Information ow of spontaneous and auditory-evoked neural activity in the rat thalamo-cortical system

The interaction between the thalamus and sensory cortex plays critical roles in sensory processing. Previous studies have revealed pathway-specic synaptic properties of thalamo-cortical connections. However, few studies to date have investigated how each pathway routes moment-to-moment information. Here, we simultaneously recorded neural activity in the auditory thalamus (or ventral division of the medial geniculate body; MGv) and primary auditory cortex (A1) with a laminar resolution in anesthetized rats. Transfer entropy (TE) was used as an information theoretic measure to operationalize “information ow”. Our analyses conrmed that communication between the thalamus and cortex was strengthened during presentation of auditory stimuli. In the resting state, thalamo-cortical communications almost disappeared, whereas cortico-cortical communications were strengthened. The predominant source of information was the MGv at the onset of stimulus presentation and layer 5 during spontaneous activity. In turn, MGv was the major recipient of information from layer 6. TE suggested that a small but signicant population of MGv-to-A1 pairs was “information-bearing,” whereas A1-to-MGv pairs typically exhibiting small effects played modulatory roles. These results highlight the capability of TE analyses to unlock novel avenues for bridging the gap between well-established anatomical knowledge of canonical microcircuits and physiological correlates via the concept of dynamic information ow.


Introduction
The interaction between the thalamus and cortex is thought to play critical roles in sensory processing [1,2]. Anatomically, the middle cortical layer is the predominant recipient of thalamocortical projections, whereas the deep cortical layer is the source of cortico-thalamic projections [3][4][5][6][7][8][9]. This general structural pattern is observed across different thalamo-cortical systems and mammalian species, and is thus considered a canonical microcircuit in the thalamo-cortical system. These hodological motifs suggest that feedforward pathways originate principally from the supragranular layer (L2/3) and terminate in the granular layer (L4), whereas feedback pathways originate from the infragranular layers (L5/6) and avoid terminating in L4 [10][11][12][13][14][15][16]. Information ow within these anatomical circuits is thought to be dynamic, with moment-to-moment variation in active pathways [17][18][19][20][21]. For example, communication between the thalamus and cortex is expected to be strengthened during stimulus presentation, whereas communication within the cortex is likely to be strengthened in the resting state in the absence of overt sensory processing. Nevertheless, these differences have yet to be characterized comprehensively in the thalamo-cortical sensory system, especially at the level of neuronal spiking.
Beyond layer-based categorization, further subdivisions of thalamo-cortical pathways have been proposed based on synaptic properties, which may delineate speci c roles in information transmission. For example, glutamatergic pathways in the thalamo-cortical system can be classi ed into either Class 1 or Class 2 (previously termed driver or modulator, respectively). Class 1 inputs express ionotropic glutamate receptors and constitute the main information-bearing pathway, whereas Class 2 projections express metabotropic receptors and modulate the transmission of Class 1 inputs [22][23][24]. In the auditory system, Class 1 constitutes the main pathway from the ventral division of the medial geniculate body (MGv) to L4-L6 in the primary auditory cortex (A1). Class 2 projections are observed from the MGv to L2/3 and from L5/6 to the MGv [25,26]. Within the cortex, Class 1 and Class 2 are likely intermingled [27][28][29][30]. However, this synapse-based pathway characterization has yet to be validated by physiological neural recordings paired with information theoretical analyses, which will enable dynamic and quantitative determination of the nature of information ow.
To characterize the electrophysiological responses in the auditory thalamo-cortical system, we previously designed a microelectrode array that enabled simultaneous neural measurements in the MGv and every layer in A1 [31]. In the present study, we used transfer entropy (TE) to characterize pathway-speci c information ow in the MGv-A1 system [32][33][34][35][36]. TE is a metric based on information theory that statistically quanti es the directed in uence between two sets of time-series data.
Here, we rst con rm that information ow during spontaneous activity is distinct to that during stimulusdriven activity. Next, we demonstrate that TE is able to capture feedforward/feedback ow during stimulus-evoked activity, which is consistent with well-established canonical microcircuits in the thalamocortical system. Our ndings provide a robust link between neuroanatomical knowledge of canonical microcircuits and physiological observations via the concept of dynamic information ow.

Animals
This study was performed in strict accordance with the "Guiding Principles for the Care and Use of Animals in the Field of Physiological Science" published by the Japanese Physiological Society, and the recommendations in the ARRIVE guidelines (https://arriveguidelines.org/). The experimental protocol was approved by the Committee on the Ethics of Animal Experiments at the Research Center for Advanced Science and Technology, University of Tokyo (Permit Number: RAC 130107). All surgeries were performed under iso urane anesthesia. All efforts were made to minimize animal suffering. Following the experiments, animals were euthanized with an overdose of pentobarbital sodium (160 mg/kg, i. p.).
Four male Wistar rats were used in this study (11-13 weeks old; body weight, 290-330 g). The protocols for animal preparation and neural recordings have been described elsewhere [20,31,37]. Brie y, the rats were anesthetized with iso urane and air at a concentration of 3% for induction and 1% for maintenance during the surgery and experiments. Animals were held in place with a custom-made head-holding device. Atropine sulfate (0.1 mg/kg) was administered pre-and post-surgery to reduce the viscosity of bronchial secretions. A skin incision was made at the start of surgery under local anesthesia using lidocaine (0.3-0.5 mL). A needle electrode was subcutaneously inserted into the right forepaw and used as the ground. A small craniotomy was performed close to bregma in order to embed a 0.5 mm-thick integrated circuit socket as a reference electrode, with electrical contact to the dura mater. The right temporal muscle, cranium, and dura overlying the auditory cortex were surgically removed. The exposed cortical surface was perfused with saline to prevent desiccation. Cisternal cerebrospinal uid drainage was performed to minimize cerebral edema. The right eardrum, ipsilateral to the exposed cortex, was ruptured and waxed to ensure unilateral sound inputs from the ear contralateral to the exposed cortex. Respiratory rate, heart rate, and hind-paw withdrawal re exes were monitored throughout the surgery to ensure maintenance of stable and su cient anesthesia. For acoustic stimulation, a speaker (Technics EAS-10TH800, Matsushita Electric Industrial Co. Ltd., Japan) was positioned 10 cm from the left ear (contralateral to the exposed cortex). Test stimuli were calibrated at the pinna with a 0.25-inch microphone (4939, Brüel & Kjaer, Denmark) and spectrum analyzer (CF-5210, Ono Sokki Co., Ltd., Japan). Stimulus levels were presented in dB SPL (sound pressure level in decibels with respect to 20 µPa).

Electrophysiology
We used a surface microelectrode array and depth electrode array (NeuroNexus, Ann Arbor, MI, USA) to simultaneously measure neural activity in the auditory cortex and thalamus, as previously described [28] ( Fig. 1a). The surface microelectrode array comprising a 10 × 7 grid within 4 × 3 mm 2 mapped local eld potentials (LFPs) in the right temporal cortex to identify the location of the primary auditory cortex (A1) [37]. The depth microelectrode array was then inserted perpendicular to the cortical surface in A1. The array comprised three shanks (6 mm in length), each of which constituted 15 distal recording sites for MGv and 17 proximal sites for A1. The array simultaneously measured multi-unit activity (MUA) and LFPs from the MGv and A1. The diameter of recording sites was 30 µm. The center-to-center interelectrode distance was 120 µm. The most distal site was placed 100 µm from the tip of the shank, and the distance between the most proximal site in the MGv and the most distal site in A1 was 1,200 µm.
Each electrode was composed of iridium oxide and coated with platinum black.
Neural signals were ampli ed with a gain of 1,000 (Cerebus Data Acquisition System; Cyberkinetics Inc. Salt Lake City, UT, USA) software. The digital lter bandpass was 0.3-500 Hz for LFP and 250-7500 Hz for MUA. The sampling rates for LFPs and MUA were 1,000 Hz and 30 kHz, respectively. Multi-unit spikes were detected online from MUA by threshold-crossing (-5.65 times root mean square of MUA).
Spontaneous activity was rst characterized as MUA in a silent environment for 5 min. Auditory-evoked activity was then characterized in response to clicks and tone bursts. Clicks were presented at a rate of 1 Hz. Tone bursts were used to characterize the characteristic frequency (CF) at each recording site. CF was determined as the frequency at which test tones evoked MUA with the lowest intensity or the largest response at 20 dB SPL (the minimum intensity used in this study). Test frequencies ranged from 1.6 to 6.4 kHz with an increment of 1/3 octaves and intensities from 20 to 80 dB SPL with an increment of 10 dB. Each test tone was repeated 20 times in a pseudorandom order with an inter-tone interval of 600 ms. Recording sites at which CF was identi ed were de ned as either MGv or A1, whereas those at which CF was not identi ed were excluded from further analyses.
For the grand average of 240-trial click-evoked LFPs from the depth array, one-dimensional current source density (CSD) analysis ( Fig. 1b) was conducted, as described previously [28,38,39]. Brie y, twice the potential at a given depth (V 0 ) was subtracted from the sum of the potentials at the upper and lower adjacent sites of a given depth (V u and V l ), and then divided by the square of the distance (Δx) between the recording sites (120 µm): Each layer was de ned based on the CSD results as follows: L4 was rst de ned as the site with the earliest sink and adjacent sites as sinks and no source. L2/3 was de ned as sites above L4 with sinks, followed by short sources. L5 was de ned as two successive sites with sources below L4. Weak sinks were identi ed in deeper sites, of which the second deeper site was de ned as L6.

Transfer entropy
TEs of either thalamo-cortical, cortico-cortical, or cortico-thalamic projections were derived from MUA data of either spontaneous activity or click-evoked activity in a pairwise manner. TE was estimated from MUA data binarized with a bin of 1 ms (Fig. 2a). Bins with spikes were labeled as 1; those without spikes were labeled as 0. None of the bins contained two or more spikes. The TE of Y to X or was de ned as follows: 1 where H(A|B) represents the conditional entropy in information theory, which indicates the unpredictability of A when information on B is known.
estimates how spikes at electrode Y ( ) improve the prediction of spikes at electrode X ( ), beyond the prediction based on past data of X ( ). Here, was calculated as follows: 2 where t, lag, and d represent the time, transfer lag, and delay, respectively, between the future and past. represents the past state of electrode Y ( ). and represent the future and past states, respectively, of electrode X ( and ). The past data of X were obtained from d bins before a given time point of (t + lag), which were optimized as follows, assuming that X t depends predominantly on past X t−d : 3 According to Eq. (1), we quanti ed for given electrode pairs with either a short window (15 ms) or long window (10 s) (Fig. 2b).
(i) Long-window TE with and without stimuli (long-window TEstim and long-window TEspon, respectively) Long-window TE was derived using 10-s windows to assess if information transmission differed depending on the state of the thalamo-cortical system (i.e., during sensory processing vs. resting state). Long-window TEstim was derived from MUA over a continuous period of 240 s, during which clicks were presented every second. Long-window TEspon was derived from a separate 240-s time period of data during which no stimulus was delivered. Ten sets of 10-s and spike trains were randomly selected to derive the joint probability, . Based on Eq. (1), 10 sets of TE were then estimated in the transfer lag ranging between 1 and 30 ms. Long-window TEs were ultimately de ned as the median across 10 sets for each lag.
(ii) Short-window TE Short-window TE was computed using 15-ms windows to characterize information transmission in the thalamo-cortical system during the time window surrounding stimulus onset. The time course of information transmission for short-window TE was investigated using moving window analysis.
For trial i (= 1, …, 240), in response to a click delivered at time s i , spike trains within 15-ms post-stimulus latency were used to derive short-window TE. Based on short-window TE at stimulus onset (short-window TEonset), we rst identi ed signi cant information transmission and the optimal lag of TE for a given electrode pair. For , the joint probability, was obtained to derive at a given lag according to Eq. (2). Shortwindow TEonset was ultimately de ned as the median across 240 trials for each lag.
We next characterized the time-course of short-window TE, i.e., how TE evolved over time in the thalamocortical system during the time window surrounding stimulus onset. We computed the short-window TE for , where T ranged from s i − -10 to s i + 40 and the lag was the optimal value in the short-window TEonset. When t was not an integer, t was rounded off to the nearest integer. The time course of short-window TE was ultimately de ned as the median across 240 trials for each T. The earliest T when TE > 0 after bias correction (see the next section) was de ned as the onset latency of information transmission.

Statistical analyses for identi cation of signi cant information transfer
To identify electrode pairs with signi cant information transfer, we compared the above TEs derived from experimental data with those derived from shu ed data (TE shu ed ). To generate the shu ed data, we randomly shu ed the inter-spike intervals (ISIs) of X t and Y t without changing the ISI distribution.
Shu ing disrupted the temporal structure underlying functional connectivity between X t and Y t .
To assess statistical signi cance of information transfer, we estimated p-values as the rank order of empirically identi ed TE values among the null distributions arising from 100 TE shu ed . For example, if the empirical TE was larger than the top 5% of 100 sets of TE shu ed , we regarded the p-value to be less than 0.05 [38]. We corrected for multiple comparisons across transfer lags (1-30 ms) using the false discovery rate (FDR) method [39]. Further, we de ned a pair of functionally connected electrodes as those with signi cant information transfer within a time window of 5 ms or more (Fig. 3).
When quantifying the amount of information transfer, we considered the degree of positive bias caused by a limited amount of sample data. Theoretically, TE shu ed must become 0 because shu ing should disrupt any causality between X and Y. However, the actual TE shu ed was larger than 0 due to biases, which were removed by subtracting the median TE shu ed from the TE. When TE was smaller than TE shu ed , no information transfer was assumed (i.e., TE = 0).

Normalized TE (nTE)
Mean ring rates of evoked activity was substantially higher than those of spontaneous activity ( Fig. 1c and 1d). To eliminate the bias due to differences in mean ring rate, we introduced the nTE. This normalization was necessary when comparing TEs derived from evoked and spontaneous states with different probability densities as follows: Practically, the bias of nTE was corrected as 4 Information transmission in a given pathway We characterized the information transmission in each pathway as the average of the peaks of nTE among pairs with signi cant information transfer: where n is the number of pairs with signi cant information transfer, and is the number of possible pairs of electrodes.

Role of a given region in information transmission
Based on the average of the nTE peaks de ned above, we quanti ed whether each region (X) served as either a receiver ( ) or a sender ( ). The metrics and were de ned as the summation of the average of the nTE peaks as follows: : one of (MGv, L2/3, L4, L5, and L6) with the exception of where the average of nTE peaks(pathway) is the average of nTE peaks in a given pathway, as de ned in equation (5). We then characterized each region X using the SR ratio: A positive SR ratio indicated that region X served as a sender, whereas a negative SR ratio indicated that region X served as a receiver.

Results
In the four rats tested, 96 sites in the MGv and 138 sites in A1 exhibited tone-evoked MUA, which exhibited de nable auditory responses and a CF. Among these sites, we simultaneously measured clickevoked and spontaneous MUA in the MGv and A1 (Fig. 1c and 1d). We derived long-window TE and shortwindow TE between all possible pairs among available sites. Signi cant information transfer was identi ed in 11483 pairs (98% of all possible pairs) in long-window TEstim, 3964 pairs (36%) in longwindow TEspon, and 5246 pairs (45%) in short-window TEonset. These signi cant pairs were further characterized as follows. Among these pairs, 2430, 107, and 1578 pairs (83%, 4%, and 54% of pairs, respectively) transmitted information from the MGv to A1; 2005, 108, and 743 pairs (69%, 4%, and 25% of pairs, respectively) transmitted information from A1 to MGv; and 2671, 2548, and 1405 pairs (75%, 71%, and 40% of pairs, respectively) transmitted information from A1 to A1 in long-window TEstim, longwindow TEspon, and short-window TEonset, respectively.

Long-window TE
For long-window TE, we compared information ow with and without stimulus inputs. The nTE of (i) longwindow TEstim and (ii) long-window TEspon as a function of transfer lag is presented in Fig. 4a. TEstim decayed abruptly at a transfer lag of approximately 15 ms in the MGv-to-A1 (red) direction, whereas it decayed smoothly in the A1-to-MGv (blue) direction (Fig. 4a-i). TEstim in the feedforward pathway from MGv to A1 and cortico-cortical pathway was larger than that in the feedback pathway from A1 to MGv. TEspon (Fig. 4a-ii) indicated that the forward transmission of information from MGv to A1 almost disappeared during spontaneous activity compared to TEstim. These results support the notion that stimulus-driven information ow is distinct to spontaneous information ow and that each pathway possesses different information ow properties.
We further subdivided the cortical recording sites into four different layers (L2/3, L4, L5, and L6) according to the CSD analysis (Fig. 1b) and characterized information ow as the averages of nTE peaks in 24 pathways (Fig. 4b), as de ned in Eq. (5). To evaluate the consistency of this measure across subjects, the relationships of information ow patterns of 24 pathways were analyzed across subjects. The correlation coe cients for TEstim and TEspon were 0.665 0.185 and 0.954 0.013 (average SD), respectively (t-test: TEstim, P < 0.05 in 5 out of 6 pairs of test rats; TEspon, P < 10 −11 for all pairs). The moderately high correlation coe cient in the presence of stimuli (0.665) and high correlation coe cient during spontaneous activity (0.954) veri ed the ability of our analyses to capture general patterns of information ow.
We next investigated the directionality of information ow. Based on TEstim and TEspon, the thick symbols in Fig. 4c-i and 4c-ii indicate that information ow was signi cantly larger than that in the opposite direction (one-sided Student's t-test; P < 0.05). These pathways are highlighted in the schematic diagram of the thalamocortical system in Fig. 4c. Neural activity originating from L5 spread to other cortical layers and terminated in L2/3 in both TEstim and TEspon. In the presence of stimulation, additional activity originated from the MGv and L4. These trends of information ow were supported by data presented in Fig. 4d, which characterized the role of each region as either a sender or receiver based on the SR ratio as de ned by Eq. (8).

Short-window TE
Based on the above analysis, we conjectured that sender/receiver characteristics may change as a function of time during the temporal window surrounding stimulus onset, and that feedforward and feedback information transmission may be temporally segregated. For each signi cant electrode pair in the short-window TEonset, the time course of short-window TE was derived using moving window analyses (Fig. 2b). We rst compared information transmission in the feedforward (i.e., 1578 pairs from MGv to A1; 54% of all possible pairs) and feedback direction (i.e., 743 pairs from A1 and MGv; 25% of all possible pairs). For both the average of all test pairs (Fig. 5a) and individual test pairs (Fig. 5b), we observed that feedforward information transmission was larger in amount and earlier in latency compared to feedback information transmission.
To test whether feedforward pathways (Fig. 5b-i) were activated earlier than feedback pathways (Fig. 5bii), we quanti ed the onset latency of (i) feedforward and (ii) feedback information transmission in each layer-speci c pathway (Fig. 5c). L4 received feedforward information with the earliest onset from MGv, whereas L2/3 received information with the latest onset (Fig. 5c-i; Kruskal-Wallis followed by Tukey-Kramer test, P < 0.05 for all pairs). In contrast, feedback information transmission from A1 to MGv was initiated in L6 (Fig. 5c-ii). These properties of feedforward information transmission were consistent with previous physiological ndings [10][11][12][13][14][15][16], underscoring the major bene ts of the estimates of feedback transmission in our analyses.
Similar to the analysis for long-window TE, we characterized information transmission based on shortwindow TE in 24 pathways (Fig. 5e). The correlation coe cient of information transmission patterns of the 24 pathways across subjects was 0.629 ± 0.148 (P < 0.05 for all test pairs). The thick symbols in Fig. 5e indicate that information transmission was signi cantly larger than that in the opposite direction (one-sided Student's t-test; P < 0.05). These pathways are highlighted in the schematic diagram of the thalamocortical system in Fig. 5f. Neural activity originated from MGv, L4, and L5, and terminated in L2/3. As depicted in Fig. 5g, MGv served as a sender and L2/3 as a receiver, whereas L4 and L5 served as relay stations with inward and outward transmission of similar magnitude. The difference between longwindow TEstim (Fig. 4c-i) and short-window TE (Fig. 5f), i.e., L5 not always serving as a sender, indicated that the MGv was the origin of stimulus-driven information transmission in the thalamo-cortical system.

Discussion
In this study, we performed TE analysis to characterize information ow in the thalamo-cortical pathway between the MGv and A1 in anesthetized rats. We simultaneously recorded MUA in both the MGv and A1 (Fig. 1), and estimated TE using two different sampling windows (Fig. 2). We employed long-window TE to compare information ow with and without stimulus presentation, and short-window TE to scrutinize feedforward and feedback transmission around the time of stimulus onset. Our analyses were consistent with well-established neuroanatomical literature and demonstrated that information ow was dynamic, with moment-to-moment variability in active pathways depending on the mode of stimulus processing.
Short-window TE revealed that the MGv acted as a sender to L4 and L5 in response to a click stimulus. Long-window TEspon indicated that thalamo-cortical information ow almost disappeared and corticocortical communication became dominant during spontaneous activity. The time-course of short-window TE demonstrated that feedforward thalamo-cortical information ow preceded feedback cortico-thalamic information ow. L4 received the greatest in uence and earliest latency from the MGv in the feedforward direction. These results are in accordance with well-characterized anatomical structures and canonical microcircuits in the thalamo-cortical system, thus con rming the validity of our analyses [3][4][5][6][7][8][9][10][11][12][13][14][15].
Our analyses revealed several differences in aspects of information ow between the MGv and A1. First, the information transfer window was approximately 15 ms for MGv-to-A1 transmission, which was narrower than that in the A1-to-MGv direction (Fig. 4a-i). This order of transfer lags for information transmission to A1 is consistent with previous ndings [40]. Such transfer lags are substantially longer than the cortical synaptic delay of 2 ms [41,42] and are therefore likely generated within abundant recurrent connections in A1, but not in MGv [11-13, 43, 44]. Further, the order of the time window for integration is reminiscent of a cycle of gamma-band oscillations, which are generated via the interaction between pyramidal and inhibitory interneurons [45][46][47] and subserve information integration [48][49][50][51][52]. Second, the in uence of A1-to-MGv nodes was typically small (Fig. 5d-ii), whereas the in uence of MGvto-A1 nodes varied considerably (Fig. 5d-i). Furthermore, the outliers in Fig. 5d-i imply that a small proportion of MGv-to-A1 information transmission was signi cantly higher than the average value, indicative of high e ciency in driving post-synaptic neurons. These information-bearing nodes are likely to be classi ed as Class 1 projections, which express glutamatergic ionotropic receptors. In contrast, other nodes may be classi ed as Class 2 projections, which express metabotropic receptors [22][23][24][25][26][27][28][29][30]. The MGv may comprise more Class 1 pathways compared to A1, enabling transfer of external stimulus information (outliers in Fig. 5d-i vs. 5d-ii).
Our analyses demonstrated that communication between the thalamus and cortex is strengthened during stimulus presentation, whereas communication within the cortex is strengthened during spontaneous activity. Furthermore, differences and similarities between long-window TEstim and long-window TEspon provide critical insight into cortical computational processes. For example, our results are consistent with past reports in that the major source of information ow during spontaneous activity likely originated from L5 [17,53]. L5 is more likely to serve as the source of spontaneous activity in A1 because L5 exhibits more depolarized membrane potentials and higher ring rates compared to other layers [44,54,55] with less inhibition [56,57]. During both spontaneous and evoked activity, L2/3 constituted more pathways for information in ow than for information out ow. These patterns of information ow suggest that L2/3 has a higher-dimensional space of representation compared to L4, corroborating the conceptual framework of sparse coding formation in L2/3 from high activity in L4 [17,44,54,58,59].
Our pairwise estimation of information ow has several limitations, which may complicate the interpretation of our results. Although monosynaptic connections in L5 were previously identi ed in 0.25% of test pairs [41], signi cant information ow was observed for 25-49% of possible test pairs in our analyses, suggesting that TE based on MUA is distinct to monosynaptic connectivity. Furthermore, false-positive TE may have been obtained for a pair of nodes which both receive projections from a common origin [33,34,60]. For example, a proportion of information ow between L4 and L5 may have been false positives, because both L4 and L5 receive dense projections from the MGv [3, 4, 7, 8, 11-13, 15, 26, 61-63]. This false positive information ow may be more frequently observed in the L4-to-L5 direction than in the opposite direction because click-evoked responses occur earlier in L4 than in L5. To overcome these limitations, conditional mutual information methods such as momentary TE [64] should be employed to estimate direct causality by conditioning out the effects of possible common drivers. Other alternatives to reduce the effects of common drivers exist [65-68]. Nevertheless, these techniques share the issue of estimation of neural interactions when the number of nodes in the analysis is large. There are currently no known techniques to address the problem of combinatorial explosion.
In conclusion, we simultaneously measured MUA in the MGv and A1 in rats and harnessed TE analyses to characterize information ow in the auditory thalamo-cortical system. Long-window TE revealed that communication between MGv and A1 was strengthened during stimulus presentation, whereas thalamocortical communications almost disappeared and cortico-cortical communications were strengthened during spontaneous activity. Short-window TE indicated that feedforward (thalamo-cortical) information transmission was followed by feedback (cortico-thalamic) transmission at stimulus onset, and L4 exerted the largest in uence with the earliest latency from the MGv in the feedforward direction, corroborating anatomical reports on thalamo-cortical projections. Furthermore, consistent with the notion of Class 1 and Class 2 synaptic properties, a small but signi cant population of MGv-to-A1 pairs was informationbearing, whereas A1-to-MGv pairs that typically exhibited a small in uence were likely to play modulatory roles. Our results highlight the capability of TE analyses to unlock novel avenues for bridging the gap between well-established anatomical knowledge of canonical microcircuits and physiological ndings via the concept of dynamic information ow.

Declarations Data Availability Statement
The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.   Identi cation of signi cant information transfer. We shu ed the raw spike data to generate 100 datasets without changing inter-spike interval (ISI) distribution (a). Transfer entropy (TE) values were derived from either raw spike data (red line; TEraw) or shu ed data (blue dots; TEshu ed) at every transfer delay (b).
Signi cant information transfer was de ned as TEraw exceeding the top 5% of 100 TEshu ed (black dotted lines). An electrode pair was de ned as having functional connectivity if signi cant information transfer was observed for ve successive 5-ms transfer delays.

Figure 4
Long-window TE differentiates information ow with and without stimulus presentation. (a) Normalized TE (nTE) as a function of transfer lag: (i) nTE during stimulus presentation (long-window TEstim) and (ii) nTE during spontaneous activity (long-window TEspon). Average and standard error across subjects are presented. Colors indicate pathways: red, MGv-to-A1; blue, A1-to-MGv; green, A1-to-A1 pathways. (b) Information ow in a given pathway. Recording sites in A1 were classi ed into Layers 2/3, 4, 5, and 6.
TEstim and TEspon were characterized in four thalamo-cortical pathways, four cortico-thalamic pathways, and 16 cortico-cortical pathways between cortical nodes (L2/3, L4, L5, and L6) and nodes in the MGv. Information ow in each pathway was quanti ed as the average of nTE peaks (see text for details). The thick symbols indicate that information ow was signi cantly larger than that in the opposite pathway (one-sided Student's t-test, P < 0.05). (c) Schematic diagram of information ow.
Signi cant directional in uences in (b) are depicted. (d) Sender/receiver (SR) ratio at each region. SR ratios were derived for TEstim and TEspon. Regions with a positive SR ratio served as senders, whereas those with a negative ratio served as a receiver.