JNEUROSCI.1729-23.2024.full-html.html

Research Articles | Systems/Circuits

Basic Properties of Coordinated Neuronal Ensembles in
the Auditory Thalamus

https://doi.org/10.1523/JNEUROSCI.1729-23.2024

Received: 13 September 2023
Revised: 4 March 2024
Accepted: 11 March 2024

This Early Release article has been peer reviewed and accepted, but has not been through
the composition and copyediting processes.The final version may differ slightly in style or
formatting and will contain links to any extended data.

Alerts: Sign up at

www.jneurosci.org/alerts

to receive customized email alerts when the fully

formatted version of this article is published.

Title: Basic Properties of Coordinated Neuronal Ensembles in the Auditory Thalamus

Abbreviated title: cNEs in MGB

Congcong Hu

1,2,3

, Andrea R. Hasenstaub

1,2,3

, Christoph E. Schreiner

1,2,3

John & Edward Coleman Memorial Laboratory,

Neuroscience Graduate Program,

Department

of Otolaryngology-Head and Neck Surgery, University of California-San Francisco, San

Francisco, California 94158, USA

Corresponding author: Christoph E. Schreiner, christoph.schreiner@ucsf.edu

Number of pages: 51

Number of figures: 9

Number of all Words: 13940

JNeurosci Accepted Manuscript

Abstract (187), Significance Statement (115), Introduction (631), Discussion (1456)

Conflict of interest statement: The authors declare no competing financial interests.

Acknowledgements: We thank Drs. Jermyn See and Natsumi Homma for their help with

software and data collection. This work was supported by the National Institutes of Health

(DC002260 and DC017396 to C.E.S., and DC014101, NS116598, MH122478, and EY025174 to

A.R.H.), the Klingenstein Foundation (A.R.H.), PBBR Breakthrough Fund (A.R.H.), the

Coleman Memorial Fund (A.R.H., C.E.S.), and Hearing Research Inc. (C.E.S., A.R.H).

JNeurosci Accepted Manuscript

Abstract

Coordinated neuronal activity has been identified to play an important role in information

processing and transmission in the brain. However, current research predominantly focuses on

understanding the properties and functions of neuronal coordination in hippocampal and cortical

areas, leaving subcortical regions relatively unexplored. In this study, we use single-unit

recordings in female Sprague-Dawley rats to investigate the properties and functions of groups

of neurons exhibiting coordinated activity in the auditory thalamus -- the medial geniculate body

(MGB). We reliably identify coordinated neuronal ensembles (cNEs), which are groups of

neurons that fire synchronously, in the MGB. cNEs are shown not to be the result of false

positive detections or byproducts of slow state oscillations in anesthetized animals. We

demonstrate that cNEs in the MGB have enhanced information encoding properties over

individual neurons. Their neuronal composition is stable between spontaneous and evoked

activity, suggesting limited stimulus-induced ensemble dynamics. These MGB cNE properties

are similar to what is observed for cNEs in the primary auditory cortex (A1), suggesting that

ensembles serve as a ubiquitous mechanism for organizing local networks and play a

fundamental role in sensory processing within the brain.

JNeurosci Accepted Manuscript

Significance Statement

Temporal coordination of neuronal activity has been widely observed in various cortical areas

and has been shown to be important for signal processing and information transmission in the

brain. However, it remains unclear whether neuronal coordination is exclusive to cortical local

networks or if it also holds significance in subcortical regions. We conducted single-unit

recordings to investigate coordinated neuronal ensembles (cNEs), which are groups of neurons

with synchronous firing, in both the auditory thalamus and cortex. We demonstrated the

existence of cNEs in the auditory thalamus, which have similar properties to cNEs in the

auditory cortex. This provides evidence that subcortical neuronal coordination can serve as a

fundamental mechanism for organizing and processing neural signals.

JNeurosci Accepted Manuscript

Introduction

The function of coordinated neuronal activity in cognitive processes has long been a subject of

interest in systems neuroscience (Konorski, 1948; Hebb, 1949). Initially, such activity was

difficult to observe experimentally. However, recent technological advancements in large-scale

recording, such as two-photon imaging and high-density multi-channel probes, have facilitated

extensive investigations into the properties and functions of coordinated neuronal firing,

primarily within the hippocampus and neocortex (Laubach et al., 2000; Baeg et al., 2003; Harris

et al., 2003; Bizley et al., 2010; Buzsáki, 2010; Bathellier et al., 2012; Oberto et al., 2021;

Boucly et al., 2022; Domanski et al., 2023). These studies have revealed temporal coordination

among neurons in several brain areas, shedding light on their potential roles in various cognitive

processes, such as perception, memory formation, and decision making. Indeed, neuronal

ensembles have been proposed as the fundamental units for information processing and

transmission (Buzsáki, 2010; Yuste, 2015).

In sensory systems in particular, temporal coordination among neurons has been proposed as a

mechanism to enhance information processing (Kreiter and Singer, 1996; Dan et al., 1998; See et

al., 2018, 2021) and facilitate communication within and between brain regions (Zandvakili and

Kohn, 2015; Oberto et al., 2022). Neuronal coordination allows more reliable and specific

representation of stimuli (See et al., 2018; Yoshida and Ohki, 2020; See et al., 2021; Ebrahimi et

al., 2022), and considering neuronal coordination allows identification of emergent stimulus

encoding properties (deCharms and Merzenich, 1996; Shahidi et al., 2019). Additionally,

elevated coordination in neuronal activity in an output or sender area often precedes activity in a

JNeurosci Accepted Manuscript

target or receiver area (Zandvakili and Kohn, 2015). Thus, it is crucial to investigate the

expression of cNE structure and function at various stages along the sensory pathway.

In the auditory system, A1 contains pairs of neurons with correlated activity (Brosch and

Schreiner, 1999; Eggermont, 2000; Atencio and Schreiner, 2013) as well as larger groups of

neurons with correlated firing. These groups contribute to the representation of auditory stimuli

(Kreiter and Singer, 1996; Miller and Recanzone, 2009; Ince et al., 2013; See et al., 2018). While

characteristics of coordinated activity within the cortex have recently been studied (Bathellier et

al., 2012; Chamberland et al., 2017; See et al., 2018, 2021), our understanding of the

organization and functional significance of neuronal ensembles in subcortical regions, such as

the thalamus, remain unknown. The thalamus is of particular interest since it is the gateway and a

direct intermediary between the peripheral sensory system and the cortex (Winer et al., 2005;

Smith et al., 2012; Bartlett, 2013). The auditory thalamus MGB and A1 are highly

interconnected, with structured connections linking neurons which share similar spectral and

temporal response properties (Miller et al., 2002; Bartlett and Wang, 2007). Considering the

strong connections between the thalamus and cortex, investigating the shared characteristics of

neuronal ensembles in both regions will help us better understand the role these ensembles may

play in auditory information processing and transmission, and sensory processing in general.

In this study, we aimed to identify and characterize coordinated neuronal ensembles (cNEs) in

the MGB. We reliably detected cNEs, defined as groups of neurons exhibiting temporally highly

coordinated activity, in the MGB. The applied detection method showed robustness, consistently

identifying cNEs across different time bin sizes. Importantly, we observed a high degree of

similarity between cNEs derived from spontaneous and evoked activity, suggesting that these

JNeurosci Accepted Manuscript

ensembles represent functional networks that can operate, to a substantial degree, independently

of specific sensory stimuli. Furthermore, cNEs in the MGB and A1 shared key characteristics. In

both structures, spikes associated with cNEs reflect auditory information more reliably than

random spikes from the same neurons. These findings support the hypothesis that cNEs serve as

a ubiquitous mechanism for organizing local networks and function as fundamental units for

100

sensory processing in the brain.

101

Materials and Methods

102

Animals

103

All experimental procedures were approved by the Institutional Animal Care and Use Committee

104

at the University of California, San Francisco (UCSF), and followed the guidelines of the

105

National Institute of Health for the care and use of laboratory animals. Twenty-four female

106

Sprague-Dawley rats (wild type, 250-350g, 2-4 months; RRID: MGI: 5651135), sourced from

107

Charles River, were used in this study.

108

Surgery

109

The detailed procedures were as described in previous studies (See et al., 2018; Homma et al.,

110

2020). Briefly, anesthesia was induced with a combination of ketamine (100 mg/kg, Ketathesia,

111

HenrySchein) and xylazine (3.33 mg/kg, AnaSed, Akorn), along with atropine (0.54 mg/kg,

112

AtroJectSA, HenrySchein), dexamethasone (4 mg/kg, Dexium-SP, Bimeda), and meloxicam (2

113

mg/kg, Eloxiject, HenrySchein). Additional doses of ketamine (10-50 mg/kg) and xylazine (0-20

114

mg/kg) were given as needed to maintain anesthesia. Local anesthesia was provided using

115

lidocaine (Lidoject, 2%, HenrySchein) prior to making incisions. The respiratory rate, heart rate,

116

JNeurosci Accepted Manuscript

and depth of anesthesia were continuously monitored, and anesthesia was adjusted as needed.

117

The body temperature was monitored and maintained at 37°C using a homeothermic blanket

118

system (Harvard Apparatus 55-7020). Lubricant ophthalmic ointment (Artificial Tears,

119

HenrySchein) was applied to protect the eyes. A tracheotomy was performed to ensure stable

120

breathing during recording. To access the brain, the skin, muscle, skull, and dura over the right

121

temporal lobe were removed, and silicone oil (Sigma-Alderich) was applied to cover the cortex.

122

A bone rongeur was used to widen the craniotomy window and provide dorsal access to the

123

MGB. A cisternal drain was performed to prevent brain swelling.

124

Electrophysiology

125

The frequency organization of auditory cortex was first mapped using Tungsten electrodes. A1

126

was identified as the area with a high to low frequency preference gradient on the rostral-causal

127

axis and short latency response to pure tones (Polley et al., 2007). Then, electrophysiological

128

recordings were performed using a linear silicon probe with 64 channels (H3, 20µm channel

129

distance, Cambridge NeuroTech) in the MGB and a 2-shank probe with 64 channels (H2, 25µm

130

channel distance, Cambridge NeuroTech) in A1. The ventral division of the MGB is

131

characterized by a low to high frequency gradient on the dorsal-ventral axis (Morel et al., 1987;

132

Anderson and Linden, 2011). The probes were inserted using microdrives (David Kopf

133

Instruments) at a rate of 25 µm/s to a depth of 4500 to 6000 µm from the surface of the cortex to

134

reach MGB (Figure 1A) and 900 to 1300 µm in A1 along the columnar structure (Figure 7A),

135

respectively. Extracellular voltage traces were recorded at a sampling rate of 20 kHz with an

136

Intan RHD2132 Amplifier system (Intan Technologies). Multi-unit (MU) activities (Figure 1A

137

and 7A) were defined as negative peaks crossing 4 standard deviations from the mean in the

138

JNeurosci Accepted Manuscript

extracellular voltage trace filtered between 300 and 6000 Hz. Single unit activities were obtained

139

by spike sorting using Kilosort 2.5 (Steinmetz et al., 2021; Pachitariu et al., 2023), followed by

140

manual curation using Phy (

https://github.com/cortex-lab/phy

). In the manual curation process,

141

we visually evaluated individual clusters by examining auto-correlograms, spike waveforms, the

142

stability of spike amplitude over time, the persistence of activity over time, the cluster's

143

separation from noise in the feature space, and other visual aids provided by phy2 for

144

distinguishing single-unit clusters from multi-unit or noise clusters. Subsequently, the identified

145

units underwent filtering based on specific criteria: inter-spike interval (ISI) violation within 2ms

146

< 1.5%. The majority of single units exhibited an ISI violation lower than 0.25%; peak signal-to-

147

noise ratios (SNRs) of the waveform > 1.5 (median peak SNR: MGB: 4.52, A1: 3.77); and firing

148

rates > 0.1 Hz to eliminate potential multi-units. To assess the reliability of activity of the single

149

units across the entire recording duration we obtained the presence ratio. This is the ratio of the

150

number of blocks where the unit showed activity and the total number of blocks in a recording

151

session. To calculate the presence ratio, the entire recording was divided into 100 equal time

152

blocks. The majority of the obtained single units were active in more than 95% of the time

153

blocks. This sorting resulted in single units that exhibit low ISI violations, high peak SNR, firing

154

rates with a log-normal distribution, and a more consistent presence during recording when

155

compared to all clusters generated by Kilosort. To identify oscillatory response in MGB or A1,

156

single units on the same electrode were combined to form multi-units (Figure 8).

157

Stimuli

158

To measure frequency tuning, we presented pure tones with frequencies ranging from 0.5 to 32

159

kHz in 0.13 octave steps and sound levels from 0 to 70 dB in 5 dB steps (50ms, 5ms ramps).

160

JNeurosci Accepted Manuscript

Each frequency-sound level combination was presented once in a pseudo-random order, with an

161

inter-stimulus interval of 250ms. To assess spectrotemporal receptive fields (STRFs), we used a

162

15-minute dynamic moving ripple (DMR) (Escabí and Schreiner, 2002). The DMR consisted of

163

40 sinusoidal carrier frequencies per octave in the range of 0.5 to 40 kHz, each with a random

164

phase. The carriers were slowly modulated (maximum rate of change



3Hz) by a

165

spectrotemporal envelope with a maximum spectral modulation rate of 4 cycles/octave, a

166

maximum temporal modulation rate of 40 cycles/s, and a maximum modulation depth of 40 dB.

167

The mean intensity of the DMR was set at 70 dB sound pressure level. We selected DMR as the

168

stimulus for analyzing cNEs’ response to sound stimuli, as overt onset response effects for the

169

15-minute continuous stimulus is negligible. Additionally, DMR has been observed to reduce

170

oscillatory states in the neural population (Miller and Schreiner, 2000). All auditory stimuli were

171

generated using MATLAB (MathWorks) and calibrated using a 1/2-inch pressure field

172

microphone (Type 4192, Bruel and Kjær). The stimuli were delivered contralaterally from the

173

recording site using a closed-field electrostatic speaker (EC1, TuckerDavis Technologies) at a

174

sampling rate of 96 kHz.

175

Detecting coordinated Neural Ensembles (cNEs)

176

To identify groups of neurons that exhibit synchronized co-activation, referred to as “cNEs”, we

177

used a method combining principal component analysis (PCA) and independent component

178

analysis (ICA) (Lopes-dos-Santos et al., 2013; See et al., 2018). We selected a bin size of 10ms

179

as a standard synchronization span because it represents the most appropriate time window to

180

capture the synaptic integration window of most cortical neurons (Léger et al., 2005; D’amour

181

and Froemke, 2015). First, the individual spike trains of simultaneously recorded neurons were

182

JNeurosci Accepted Manuscript

binned and z-scored. Next, the z-scored spike matrix underwent PCA to obtain the eigenvalues

183

of the spike train correlation matrix. To determine the number of cNEs, we took eigenvalues to

184

be significant if their value exceeded the 99.5

percentile of the Marchenko-Pastur distribution,

185

which describes the probability density function of eigenvalues of large rectangular random

186

matrices (Marchenko and Pastur, 1967; See et al., 2018) (Figure 2A-ii). We then performed ICA

187

(FastICA) on the subspace spanned by the eigenvectors corresponding to the significant

188

eigenvalues. The resulting independent components (ICs) represent groups of neurons with

189

shared spiking events. The weight of each neuron on an IC indicates the neuron’s contributions

190

to the cNE (Figure 2A-iii). As the signs of IC weights were arbitrary, for each IC, the direction

191

with the largest absolute weight was rendered positive. The length of each IC was normalized to

192

one, making an IC with equal contribution from all neurons have weights of 1/√N, where N was

193

the number of neurons in the recording. Neurons with weights over 1/√N were referred to as

194

“cNE members” (Oberto et al., 2021) (Figure 2A-iv).

195

The strength of the cNE activation at each time point was measured by the similarity between the

196

activity of cNE members as well as the cNE pattern, i.e., which neurons in a penetration were

197

cNE members. The similarity can be measured as the square of the weighted sum of the z-scored

198

spike counts,

𝑠 = 𝑧

𝑇

𝑤𝑤

𝑇

𝑧 = 𝑧

𝑇

𝑃𝑧

, where

is the z-scored spike counts of cNE members at

199

each time point,

is the IC weights of cNE members, and the projection matrix

is the outer

200

product of

. To consider only co-activation of multiple cNE members, we set the diagonal of

201

the projection matrix to zero and obtained the modified projection matrix

. cNE activity

202

strength was calculated as

𝑠 = 𝑧

𝑇

𝑃

∗

𝑧

. A null distribution of cNE activity was obtained by

203

projecting a circularly shifted spike matrix, where the temporal relationship of neurons was

204

JNeurosci Accepted Manuscript

disrupted, to the template matrix (See et al., 2018). This process was iterated 50 times, and the

205

threshold of cNE activation was defined as the 99.5th percentile of the null distribution (Figure

206

2A-v). The spikes of cNE members within the selected time bin where the cNE was active were

207

referred to as “cNE spikes”.

208

Matching cNEs across different bin sizes

209

We used the correlation between IC weights of all neurons in a penetration to assess the

210

similarity of cNE patterns across different synchronization windows (i.e., time bin width: 2, 5,

211

10, 20, 40, 80 and 160ms). To visualize the similarity of cNEs identified using 10ms bins to

212

those identified using other bin sizes, we calculated the cNEs for the same recording using other

213

bin sizes. We displayed the cNEs whose IC weights were best correlated with the 10ms cNE

214

(Figure 3A). To measure the variability of cNE identities across bin sizes (Figure 3B, C), we

215

matched each cNE to the most similar cNE calculated using reference bins (i.e., when using

216

10ms as the reference bin size, to the 10ms cNE which had the best-correlated IC weights). The

217

proportion of shared members with a reference cNE was then calculated by dividing the number

218

of members in a cNE that were also identified as members in its matching reference cNE by the

219

total number of unique members in both cNEs combined.

220

The significance of the match was determined based on the null distribution of IC weight

221

correlations between matched cNEs. For example, to determine the significance of the

222

correlation between the IC weights of a 10ms cNE with its most correlated 160ms cNE, we first

223

generated a null distribution of IC weight correlations. We circularly shifted spike trains and then

224

applied PCA/ICA to identify sham cNEs using the shuffled spike matrices binned at 160ms,

225

JNeurosci Accepted Manuscript

maintaining the same number of cNEs as the original 160ms cNEs. Then we identified the most

226

correlated sham 160ms cNE for the 10ms cNE. This process was repeated 1000 times to generate

227

the null distribution of correlation values. The significance threshold was set at p < 0.01.

228

Assessing stability of cNEs

229

To assess the stability of cNEs during and across spontaneous and stimulus-driven activity, we compared

230

the cNEs from adjacent recording segments (Figure 4A).

To match the IC weights of cNEs identified

231

from the different recording segments, we used an iterative process that involved selecting cNE

232

pairs from the two segments with the highest correlations (Spearman’s r) (Oberto et al., 2021).

233

First, we computed the correlations between all possible pairs of cNEs that were generated from

234

the two segments. Then, the pair with the highest correlation was set aside, and the same process

235

was repeated with the remaining cNEs until all cNEs were paired. If there were any remaining

236

cNEs that did not have a match due to a difference in the number of cNEs between the two

237

segments, they were left unmatched.

238

To generate a null distribution of IC weight correlations between matched cNEs from two

239

recording segments (Figure 4D), we circularly shifted spike trains within each activity block. We

240

then applied PCA/ICA to identify sham cNEs using the resulting shuffled spike matrices. As the

241

shuffling disrupted correlations between neurons, very few eigenvalues exceeded the upper

242

bounds of the Marchenko-Pastur distribution. To address this, we maintained the number of

243

sham cNEs in the shuffled data equivalent to the number of significant eigenvalues obtained

244

from the original spike matrix. The sham cNEs from adjacent activity blocks were then matched

245

following the procedure described above. This iterative process was repeated 1000 times to

246

JNeurosci Accepted Manuscript

establish a null distribution of IC weight correlations for the matched cNEs. The 99.5th

247

percentile of each null distribution was set as the significance threshold.

248

False positive detection of cNEs

249

To assess the potential for false positive cNE detection, we applied the cNE detection algorithm

250

to shuffled data, using the same criteria as those applied to the real dataset. This process was

251

repeated 10 times, resulting in an average count of false positive cNEs across the circularly

252

shifted data (Figure 4F). Despite conducting 10 iterations, false positive cNEs were not

253

consistently identified in neighboring blocks. In cases where a false positive cNE was detected

254

(i.e., when an eigenvalue computed from shuffled data exceeded the Marchenko-Pastur

255

distribution), we evaluated its stability by measuring the highest correlation of its IC weights

256

with those of real cNEs in the adjacent block (Figure 4E). The significance of false positive cNE

257

IC weight correlations was determined using the same threshold established for real cNEs.

258

STRF analysis

259

For analysis, we down-sampled DMR to a resolution of 0.1 octaves in frequency and 5ms in

260

time. We used the reverse correlation method to obtain the STRFs of the units (Theunissen et al.,

261

2000; Escabí and Schreiner, 2002). To derive the STRFs, we averaged the spectrotemporal

262

envelopes of the stimulus over a period of 100ms preceding spikes (Figure 1B and 6A). Positive

263

(red) values on a STRF indicate that the sound energy at that frequency and time tends to

264

increase the firing rate of the unit, while negative (blue) values indicate where the stimulus tends

265

to decrease the firing rate of the unit. The frequency corresponding to the highest absolute value

266

JNeurosci Accepted Manuscript

of the STRF is considered the best frequency (BF) of the unit (Miller et al., 2002). We also

267

determined the peak-to-trough difference (PTD) as a measure of STRF strength.

268

A STRF was considered significant if it reliably described a neuron's response to DMR sound.

269

To assess the reliability of a STRF, we divided a neuron's spikes into two equal halves and

270

generated two corresponding STRFs (STRF A and B) using each half (Qiu et al., 2003). The

271

similarity between STRF A and B was computed using Pearson's correlation. This process was

272

reiterated 1000 times, and the average STRF similarity across these iterations was used as the

273

measure of reliability. To determine the statistical significance of a STRF's reliability, we

274

constructed a null distribution by reversing the neuron's spike train, thus disrupting the temporal

275

correlation of neural responses to the stimulus. We considered STRFs with a reliability

276

surpassing a z-score of 2.58 to be significant.

277

We used mutual information (MI) as the metric to quantify the amount of information we can

278

obtain about the stimulus by observing spikes of neurons or cNEs (Atencio and Schreiner, 2008;

279

See et al., 2018). The stimulus segment

𝑠

preceding each spike was projected onto the STRF via

280

the inner product

𝑧 = 𝑠 ∗ 𝑆𝑇𝑅𝐹

.The projection values were then binned to get the probability

281

distribution

𝑃(𝑧|𝑠𝑝𝑖𝑘𝑒)

. The

a priori

distribution of stimulus projection values,

𝑃(𝑧)

, was

282

calculated by projecting all stimulus segments of DMR onto the STRF, regardless of spike

283

occurrence. Both distributions

𝑃(𝑧)

and

𝑃(𝑧|𝑠𝑝𝑖𝑘𝑒)

were normalized relative to the mean

𝜇

and

284

standard deviation

𝜎

𝑃(𝑧)

, by

𝑥 =

(𝑧−𝜇)

𝜎

, resulting in

𝑃(𝑥)

and

𝑃(𝑥|𝑠𝑝𝑖𝑘𝑒)

. The MI between

285

STRF projection values and single spikes was computed according to

𝐼 =

286

∫ 𝑑𝑥𝑃(𝑥|𝑠𝑝𝑖𝑘𝑒) log

𝑃(

𝑥

𝑠𝑝𝑖𝑘𝑒

)

𝑃(𝑥)

287

JNeurosci Accepted Manuscript

STRF comparisons between cNEs and non-cNE groups of neurons

288

To control for the potential influence of population synchrony on a cNE due to independent

289

neuron activity, we compared STRFs derived from cNEs and non-cNE groups of neurons. If less

290

than half of members had significant STRFs, the cNE was excluded from analysis.

291

First, we compared the group STRFs of cNEs and non-cNE groups of neurons (Figure 6B-ii and

292

C-ii) (See et al., 2018). The group STRF was calculated using all spikes from neurons within a

293

group. To generate non-cNE groups of neurons relative to a cNE, we first selected one neuron

294

from the cNE and then sampled from the remaining neurons with significant STRFs within a

295

penetration, forming a group with the same number of neurons as the cNE for comparison. This

296

process excluded other member neurons of the cNE. This procedure was repeated for all

297

members of the cNE, generating all possible combinations of neurons, each including exactly

298

one member neuron from the cNE under examination. Combinations of neurons that included

299

more than one neuron from any other cNEs in the same recording were then also excluded. For

300

each cNE/non-cNE group comparison, we subsampled the spikes in the cNE and the non-cNE

301

groups to the same number. Subsequently, STRF peak-to-trough difference (PTD) and mutual

302

information (MI) of the cNE group were compared to the median values of the non-cNE groups.

303

We also compared the STRFs of cNE spikes with those of coincident spikes from a single neuron

304

(Figure 6B-iii and C-iii) to assess the influence of random coincidence on stimulus preference.

305

To obtain coincident spikes from a specific neuron, we first sampled neurons from the recorded

306

population that do not share membership with the neuron under examination in any cNE to

307

create a non-cNE group. We kept the number of neurons in the non-cNE group the same as the

308

JNeurosci Accepted Manuscript

cNE to which the neuron being examined belongs. This sampling process was restricted to

309

neurons exhibiting significant STRFs. The coincident spikes of the cNE member refer to spikes

310

within 10ms of spikes from other neurons within the non-cNE group. We repeated this procedure

311

to generate all possible combinations of non-cNE groups, each containing the cNE member and

312

excluding any neuron that shares membership with the neuron under scrutiny. Coincident spike

313

trains with less than 100 events were discarded. For each cNE spike/non-cNE spike comparison,

314

we subsampled the cNE spikes and spikes from the non-cNE groups to the same number.

315

Subsequently, the cNE spike STRF PTD and MI were compared to the median values of the

316

random spike STRF PTD and MI from the non-cNE groups.

317

Quantifying slow oscillations in neural activity

318

To determine whether the neural activity in a recording showed a prominent pattern of slow

319

oscillations, we measured silence density and the coefficient of variation (CV) of MU firing rate.

320

Silence density was defined as the fraction of 20ms time bins with no population activity (zero

321

spikes) (Mochol et al., 2015). The CV of MU firing rate was calculated as CV =

𝜎

𝜇

, where

𝜇

322

the mean firing rate and

𝜎

is the standard deviation of the firing rate binned at 20ms time bins.

323

Permutation test

324

We used permutation tests to determine the statistical significance of differences in cross-

325

correlograms (CCGs) among neurons based on their membership (Figure 2C and 3D), as well as

326

to assess differences in the proportion of stable cNEs between different stimulus conditions

327

(Figure 4E and 7E). For example, to assess the difference in CCGs between member pairs and

328

non-member pairs (Figure 2C), we shuffled the membership labels of the CCGs and calculated

329

JNeurosci Accepted Manuscript

the difference between the average CCGs of member and non-member pairs. We repeated this

330

process 10,000 times to generate null distributions of the CCG difference for each data point.

331

The 0.5

and 99.5

percentiles of the null distribution were taken as the cutoffs for significance.

332

We considered consecutive time bins around 0ms-lag with p < 0.01 to be significant. To assess

333

the difference in the proportion of stable cNEs, we shuffled the stimulus condition label

334

(spontaneous [‘spon’], DMR [‘dmr’], or cross condition comparison [‘cross’]) and repeated this

335

process 10,000 times to generate a null distribution of the difference in proportion. The

336

significance level was then determined based on the null distribution.

337

Statistics

338

Statistical analyses were performed in Python. To compare two unpaired groups (e.g., Figure

339

2B), we used Mann–Whitney U tests. To compare two paired groups (e.g., Figure 4F), we used

340

Wilcoxon signed-rank tests. Permutation tests (e.g., Figure 2C), and Monte Carlo methods (e.g.,

341

Figure 4D) were used as described above. To determine if two samples are drawn from the same

342

distribution, we used Kolmogorov-Smirnov test (Figure 3D). The specific applications of these

343

tests are explained in the results section and figure legends. Significance levels are noted as n.s.

344

(p >= 0.05), * (p

0.05), ** (p

0.01) and *** (p

0.001).

345

346

Results

347

Auditory responses in MGB

348

We conducted extracellular recordings in the rat MGB (Figure 1A) using a 64-channel linear

349

probe, which allowed us to cover most of its span along the dorsal-ventral axis. To obtain the

350

JNeurosci Accepted Manuscript

tonal response properties of the recording sites, we presented pure tones of various frequencies

351

and intensities. In the MGB, we usually observed a gradient in the frequency preference of multi-

352

unit (MU) responses from low to high along the dorsal-ventral axis, which could vary gradually

353

(Figure 1A-i) or abruptly (Figure 1A-ii) depending on the probe’s location. Responses on most

354

channels in the tonotopic region exhibited clear frequency tuning (between the red lines in Figure

355

1A), which likely reflect activities in the ventral MGB, the primary input station to the A1. We

356

included all SUs from the MGB in our analysis after spike sorting, without distinguishing

357

between sub-regions although the vast majority was likely from the ventral nucleus according to

358

its tonotopic organization.

359

To estimate the STRFs of SUs, we used a 15-minute DMR stimulus, which is a broadband noise

360

with varying spectral and temporal modulation (Escabí and Schreiner, 2002). The STRFs of

361

MGB neurons also showed a clear gradient in frequency preference from low to high along the

362

dorsal-ventral axis (Figure 1B), consistent with the MU responses to pure tones. We then

363

examined the firing correlations between pairs of simultaneously recorded SUs. MGB neuron

364

pairs showed widely different correlations in their firing activity, even if they were close in

365

proximity and had similar STRFs. For example, Neurons #1, #2, and #3 had similar receptive

366

fields (Figure 1B). While neurons #1 and #3 showed correlated firing in both stimulus-driven

367

and spontaneous activity, neurons #2 and #3 showed no significant correlation in their activity

368

despite similar STRFs (Figure 1C). This diversity of correlation patterns, even among neurons

369

with similar receptive fields, parallels what was previously observed in the cortex (Brosch and

370

Schreiner, 1999; Eggermont, 2000; Atencio and Scheiner, 2013; See et al., 2018; Mogensen et

371

al., 2019; Wahlbom et al., 2021). Although the role of neuronal coordination in information

372

JNeurosci Accepted Manuscript

processing in the cortex has been extensively proposed and studied (e.g., Paninski et al., 2004;

373

Bizley et al., 2010; Buzsáki, 2010; Carrillo-Reid et al., 2015; See et al., 2018), less is known

374

about the organization of neuronal ensembles in subcortical regions. Therefore, we aimed to

375

identify clusters of neurons that exhibit consistent synchronized firing in the MGB and compared

376

the properties of these ensembles with those in A1.

377

Identifying groups of neurons with coordinated firing in MGB

378

To identify cNEs, i.e., groups of neurons with synchronous firing, we performed a combined

379

principal- and independent-component analysis (PCA-ICA) (Lopes-dos-Santos et al., 2013; See

380

et al., 2018). The procedure for detecting cNEs in a population of neurons is demonstrated in

381

Figure 2A using a recording of spontaneous activity from the MGB (Figure 2A). Among the 20

382

isolated single units in the recording, some pairs of neurons had highly correlated firing with

383

each other, as shown in dark red in the correlation matrix, while others showed low correlation

384

(Figure 2A-i). We performed PCA on the correlation matrix of 10ms-binned spike trains,

385

resulting in 20 eigenvalues and corresponding eigenvectors or principal components (PCs)

386

(Figure 2A-ii). These eigenvalues describe the contribution of each PC to the variance in the

387

neural population activity. To determine the significance of the patterns extracted by PCA, we

388

compared the eigenvalues to a threshold drawn based on the Marchenko-Pastur distribution

389

(Peyrache et al., 2010; Lopes-dos-Santos et al., 2013) (Figure 2A-ii). In this example recording,

390

we observed four significant eigenvalues above the threshold, indicating the presence of four

391

detectable cNEs in the recorded population. Although PCA efficiently extracts ensemble

392

patterns, it has some limitations due to its variance maximization framework. When two

393

ensembles account for similar variance in the data on their corresponding axis, the first PC will

394

JNeurosci Accepted Manuscript

represent the average of the two instead of an individual ensemble. This problem is even more

395

pronounced when ensembles share neurons. To overcome these limitations, we applied ICA to

396

the subspace spanned by the significant PCs (Figure 2A-iii). This approach is not constrained by

397

the orthogonality requirement of PCA, allowing for a more precise identification of individual

398

cNEs. After the PCA-ICA procedure, we obtained the weights of neurons on the axes that define

399

cNEs in the neural population, which were color-coded as columns in Figure 2A-iii. Neurons

400

were considered members of a cNE if their IC weights were higher than what would be expected

401

from an even weight contribution from all neurons (Figure 2A-iv). The activity resulting from

402

co-activation of cNE members can be obtained by projecting the spike matrix on the

403

corresponding IC weights of the cNE. To determine the significance of cNE activity magnitude,

404

we generated a null distribution of the cNE activity values by circularly shuffling spike trains

405

and set the significance criteria at 99.5% (Figure 2A-v). For example, when cNE #1 was active,

406

multiple member neurons (2-5 out of 6) fired together. The combined PCA-ICA approach

407

provided a useful framework to investigate the organization and function of coordinated neuronal

408

activity in the MGB.

409

To provide evidence that cNEs captured groups of neurons with correlated firing, we compared

410

the correlations of 10ms-binned spike trains based on their cNE membership. Pairs of neurons

411

that participated in the same cNE (“member pairs”) exhibited significantly higher correlations

412

compared to pairs of neurons that did not share membership in any cNE (“non-member pairs”)

413

(Figure 2B). To examine the correlation between member and non-member pairs at a finer

414

timescale, we cross-correlated the spike trains using 1ms bins. The correlation among cNE

415

member pairs was significantly higher compared to non-member pairs within the [-50, 40] ms

416

JNeurosci Accepted Manuscript

lag window (red-shaded area in Figure 2C, bottom left). These results provided evidence that

417

groups of neurons with coordinated firing exist in the MGB, and that their coordination was

418

captured by the PCA-ICA procedure, resulting in the identification of cNEs.

419

Variability in cNE identity for different bin sizes

420

The temporal frame used to identify neuronal ensembles plays a critical role in shaping our

421

understanding of their nature and function (Buzsáki, 2010). To investigate how different choices

422

of timescale affect the identification of cNEs, we calculated cNEs using various spike train bin

423

sizes, ranging from 2ms to 160ms (Figure 3). Some cNEs showed consistent IC weights across

424

different time bin sizes, with a high correlation to IC weights obtained using 10ms bins (Figure

425

3A-i). Other cNEs could only be consistently identified using smaller time bin sizes but deviated

426

from those when assessed with larger bin sizes (Figure 3A-ii). We matched cNEs calculated

427

using different bin sizes to evaluate their similarity (Figure 3B). With only small change in the

428

bin size, the cNEs identified were highly similar. For example, 96% of cNEs identified using

429

10ms bins had a significant match with 20ms cNEs. However, when compared to cNEs

430

calculated with larger differences in bin sizes, their identities could vary substantially. For

431

example, 43% of 10ms cNEs did not show a significant match with 160ms cNEs (Figure 3B).

432

The remaining 57% of 10ms cNEs that significantly matched 160ms cNEs exhibited high

433

correlation (> 0.6) in their IC weights (Figure 3C-i). Moreover, the majority of significantly

434

matched cNEs shared more than half of their neuron membership. Nonetheless, approximately

435

25% of 10ms cNEs had no common members with the 160ms cNEs (Figure 3C-ii). There was no

436

significant difference in the firing rate of MGB neurons participating in 10ms and 160ms cNEs

437

(Figure 3D). In summary, small variations in time bin sizes have a limited effect on cNE identity.

438

JNeurosci Accepted Manuscript

However, using large time bin sizes to identify cNEs, such as 160ms, may result in the loss of

439

half or more of the cNEs identified using small bin sizes, such as 10-20ms.

440

In cases where differences in cNE membership arise due to different bin sizes, we investigated

441

the firing correlations between neurons that had shifted in or out of the ensemble. We compared

442

the membership of neurons in cNEs identified using 10ms and 160ms bin sizes (“10ms cNEs”

443

and “160ms cNEs”, Figure 3E) and categorized neurons in each cNE as the following: members

444

in both 10ms and 160ms cNEs, members only in the 10ms cNE, or members only in the 160ms

445

cNE. Neurons sharing memberships in both 10ms and 160ms cNEs (stable members) had

446

positive correlations in their firing (Figure 3E-i). Neurons only in 160ms cNEs were positively

447

correlated with stable members, although the correlation was significantly weaker in the [-17,

448

10] ms window (Figure 3E-ii) compared to the correlation among stable members (Figure 3E-i).

449

Furthermore, some members of 10ms cNEs were not identified as members in 160ms cNEs.

450

These neurons showed no significant difference in their correlations with stable members (Figure

451

3E-iii) compared to the correlation among stable members (Figure 3E-i). Therefore, using wider

452

bin sizes to identify cNEs results in neurons with weak correlations being included in the

453

ensemble, as well as neurons with strong correlations being omitted.

454

Variability in cNE structures across spontaneous and evoked activity

455

Several studies have shown that cortical neuronal ensembles have stable structures across

456

spontaneous and stimulus-driven activity, suggesting a consistent local network organization

457

utilized in processing stimulus information (Jermakowicz et al., 2009; Luczak et al., 2009; See et

458

al., 2018; Filipchuk et al., 2022). To investigate if this property also exists in cNEs in the MGB,

459

JNeurosci Accepted Manuscript

as was observed in A1 (See et al., 2018), we recorded continuous segments of neural activity in

460

the absence of sound (spontaneous, hereafter ‘spon’) and during the presentation of the DMR

461

stimulus (‘dmr’). We divided each activity type into two segments and detected cNEs in each

462

segment separately. We then compared the stability of the cNEs within and across stimulus

463

conditions, measured by the correlation of IC weights between adjacent segments (Figure 4A and

464

B). We observed that while some cNEs exhibited high stability across stimulus conditions, with

465

IC weights correlation comparable to that within the same stimulus condition (Figure 4C-i),

466

others showed structures that were less stable across stimulus conditions compared to within a

467

stimulus condition (Figure 4C-ii). Using null distributions generated by circularly shuffling

468

spikes, we determined the significance of the IC weight correlations and found that both

469

examples in Figure 4C were significantly stable across stimulus conditions, although one was

470

slightly more stable than the other (Figure 4D). In MGB, within spontaneous or stimulus-driven

471

activity, around 80% of cNEs exhibited stable structures across adjacent activity blocks (Figure

472

4E). Significantly fewer cNEs (54.8%) were stable across stimulus conditions than within a

473

stimulus condition (spon vs cross, p = 0e-4; dmr vs cross, p = 6e-4; spon vs dmr, p = 1.0,

474

permutation test with Bonferroni correction). The results provide evidence for the stability of

475

cNEs in the MGB, during both spontaneous and stimulus-driven activity, although fewer cNEs

476

exhibit stable structures across different stimulus conditions than within the same stimulus

477

condition.

478

To test the possibility of false positive detection of cNEs, we generated shuffled data on each

479

segment by circularly shifting spike trains to disrupt their temporal correlations. We then applied

480

the cNE detection algorithm to the shuffled data using the same criteria as for the real data. Our

481

JNeurosci Accepted Manuscript

analysis revealed a drastically lower number of cNEs identified in shuffled segments (real data

482

vs. shuffled data, spon1: 3.2 ± 0.9 vs 0.2 ± 0.3; spon2: 3.1 ± 0.7 vs 0.3 ± 0.4; dmr1: 3.1 ± 0.9 vs

483

0.1 ± 0.2; dmr2: 3.2 ± 1.1 vs 0.1 ± 0.1, mean ± SD) (Figure 4F), suggesting that the chance of

484

false positive detection of cNEs is quite low. Furthermore, any cNEs identified in the shuffled

485

data did not exhibit the same stability across stimulus conditions observed in real data (Figure

486

4E). In summary, our findings indicate that the detection of cNEs in the MGB is reliable and not

487

susceptible to false positives. Moreover, the properties of cNEs we observe, such as their

488

stability across stimulus conditions, are genuine and not artifacts of random data.

489

cNE properties in MGB

490

We determined some basic structural properties of MGB cNEs. The spontaneous activity in 34

491

MGB recordings revealed 115 cNEs with 3.4 ± 0.9 cNEs per penetration. More cNEs were

492

observed in penetrations that captured a higher number of isolated single units (Figure 5A). The

493

mean cNE size was 4.3 ± 1.5 members, dependent on the number of isolated neurons (Figure

494

5B). Of the 407 neurons isolated in MGB, the majority (78.6%) belonged to a single cNE, 11.5%

495

did not belong to any cNE, and 9.8% belonged to multiple cNEs (Figure 5C).

496

Next, we investigated whether cNE members were physically and functionally closer to each

497

other than non-member pairs of neurons. The pairwise spatial distance of cNE members was

498

significantly smaller than that of non-member pairs of neurons in MGB (Figure 5D-i). Moreover,

499

the span of cNEs, defined as the longest pairwise distance among all members, was shorter than

500

that of randomly selected groups of neurons in the recording (Figure 5D-ii). The tuning of cNE

501

members was also closer to each other, as the difference in the BF between cNE member pairs

502

JNeurosci Accepted Manuscript

was smaller than that of non-member pairs (Figure 5E-i). The BF span of cNEs, defined as the

503

largest difference in BF among all members, was smaller than that of randomly selected groups

504

of neurons (Figure 5E-ii). Our results demonstrate that cNEs in the MGB are composed of

505

neurons that are physically and functionally closer to each other than non-member pairs,

506

suggesting a pattern of local circuit organization as well as local functional congruence.

507

While we observed similarity in the tuning of cNE members (Figure 5E-ii), cNEs are not limited

508

to groups of neurons with similar receptive fields. The distribution of the pairwise difference in

509

the BF of cNE members exhibits a long tail, where cNE members could differ in their BF by

510

more than 4 octaves (Figure 5E-i). This suggests that neurons with largely different tuning

511

properties also exhibit synchronous activities and participate in the same cNE.

512

Previous studies have proposed a manifestation of ensemble coding based on neuronal groups

513

with covarying firing rate (Wills et al., 2005; Niessing and Friedrich, 2010; Aschauer et al.,

514

2022), e.g., groups of neurons that jointly increased their firing rates in response to various pure

515

tones (Aschauer et al., 2022). Assessing coactivation based on firing rate in response to various

516

stimulus, however, is a limited basis for the identification of neurons that functionally cooperate.

517

Such methods may simply reflect coactivation due to the overlap between receptive fields along

518

the tonotopic axis, and distinct groups of neurons may emerge due to best-frequency

519

discontinuities in the tonotopic organization (see our Figure 1A; Imaizumi et al., 2004). A more

520

stringent criterion, that of tight temporal synchrony, as utilized here, can help differentiating

521

between neuron groups based on coincidental coactivation and groups based on synchronous

522

coactivation.

523

JNeurosci Accepted Manuscript

cNEs enhance stimulus information encoding

524

Synchronization of neuronal spikes in the cortex has been found to enhance information

525

encoding about a stimulus compared to the participating neurons alone (Dan et al., 1998; Atencio

526

and Schreiner, 2013; See et al., 2018). This is consistent with the multiplexed nature of an

527

individual spike train, whereby spikes representing distinct stimulus aspects are mixed but can be

528

separated based on their synchrony with other neurons (Lankarany et al., 2019; See et al., 2021).

529

To investigate whether spikes from individual neurons that participate in cNEs also exhibit

530

differential coding compared to the neuron’s entire spike train, we compared the STRFs

531

calculated using all the spikes emitted by a neuron to STRFs only based on the subset of spikes

532

that contributed to cNE events (‘cNE spikes’; Figure 6A). The spike trains were subsampled to

533

ensure an equal number of spikes across conditions. Our analysis revealed that the STRFs of

534

cNE spikes exhibit stronger excitatory and inhibitory fields compared to the STRFs of all spikes

535

from the same neuron, as evidenced by the larger peak-trough difference (PTD) of the cNE

536

STRFs (Figure 6B-i), which quantifies the difference between the largest and smallest value in

537

the STRF. Given that PTD only considers two extreme values in the STRF, we further evaluated

538

the reliability of cNE spikes relative to all spikes in encoding the sound features represented by

539

their STRFs by calculating the MI between the stimulus and the spikes. Our results demonstrate

540

that cNE-spike STRFs have higher MI than STRFs constructed from all spikes (Figure 6C-i).

541

To demonstrate that the increased information conveyed by cNE spike STRFs was not simply

542

because cNEs integrate signals over multiple neurons, thus must enhanced information through

543

population encoding, we compared STRFs derived from cNE member and non-member neurons.

544

First, we compared the multi-unit STRFs of cNE member neurons (cNE group STRF) with those

545

JNeurosci Accepted Manuscript

of a neuronal group devoid of neurons sharing membership from any cNE (non-cNE group

546

STRF). The cNE group STRFs exhibited a significantly higher MI than that of the non-cNE

547

STRFs, although no significant difference in PTD was observed (Figure 6B-ii and C-ii). As the

548

group STRFs did not take spike synchrony into account, we further compared cNE spikes and

549

coincident spikes of a cNE member. The coincident spikes refer to spikes that occurred within a

550

10ms window relative to firing of other neurons not sharing membership with the neuron under

551

examination. Both the PTD and MI of cNE spike STRFs were significantly higher than that of

552

the coincident spike STRFs (Figure 6B-iii and C-iii).

553

Collectively, these findings suggest that cNE spikes can enhance information processing by

554

increasing the signal-to-noise ratio and promoting more consistent encoding of certain stimulus

555

features compared to including all spikes from the neuron. Furthermore, the enhanced

556

information encoding of cNEs is not a trivial result of population encoding but rather hinges on

557

the identity of cNE members and synchronous spike events.

558

MGB and A1 cNEs have similar properties

559

The properties and functions of cNEs have previously been explored within the primary auditory

560

cortex (A1; See et al., 2018, 2021), whereas investigations into cNEs in subcortical regions are

561

limited. Hence, we aim to determine whether the properties of cNEs in the MGB differ

562

substantially from those observed in A1 cNEs, or if they share similarities. To target A1, we used

563

a 2-shank probe with 64 channels. The MU responses to pure tones from the two shanks of the

564

probe exhibited similar frequency tuning, as the shanks sampled nearby cortical columns (Figure

565

7A). The responses on each shank showed small variation in their frequency preference along the

566

JNeurosci Accepted Manuscript

depth of the probe, as neurons in the same cortical column have consistent characteristic

567

frequencies across the active middle and deep cortical layers (Atencio and Schreiner, 2010;

568

Merzenich et al., 2013). Much like cNEs in MGB, cNE spike STRFs in A1 exhibited higher

569

peak-trough differences (PTD) and mutual information (MI) compared to all spike STRFs

570

(Figure 7B-i and C-i). Compared to A1 neurons, MGB neurons displayed significantly higher

571

STRF PTD (all spike STRF PTD, p = 3.1e-42; cNE spike STRF PTD, p = 8.6e-36) and mutual

572

information (all spike STRF MI, p = 1.9e-25; cNE spike STRF MI, p = 4.2e-22). We did not

573

observe, however, a significant difference between MGB and A1 cNEs regarding their gain in

574

cNE spike STRF PTD and MI values over that of member neuron spiking (Figure 7B-ii and C-

575

ii).

576

Addressing concerns of potential false positive detections in A1, we compared the number of

577

cNEs detected on real and shuffled activities. A substantially smaller number of cNEs were

578

detected on shuffled data compared to real data (real data vs. shuffled data, spon1: 4.6 ± 1.5 vs

579

0.9 ± 0.6; spon2: 4.2 ± 1.2 vs 1.1 ± 0.6, dmr1: 4.3 ± 1.5 vs 0.9 ± 0.6, dmr2: 4.2 ± 1.2 vs 1.0 ± 0.8,

580

mean ± SD) (Figure 7D). Moreover, A1 cNEs were mostly stable across stimulus conditions,

581

similar to MGB cNEs, whereas false positive cNEs did not show such stability (Figure 7E). The

582

similarity between MGB and A1 cNEs in their stability across stimulus conditions and enhanced

583

information encoding provides support for the concept of cNEs serving as a universal

584

mechanism for neuronal organization and information processing.

585

JNeurosci Accepted Manuscript

cNE formation does not rely on strong slow oscillations

586

Slow-wave oscillations, characterized by alternating periods of large and sustained network

587

activity (UP states) and neural quiescence (DOWN states), are frequently observed in the cortex

588

and thalamus during anesthesia (Steriade et al., 1993; Contreras et al., 1996; Sanchez-Vives and

589

McCormick, 2000; Hasenstaub et al., 2007; Chauvette et al., 2011; Neske, 2016). To quantify the

590

level of slow oscillations in the recording and their potential influence on cNE properties, we

591

used two measurements: silence density and the coefficient of variation (CV) of the multiunit

592

(MU) firing rate. Silence density represents the proportion of recording time when no spike was

593

fired by the neural population, which is characteristic of the DOWN state in slow oscillations. In

594

brains without strong slow oscillations, the population of neurons fires continuously, resulting in

595

low silence density. The MU firing rate CV measures the level of variation in the firing rate of

596

the MU, which is high for neurons going through UP-DOWN state cycles, but small for neural

597

populations with less synchronized oscillatory activity. Recordings with high silence density and

598

high MU firing rate CV showed prominent slow oscillations, with epochs of synchronous firing

599

of neurons and epochs of quiescence with no spikes (Figure 8A-i). In contrast, recordings with

600

low silence density and low MU firing rate CV did not exhibit strong slow oscillations,

601

displaying relatively stable and continuous firing (Figure 8A-ii). Recordings with moderate

602

silence density and MU firing rate CV exhibited moments of elevated firing, although not as

603

synchronized as in recordings with strong oscillations (Figure 8A-iii). By utilizing silence

604

density and MU firing rate CV, we were able to differentiate recordings without strong slow

605

oscillations from those with strong slow oscillations.

606

JNeurosci Accepted Manuscript

The relationship between cNEs and slow oscillations was investigated by applying the cNE

607

detection algorithm to recordings without strong oscillations in response to DMR during 15-

608

minute recordings. Recordings exhibiting low silence density and low CV of the MU firing rate

609

were selected to ensure the absence of strong slow oscillations (n(animals): MGB = 8, A1 = 18;

610

n(recordings): MGB = 13, A1 = 10). To determine whether cNEs were solely a byproduct of

611

slow oscillations, we compared the number of cNEs detected in these recordings against the

612

expected false positive rate. We found a significantly higher occurrence of cNEs in activity

613

characterized by low silence density and low MU firing rate CV in both MGB (n(cNE) = 3.9 ±

614

1.0, mean ± SD; p = 2.4e-4, Wilcoxon signed-rank test) and A1 (n(cNE) = 5.2 ± 1.9, p = 0.002)

615

when compared to shuffled data (as in Figure 5). This supports the notion that cNEs are not a

616

byproduct of slow oscillations.

617

Slow oscillations in thalamic and cortical firing rates are commonly observed and can be related

618

to synchronized and desynchronized states of the system (Metherate and Ashe, 1993; Steriade et

619

al., 1993; Cowan and Wilson, 1994; Sanchez-Vives and McCormick, 2000; Hasenstaub et al.,

620

2007). However, the effect of firing rate changes on the information carried by cNEs is not

621

known. Therefore, we tested whether cNE spikes exhibit enhanced information encoding in

622

recordings without strong oscillations in response to the stimulus. The results showed that cNE

623

spike STRFs have higher MI compared to all spikes in both MGB and A1 (Figure 8B), indicating

624

that enhanced information encoding is not specific to synchronized states under slow

625

oscillations. Additionally, a subset of recordings did not show strong slow oscillations in either

626

stimulus-driven or spontaneous activity (n(animals): MGB = 4, A1 = 6; n(recordings): MGB = 5,

627

A1 = 8). We also examined the stability of cNEs across and within stimulus conditions in these

628

JNeurosci Accepted Manuscript

recordings. The majority of cNEs were stable both within and across stimulus conditions (Figure

629

8C). In summary, cNEs exist in both MGB and A1 without the presence of strong slow

630

oscillations in neural activity. Their enhanced information encoding and stability across stimulus

631

conditions are not due to a special behavior of neurons in synchronized states under slow

632

oscillations.

633

634

Discussion

635

This study aimed to investigate whether the auditory thalamus (MGB) contains coordinated

636

neural ensembles (cNEs) with enhanced information properties, similar to those observed in A1.

637

Our results confirm the presence of cNEs in the MGB, with consistent compositions across

638

various, but especially smaller, bin sizes and stable structures across different stimulus

639

conditions. Importantly, coordinated spikes among cNE member neurons exhibit higher

640

reliability and convey more stimulus-related information than individual neurons. Neuronal

641

groups formed by shared firing-rate changes to stimuli appear not to be congruent with cNEs.

642

Furthermore, our findings demonstrate that cNEs are not the result of false positive detections or

643

byproducts of slow state oscillations in anesthetized animals. These findings provide support for

644

the notion that synchronized neuronal ensembles represent a general principle of local

645

organization for information processing in the auditory forebrain.

646

cNEs are ubiquitous in local circuit organization

647

Neuronal ensembles were proposed as fundamental units for information processing in the brain

648

(Hebb, 1949; Buzsáki, 2010), supported by evidence of precise temporal coordination in cortical

649

JNeurosci Accepted Manuscript

columns (Atencio and Schreiner, 2013; See et al., 2018; Lankarany et al., 2019). Cortical

650

columns consist of neurons with fairly homogeneous properties maintained through intracortical

651

processing and shared afferent input (Mountcastle, 1997). This raises the question of whether

652

cNEs are unique to cortical organization or represent a general organizational and information

653

processing unit along sensory pathways. In the auditory system, reciprocal connectivity exists

654

between the MGB and A1, with convergence of frequency tuning and spectral and temporal

655

modulation preferences, preserving topographic organization in both regions (Miller et al., 2001,

656

2002; Bartlett and Wang, 2007; Read et al., 2011). Therefore, investigating neuronal

657

coordination in the MGB, where neurons possess similar properties but differ in their

658

organizational and cytoarchitectonic patterns from A1 (Winer, 2010), can provide insights into

659

whether cNEs are general organizational principles of local circuits or specialized units specific

660

to the cortical circuit composition.

661

We conducted recordings of neuronal activity across multiple iso-frequency layers of the MGB

662

and were able to reliably detect cNEs in MGB (Figure 2). Neurons within the same cNE

663

displayed closer spatial proximity and shared more similar tuning properties (Figure 5D and E),

664

indicating functional coherence within cNEs. It is important to note that our recordings were

665

limited to relatively small populations of 10-30 neurons due to the techniques employed.

666

Therefore, the confinement of spatial and frequency tuning properties within cNEs may vary

667

when larger populations with hundreds or thousands of neurons are recorded.

668

We further demonstrated that the identification of cNEs relies on the temporal coordination

669

among neurons. When the original temporal order among neurons was disrupted through circular

670

shuffling of spike trains, a significantly lower number of cNEs were detected in both the MGB

671

JNeurosci Accepted Manuscript

and A1 (Figure 4F and 7D). Moreover, the few false positive cNEs that were identified did not

672

exhibit the properties observed in cNEs identified in the real data, such as stability across

673

different stimulus conditions (Figure 4E and 7E). These findings provide strong support for the

674

critical role of temporal coordination in the formation and characterization of cNEs in the

675

auditory thalamus and cortex.

676

Time scale of cNEs

677

Previous studies have investigated neuronal synchrony and coordination across widely differing

678

timescales, ranging from a few milliseconds (Lankarany et al., 2019; Shahidi et al., 2019; El-

679

Gaby et al., 2021) to several hundred milliseconds (Miller et al., 2014; Tremblay et al., 2015;

680

Filipchuk et al., 2022). The selection of a specific timescale in these studies was influenced by

681

various factors, including the temporal resolution of the recording methods used, the targeted

682

functional timescale, and the inter-neuronal distance under investigation. In the context of

683

auditory processing, where information changes rapidly within tens of milliseconds (Rosen,

684

1992; Lewicki, 2002), we specifically chose a temporal resolution of 10ms. This choice aligns

685

with the timescale at which auditory information operates and holds relevance for synaptic

686

integration. Selecting an appropriate timescale is crucial for future investigations into the

687

functional role of cNEs in synaptic transmission within the auditory thalamocortical system.

688

We have demonstrated the robustness of cNE identification across different time bin sizes

689

(Figure 3), which can be attributed to the sparse nature of neural activity. Since most

690

synchronized neuronal firing occurs at frequencies below 10 Hz (O’Connor et al., 2010), the

691

choice of time windows, whether 10ms or 20ms, has minimal impact on the observed

692

correlations among neurons. However, it is important to note that cNEs identified with longer

693

JNeurosci Accepted Manuscript

time resolutions (hundreds of milliseconds) may significantly differ from those identified with

694

shorter resolutions (tens of milliseconds). Specifically, cNEs identified with larger time bins may

695

falsely include 'synchronous' events from bursting or rebound activity rather than from an initial

696

period that dominates the transmission of stimulus-triggered information. Long synchronization

697

windows may also include neurons displaying weaker synchronization within short time

698

windows, while potentially de-emphasizing neurons with high temporal precision in synchrony

699

(Figure 3D). Hence, the chosen temporal resolution influences the composition and properties of

700

cNEs, emphasizing the importance of selecting the appropriate timescale for studying neural

701

ensembles.

702

Stability of cNEs for spontaneous and evoked activity

703

Our study revealed that a significant proportion of cNEs (55% in MGB and 76% in A1)

704

maintained a consistent composition during both spontaneous and sensory-evoked neural activity

705

(Figure 4E and 7E). This suggests that cNEs generally represent stable configurations within

706

local circuits that can manifest independently of stimulus-driven synchrony. These findings align

707

with previous research demonstrating similarities between patterns observed in spontaneous and

708

stimulus-driven activity (Luczak et al., 2009). Moreover, the similarity between MGB and A1

709

cNEs indicates that functional network units are not limited to cortical organization but likely

710

exist as a common modality across multiple stages of the sensory pathway.

711

cNEs enhance stimulus encoding

712

Considering that cNEs were observed in both spontaneous and stimulus-driven activity, some

713

argue that they are merely a reflection of background activity and not involved in stimulus

714

JNeurosci Accepted Manuscript

encoding and even potentially impairing it (Zohary et al., 1994; Abbott and Dayan, 1999;

715

Jermakowicz et al., 2009). Contrary to this notion, our observations revealed that cNE spikes

716

exhibit a higher signal-to-noise ratio and convey more information per spike when compared to

717

the entire spike train (Figure 6B and C). This suggests that cNE events are more stimulus-

718

selective than the contributing neurons (See et al., 2021) and exhibit a more reliable response to

719

the stimulus features represented by the cNE STRF. The stable connectivity pattern revealed by

720

spontaneous, intrinsic activity likely reveals aspects that have been imprinted by extensive

721

experience and the behavioral relevance of the associated functional preferences.

722

Additionally, we observed an enhanced information encoding in cNE groups and cNE spikes

723

when compared to non-cNE groups or coincident spikes, with control of the total number of

724

neurons in the groups (Figure 6B and C). This observation suggests that the information increase

725

relies on the coordination among cNE member neurons, rather than being a simple result of

726

independent population coding (deCharms, 1998; Hatsopoulos et al., 1998).

727

Correlated spikes can enhance the transmission of salient auditory information by synchronously

728

converging onto their targets (Stevens and Zador, 1998; Zandvakili and Kohn, 2015).

729

Additionally, neurons exhibit a multiplexed nature of stimulus encoding, where spikes from the

730

same neuron can carry information related to distinct stimulus aspects (Walker et al., 2011;

731

Lankarany et al., 2019; See et al., 2021). The function of cNEs may involve selectively choosing

732

spikes from member neurons that are most relevant for a specific target information and

733

enhancing information propagation while excluding functionally irrelevant spikes of the same

734

neurons. This mechanism significantly improves both the robustness and capacity of information

735

encoded within a population of neurons (Walker et al., 2011; See et al., 2021). Thus, the

736

JNeurosci Accepted Manuscript

presence of cNEs and their coordination within a neuronal population can facilitate efficient

737

information processing and transmission in the auditory system. Future studies involving

738

simultaneous recordings from two stations along the auditory pathway will be necessary to test

739

this hypothesis.

740

cNE formation does not depend on strong slow oscillations

741

Slow oscillations are commonly observed in neural activity during anesthesia (Chauvette et al.,

742

2011; Dasilva et al., 2021) and have been shown to influence stimulus encoding (Pachitariu et

743

al., 2015). Concerns have been raised regarding whether cNEs are solely a result of anesthesia-

744

induced synchrony. However, our research findings refute this notion. We focused on a distinct

745

time scale of synchronization unrelated to anesthesia-induced slow oscillations and successfully

746

detected cNEs in recordings without strong slow oscillations. These cNEs exhibited stable

747

structures and enhanced information properties, indicating that they are not solely a byproduct of

748

anesthesia-induced synchrony. While we ruled out slow oscillations as the primary force

749

underlying cNE formation, it is important to consider their potential interaction with other

750

oscillatory activity, such as gamma rhythms (Oberto et al., 2021). Further research is needed to

751

explore the interplay between cNEs and different types of brain oscillations.

752

753

JNeurosci Accepted Manuscript

References

755

Abbott LF, Dayan P (1999) The effect of correlated variability on the accuracy of a population

756

code. Neural Comput 11:91–101.

757

Anderson LA, Linden JF (2011) Physiological differences between histologically defined

758

subdivisions in the mouse auditory thalamus. Hear Res 274:48–60.

759

Aschauer DF, Eppler JB, Ewig L, Chambers AR, Pokorny C, Kaschube M, Rumpel S (2022)

760

Learning-induced biases in the ongoing dynamics of sensory representations predict

761

stimulus generalization. Cell Rep 38:110340.

762

Atencio CA, Schreiner CE (2008) Spectrotemporal processing differences between auditory

763

cortical fast-spiking and regular-spiking neurons. J Neurosci 28:3897–3910.

764

Atencio CA, Schreiner CE (2010) Laminar diversity of dynamic sound processing in cat primary

765

auditory cortex. J Neurophysiol 103:192–205.

766

Atencio CA, Schreiner CE (2013) Auditory cortical local subnetworks are characterized by

767

sharply synchronous activity. J Neurosci 33:18503–18514.

768

Baeg EH, Kim YB, Huh K, Mook-Jung I, Kim HT, Jung MW (2003) Dynamics of population

769

code for working memory in the prefrontal cortex. Neuron 40:177–188.

770

Bartlett EL (2013) The organization and physiology of the auditory thalamus and its role in

771

processing acoustic features important for speech perception. Brain Lang 126:29–48.

772

Bartlett EL, Wang X (2007) Neural representations of temporally modulated signals in the

773

auditory thalamus of awake primates. J Neurophysiol 97:1005–1017.

774

JNeurosci Accepted Manuscript

Bathellier B, Ushakova L, Rumpel S (2012) Discrete neocortical dynamics predict behavioral

775

categorization of sounds. Neuron 76:435–449.

776

Bizley JK, Walker KMM, King AJ, Schnupp JWH (2010) Neural ensemble codes for stimulus

777

periodicity in auditory cortex. J Neurosci 30:5078–5091.

778

Boucly CJ, Pompili MN, Todorova R, Leroux EM, Wiener SI, Zugaro M (2022) Flexible

779

communication between cell assemblies and ‘reader’neurons. bioRxiv:2009–2022.

780

Brosch M, Schreiner CE (1999) Correlations between neural discharges are related to receptive

781

field properties in cat primary auditory cortex. Eur J Neurosci. 11:3517-3530.

782

Buzsáki G (2010) Neural Syntax: Cell Assemblies, Synapsembles, and Readers. Neuron 68:362–

783

385.

784

Carrillo-Reid L, Miller JEK, Hamm JP, Jackson J, Yuste R (2015) Endogenous sequential

785

cortical activity evoked by visual stimuli. J Neurosci 35:8813–8828.

786

Chamberland S, Yang HH, Pan MM, Evans SW, Guan S, Chavarha M, Yang Y, Salesse C, Wu

787

H, Wu JC (2017) Fast two-photon imaging of subcellular voltage dynamics in neuronal

788

tissue with genetically encoded indicators. Elife 6:e25690.

789

Chauvette S, Crochet S, Volgushev M, Timofeev I (2011) Properties of slow oscillation during

790

slow-wave sleep and anesthesia in cats. J Neurosci 31:14998–15008.

791

Contreras D, Timofeev I, Steriade M (1996) Mechanisms of long-lasting hyperpolarizations

792

underlying slow sleep oscillations in cat corticothalamic networks. J Physiol 494:251–264.

793

Cowan RL, Wilson CJ (1994) Spontaneous firing patterns and axonal projections of single

794

JNeurosci Accepted Manuscript

corticostriatal neurons in the rat medial agranular cortex. J Neurophysiol 71:17–32.

795

D’amour JA, Froemke RC (2015) Inhibitory and excitatory spike-timing-dependent plasticity in

796

the auditory cortex. Neuron 86:514–528.

797

Dan Y, Alonso JM, Usrey WM, Reid RC (1998) Coding of visual information by precisely

798

correlated spikes in the lateral geniculate nucleus. Nat Neurosci 1:501–507.

799

Dasilva M, Camassa A, Navarro-Guzman A, Pazienti A, Perez-Mendez L, Zamora-López G,

800

Mattia M, Sanchez-Vives M V (2021) Modulation of cortical slow oscillations and

801

complexity across anesthesia levels. Neuroimage 224:117415.

802

deCharms RC (1998) Information coding in the cortex by independent or coordinated

803

populations. Proc Natl Acad Sci 95:15166–15168.

804

deCharms RC, Merzenich MM (1996) Primary cortical representation of sounds by the

805

coordination of action-potential timing. Nature 381:610–613.

806

Domanski APF, Kucewicz MT, Russo E, Tricklebank MD, Robinson ESJ, Durstewitz D, Jones

807

MW (2023) Distinct hippocampal-prefrontal neural assemblies coordinate memory

808

encoding, maintenance, and recall. Curr Biol 33:1220–1236.

809

Ebrahimi S, Lecoq J, Rumyantsev O, Tasci T, Zhang Y, Irimia C, Li J, Ganguli S, Schnitzer MJ

810

(2022) Emergent reliability in sensory cortical coding and inter-area communication. Nature

811

605:713–721.

812

Eggermont JJ (2000) Sound-induced synchronization of neural activity between and within three

813

auditory cortical areas. (2000) J. Neurophysiol.83:2708-2722.

814

JNeurosci Accepted Manuscript

El-Gaby M, Reeve HM, Lopes-dos-Santos V, Campo-Urriza N, Perestenko P V, Morley A,

815

Strickland LAM, Lukács IP, Paulsen O, Dupret D (2021) An emergent neural coactivity

816

code for dynamic memory. Nat Neurosci 24:694–704.

817

Escabí MA, Schreiner CE (2002) Nonlinear Spectrotemporal Sound Analysis by Neurons in the

818

Auditory Midbrain. J Neurosci 22:4114–4131.

819

Filipchuk A, Schwenkgrub J, Destexhe A, Bathellier B (2022) Awake perception is associated

820

with dedicated neuronal assemblies in the cerebral cortex. Nat Neurosci 25:1327–1338.

821

Harris KD, Csicsvari J, Hirase H, Dragoi G, Buzsáki G (2003) Organization of cell assemblies in

822

the hippocampus. Nature 424:552–556.

823

Hasenstaub A, Sachdev RNS, McCormick DA (2007) State changes rapidly modulate cortical

824

neuronal responsiveness. J Neurosci 27:9607–9622.

825

Hatsopoulos NG, Ojakangas CL, Paninski L, Donoghue JP (1998) Information about movement

826

direction obtained from synchronous activity of motor cortical neurons. Proc Natl Acad Sci

827

U S A 95:15706–15711.

828

Hebb DO (1949) The organization of behavior. Wiley.

829

Homma NY, Hullett PW, Atencio CA, Schreiner CE (2020) Auditory Cortical Plasticity

830

Dependent on Environmental Noise Statistics. Cell Rep 30:4445-4458.e5.

831

Ince RAA, Panzeri S, Kayser C (2013) Neural codes formed by small and temporally precise

832

populations in auditory cortex. J Neurosci 33:18277–18287.

833

Jermakowicz WJ, Chen X, Khaytin I, Bonds AB, Casagrande VA (2009) Relationship between

834

JNeurosci Accepted Manuscript

spontaneous and evoked spike-time correlations in primate visual cortex. J Neurophysiol

835

101:2279–2289.

836

Konorski J (1948) Conditioned reflexes and neuron organization.

837

Kreiter AK, Singer W (1996) Stimulus-dependent synchronization of neuronal responses in the

838

visual cortex of the awake macaque monkey. J Neurosci 16:2381–2396.

839

Lankarany M, Al-Basha D, Ratté S, Prescott SA (2019) Differentially synchronized spiking

840

enables multiplexed neural coding. Proc Natl Acad Sci U S A 116:10097–10102.

841

Laubach M, Wessberg J, Nicolelis MAL (2000) Cortical ensemble activity increasingly predicts

842

behaviour outcomes during learning of a motor task. Nature 405:567–571.

843

Léger JF, Stern EA, Aertsen A, Heck D (2005) Synaptic integration in rat frontal cortex shaped

844

by network activity. J Neurophysiol 93:281–293.

845

Lopes-dos-Santos V, Ribeiro S, Tort ABL (2013) Detecting cell assemblies in large neuronal

846

populations. J Neurosci Methods 220:149–166.

847

Luczak A, Barthó P, Harris KD (2009) Spontaneous Events Outline the Realm of Possible

848

Sensory Responses in Neocortical Populations. Neuron 62:413–425.

849

Marchenko VA, Pastur LA (1967) Distribution of eigenvalues for some sets of random matrices.

850

Mat Sb 114:507–536.

851

Merzenich MM, Knight PL, Roth GL, Linn G (2013) Representation of cochlea within primary

852

auditory cortex in the cat of Cochlea in the Cat. :231–249.

853

JNeurosci Accepted Manuscript

Metherate R, Ashe JH (1993) Nucleus basalis stimulation facilitates thalamocortical synaptic

854

transmission in the rat auditory cortex. Synapse 14:132–143.

855

Miller JEK, Ayzenshtat I, Carrillo-Reid L, Yuste R (2014) Visual stimuli recruit intrinsically

856

generated cortical ensembles. Proc Natl Acad Sci U S A 111:E4053–E4061.

857

Miller LM, Escabí MA, Read HL, Schreiner CE (2001) Functional convergence of response

858

properties in the auditory thalamocortical system. Neuron 32:151–160.

859

Miller LM, Escabí MA, Read HL, Schreiner CE (2002) Spectrotemporal receptive fields in the

860

lemniscal auditory thalamus and cortex. J Neurophysiol 87:516–527.

861

Miller LM, Recanzone GH (2009) Populations of auditory cortical neurons can accurately

862

encode acoustic space across stimulus intensity. Proc Natl Acad Sci 106:5931–5935.

863

Miller LM, Schreiner CE (2000) Stimulus-based state control in the thalamocortical system. J

864

Neurosci 20:7011–7016.

865

Mochol G, Hermoso-Mendizabal A, Sakata S, Harris KD, De La Rocha J (2015) Stochastic

866

transitions into silence cause noise correlations in cortical circuits. Proc Natl Acad Sci U S

867

A 112:3529–3534.

868

Mogensen H, Norrlid J, Enander JMD, Wahlbom A, Jörntell H (2019) Absence of repetitive

869

correlation patterns between pairs of adjacent neocortical neurons in vivo. Front Neural

870

Circuits 13:1–11.

871

Morel A, Rouiller E, de Ribaupierre Y, de Ribaupierre F (1987) Tonotopic organization in the

872

medial geniculate body (MGB) of lightly anesthetized cats. Exp Brain Res 69:24–42.

873

JNeurosci Accepted Manuscript

Mountcastle VB (1997) The columnar organization of the neocortex. Brain a J Neurol 120:701–

874

722.

875

Neske GT (2016) The slow oscillation in cortical and thalamic networks: Mechanisms and

876

functions. Front Neural Circuits 9:1–25.

877

Niessing J, Friedrich RW (2010) Olfactory pattern classification by discrete neuronal network

878

states. Nature 465:47–52.

879

O’Connor DH, Peron SP, Huber D, Svoboda K (2010) Neural activity in barrel cortex underlying

880

vibrissa-based object localization in mice. Neuron 67:1048–1061.

881

Oberto VJ, Boucly CJ, Gao H, Todorova R, Zugaro MB, Wiener SI (2022) Distributed cell

882

assemblies spanning prefrontal cortex and striatum. Curr Biol:1–13.

883

Pachitariu M, Lyamzin DR, Sahani M, Lesica NA (2015) State-dependent population coding in

884

primary auditory cortex. J Neurosci 35:2058–2073.

885

Pachitariu M, Sridhar S, Stringer C (2023) Solving the spike sorting problem with Kilosort.

886

bioRxiv:2023.01.07.523036.

887

Paninski L, Shoham S, Fellows MR, Hatsopoulos NG, Donoghue JP (2004) Superlinear

888

population encoding of dynamic hand trajectory in primary motor cortex. J Neurosci

889

24:8551–8561.

890

Peyrache A, Benchenane K, Khamassi M, Wiener SI, Battaglia FP (2010) Principal component

891

analysis of ensemble recordings reveals cell assemblies at high temporal resolution. J

892

Comput Neurosci 29:309–325.

893

JNeurosci Accepted Manuscript

Polley DB, Read HL, Storace DA, Merzenich MM (2007) Multiparametric auditory receptive

894

field organization across five cortical fields in the albino rat. J Neurophysiol 97:3621–3638.

895

Qiu A, Schreiner CE, Escabí MA (2003) Gabor analysis of auditory midbrain receptive fields:

896

Spectro-temporal and binaural composition. J Neurophysiol 90:456–476.

897

Read HL, Nauen DW, Escabí MA, Miller LM, Schreiner CE, Winer JA (2011) Distinct core

898

thalamocortical pathways to central and dorsal primary auditory cortex. Hear Res 274:95–

899

104.

900

Rosen S (1992) Temporal information in speech: acoustic, auditory and linguistic aspects. Philos

901

Trans R Soc London Ser B Biol Sci 336:367–373.

902

Sanchez-Vives MV, McCormick DA (2000) Cellular and network mechanisms of rhythmic

903

recurrent activity in neocortex. Nat Neurosci 3:1027–1034.

904

See JZ, Atencio CA, Sohal VS, Schreiner CE (2018) Coordinated neuronal ensembles in primary

905

auditory cortical columns. Elife 7:1–33.

906

See JZ, Homma NY, Atencio CA, Sohal VS, Schreiner CE (2021) Information diversity in

907

individual auditory cortical neurons is associated with functionally distinct coordinated

908

neuronal ensembles. Sci Rep 11:1–15.

909

Shahidi N, Andrei AR, Hu M, Dragoi V (2019) High-order coordination of cortical spiking

910

activity modulates perceptual accuracy. Nat Neurosci 22:1148–1158.

911

Smith PH, Uhlrich DJ, Manning KA, Banks MI (2012) Thalamocortical projections to rat

912

auditory cortex from the ventral and dorsal divisions of the medial geniculate nucleus. J

913

JNeurosci Accepted Manuscript

Comp Neurol 520:34–51.

914

Steinmetz NA et al. (2021) Neuropixels 2.0: A miniaturized high-density probe for stable, long-

915

term brain recordings. Science (80- ) 372.

916

Steriade M, McCormick DA, Sejnowski TJ (1993) Thalamocortical oscillations in the sleeping

917

and aroused brain. Science (80- ) 262:679–685.

918

Stevens CF, Zador AM (1998) Input synchrony and the irregular firing of cortical neurons. Nat

919

Neurosci 1:210–217.

920

Theunissen FE, Sen K, Doupe AJ (2000) Spectral-temporal receptive fields of nonlinear auditory

921

neurons obtained using natural sounds. J Neurosci 20:2315–2331.

922

Tremblay S, Pieper F, Sachs A, Martinez-Trujillo J (2015) Attentional filtering of visual

923

information by neuronal ensembles in the primate lateral prefrontal cortex. Neuron 85:202–

924

215.

925

Wahlbom A, Mogensen H, Jörntell H (2021) Widely Different Correlation Patterns Between

926

Pairs of Adjacent Thalamic Neurons In vivo. Front Neural Circuits 15:1–10.

927

Walker KMM, Bizley JK, King AJ, Schnupp JWH (2011) Multiplexed and robust

928

representations of sound features in auditory cortex. J Neurosci 31:14565–14576.

929

Wills TJ, Lever C, Cacucci F, Burgess N, O’Keefe J (2005) Attractor dynamics in the

930

hippocampal representation of the local environment. Science (80- ) 308:873–876.

931

Winer JA (2010) Neurochemical organization of the medial geniculate body and auditory cortex.

932

In: The auditory cortex, pp 209–234. Springer.

933

JNeurosci Accepted Manuscript

Winer JA, Miller LM, Lee CC, Schreiner CE (2005) Auditory thalamocortical transformation:

934

Structure and function. Trends Neurosci 28:255–263.

935

Yoshida T, Ohki K (2020) Natural images are reliably represented by sparse and variable

936

populations of neurons in visual cortex. Nat Commun 11:872.

937

Yuste R (2015) From the neuron doctrine to neural networks. Nat Rev Neurosci 16:487–497.

938

Zandvakili A, Kohn A (2015) Coordinated Neuronal Activity Enhances Corticocortical

939

Communication. Neuron 87:827–839.

940

Zohary E, Shadlen MN, Newsome WT (1994) Correlated neuronal discharge rate and its

941

implications for psychophysical performance. Nature 370:140–143.

942

Figure Legends

943

Figure 1. In vivo recordings in rat MGB

944

(A) Left: Schematic of the recording setup in the MGB using a linear 64-channel probe. (i) and

945

(ii): two electrode penetrations with multi-unit (MU) recordings from the MGB. Left: Stacked

946

firing rate (color coded) of pure-tone frequency response areas. Right: Characteristic frequencies

947

(CF, the frequency at which the response threshold is the lowest). The red dashed lines indicate

948

the potential boundaries of the ventral MGB. (B) Example STRFs of single units (SUs) from a

949

recording in the MGB. Unit numbers 1 to 3 indicate the positions and STRFs of pairs of neurons

950

whose CCGs are plotted in (C). (C) Example CCGs from two pairs of neurons (#1 - #3 and #2 -

951

#3). The black bars represent the CCGs of stimulus-driven activity, while the grey lines represent

952

the CCGs of spontaneous activity. The baseline is estimated by averaging the counts in 5ms

953

windows at the shoulders of the CCGs and is indicated by dashed red lines.

954

JNeurosci Accepted Manuscript

Figure 2 Groups of neurons with coordinated activities exist in MGB

955

(A) Procedures for detecting cNEs in a thalamic penetration. (i) Correlation matrix of spike

956

trains. (ii) Eigenvalues of the correlation matrix shown in (i). The dashed red line represents the

957

99.5th percentile of the Marchenko–Pastur distribution, which was used as the significance

958

threshold for eigenvalues. The top four eigenvalues are significant and represent the number of

959

detected cNEs. (iii) IC weights of neurons for each cNE. The green dots represent neurons that

960

are members of a cNE. (iv) cNE members (red stems) are neurons with IC weights exceeding the

961

threshold (1

√

) shown as grey areas. (v) Example of cNE activation. Top: Activity trace of cNE

962

#1. The red line shows the threshold estimated using Monte Carlo methods. The peaks crossing

963

the threshold indicate cNE events when multiple cNE member neurons fire jointly. Bottom:

964

Spike raster of neurons, with red ticks indicating spikes that contribute to instances of cNE

965

events, which were referred to as cNE spikes. Shaded areas show member neurons. (B)

966

Correlations (10ms bin) of neuron pairs that were both members of the same cNE (members) or

967

neuron pairs that were not members of the same cNE (non-members) in MGB (p = 6.9e-235,

968

Mann–Whitney U test). (C) Z-scored CCGs (1ms bin) of member pairs (left) and non-member

969

pairs (right) in MGB. Top: Stacked CCGs ordered by the peak delay. Bottom: Average of the

970

data above (mean ± SD; shaded area: p < 0.01, permutation test, shuffling the members/non-

971

members labels).

972

Figure 3. Variability of IC weights across different time bin sizes.

973

(A) Example of two MGB cNEs whose member neurons are either consistently identified across

974

different bin sizes (i) or only detected using smaller bin sizes (ii). (B) Proportion of significantly

975

matched cNEs. Using different bin sizes as reference bin sizes (row), we calculated the

976

JNeurosci Accepted Manuscript

proportion of cNEs on different bin sizes that have a significantly matched cNE on the reference

977

bin size. The red square highlights the proportion of 10ms cNEs that have significant matches

978

with 160ms cNEs and is further analyzed in (C). (C) Left: Correlation of IC weights of 10ms

979

cNEs with the most correlated IC weights of 160ms cNEs. Right: Proportion of shared

980

membership between 10ms cNEs with their most correlated 160ms cNEs. (D) Firing rate of

981

member neurons in 10ms-only cNEs, 10ms cNEs without a significant match with 160ms cNEs,

982

and 160ms-only cNEs (p = 0.57, Kolmogorov-Smirnov test). (E) Top: Schematic of membership

983

of neurons for different bin sizes. Bottom: Mean z-scored CCGs (1ms bin, mean ± SD) between

984

(i) member neurons in both 10ms and 160ms cNEs, (ii) member neurons only in the 160ms cNE

985

and member neurons in both 10ms and 160ms cNEs (shaded area: p < 0.01, permutation test,

986

shuffling neuron pair labels of i and ii), and (iii) member neurons only in the 10ms cNE and

987

member neurons in both 10ms and 160ms cNEs (no time bin showed significant difference from

988

i).

989

Figure 4. cNEs identified in spontaneous activity are mostly preserved in stimulus-driven

990

activity.

991

(A) Diagram illustrating the recording sequence and partitioning of spontaneous (yellow/orange)

992

and DMR-evoked (dark green/light green) activity. The four blocks allowed the comparison of

993

cNEs obtained within stimulus conditions and across stimulus conditions. (B) Absolute

994

correlation values of the IC weights calculated on adjacent blocks from an MGB recording

995

including the two examples (i and ii) shown in (C). (C) Example IC weights on adjacent blocks

996

with high (i) and moderate (ii) correlation values across stimulus conditions. The dashed lines

997

show the threshold to determine the membership of the neurons. (D) The two cNE examples in

998

JNeurosci Accepted Manuscript

999

null distribution was generated. The significance threshold for the correlation values was set at p

1000

= 0.01 (red dashed line, 99.5th percentile of the null distribution). The brown (i) and blue (ii)

1001

solid lines represent the two examples in (C). (E) Correlations of IC weights identified on

1002

adjacent activity blocks for real (red) and circularly shifted data (pink). The hollow histograms

1003

show all correlations of matched cNEs on adjacent blocks; the histograms show significant

1004

correlations based on the test shown in (D). The inset numbers show the percentage of cNEs with

1005

significantly matched IC weights on adjacent blocks. The triangles show the median of all IC

1006

weight correlations (real data vs shuffled data: spon, p = 5.6e-36; dmr: p = 1.3e-28; cross, p =

1007

2.0e-22, Mann–Whitney U test with Bonferroni correction). (F) Number of cNEs detected using

1008

real and circularly shifted activities on the four recording blocks (spon1, p = 4.7e-10; spon2, p =

1009

4.7e-10; dmr1, p = 4.7e-10; dmr2, p = 4.7e-10, Wilcoxon signed-rank test with Bonferroni

1010

correction).

1011

Figure 5. Properties of cNEs in MGB.

1012

(A) Number of cNEs detected in any given penetration increases with the number of recorded

1013

neurons. (B) cNE size increases with the number of recorded neurons. (C) The number of cNEs a

1014

neuron belongs to. (D) Spatial distribution of cNE members. (i) Pairwise distance of neurons in

1015

the same cNE (colored bar) or neurons not in the same cNE (black line). (ii) Spatial span of cNE

1016

members (colored bar) and random groups of neurons with the same number of neurons as cNEs

1017

(black line). (E) Frequency tuning distribution of cNE members. (i) Pairwise difference in the

1018

best frequencies (BF) of neurons. (ii) The largest difference in the BF among cNE members or

1019

random groups of neurons. (Mann–Whitney U test).

1020

JNeurosci Accepted Manuscript

Figure 6. MGB cNEs can refine sound features encoded by member neurons.

1021

(A) Two example STRFs of MGB neurons calculated with all spikes (left) and cNE spikes

1022

(right). All spikes are subsampled to have an equal number as cNE spikes. (B) (i) STRF peak-

1023

trough difference (PTD) for cNE spikes or all spikes of neurons. (ii) STRF PTD for groups of

1024

cNE members or non-members. (iii) STRF PTD for cNE spikes and coincident spikes of a

1025

neuron. The coincident spikes refer to instances where a neuron's firing occurs within a 10ms

1026

timeframe of another neuron's firing in a group of non-member neurons. This group is designed

1027

to match the number of neurons present in the cNE. (C) (i) Mutual information (MI) between

1028

stimulus and cNE spikes or all spikes of neurons. (ii) STRF MI for groups of cNE members or

1029

non-members. (iii) STRF MI for cNE spikes and coincident spikes of a neuron. (Wilcoxon

1030

signed-rank test).

1031

Figure 7. MGB and A1 cNEs have similar properties.

1032

(A) Left: schematic of the recording setup in A1 using a 2-shank probe with 64 channels. Right:

1033

MU responses to pure tones as in Figure 1A. (B) (i) PTD of STRFs calculated using all spikes or

1034

only cNE spikes from a neuron in A1. (ii) Difference between cNE spike STRF PTD and all

1035

spike STRF PTD in MGB and A1 (p = 0.72, Mann–Whitney U test). (C) (i) MI of STRFs

1036

calculated using all spikes or only cNE spikes from a neuron in A1. (ii) Difference between cNE

1037

spike STRF MI and all spike STRF MI in MGB and A1 (p = 0.92, Mann–Whitney U test). (D)

1038

Number of cNEs detected using real and circularly shifted activities on the four recording blocks

1039

in A1, as shown in Figure 4F for MGB. (spon1: p = 6.1e-5, spon2: p = 6.1e-5, dmr1: p = 6.1e-5,

1040

dmr2: p = 6.1e-5, n = 17 recordings, Wilcoxon signed-rank test with Bonferroni correction). (E)

1041

JNeurosci Accepted Manuscript

Correlations of IC weights identified on adjacent activity blocks for real (blue) and circularly

1042

shifted data (light blue) in A1. The hollow histograms and histograms show the distribution of

1043

correlations and significant correlations as in Figure 4E. The triangles show the median of IC

1044

weight correlations (spon: p = 1.4e-33, dmr: p = 4.6e-34, cross: p = 4.8e-28, Mann–Whitney U

1045

test with Bonferroni correction).

1046

Figure 8. cNE events do not rely on slow oscillation in neural activity.

1047

(A) Silence density and firing rate (FR) coefficient of variation (CV) of population activity in

1048

response to DMR. Some recordings show strong slow oscillation in population activity with

1049

pronounced silent period between highly active moments (i), while others show little (ii) or

1050

moderate (iii) levels of slow MU firing rate oscillation. Recordings with silence density

0.4

1051

and MU firing rate coefficient of variance

0.8 (dashed lines) did not show prominent slow

1052

oscillation in population activity and were included in (B). (B) STRF MI with all spikes and cNE

1053

spikes in recordings without prominent slow oscillations (Wilcoxon signed-rank test). (C)

1054

Correlation values of cNE IC weights on adjacent activity blocks from recordings with no

1055

prominent slow oscillations in both spontaneous and stimulus-driven activities. The inset

1056

numbers show the percentage of cNEs with significantly matched IC weights on adjacent blocks

1057

in MGB (red) and A1 (blue). The hollow histograms and histograms show all correlation values

1058

and significant correlation values with the same presentation scheme as in Figure 4E.

1059

JNeurosci Accepted Manuscript