Abstract
How information in the nervous system is encoded by patterns of action potentials (i.e. spikes) remains an open question. Multineuron patterns of single spikes are a prime candidate for spike time encoding but their temporal variability requires further characterisation. Here we show how known sources of spike count variability affect stimulusevoked spike time patterns between neurons separated over multiple layers and columns of adult rat somatosensory cortex in vivo. On subsets of trials (clusters) and after controlling for stimulusresponse adaptation, spike time differences between pairs of neurons are “timewarped” (compressed/stretched) by trialtotrial changes in shared excitability, explaining why fixed spike time patterns and noise correlations are seldom reported. We show that predicted cortical state is correlated between groups of 4 neurons, introducing the possibility of spike time pattern modulation by populationwide trialtotrial changes in excitability (i.e. cortical state). Under the assumption of statedependent coding, we propose an improved potential encoding capacity.
Similar content being viewed by others
Introduction
The neural coding problem addresses how information is encoded by, and decoded from, patterns of action potentials (i.e. spikes) in the nervous system^{1,2,3}. Whether information is encoded by spike counts or spike times, remains the subject of debate and study^{4,5,6}. Responses following sensory stimulation can be analysed to determine principles of neural coding. Precise and reliable temporal spike patterns have been observed over repeated trials of single stimuli^{7,8,9,10,11,12} and can improve coding (over spike counts) in primary sensory cortices^{6,13}. Nevertheless, a large proportion of spike time variability remains unexplained^{14}. With the definitive form of spike time encoding undetermined, population spike count (i.e. mean firing rate) encoding is commonly assumed, in which spike count variability averages out over a population^{15,16,17,18}. Spike time encoding, however, offers the possibility of a fast, efficient and higher capacity coding form, that could leverage the known temporal sensitivity of neural integration^{19,20,21} and plasticity^{22}. This paper aims to better characterise shared spike time variability in the somatosensory cortex over repeated trials of single sensory stimuli, as this constrains the form in which spike times encode sensory information.
As single cortical neurons can generate highly reliable spike trains to direct stimulation in vitro^{23}, in vivo response variability largely reflects cortical network dynamics^{24} and synaptic interactions^{25,26}. Analysis of response variability shared between neurons over repeated trials of a single stimulus (“noise correlations”) allows the effect of underlying variability sources to be investigated. For spike counts, shared variability may be studied for pairs of neurons and at the populationlevel^{15} (Fig. 1a, left and centre). The latter includes analysis of latent factors and trialtotrial variability in the lowdimensional trajectory of population activity^{27,28} (Fig. 1a, right). In the visual cortex, shared spike count variability is modulated by shared trialtotrial variability in the excitability of neurons across a population. Particularly, neurons across the population respond with either higher or lower spike counts depending on the population excitability level on a single trial (Fig. 1a, centre). Such shared variability has been explained by an additive interaction of stimulusevoked activity with spontaneous background activity^{15,29,30} and an additional multiplicative interaction^{15}. A population’s state at the start of a trial therefore affects the representation of a stimulus. Analysis of population dynamics in the motor cortex also supports the dependence of neural processing on a population’s state at the start of a trial^{27,28} (Fig. 1a, right). Here we characterise how precise spike time representations are modulated by the shared excitabilitylevel of spatially separated neurons on single trials.
Multineuron first spike patterns following stimulus onset offer the simplest opportunity to study spike time encoding and are supported as a fundamental form of representation^{12,13,31}, which may generalise to multineuron single spike patterns (Fig. 1b). For example, the majority of neurons in the barrel field of the rodent somatosensory cortex (“barrel cortex”) respond to single whisker deflections with 0 or 1 spike^{12} and first spike patterns in primary visual cortex are sufficient to enable visual discrimination^{31}. With regards to the timing characteristics of multineuron single spike patterns, previous analysis of the dataset used here found millisecond precise multineuron spike patterns were organised intra and interlaminar, as well intra and intercolumnar in the barrel cortex^{12}, and that spike time patterns and (weakly noise correlated) spike counts encoded stimulus information (stimulation frequency and location)^{32}. It has also been shown that first spike times can improve decoding of whisker stimulus information by 44% over spike counts^{13}. Precise relative differences between the timings of spikes do not require reference to an onset signal to be decoded by downstream neurons (Fig. 1b) and have been observed experimentally^{8,9}, including for this dataset^{12}. Infrequent reporting of precise relative spike time differences, however, is a key argument against spike time coding. We propose that the temporal variability of spike time patterns requires more detailed characterisation.
The topographic representation of each whisker on the snout of a rodent in the barrel cortex, represents an excellent model to study timing characteristics of multineuron single spike patterns in response to controlled sensory stimulation^{33}. Using a large dataset of spike sorted recordings in anesthetized adult rat barrel cortex (Methods), this paper aims to characterise the nontrivial timing variability of multineuron first spike patterns for groups of neurons separated across layers and columns of the barrel cortex. Firstly, for each stimulus condition and each pair of neurons, the first spike times of the first and second neuron in the pair were plotted against each other for trials of the stimulus condition on which both neurons spiked (Fig. 1c). Within such ‘first spike pair stimulusresponse distributions’, spike time clusters (isolated regions of high response probability) were found to be positively correlated above chancelevels. Such positively correlated spike time clusters represent spike time noise correlations (i.e. comodulation of first spike times by some underlying latent process(es); Fig. 1c).
To the authors’ knowledge, all previous investigations of precise spike time noise correlations have assumed fixed spike time differences, either with a zero^{34} or nonzero^{35} delay, which would correspond with correlation angles of \(45^{\circ }\) (Fig. 1c, upper right) in the stimulusresponse distribution. In contrast, we find that cluster correlation angles are often different from \(45^{\circ }\). This shows that first spike times of neuron pairs covary with respect to some underlying latent process(es), such that the relative firstspike time difference varies with respect to this process(es) (Fig. 1c, lower right). This explains why precise fixed relative spike time patterns are seldom reported^{8,9,12}.
After controlling for neural stimulusresponse adaptation, which causes spike time latencies to increase over rapidly repeated whisker deflections^{26,32,36,37} (Fig. 1d), many clusters remained positively correlated with non\(45^{\circ }\) correlation angles. This cannot be explained by variability in stimulus or cortical onset, which would correspond with a fixed relative difference (\(45^{\circ }\) correlation angle; Fig. 1d). Non\(45^{\circ }\) correlation angles suggest that relative first spike time differences are ‘warped’ (i.e. stretched/compressed) depending on the shared excitabilitylevel of the two neurons on a single trial (Fig. 1d,e) and that the shared excitabilitylevel varies from trialtotrial due to a process other than adaptation. Comodulation of spike times for neurons separated over multiple cortical columns and layers introduces the poss ibility that spike time patterns are modulated by lowdimensional trialtotrial populationwide changes in neuronal excitability not caused by adaptation. In support of this (although not conclusive), predictions of single trial excitability were correlated for a large proportion of pairs of clusters, which did not share a common neuron (Fig. 1f). If decoding neurons were also comodulated by populationwide changes in neuronal excitability, they could integrate spike time patterns in an excitabilitydependent (or ‘statedependent’) manner. An improved potential encoding capacity is predicted under the classical and novel assumptions of ‘fixedstate’ and ‘statedependent’ decoding, where state defines the shared population excitabilitylevel at the start of a trial.
Results
Single whisker deflections (sweeps) were made to single whiskers in blocks of typically 200 trials (Methods). For each block, single deflections were made at a certain frequency (between 0.066–10 Hz), such that the intertrialinterval was 1/f.
Confirmation of single spike patterns following stimulus onset
Each neurontype and layer group demonstrates a clear initial response to single whisker deflections (Fig. 2a). Moreover, the majority of responding neurons spike only once (Fig. 2b), suggesting low spike count encoding capacity. This also supports previous studies^{13} and analyses of the current dataset^{12} in implicating multineuron first spike patterns as a prime encoding form. This paper aims to characterise the spike time variability of such multineuron single spike patterns. Low response reliability (Fig. 2c) complicates analysis of multineuron single spike patterns. For example, a matrix of first spike times (for a single stimulus condition) by neuron and trial index would contain a large number of missing values.
Confirmation of neuronal stimulusresponse adaptation for single neurons
Consistent with previous reports^{26,36,37}, including for the current dataset^{32}, depressive spike count and spike time latency adaptation were confirmed at stimulation frequencies \(\ge 5Hz\) and \(\ge 1Hz\) respectively, with the majority of adaptation occurring in the first 20 stimulus trials (Fig. 2d; Supplementary Fig. 1). Such adaptation was confirmed for all neuron groups, excluding spike time adaptation for the undersampled L2/3 I group (Supplementary Fig. 2a). Such spike time latency adaptation can underlie spike time noise correlations and therefore must be controlled for when testing for noise correlations caused by nonadaptative trialtotrial changes in shared excitability. Controlling for adaptative trends must consider the heterogeneity of responses and adaptation over stimulus trials (Supplementary Fig. 3). Autocorrelation and partial autocorrelation analysis of the the trialtotrial sequence of first spike times of single neurons supported the presence of general adapatative trends without additional autocorrelative structure in the trialtotrial sequence of single neuron first spike times (Supplementary Fig. 4).
Detection of spike time noise correlations
Analysing the joint responses of two neurons allows noise correlations caused by adaptation and nonadaptative trialtotrial changes in shared excitability to be discerned. The first spike times of pairs of neurons were plotted against each other for trials of a single stimulus condition on which both neurons spiked. For example, for repeated 4Hz stimulation of a single whisker, Fig. 3a shows the first spike times of two neurons (L5AI and L4I) in the barrel column corresponding to the whisker (i.e. principal column) for trials on which both neurons spiked. First spike pairs appear to come from one of two clusters (isolated regions of high response probability) in the ‘firstspike pair stimulusresponse distribution’. Moreover, the cluster of green first spike pairs appears to be correlated with a non\(45^{\circ }\) angle , such that the relative difference between the spike times of the two neurons varies with respect to some underlying latent process(es) (e.g. adaptation or nonadaptative trialtotrial changes in shared excitability).
Such correlated clusters with apparent non\(45^{\circ }\) angles were observed in many first spike pair distributions (Supplementary Video 1), representing isolated regions of high response probability, which are noise correlated. The occurrence and angle of spike time noise correlations, and the nature of the underlying latent process(es), were quantified both for (i) isolated clusters and (ii) the original unclustered first spike pair response distributions (Fig. 3b).
For the clustered case, a twostage algorithm was developed to extract isolated clusters which did not contain smaller isolated subclusters. Stage 1 of the algorithm utilised ‘densitybased spatial clustering of applications with noise’ (DBSCAN)^{38,39}, an established clustering algorithm which detects clusters based on the underlying probability density function. Stage 1 applied two additional steps, firstly discarding DBSCAN clusters, which (like the cluster of orange points in Fig. 3a) were likely to be nonisolated subparts of larger clusters, and secondly including points within a flat \(4\sigma\) ellipse calculated for the DBSCAN cluster points so that correlation tests were unbiased (Methods).
Stage 1 extracted extracted 3929 well isolated clusters defined by the \(4\sigma\) boundary of a flat 2D Gaussian ellipse (Methods; Fig. 3a). Stage 1 did not check for repetitions or similar clusterings produced by different values of the DBSCAN epsilon parameter (to avoid bias). Stage 1 clusters were statistically significantly correlated above chance levels (Fig. 3c, left; Fisher’s method: \({\mathcal {X}}^2\)(7858; N = 3929) = 17308:0; p < :001) as were the unclustered response distributions (Fig. 3c, right; Fisher’s method: \({\mathcal {X}}^2\)(8676; N = 4338) = 14863:4; p < :001).
The clustering algorithm found correlated clusters at chancelevels for two surrogate control datasets and was compared to Gaussian Mixture Model approaches (Supplementary Fig. 5; Methods).
An angled ‘Stage 2’ ellipse was calculated to better estimate the distribution underlying each correlated \((p<0.005)\) Stage 1 cluster (Methods; Fig. 3a). Each Stage 2 cluster came from a unique first spike pair response distribution. 247 unique Stage 2 clusters were extracted, 98.0% (242/247) of which remained correlated above chance \((p < 0.005)\).
Unclustered response distributions were also further analysed, and 243 unclustered response distributions were determined as correlated \((p<0.005)\). The detected Stage 2 clusters and correlated unclustered response distributions are shown in (Supplementary Videos 1 and 2 respectively). In the following sections, the same analyses are performed on the correlated Stage 2 clusters and correlated unclustered response distributions. For a subset of first spike pair response distributions, correlations \((p<0.005)\) were detected for both the clustered and unclustered case (Fig. 3d, top).
98.0% (242/247) of Stage 2 clusters and 99.2% (241/243) of correlated unclustered response distributions were positively correlated (Fig. 3d) suggestive of modulation by spike time latency adaptation or trialtotrial changes in pairwise shared excitability.
Linear relationships explained a large proportion of spike time variance within Stage 2 clusters, but not within the response distributions from which Stage 2 clusters were extracted (Fig. 3e). In fact, clustering was necessary to uncover a large proportion of spike time noise correlations detected through clustering (Fig. 3e). This result partly explains why precise spike time noise correlations are rarely reported. Figure 3f confirms that noise correlated clusters consisted of responses from a subset of stimulus trials.
More variance was explained by the linear relationship between the spike time pairs than by linear models fit to predict the joint spike time pairs from trial number, suggesting that clustered and unclustered noise correlations are modulated more by trialtotrial changes in pairwise shared excitability than by adaptation (Fig. 3g; tested further below).
Non\(45^{\circ }\) spike time noise correlation angles
For Stage 2 clusters and correlated unclustered response distributions, relative spike time differences were correlated with the spike time of the first neuron in the pair (Fig. 4a). Correlations are therefore not explained by \(s_1 = s_0 + d + independent\_noise\) where d is a fixed difference and \(s_0\) (the spike time of neuron 0) varies from trialtotrial.
To test whether cluster correlation angles were different from \(45^{\circ }\), the correlation angle \(\theta\) of each correlated Stage 2 cluster and correlated unclustered response distribution was estimated (defined by the bootstrapped angle that the first principal component made with the positive xaxis; Methods). \(\theta _{45} = min(\theta , 90\theta )\) was used, as only the ordering of the neurons determined whether \(\theta\) was above or below \(45^{\circ }\). It is important to note that the Henze–Zirkler null hypothesis of normality was rejected \((p < 0.05)\) for 71.7% (177/247) of Stage 2 clusters and 76.5% (186/243) of correlated unclustered response distributions.
73.7% (182/247) and 47.7% (116/243) of \(\theta _{45}\) angles were between and significantly different from \(0^\circ\) (pvalue < 0.025) and of \(\theta _{45}\) angles were between and significantly different from \(0^{\circ }\) (pvalue < 0.025) and \(45^{\circ }\) (pvalue < 0.025) for Stage 2 clusters and correlated unclustered response distributions respectively (Fig. 4b). This suggests that the relative difference representation of a stimulus is changed by an underlying latent process(es). The clustering algorithm found correlated clusters with correlation angles at \(45^{\circ }\) for two additional surrogate control datasets (Supplementary Fig. 6; Methods). The clustering algorithm enabled estimations of correlation angles with smaller confidence intervals compared to the unclustered response distributions (Fig. 4c).
Stationary correlations
The next sections control for neuronal stimulusresponse adaptation to determine when non\(45^{\circ }\) correlations are due to shared trialtotrial changes in the excitability of neuronal pairs. For Stage 2 clusters and correlated unclustered response distributions, spike times were either not modulated by adaptation, or the cluster spike times of one or both of the neurons were modulated by adaptation (Supplementary Fig. 7). Clusters or response distributions were defined as ‘stationary’ if the spike times of both neurons were not modulated by adaptation, and ‘nonstationary’ otherwise (see Methods stationarity criteria). Figure 4a shows example stationary and nonstationary clusters. The nonstationary cluster shows a clear trial dependence.
40.5% (100/247) of Stage 2 clusters and 43.2% (105/243) of correlated unclustered response distributions were determined as stationary.
Figure 4e shows the proportion of variance explained by the correlations of these stationary Stage 2 clusters and stationary correlated unclustered response distributions.
For stationary Stage 2 clusters, cluster spike times of the first neuron on cluster trials t, were not correlated with the cluster spike times of the second neuron on cluster trials \(t+1\) (i.e. triallag1 crosscorrelation; Supplementary Fig. 8). The same was true for stationary correlated unclustered response distributions. This showed that removal of single trial correlations removed the correlations of the stationary clusters and unclustered response distributions, which would not be the case if general adaptative trends underlied the stationary correlations. Moreover, chancelevel autocorrelations and crosscorrelations for multiple triallags confirmed that the stationary correlations were not affected by adaptative changes or autocorrelative structure (Supplementary Fig. 8), providing an important control test.
77.0% (77/100) of stationary Stage 2 clusters and 43.8% (46/105) of stationary correlated unclustered response distributions had \(\theta _{45}\) angles between and significantly different from \(0^{\circ }\) and \(45^{\circ }\) (Fig. 4f) showing that relative first spike time differences are warped by shared trialtotrial changes in the excitability of neuronal pairs. It is important to note that the Henze–Zirkler null hypothesis of normality was rejected \((p < 0.05)\) for 75.0% (75/100) of stationary Stage 2 clusters and 76.2% (80/105) of stationary correlated unclustered response distributions.
Fixedstate and stateconditioned relative difference response distributions
The paper compares the potential encoding capacity of the characterised clusters under three alternative relative spike time difference coding regimes named Relative Difference Regimes (RDR) 13. It is first necessary to consider how downstream neurons discriminate between stimuli. If two sensory stimuli elicit a behaviourally different response, downstream neurons must be able to respond differently to samples from the response distribution of either stimulus. This is possible if the response distributions of the two stimuli are nonoverlapping in dimensions that downstream neurons are sensitive to, such as spike counts or relative spike time differences (Fig. 5a). The smaller the region of stimulusresponse distribution space that single stimulusresponse distributions occupy, the higher the capacity for encoding the presence of different stimuli reliably.
Two possible relative difference coding regimes are firstly considered under the classical assumption of fixedstatedecoding (i.e. decoding neurons are not modulated by processes of adaptation or shared trialtotrial changes in the excitability of neuronal pairs). Relative Difference Regime 1 (RDR 1) describes the assumption that pairs of neurons encode information through a fixed relative spike time difference. The standard deviation of relative differences over all spike pairs (\(RDR1\_\sigma _{diff}\)) offers a quantification and prediction of the potential encoding capacity for relative differences under RDR 1. The assumption, however, does not capture the quantified structure of the observed response distributions (Fig. 5b; Supplementary Video 1).
Observation of first spike stimulusresponse distributions (Fig. 5b; Supplementary Video 1) shows that clusters correspond to regions of high response probability, which better estimate the underlying stimulusresponse distributions. Relative Difference Regime 2 (RDR 2) describes the notion that correlated clusters may represent trials on which the neuron pair’s relative first spike time difference reliably contributes to the representation of the stimulus. Under the classical assumption of fixedstatedecoders, the standard deviation of relative differences (\(RDR2\_\sigma _{diff}\)) calculated for samples in a stationary Stage 2 cluster quantifies the size of the relative difference response distribution and the potential encoding capacity for relative differences under RDR 2. As stationary clusters are not affected by adaptation, all cluster samples can be used in estimating \(RDR2\_\sigma _{diff}\). The smaller first spike pair relative difference standard deviations (\(\sigma _{diff}\)) under RDR 2 compared to RDR 1 demonstrates an improved potential encoding capacity for RDR 2 (Fig. 5b,c; Supplementary Video 1).
RDR 3 considers that downstream decoding neurons may be modulated by the same shared trialtotrial changes in the excitability that modulate cluster correlations (i.e. statedependentdecoding). For example, a higherlevel of excitability (i.e. continuous cortical state) on a single trial would decrease the latency of both neurons. Spike times of both neurons would then come from a smaller stateconditioned stimulusresponse distribution. If the cluster correlation is different from \(45^{\circ }\), as in Fig. 5b, different state values would correspond with different distributions of relative differences. A downstream neuron modulated by the same trialtotrial changes in excitability (i.e. continuous cortical state) could respond in a statedependent manner. Such (smaller) stateconditioned first spike difference stimulusresponse distributions would provide a higher coding capacity under statedependent decoding.
To first estimate the size of statedependent first spike response distributions for each stationary Stage 2 cluster, the following linear single factor model was fit to each zerocentred stationary Stage 2 cluster: \(s_i = \lambda _i \eta + \sigma _i\) for \(i \in [0, 1]\) (Methods). \(\eta\) represents the value of the continuous underlying state (i.e. shared level of excitability) with most values between \([3, 3]\). \(\lambda _i\) is the gradient at which the spike time of neuron i varies w.r.t. the state \(\eta\), and \(\sigma _i\) is the predicted standard deviation of the spike times \(s_i\) given \(\eta\). Figure 5b illustrates statedependent first spike response distributions for example values of \(\eta = [2.0, 1.0, 0.0, 1.0, 2.0]\).
The estimated size of stateconditioned distributions of decodable relative spike time differences (RDR 3) was then calculated from the factor analysis model as \(\sigma _{diff} = \sqrt{\sigma _{0}^2 + \sigma _{1}^2}\) (see Methods). The estimated standard deviations of stateconditioned distributions of relative differences (\(RDR3\_\sigma _{diff}\)) offers an improved potential coding capacity over RDR 2 (Fig. 5b,c; Supplementary Video 1) and thus also over RDR 1, but would require confirmation of modulation by a populationwide state, which varies from trialtotrial. Figure 5d shows the predicted improvement of \(RDR3\_\sigma _{diff}\) over RDR 2 for the correlated unclustered response distributions.
Nonstationary clusters also modulated by shared trialtotrial changes in the excitability of neuronal pairs
To test whether shared nonadaptative trialtotrial changes in the excitability of neuronal pairs also modulate the remaining nonstationary Stage 2 clusters and nonstationary correlated unclustered response distributions, and to estimate the size of underlying relative difference response distributions at a single point in time under RDR 2 and 3, single neuron adaptative trends of nonstationary clustered and unclustered spike times were removed using time series modelling. Two techniques with different advantages were applied to nonstationary spike times (Fig. 6a). For the clustered case, adaptative trends were not removed from spike times before clustering so as to ensure that clusters and spike time correlations were not introduced by such preprocessing.
Simple parametric models were not used to remove trends because nonstationarities in the trialtotrial sequence of single neuron and cluster spike time did not always follow standard adaptative/nonstationary trends (Supplementary Fig. 83; Methods). In these cases, standard parameteric models would not remove the nonstationarity and would risk introducing correlations were none where present.
Firstly, first order ‘differencing’^{40} was applied to remove adaptative trends of single neuron spike times determined as nonstationary by the stationarity criteria (Fig. 6a; Methods). Particularly, spike times on trials t were subtracted from spike times on trials \(t+1\) (for trials that contributed samples to the cluster or unclustered response distribution). As differencing increases the standard deviation of the time series, the differenced time series were rescaled appropriately (Methods). Negative autocorrelations known and shown here to be introduced by differencing (Supplementary Fig. 9) were ignored, because the resulting spike times allowed direct assessment of correlation angles after trend removal without further processing.
76.2% (112/147) of nonstationary clusters and 80.4% (111/138) of nonstationary correlated unclustered response distributions had differencing applied to at least one neuron. After differencing, 29.5% (33/112) and 33.3% (37/111) of these clusters and unclustered response distributions fulfilled the stationarity and correlation criteria (see Methods), of which 72.7% (24/33) and 54.1% (20/37) had \(\theta _{45}\) angles between and significantly different from \(0^{\circ }\) (pvalue < 0.025) and less than \(45^{\circ }\) (pvalue < 0.025; Fig. 5b). It is important to note that the Henze–Zirkler null hypothesis of normality was rejected \((p < 0.05)\) for 21.2% (7/33) and 62.2% (23/37) of the criteria fulfilling differenced Stage 2 clusters and unclustered response distributions.
The second technique fitted autoregressive (AR) and/or moving average (MA) models appropriately to the differenced time series (i.e. ‘ARIMA’ modelling^{41,42}), in order to remove the autocorrelations present in the time series and/or introduced by differencing, and to return the time series to the original scale without rescaling (Methods). The ’I’ in ARIMA refers to the differencing applied previously in the first step. The AR model component regresses the values of the time series on values of the time series on previous trials, where the time series is a trialtotrial sequence of first spike times. The MA model component regresses the value of a time series on a moving average of the previous values. The order of the AR and MA components determines the number of previous trial first spike times that are used. An order of 0 means that no AR or MA component was applied. Orders were selected automatically (Methods).
96.4% (108/112) and 100.0% (111/111) of these differenced clusters and differenced unclustered response distributions had an AR and/or MA model applied to at least one neuron.
The majority of fitted AR or MA models were of the MA type with a trial lag of 1 (Fig. 6c, Supplementary 93), as expected for removal of the negative autocorrelations introduced by differencing. That is, the algorithm automatically selects to remove the dependence of a spike time on the previous spiking trial’s spike time (a dependence that is introduced by the differencing operation).
37.0% (40/108) and 39.6% (44/111) of ARIMA modelled clusters and unclustered response distributions fulfilled the criteria, of which 72.5% (29/40) and 56.8% (25/44) had \(\theta _{45}\) angles between and significantly different from \(0^{\circ }\) (pvalue < 0.025) and less than \(45^{\circ }\) (pvalue < 0.025; Fig. 6b, Supplementary Fig. 93). The Henze–Zirkler null hypothesis of normality was rejected \((p < 0.05)\) for 21.2% (7/33) and 62.2% (23/37) of the criteria fulfilling ARIMA modelled Stage 2 clusters and unclustered response distributions. Autocorrelations and cross correlations were below chance levels for the criteria fulfilling ARIMA modelled clusters (Supplementary Fig. 10), showing that AR and MA modelling successfully removed the negative autocorrelations introduced by differencing.
These results demonstrated that the spike times and relative differences of a proportion of nonstationary Stage 2 clusters and unclustered response distributions are also modulated by shared nonadaptative trialtotrial changes in the excitability of neuronal pairs. Figure 6d and Supplementary Fig. 93 show the variance explained by the correlations after trend removal (differencing and ARIMA). The angles of the differenced and ARIMA modelled spike times corresponded with the correlation angles before the application of differencing and ARIMA modelling (Supplementary Fig. 11). This supports comodulation by shared nonadaptative trialtotrial changes in the excitability of neuronal pairs and shows that the non\(45^{\circ }\) correlation angles are not artifacts of time series analysis.
The trend removed clusters allowed all cluster samples to be used in the estimation of the underlying cluster and corresponding distribution of relative differences (RDR 2) at a single point in time. Factor analysis was applied to the trend removed clusters to estimate statedependent distributions of relative differences at a single point in time (RDR 3). Again, RDR 2 gave an improved potential encoding capacity over RDR 1, whilst RDR 3 gave an improved potential encoding capacity over both RDR 1 and 2 (Fig. 6c,d).
Frequency dependence and spatial distribution of correlated clusters
Stationary correlations were found at all stimulation frequencies, but mostly at lower frequencies (Fig. 6e, Supplementary Fig. 93), showing that lower frequencies are better suited for stimulus response distribution characterisation as adaptation occurs to a lesser extent. ARIMA modelled clusters were found for all frequencies \(\le 8Hz\) (Fig. 6e, Supplementary Fig. 93), demonstrating that the techniques can characterise noise correlations caused by shared nonadaptative trialtotrial changes in the excitability of neuronal pairs in the presence of adaptation and nonstationarity, which occur both at lower frequencies suited for response distribution characterisation (0.0–0.2 Hz), and behavioural whisking frequencies (1–20 Hz^{43}).
Spike time correlations (stationary and ARIMA trend removed) were found for E–E, I–I and E–I neuron pairs separated over multiple (up to 4) cortical columns and layers (Fig. 7a, Supplementary Fig. 112), demonstrating shared nonadaptative trialtotrial changes in the excitability of neuronal pairs over a large spatial distance. This introduces the possibility that spatially broad populationwide changes in neuronal excitability could modulate decoding neurons and allow them to respond in an statedependent manner. The most correlations were found between Etype neurons in L5A and L5B/6 of the principal and neighbouring columns, although differences in the number of spike sorted neurons extracted by layer and E/I type must be considered (Supplementary Fig. 2). Low numbers of correlated clusters were found for neuron pairs within the same group, which could be a consequence of the local spike overlapping problem (i.e. spike detection difficulties when two or more neurons fire in close temporal and spatial proximity). For 32 neuron pairs, at least one neuron lied in a septum (i.e. between barrels).
Correlation of predicted cortical state
Comodulation of spike times for pairs of neurons separated over multiple cortical columns and layers introduced the possibility that spike time patterns are modulated by lowdimensional nonadaptative trialtotrial populationwide changes in neuronal excitability. For the remaining analyses, Stage 2 clusters and correlated unclustered response distributions, which were stationary or made stationary, were pooled. If a cluster in the pool came from a correlated unclustered response distribution also in the pool, the correlated unclustered response distribution (rather than the cluster) was removed from the pool so that there was only one example per response distribution in the pool.
As an initial test, the value of each spike pair along the cluster’s (or unclustered response distribution’s) first principal component enabled a prediction of populationwide state (i.e. excitability) on a single trial (Fig. 7b). Such predicted single trial states were correlated over trials of single stimulus conditions for a large proportion of pairs of clusters and unclustered response distributions, which did not share a common neuron (Fig. 7c, top). Moreover, all statistically significant correlations \((p<0.05)\) were positive (Fig. 7c, bottom), providing evidence that trialtotrial changes in excitability modulate the spike times of up to 4 neurons. Figure 7d also shows that the first PCA dimension explains \(\sim 70\%\) of explainable spike time variance between groups of 4 neurons (on trials that all 4 neurons spiked). Together, these results strengthen the possibility that spike time patterns are modulated by lowdimensional nonadaptative trialtotrial populationwide changes in neuronal excitability.
Spike time and spike time difference prediction
Over trials of single stimulus conditions, the predicted state of single correlated neuron pairs is correlated with the spike times (Fig. 8a) and spike time differences (Fig. 8b) of other correlated neuron pairs, which do not share a common neuron. Figure 8c (left) demonstrates a reduction in variability when spike times are conditioned on the predicted state of other single correlated neuron pairs. In some cases, such stateconditioned spike time variability was on a similar order to that predicted by the factor analysis model above (Fig. 8c, right). A reduction in spike time difference variability is also seen when non\(45^{\circ }\) correlations are conditioned on the predicted state of other correlated single neuron pairs (Fig. 8d, left), although variability was mostly higher than the variability predicted by the factor analysis model under RDR 3. It is important to note that there is inherent noise in the state predicted from single pairs of neurons, which hampers spike time and spike time difference prediction. Future work may explore improved estimates of single trial state.
Discussion
‘Timewarped multineuron single spike patterns’ describes the novel finding that, after clustering and control for adaptation, pairwise relative spike time differences are ‘warped’ by trialtotrial changes in the excitability of neuron pairs (Figs. 4f, 6b) separated over multiple cortical layers and columns (Fig. 7a). The results show that clustering, non\(45^{\circ }\) noise correlations and adaptation, together, explain a large proportion of spike time variability and should be considered collectively when aiming to detect spike time relative difference representations. Correlations between the predicted states of pairs of clusters not sharing a common neuron (4 neurons total; Fig. 7c), strengthen the possibility that lowdimensional nonadaptative trialtotrial populationwide changes in neuronal excitability modulate spike time patterns.
The study differs from revealment of oscillatory activity and spike train alignment to behavioural events through optimisation of populationwide shift and/or stretch functions^{44}, firing rate timewarp modelling^{45} and spike train alignment to behavioural events^{46}. Here we characterise the millisecond and submillisecond structure of spiketiming stimulusresponse distributions, whilst accounting for clustering and nonstationarity. A novel form of sensory representation (timewarped multineuron single spike patterns) is proposed with a potential for submillisecond relative spike time difference encoding (Fig. 5). Non\(45^{\circ }\) correlation angles also demonstrate different timewarp functions for the spike times of different neurons. One previous study modelled timewarpinvariant integration^{47}, which differs from statedependent processing.
Sources of noise correlations are reviewed in^{48}. Subthreshold membrane potential correlations are a prime candidate to underlie spike time noise correlations caused by trialtotrial changes in shared excitability, as they are correlated across areas^{49} and affect action potential initiation^{50}. Heterogeneous relationships between membrane potential changes and spike time latency could underlie non\(45^{\circ }\) correlations. The results open the opportunity to study the relationship between stimulusevoked spike time noise correlations and noise correlations during spontaneous activity (previously characterised for the dataset^{32}). We suggest that the purpose of relative spike time difference modulation, may be to maintain spike time patterns/relationships, by balancing spiketimingdependent synaptic depression and potentiation (Fig. 8e).
Single trial variability in population dynamics is thought to reflect differences in population excitability or state at the start of a trial^{27}. As such, population dynamics may offer a proxy for a cortical state, which modulates spike time patterns. Furthermore, the characterised spike timing variability may reflect the relationship between variability in population dynamics and single neuron spike time (and pairwise spike time difference) variability. In another sense, the characterised spike timing variability may reflect variability in the trajectory of population dynamics at a single neuron level and at high temporal resolution (Fig. 8f). Indeed, the characterisation of shared variability on a continuous spike timing scale could offer the opportunity to study population dynamics without loss of information through discretisation of neuronal responses, which typically uses bins of 1020ms. The precise structure and latent factor dependence of relative difference representations uncovered here suggests that such representations may encode information that is additional to information decodable from classical binned spike count analysis (including classical modelling of population dynamics).
The results offer a large potential improvement for discriminable encoding under fixedstatedecoding and statedependentdecoding regimes. Testing of decoding performance was limited by low cluster counts per stimulus condition and variability in cluster detection between stimulus conditions. For example, most clusters were detected at 0–4 Hz stimulations (Fig. 6e) and clusters were unlikely to be detected for many stimulus conditions for the same neuron pair, likely due to more nonspiking trials for higher stimulation frequencies due to spike count adaptation above 4 Hz^{26,32,36,37} (Fig. 2d). This does not imply the characterised form of representation is not utilised at higher frequencies; but means that fewer samples are available for characterisation at higher frequencies.
The detection of noise correlations over multiple cortical layers and columns (Fig. 7a) between groups of up to 4 neurons (Fig. 7c) suggests that trialtotrial populationwide changes in neuronal excitability could be lowdimensional. This would increase the likelihood that downstream neurons were modulated by the same changes and could respond in a statedependent manner. Determining the dimensionality of excitability changes requires characterisation of noise correlations over larger groups of neurons. Determining the dimensionality of single spike pattern noise correlations is made challenging by the observed clustered nature of pairwise spike time response distributions and low spiking reliability for single neurons (Fig. 2c). The observed clusters, however, offer the opportunity to to study a reliable part of stimulus representations, and to explore the dimensionality of spike time pattern modulation.
Future exploration of the characterised phenomena, including determining dimensionality of excitability changes and assessing encoding performance, would benefit from high trial counts and experimental paradigms designed to minimise adaptation^{51}. Use of dynamic nonstationary and nonparametric clustering may also be beneficial both in the presence and absence of adaptation (Methods).
As the central goal of the manuscript was to extract isolated regions of high response probability, we chose to use DBSCAN as this is known to perform more robustly than GMMs at capturing the structure of the underlying probability density function when regions are nonGaussian^{39}. For example, because adaptation was likely to affect the shape of these regions by increasing latencies over trials, it was important to extract regions of high density for further analysis independently of an assumption of their structure. The nonGaussianity of the isolated regions of high response probability was confirmed for a proportion of the extracted clusters. The first principal component and its angle offered a first approximation of the relationship of spike times and spike time differences with trialtotrial changes in shared excitability (for the isolated regions of high response probability). Moreover, this approximation was able to explain the variance in spike times of other neuron pairs (Fig. 8a). Compared to the unclustered case, clustering better captured isolated regions of highresponse probability for application of PCA, leading to novel detection of correlations and more precise cluster angle estimation (Fig. 4c). Moreover, a large proportion of clusters determined as Gaussian had correlation angles different from \(45^{\circ }\) . Future work may aim to model nonlinear relationships with underlying state, which in turn could enable more precise estimations of single trial state.
The encoding performance of RDR 2 and 3 would be best assessed over a range of similar stimuli. For example, small changes in whisker stimulation angle are likely to elicit slight changes to the response distributions, whilst not drastically changing the number of spiking trials per neuron. This would improve similarity in the number of clusters detected per stimulus and neuron pair. It can then be tested which response distribution properties improve decoding of transformation and correlate with behavioural performance. It is predicted that cluster (RDR 2) or statedependent relative difference distributions (RDR 3) may change continuously with continuous stimulus transformations (Fig. 8g). More broadly, we propose that continua may be encoded in the cortex by timewarped relative spike time differences.
Methods
Neurophysiology
The dataset used for this paper is reused from earlier studies, of which analyses, data collection and spike sorting methods have previously been reported^{5,12,32,52}.
In summary, extracellular recordings were performed in 18 anaesthetised adult Wistar rats using 16 or 128 electrode silicon probes (see below), inserted perpendicularly into the cortical representation of the whiskers in the primary somatosensory cortex, the socalled barrel cortex. Anaesthesia minimised the impact of motor signals and controlled for movement, which is a known source of variability in sensory cortices^{53}. Neural latencies in the principal barrel column are unaffected by the level of anaesthesia^{54}. Single whisker deflections were made at various frequencies (0.06610Hz) in blocks, each containing typically 200 trials (sweeps) with an inter trial interval of 1/f. This was repeated for each of the selected target whiskers (maximum 4). Experimental sessions varied in terms of electrode placement, stimulated whiskers, stimulation frequencies and number of trials per unique stimulus. 12 experimental sessions stimulated at all frequency groups (0.066–0.2, 1, 2, ..., 10 Hz), 2 at low frequencies and a subset of higher frequencies (0.066–0.2, 1, 2, 4, 7 and 10 Hz), and 4 at low frequencies only (0.066–0.2 Hz).
Data from all experimental sessions were used in all analyses. 8 experimental sessions used single shanks with 16 electrodes, whilst 10 experimental sessions used 8 neighbouring shanks with 16 electrodes per shank covering 2–4 barrelrelated columns (median 4). Neurons were classified as excitatory or inhibitory and assigned to a cortical layer L2/3, L4, L5A or L5B/6. Spiking activity is analysed following cortical activation onset, determined as 5.4 ms following stimulus onset (Supplementary Fig. 12).
Spikes were detected using groups of 2 to 4 channels defined as ‘virtual tetrodes’^{55,56}. Spike detection was performed using amplitudethresholding^{57}. Spike sorting was performed leveraging waveform feature engineering and open source software^{55,58}. Neuron quality was verified using several additional steps^{57}. Cells were classified as putative excitatory (E) and inhibitory (I) neurons based on the mean spike waveform^{59,60,61}. Voltagesensitive dye (VSD) imaging, currentsource density analyses and histology were used to identify the appropriate whiskerrelated cortical column and to determine the threedimensional location of electrodes so that each neuron could be assigned to one cortical layer (L2/3, L4, L5A, L5B/6).
Stage 1 clusters
1. DBSCAN clustering. A Python implementation^{62} of DBSCAN^{38,39} was used to detect clusters in first spike pair response distributions. Increasing the epsilon parameter increases the resulting cluster sizes. Clustering was performed using the following epsilon values sequentially: [0.4, 0.45, ..., 5.0]. The colour of points in Supplementary Fig. 13 represents the clusters to which points were assigned for epsilon = 1.35. If the previous lower epsilon value produced the same clustering, then the new clustering was discarded. Stage 1 clusters with less than 30 elements were also discarded, although these data points may be part of larger clusters extracted for higher epsilon values. Checks for similar or identical individual clusters were not made at this stage to avoid sampling bias, which could affect the uniformity of surrogate control tests.
2. Isolation test. This step aimed to ensure that clusters were not subparts of larger clusters but were instead well isolated clusters. For each cluster produced in step 1, an ‘intersection ellipse’ was created (dashed ellipses in Supplementary Fig. 13), with centre equal to the cluster mean and widths equal to \(w_i = 3*max(0.5, \sigma _i)\) for neurons \(i=0,1\), where \(\sigma _i\) is the standard deviation of samples in the cluster for neuron i. Any cluster with an intersection ellipse within or intersecting with another intersection ellipse was excluded from further analysis. In Supplementary Fig. 13, this corresponded to the blue, red and light blue dashed clusters.
3. Stage 1 clusters. For each remaining cluster, a flat bounding ellipse was created (solid green ellipse in Supplementary Fig. 13) with centre equal to the cluster mean, and widths equal to \(w_i = 4*\sigma _i\) for neurons \(i=0,1\), where \(\sigma _i\) is the standard deviation of samples in the cluster for neuron i. All elements within the bounding ellipse were assigned to the Stage 1 cluster. The use of nonangled ellipses was necessary for a controlled test of cluster correlations.
As 94.6% of Stage 2 clusters extracted over 90ms following cortical activation onset lay within 50ms following onset (Supplementary Fig. 12), analysis was constrained to spikes in the first 50ms following activation onset.
Correlation surrogate control datasets
To test that correlations were not artifacts of the clustering algorithm and were above chance levels, the algorithm was applied to the following surrogate control datasets in which correlated clusters appeared at chance levels (Supplementary Fig. 51):
(i) All shuffled. Pairwise shuffling of elements of first spike pair distribution. For each first spike pair stimulusresponse distribution, first spike pairs are shuffled so that each spike time of one neuron is matched with a different spike time of the other neuron.
(ii) Single Gaussian. Spike pairs drawn from an independent bivariate normal distribution with the same mean and standard deviation as the original spikes. The number of samples drawn from the distribution was equal to the number of original samples.
Stage 2 clusters
4. Angled ellipses. An angled ‘Stage 2’ ellipse was calculated to better estimate the distribution underlying each correlated \((p<0.005)\) Stage 1 cluster using a 3stage bootstrap procedure. In the bootstrap step 1, a number of samples were taken with replacement from the Stage 1 cluster equal to the number of samples in the Stage 1 cluster. PCA was performed on the sample to attain a sample mean and covariance matrix. 10,000 such bootstrap iterations were made. The sample means and covariance matrices were then averaged to construct a new bounding ellipse (similar to the rotated green ellipse in Supplementary Fig. 13) with centre equal to the bootstrapped mean, rotation equal to the angle that the principal component made with the positive xaxis, and widths equal to \(4\sigma _i\) where \(\sigma _i\) is the square root of the variance explained by each principal component.
The bootstrap procedure was then repeated by sampling the elements within this new ellipse (bootstrap step 2), with the averaged mean and covariance matrix used to construct a third ellipse for sampling in bootstrap step 3. Elements within the final bounding ellipse defined by the averaged mean and covariance matrix of bootstrap step 3 were assigned to the Stage 2 cluster.
5. Cluster similarity check. Angled clusters which shared at least 15 elements to previous angled clusters (produced by lower epsilon values for the same response distribution) were excluded from further analysis. This means that no clusters were included, which contained well isolated ‘subclusters’ produced by lower epsilon values. As a result, all extracted Stage 2 clusters were from unique response distributions, guaranteeing that there were no repeated clusters.
Correlation angle
The 10,000 sample first principal components produced in bootstrap stage 3 were used to estimate the angle that the first principal component makes with the positive xaxis (with empirical confidence intervals). To account for symmetry, sample principal components with negative ycomponents were replaced by their anticorrelated vector so that all vectors had angles with the positive xaxis between 0 and \(180^{\circ }\). To allow averaging and angular difference calculations without the effects of discontinuity, the vectors were rotated anticlockwise about the origin such that the anticlockwise angles of the vectors with the positive xaxis were between 0 and \(360^{\circ }\) (double their original angle). A mean first principal component vector was then calculated in this space by averaging over the bootstrap samples. For each cluster, the 250th and 9750th largest bootstrap sample angles were used to define a 95% empirical confidence interval (in this \(360^{\circ }\) space). The resulting mean vector was then rotated clockwise by half of its anticlockwise angle with the positive xaxis so that it was between 0 and \(180^{\circ }\). This was used as the final \(\theta\) angle for the cluster. The confidence interval bounds were divided by 2 to create angle confidence intervals for the first principal component on the correct scale.
Cluster angle surrogate control datasets
The following two angle surrogate control datasets were created to verify the main clustering algorithm and compare it to the unclustered case:
(i) \(45^{\circ }\) control. To validate that differences from \(45^{\circ }\) were not artifacts of the algorithm, a surrogate control dataset was created by replacing each Stage 2 cluster and all samples within a \(6\sigma\) circle of its mean (where \(\sigma\) was equal to the mean of the cluster spike time standard deviations of the two neurons) with samples drawn from a \(45^{\circ }\) 2D Gaussian. The algorithm found correlated clusters with angles different from \(45^{\circ }\) at chancelevels (Supplementary Fig. 6).
(ii) 2nd distractor cluster. An additional surrogate control dataset was created to demonstrate that the clustering algorithm, in the presence of a second uncorrelated cluster, detects the first correlated cluster and calculates its correct correlation angle (Supplementary Fig. 62). For each response distribution in which a Stage 2 cluster was extracted by the main clustering algorithm, a new response distribution was created with two new clusters. The first cluster was created by sampling a number of samples equal to the number of samples in the original Stage 2 cluster, drawn from a \(45^{\circ }\) Gaussian centred at the mean of the original Stage 2 cluster, with the new cluster spike time standard deviations of the two neurons equal to the mean of their original standard deviations. The second cluster was created by sampling a number of samples equal to the number of samples in the original Stage 2 cluster, drawn from an uncorrelated 2D Gaussian centred to the right at a \(15^{\circ }\) angle from the first cluster with the cluster spike time standard deviations of the two neurons equal to those of the original Stage 2 cluster. The x coordinate of the mean point was equal to the mean of the original Stage 2 cluster plus 10 standard deviations of the xaxis neuron’s cluster spike times. For this surrogate control dataset, the clustering algorithm detected clusters with correlation angles different from \(45^{\circ }\) at chancelevels (Supplementary Fig. 62).
This differed to the unclustered approach, which found correlations with angles \(\sim 15^{\circ }\), highlighting an advantage of the clustering algorithm over the unclustered approach.
The Henze–Zirkler null hypothesis of normality was rejected \((p < 0.05)\) for 6.5% (11/169) and 7.2% (12/166) of the Gaussian sampled clusters detected in the two angle control datasets, verifying that the algorithm and Henze–Zirkler test could detect normally distributed clusters.
Comparison to alternative clustering methods
The clustering algorithm was compared to the following two Gaussian Mixture Model approaches:
(i) Standard GMM. The first approach uses the classical Bayesian information criterion (BIC) to select an optimal number of Gaussians (between 1 and 10) for the mixture model of each pairwise response distribution. This approach was therefore not constrained to find isolated regions of high response probability. Indeed, for the initial Stage 1 clustering test, this algorithm detected clusters, which were fragmentations of regions of highresponse probability (e.g. Supplementary Fig. 52), rather than isolated regions of high response probability (which were the focus of the study). In fact, the algorithm was biased to detect correlations, such that it found abovechance correlations for the two correlation control datasets (Supplementary Fig. 52). This made it unsuitable for detecting correlated clusters in the main dataset.
(ii) Custom GMM. An augmented GMM approach was used in which steps 24 of the main clustering algorithm were additionally applied after the initial BIC based GMM clustering step (i). This clustering type therefore included checks for isolation. Supplementary Fig. 53a shows that correlations were found at chancelevels after the addition of these steps. However, almost all of the detected clusters included all samples from their originating response distribution (Supplementary Fig. 53b), showing that the algorithm did not find isolated regions of highresponse probability that contained less than 95% of the samples in the original sample distribution. The algorithm did, however, find clusters with correlation angles different from \(45^{\circ }\) both before and after checks for stationarity (Supplementary Fig. 53c), although considerably fewer than the main clustering algorithm and unclustered approach. The Henze–Zirkler null hypothesis of normality was rejected \((p < 0.05)\) for only 31.9% (15/47) of Stage 2 clusters extracted by the Custom GMM algorithm, confirming that the Custom GMM algorithm was biased towards detecting clusters likely to be Gaussian distributed. The algorithm was also verified for the two angle control datasets (Supplementary Fig. 63). Moreover, the Henze–Zirkler null hypothesis of normality was rejected \((p < 0.05)\) for 3.8% (1/26) and 3.3% (1/30) of the Gaussian sampled clusters detected in the two angle control datasets, further verifying the Henze–Zirkler test of normality when used in combination with clustering.
Stationary cluster and factor analysis criteria
Clusters were determined to be stationary and sufficiently correlated such that trialtotrial changes in shared excitability could underlie the covariation if they passed the following tests:

(1)
Cluster spike times were correlated: linear regression pvalue < 0.005, rvalue > 0.3.

(2)
Single neuron cluster spike times were not linearly correlated with trial index (pvalue > 0.05).

(3)
The cluster correlation was not explained by trial index. Particularly, a linear regressor predicting spike pairs from trial index must have \(r^2 < 0.05\).

(4)
Both neurons KPSS stationary: pvalue > 0.05. The Kwiatkowski–Phillips–Schmidt–Shin trend stationarity test (KPSS) tested whether cluster first spike times of single neurons were stationary around a deterministic trend.

(5)
Cluster spike times of both neurons have a triallag1 autocorrelation pvalue > 0.05 (this requirement was relaxed for the differenced clusters but not ARIMA clusters). The triallag1 autocorrelation describes the following. Consider the cluster spike times \(s_{j,t}\) of neurons j = 0,1 on cluster trials t = 1,...,T where T is the number of cluster trials for the cluster. Triallag1 autocorrelations tests the correlations between single neuron cluster spike times on trials t with spike times on trial t+1.

(6)
Bartlett sphericity pvalue > 0.05. Bartlett sphericity tests that the correlation matrix is significantly different from the identity, and is a standard pretest for factor analysis.

(7)
The factor correlation matrix had only one eigenvalue greater than 1, suggesting that a single factor underlies each cluster correlation.
Factor analysis angle and difference distribution
\(\theta\) and \(\theta _{45}\) angles of factor analysis mean lines were estimated using the same technique as for PCA angles. Factor analysis \(\theta\) angles were similar to the PCA \(\theta\) angles for stationary, differenced and ARIMA modelled clusters (Supplementary Fig. 11).
As the factor analysis model assumes that noise for each neuron given \(\eta\) is normally distributed and independent between neurons, the statedependent spike time relative difference standard deviation predicted by the model is equal to \(\sigma _{diff} = \sqrt{\sigma _{0}^2 + \sigma _{1}^2}\) where \(\sigma _i\) is the statedependent standard deviation of neuron i predicted by factor analysis.
Differencing
Differencing was applied to single neuron cluster spike times if any of the following stationarity criteria were not met for the cluster spike times of either neuron: (1) Trial index vs spike time linear regression pvalue > 0.05, (2) KPSS stationarity pvalue > 0.05, (3) Triallag1 autocorrelation pvalue > 0.05.
Differencing a series of first spike times subtracts the first spike time on trial t1 from the first spike time trial t. Differencing reduces the number of samples by 1. For clusters for which the spike times of only one neuron was differenced, the first sample of the undifferenced neuron was removed, so that the number of samples was equal for both neurons. In this case a spike time for an undifferenced trial t was matched with the difference between the spike times on trial t and trial t−1.
Differencing increases the standard deviation of a time series \(S = (s_1, \ldots , s_t)\) by approximately \(\sqrt{2}\). This is shown by considering differencing as approximate to negating two independent random variables \(S_{(1, t1)} = (s_1, \ldots , s_{t1})\) and \(S_{(2, t)} = (s_2, \ldots , s_t)\), where t is the number of samples in the original time series. As the standard deviation of \(S_{(0, t1)}\) and \(S_{(1, t)}\) are approximate to the standard deviation of the original series S, the standard deviation of the differences is on the order of \(\sigma _{differenced} \approx \sqrt{\sigma _{(1, t1)}^2 + \sigma _{(2, t)}^2} \approx \sqrt{2\sigma ^2} = \sqrt{2}\sigma\) where \(\sigma\) is the standard deviation of the original time series. Differenced time series were therefore divided by \(\sqrt{2}\) to return the differenced time series to the scale of the original time series, which comparison to ARIMA models shows to be reasonable (Fig. 4f). The assessment of ‘raw’ differences, rather than the ARIMA residuals, however, allows a direct assessment of cluster angles after trend removal.
Differencing was applied to the spike times of clusters rather than to all first spike times of a single neuron in response to a single stimulus. This was to ensure that differences were calculated between spike times of the same trials for both neurons in a pair. That is, the difference between spike times on trial X and spike times on trial Y were compared for neuron pairs. Differencing was not alternatively applied to all samples of a first spike stimulusresponse distribution, as this could give rise to correlations caused by switching between clusters.
Automated BoxJenkins method
To determine the order of autoregressive (AR) and moving average (MA) components, the BoxJenkins method is classically applied manually based on observation of autocorrelations and partial autocorrelations. To deal with the negative autocorrelations introduced by differencing and to account for other autocorrelations, an automated and simplified version was developed based on the description of the BoxJenkins method found at^{63}.
For determination of the MA order, if the triallag1 autocorrelation is negative and significant, the series is slightly overdifferenced and the BoxJenkins method checks for a sudden cut off in the autocorrelation plot. The lag at which there is a cutoff is used as the MA order. To automate this, the algorithm checked if the triallag1 autocorrelation was negative (\(r < 0.1\)). If so, the pvalues of lagged autocorrelations were tested (lags 1–5). If there was a sudden increase in pvalue from below 0.05, from one lag to another, by more than 0.15, then the value of the last significant autocorrelation was used as the MA order.
For determination of the AR order, if the triallag1 autocorrelation is positive (\(r > 0.1\)) and significant, the series is slightly underdifferenced, and the BoxJenkins method checks similarly for a sudden cutoff in the partialautocorrelation plot to determine the AR term.
Figure 5c confirms that the majority of ARIMA applications were of simple loworder MA components, as expected for the differenced time series. Moreover, autocorrelations and crosscorrelations were reduced to chancelevels by ARIMA application, and ARIMA estimated cluster angles were similar to the original clusters (Supplementary Fig. 11).
Further discussion of methods
To control the possible illposedness and the robustness problem of the underlying clustering problem in nonGaussian situations with strong noise and small data samples, a robust clustering pipeline was designed which took into account the level of noise present in the data and the low number of samples in each response distribution. It follows that surrogate control datasets should match the general statistics of the response distributions. For these reasons surrogate control datasets were created by transforming preexisting response distributions.
The joint roles of adaptation and trialtotrial changes in excitability promote the use of dynamic clustering techniques in future work where adaptation is present^{64}. The multiscale character of the underlying neural processes introduces the possibility of lowdimensional temporally persistent latent regimeswitching processes, which are a common source of nonstationarity across nature^{64} and are also commonly ignored by analysis techniques^{65}. Such processes are likely to contribute to the complexity of the observed time series, including clusters.
Techniques, which address these issues for limited sample sizes (i.e.^{66}) should be applied in future research where adaptation is present to (i) characterise the multiscale nature of regime switching processes, (ii) estimate stimulus representation independently of adaptation and invisible biases caused by multiscale processes beyond depressive spike time latency adaptation^{65} and (iii) characterise the dimensionality of trialtotrial changes in excitability . Understanding of (i) and (ii) will allow stricter control of adaptative and multiscale processes for the understanding of (iii). This paper aimed to take a first step towards understanding such dynamical processes by characterising a reliable subset of responses (clusters).
Data availability
All code and data needed to reproduce all figures of the manuscript are available at https://github.com/jamesisbister/spikewarp.git and https://github.com/jamesisbister/IsbisterEtAlSciRep2021Data.git.
References
Gerstner, W., Kreiter, A. K., Markram, H. & Herz, A. V. Neural codes: Firing rates and beyond. Proc. Natl. Acad. Sci. 94, 12740–12741 (1997).
Johnson, K. O. Neural coding. Neuron 26, 563–566 (2000).
Gerstner, W., Kistler, W. M., Naud, R. & Paninski, L. Neuronal Dynamics: From Single Neurons to Networks and Models of Cognition (Cambridge University Press, 2014).
Brette, R. Philosophy of the spike: Ratebased vs. spikebased theories of the brain. Front. Syst. Neurosci. 9, 151 (2015).
ReyesPuerta, V. et al. Longrange intralaminar noise correlations in the barrel cortex. J. Neurophys. 113, 3410–3420 (2015).
Zuo, Y. et al. Complementary contributions of spike timing and spike rate to perceptual decisions in rat S1 and S2 cortex. Current Biol. 25, 357–363 (2015).
Uzzell, V. & Chichilnisky, E. Precision of spike trains in primate retinal ganglion cells. J. Neurophysiol. 92, 780–789 (2004).
Johansson, R. S. & Birznieks, I. First spikes in ensembles of human tactile afferents code complex spatial fingertip events. Nat. Neurosci. 7, 170–177 (2004).
Gollisch, T. & Meister, M. Rapid neural coding in the retina with relative spike latencies. Science 319, 1108–1111 (2008).
Storchi, R., Bale, M., Biella, G. & Petersen, R. Comparison of latency and rate coding for the direction of whisker deflection in the subcortical somatosensory pathway. J. neurophysiol. 108(7), 1810–21 (2012).
Reinagel, P. & Reid, R. C. Precise firing events are conserved across neurons. J. Neurosci. 22, 6837–6841 (2002).
ReyesPuerta, V., Sun, J.J., Kim, S., Kilb, W. & Luhmann, H. J. Laminar and columnar structure of sensoryevoked multineuronal spike sequences in adult rat barrel cortex in vivo. Cerebral Cortex 25, 2001–2021 (2014).
Panzeri, S., Petersen, R. S., Schultz, S. R., Lebedev, M. & Diamond, M. E. The role of spike timing in the coding of stimulus location in rat somatosensory cortex. Neuron 29, 769–777 (2001).
Shadlen, M. N. & Newsome, W. T. The variable discharge of cortical neurons: Implications for connectivity, computation, and information coding. J. Neurosci. 18, 3870–3896 (1998).
Lin, I.C., Okun, M., Carandini, M. & Harris, K. D. The nature of shared cortical variability. Neuron 87, 644–656 (2015).
Hong, H., Yamins, D. L., Majaj, N. J. & DiCarlo, J. J. Explicit information for categoryorthogonal object properties increases along the ventral stream. Nat. Neurosci. 19, 613 (2016).
Chang, L. & Tsao, D. Y. The code for facial identity in the primate brain. Cell 169, 1013–1028 (2017).
Stringer, C., Michaelos, M., Tsyboulski, D., Lindo, S. E., & Pachitariu, M. Highprecision coding in visual cortex. Cell, 184(10), 2767–2778 (2021).
Gasparini, S. & Magee, J. C. Statedependent dendritic computation in hippocampal CA1 pyramidal neurons. J. Neurosci. 26, 2088–2100 (2006).
Branco, T., Clark, B. A. & Häusser, M. Dendritic discrimination of temporal input sequences in cortical neurons. Science 329, 1671–1675 (2010).
Branco, T. & Häusser, M. Synaptic integration gradients in single cortical pyramidal cell dendrites. Neuron 69, 885–892 (2011).
Markram, H., Lübke, J., Frotscher, M. & Sakmann, B. Regulation of synaptic efficacy by coincidence of postsy502 naptic APs and EPSPs. Science 275, 213–215 (1997).
Mainen, Z. F. & Sejnowski, T. J. Reliability of spike timing in neocortical neurons. Science 268, 1503–1506 (1995).
Carandini, M. Amplification of trialtotrial response variability by neurons in visual cortex. PLoS Biol. 2, e264 (2004).
Branco, T. & Staras, K. The probability of neurotransmitter release: variability and feedback control at single synapses. Nat. Rev. Neurosci. 10, 373–383 (2009).
Chung, S., Li, X. & Nelson, S. B. Shortterm depression at thalamocortical synapses contributes to rapid adaptation of cortical sensory responses in vivo. Neuron 34, 437–446 (2002).
Kao, J. C. et al. Singletrial dynamics of motor cortex and their applications to brainmachine interfaces. Nat. Commun. 6, 1–12 (2015).
Pandarinath, C. et al. Inferring singletrial neural population dynamics using sequential autoencoders. Nat. Methods 15, 805–815 (2018).
Arieli, A., Sterkin, A., Grinvald, A. & Aertsen, A. Dynamics of ongoing activity: Explanation of the large variability in evoked cortical responses. Science 273, 1868–1871 (1996).
Schölvinck, M. L., Saleem, A. B., Benucci, A., Harris, K. D. & Carandini, M. Cortical state determines global variability and correlations in visual cortex. J. Neurosci. 35, 170–178 (2015).
Resulaj, A., Ruediger, S., Olsen, S. R. & Scanziani, M. First spikes in visual cortex enable perceptual discrimina tion. Elife 7, e34044 (2018).
ReyesPuerta, V. et al. High stimulusrelated information in barrel cortex inhibitory interneurons. PLoS Comput. Biol. 11, e1004121 (2015).
Feldmeyer, D. et al. Barrel cortex function. Progr. Neurobiol. 103, 3–27 (2013).
Abeles, M. Local Cortical Circuits: An Electrophysiological Study (Springer, 2012).
Izhikevich, E. M. Polychronization: Computation with spikes. Neural Comput. 18, 245–282 (2006).
Khatri, V., Hartings, J. A. & Simons, D. J. Adaptation in thalamic barreloid and cortical barrel neurons to periodic whisker deflections varying in frequency and velocity. J. Neurophysiol. 92, 3244–3254 (2004).
Wang, Q., Webber, R. M. & Stanley, G. B. Thalamic synchrony and the adaptive gating of information flow to cortex. Nat. Neurosci. 13, 1534 (2010).
Ester, M. et al. A densitybased algorithm for discovering clusters in large spatial databases with noise. KDD 96, 226–231 (1996).
Rodriguez, A. & Laio, A. Clustering by fast search and find of density peaks. Science 344, 1492–1496 (2014).
Nau, R. Identifying the order of differencing in an ARIMA model. Retrieved July 13, 2021, from https://people.duke.edu/~rnau/411arim2.htm (2020).
Ho, S. & Xie, M. The use of ARIMA models for reliability forecasting and analysis. Comput. Ind. Eng. 35, 213–216 (1998).
Box, G. E., Jenkins, G. M., Reinsel, G. C. & Ljung, G. M. Time Series Analysis: Forecasting and Control (John Wiley & Sons, 2015).
Carvell, G. E. & Simons, D. J. Biometric analyses of vibrissal tactile discrimination in the rat. J. Neurosci. 10, 2638–2648 (1990).
Williams, A. H. et al. Discovering precise temporal patterns in largescale neural recordings through robust and interpretable time warping. Neuron 105, 246–259 (2020).
Lawlor, P. N., Perich, M. G., Miller, L. E. & Kording, K. P. Linearnonlineartimewarppoisson models of neural activity. J. Comput. Neurosci. 45, 173–191 (2018).
Shusterman, R., Sirotin, Y. B., Smear, M. C., Ahmadian, Y., & Rinberg, D. Sniff invariant odor coding. Eneuro, 5(6). (2018).
Gütig, R. & Sompolinsky, H. Timewarpinvariant neuronal processing. PLoS Biol. 7, e1000141 (2009).
Doiron, B., LitwinKumar, A., Rosenbaum, R., Ocker, G. K. & Josic, K. The mechanics of statedependent neural correlations. Nat. Neurosci. 19, 383–393 (2016).
Poulet, J. F. & Petersen, C. C. Internal brain state regulates membrane potential synchrony in barrel cortex of behaving mice. Nature 454, 881 (2008).
Petersen, C. C., Hahn, T. T., Mehta, M., Grinvald, A. & Sakmann, B. Interaction of sensory responses with spontaneous depolarization in layer 2/3 barrel cortex. Proc. Natl. Acad. Sci. 100, 13638–13643 (2003).
Harris, K. D. Nonsense correlations in neuroscience. bioRxiv, 202011. (2021).
ReyesPuerta, V. et al. Propagation of spontaneous slowwave activity across columns and layers of the adult rat barrel cortex in vivo. Brain Struct. Funct. 221, 4429–4449 (2016).
Stringer, C. et al. Spontaneous behaviors drive multidimensional, brainwide activity. Science 364, 6437 (2019).
Simons, D. J., Carvell, G. E., Hershey, A. E. & Bryant, D. P. Responses of barrel cortex neurons in awake rats and effects of urethane anesthesia. Exp. Brain Res. 91, 259–272 (1992).
Harris, K. D., Henze, D. A., Csicsvari, J., Hirase, H. & Buzsaki, G. Accuracy of tetrode spike separation as determined by simultaneous intracellular and extracellular measurements. J. Neurophysiol. 84, 401–414 (2000).
Einevoll, G. T., Franke, F., Hagen, E., Pouzat, C. & Harris, K. D. Towards reliable spiketrain recordings from thousands of neurons with multielectrodes. Current Opin. Neurobiol. 22, 11–17 (2012).
Gray, C. M., Maldonado, P. E., Wilson, M. & McNaughton, B. Tetrodes markedly improve the reliability and yield of multiple singleunit isolation from multiunit recordings in cat striate cortex. J. Neurosci. Methods 63, 43–54 (1995).
Hazan, L., Zugaro, M. & Buzsáki, G. Klusters, NeuroScope, NDManager: a free software suite for neurophysio logical data processing and visualization. J. Neurosci. Methods 155, 207–216 (2006).
Sirota, A. et al. Entrainment of neocortical neurons and gamma oscillations by the hippocampal theta rhythm. Neuron 60, 683–697 (2008).
Sakata, S. & Harris, K. D. Laminar structure of spontaneous and sensoryevoked population activity in auditory cortex. Neuron 64, 404–418 (2009).
Royer, S. et al. Control of timing, rate and bursts of hippocampal place cells by dendritic and somatic inhibition. Nat. Neurosci. 15, 769–775 (2012).
Pedregosa, F. et al. Scikitlearn: Machine Learning in Python. J. Mach. Learn. Res. 12, 2825–2830 (2011).
Nau, R. Identifying the orders of AR and MA terms in an ARIMA model https://people.duke.edu/~rnau/ 411arim3.htm. Accessed 28 Oct 2020. (2020).
Metzner, P., Putzig, L. & Horenko, I. Analysis of persistent nonstationary time series and applications. Commun. Appl. Math. Comput. Sci. 7, 175–229 (2012).
DeWiljes, J., Putzig, L. & Horenko, I. Discrete nonhomogeneous and nonstationary logistic andMarkov regression models for spatiotemporal data with unresolved external influences. Commun. Appl. Math. Comput. Sci. 9, 1–46 (2014).
Horenko, I. On a scalable entropic breaching of the overfitting barrier for small data problems in machine learning. Neural Comput. 32, 1–17 (2020).
Acknowledgements
JBI would like to thank family, friends and colleagues who kindly read and commented on the manuscript, as well as friends and colleagues in the Oxford Centre for Theoretical Neuroscience and Artificial Intelligence, EPFL Blue Brain Project and MIT Centre for Brains, Minds and Machines for their useful discussions and support. JBI would like to thank Dr. Simon Stringer for inspiring him to explore spike time coding. JBI’s contribution to this work was supported by the Oxford Foundation for Theoretical Neuroscience and Artificial Intelligence, the UK Economic and Social Science Research Council (ESRC) (Grant No. ES/J500112/1) and the UK Engineering and Physical Science Research Council (EPSRC) (Grant No. EP/N509711/1), and funding granted by the Swiss government’s ETH Board of the Swiss Federal Institutes of Technology to the Blue Brain Project, a research centre of the École Polytechnique Fédérale de Lausanne (EPFL). HJL and IH are supported by Deutsche Forschungsgemeinschaft and CarlZeiss grant EAI to the University of Mainz.
Author information
Authors and Affiliations
Contributions
J.B.I. conceptualised the study and main ideas. J.B.I. designed, implemented and performed the analysis and methodology. J.B.I. wrote the manuscript, which was edited by V.R.P., H.J.L. and J.J.S. Data collection and curation was performed by V.R.P., H.J.L. and J.J.S. Access to the dataset was facilitated by V.R.P. and H.J.L. The applied methodology was checked and commented on by I.H., V.R.P., H.J.L. and J.J.S.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher's note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Isbister, J.B., ReyesPuerta, V., Sun, JJ. et al. Clustering and control for adaptation uncovers timewarped spike time patterns in cortical networks in vivo. Sci Rep 11, 15066 (2021). https://doi.org/10.1038/s41598021940020
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41598021940020
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.