A further comparison between the template search method and the JPM method was carried out by splitting the recorded session into two halves. I tested whether the spike sequence identified by the template method in the first half of the session can predict the excessive occurrence of the same triplet in the second half as identified by the JPM method. For this test, a wheel running session was selected and I assumed that the behavior of the rat and the environmental effects on the animal were similar throughout the recording time (continuous 20 min). The first half was scanned by the template matching algorithm. The most frequently repeating sequence pattern was a spike triplet of neurons 2, 4 and 0 (2,4,0;30-40, 120-130; Fig. 3.8a). A JPM was constructed from the second half of the recording session. The significant pixels, detected at 40 msec (x = cell 4-2) and 125 m25 msec (y = cell 2-0), matched the spike detected from the first half of the session (Fig. 3.8b). In addition, significant pixels detected from the first and second halves of the recording sessions were superimposed (Fig. 3.8c). The overlap of the significant pixels, identified from the first and second halves of the recording session, also indicated that spike sequences occurred relatively homogeneously throughout the recording session rather than confined to certain unusual events during the recording.

The majority of spike sequences were either shorter than 50 msec or longer than 100 msec. In general, short sequences dominated in slow wave sleep whereas the longer sequences occurred in the awake animal or during REM sleep (e. g., Fig. 3.2). To quantify this observation, I examined the EEG correlates of short and long duration sequences of the same neurons. Two epochs were extracted from the EEG including the first spike of the sequence as time zero. The shorter epoch (204.8 msec) provided a more precise estimate of thete of the exact EEG state, whereas the longer one (-819.2 msec to 2,457.6 msec) was used to assess the EEG power at lower frequencies. Power spectra, calculated from the short and long EEG epochs (0 to 300 Hz and 0 to 20 Hz, respectively), revealed that short spike sequences were associated with a significant peak at 140-200 Hz, corresponding to field "ripples" in the EEG (Buzsáki et al., 1992). Conversely, the long spike sequences were associated with oscillatory fields at theta frequency (Fig. 3.9).

For the statistical analysis of the behavioral effect on sequence generation and persistence, the significance of all possible triplets (n=72) were compared across the three successive behavioral states: sleep phase1 (SLEEP1), wheel running behavior (RUN) and sleep phase2 (SLEEP2), 11 min each. The number of significant triplets were calculated for each variations by using Fisher's exact probability test. All together 72 triplet variation were investigated with usingw=200 msec time window and 5 msec bin size. I compared the non-parametric distributions of the significant triplets between SLEEP1 and RUN and between RUN and SLEEP2. The chi-square test resulted in significant difference between the distribution of SLEEP1 and RUN triplets. In contrast there was no significant difference in the distribution of the significant triplets between RUN and SLEEP2 (Fig. 3.10). The result is consistent with the hypothesis that (1) sequences are altered by experience, (2) sequences which were generated during active behavior tend to persist during subsequent sleep behavior. Also, as consistent with the power spectra analysis, triplets detected during sleep (including slow wave sleep) were compressed at 2-5 times relative to the order-identitical sequences occurred during run.

I have presented evidence for temporal coordination in multi-neuronal spike trains. Significant repetitions of spike configurations were observed in the original spike trains when compared to shuffled surrogates and Poissonian combinatonian combination of cross-correlograms. The consistency of spike sequences within the same behavioral condition has been demonstrated by "split half" test. I have also illustrated the concordance of the different sequence detection methods on the same database. Sequences were present regardless of the method of detection. Finally, I have demonstrated the dependence of sequence preservation on successive behavioral states. Sequence preservation in association with sequence compression occurs when sleep states follow running states.

Temporal coordination of spikes, in relation to the stimulus presentation, has been described in the locust's antennal lobe (Laurent, 1996), in the crab somatogastric ganglion (Marder & Calabrese, 1996), in the crayfish (Dayhoff and Gerstein, 1983a-b) and in various neocortical areas of cats (Frostig et al., 1990; Baranyi & Feher, 1981; Kirkwood and Sears, 1991; Singer & Gray, 1995; Gray et al., 1992; Ts’o et al., 1986; Eckhorn, 1991) and monkeys (Abeles, 1993; Abeles et al., 1993; Aertsen et al., Vaadia & Abeles, 1987; Strehler and Lestienne, 1986; Riehle et al., 1997; Mechler et al., 1998).

Repeating spike sequence patterns mostly have been reported in stimulus presentation paradigms (Laurent, 1996, Abeles, 1993, Dayhoff and Gerstein et al., 1983; Riehle, 1997). To date, spontaneous occurrences of invariant spatio-temporal patterns (sequences) have only been detected in basal forebrain areas associated with cardiac and respiratory activity (Frostig,1990; Riehle, 1997). Spike sequences may recur by chance or can be generated by causal mechanism oof neuronal interactions. Separation between these two possibilities is complicated by the non-stationary nature of neuronal spike trains (Abeles and Gerstein, 1988; Softky and Koch, 1992). For example, periodic modulation of spike frequency in cell assemblies can lead to the emergence of repeating spatio-temporal discharges of the participating neurons. As a result, statistical tests based on the stationarity of the events may reveal spurious sequences. In the experiments described above, I examined whether recurring spike sequences are present in the awake and sleeping animal. Since hippocampal pyramidal neurons are considered to be "place" cells (O’Keefe and Nadel, 1978), it is expected that neurons are activated sequentially when the animal visits different locations or performs a task in certain temporal order in a structured environment (Wilson and McNaughton, 1994). During sleep, on the other hand, there is no external reference or motor behavior to drive hippocampal cells. Therefore, if recurring spike sequences are present during sleep, they are likely to be generated internally. Simultaneously recorded pyramidal cell assemblies were recorded in rats, trained to run in a wheel, and in sleeping animals. Spatio-temporal sequences of spike patterns were detected by either a "template matching" method or by the joint probabilities of spikes. The significance of sequence repetition was tested by various Monte Carlo shuffling methods and by calculating the chance probabilities of spike sequences from spike doublets. The results indicate that repeating spike sequences are present in both the awake and sleeping animal in excess of what might be predicted by random variations.

A critical issue that must be addressed is whether the repeating spike sequences shown here are generated by biological mechanisms or emerged simply as a result of random coincidences of spike trains. Repeating spike patterns have been reported in several cortical and subcortical structures (Lindsey et al., 1997). However, validation of statistical significance of these patterns has been difficult due to a variety of factors. Although Monte Carlo significance tests can reliably examine a null hypothesis, the choice of the proper shuffling protocol is critical. The ideal shuffling protocol should maintain the discharge frequency of individual spike train constant and should not alter the population dynamics of the parallel-recorded neurons.

The number of spuriously occurring spike sequence patterns increases rapidly with discharge frequency (Abeles and Gerstein, 1988). Therefore, proper comparison of the original and shuffled spike trains requires that the discharge frequency of neurons should not be altered significantly by tby the shuffling procedure. If the dynamics of cortical neurons could be described by a Poisson process (Bair and Koch, 1996; Shadlen and Newsome, 1998), within-spike train randomization of spike occurrences would be appropriate since this procedure does not alter the average firing rate of the individual neurons. Unfortunately, random shuffling within the same spike train may alter the correlation structure of the parallel spike trains. This issue is very important in the hippocampal formation, since firing patterns of pyramidal cells are dynamically modulated by behavior. In the exploring animal population synchrony is periodically modulated at theta frequency. In the absence of rhythmic theta waves, irregularly occurring sharp wave bursts are present during which population synchrony of neurons increases 4 to 8-fold (Csicsvari et al., 1998). In order to retain population behavior and associated firing rate changes, three additional shuffling protocols were used. First, spikes were randomly shuffled across spike trains (Fig. 3.2d). Although this method preserves population dynamics perfectly, it tends to decrease the firing rate differences of individual neurons in the original spike trains. In the second (Fig. 3.2b), third (Fig. 3.2c) and forth (Fig. 3.5a) methods, shuffling was carried out within the same spike trains. In the forth method, shifting spikes was always phase-locked to phase of theta activity and all neighboring spikes within a theta cycle were shifted together. This shuffling procedure retained spike dynamics both within and across spike trains. However, the method can be applied only to relatively stationary data sets (e. g., continuous theta epochs). The third method (Fig. 3.2c) shifted spikes randomly within 50 msec with the goal of retaining the population synchrony across spike trains during theta waves and sharp waves. Independent of the shuffling method used, significant repetition of spike sequences were found in each animal.

The joint probability map was a complementary method for the extraction of repeating spike sequences in parallel-spike trains. The joint probability map can be considered as a multineuronal extension of the cross-correlation method. It searches for precisely repeating spike triplets and provides information that is not available by using a cross-correlation method (see Chapter 2). Sequences that emerge as random occurrences of spike doublets are eliminated and the excessive number of triplets are determined Fischer’s exact probability statistics. The window method and the joint probability method detected similar spike sequences. Importantly, the sequences detected by one method in the first part of the recording session reliably predicted the sequences detected by the other method in the remaining part of the session.

In summary, repeating spike sequences were found more frequently in our parallel-recorded spike trains than expected by chance, independent of the sequence detection method or the shuffling procedure. It should be emphasized, however, that each method used in this study has shortcomings, as discussed above. Therefore, the conclusions based on these methods should remain tentative until more reliable mathematical tools become available for the analysis of the temporal relationship among coactive neurons. Nevertheless, it should also be emphasized that the number of different sequences, the number of repeating spike sequences and the number of spikes within a given sequence (complexity) varied even within the same data set depending on the neuron which served as a sequence-initiator. These observations, together with the behavioral modification of spike sequences, indicate that the observed spike sequences could not be fully accounted for by random coincidences of neuronal discharges of hippocampal cells.

Spike sequences were observed in both the awake and sleeping animal. The structured discharges during sleep could be a consequence of some hard-wiring (Deadwyler et al., 1996) or may reflect synaptiynaptic changes as a result of learning in the awake animal. We have hypothesized earlier that the behavior-dependent electrical changes in the hippocampal formation (theta and sharp SPW-associated states) might subserve a two-stage process of information storage (Buzsáki, 1989). SPWs bursts, which invade the CA1 region and deep layers of the entorhinal cortex, are hypothesized to transfer the stored representations to neocortical networks. Explicit predictions of this model are that neuronal patterns during SPW bursts are a consequence of learning in the awake brain and that the learned sequences of neuronal discharges are replayed during SPWs in a time-compressed manner. In accordance with the model, Wilson and McNaughton (1994) and Skaggs and McNaughton (1995) found that neuron pairs, which represented similar parts of the environment, and therefore fired together during exploration, showed an increased correlation during the subsequent slow wave sleep episode compared to a preceding sleep episode. Cells that had non-overlapping spatial firing or were inactive during exploration did not show increased correlation. The present finding confirm and extend these observations. Using a similar design, I found that spike sequences, which were activated during behavioral training in the wheel-running task, were preserved in the subsequent slow wave sleep episode. Furthermore, the spike sequences during sleep were compressed in time and were associated with SPW-ripples. What could be the physiological importance of these events? As hypothesized elsewhere, the SPW population bursts in the CA3-CA1-subiculum-entorhinal cortex axis could provide a necessary depolarization force required for synaptic modification of their neocortical targets (Chrobak and Buzsáki, 1994).

Precise temporal coordination of APs and repetition are empirically inseparable properties of the spike trains. Precise coordination in an absence of external trigger event is detectable only on the basis of repetition. Although repetition and temporal coordination are co-occurring properties they may have different biological determinants. First we consider the determinants of AP timing and coordination.

Assuming that the statistical reliability of the excessive spike sequences in the hippocampus will withstand further scrutiny, the important question to ask is whether they serve any important physiological role (Lisman, 1998). Precisely coordinated spike timing has been observed previously in both cortical and subcortical systems (Laurent, 1996; Marder & Calabrese, 1996; Dayhoff and Gerstein, 1983a-b; Frostig et al., 1990; Baranyi & Feher, 1981; Ki981; Kirkwood and Sears, 1991; Singer & Gray, 1995; Gray et al., 1992; Ts’o et al., 1986; Eckhorn, 1991; Abeles, 1993; Abeles et al., 1993; Aertsen et al., Vaadia & Abeles, 1987; Strehler and Lestienne, 1986; Riehle,et al., 1997). In many of these studies, however, the spike sequences are likely to have involved both principal cells and inhibitory interneurons. Separation of principal and interneuron classes is important because (1) it has been suggested that coincident discharge of presynaptic neurons is critical for the emergence of precisely repeating spike sequences (Abeles, 1982; Softky and Koch, 1993), (2) interneurons may function as coincidence detectors (Geiger et al., 1997) and (3) it has been argued that cortical pyramidal cells perform pure algebraic averaging (Shadlen and Newsome, 1995; 1998). The latter authors went further and suggested that in cortical structures, spike times and sequences do not convey any information. This conclusion is based on the assumptions that principal cells are linear devices, the distribution of interspike intervals resemble a random (Poisson) point process and that net excitatory and inhibitory inputs are balanced (Shadlen and Newsome, 1998). I have to note that neither of the latter assumptions are supported by recent empirical data.

Recent studies on individual pyramidal neurons and their network interactions suggest that t pyramidal cells are highly non-linear neurons (Yuste and Denk, 1995; Spruston et al., 1995; Magee and Johnston, 1997; Markram et al., 1997; Kamondi et al., 1998a) and they are equipped with intrinsic oscillatory properties (Llinas, 1988; Huguenard and McCormick, 1992; Penttonen et al., 1998; Kamondi et al., 1998b). Hippocampal pyramidal cells are embedded in an oscillatory network of interneurons (Buzsáki and Chrobak, 1995; Jefferys et al., 1996). The timing of pyramidal cell action potentials is modulated by theta (5-10 Hz), gamma (40-100 Hz) and ultrafast (200 Hz "ripples") rhythms rather than by a random process. Within these oscillatory patterns, the ratio of excitation to inhibition can vary substantially (Rudell and Fox, 1984; Buzsáki et al., 1981). During the time window of the ripple, the relative ratio of excitation to inhibition transiently can increase up to 300 percent (Csicsvari et al., 1998). Furthermore, the occurrence of action potentials on the hyperpolarizing phase of the intracellular theta cycle is much less likely than during the rising phase, most likely due to the slow modulation of a K⁺ current (Kamondi et al., 1998b). Interestingly, place cells in the hippocampus tend to discharge on the theta phase opposite to the discharge of the "background" population (O’Keefe and Recce, 1993; Skaggs et al., 1996). The expected consequence of these intrinsic and network effects is that temporal coordination of presynaptic spikes is an important factor that affects the probability of the postsynaptic response. In short, oscillatory systems provide " temporal windows of opportunity" to selectively suppress or enhance the effectiveness of presynaptic activity. As a result, members of a given spike sequence, as shown here, could exert a differential impact on their postsynaptic targets.

Spike sequences can be generated due to nonlinear interactions between the input and the actual state of the cell at a cellular level. As a result, EPSPs occuring within a certain interval are amplified relative to those which do not occur in that time range. The consequence is that the output is not the actual arithmetic sum of the input but rather a function of the temporal pattern of the input. The network-level effect of this type of temporal filtering is that APs are precisely scheduled relative to the ongoing network activity.

A network-level coordination of spikes may occur due to a reverberation of the APs or a consistent external drive. The reverberation hypothesis was first proposed by Hebb (1949). According to the this hypothesis APs propagate in a circuit of interconnected cells as in a closed-loop. In a sparse-connectivity network, many closed-lood-loops can be formed on a random basis. Recently, sparsely connected networks became a subject of intensive study (e.g. Amit, 1989). Attractor dynamics are likely to occur within sparsely connected network under certain conditions (Amit, 1989). The network dynamics will converge to a sequence of steady states i.e. to the limit cycle attractor which best supports the coincident input activity to specific cells. More precisely, only those feed-back loops will be stabilized which converge back to the already active cells. According to computer simulations, the conditions necessary for repetition to occur are as follows: (1) spontaneous activity of some cells, (2) fixed refractory periods and (3) feed-back loops in the network (Nadasdy, 1998). The anatomy of the hippocampus CA3 recurrent collateral system is consistent with a sparsely connected network with feed-back loops. In a recurrent collateral system, no rhythmic external stimulation is necessary to generate and maintain the AP sequences. The maintenance is provided by the spontaneous pacemaker activity of certain tonically active cells. Alternatively, interneurons may support a rhythmic hyperpolarization of the pyramidal cells making them coherently exit from hyperpolarized states (Buzsaki, Chrobak, 1995). A small fraction of cells with intrinsic oscillatory properties (<10%) is sufficient to maintain the spatio-temporal pattern in the network (unpublished observation). Since there are many steady states of oscillations are supported by a sparsely connected random network the external input can switch between these states. Environmental stimulation thus can serve as a perturbation which forces the network to settle down in one of several oscillatory states.

Alternatively, in the absence of intrinsic pacemaker activity, a congruent external drive is capable generating AP sequences in a sparsely connected network. During a consistent stereotypical behavior state, the spike emission may be coordinated and the pattern of spikes may repeat as many times as the same sensory cues occur. When the animal is taking the same path in a maze or running in a wheel, sensory input such as the visual flow follows the same temporal order. As a result, the sensory drive to the hippocampal cells is also consistently repeated. The repetitive activation of the pyramidal cells within the same context of environmental stimuli may cause cells to fire in the same temporal relationship. This explanation can be tested within a behavioral paradigm, where the running path of the animal is continuously monitored and the locomotion could be correlated with the emergence of spike sequences.

The stThe statistical significance of sequence repetition does not necessarily imply functional significance. Sequences can be epiphenomena that are products of cellular interactions but never are utilized by the nervous system. However, precise temporal resolution and comparison of spikes on a sub-millisecond time scale have been described in relationship to sound localization (Konishi et al., 1988). We know from studies on the barn owl auditory system that sub-millisecond ranges of relative delays between APs can precisely be discriminated by neurons which have a much longer time constant. The basis for the fine temporal discrimination is a set of neuronal delay circuits. The inter Aural delays are converted to a topographic map of the cells which receive synchronous inputs from the both ears. The active cell then presumable indicates the relative direction of the sound source in three dimensional space (Konishi et al., 1988).

Whether hippocampal pyramidal cells make use of any available temporal coding or not we do not know. The behavioral modulation (induction, persistence and compression) suggests that spike sequences may represent a type of population codes that efficiently encodes different information by different temporal patterns that are superimposed on the same population of cells.

Functionally the hippocampus can be consideridered as a "filter" which extracts relevant information from the sensory space for permanent storage in the neocortex. Any kind of population-level temporal organization of spikes that is (1) specific for the new experiences and (2) persistent across active behavior and sleep is a candidate for memory trace. If sequences can be correlated with sensory configurations or with the presence of objects in the environment, like place fields are, then the question of memory representation by sequences can be addressed more precisely. Until then, we can investigate the sequence preservation across different behavioral states. This can be accomplished by comparing the statistical significance of the spike sequences across unrelated behavioral states (such as sleep and run) and related behavioral states (such as run and sleep). As I have demonstrated the spectral distribution of significant sequences detected during run is independent from the spectral distribution of sequences detected during preceding sleep state. In contrast the spectral distribution of sequences during continuous running behavior was not independent from the sequences observed during the subsequent sleep state. This finding suggests that the information represented by the sequences during active exploration are transferred to the sleep phase for further consolidation. Independent of this analysis, I found that sequences are repeated in a compressed manner during sharp wave episodes, which are the dominant field events during slow wave sleep. Based on these two findings, I assume that those sequences which were generated during active exploratory behavior are transferred in a compressed format to the subsequent sleep state and being replayed during sharp-waves. The functional relevance of sharp-waves is that they are associated with high frequency population discharges which are hypothesized to reactivate synaptic connections which have been pre-potentiated earlier during theta activity (Buzsaki, 1987). As a result sharp-waves induce permanent synaptic modifications in the entorhinal and neo-cortical targets. The frequency range of sharp wave associated discharges (recorded as 200 Hz ripple) is the same as the frequency range of LTP induction. As such, sharp waves are potential candidates for the biological induction of LTP-type of plasticity (Buzsaki, 1987).

Assuming that hippocampal pyramidal cells are linked to different functional assemblies which emit APs in a deterministic sequence, we can ask what is the probability of detecting recurring sequences. If population coding is stochastic and precise timing is an exception rather than a rule, the chance of recording from the fraction of cells which shows this property is rather low. However, the fact that sequences repeated significantly more than by chance suggests that all cells are i indeed engaged in functional assemblies and the probability of observing sequences depends on how many cells we are able to monitor. Since there is no known topography in the distribution of hippocampal cell assemblies, the probability of detecting functionally connected cells is proportional to the number of cells recorded. Since I have shown significant sequence repetition generated by as few as four pyramidal cells I assume that the engagement of individual cells in a larger functional network is rather strong. Those spikes which were not identified as a part of a sequence can be considered to be a part of sequences outside of the scope of our recording.

In the near future, using more multiple electrode arrays we should be able to record hundreds of cells simultaneously. Based on this technique, we will be able to infer the spatio-temporal distribution of the population code over the cell assemblies and the estimate the extent that neurons are associated with each other.