The Journal of Neuroscience, April 21, 2021 • 41(16):3665–3678 • 3665

Systems/Circuits

Cortex-Wide Dynamics of Intrinsic Electrical Activities:

Propagating Waves and Their Interactions
Yuqi Liang,1,2 Chenchen Song,3 Mianxin Liu,1,8 Pulin Gong,4,5 Changsong Zhou,1,2,6,7 and Thomas Knöpfel3
1
Department of Physics, Centre for Nonlinear Studies and Beijing-Hong Kong-Singapore Joint Centre for Nonlinear and Complex Systems (Hong
Kong), Institute of Computational and Theoretical Studies, Hong Kong Baptist University, Kowloon Tong, Kowloon, Hong Kong, People’s Republic
of China, 2The HKBU Institute of Research and Continuing Education, Shenzhen 518000, People's Republic of China, 3Laboratory for Neuronal
Circuit Dynamics, Imperial College London, London SW7 2AZ, United Kingdom, 4School of Physics, University of Sydney, Sydney 2006, New South
Wales, Australia, 5Australian Research Council Centre of Excellence for Integrative Brain Function, University of Sydney, Sydney 2001, New South
Wales, Australia, 6Department of Physics, Zhejiang University, Hangzhou 310027, People’s Republic of China, 7Beijing Computational Science
Research Center, Beijing 100193, People’s Republic of China, and 8School of Biomedical Engineering, Shanghai Tech University, Shanghai 201210,
People’s Republic of China

Cortical circuits generate patterned activities that reflect intrinsic brain dynamics that lay the foundation for any, including
stimuli-evoked, cognition and behavior. However, the spatiotemporal organization properties and principles of this intrinsic
activity have only been partially elucidated because of previous poor resolution of experimental data and limited analysis
methods. Here we investigated continuous wave patterns in the 0.5–4 Hz (delta band) frequency range on data from high-
spatiotemporal resolution optical voltage imaging of the upper cortical layers in anesthetized mice. Waves of population activ-
ities propagate in heterogeneous directions to coordinate neuronal activities between different brain regions. The complex
wave patterns show characteristics of both stereotypy and variety. The location and type of wave patterns determine the dy-
namical evolution when different waves interact with each other. Local wave patterns of source, sink, or saddle emerge at
preferred spatial locations. Specifically, “source” patterns are predominantly found in cortical regions with low multimodal
hierarchy such as the primary somatosensory cortex. Our findings reveal principles that govern the spatiotemporal dynamics
of spontaneous cortical activities and associate them with the structural architecture across the cortex.
Key words: cortex; dynamics; fluorescence; GEVI; imaging; mouse

Significance Statement
Intrinsic brain activities, as opposed to external stimulus-evoked responses, have increasingly gained attention, but it remains
unclear how these intrinsic activities are spatiotemporally organized at the cortex-wide scale. By taking advantage of the high
spatiotemporal resolution of optical voltage imaging, we identified five wave pattern types, and revealed the organization
properties of different wave patterns and the dynamical mechanisms when they interact with each other. Moreover, we found
a relationship between the emergence probability of local wave patterns and the multimodal structure hierarchy across corti-
cal areas. Our findings reveal the principles of spatiotemporal wave dynamics of spontaneous activities and associate them
with the underlying hierarchical architecture across the cortex.

Received Mar. 16, 2020; revised Feb. 18, 2021; accepted Feb. 22, 2021. Introduction
Author contributions: T.K., C.Z. and P.G. designed research; Y.L., C.S., and T.K. performed research; Y.L., M.
L., P.G., C.Z., and T.K. analyzed data; Y.L., C.Z., T.K., P.G., C.S. wrote the paper.
During anesthesia and sleep but also in quiet wakefulness, the
This work was supported by Hong Kong Baptist University Strategic Development Fund, the Hong Kong brain exhibits spontaneous activities that reflect self-organized
Research Grant Council (Grant GRF12200217), and the National Science Foundation of China (Grant 11975194) intrinsic brain dynamics, which are independent from external
to C.Z.; National Institutes of Health BRAIN Initiative Grants 1U01-MH-109091 and 5U01-NS-099573 to T.K.; stimuli application or task performance and, at least in the case
and the Australian Research Council (Grant DP160104316) to P.G.
The authors declare no competing financial interests.
of general anesthesia, do not relate to motor behavior. In recent
Correspondence should be addressed to Thomas Knöpfel at [email protected] or Changsong Zhou decades, spontaneous activity has received increasing attention
at [email protected]. and has been described in various species (Xiao et al., 2017;
https://doi.org/10.1523/JNEUROSCI.0623-20.2021 Smith et al., 2018). Spontaneous activities can influence evoked
Copyright © 2021 Liang et al.
This is an open-access article distributed under the terms of the Creative Commons Attribution 4.0
responses (Azouz and Gray, 1999; Davis et al., 2020); contribute
International license, which permits unrestricted use, distribution and reproduction in any medium provided to observable variability in cognition and behavior (Romano et
that the original work is properly attributed. al., 2015); and play important roles in the consolidation of new
3666 • J. Neurosci., April 21, 2021 • 41(16):3665–3678 Liang et al. · Spatiotemporal Properties of Cortex-Wide Intrinsic Activities

memories, motor learning (Marshall et al., 2006), and sleep- communications and reflect details of the underling neuronal
dependent enhancement in visuomotor performance circuit mechanism.
(Landsness et al., 2009). Studying the essential organization
properties of spontaneous activity is thus vital for under- Materials and Methods
standing brain functions (Deco et al., 2014; Capone et al., Experimental design and statistical analyses. We used mesoscopic
2019; Chen and Gong, 2019). transcranial voltage imaging datasets for the analysis. Data acquisition
The traditional way of characterizing spontaneous activity is was as described previously (Akemann et al., 2010, 2012; Scott et al.,
based on correlation analysis. For instance, based on cross-corre- 2014; Song et al., 2018). Briefly, CaMK2A-tTA;tetO-chiVSFP transgenic
lations between a predefined seed region and other brain regions, animals expressed the GEVI chimeric VSFP Butterfly (Mishina et al.,
whole-brain spontaneous activity exhibits spatially structured, 2014; Song et al., 2018) in pyramidal neurons across all cortical layers.
patchy correlation patterns (Fox et al., 2005). However, increas- The epifluorescence imaging approach we used here restricts optical
ing evidence has shown that spontaneous brain activity exhibits access and signal detection to the superficial cortical layers (layer 2/3).
Under surgical anesthesia, animals were implanted with a transcranial
far richer spatial and temporal dynamics. Particularly, spontane-
cortical window through a thinned but otherwise fully intact skull and
ous activity is organized as propagating waves, which have been
a head-fixation plate. Image acquisition was performed using a dual-
found at different neural levels by using different recording tech- emission wide-field epifluorescence macroscope equipped with two
niques, including multielectrode arrays, voltage-sensitive dye synchronized CMOS (complementary metal oxide semiconductor)
(VSD) imaging, electroencephalography (EEG), electrocorticog- cameras, using high-power halogen lamps for fluorescence excitation
raphy, magnetoencephalography, and functional magnetic reso- (Moritex/BrainVision) and the following optics (Semrock): mCitrine
nance imaging (Freeman and Barrie, 2000; Garaschuk et al., (donor) excitation 500/24; mCitrine emission FF01-542/27; mKate2
2000; Lee et al., 2005; Rubino et al., 2006; Wu et al., 2008; Huang emission BLP01-594R-25; excitation beam splitter 515LP; and detection
et al., 2010; Patten et al., 2012; Alexander et al., 2013; Zanos et beam splitter 580LP. The ratio of changes of gain-equalized (Akemann et
al., 2015). Several studies have described the origins, pathways, and al., 2012) fluorescence intensities acquired with the two cameras reflects
recruitment of various cortical areas during the propagation of the spatiotemporal dynamics of spontaneous membrane voltage fluctua-
waves of activities (Nir et al., 2011; Shimaoka et al., 2017). However, tions of populations of pyramidal neurons.
recording techniques such as intracellular recordings and extracellu- Datasets were acquired at a 150 Hz acquisition frame rate and over a
cortex-wide two-hemisphere field of view. The original spatial resolution
lar multiple-unit recordings have limited spatial coverage, and EEG
of the signal we extracted from the camera is 29
29 mm. During the
has limited spatial resolution, hence the recruitment of neuronal imaging session (several trials of 180 s duration), mice were under anes-
populations during wave patterns has been evaluated mainly on the thesia/sedation (induced by a bolus injection of 30 mg/kg pentobarbital
basis of the timing of local oscillation peaks. Thus, detailed spatio- sodium). With this protocol, animals are lightly anesthetized initially,
temporal organization properties of spontaneous activity and the and then recover over a state of sedation to wakefulness. This progres-
underlying mechanisms remain unexplored. sion was monitored using the heart rate as a proxy. A brain state charac-
Optical voltage imaging data using genetically encoded volt- terized by an absolute lack of spontaneous limb and whisker movements
age indicators (GEVIs) offers a high-spatiotemporal resolution is referred to as “anesthesia.” We analyzed 13 experimental trials over
readout of population membrane voltage in superficial layers five mice (mice 1–4 are males and mouse 5 is female), and each mouse
across a large portion of the cerebral cortex of living mice provided several trials under the anesthesia condition (mouse 1, three
(Knöpfel, 2012). This approach hence provides an opportunity trials; mouse 2, two trials; mouse 3, three trials; mouse 4, three trials;
mouse 5, two trials. Different mice had different number of trials satisfy-
to investigate the large-scale neural dynamics at sufficient spatio-
ing the condition of anesthesia). Each trial recorded spontaneous voltage
temporal resolution and coverage. Recent studies using a similar
activity for 180 s continuously.
VSD imaging method typically focus on specific sensory-evoked MATLAB is used for data processing and analysis. Data and codes
and corresponding spontaneous activity motifs (Mohajerani et used to create all the plots are available on request.
al., 2013) or the existence of wave propagation (Muller et al., Data preprocessing. We extracted voltage signal from raw fluores-
2014). However, details of how waves are spatiotemporally cence signals as previously described (Akemann et al., 2012; Shimaoka et
organized are still missing. Using GEVI imaging of spontaneous al., 2017; Song et al., 2018). Reliable computation of the phase velocity
voltage activity from pyramidal neurons in mice, here we fields (PVFs) was achieved after reduction of spatial noise by 2 times
addressed the following questions. (1) Are there any regularities coarse graining using bicubic interpolation (weighted average of pixels
in the seemingly random occurrence of complex spatiotemporal in the nearest 4-by-4 neighborhood; function imresize, scale = 0.5,
MATLAB, MathWorks). The final spatial data size for the imaged field
wave patterns at different spatial scales? (2) Can different types
of view was a 44
52 matrix. The analysis method based on the phase
of waves coexist, and how do these coincident waves interact? (3)
wave requires that signals across multiple neighboring pixels occur with
How does the formation of cortex-wide wave patterns relate to similar spectral properties. Thus, we focused on narrowband delta oscil-
the underlying brain structure? lation 0.5–4 Hz by bandpass temporal filtering (Chebyshev type II, func-
We address these questions on delta waves (0.5–4 Hz) with tion filtfilt, MATLAB, MathWorks) to improve the signal-to-noise ratio
an advanced wave analysis method, which extracts both the according to the signal power of the wavelet transform. Periods with
spatial and temporal features of continuous wave patterns large temporal fluctuations (amplitude, .3 SDs) of the filtered voltage
(Townsend et al., 2015). Although propagating patterns show a signals were identified as movement artifacts and were excluded from
large variation between individual events, principal component further analysis. The phase velocity fields were calculated on the spatially
analysis revealed that the top five principal modes already con- and temporally filtered data.
tribute to .80% of the variance. We found that the emergence of PVF. We characterized the cortex-wide spatiotemporal patterns
using phase velocity field analysis, which was adapted from physical the-
sources and sinks can reverse the propagating direction of large-
ories of turbulence and validated for data recorded with different techni-
scale traveling waves, and their interactions can generate saddle ques and modalities (Townsend et al., 2015; Townsend and Gong, 2018).
wave patterns. Moreover, sources, sinks, and saddles are prefer- Briefly, the method assumes that contours (isolines) of the phase of brain
entially formed at specific locations, intricately related to the waves propagate monotonically spatiotemporally. First, we used Hilbert
cortical connectivity. These findings underpin the idea that spon- transform on the 0.5-4Hz bandpass-filtered voltage signal to extract in-
taneous activity plays an important role in intracerebral stantaneous phase on each pixel f ðx; y; tÞ. Phase velocity fields
Liang et al. · Spatiotemporal Properties of Cortex-Wide Intrinsic Activities J. Neurosci., April 21, 2021 • 41(16):3665–3678 • 3667

v w ðx; y; tÞ ¼ ðuðx; y; tÞ; vðx; y; tÞÞ can be calculated by adapting the opti- the corners of the four pixels around the critical point were then used to
cal flow estimation method, which is based on the changes of phases further classify the pattern types into source, sink, or saddle, as follows:
between two consecutive frames, with phase constancy and spatial
smoothness constraints. Both constraints are expressed together as a sin- 0 1
@u @u
gle minimization problem, which is described as the Euler–Lagrange B C
B @x @y C
J¼B :
@v C
equations (Bruhn et al., 2005). The toolbox NeuroPatt implementing the (3)
above method can be found on the website https://github.com/rorygt/ @ @v A
NeuroPattToolbox. @x @y
To further validate our PVF method, we shuffled data by randomiz-
ing raw voltage signal temporally (keeping power spectrum on each pixel Based on the trace (t ) and determinant (D) of the Jacobian matrix,
but randomizing the phase of the Fourier components) and spatially source (unstable point, t . 0), sink (stable point; t , 0) and saddle
(randomly shuffling the pixels). Then we applied the same data-process- (D , 0) patterns were determined. If both eigenvalues are real and of the
ing procedure to the shuffled data as for the real data. same sign, the hyperbolic equilibrium is a “node.” When eigenvalues are
Gradients of PVF. Spatiotemporal patterns of neural activities complex-conjugate, the hyperbolic equilibrium is a “focus.” Different from
should be distinguishable from patterns obtained with random shuffled previous definitions (Townsend et al., 2015), we counted each unstable node
data in terms of continuous spatial
gradients in the PVF.  We measured or focus as a source event, and stable node or focus as a sink event. We used
@v w ðx; y; tÞ @v w ðx; y; tÞ the singularity point of PVF with near-zero velocity to define the locations of
the gradients rv w ðx; y; tÞ ¼ ; at each pixel
@x @y a source or sink. The patterns were detected in the analysis window by the
(Chebyshev type II, function gradient, MATLAB, MathWorks). The par- NeuroPatt toolbox (https://github.com/rorygt/NeuroPattToolbox).
tial differentiation is obtained as the difference of the x-projection (or To validate our method to detect sources, sinks, and saddles (local
y-projection) of the velocity vectors at two neighboring pixels. Small gra- waves), we compared the probability of local waves detected in real data
dients indicate high coherence in the speed and direction, as expected and shuffled data as a function of a detection threshold. This detection
for an organized (nonrandom) activity pattern. At each time t, gradient threshold was defined as a pair of values (d, r), where d is the duration
of PVF was defined as spatial average of pixel-wise gradients, as follows: (number of time steps) of the lifetime of the same local wave pattern and r
is the minimum radius (number of pixels from the singularity) of the local
1 X  wave pattern. Probability at detection threshold (d, r) is defined as the
jrv w jðtÞ ¼ rvw ðx; y; tÞj; (1) number of local wave patterns with detection threshold (d, r), divided by
N x;y
the number of local wave patterns with detection threshold (1, 1).
Singular value decomposition of wave patterns. We applied singular
where N is the number of pixels in the analysis window. value decomposition (SVD) to all phase velocity fields v w ðx; y; tÞ ¼
Identification of wave patterns. We followed the method of identify- ðuðx; y; tÞ; vðx; y; tÞÞ to identify the principal components of the wave pat-
ing wave patterns as previously described (Townsend et al., 2015) with terns (Townsend and Gong, 2018). We first reorganized all phase velocity
some modifications to detect different types of wave patterns, including fields into a standard form and combined them into matrix w: At every
plane wave, and source and sink patterns. We detected plane waves by time step, we had a 2D matrix A of PVF vectors [44
52]. In step 1, we
using the order parameter of the PVF in the analysis window, as follows: extracted the real part of the matrix A and reshaped them into a 1D matrix
B [1
2288]. We also extracted the imaginary part of the matrix A and

j j
1 X reshaped them into a 1D matrix C [1
2288]. In step 2, we concatenated C
v w ðt Þ ¼ vw ðx; y; tÞ ; (2) to the end of B to get matrix D [1
4576]. In step 3, we repeated step 1 and
Nv0 ðtÞ x;y
step 2 on every time step and concatenated all the matrices D for the time
length L to organize matrix w [L
4576]. L equals the time steps of one 3
where N is the number of pixels in the analysis window, v0 is the average min data trial. The singular value decomposition can be defined as follows:
magnitude of the velocity over all pixels, and vw is the phase velocity
v w ðx; y; tÞ. The order parameter v w ranges from 0 to 1, with 1 representing X
the case where the velocity vectors are parallel. We set a threshold v w ðtÞ  w¼T Rp ; (4)
0.85 to identify plane waves in the examined areas. Variation of the thresh-
old value between 0.8 and 0.9 did not substantially change the results. where T and R are unitary matrices, p denotes the conjugate transpose,
Relaxing the threshold to smaller values allowed us to capture largely coher- and + is a rectangular diagonal matrix of singular valuess . The kth spatial
ent propagating waves within the analysis window that can coexist with
regions of heterogeneous propagation directions corresponding to complex Xcolumn of R, has a propor-
mode, defined by the velocity field in the kth
tion of the overall variance given by s 2k = s 2i . The top three spatial
local wave patterns such as sources, sinks, and saddles (see below). i
We defined standing waves (synchrony) as periods where there is no modes with the largest variance accounted for 70.4% of the total variance
apparent wave velocity (i.e., propagation) across the analysis window. in the data of one representative mouse shown in Results section. Across
The criterion for standing waves was an average magnitude of the veloc- the data from the five mice analyzed the three spatial modes with the larg-
ity fields 2 SDs below the mean value across the analyzed time period. est variance accounted for 44.5 6 17.9% of the total variance.
We first asked whether the PVF across the analysis window is sta- Then we projected the instantaneous representative phase velocity
tionary (standing wave), and only if not, applied the criteria to detect fields on the principal modes, as follows:
plane and complex waves. Plane waves are detected when the wave prop-
agation directions are narrowly distributed within the analysis window.
w
We examined how the emergence of plane waves depends on the size of M¼ ; (5)
the analysis window (see Fig. 5). In this context, we note that in a very R
small analysis window most nonstationary wave pattern would be
detected as a plane wave. Our smallest analysis window has 16 regions of where M is the weight matrix of every principal mode contributed to all
interest, and the largest probability of plane wave is 81% among five phase velocity fields. From this, we can obtain the projection variance of
mice; hence, even with our smallest window plane waves are not the mth spatial mode on thenth phase velocity fields, as follows:
detected in all PVFs. We also like to note that if waves were all (domi- X
nantly) large-scale waves, their detection probability would not increase M2m;n = 2
Mi;n :
(or increase a little) with decreasing window size. i
Local complex wave patterns are organized around the critical points,
which were identified by the intersections of two bilinearly interpolated Probability of local wave patterns. To investigate the spatial distribu-
null clines of the phase velocity field. Eigenvalues of the Jacobian matrix at tion of the emergence of local wave pattern types (source, sink, and
3668 • J. Neurosci., April 21, 2021 • 41(16):3665–3678 Liang et al. · Spatiotemporal Properties of Cortex-Wide Intrinsic Activities

Figure 1. Spontaneous voltage activity as propagating waves characterized by the phase velocity field. A, Power spectrum of spatially averaged voltage activity amplitude. B, Time–fre-
quency spectrum of voltage signals taken from the position marked by the blue spot in C. Warmer color represents higher power. C, Imaged cortical area for mouse 1. Background is the voltage
map at the beginning of the time period shown in D. D, Example of GEVI-based voltage imaging data along a rostral–caudal line across the left hemisphere as indicated in C; the colors repre-
sent the voltage amplitude (color bar is the same as in F). E, Voltage amplitude oscillation measured at the three spots; colors are as indicated in C. The black line is the spatial averaged volt-
age amplitude. F, Example of snapshots of phase velocity fields. The arrows are oriented in the wave-propagating direction, and their length indicates propagating speed. For clarity of
visualization, only each second calculated vector is shown. The background colors represent the voltage amplitudes. G, PVF of spatiotemporally shuffled data. Background colors are the same
as in F. H, Probability distribution of gradients of PVFs obtained from real data and shuffled data. I, Probability distribution of wave directions.

saddle), we calculated the center location of every source, sink, and sad- as a measure of the experimental probability of pattern emergence of
dle pattern and the wave duration (the number of time steps that the that region.
same location is occupied by the same singularity). For each pixel and Functional registration of data into Allen Mouse Brain Atlas. To
wave pattern type, we calculated the probability of pattern emergence as study the relationship between wave patterns and the anatomic proper-
the cumulative time the pattern existed. After obtaining the probability ties of the underlying circuitries (see Fig. 9), we used the Allen Institute
of pattern emergence, we averaged across pixels for each cortical region Mouse Brain Atlas (www.brain-map.org) projected to our plane of
Liang et al. · Spatiotemporal Properties of Cortex-Wide Intrinsic Activities J. Neurosci., April 21, 2021 • 41(16):3665–3678 • 3669

imaging. The two cortical hemispheres were registered into the atlas by
aligning the visual and whisker stimuli response centers in voltage imag-
ing data (Song et al., 2018) to the center of primary visual area and pri-
mary somatosensory area-barrel field of the atlas correspondingly
(function imregtform, imwarp, MATLAB, MathWorks).
Correlation of complex wave patterns to the mouse cortical architec-
ture. Next, we examined how the occurrence probability of the complex
wave patterns relates to the underlying cortical architecture. We first
calculated the singularity probability for each wave pattern type and for
each pixel. Then, we averaged across cortical region, pooling across the
two hemispheres to generate a regional probability map. To relate these
regional probabilities to the underlying cortical architecture, we used
the hierarchical gradients of the mouse cortex (Fulcher et al., 2019) as a
structural estimate. We calculated the Pearson’s linear correlation coef-
ficient between the average (N = 13 trials; five mice) probability of each
wave type against the hierarchical gradient index of the corresponding
cortical region (a total 21 cortical regions common to both data types).
To investigate whether this correlation was by chance, we randomly
shuffled the locations of singularity probabilities of each pixel and of
each pattern type, averaged across cortical regions and pooled across the
two hemispheres to generate randomly shuffled regional probability
maps. For each randomly shuffled probability map, we calculated the Movie 1. Complex wave dynamics during spontaneous activity. [View online]
Pearson’s linear correlation coefficient with the hierarchical gradient
index. The correlation coefficient calculated with the actual data lies shuffled data do not exhibit a preference at all. In a previous analy-
clearly outside the distribution of the correlation coefficients obtained sis based on time delays in the positive peaks of LFP (Vyazovskiy
from 1000 shuffled datasets (see Fig. 9E). This analysis strongly suggests et al., 2009), traveling waves appeared most frequently in anterior-
that the correlation between localizations of small-scale activity waves to-posterior and posterior-to-anterior directions. Measurement by
and cortical architecture, as reflected in the hierarchical gradient index,
time delays in peaks is consistent with our analysis of phase veloc-
is not by chance.
ity fields, but it only captures the dynamics at the peak of the oscil-
lations. Phase velocity fields extend the characterization of wave
Results propagation temporally to the whole time course of the activity
Complex spatiotemporal wave patterns revealed by phase wave and spatially to every imaged pixel and reveals far more
velocity field analysis complex propagation dynamics of waves during spontaneous ac-
We recorded voltage signals across the dorsal cerebral cortex tivity (Movie 1). As the propagation path is consistent with the
using GEVI imaging in anesthetized mice. Consistent with previ- spatially organized sequential activation local circuits, wave propa-
ous studies, voltage power spectra indicate the presence of slow gations denote the information flow that could be validated by
waves with a frequency of ;2 Hz (Fig. 1A,B). For subsequent Granger causality (Granger, 1969; Seth, 2010). By calculating the
analysis, we bandpass filtered the signal from 0.5 to 4 Hz to pairwise Granger causalities, we found that neurons in the earlier
increase the accuracy in the calculation of the phase velocity field part of a wave would have a causal role on neurons in its later part
(Fig. 1F). The spatiotemporal patterns of these spontaneous (Fig. 2). However, because these spontaneous wave patterns are
activities are well resolved when plotting the voltage signals from ever changing, they may change directions in the next few time
pixels residing along a rostral–caudal line in the imaged cortical steps. This dynamical property makes the detection of a relatively
space (Fig. 1C,D). These maps illustrate the occurrence of posi- stable causality between different areas unrealistic.
tion-dependent time shifts, suggesting the existence of propagat-
ing waves. These plots alone may also indicate the presence of Features of spontaneous large-scale activity waves
standing (nontraveling) waves (Volgushev et al., 2006), as sug- Although the phase velocity fields displayed very rich and vari-
gested by the voltage signals measured at three positions on the able wave propagation dynamics, we wondered whether they ex-
line across the cortex (Fig. 1E). However, phase velocity field anal- hibit features that are nonrandom, and can be consistently
ysis reveals continuous and complex wave propagation (Fig. 1F), observed across many waves detected across animals. To address
even in cases where the voltage signal (Fig. 1D) appears to be this question, we performed SVD on the PVFs (see Materials and
largely synchronized across distant locations. To validate our Methods). The most frequent (principal) SVD modes reveal the
method of PVFs, we performed spatiotemporal shuffling (see typical wave propagation pathways, while the large number of
Materials and Methods) on the raw voltage signals. Application of small modes reflects the rich dynamics. The top five most domi-
the identical data-processing pipeline resulted in unstructured nant modes accounted for ;80% of the total variance (Fig. 3A),
PVFs (Fig. 1G). As the PVFs of shuffled data display less coher- and variance distribution of the modes decreased sharply (Fig.
ence in wave directions and speeds, gradients of PVFs (see 3B). Each SVD mode represents patterns with a dominant direc-
Materials and Methods) were used to quantify the distinction of tion (most frequently observed; illustrated in Fig. 3A) and the op-
PVFs between real data and shuffled data (Fig. 1H). Smaller gra- posite direction (reversed for all pixels). The variance for each
dients indicate a larger coherence of local wave direction and mode (both directions included) along with the relative propor-
speed. The clear separation of the distributions of gradients of tion for each direction is indicated in Figure 3C. The first SVD
PVFs from original and shuffled data demonstrates that the PVFs mode accounts for nearly 50% of the variance and represents
of unshuffled data do not emerge as a result of data processing but large-scale plane waves that are symmetrical across hemispheres.
represent features of organized cortical activity. In the unshuffled The less frequent mode 2 represents waves that, for their domi-
data, the statistics of wave directions (Fig. 1I) showed that most of nant direction, spread out from the somatosensory areas toward
the waves propagate from rostral cortex to caudal cortex, while rostral and caudal cortices. The illustrated mode 3 waves spread
3670 • J. Neurosci., April 21, 2021 • 41(16):3665–3678 Liang et al. · Spatiotemporal Properties of Cortex-Wide Intrinsic Activities

Figure 2. Granger causality of a traveling wave. A, The phase velocity fields on a traveling wave propagate from posterior to anterior. The background color denotes the voltage amplitudes.
B, Granger causalities between a seed (red dot) and other pixels during period in A. The background color denotes the value of F-Critical values in F statistic, for the 0.05 significance level,
with F  0 referring to the absence of Granger causality, and F . 0 denoting the presence of Granger causality. Larger values can be considered as stronger causality. We set the maximum
lag of the Granger analysis as five time steps.

Figure 3. Principal modes of the phase velocity fields. A, Top five modes calculated by singular value decomposition of all phase velocity fields from mouse 1 (each vector for each second
pixel shown). B, Variance distribution of the top 20 modes. C, Proportion of the dominant and opposite directions of the top five modes. D, Similarity of the top 20 aligned modes between dif-
ferent mice/trials (N = 13 trials, five mice).

from one hemisphere to the other hemisphere, while mode 4 values of the Pearson correlation of reranked modes decreased
waves have opposite directions of traveling waves in the two with mode index (Fig. 3D). The high correlation coefficients of the
hemispheres. The illustrated mode 5 represents a sink (or source top three modes indicate consistent features across animals and
for opposite direction) in the somatosensory cortex on both trials, and the sharp decrease in correlations across modes indicat-
hemispheres. Across different mice, modes 1 and 2 are consis- ing increasing individual variability of the high-order modes.
tently found to be the most frequent, but the frequency rank of
the less frequent modes varies across animals. We used the Classification and detection of wave patterns at smaller
modes shown in Figure 3A (mouse 1, trial 2) as a reference and scales
aligned the top 20 modes in the other mice/trials according to Using the SVD method, we found consistent propagation path-
their similarity to this reference, and the average of the absolute ways for the cortex-wide activity waves. However, except for the
Liang et al. · Spatiotemporal Properties of Cortex-Wide Intrinsic Activities J. Neurosci., April 21, 2021 • 41(16):3665–3678 • 3671

Figure 4. Classification and detection of specific wave patterns at the cortex-wide scale. A, Examples of wave patterns: plane wave, standing wave, source, sink, and saddle. Vectors are
shown for each pixel to facilitate the identification of individual patterns. B, Probability of local wave pattern detection as a function of detection threshold [defined as a function of duration
(d) and radius (r) threshold values]. Colors represent the probability that a local wave pattern is detected. Red line in shuffled data is the 95% confidence line for unshuffled data to be because
of spatiotemporally organized activity. Red dot indicates the local wave detection threshold (5, 3) used in the following analysis. C, Time course of the wave patterns from an example trial.
Note that, when defined at the large scale, plane and standing waves occupy most of the imaged cortical space but are rare, whereas sources, sinks, and saddles may be locally restricted, but
frequently occur. Color bar shows the number of waves of a certain type existing at time of a snapshot.

top three SVD modes, the other modes are quite complex and the probability in the real data decreased less steeply. A detection
capture less frequent events. A closer inspection of these SVD threshold of (5, 3) (Fig. 4B, red line) indicates 96.5% confidence
modes suggested that individual cortex-wide spatiotemporal that detected waves are not because of randomness that is also
wave patterns likely contain many smaller-scale features. To dis- contained in the shuffled data. This detection threshold is used
tinguish specific smaller-scale wave patterns, we defined the fol- for the analysis presented in Figure 4 (and see Figs. 6, 7, 8, 9).
lowing five wave types (Fig. 4A; see Materials and Methods): Both wave pattern snapshots (Fig. 4A) and analysis of pattern ac-
plane wave (a traveling wave mainly in one direction); standing tivity occurrence in time (Fig. 4C), reveal that source, sink, and
wave (a nonpropagating wave, effectively synchrony); source; saddle wave patterns frequently coexist.
sink; and saddle. If we did not detect any of these wave patterns
in a given PVF frame, we counted such patterns as unclassified. Wave propagation across spatial scales
Plane (traveling) and standing waves that occupy the whole The high spatiotemporal resolution of our recordings allows us
imaged region occurred much less frequently than (smaller- to test a key question, namely whether the plane waves and
scale) source, sink, and saddle waves (Fig. 4C). According to standing waves occur only at the scale of our whole imaging win-
our classification criteria, waves analyzed across the whole dow, or also more locally. To address this question, we divided
imaged cortical area cannot be identified as plane wave and the imaged field of view into smaller spatial fields with 2, 4, or 16
standing wave at the same time. However, plane waves and regions of interest (ROIs; Fig. 5) and observed that the probabil-
standing waves often occurred within smaller areas of the cortex ity of detecting plane waves and standing waves sharply
and, when considering those smaller areas, can coexist with other decreased with increasing size of the ROI. This indicates that
wave patterns (sink, source, and saddle). For example, the large-scale waves represent only a small portion of detectable
unclassified pattern at the cortex-wide scale in Figure 4A con- waves. Compared with plane waves, standing waves (including,
tains a mixture of plane wave (rostral cortex) and standing wave according to our detection algorithm, also time points with no
(caudal cortex) in the left hemisphere. Thus, the classification of up-state-like activity in the detection window) are more sensitive
waves as plane or standing depended on the size of the cortical to the detection window size (Fig. 5B), indicating that they very
area analyzed. Because the source, sink, and saddle typically rarely occur globally. All these results are robust over the five
occupied small areas, we regard them as local wave patterns. mice analyzed. This observation underlines the notion that com-
Probability comparison of local wave patterns detected in real plex local waves contain rich information that may be missed if
data and shuffled data (see Materials and Methods) indicates the recording resolutions are not high enough.
that large/long-lived local wave patterns are distinguishable from
random nonorganized events. As shown in Figure 4B, with the Direction reversal of large-scale propagating waves by local
increase of the detection threshold (defined by radius and dura- wave patterns
tion of the detected event), the detection probability of localized Previous studies and our present data show that large-scale trav-
waves decreased sharply when analyzing the shuffled data, while eling waves preferentially propagate in an anteroposterior
3672 • J. Neurosci., April 21, 2021 • 41(16):3665–3678 Liang et al. · Spatiotemporal Properties of Cortex-Wide Intrinsic Activities

Figure 5. A, B, Probabilities of detecting plane wave (A) and standing wave (B) decay with increasing ROI window size. The cortical image was divided into 16 parts (light blue), 4 parts (green),
2 parts (left and right, purple; rostral and caudal, orange and yellow), and 1 part (dark blue). The color of filled dots corresponds to coloration of the segments. The x-axis denotes the size of the
cortical segments normalized to the total imaged area. Shown are the results of the three experimental trials from mouse 1. Similar results were obtained from the data of the other mice.

direction (Fig. 6A), but can occasionally reverse direction source emerges in the somatosensory cortex (Fig. 6H, blue
(Sanchez-Vives and McCormick, 2000; Massimini et al., 2004). square), accompanied by order reversing of the instantaneous
However, the mechanism underlying this direction reversal is phases between the ROI in the somatosensory cortex and ROI in
not known. We hypothesize that local complex waves may affect the visual cortex. It appears that enhanced excitation in the
the direction of propagation of large-scale waves. As shown in somatosensory cortex accelerates the slope of the local depolari-
Figure 5A, global plane waves rarely happened, let alone in the zation (blue curve) and the emergence of a source in the PVF.
anteroposterior direction. Thus, we identified relatively large The observed activity pattern also suggests that the prolongation
(but not necessarily whole field of view covering) traveling waves of the activity in the visual cortex (Fig. 6H, middle, orange curve)
by setting an appropriate order parameter threshold (for details, is caused by excitatory inputs originating in the somatosensory
see Materials and Methods). By lowering the order parameter to source. These effects lead to phase order reversal and reversal of
0.5, we identified still relatively coherent, large-scale traveling the propagation direction in visual cortex. As shown in Figure
waves, but allowing their temporal coexistence with other waves. 6I, a large-scale wave initially propagated in the posteroanterior
As a measure of the instantaneous wave propagation direction, direction, but the activity wave in the somatosensory cortex (blue
we calculated the spatially averaged propagation direction u square) slowed down relative to the frontal cortical regions
(Fig. 6B,C). This allowed us to identify waves propagating in the (green diamond), inducing phase order reversal and the changed
anteroposterior (5p =4,u , 7p =4Þor posteroanterior (p =4 , wave propagating direction in the primary cortex from poster-
u , 3p =4Þ directions. The direction measure u together with oanterior to anteroposterior.
the order parameterv w .0:5 identify relatively coherent travel- Having observed many cases like these two examples, we
ing waves with either anteroposterior or posteroanterior direc- hypothesized that a sufficiently strong source and sink can change
tions. We classified other PVF patterns as disordered (Fig. 6E), the direction of nearby ongoing traveling waves. As schematically
but we shall note that disorder at the larger scale does not depicted in Figure 7A, we more specifically hypothesized that
exclude order at a smaller scale. The time courses of u and the waves of activity that propagate out from a source or toward a
order parameter v w (see Materials and Methods) showed inter- sink in the direction of a traveling wave will not affect the latter;
related fluctuations (Fig. 6C,D), as expected from a preferred otherwise, the direction of the traveling wave will be reversed
propagation direction of coherent waves. Disorder at the large through interaction with local sources and sinks (Fig. 6H,I).
scale was found much more frequently than coherent PVFs, To test this hypothetical scheme by statistical analysis, we
while in the coherent PVFs, the anteroposterior direction is considered the wave directions of two successive large-scale trav-
more frequent than the reverse direction (Fig. 6F). These coher- eling waves interlaced by a large-scale disordered state. The
ent waves in the anteroposterior or posteroanterior directions directions of the large-scale wave can either remain the same
can coexist with local complex waves, while large-scale disor- [anteroposterior followed by anteroposterior (a-a) or posteroan-
dered PVF states can involve many local waves (Fig. 6G). terior followed by posteroanterior (p-p)] or can change [antero-
Next, we analyzed the formation of sources and sinks in the posterior followed by posteroanterior (a-p) or posteroanterior
presence of coherent propagating waves. A source is defined as followed by anteroposterior (p-a)]. We counted the occurrence
the origin of a wave with an oscillation phase that leads relative of sources and sinks that occurred during the last quarter interval
to the neighborhood. An example is shown in Figure 6H. of the previous disordered state or during the first quarter inter-
Initially, oscillation phase over the visual cortex (Fig. 6H, orange val of the following large-scale wave for each of the conditions a-
triangle) is ahead, and the wave propagates from visual cortex to a, a-p, p-a, and p-p. Based on our hypothesis that the direction of
prefrontal cortex. At a later time, the visual cortex falls behind in the large-scale traveling wave depends on the nature and local-
phase along with a long period of increased activation, and a ization of preceding local waves, we combined a-a/p-a as one
Liang et al. · Spatiotemporal Properties of Cortex-Wide Intrinsic Activities J. Neurosci., April 21, 2021 • 41(16):3665–3678 • 3673

Figure 6. Reverse of propagation direction of traveling waves through interaction with local activity sources and sinks. A, Anterior–posterior flow starts with anterior sources and ends with
posterior sinks. B, Schematic depiction of relative concordant directions (left) and disordered directions (right). Black dots on the circle represent local wave directions, the length of the blue
vector is the order parameter and the angle of the vector is the average directionu . C, D, The temporal evolution of u andvw : We set p =4 , u , 3p =4 as posteroanterior direction
(purple) and 5p =4 , u , 7p =4 (green) as the anteroposterior direction, and considered times of v w , 0.5 as (large-scale) disordered. E, The temporal evolution of wave propagation
direction, according to the thresholding schemes in C and D. F, Probabilities of the anteroposterior, disorder, and posteroanterior states (N = 13 trials from five mice; mean 6 SD).
G, Probabilities of occurrence of local complex waves during each large-scale state. The color code indicates the number of coexisting complex waves (total across sources, sinks, and saddles).
H, I, Examples of the interaction of large-scale traveling waves and a local source (H) and sink (I). Top, Snapshots of the phase velocity fields (vectors for each second pixel are shown), at times
indicated by vertical dashed lines in the middle and bottom panels. Color codes represent the instantaneous voltage amplitude. Middle, Instantaneous voltage amplitude (average of 3
3 pix-
els) recorded at the three ROIs are indicated on the inset. Bottom, Instantaneous phases corresponding to the 3
3 pixel-averaged signals for the three indicated ROIs. The reversal of phase
orders is indicated by black circles.

scenario and p-p/a-p as another scenario. Because it is unlikely sources and sinks in the anterior area (Fig. 7C, red box) and pos-
that sources or sinks localized in the center of the cortex would terior area (Fig. 7C, black box) of the cortex. As shown in Figure
influence the direction of the large-scale waves emerging in the 7C, we plotted the average spatial distribution of sources and
posterior or anterior portion of the cortex, we only counted the sinks for each of the scenarios (five mice, 13 experimental trials).
3674 • J. Neurosci., April 21, 2021 • 41(16):3665–3678 Liang et al. · Spatiotemporal Properties of Cortex-Wide Intrinsic Activities

Figure 7. Interaction of plane wave and local waves induces transition of the wave-propagating directions. A, Schematic illustration showing that the location and type of local waves can
strongly determine the final wave propagation direction (Fig. 6A, example for the a-a condition). The four boxes represent different situations. The colored background indicates the initial plane
wave-propagating direction (from red to blue), which is consistent with the arrow direction on the left. The arrow on the right side shows the final direction of the wave after the disordered
period. The top (bottom) panel shows that the final direction tends to be anteroposterior (posteroanterior) regardless of the initial direction, if there is a rostral source (sink) and/or a caudal
sink (source). B, Probability of local wave patterns (sources and sinks in rostral and caudal cortex) during the large-scale disorder states on different conditions, grouped to compare the antero-
posterior direction (a-a/p-a) or posteroanterior direction (p-p/a-p) in the final coherent waves following disordered interval. The legend “sourceR (sourceC)” denotes a source located in rostral
(caudal) cortex, and analogous for sink. The bars are the mean probability across five mice, and the error bar is the SD. The p values were calculated by Wilcoxon test: *p , 0.05,
**p , 0.005. C, Spatial distribution of the number of sources or sinks for different combinations of preceding and following large-scale wave directions (pooled data from five mice, 13 trails).
The color shows the average number of the sources/sinks that occurred in the corresponding condition.

For statistical analysis, we first normalized the small wave counts coexisted, the most frequent combination was source–saddle
for the a-a/p-a scenario and the p-p/a-p scenario by the total num- pairs. Coincidence of three or more local wave patterns was also
ber of anteroposterior waves and posteroanterior waves (Fig. 6F). observed. To gain further insights into the spatial characteristics
We also normalized the number of sources and sinks because of the interaction of local waves, we analyzed the representative
sources were more frequently detected than sinks (Fig. 7B). When phase velocity fields for each combination of coinciding local
examining the scenario a-a/p-a, we noted that in the rostral cortex wave patterns (Fig. 8B). This analysis revealed that the saddle
there is a much higher probability of sources than sinks, while in pattern typically emerged from the interaction of two complex
the caudal cortex there is a higher probability of sinks than sour- waves regardless of wave type (source or sink). Moreover, two
ces. Furthermore, rostral sources were found more frequently in sources or two sinks would produce a saddle between them (Fig.
the scenario a-a/p-a than in the scenario p-p/a-p (Fig. 7B). 8B, highlighted by black dot rectangle), while one source and one
Consistent with our hypothesis, these results suggest that complex sink would produce a saddle in a location away from the line seg-
waves during initialization and early progression of the large trav- ment and form a triangle (Fig. 8B, highlighted by the black dot-
eling waves can influence the direction of the latter. ted triangles).
Is there an organization principle that governs the interac-
tions of local wave patterns? According to the Poincaré–Hopf
Interaction of local wave patterns theorem of vector flow fields (Hopf, 1927), if the vectors along
Most PVF snapshots show only a single source, sink, or saddle the boundary do not change in topology, the emergence of local
(Fig. 8A). Among all occasions where two local complex waves wave patterns within the boundary pairwise maintain the total
Liang et al. · Spatiotemporal Properties of Cortex-Wide Intrinsic Activities J. Neurosci., April 21, 2021 • 41(16):3665–3678 • 3675

Figure 8. Interaction of waves. A, Coexistence of local wave patterns. We considered the following seven combinatory situations: only source; only sink; only saddle; source and sink; source
and saddle; sink and saddle; source, sink, and saddle. Shown are the probabilities of occurrence/co-occurrence in field of view with the number of detected waves indicated by the color code
(N = 13 trials from five mice, mean 6 SD). B, Top row, Examples of phase velocity fields for every situation counted in A. C, Phase velocity fields with paired appearance and disappearance of
local wave patterns with Poincaré index (11 for source or sink, 1 for saddle). Background colors denote the voltage amplitude. D, The probability of changes of total Poincaré index between
successive time steps.

Poincaré index (11 for source or sink, 1 for saddle). An exam- different mice (Fig. 9B), conferring some degree of individuality.
ple of the emergence and disappearance of a sink–saddle pair is Despite this, the high similarity (Fig. 9B) of spatial distribution of
shown in Figure 8C. We used the difference of the Poincaré singularity (wave center) across mice suggests that the preferential
index at time t 1 1 and t to monitor the (balanced) emergence of spatial locations of the source, sink, and saddle patterns may be
complex waves. The statistical analysis of data from 13 180 s re- related to the underlying anatomic architecture of the brain. After
cording sessions across five mice confirmed that the total registration of the wave localization maps onto the Allen Mouse
Poincaré index remains mostly unchanged (Fig. 8D). This in Brain Atlas, we generated the averaged the probability values
turn confirms that the appearance or disappearance of sources or among pixels in each brain area (Fig. 9C). When comparing these
sinks is accompanied by the emergence of a saddle. The events wave pattern probability maps against the common hierarchical
with nonzero changes of total Poincaré index are still consistent gradient (as a choice of structural measure; Fulcher et al., 2019),
with the Poincaré–Hopf theorem since the topology of the vec- we observed significant negative Pearson linear correlations
tors along the boundary change when the complex waves move between the regional probabilities of sources (correlation = 0.76,
across the boundary. p = 1.1e-4) and the hierarchical index across the cortical areas,
while sinks (correlation = 0.31, p = 0.19) and saddles (correlation
Relating preferred locations of sources, sinks, and saddles to = 0.41, p = 0.074) have some trends without significance (Fig.
cortical hierarchy 9D). No dependencies were observed in the spatially randomized
Finally, we investigated whether there are preferred locations for singularity distribution (Fig. 9E).
sources, sinks, and saddles, and, if so, whether they relate to
structural features of the mouse brain. From the maps of cumu-
lative localization for each wave pattern type (Fig. 9A; N = 13 tri- Discussion
als from five mice), the probabilities of source, sink, and saddle By taking advantage of high-resolution voltage imaging of the
occurrences all show uneven spatial distribution that are similar mouse cortex, here our work identifies key spatiotemporal orga-
across individual mice (Fig. 9B). The similarity between different nization properties of spontaneous activity at the cortex-wide
datasets from a given mouse is higher than the similarity among level. These properties include the following. (1) Waves of
3676 • J. Neurosci., April 21, 2021 • 41(16):3665–3678 Liang et al. · Spatiotemporal Properties of Cortex-Wide Intrinsic Activities

Figure 9. Spatial distributions of singularities and their relationship with cortical structural hierarchy. A, Pixel-wise distribution of the average occurrence probability for each wave pattern type
(source, sink, and saddle), total complex waves (average), and one example of the shuffled data from the source pattern (randomized). N = 13 trials across five mice. B, Correlation coefficients of
the spatial distribution of local wave patterns. The x-axis and y-axis indicate the experiment index: mouse 1, 1–3; mouse 2, 4–5; mouse 3, 6–8; mouse 4, 9–11; mouse 5, 12–13. Black box shows
the correlation of trails within the same mouse. C, Left, Voltage-imaging data after functional registration into the Allen Mouse Brain Atlas. Middle, Cortical region maps (left hemisphere) of cortex-
wide average probability (pooling across the two hemispheres) of source, sink, and saddle wave patterns. Right, Cortical region maps (left hemisphere) of hierarchy index derived from the study by
Fulcher et al. (2019). Regions with black colors are not covered by the field of view of the voltage-imaging datasets used. D, Scatter plot of the correlation between regional source/sink/saddle prob-
ability and the hierarchy index of the corresponding cortical regions. E, Histogram distributions of the correlation coefficients between the hierarchy index and spatially randomized singularity data.

activity propagate along many different pathways but a few dom- occur either in isolation or in combination, as expected from the
inant principal velocity field modes account for a large propor- Poincaré–Hopf theorem. The direction of waves can be reversed
tion of variance (Figs. 1, 3). (2) Global plane waves and standing as a consequence of their interactions (Figs. 6, 7), and interac-
waves are rare (Fig. 4). (3) Different complex wave patterns can tions of two local waves, no matter whether source or sink,
Liang et al. · Spatiotemporal Properties of Cortex-Wide Intrinsic Activities J. Neurosci., April 21, 2021 • 41(16):3665–3678 • 3677

generate a saddle pattern (Fig. 8). (4) The spatial distributions of both frameworks, spontaneous activities could serve as a cortex-
source, sink, and saddle are similar across individual mice, and wide coordinating mechanism for performing distributed dy-
the incidence of local wave patterns correlates with features of namical computation (Gong and van Leeuwen, 2009).
the underlying cortical hierarchical architecture, as characterized We observed that different types of local waves can exist alone
by the covariation of diverse measurements including gene (Fig. 6) or in combination (Fig. 8) at any location of the imaged
expression, intracortical axonal connectivity, the ratio of T1- cortex. Saddles typically emerge from the interaction of two local
weighted to T2-weighted (T1w:T2w ratio) (myelination con- waves, similar to observations in the middle temporal area of
tents), and interneuron cell density (Fulcher et al., 2019). In par- marmoset cortex (Townsend et al., 2015), and such interactions
ticular, the spatial distribution of sources is highly related to the might endow neural circuits with the computational capacity of
hierarchical gradients derived from the structural variation (Fig. integrating distributed events happening at different space and
9). time. We also indicated that the emergence and disappear-
The SVD shows that the top two principal modes reflecting ance of local wave patterns are in agreement with predictions
two typical wave-propagating pathways are highly conserved of the Poincaré–Hopf theorem. The observed small fluctua-
across experimental trials and animals (Fig. 3). The first domi- tions in the total Poincaré index may be explained by several
nant principal SVD modes represent cortex-wide traveling waves factors including the fact that (1) the imaged portion of the
that propagate preferentially along an anterior–posterior (with a mouse cortex is not perfectly flat, which complicates geomet-
smaller medial–lateral component) axis with hemispheric sym- rical analysis, and that (2) our spatiotemporal resolution and
metry. This result is consistent with intracellular and extracellu- the signal-to-noise ratio may not be high enough to resolve
lar recordings that revealed that spontaneous slow waves under all the local wave patterns.
anesthesia preferentially propagate in the anterior–posterior At present, it is unclear whether global and local waves are
direction and sometimes propagate reversely (Sanchez-Vives and mechanistically distinct. However, their interactions described
McCormick, 2000). The delta band calcium signals under anes- here suggest that they, at least, share or compete for the same cel-
thesia also provided evidence of traveling waves in directions lular mechanisms. Indeed, large-scale waves propagate out from
along the anterior–posterior axis of the mouse cortex (Mitra et al., sources and into sinks; a rostral source or a caudal sink or their
2018). Corresponding large-scale propagation directions were combination favors wave propagation in the anteroposterior
shown using high-density EEG recordings in humans (Murphy et direction (Fig. 7). If the local waves are of low amplitude (i.e.,
al., 2009), in cats (Volgushev et al., 2006), and, in part, in rodents recruiting only a small fraction of local neurons), larger global
(Vyazovskiy et al., 2009). waves may be simply sweeping through. Otherwise, if a local
The second principal mode appears as a source localized complex wave is of larger amplitude (i.e., recruiting a larger frac-
medially halfway along the rostral–caudal axis. The third princi- tion of neurons), the complex wave will expand and eventually
pal mode represents sweeps of activity from one hemisphere to dominate a larger fraction of the cortical space. Meanwhile,
the other. This contrasts to the much more frequent homeotopic global plane waves may also propagate from sources to sinks
patterns and may relate to left–right asymmetry (lateralization). with the sources and sinks outside our optical imaging window.
The wave propagation represents sequential activation of All this evidence supports the idea that local and brain-wide
neural circuits along the propagation pathway; thus, it generally waves share the same cellular and synaptic mechanisms.
denotes the information flow in the brain. In the special case of The organization properties of brain waves is a long-standing
anesthesia, the orderly propagation of neuronal activity along interest in neuroscience (Muller et al., 2018), and recent studies
typical pathways gives rise to sequence-like activity, which gives have begun extending the classical description of brain waves in
rises to causality in the dynamics. Similar dynamical patterns terms of temporal synchrony (Nir et al., 2011) to the spatial and
during slow-wave sleep may interplay with spike timing-depend- temporal domains (Takagaki et al., 2011; Zhang et al., 2018;
ent synaptic plasticity required for memory consolidations or Davis et al., 2020). The advance in our current work is the appli-
synaptic downscaling (Tononi and Cirelli, 2003). At least some cation of the analysis on data of high spatial and temporal resolu-
of these processes are likely to occur during slow-wave sleep. tion as well as of large coverage (the major part of dorsal cortical
Moreover, spontaneous traveling cortical waves in the local corti- hemispheres), revealing a variety of wave organization properties
cal region have been reported to gate perception in behaving pri- in a quantitative way. In this study, we focused on the anesthe-
mates (Davis et al., 2020) and to modulate neural excitability by tized state, and we expect that during the fully awake state, the
locking to spikes (Townsend et al., 2015). It is therefore plausible cortex-wide dynamics could exhibit more localized wave patterns
to consider that cortex-wide waves would modulate large-scale with complex interactions; this prediction can be tested in the
propagation of stimulus-induced neuronal activities, thus coordi- future. In this study, we mainly investigated the wave patterns of
nating activities at different cortical areas. Therefore, an impor- slow oscillations. It has been widely observed that the phases of
tant issue to be addressed in future studies is to which extent low-frequency oscillations are coupled to the amplitudes of high-
these specific wave directions/trajectories determine activities frequency oscillations in the brain (Canolty and Knight, 2010).
observed in awake states. Therefore, it would be interesting to investigate how this cross-
We noted that the spontaneous activities propagate across frequency coupling influences or is influenced by the underlying
cortical space without being affected by the boundary of anatom- wave patterns (Townsend and Gong, 2018). By revealing the or-
ically and functionally defined cortical areas while sensory stim- ganization properties of propagating waves at the cortex-wide
uli-driven activities, at least initially, are (Petersen et al., 2003; Xu level, the present study lays a ground for further exploring these
et al., 2007; Mohajerani et al., 2013; Song et al., 2018). Stimuli- key questions in future studies.
driven activities may be confined to one brain region, for
instance in sensory representations that trigger no behavioral
response. However, the predictive coding theory and simple
References
Akemann W, Mutoh H, Perron A, Rossier J, Knöpfel T (2010) Imaging brain
stimulus response models would require the involvement and electric signals with genetically targeted voltage-sensitive fluorescent pro-
coordination of distributed brain systems (Arieli et al., 1996). In teins. Nat Methods 7:643–649.
3678 • J. Neurosci., April 21, 2021 • 41(16):3665–3678 Liang et al. · Spatiotemporal Properties of Cortex-Wide Intrinsic Activities

Akemann W, Mutoh H, Perron A, Park YK, Iwamoto Y, Knöpfel T (2012) activity alternates between motifs defined by regional axonal projections.
Imaging neural circuit dynamics with a voltage-sensitive fluorescent pro- Nat Neurosci 16:1426–1435.
tein. J Neurophysiol 108:2323–2337. Muller L, Reynaud A, Chavane F, Destexhe A (2014) The stimulus-evoked
Alexander DM, Jurica P, Trengove C, Nikolaev AR, Gepshtein S, Zvyagintsev population response in visual cortex of awake monkey is a propagating
M, Mathiak K, Schulze-Bonhage A, Ruescher J, Ball T, van Leeuwen C wave. Nat Commun 5:3675.
(2013) Traveling waves and trial averaging: the nature of single-trial and Muller L, Chavane F, Reynolds J, Sejnowski TJ (2018) Cortical travelling
averaged brain responses in large-scale cortical signals. Neuroimage waves: mechanisms and computational principles. Nat Rev Neurosci
73:95–112. 19:255–268.
Arieli A, Sterkin A, Grinvald A, Aertsen A (1996) Dynamics of ongoing activ- Murphy M, Riedner BA, Huber R, Massimini M, Ferrarelli F, Tononi G
ity: explanation of the large variability in evoked cortical responses. (2009) Source modeling sleep slow waves. Proc Natl Acad Sci U S A
Science 273:1868–1871. 106:1608–1613.
Azouz R, Gray CM (1999) Cellular mechanisms contributing to response var- Nir Y, Staba RJ, Andrillon T, Vyazovskiy VV, Cirelli C, Fried I, Tononi G
iability of cortical neurons in vivo. J Neurosci 19:2209–2223. (2011) Regional slow waves and spindles in human sleep. Neuron
Bruhn A, Weickert J, Schnörr C (2005) Lucas/Kanade meets Horn/Schunck: 70:153–169.
combining local and global optic flow methods. Int J Comput Vis 61:1– Patten TM, Rennie CJ, Robinson PA, Gong P (2012) Human cortical travel-
21. ing waves: dynamical properties and correlations with responses. PLoS
Canolty RT, Knight RT (2010) The functional role of cross-frequency cou- One 7:e38392.
pling. Trends Cogn Sci 14:506–515. Petersen CCH, Hahn TTG, Mehta M, Grinvald A, Sakmann B (2003)
Capone C, Rebollo B, Muñoz A, Illa X, Giudice P, Del Sanchez-Vives MV, Interaction of sensory responses with spontaneous depolarization in layer
Mattia M (2019) Slow waves in cortical slices: how spontaneous activity is 2/3 barrel cortex. Proc Natl Acad Sci U S A 100:13638–13643.
shaped by laminar structure. Cereb Cortex 29:319–335. Romano SA, Pietri T, Pérez-Schuster V, Jouary A, Haudrechy M, Sumbre G
Chen G, Gong P (2019) Computing by modulating spontaneous cortical ac- (2015) Spontaneous neuronal network dynamics reveal circuit’s func-
tivity patterns as a mechanism of active visual processing. Nature com- tional adaptations for behavior. Neuron 85:1070–1085.
munications 10:1–15. Rubino D, Robbins KA, Hatsopoulos NG (2006) Propagating waves mediate
Davis ZW, Muller L, Trujillo JM, Sejnowski T, Reynolds JH (2020) information transfer in the motor cortex. Nat Neurosci 9:1549–1557.
Spontaneous traveling cortical waves gate perception in behaving prima- Sanchez-Vives MV, McCormick DA (2000) Cellular and network mecha-
tes. Nature 587:432–436. nisms of rhythmic recurrent activity in neocortex. Nat Neurosci 3:1027–
Deco G, Hagmann P, Hudetz AG, Tononi G (2014) Modeling resting-state 1034.
functional networks when the cortex falls asleep: local and global changes. Scott G, Fagerholm ED, Mutoh H, Leech R, Sharp DJ, Shew WL, Knöpfel T
Cereb Cortex 24:3180–3194.
(2014) Voltage imaging of waking mouse cortex reveals emergence of
Fox MD, Snyder AZ, Vincent JL, Corbetta M, Van Essen DC, Raichle ME
critical neuronal dynamics. J Neurosci 34:16611–16620.
(2005) The human brain is intrinsically organized into dynamic, anticor-
Seth AK (2010) A MATLAB toolbox for Granger causal connectivity analysis.
related functional networks. Proc Natl Acad Sci U S A 102:9673–9678.
J Neurosci Methods 186:262–273.
Freeman WJ, Barrie JM (2000) Analysis of spatial patterns of phase in neo-
Shimaoka D, Song C, Knöpfel T (2017) State-dependent modulation of slow
cortical gamma EEGs in rabbit. J Neurophysiol 84:1266–1278.
wave motifs towards awakening. Front Cell Neurosci 11:108.
Fulcher BD, Murray JD, Zerbi V, Wang XJ (2019) Multimodal gradients
Smith GB, Hein B, Whitney DE, Fitzpatrick D, Kaschube M (2018)
across mouse cortex. Proc Natl Acad Sci U S A 116:4689–4695.
Distributed network interactions and their emergence in developing neo-
Garaschuk O, Linn J, Eilers J, Konnerth A (2000) Large-scale oscillatory cal-
cortex. Nat Neurosci 21:1600–1608.
cium waves in the immature cortex. Nat Neurosci 3:452–459.
Song C, Piscopo DM, Niell CM, Knöpfel T (2018) Cortical signatures of
Gong P, van Leeuwen C (2009) Distributed dynamical computation in neural
wakeful somatosensory processing. Sci Rep 8:11977.
circuits with propagating coherent activity patterns. PLoS Comput Biol 5:
e1000611. Takagaki K, Zhang C, Wu J-Y, Ohl FW (2011) Flow detection of propagating
Granger CW (1969) Investigating causal relations by econometric models waves with temporospatial. J Neurosci Methods 200:207–218.
and cross-spectral methods. Econometrica 37:424–438. Tononi G, Cirelli C (2003) Sleep and synaptic homeostasis: a hypothesis.
Hopf H (1927) Vektorfelder inn-dimensionalen mannigfaltigkeiten. Math Brain Res Bull 62:143–150.
Ann 96:225–249. Townsend RG, Gong P (2018) Detection and analysis of spatiotemporal pat-
Huang X, Xu W, Liang J, Takagaki K, Gao X, Wu J young (2010) Spiral wave terns in brain activity. PLoS Comput Biol 14:e1006643.
dynamics in neocortex. Neuron 68:978–990. Townsend RG, Solomon SS, Chen SC, Pietersen ANJ, Martin PR, Solomon
Knöpfel T (2012) Genetically encoded optical indicators for the analysis of SG, Gong P (2015) Emergence of complex wave patterns in primate cere-
neuronal circuits. Nat Rev Neurosci 13:687–700. bral cortex. J Neurosci 35:4662–4667.
Landsness EC, Crupi D, Hulse BK, Peterson MJ, Huber R, Ansari H, Coen Volgushev M, Chauvette S, Mukovski M, Timofeev I (2006) Precise long-
M, Cirelli C, Benca RM, Ghilardi MF, Tononi G (2009) Sleep-dependent range synchronization of activity and silence in neocortical neurons dur-
improvement in visuomotor learning: a causal role for slow waves. Sleep ing slow-wave oscillations [corrected]. J Neurosci 26:5665–5672.
32:1273–1284. Vyazovskiy VV, Faraguna U, Cirelli C, Tononi G (2009) Triggering slow
Lee SH, Blake R, Heeger DJ (2005) Traveling waves of activity in primary vis- waves during NREM sleep in the rat by intracortical electrical stimula-
ual cortex during binocular rivalry. Nat Neurosci 8:22–23. tion: effects of sleep/wake history and background activity. J
Marshall L, Helgadóttir H, Mölle M, Born J (2006) Boosting slow oscillations Neurophysiol 101:1921–1931.
during sleep potentiates memory. Nature 444:610–613. Wu JY, Huang X, Zhang C (2008) Propagating waves of activity in the neo-
Massimini M, Huber R, Ferrarelli F, Hill S, Tononi G (2004) The sleep slow cortex: what they are, what they do. Neuroscientist 14:487–502.
oscillation as a traveling wave. J Neurosci 24:6862–6870. Xiao D, Vanni MP, Mitelut CC, Chan AW, Ledue JM, Xie Y, Chen ACN,
Mishina Y, Mutoh H, Song C, Knopfel T (2014) Exploration of genetically Swindale NV, Murphy TH (2017) Mapping cortical mesoscopic networks
encoded voltage indicators based on a chimeric voltage sensing domain. of single spiking cortical or sub-cortical neurons. Elife 6:e19976.
Front Mol Neurosci 7:78. Xu W, Huang X, Takagaki K, Wu J. y (2007) Compression and Reflection of
Mitra A, Kraft A, Wright P, Acland B, Snyder AZ, Rosenthal Z, Czerniewski Visually Evoked Cortical Waves. Neuron 55:119–129.
L, Bauer A, Snyder L, Culver J, Lee JM, Raichle ME (2018) Spontaneous Zanos TP, Mineault PJ, Nasiotis KT, Guitton D, Pack CC (2015) A sensori-
infra-slow brain activity has unique spatiotemporal dynamics and lami- motor role for traveling waves in primate visual cortex. Neuron 85:615–
nar structure. Neuron 98:297–305. 627.
Mohajerani MH, Chan AW, Mohsenvand M, Ledue J, Liu R, McVea DA, Zhang H, Watrous AJ, Patel A, Jacobs J (2018) Theta and alpha oscillations
Boyd JD, Wang YT, Reimers M, Murphy TH (2013) Spontaneous cortical are traveling waves in the human neocortex. Neuron 98:1269–1281.