation of many patterns 48 and thus to confusion or confabulation. To avoid these interferences, the Hebb-Marr framework 49 proposes that redundancy between input patterns is reduced before they are stored. This process 50 of transforming similar input patterns into less similar output patterns is termed pattern 51 separation (O' Reilly & McClelland, 1994;Santoro, 2013). 52 Theoretical models suggest that the dentate gyrus (DG) performs pattern separation of 53 cortical inputs before sending its differentiated outputs to CA3 (Rolls, 2010;Treves et al., 2008). 54 Indeed, DG is ideally located to do this, receiving signals via the major projection from  (Figure 2A-B). 116 For every recording set, R output was lower than the R input of the associated input set, 117 indicating a decorrelation of the output spiketrains compared to their inputs ( Figure 2C). This 118 was also the case in GCs recorded in slices from adult mice (35 recording sets from 14 neurons) 119 ( Figure 9D). These results are the first direct experimental evidence that single GCs, the output 120 neurons of DG, exhibit pattern separation. The effective decorrelation, defined as the difference 121 between R input and R output , was statistically significant for every input set ( Figure 2D left), but 122 was larger when input spiketrains were highly correlated ( Figure 2D). This is consistent with the 123 role of DG in discriminating between similar memories more than already dissimilar ones 124 (Kesner & Rolls, 2015). Note, however, that the decorrelation normalized to R input is invariant: 125 whatever the input set, the output trains were always decorrelated to ~70% of R input (Figure 2E). 126 Such invariance suggests that the same decorrelating mechanism is used on all input sets. 127 These results constitute the first demonstration that input spiketrains are decorrelated in 128 the DG at the level of single GCs.

130
Temporal pattern separation in other cell types of the DG network 131 Any channel processing inputs and returning outputs, and thus any brain network, performs 132 either pattern separation or pattern convergence to some degree (Santoro, 2013). Thus, GCs are 133 unlikely to be the only neurons to exhibit temporal pattern separation of spiketrains. However, 134 we would expect pattern separation to be at its greatest in GCs, at least among DG cells, because 135 they are the output neurons of the DG and thus should provide the most separated patterns to 136 CA3 before they are stored. To test this hypothesis, we performed the same pattern separation 137 assay while recording from fast-spiking interneurons (FS, 20 recording sets) ( Figure 3A) or hilar 138 mossy cells (HMC, 18 recording sets) ( Figure 3B). 139 At the 10 ms timescale, the distributions of average (R input , R output ) were significantly 140 different between FSs and GCs, with the R output of simultaneously recorded GCs always lower 141 than their corresponding FS ( Figure 3A4). This indicates that FSs perform lower levels of   156 and adding 100 nM of gabazine to the bath, which only slightly decreases the amplitude of Temporal pattern separation at different timescales across celltypes 170 Because the timescales meaningful for the brain remain uncertain, it is important to assess the 171 separation of spiketrains at different timescales. We therefore binned spiketrains using a range of 172 τ w from 5 ms to 250 ms and performed a finer grained analysis using pairwise R input and 173 associated pairwise R output instead of the average across the ten pairs of input trains ( Figure 5A). 174 We discovered that pattern separation levels can dramatically change as a function of τ w . In GCs, 175 the larger the timescale the less they exhibit decorrelation of their input spiketrains, especially at 176 high R input ( Figure 5B). Nonetheless, GCs still exhibit relatively high levels of pattern separation 177 of highly similar input spiketrains, even at long timescales ( Figure 5B). 178 This analysis confirmed that, at short timescales, FSs exhibit less pattern separation than 179 GCs ( Figure 5C) and revealed significant differences between HMCs and GCs, especially for 180 pairs of input spiketrains with low R input (Figure 5D), which were not as obvious in our previous 181 coarse analysis (Figure 3B3). At longer timescales, the variability across neurons has a tendency 182 to increase for all celltypes, but the average levels of decorrelation of both FSs and HMCs stayed 183 highly significantly different from those of GCs (Figure 5C-D). Interestingly, the difference 184 between FSs and GCs was larger at short timescales, whereas the difference between HMCs and 185 GCs increased with larger timescales. Indeed, at τ w above 100 ms, HMCs can often exhibit 186 pattern convergence instead of pattern separation, especially for pairs of already dissimilar input 187 spiketrains (low or negative pairwise R input ), whereas FSs just show weak or no pattern 188 separation.

189
Concerning CA3 PCs, they still exhibit high levels of temporal pattern separation at long 190 timescales and the difference with GCs increases (Figure 5E). Interestingly, in contrast to all 191 other tested celltypes and conditions ( Figure 5B-D), PCs even show a dramatic increase of their 192 average levels of decorrelation at the 250 ms timescale ( Figure 5E). 193 Overall, these results show that among the tested DG celltypes, GCs exhibit the highest 194 levels of temporal pattern separation of cortical spiketrains across all timescales. Our findings 195 also suggest that the high level of separation by the DG is amplified in CA3.

197
Mechanism of temporal pattern separation: neural noise 198 To determine what mechanisms might support temporal pattern separation in GCs, it is necessary 199 to understand its dynamics first. Limiting our analysis to the first presentation of an input set 200 revealed that outputs were already significantly decorrelated ( Figure 6A-B). This shows that the 201 separation mechanism is fast, consistent with the fact that the brain generally does not have the 202 opportunity to average repeated signals and that separation must happen immediately during 203 encoding. In addition, analysis of the last presentation revealed only modestly more separation 204 than for the first one, and only for high input correlations (Figure 6C), suggesting that learning 205 to recognize the input pattern is not critical.

206
Because the mechanism for temporal pattern separation is fast and does not require 207 learning, we asked first whether intrinsic properties of GCs could play a role. Linear regression 208 analysis revealed that the membrane capacitance, resistance, time constant as well as the resting 209 membrane potential are not predictors of decorrelation in GCs (see low R 2 in Table 1). Another 210 hypothesis is that randomness in neuronal responses drives the decorrelation. Indeed, when the 211 same input spiketrain is repeated (e.g. R input = 1) the output spiketrains are not well correlated  (Aimone et al., 2009), which suggests that "neural noise" is a likely contributor 217 to any form of pattern separation. However, because "neural noise" can cover multiple different 218 definitions and phenomena (Faisal et al., 2008), determining its role in a complex computation is 219 not trivial.

220
The noisiness in neural communication is often understood as the unreliability of spiking 221 after a single input spike, and the jitter of the delay between an input spike and an output spike delay and jitter of GCs were slightly higher in our experiments than in a previous report (but 226 recording and analysis methods were different), the variability between GCs was consistent 227 (Mongiat et al., 2009). We then asked whether these spike-wise noise parameters could predict 228 the degree of decorrelation by GCs. First, linear regression analysis shows no good relationship 229 with any parameter, the SP being a mediocre predictor at best ( Figure S2A and Table 2). 230 Moreover, the average firing rate of a GC output set (a measure dependent on SP) is not well 231 correlated with the degree of decorrelation either (Table 3 and Figure S7A). Temporal pattern 232 separation in GCs seems to not be achieved merely because their output spiketrains are a sparser 233 and jittered version of their inputs.

234
To more carefully test the hypothesis that random spiking failures and random delays 235 support fast temporal pattern separation, we produced a shuffled dataset and a simulated dataset 236 only governed by spike-wise noise statistics comparable to the original data ( Figure S1, S2B, 237 Table 2 and Methods -Computational models). The distribution of (R input , R output ) was 238 significantly higher in the original data (Figure 7 and S4), showing that simple random 239 processes yield greater levels of separation than real GCs, especially for highly similar inputs 240 (Figure 7E and S4A). 241 In addition to the spike-wise noise, we considered neural noise at the level of spiketrains 242 by computing the average correlation coefficient R w between "children" output spiketrains from 243 the same "parent" input train ( Figure 7A). R w characterizes the more complex notion of 244 spiketrain reliability, that is the ability of a neuron to reproduce the same output spiketrain in 245 response to repetitions of the same input spiketrain. R w is not dependent on intrinsic cellular 246 properties ( Table 1) and only moderately determined by spike-wise noise parameters (Table 2), 247 suggesting that the rather low R w of GCs is the expression of more complex noisy biophysical 248 processes. Consistently, R w was significantly lower for shuffled and simulated data than in real 249 GCs ( Figure 7D-E and S4B). This indicates that the output spiketrains of GCs are more reliable 250 than if their output was entirely determined by simple random processes. Overall, the lower 251 R output and R w distributions of random datasets compared to GCs (Figure 7) clearly show that 252 simple noise cannot fully underlie the operations performed by GCs on input spiketrains.

253
The analysis above also suggests that there might be an unavoidable trade-off between 254 achieving pattern separation and reliable information transmission about input spiketrains. To 255 further investigate this, we looked at the relationship between the spiketrain reliability R w and the 256 decorrelation levels in GC recordings and found a strong anticorrelation ( Figure 8A and Table   257 3). This linear relationship is clear evidence that a biological process leading to sweep-to-sweep 258 variability is a powerful mechanism for temporal pattern separation in DG.

259
However, spiketrain unreliability is not the only reason GCs output spiketrains are 260 decorrelated. Indeed, three lines of evidence supports the idea that even if GCs were perfectly 261 reliable from sweep to sweep, their output spiketrains would still be less similar than their inputs.

262
First, when averaging out the variability between spiketrains associated to the same input (i.e. by 263 computing the PSTH), a significant level of decorrelation is still detected (Figure 8B-D). 264 Second, the linear model describing the relationship between R w and decorrelation levels suggest 265 that even at R w = 1, ~10% of decorrelation would still be achieved ( Figure 8A). Finally, high 266 levels of pattern separation are performed upon the first presentation of an input set ( Figure 6). 267 We can thus conclude that temporal pattern separation is supported by mechanisms in addition to 268 spiketrain unreliability.

269
Taken together, these results suggest that complex but noisy biophysical mechanisms 270 allow GCs to balance temporal pattern separation and reliable signaling about their inputs.

272
Mechanism of temporal pattern separation: short-term synaptic dynamics 273 The fact that intrinsic properties of GCs do not predict their decorrelation levels ( Table   274 1) suggests that temporal pattern separation comes from the DG network in which each recorded 275 GC is embedded. We thus hypothesize that the specific short-term dynamics of synapses in the 276 DG network implement temporal pattern separation in GCs. This is a likely mechanism because:

296
To further understand the mechanisms behind temporal pattern separation in single hippocampal 297 cells, we investigated the sources of the difference in decorrelation levels between the tested 298 celltypes. We considered differences in intrinsic cellular properties (Table 1), firing rate and 299 probability of bursting ( Figure S6-7), spike-wise noise (Figure S2-3) and spiketrain reliability 300 ( Figure 8A and S5). 301 Both FS and HMC recordings displayed bursts of spikes (defined as more than one output 302 spike between two input spikes), which was very rarely seen in our GC and CA3 recordings 303 (Figure 3-4, S6C-E, and S7C, E). As a result, both FSs and HMCs had significantly higher 304 firing rates than GCs (Figure S6A-B) (although the effect size was smaller for HMCs, because 305 they burst less often than FSs and they have generally less spikes per bursts. Figure S6D-E). 306 Then, are the firing rate or the probability of bursting predictive of differences in pattern 307 separation? The relationships are unclear, but it seems that both could partially explain the lower 308 decorrelations observed in some HMCs and FSs (Figure S7A-B). To more directly test the role 309 of bursting, we processed all FS and HMC recordings by removing all but the first spike in each separation ability between FSs and GCs, although, to our knowledge, they have never been 321 formally compared. We thus first confirmed the idea that FS show much less spike-wise noise 322 than GCs ( Figure S3). Then, linear regressions revealed that spiking probability (SP) is a good 323 predictor of both spiketrain unreliability (1 -R w ) and decorrelation performance in FSs ( Figure   324 S2B and Table 2). Surprisingly, the membrane resistance of FSs was also a good predictor 325 (Table 1). Thus, contrarily to GCs, FS pattern separation behavior is strongly and linearly 326 determined by some intrinsic and spike-wise properties, even though it is in principle hazardous 327 to anticipate complex neuronal operations from such low-level characteristics, as our previous 328 analysis on GCs illustrated. Indeed, spike-wise noise parameters of HMCs were very close to 329 those of GCs ( Figure S3) and they showed striking pattern separation differences nonetheless.

330
Across all celltypes, it is clear that higher SP correlates with lower decorrelation levels 331 ( Figure S2D). This suggests that sparseness can be a mechanism partially supporting temporal 332 pattern separation. Only partially, because the correlation is mediocre (R 2 = 57%), and although 333 it is a better predictor in FSs, FSs decorrelation levels are consistently lower than expected by the 334 linear model fitting all celltypes ( Figure S2D). 335 In the end, the best (but partial) predictor of decorrelation levels across all celltypes and 336 conditions is R w , the spiketrain reliability ( Figure 8A   We report that similar cortical input spiketrains are transformed in the DG network, leading to 346 less similar output spiketrains in GCs. Our findings provide the first experimental demonstration 347 that a form of pattern separation is performed within the DG itself and exhibited at the level of 348 single neurons at different timescales. This computation arises from noisy but specific 349 biophysical processes (e.g. synaptic dynamics) in the DG network, where interneurons do not 350 exhibit as much temporal pattern separation as the final DG output. In turn, the CA3 network 351 seems to amplify this separation even more at the level of single PCs, suggesting that, at least in 352 the hippocampus, it is not a computation specific to the DG.  ). Yet, our study was the first to investigate pattern separation 396 at a range of timescales. We found that temporal pattern separation in the DG output was best at 397 short timescales. This relationship was generally conserved across celltypes, with HMCs even 398 achieving the opposite of pattern separation, pattern convergence, at timescales above 100 ms. 399 Note, however, that temporal pattern separation is not necessarily a monotonically decreasing 400 function of the time resolution, as CA3 PCs exhibited a surprising sharp increase of temporal 401 pattern separation for low input similarity at 250 ms ( Figure 5). 402 Our study of pattern separation is the first to focus on temporal patterns, as opposed to 403 spatial ones; but neural activity patterns are spatiotemporal. More work is clearly needed to test 404 whether the DG is a pattern separator at the spatiotemporal, population level, but the discovery  separation, our data show that it hardly benefits from the repetition of input patterns (Figure 6). 424 We also offer indirect evidence that non-adaptive decorrelation processes support temporal 425 pattern separation because output patterns are always decorrelated to the same proportion 426 (Figure 2E), a feat that a simple random process can achieve (Figure S4C), suggesting that  Adaptive or not, what is the biological source of the temporal decorrelation we observed? 434 We first determined that intrinsic membrane properties do not predict decorrelation levels (Table   435 1), and that celltypes ability to fire bursts only moderately affects pattern separation ( Figure S7). 436 Simple randomness was not sufficient to reproduce our results, even though the spiking 437 probability, a form of spike-wise neural noise, plays a partial role (Figure 8 and S2). In the end, 438 it seems that the history-dependent noise of the specific presynaptic dynamics targeting GCs is 439 the strongest candidate to support temporal pattern separation (Figure 9). Indeed, GCs are  The role of sweep-to-sweep variability 464 Because the brain needs to be able to recognize when situations are exactly the same, our finding 465 that pattern separation occurs even when the same input pattern is repeated ( Figure 7A) might 466 seem counter-intuitive at first. However, in theory, the separation and the recognition functions 467 do not have to be supported by the same network. The Hebb-Marr framework actually 468 hypothesizes that the CA3 recurrent, auto-associative network is able to recall the original 469 pattern from a noisy input from DG. Even though most computational models that tested the

474
In the cortex, the well-known variability of single neuron activity between trials is often 475 supposed to be "averaged out" at the population level so that the output of the population is 476 reliable (Faisal et al., 2008). It is thus conceivable that considering an ensemble of GCs would 477 increase the signal-to-noise ratio. In fact, when we average out the sweep-to-sweep variability, 478 GCs exhibit pattern separation for highly similar patterns but almost no separation for identical 479 ones ( Figure 8D). 480 However, this variability, or "noise", is not necessarily meaningless (Faisal et al., 2008).

481
Our results suggest it might be a mechanism amplifying pattern separation (Figure 8). The 482 variability might even be just apparent, if we consider that when the same input is repeated it is 483 at different points in time: each repetition could be considered as a different event that needs to 484 be encoded slightly differently. The DG would thus meaningfully add some noise to transform 485 input spiketrains so that cortical information about an event is stored in the hippocampus with a   Assuming that pattern completion is a process realized by CA3, this implies that, when 503 presented with different but similar partial cues of the same initial memory, the final output of 504 CA3 should converge towards the same representation. Our finding that CA3 PCs exhibit high 505 levels of temporal pattern separation might then come as a surprise (Figure 4-5). Several lines of   (aCSF + 100 nM gabazine) were: V rest = -72.7 ± 2.2 mV; R i = 186 ± 12 MΩ and C m = 36 ± 2 pF.

606
The intrinsic properties of GCs recorded under the same pharmacological conditions were: V rest 607 = -77.3 ± 1.9 mV; R i = 244 ± 13 MΩ and C m = 25 ± 2 pF.  Figure   618 4 (CA3 and GC control), the mean frequency of input trains was set close to 10 Hz (11.9 ± 0.7 619 Hz). This input firing rate was chosen to be consistent with the frequency of EPSCs recorded in  The responses of one or two DG neurons were recorded in whole-cell mode while stimuli 623 were delivered to the outer molecular layer (OML). Stimulus current intensity and location were 624 set so that the recorded neuron spiked occasionally in response to electrical impulses (see range 625 of spike probability in Figure S1-3) and the stimulation electrode was at least 100 µm away 626 from the expected location of the dendrites of the recorded neuron. Once stimulation parameters 627 were set, a pattern separation protocol was run: the five trains of a given input set were delivered 628 one after the other, separated by 5 s of relaxation, and this was repeated ten times. The ten 629 repetitions of the sequence of five patterns were implemented to take into account any potential 630 variability in the output, and the non-random sequential scheme was used to avoid repeating the 631 same input spiketrain close in time. Each protocol yielded a recording set of fifty output 632 spiketrains, each associated with one of the five input trains of an input set ( Figure 1C). A given 633 cell was recorded in response to up to five input sets with different R input (i.e. a recorded cell 634 produced between one and five data points on Figure 2C). 635 The membrane potential baseline was maintained around -70mV during both current- For each recording set, the similarity between pairs of spiketrains was computed as the Pearson's 667 correlation coefficient between the spiketrains rasters binned at a τ w timescale. Sweeps without 668 spikes were excluded from further analysis. 669 We did not use separate protocols to assess the firing rates, probability of bursting and 670 spike-wise noise parameters (spike probability, delay, jitter), but computed them directly for each 671 recording set of spiketrains from a pattern separation experiment. The mean firing rate was 672 computed as the average firing rate across all fifty output spiketrains. A burst was defined as the 673 occurrence of more than one output spike in the interval of two input spikes (see Figure S7). 674 We define the spike-wise neural noise as the probability of spiking at least once after an 675 input spike, the delay of an output spike after an input spike and its average jitter. To assess these  To assess the role of spike-wise neural noise in pattern separation, we generated two data sets.

691
First, we simulated output spiketrains in response to our input sets (for each input set, we 692 simulated ten output sets of fifty synthetic spiketrains). This simulation was entirely based on the 693 average spike-wise noise parameters computed from the original GC recordings (see above): the 694 matrix of input spike times was replicated ten times, and for each of the fifty resulting sweeps, 695 spikes were deleted randomly following a binomial distribution B(Nspk, F), where Nspk is the 696 number of input spikes in a sweep, and F the probability of not spiking (F = 1-mean SP = 1-697 0.42). A random delay, sampled from a Gaussian distribution N(µ,σ), was added to each 698 resulting spike times, with µ and σ being respectively the mean delay and mean jitter in the 699 original recordings. The noise statistics of the resulting simulated data set is shown in Figure   700 S1C. Second, we created a surrogate data set by randomly shuffling the output spikes of the 701 original GC recordings: the delay of each spike was conserved but it was relocated to follow a 702 randomly selected input spike in the same input train (from a uniform distribution). This strategy 703 yielded a data set with noise statistics closer to the original data ( Figure S1D). 704 To test whether probabilistic synaptic dynamics is a potential mechanism of temporal 705 pattern separation, we used an Izhikevich model of a regular spiking neuron (Izhikevich, 2003)

806
In all graphs, τ w = 10 ms. Means and SEM in black.     Figure 2E) for all celltypes and conditions. Notice that, despite the strong anticorrelation, the 947 intercept of the linear model at R W = 1 predicts that even a perfect reliability could still allow 948 10% of decorrelation. See Table 3 for linear regressions on single celltypes.

949
(B) To assess the amount of decorrelation not due to spiketrain unreliability, the ten children    Figure S7). This suggests that a low SP can be a potent mechanism for decorrelation, and that FS  Figure 3A) and HMC interneurons compared to a different set of GC recordings associated with 1030 the same input sets than HMC (bottom) (18 HMC recordings, 22 GC recordings, see Figure 3B).

1031
A U-test was applied to each pair of comparison, showing that FS are much faster and less noisy 1032 than GCs, whereas HMC are similar to GCs (slightly larger jitter, but slightly higher SP).

. Differences in pattern separation between GC and DG
that under this analysis, nbHMCs and GCs are also significantly different at lower timescales (τ w 1101 = 100 and 50 ms) but with a lower effect size as τ w decreases. correspond to columns, and the variables to be explained (y-axis) correspond to rows. Red highlights 1105 significant regressions that explain more than 50% of the variance (R 2 > 50%). Blue highlights 1106 regressions that are significant (p < 0.01) but that explain less than 50% of the variance. The values used 1107 for Normalized Decorrelation, i.e. (R input -R output )/ R input , and for Spiketrain Reliability (Rw) were 1108 computed with a binning window of 10 ms, unless specified. Abbreviations: GC yo = from young mice, 1109 GC ad = from adult mice, ALL = dataset pooling all celltypes and conditions recorded in young mice 1110 1111