Introduction

Information processing in a mammalian brain is distributed across many anatomically and functionally distinct areas¹. The inter-areal functional and anatomical connectivity can be reconstructed using electron microscopy², gene co-expression³, projection tracing methods^4,5, and functional neuro-imaging⁶. Further analysis of the temporal variation of neural activity across this connectome⁷ using functional neuroimaging^8–11 have indicated that complex cognitive functions rely on highly dynamic routing of neural activity within this anatomically constrained structural backbone¹². Neuroimaging, however, captures activity averaged over large populations (over 500K neurons per voxel) at relatively slow time scales (minutes to hours) that typically shows relatively low dimensionality¹³ hence hindering mechanistic explanation. Further insights into dynamic routing of information flows across the brain are expected with adoption of recent technological advances in multielectrode arrays^14,15 and volumetric 2-photon imaging¹⁶ that enable simultaneous recording from large neuronal populations at a cellular level with single spike temporal resolution that reveals rich and high-dimensional dynamic interactions^17,18. Descriptive statistical approaches^15,19 that summarize the pairwise millisecond-timescale synchrony of spiking within recorded neuronal population indicate a directional flow of functional connectivity¹⁵ ascending along the anatomical hierarchy⁵ and reveal distinct engaged modules across the cortical hierarchy^19–21. However, both descriptive statistical^15,19 as well as modern probabilistic models^22–24 typically extract interactions as time-independent “static” variables^15,19,25 or consider temporal dynamics of the whole population^{20,21,26–29}. Temporal changes in the magnitude of pair-wise correlations calculated with descriptive statistical methods have been used to track the dynamics of synaptic interactions during behavior and formation of place fields in hippocampal networks^30–32. Temporal fluctuations of the synaptic efficacy were also inferred from pair-wise correlations in spike trains using various extended generalized linear models (GLM)^33–37. However, the synaptic efficacy is typically calculated separately for each neuron pair while in brain networks each neuron is connected to thousands of others.

Here, to uncover the dynamic routing of information in cortical networks, we develop an unsupervised probabilistic model - Dynamic Network Connectivity Predictor (DyNetCP). As opposed to previous approaches that focus on extracting pair-wise correlations^30–37, we aim at learning dynamic changes in connection strength for all recorded neurons in the network. Our approach is also fundamentally different from most of latent variable models that infer the dynamics of the entire recorded population in a low-dimensional latent feature space^26–29. Instead, to enable interpretation as directed time-dependent network connectivity across distinct cellular-level neuronal ensembles, DyNetCP extracts latent variables that are assigned to individual neurons.

Model

Model architecture

Classical machine learning approaches like variational autoencoders^26,38 successfully train dynamic models of brain activity across entire recorded neural population. However, development of such dynamic models to uncover cellular-level dynamic connectivity structure, is facing a major problem of the virtual absence of ground-truth (GT) data either as verified experimental or synthetic datasets with known time-varying connectivity. In this situation optimization of the model hyper-parameters, control of overfitting, and model verification is largely becoming a surmise. Therefore, to address this challenge, here, learning of the directed dynamic cellular-level connectivity structure in a network of spiking neurons is divided into two stages (Fig.1A). First stage, the static component, infers the static connectivity weight matrix W_st (Fig.1A, left) via a generalized linear model (GLM). At this stage, the classical descriptive statistical approaches like pair-wise jitter-corrected cross-correlograms^15,30 can be used as a GT to verify the model performance. Next, in the second stage, W_st is used to influence the dynamic offset weight term W_off learned via a recurrent long-short-term-memory (LSTM) algorithm (Fig.1A, right). The final dynamic weight W_dyn = W_st + W_off is averaged across all trials. At this stage, the model W_dyn can be verified against another classical descriptive statistical method - pairwise Joint Peristimulus Time Histogram³⁹ (JPSTH) that represents the time variation of joint correlations. Therefore, dividing the problem into two stages enables independent verification of static weights and dynamic offsets against GT synthetic and experimental data as well as a direct comparison of DyNetCP performance to existing models.

Figure 1.

Fig.1 - figure supplement 1.

The model operates on trial-aligned binned neural spikes. We collect N neurons recorded for T time bins binned at d bin width across M trials in tensor X ∈ {0,1}^M×N×T, where each entry represents the presence or absence of a spike from neuron j at time t in trial i. The spiking probability of neuron i at time t is modelled as follows:

where σ is the sigmoid function, E_static ∈ ℝ and E_dyn ∈ ℝ are static and dynamic connectivity logits depending on the trainable weight matrix W_st and the dynamic offset model W_off respectively. Given spiking data as input, the model produces two sets of output weight matrices. The first set, W_st, encodes an estimate of the static network structure and contains one weight per pair of neurons averaged across the trial time. These represent the time-averaged influence the spikes of the whole population have on spiking probability of a given neuron. The second set, W_dyn, contains a single weight for every pair of neurons for every point in time during the experimental trial. These weights, which are a function of the input spiking data of all neurons in a population, represent a time-varying offset to the static weights. Hence, they encode temporal changes in activity patterns, and are used to determine when a given connection is being active and when it is unused. Therefore, in this formulation, the model learns time-varying connectivity patterns that can be directly linked to the underlying network structure at a cellular level.

Static connectivity inference

The static component of DyNetCP (Fig.1A, left), i.e.,

learns static connectivity weights that encode static network structure averaged over the time of the trial t, and b_i ∈ ℝ are biases. We adopt a multivariate autoregressive GLM model²³ without temporal kernel smoothing that learns W_st as a function of a time lag d (Fig.1B). Sharp peak at short time lags implies increased probability of spiking of neuron j when neuron i is producing a spike that can be an indication that these two neurons are synaptically connected. An apparent time lag of the low-latency peak indicates a functional delay and defines the directionality of the putative connection. The static component of DyNetCP is implemented for all neurons jointly using a 1-dimensional convolution. Each convolutional filter contains the static weights used to predict the spiking probability for a single neuron as a function of the spiking history of all other neurons. The bias term b_i in Eq.2 is initialized to the logit function σ (i.e., inverse sigmoid) applied to the spike rate of each neuron calculated for a given time bin.

Dynamic connectivity inference

The dynamic component of DyNetCP (Fig.1A, right), i.e.,

learns the dynamic connectivity weights which are a function of input spike trains and all other . Each of these weights represents an offset to the time-averaged static weight at time t for a connection at time lag d (Eq.2). Hence, assuming the spikes of neuron i at a time lag d are influenced by spikes from all , the total influence applied to the spiking of neuron i from neuron j at time t is . Unlike the static weights, these dynamic weights are allowed to change throughout an experimental trial as a function of the spiking history of the connected neurons.

The architecture of the dynamic part (Fig.1C) consists of several stages including generation of Fourier features, recurrent long-short-term-memory (LSTM) model, multilayer perceptron (MLP), embedding, and positional encoding. First, we shift each spike train such that each entry lies in {−1, 1} and then represent each spike train using Fourier features⁴⁰ that were recently introduced in the computer vision domain to facilitate the learning of high-frequency functions from low-dimensional input. Specifically, we encode every spike using the following transformation:

Here, we use m = 64 features, where coefficients b₁, …, b_m are linearly spaced between –1 and 1. We evaluate the use of Fourier features within the model on synthetic data consisting of a simulated network of neurons, where one neuron induces spiking in a second neuron during the middle 250 ms of each trial (Fig.1 - figure supplement 1B). DyNetCP without Fourier features is unable to properly localize the period where induced spiking is occurring, while the model with Fourier features can do so correctly.

The spike train features for each neuron are then input into separate LSTMs to produce an embedding per neuron per time step. Each of these LSTMs uses a separate set of weights and has a hidden dimension of 64. The LSTM embeddings are passed through two MLPs, f_send and f_recv, where f_send produces an embedding which represents the outgoing neural signal to other neurons and f_recv produces an embedding which represents the incoming signal from other neurons. Each of these MLPs consists of one linear layer with output size 64 and a ReLU activation. For each pair n_j → n_i of neurons being modeled, the send embedding ℎ_s,j(t) is concatenated with the receive embedding ℎ_r,i(t − 1) as well as a time encoding ℎ_T(t) which represents the trial time and is represented as

where D_τ is the size of the positional encoding which we set to 128. Note that the send embedding contains spiking information through time t, i.e., it contains information about the most recent spikes, which is necessary to properly represent synchronous connections. Importantly, the receive embedding only contains spiking information through time t − 1 such that the output spiking probability is not trivially recoverable. Finally, this concatenated embedding is passed through a 2-layer MLP with hidden size 384 and a Rectified Linear Unit (ReLU) activation, producing the dynamic weight vector which contains dynamic weight offsets for each delay d ∈ {1, …, M}.

Model training

The loss used to train the model consists of two terms. The first is a negative log likelihood over the predicted spike probabilities, which encourages the weights to accurately encode the influence of spiking from other neurons on the spiking of the neuron in question. The entire loss is written as

where w represents all the predicted static and dynamic weights, and λ is a coefficient used to balance the magnitudes of the losses. Loss optimization during training results in predicted spike rate (PSTH) that is approaching the actual spike rate (Fig.1 – figure supplement 1A).

We introduce a squared L2 penalty on all the produced static and dynamic weights as the second term to discourage the network from overpredicting activity for times when no spikes are present and the weight multiplied by zero is unconstrained. In these cases, adding the penalty provides a training target. For all experiments we used an L2 weight between 0.1 and 1 that produced the best fit to the descriptive statistics results discussed in detail below (Methods).

Overall, once the architecture of Fig.1C is established, there are just a few parameters to optimize including L2 regularization and loss λ coefficient. Other variables, common for all statistical approaches, include bin sizes in the lag time and in the trial time. Decreasing the bin size will improve time resolution while decreasing the number of spikes in each bin for reliable inference. Therefore, the number of spikes threshold and other related thresholds need to be adjusted as discussed in detail in the Methods, Section 4.

Results

The DyNetCP model is designed to be trained in a single step to produce both static and dynamic weight matrices simultaneously for the whole dataset. However, for the purpose of this work, we are interested in verification and detailed comparative analysis of the model performance that enables clear justification for the choice of model parameters. Therefore, here, training of the model to analyze a new dataset is implemented in two separate stages. In the first stage, only the static weights W_st are trained for all the valid neurons while the dynamic weights W_dyn are all set to zero. In the second stage, the static model parameters and static weights are fixed, and we train only the dynamic component of the model.

This two-stage architecture enables detailed comparison of static and dynamic matrices to classical descriptive statistical methods as well as to modern probabilistic models. In the virtual absence of synthetic or experimental datasets with known GT such comparisons are becoming crucial for tuning the model parameters and for validation of the results. It is especially important for validation of the dynamic stage of the DyNetCP since it predicts W_dyn that describes temporal variation of the connection strength that is assigned to individual neurons, and, therefore, could not be directly compared to most of the population dynamics models that produce temporal trajectories in a latent feature space of the whole population^26–29. In what follows we consider these comparisons separately.

Static DyNetCP can faithfully reproduce descriptive statistics cross-correlogram

The DyNetCP formulation of its static stage (Eq.2) is similar to classical cross-correlogram^15,30 (CCG) between a pair of neurons n₁ and n₂ that is defined as:

where τ represents the time lag between spikes from n₁ and n₂, λ_j is the spike rate of neuron j, and θ(τ) is a triangular function which corrects for the varying number of overlapping bins per delay. However, instead of constructing the statistical CCG histogram for each pair individually, the model given in Eq.2 infers coupling terms from spiking history of all N neurons jointly. Typically, to minimize the influence of a potential stimulus-locked joint variation in population spiking that often occurs due to common input from many shared presynaptic neurons, a jitter correction^{15,30,41–43} is applied. To directly compare the results of static DyNetCP with jitter-corrected CCG, we train two static models (Methods) – one on original spike trains and the second on a jitter-corrected version of the data generated using a pattern jitter algorithm⁴¹ with a jitter window of 25ms.

For this comparison we used a subset of the Allen Brain Observatory-Visual Coding Neuropixels (VCN) database¹⁵ that contains spiking activity collected in-vivo from 6 different visual cortical areas in 26 mice (Methods, Table 1) passively viewing a “drifting grating” video (top left inset in Fig.1A). For a pair of local units (spatially close units from Layer 5 of the VISam area) comparison of static DyNetCP (Fig.2A, red curve) with jitter-corrected CCG (Fig.2A, black curve) shows close correspondence (Pearson correlation coefficient r_p=0.99 and p-value P_p<0.0001). Residual analysis (Fig.2A, bottom panel) shows a mean value of 0.01±0.02 SEM (Fig.2A, top right panel) that is smaller than the CCG values on the flanks (bottom right panel) (mean:0.11±0.06 SEM). For such a local connection the peak is narrow (∼2ms) with a functional delay of −2.3ms. For a pair of spatially distant units (units in VISam and VISp areas) a comparison of static DyNetCP (Fig.2B, red curve) with jitter-corrected CCG (Fig.2B, black curve) also shows close correspondence (r_p=0.99 and P_p<0.0001). For such a distant connection the peak is wider (∼9ms) with a functional delay of −5.8ms. Further examples of comparisons of both models for all the units in the session showed similar results (see accompanying Jupyter Notebook).

Figure 2.

Fig.2 - figure supplement 1.

Table. 1.

Following the common formulation, a putative excitatory (or inhibitory) synaptic connection can be considered statistically significant for both methods when the magnitude of the short-latency peak (or trough) exceeds the peak height threshold α_w (lower than −α_w) compared to the standard deviation (SD) of the noise in the flanks far from the peak^15,19,25,30 (whisker plots on top right panels in Fig.2A,B).

Static DyNetCP reproduces GT connections on par with alternative probabilistic method

We compare the performance of DyNetCP to a recent GLM-based cross correlation method (GLMCC)²⁵ that detects connectivity by fitting a parameterized synaptic potential filter (Fig.2C, magenta). For these experiments, we use the synthetic dataset²⁵ consisting of 1000 Hodgkin-Huxley (HH) neurons with known GT synaptic connections (Methods). We focus specifically on recovering ground-truth (GT) excitatory connections. The W_st inferred by DyNetCP for a representative pair of neurons (Fig.2C, red) reproduces well the short-latency narrow peak and the flanks of the postsynaptic potential (PSP) filter learned by GLMCC (Fig.2C, magenta). Comparing to CCG histogram (Fig.2C, black), the DyNEtCP is performing much better than GLMCC (r_p=0.67 and P_p<0.001), however the results are not as directly interpretable as GLMCC that outputs synaptic weight from the PSP fitting. Both models, however, can infer presence or absence of the connection by examining the height of the low latency peak.

To facilitate a performance comparison, we use the Matthews Correlation Coefficient (MCC), defined as

where TP, TN, FP, and FN are true positive, true negative, false positive, and false negative pair predictions, respectively. MCC optimization balances between maximizing the number of TP pairs recovered while minimizing the number of FP pairs which are predicted.

Performance of all three methods, classical CCG and probabilistic GLMCC and DyNetCP, will be strongly affected by the total number of spikes available for inference. To ensure the statistical significance of connections inferred by all 3 methods a large enough number of spikes should be present in each time bin (Methods) that is tracked by two threshold parameters: a unit average spike rate α_s, and a pair spike parameter α_p that is the number of spikes that both units generate in a time window of interest (Methods). Fig.2D compares the performance of DyNetCP and GLMCC on a synthetic dataset, while the number of spikes present for learning is varied via a threshold pair spike parameter α_p. Increasing α_p ensures statistically significant inference about the presence or absence of a connection²⁵ (Methods). As α_p increases, the MCC for both models shows a maximum around α_p=80 corresponding to the optimal balance. For a conservative value of α_w=7SD (blue), the MCC for the DyNetCP model is 63%, approaching the performance of GLMCC (68%, black). When α_w is relaxed to 5SD (red), DyNetCP recovers a larger number of TP with a reasonable increase of FP connections (Fig.2 - figure supplement 1), increasing MCC up to 74%, thus outperforming GLMCC. Overall, as long as DyNetCP has access to a sufficient number of spikes for inference, it can restore GT static connectivity on par with or even better than GLMCC (Fig.2E).

Though GLMCC performs well in recovering static network structure and it produces directly biologically interpretable results, we emphasize that GLMCC is not compatible with modeling the network dynamics and thus cannot be integrated into DyNetCP. Specifically, our goal is to model the correlations of spiking of a network of neurons and track the changes of these correlations across time. GLMCC learns correlations as a function of the time lag relative to the spiking time of a reference neuron. It is hence unable to link correlations to specific moments in a trial time. Therefore, it cannot represent the dynamics of correlations, preventing its use in analysis of network activity relative to subject behavior. In contrast, the formulation of static DyNetCP model of Eq.2 provides a weight matrix W_st which is compatible with the subsequent inference of time-dependent dynamic weights W_dyn.

Static connectivity inferred by the DyNetCP from in-vivo recordings is biologically interpretable

Recent reconstruction of the brain-wide connectome of the adult mouse brain using viral tracing of axonal projections^4,5 enables assignment of anatomical hierarchical score values to cortico-cortical projections across visual cortices⁴ (Fig.3A). Similarly, measuring the time lags of the maximum peaks in the CCG of recorded neuron pairs (e.g. as in Fig.1B) can be used to assess a hierarchy of functional connectivity between different cortical areas. Such analysis of functional delays extracted from massive recording of spiking activity across visual cortices¹⁵ revealed strong correlation of area-averaged functional delays and the hierarchical score values indicative of the hierarchical processing of visual information. Classical CCG (as well as GLMCC), however, recovers static network structure by modeling pairwise correlations and hence must be run separately for each pair in the dataset without considering the relative influence of other pairs in the network. Each neuron, however, is connected to thousands of others and is potentially contributing to multiple networks simultaneously. In contrast, DyNetCP learns connections of an entire network jointly in one pass and can track multiple connections beyond pair-wise. It is, therefore, instructional to compare the hierarchical interactions between visual cortices inferred by DyNetCP to CCG results using the same large-scale in-vivo VCN dataset¹⁵. Following the previous methodology^15,19, functional delays corresponding to the time lags of jitter-corrected CCG peaks are determined for all recorded pairs across all six cortical areas with peak amplitudes above α_w of 7SD (5023 pairs, n=26 animals). The median time lags averaged across pairs within a given visual area (Fig.3B) are identified as the area’s functional delay. When plotted against the hierarchical score difference (Fig.3C, red triangles), the area median functional delays exhibit strong correlation (r_p=0.76 and P_p<0.0001).

Figure 3.

The DyNetCP trained on the same 5023 pairs with W_st peak amplitudes above α_w of 7SD, whose shape is statistically indistinguishable (typical r_p>0.9 and P_p<0.0001) from the corresponding jitter-corrected CCG, produce similar functional delays (Fig.3D). The median functional delays predicted by DyNetCP by learning from all units (Fig.3C, blue triangles) also show strong correlation (r_p=0.73, P_p<0.0001) with the visual area anatomical hierarchical score difference⁵, very closely replicating (r_p=0.99, P_p<0.0001) the correlation observed with the descriptive pairwise CCG-based analysis. Therefore, the static DyNetCP is able to confirm previous observation of hierarchical organization of inter-area functional delays¹⁵.

Comparison of inferred dynamic connectivity to descriptive statistical methods

Once the static connectivity W_st is established and verified, the dynamic stage of DyNetCP learns W_dyn that is strongly temporally modulated (Video 1). While no GT data are currently available to verify the model performance, model predictions can be compared to a classical pairwise Joint Peristimulus Time Histogram³⁹ (JPSTH) that represents the joint correlations of spiking between two neurons and their variation across time. Specifically, given trial-aligned data with a trial length of T bins, the JPSTH is a T × T matrix where each entry (a, b) represents the percent of trials where neuron i spiked at time a and neuron j spiked at time b. For an example pair of excitatory spatially close units located in Layer 5 of the VISp area (session 766640955) corresponding JPSTH is shown in Fig.4A. Averaging over the columns of the JPSTH recovers the peristimulus time histogram (PSTH) of neuron i (left panel in Fig.3A), while averaging over the rows of the JPSTH recovers the PSTH for neuron j (bottom panel in Fig.3A). Since DyNetCP learns the conditional probability of the spiking of one neuron conditioned on others, we cannot directly compare dynamic weights to a standard JPSTH, which models joint correlations. Instead, for a proper comparison, we use a conditional joint peristimulus time histogram (cJPSTH), where we divide the columns of the JPSTH by the PSTH of neuron i (Fig.4B). This changes the meaning of each entry (a, b) in the cJPSTH that represents the conditional probability that neuron j spikes at time b, conditioned on neuron i having spiked at time a (Fig.4C).

Figure 4.

Fig.4 – figure supplement 1.

The dynamic weights W_dyn matrix inferred by DyNetCP for the same example pair (Fig.4D, see also (Video 1) reveal strong temporal modulation peaking at around 70ms, 220ms, 300ms, and 480ms indicating increased connection strength. Comparison with the cJPSTH matrix (Fig.4E) shows similar temporal variation. This comparison can be used to validate the choice of an optimal squared L2 penalty that is introduced into the model loss function (Eq.6) to compensate the sparseness and irregularity of spiking often observed in cortical recordings (Methods). For a given set of thresholds α_w, α_s, and α_p, scanning of L2 from 0.01 to 10 (Fig.4 – figure supplement 1A) demonstrates statistically significant correlation (r_p>0.77 and P_p<0.001) of the dynamic weights W_dyn matrices and cJPSTH matrices observed for all L2 values with the best results obtained for L2 between 0.1 and 1 (Fig.4 – figure supplement 1B). The two-sided paired equivalence test (TOST) (Fig.3 – figure supplement 1C) confirms that the average difference between normalized weights inferred from both models is smaller than 0.09. Therefore, the DyNetCP reliably recovers the time-dependent dynamic weights W_dyn even in the presence of strong nonstationary spiking irregularities and sparse spiking.

Model enables discovery of cellular-level dynamic connectivity patterns in visual cortex

The model architecture (Fig.1A) designed with two interacting static and dynamic parts enables justification of the choice of model parameters and verification of model inference results by comparison with the results of descriptive statistical methods. Therefore, even in the absence of experimental or synthetic datasets with known ground truth (GT) of cellular-level dynamic connectivity, the model can provide an otherwise unattainable insights into dynamic flows of information in cortical networks. To ensure statistical significance of inferred connections three thresholds (Methods) were applied simultaneously to each of the 26 sessions in the VCN database: a unit average spike rate α_s>1Hz, a pair spike rate threshold α_p>4800, and a static weight peak threshold α_w>7SD that produced 808 valid units forming 1020 valid pairs across all 6 cortical areas (Methods, Table 1).

Reconstruction of dynamic connectivity in a local circuit

To illustrate reconstruction of static and dynamic connectivity in a local circuit we use recording from a single animal (session 819701982) from a single Neuropixel probe implanted in the VISam area. Application of three thresholds α_s, α_p, and α_w to all clean units passing quality metrics (Methods) produce 6 pairs with statistically significant (α_w>7SD) and narrow (<2ms) peak in a static weight W_st (Methods, Table 2) indicating strong putative connections (Fig.5A). Units are classified as excitatory (waveform duration longer than 0.4ms) or inhibitory (less than 0.4ms) and are assigned to cortical depth and cortical layers based on their spatial coordinates (Methods). Directionality of the connection is defined by negative functional time lag in a static weight W_st (or jitter corrected CCG) binned at 1ms. Local circuit reconstructed from static weights matrices is shown in Fig.5B. Corresponding dynamic matrices W_dyn (Fig.5C) for the same pairs demonstrate strong temporal modulation (see also Video 2). Prior to presentation of a stimulus (t<0ms) the inhibitory drives (negative values of W_dyn marked by deep blue color in Fig.5C) prevail. The first positive peak of W_dyn around 70ms (positive values of W_dyn marked by bright yellow color in Fig.5C) is shared by all but one pair. Such synchronous spiking at short time lags does not necessarily indicate a TP direct connection but can also be an FP artefact induced by arrival of afferent signals⁴⁴. However, at later times, additional strong peaks appear for some of the pairs that are not accompanied by corresponding increase of the firing rates. Some of the pairs form an indirect serial chain connections (e.g., 3→2 and 2→4 Fig.5B) that can also generate confounding FP correlations that is notoriously difficult to identify using descriptive statistical models²⁴. However, the DyNetCP being a probabilistic model might lessen the impact of both the common input and the serial chain FP artefacts by abstracting out such co-variates that do not contribute significantly to a prediction. For example, the W_dyn for a potential 3→4 artefactual connection does not show peaks expected from the serial chain.

Figure 5.

Table. 2.

To aid identification of such confounding FP connections, we performed preliminary tests on a synthetic dataset that contain a number of common input and serial chain triplets and were able to identify them by comparing W_dyn values at a given trial time to corresponding W_st (Methods). Figure 6 shows temporal evolution of normalized dynamic weight values taken from Fig.5C along the trial time axis at corresponding functional time lags. We assign excitatory (inhibitory) connection to a given pair as present when the ratio W_dyn⁄W_st is larger (or smaller) than 1 (−1). Corresponding dynamic graphs for a local circuit of Fig.5 are shown in Fig.6B for characteristic trial times. Such dynamic circuit reconstruction is aided by observation of the troughs (inhibitory connection) and peaks (excitatory connection) in the W_dyn cross-sections along the lag time axis taken at a given trial time (Fig.6 – figure supplement 1, see also Video 2).

Fig. 6.

Figure 6 - figure supplement 1.

Note, that since DyNetCP, as opposed to population dynamic models^26–29, learns latent variables assigned to individual units, the anatomical and cell-type specific information in static graph of Fig.5B is preserved. As in Fig.5B, units in the circuit of Fig.6B are ordered with respect to their cortical depth, assigned to specific cortical layers, classified as inhibitory (excitatory) based on their waveform duration. Further analysis of optotagging⁴⁵ (Methods) was used to identify that none of inhibitory units in this local circuit belong to a Vasoactive Intestinal Polypeptide (Vip) cell class. DyNetCP can also be trained separately on trials when animal was running, and animal was at rest (Methods, Table 1) to explore the role of brain state on dynamic connectivity patterns. Therefore, not only DyNetCP can provide static connections in this local circuit, but, more importantly, can also track the time epochs when individual cellular-level connections are activated or inhibited that are time-aligned to the stimulus presentation, brain state, and behavior.

Dynamic connectivity in VISp primary visual area

Going beyond local circuits, the dynamic analysis of information flows during stimulus presentation can be extended to area-wide dynamic maps. However, with the aggressive choice of threshold parameters (α_s>1Hz, α_p>4800, α_w>7SD) the average number of valid pairs per session (Methods, Table 1) is relatively low to generalize (mean:39 ± 4 SEM). To analyze area-wide connectivity, an example of cellular-level dynamic interactions in primary visual area VISp learned from all 26 sessions is shown in Video 3. Still frames from this video are presented in Fig.7A for a number of characteristic trial times. Units are ordered by cortical depth and assigned to cortical layers. Circles represent the magnitude of a dynamic weight W_dyn taken at corresponding functional time lag. The circles diameter is proportional to the dynamic offset W_dyn − W_st. The dynamic weights are changing over the time of the trial (Video 3) and exhibit interactions between excitatory cells (red circles), inhibitory cells (blue circles) and mixed drives (magenta circles). While excitatory-excitatory connections are mostly activated within layer 4 and layer 5, connections involving inhibitory cells dominate inter-layer interactions between all the layers. Feedforward extra-laminar interactions (e.g., input layer 4 is driving supergranular layer 2/3 and infragranular deep layers 5 and 6) reveal formation of neuronal ensembles across all the cortical depth. Feedback interactions are also distributed across the cortical depth with nearly the same maximum weights. These observations are inconsistent with static connection efficacy extracted from CCG peaks recorded in V1 columns in anesthetized monkeys⁴⁶ that indicate the efficacy greatly diminished within <200μm from the source. It is consistent, however, with static spike transmission strength extracted from CCG peaks in V1 columns in mice⁴⁷ that reveal much strong increase of long-distance correlations between layer 2/3 and layer 5 during waking in comparison to sleep states. In general, not only brain state can modulate the connection strength, but most importantly, as shown in Video 3 and Fig.7A, the connection strength between all the layers is strongly temporally modulated during the stimulus presentation. Therefore, our results demonstrate that tracking dynamic connectivity patterns in local cortical networks can be invaluable in assessing circuit-level dynamic network organization.

Figure 7.

Dynamic connectivity along the hierarchy axis of all visual cortices

Similarly, cellular-level dynamic interactions at the cortex-wide level for all 6 cortical areas learned from all 26 sessions are shown in Video 4. Still frames from this Video for the same characteristic times are shown in Fig.7B. Inter-areal long-range interactions across various visual cortices (off-diagonal terms in Fig.7B) exhibit complex temporal evolution with rich feedforward (along the anatomical hierarchical score difference⁵) and feedback interactions. Therefore, DyNetCP can uncover the dynamics of feedforward- and feedback-dominated epochs during sensory stimulation, as well as behavioral states (e.g., epochs in the VCN dataset when animals are running or sitting), however, as opposed to population decoding models^26–29, produce cellular-level cell-type specific connectivity patterns that are otherwise hidden from analysis.

Discussion

DyNetCP can reveal cellular-level dynamic interactions within local circuits, within a single brain area, as well as across global hierarchically organized information pathways. Such detailed information, although present in the original datasets, is otherwise hidden from usual analysis that typically relies on time-averaged spike rate coding or population-averaged coding.

Biological interpretation of the extracted dynamic weights can follow the terminology of the short-term plasticity between synaptically connected neurons^25,33–37 or spike transmission strength^30–32,47. Alternatively, temporal changes in connection weights can be interpreted in terms of dynamically reconfigurable functional interactions of cortical networks^8–11,13,46 through which the information is flowing. We could also not exclude interpretation that combines both ideas. In any event our goal here is to extract these signals for a pair (video1, Fig.4), a cortical local circuit (Video 2, Fig.5), and for the whole visual cortical network (Videos 3, 4 and Fig.7). We envision that the next step here, falling beyond the scope of the current paper, would be to correlate the time dependence of the extracted dynamic weights with the stimulus presentation and the dynamics of the behavioral parameters during the decision-making process to infer where and how the information is flowing in the timescale of a decision making.

It is instructional to assess whether the temporal cellular-level interactions uncovered by DyNetCP are bringing new latent dimensions to the analysis of neural code. Dimensionality reduction methods^18,43,48 are often used to estimate the number of independent dimensions of the code that can better explain the variability of spike trains. When principal component (PC) dimensionality reduction²⁹ is applied to the dynamic weights W_dyn of all valid pairs (i.e., passed through all 3 thresholds) in a single animal (65 units, 105 pairs, session 831882777), up to 33 components are required to account for 95% of the variance (Fig.8A, black). In contrast, only 5 PC are sufficient to explain 95% variance of spike rates for the same units (Fig.8A, red). If all clean units in all visual cortices are taken (264 VIS units) the dimensionality increases, but only slightly to 8 PC (Fig.8A, blue). The normalized intrinsic dimensionality (NID) estimated as a number of major PC components divided by the number of units is between 0.03 and 0.08 for spike rates exhibiting striking simplicity consistent with an often observed trend⁴⁹. However, in contrast, the NID for dynamic weights is as large as 0.5 showing dramatic increase of dimensionality.

Figure 8.

This result holds for each of the 26 sessions with the NID in the spiking rate space for valid pairs (Fig.8B, green, mean:0.20±0.018 SEM) and for all VIS units (Fig.8B, violet, mean:0.08±0.01 SEM). In contrast, much larger NID is observed in the space of dynamic weights (Fig.8B, orange, mean:0.55±0.02 SEM) indicating that the dimensionality of a dynamic network manifold is becoming comparable to the number of recorded units. Hence the observed complexity stems from a relatively small subpopulation (valid units in pairs mean: 42.7±3.3 SEM) comprising just 20% of the total (all clean VIS units mean: 194.9±11.9 SEM) and at relatively short time scales below 30ms when it survives temporal Gaussian kernel smoothing (inset in Fig.8A). DyNetCP, therefore, reveals the dynamically reconfigurable functional architecture of cortical networks that is otherwise hidden from analysis.

In summary, DyNetCP provides a fully reproducible and generalizable processing framework that allows investigators to explore often overlooked dimensions of the neural code that are potentially rich for novel biological insights. We believe that our approach can help to provide a conceptual link between brain-wide network dynamics studied with neuroimaging methods and modern electrophysiology approaches resolving such dynamics at a cellular level.

Methods

1. Model implementation and training

Computational resources

The DyNetCP model can be trained in a single step to produce both static and dynamic weight matrices simultaneously. Alternatively, static, and dynamic training can be done sequentially to minimize the computational resources and to enable intermediate checking of the model parameters with respect to descriptive statistical results. In this case, prior to implementing the dynamic part of DyNetCP, the pairs are classified as putative connections based on a statistical analysis (see below) of the static weights, and the dynamic component is trained only on these pairs.

Computation time and memory consumption can also be optimized by the choice of the bin size d at the cost of temporal precision of the weight prediction. To better identify putative connections and to ease comparison with descriptive CCG models, we use a bin size of 1ms for training of static weights. For training the dynamic part, we use spiking data binned at a 5ms resolution. Evaluation of statistical significance of the model predictions (see below) indicates that a 5ms bin size strikes a good balance between computational efficiency, temporal resolution, and data sparsity.

All models are trained using the ADAM optimizer⁵⁰, a batch size of 16, and a learning rate of 0.0001 for 2000 epochs, where the learning rate is reduced by a factor of 10 after 1000 epochs. For each dataset, we set aside some of the data for validation. We use the validation data for early stopping: after every epoch of training, we compute the loss on the validation data, and terminate training early if the validation loss has not improved for a specific number of epochs (100 when training the static portion and 50 when training the dynamic portion). Note, to properly assess generalization, all plots in this work show DyNetCP static and dynamic weights outputs from the validation data.

When training the static part of the model on the synthetic HH dataset²⁵, we use λ = 1. For training the static part on the VCN dataset¹⁵ we use λ = 100, which we found to lead to a lower residual when comparing to CCG values. For training the dynamic weights for all models we use λ = 10.

Averaging Dynamic Weights

To extract a strong signal from the DyNetCP dynamic weights, we need to aggregate information over experimental trials. However, we do not want to simply average the weights over trials, as all weights which are not paired with an incoming spike (i.e., all weights such that ) do not contribute to the spike rate prediction and therefore are not constrained to capture the spiking correlations. As a result, when computing trial-aggregated DyNetCP weights, we average over weights that are time-aligned with spike times of corresponding unit (Fig.9A) as follows:

Fig. 9.

Additionally, to compensate for the sparsity of spiking, we average these aggregated dynamic weights over the 10 most recent bins (Fig.9B).

Jitter correction

Jittering method ^{15,30,41–43} is often used in analysis of pairwise spiking correlations to minimize the influence of a potential stimulus-locked joint variation in population spiking that often occurs due to common input from many shared presynaptic neurons. To lessen the impact of such false positive (FP) connections in DyNetCP results and to ease the comparison with classical jitter-corrected CCG, we train a second static model on a jitter-corrected version of the data generated using a pattern jitter algorithm⁴¹ with a jitter window of 25ms. When the dataset is trial-aligned (e.g. VCN dataset), we use the correction procedure⁴¹ which determines the probability of a spike being present in each bin for all possible random shuffles of spikes among trials. For data which is not trial-aligned (e.g. synthetic dataset of HH neurons²⁵), we randomly permute the spike times within the interval [−5ms, +5ms] before every weight update. New spike resampling is performed during every epoch of training. To assess putative connectivity, the jitter-corrected static weights w’_j→i produced by a jitter-corrected model are subtracted from the weights w_j→i trained on the original data to obtain and .

2. Datasets

Synthetic network of HH neurons

We use a synthetic dataset consisting of Hodgkin-Huxley (HH) 800 excitatory and 200 inhibitory neurons with known ground truth (GT) synaptic connections²⁵. The excitatory neurons influence 12.5% and the inhibitory neurons influence 25% of other neurons. We use the first 4320s of the 5400s-long data for training and the remaining 1080s for validation. Since DyNetCP requires training with trial-aligned data, we split each of these datasets into 1s “trials” consisting of 1 bins each. For training and evaluation we use spikes from the same subset of 20 excitatory neurons as was used for GLMCC validation²⁵ (Fig.2.C,D,E).

Synthetic dataset with FP connections

Synchronous spiking at short time lags does not necessarily indicate a direct connection but can also be induced by a shared common input (Fig.10A) or an indirect serial chain connection (Fig.10B) among others. While it is notoriously difficult to distinguish these confounding correlations using descriptive statistical models²⁴, the DyNetCP being a probabilistic model might lessen their impact by abstracting out such co-variates that do not contribute significantly to a prediction. To test the ability of DyNetCP to identify such confounds, we generate a synthetic dataset with known GT static and dynamic connectivity. The dataset consists of 100 excitatory neurons with their spikes generated by independent sampling from a Bernoulli distribution with spike rate between 0.002 and 0.01 per bin. The GT connection is active when spikes in neuron n₁ at time t are followed by spikes in neuron n₂ at a time t + d with a probability of 0.6. Randomly generated 1 “triplets” emulating common input and serial chain connections are added to the pool. he “triplets” are active either during the first and last 1 5 bin period or only during the middle 5 bin period. DyNetCP trained on this dataset (400 train and 200 validation trials, each 500 bins long) enables comparison of static and dynamic weights (Fig.10) that indicate that when the GT connection is present the W_dyn tends to be larger than W_st, which enables classification of connections as TP and FP for both common input (Fig.10A) and serial chain (Fig.10B) connections for this synthetic dataset.

Fig. 10.

Experimental dataset recorded in visual cortex

To demonstrate DyNetCP performance with in-vivo high-density electrophysiological recordings, we used the publicly available Allen Brain Observatory-Visual Coding Neuropixels (VCN) database¹⁵ that contains spiking activity collected across 99,180 units from 6 different visual cortical areas in 58 awake mice viewing diverse visual stimuli. We use a “functional connectivity” subset of these recordings corresponding to the “drifting grating 5 repeats” setting recorded for animals (26 sessions in Table 1). Full-field drifting grating stimuli was passively viewed by a head-fixed animal freely running on a treadmill¹⁵. Stimuli consists of 2s presentation of drifting diffraction grating (temporal frequency 2 cycles/s, spatial frequency = 0.04 cycles per degree) that is interleaved with a 1s grey period (top left inset in Fig.1A). Presentation of gratings were repeated 75 times per orientation (0, 45, 90, and 135 degrees) and per contrast (0.8 and 0.4). To maximize the number of spikes available for training we use all angles and all contrast trials.

We split the 600 total experimental trials for each session into 480 training and 120 validation trials by selecting 60 training trials from each configuration of contrast or grating angle and 15 for validation. To further investigate the influence of brain state on connectivity dynamics we identify in each session trials with the average running speed above 5cm/s that are mar ed as “running”. he rest of the trials are mar ed as “not running”. When trained and validated, the model results can then be analyzed and compared for “running” vs “non-running” conditions separately.

Models used

GLMCC model

Python code for GLMCC method is taken from the author’s version⁵¹. For these experiments, we use the synthetic dataset of interconnected HH neurons²⁵. We focus specifically on recovering ground-truth (GT) excitatory connections. To facilitate a comparison with GLMCC model²⁵, we here choose thresholds α_w, α_s, and α_p to maximize the Matthews Correlation Coefficient (MCC, Eq.9). MCC optimization balances between maximizing the number of TP pairs recovered while minimizing the number of FP pairs which are predicted.

Since a predicted connection is counted as a TP only once (i.e., i → j and j→ i do not count as two negatives if i is not connected to j) and the number of GT connections decreases when α_p increases, the MCC exhibits an increase beyond α_p=100. A more detailed analysis of TP and FP rates as a function of pair spike threshold is presented in Fig.2 - figure supplement 1.

Though GLMCC performs well in recovering static network structure, we emphasize that GLMCC is not compatible with modeling network dynamics and thus cannot be integrated into DyNetCP. Specifically, our goal is to model the correlations of spiking of a network of neurons and track the changes of these correlations across time. GLMCC learns correlations as a function of the time lag relative to the spiking time of a reference neuron. It is hence unable to link correlations to specific moments in a trial time. Therefore, it cannot represent the dynamics of correlations, preventing its use in analysis of network activity relative to subject behavior.

Conditional Joint Peristimulus Time Histogram (cJPSTH)

The Joint Peristimulus Time Histogram (JPSTH) is a classical tool³⁹ used to analyze the joint correlations of spiking behavior between two neurons and their variation across time. For a proper comparison with conditional DyNetCP results, we use a conditional joint peristimulus time histogram (cJPSTH), where we divide the columns of the JPSTH by the PSTH of neuron i (Fig.4B,C). Due to the sparsity of spiking in the data we use, it is beneficial to average the entries of the cJPSTH over short windows in time to more clearly extract the dynamics of the correlations between neurons. Specifically, we replace each entry (a, b) in the cJPSTH by the average of the T most recent bins having the same delay a − b, that is,

For most of the comparisons we use T=10 bins. Note that for most cJPSTH plots (e.g., Fig.4E), to directly compare to DyNetCP plots (e.g., Fig.4D), we only show data representing the “upper diagonal” of the full plot which represents data where neuron i spikes at the same time or before neuron j spikes. Specific implementation and comparisons of both models can be found in the accompanying Jupyter Notebook.

Dimensionality reduction

Dimensionality reduction^18,43,48 is based on calculating principal components (PC) on the dataset and estimating the data variance that can be explained by a minimal number of PC. The dimensionality reduction code following a standard PCA reduction method⁴⁸ is rewritten in Python and is used for analysis in Fig. 8A,B.

4. Statistical analysis of filtering and classification of units and pairs

The total number of units (raw units) recorded from visual cortices in 26 sessions of the VCN dataset¹⁵ is 18168. Units and pairs are filtered and classified by the following procedure.

Unit quality metrics

First, we apply standard unit quality metrics¹⁵ including amplitude cutoff < 0.1; violations of inter-spike interval (ISI) < 0.1; presence ratio > 0.95; isolation distance > 30; and signal to noise ratio (snr) > 1. This results in 5067 units (clean units).

Static weight peak threshold

Second, short-latency peaks within the first 10ms in the w^∗_j→i and w^∗_i→j of static weights are examined (example analysis for session 771160300 is shown Fig.2A,B). Following the accepted approach^{15,19,21,25,30} a putative connection is considered statistically significant when the magnitude of the peak exceeds a certain peak weight threshold α_w measured in units of standard deviations (SD) of the noise in the flanks, calculated within [51ms, 100ms] from both directions w^∗_i→j and w^∗_i→j. When pairs are filtered with increasing peak weight threshold α_w (Fig.11A), the number of corresponding valid units is fast decreasing (black curve). For all 26 sessions 5067 clean units participate in formation of 5203 pairs with noticeable weight peaks. Filtering them using α_w=7SD results in 1680 pairs corresponding to 1102 units. Close examination of these pairs (example analysis for session 771160300 is shown in Fig.11A) reveals that units not only form pairs (violet curve), but also form triplets (green), quintuplets (blue), and even participate in connections above 10-tuples (red). High radix hubs of this type are most likely FP artefacts imposed by sparse spiking (see below) and need to be filtered out. Indeed, electrophysiology spiking datasets are typically very sparse, with many interesting units often generating spikes at rates below 1Hz. Increasing the threshold α_w does help to filter out high radix connections, but still even for α_w as large as 8 there is a substantial number of units participating in connections larger than 3.

Unit spike rate

Therefore, as a third filtering step, we apply a unit spike rate threshold α_s to limit the minimal spike rate to 1Hz.

Pair spike rate threshold

It has been argued²⁵ that, to reliably infer that a connection is present (i.e. to disprove the null hypothesis that a connection is absent), each binned time interval in the CCG histogram d should contain a sufficiently large number of spikes from both participating units. This implies that the spike rate of each neuron in a pair λ_i and λ_j should exceed the threshold α_s, and the joint number of spikes in each time bin should exceed the pair spike threshold, defined as

The confidence interval for a null hypothesis that two spike trains are independent stationary Poisson processes²⁵ is then defined as

To disprove the null hypothesis and infer that a connection is statistically significant, α_p should exceed²⁵ 10. To train the dynamic DyNetCP we use recordings binned at d of 5ms repeated over M of 480 randomized trials and the dynamic weights are influenced by T=10 most recent bins (Fig.9B). This corresponds to the TMd product of 24 that implies the minimum joint pair spike rate λ_iλ_j of 0.41s⁻². This threshold, however, might be insufficient for inference of the dynamic weight since only spikes within a short trial time interval are being used for learning. Indeed, analysis of an example pair (Fig.12A) that passed the α_w=7SD threshold and has λ_iλ_j = 0.73s⁻², reveals that it exhibits on average just 2 spikes per trial or a total of 272 spikes for all 480 trials in a 500ms trial time window of interest. Corresponding jitter-corrected static weight matrix (Fig.12A, left) and cJPSTH (Fig.12A, bottom right) are very sparsely populated. The dynamic weight matrix (right top), although it is reproducing this sparseness, is not statistically convincing. Moreover, such pairs contribute significantly to high radix tuples.

By increasing the λ_iλ_j pair spike rate threshold, the number of high radix tuples is decreasing (example analysis for session 771160300 is shown in Fig.11B), however the α_p threshold alone is not sufficient to eliminate them. Hence, for a given pair to be classified as a putative connection, the classification procedure requires the use of all three thresholds: α_w, α_s, and α_p (see Jupyter Notebook for specific implementation).

Fig. 11.

Choice of thresholds

When all three thresholds α_s, α_w and α_p are applied simultaneously (Fig.11C) the number of valid units is decreasing rapidly with the increase of α_w to 7SD and λ_iλ_j beyond 10 with slower decrease at higher thresholds. Examination of an example pair with joint pair spiking rate λ_iλ_j of 50 s⁻² (Fig.12B) indicates that while the static weight matrix shows a prominent (α_w >7SD) small latency peak, the cJPSTH is still populated relatively sparsely and the dynamic weight matrix resembles it much less (r_p=0.27). In contrast, for an example pair with joint pair spiking rate λ_iλ_j of 130 s⁻² (Fig.12C) the resemblance between cJPSTH and the dynamic weight matrix is striking (r_p=0.82, P_p<0.001).

Fig. 12.

For a fixed α_s of 1Hz and α_w of 7SD the distribution of all valid pairs with respect to their joint spike rates (Fig.13A) shows a maximum around λ_iλ_j=200 s⁻² with a long tail to higher rates. To be conservative, for all the analysis of the inferred connectivity patterns in this work, we used λ_iλ_j of 200 s⁻² (α_p=4800, red arrow in Fig.13A).

Fig. 13.

Application of all 3 filters to all 26 sessions produces 808 units forming 1020 pairs (Table 1) with distribution of inferred static and dynamic weights shown in Fig.13B.

Units’ classification

Once all pairs are filtered, the directionality is assigned to each putative connection, either j → i or i → j, that is identified by the functional delay of the static weight peak binned at 1ms. Cortical depth perpendicular to cortical surface is calculated from unit coordinates in ccfv3 format (Allen Mouse Brain Common Coordinate Framework ⁵²). Layer assignment is performed following method¹⁵ using units coordinates in the ccfv3 framewor and the annotation matri with reference space ey ‘annotation/ccf_ 1’. All units are classified as excitatory if their waveform duration is longer than 0.4ms and the rest as inhibitory (Fig.13D). Since transgenic mice were used for some of the sessions in the VCN dataset, optical activation of light-gated ion channel (in this case, ChR2) in Cre+ neurons with light pulses delivered to the cortical surface generate precisely time-aligned spikes that enable to link spike trains to genetically defined cell classes. Optical tagging was used to further classify inhibitory units⁴⁵ as Parvalbumin-positive neurons (Pvalb), Somatostatin-positive neurons (Sst), and Vasoactive Intestinal Polypeptide neurons (Vip).

Potential biases

The statistical analysis and filtering procedure described above is biased towards strongly spiking units (Fig.13C) and as such can generate certain biases in connectivity patterns. For example, inhibitory units constitute about 18% of all 5067 clean units that pass unit quality metrics (Fig.13D). However, after filtering down to 808 valid units, they constitute over 50% of the population (Fig.13D). In general, inhibitory units have a higher spiking rate (for all clean units mean: 15.4Hz±0.29 SEM) than excitatory (mean: 6.3Hz±0.08 SEM). This might explain, at least partially, the over-representation of connections involving inhibitory units in the dynamic connectivity plots.

Data Availability

All of the datasets used in this article are publicly available online under the Creative Commons Attribution 4.0 International license. Synthetic data of interconnected HH neurons²⁵ is available at figshare⁵³ (https://doi.org/10.6084/m9.figshare.9637904). The Allen Brain Observatory-Visual Coding Neuropixels (VCN) database¹⁵ is available for download in Neurodata Without Borders (NWB) format via the AllenSDK. The NWB files are available as an AWS public dataset (https://registry.opendata.aws/allen-brain-observatory/). Synthetic dataset with artifactual FP connections as well as trained static and dynamic weights for all 26 sessions of VCN database are availabel at Figshare https://doi.org/10.6084/m9.figshare.24879036

Code Availability

DyNetCP introduced and used in this article is open-source software (GNU General Public License v3.0). The source code is available in a GitHub repository (https://github.com/NeuroTechnologies/DyNetCP) that contains codes for model training as well as a Jupyter notebook for generating the results used in the paper and in the SM.