Complex Oscillatory Waves Emerging From Cortical Organoids Model Early Human Brain Network Development

Short Article
Complex Oscillatory Waves Emerging from Cortical

Organoids Model Early Human Brain Network
Development
Graphical Abstract Authors
Cleber A. Trujillo, Richard Gao,
Priscilla D. Negraes, ..., Gene W. Yeo,
Bradley Voytek, Alysson R. Muotri
Correspondence
[email protected]
In Brief
Oscillatory activity is a candidate
mechanism for how neural populations
are temporally organized. Cortical
organoids exhibit periodic and highly
regular nested oscillatory network events
that are dependent on glutamatergic and
GABAergic signaling. The emerging
development of network activity
transitions to more spatiotemporally
complex activity, capturing features of
preterm infant electroencephalography.
Highlights
d Long-term, single-cell transcriptomics reveals cortical
organoid developmental dynamics
d Cortical organoids exhibit phase-amplitude coupling during

network-synchronous events
d Differential role of glutamate and GABA in initiating and

maintaining oscillations
d Network-level events are similar to the human preterm

neonatal EEG features
Trujillo et al., 2019, Cell Stem Cell 25, 558–569

October 3, 2019 ª 2019 Elsevier Inc.
https://doi.org/10.1016/j.stem.2019.08.002
Cell Stem Cell
Short Article
Complex Oscillatory Waves Emerging from Cortical

Organoids Model Early Human Brain
Network Development
Cleber A. Trujillo,1,2,10 Richard Gao,3,10 Priscilla D. Negraes,1,2,10 Jing Gu,4 Justin Buchanan,4 Sebastian Preissl,4
Allen Wang,4 Wei Wu,1 Gabriel G. Haddad,1,5 Isaac A. Chaim,2 Alain Domissy,2 Matthieu Vandenberghe,6 Anna Devor,6,7
Gene W. Yeo,2 Bradley Voytek,3,8,11 and Alysson R. Muotri1,2,8,9,11,12,*
1Department of Pediatrics/Rady Children’s Hospital San Diego, School of Medicine, University of California, San Diego, La Jolla, CA
92093, USA
2Department of Cellular & Molecular Medicine, School of Medicine, University of California, San Diego, La Jolla, CA 92093, USA
3Neurosciences Graduate Program, Institute for Neural Computation, Department of Cognitive Science, University of California, San Diego,
La Jolla, CA 92093, USA

4Center for Epigenomics, Department of Cellular & Molecular Medicine, University of California, San Diego, La Jolla, CA 92093, USA
5Department of Neurosciences, University of California, San Diego, La Jolla, CA 92093, USA
6Department of Radiology, Department of Neurosciences, University of California, San Diego, La Jolla, CA 92093, USA
7Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Harvard Medical School, Charlestown, MA 02129, USA
lu Data Science Institute, University of California, San Diego, La Jolla, CA 92093, USA
8Kavli Institute for Brain and Mind and Halıcıog
9Center for Academic Research and Training in Anthropogeny (CARTA), La Jolla, CA 92093, USA
10These authors contributed equally
11Senior author
12Lead Contact
*Correspondence: [email protected]
https://doi.org/10.1016/j.stem.2019.08.002
SUMMARY human brain. Neural oscillations, a prominent, rhythmic brain

signal found across species, robustly track behavioral and
Structural and transcriptional changes during early disease states and have long been leveraged in systems
brain maturation follow fixed developmental pro- neuroscience due to their ubiquity and accessibility (Buzsáki
grams defined by genetics. However, whether this and Draguhn, 2004; de Hemptinne et al., 2015; Fries, 2005; Hen-
is true for functional network activity remains un- riques and Davidson, 1991; Khan et al., 2013; Uhlhaas and
known, primarily due to experimental inaccessibility Singer, 2010). These complex network dynamics emerge early
in development and is unclear whether shaped exclusively by
of the initial stages of the living human brain. Here,
biological programming prenatally (Blankenship and Feller,
we developed human cortical organoids that dynam-
2010; Johnson, 2001; Power et al., 2010). In vitro and in vivo
ically change cellular populations during maturation rodent studies have shown that a conserved repertoire of
and exhibited consistent increases in electrical organized network activity, such as traveling waves, giant depo-
activity over the span of several months. The sponta- larizing potentials, and early network oscillations, develops ac-
neous network formation displayed periodic and cording to a consistent timeline prior to and immediately after
regular oscillatory events that were dependent on birth (Allène et al., 2008; Khazipov and Luhmann, 2006; Uhlhaas
glutamatergic and GABAergic signaling. The oscilla- et al., 2010). However, due to an inability to interrogate the elec-
tory activity transitioned to more spatiotemporally trophysiology of intact embryonic brains, it remains unknown
irregular patterns, and synchronous network events whether the same happens in humans. As a result, our knowl-
resembled features similar to those observed in pre- edge about human brain functional development rests upon
observations from nonhuman model systems.
term human electroencephalography. These results
Organoids generated from induced pluripotent stem cells
show that the development of structured network
(iPSCs) have emerged as a scaled-down and three-dimensional
activity in a human neocortex model may follow model of the human brain, mimicking various developmental
stable genetic programming. Our approach provides features at the cellular and molecular levels (Camp et al., 2015;
opportunities for investigating and manipulating the Cederquist et al., 2019; Giandomenico et al., 2019; Lancaster
role of network activity in the developing human and Knoblich, 2014; Lancaster et al., 2013; Luo et al., 2016; Ma-
cortex. riani et al., 2012; Paşca et al., 2015; Qian et al., 2016; Renner
et al., 2017; van de Leemput et al., 2014; Xiang et al., 2017,
2019). Despite recent advances in the understanding of their
INTRODUCTION cellular diversity, there is no evidence that these organoids
develop complex and functional neural network activity that
Diverse and hierarchical cellular networks develop into circuits resembles early human brain formation (Birey et al., 2017; Gian-
with patterns of functional spatiotemporal activity to form the domenico et al., 2019; Quadrato et al., 2017). Therefore,
558 Cell Stem Cell 25, 558–569, October 3, 2019 ª 2019 Elsevier Inc.
Figure 1. Cellular and Molecular Development of Human Cortical Organoids
(A) Overview of human neural network formation and dynamics evaluation using cortical organoids.
(B) Schematic of the protocol used to generate cortical organoids. Scale bar, 200 mm.
(legend continued on next page)
Cell Stem Cell 25, 558–569, October 3, 2019 559
researchers have not yet clearly determined whether brain orga- Figures 1D and S1). Although the initial fraction of GFAP-positive
noids are a suitable model for neural network dynamics (Kelava cells was less than 5%, this population increased to about 30%–
and Lancaster, 2016; Paşca, 2018). 40% after 6 months of differentiation (Figures 1D and 1E).
Here, we use human iPSCs to generate cortical organoids that To characterize the cellular diversity of cortical organoids dur-
exhibit evolving cellular transcriptional profile and nested oscilla- ing development, we performed single-cell RNA-seq on 1-, 3-,
tory network dynamics over the span of several months. We 6-, and 10-month organoids (Figures 1H–1M and S1; Table
subsequently investigated the molecular basis of oscillatory ac- S1). We used unsupervised clustering on the combined dataset
tivity formation, maintenance, and temporal control. Finally, we of 15,990 cells to identify clusters and their relative abundance at
applied supervised machine learning with cross-validation to distinct time points. Based on the expression gene markers, we
evaluate the similarity between electrophysiological network ac- combined smaller subclusters into five major cell classes:
tivity patterns of the in vitro model and human preterm neonatal progenitors; intermediate progenitors; glial cells; glutamatergic
electroencephalogram (EEG). Our findings suggest that orga- neurons; and GABAergic neurons. Based on this annotation,
noid models are suitable for the investigation of the physiological 1-month organoids consisted of >70% progenitor cells (express-
basis of network formation at early and late stages of the human ing SOX2 and PAX6; Figures 1J and 1M). At the 3- and 6-month
brain development. This prolonged evaluation of cortical orga- stage, cortical organoids comprised mainly glia (SLC1A3) and
noid activity expands our understanding of the emergence of glutamatergic neurons (GRIA2 and SNAP25; Figures 1J and
network-level neurodynamics in humans. 1M). The glial cells started with a small population and increased
to around 40% of cells present in the cortical organoids. Remain-
RESULTS ing populations of progenitors (around 5%) and intermediate
progenitors (around 10%) were present throughout the matura-
Generation of Functional Cortical Organoids tion. A fraction of glutamatergic neurons at the 3- and 6-month
Despite the structural and transcriptional similarities between time point expressed subunits of GABAergic receptors, such
brain organoids and the developing nervous system, the emer- as GABRB3 (Figure S1C). This expression of GABAergic recep-
gence of higher level complex network activity comparable to tors predates the appearance of interneurons.
the living human brain remains largely untested (Figure 1A). To GABAergic neurons were mainly restricted to 6- to 10-month
investigate the formation of a functional network, we promoted organoids, as indicated by expression of GAD2 (also known as
cortical specification by previously described protocols (Paşca GAD65), DLX1, and DLX5 (Figures 1J–1M, S1C, and S1D),
et al., 2015; Thomas et al., 2017; Yoon et al., 2019; Figure 1B; reaching around 15% of the total neuronal population after
see STAR Methods for details). At the beginning of differentia- 10 months of maturation, consistent with its presence later in
tion, an abundance of proliferative neural progenitor cells the in vivo development (Uylings et al., 2002). The molecular pro-
(NPCs) (expressing Ki67, SOX2, and Nestin) that self-organized file of GABAergic neurons was further evaluated by single-cell
into a polarized neuroepithelium-like structure was observed. transcriptomics (Figure S1E) and by the presence of protein
Similar to human cortical development in vivo, the proliferative markers (Figure S1F). To demonstrate the biosynthesis of
zone around a lumen delimited by b-catenin+ cells was sur- GABA during the maturation process, we employed metabolo-
rounded by progenitor cells. Progressively, the organoids mics liquid chromatography coupled to mass spectrometry
increased in size and in the proportion of mature neurons (ex- (Gertsman et al., 2014). The neurotransmitter GABA was de-
pressing NeuN and MAP2) to ultimately develop into concentric tected in the culture media of cortical organoids after 6 months
multilayer structures composed of NPCs, intermediate progeni- of maturation (Figure S1G) in a physiologically relevant concen-
tors (TBR2; also known as EOMES), and lower (CTIP2; also tration (Van Hove and Thomas, 2014). These results suggest the
known as BCL11B) and upper (SATB2) cortical layer neurons presence of the basic components for the generation of a neural
(Figures 1C–1E and S1). The neurons exhibit dendritic protru- network in a developing human cortical in vitro model.
sions and synaptic ultrastructure (Figures 1F and 1G). After
6 months, inhibitory neurons can also be observed (calretinin Emergence of Nested Oscillatory Network Activity
[CR], also known as CALB2; GABAB; NKX2.1, also known as In addition to the observed cellular diversity and expression of
TTF1; GABA; LHX6; somatostatin [SST]; and parvalbumin [PV]; synaptic markers, we interrogated the presence of functional
(C) Organoid growth during different developmental stages.

(D) Representative immunostainings showing proliferating NPCs (Ki67 and Nestin), lower (TBR1 and CTIP2) and upper (SATB2) cortical layer neurons, glial cells
(GFAP), and GABAergic (CR) neurons over time. Scale bars, 50 mm.
(E) Population analysis of specific markers indicating stages of maturation and multiple neuronal subtypes. The data are shown as mean ± SEM (n = 8).
(F) Representative image of a pyramidal neuron; dendritic structures are observed in cells transduced with the SYN:EGFP reporter (scale bar, 5 mm).
(G) Electron microscopy of synaptic ultrastructure in 4-month cortical organoids (blue). Scale bar, 200 nm.
(H) Uniform manifold approximation and projection (UMAP) plot of 15,990 cells from integrating datasets on 1-, 3-, 6-, and 10-month cortical organoids. Colors
denote cells sampled from four different time points.
(I) UMAP plot of the integrated datasets colored by seven main cell clusters: red as GABAergic neurons; orange as glutamatergic neurons; blue as glia cells; green
as intermediate progenitors; purple as progenitors; green blue as mitotic cells; and gray as others.
(J) Separate UMAP plots of integrated data by different time points. Same color scheme used for main cell clusters is shown.
(K) Dot plots showing cluster-specific gene expression across main cell clusters.
(L) UMAP plots showing expression levels of cell-type-specific markers (see Figure S1 for additional markers).
(M) Bar plots of proportion of cell types at individual time points.
560 Cell Stem Cell 25, 558–569, October 3, 2019

Figure 2. Oscillatory Network Dynamics in Long-Term Cortical Organoids
(A) Schematic of the organoid electrophysiological signal processing pipeline. Raw MEA data are analyzed as population spiking and LFP separately. Syn-
chronous network events are highlighted in yellow.
(B) Raster plot of network spiking activity after 1.5 and 8 months of maturation. A 3-s interval of activity over 5 channels is shown in the upper right corners.
(C) Cortical organoids show elevated and continuously increasing mean firing rate compared to 2D monolayer neurons (n = 8 for organoid cultures and n = 12 for
2D neurons). Inset: correlation of the firing rate vector over 12 weeks of differentiation (from 8 to 20) between pairs of cultures showing reduced variability among
organoid replicates.
(D) Temporal evolution of cortical organoid network activity. Detailed definitions and further parameters are presented in Figure S2.
(E) Time series of population spiking and LFP during network events in cortical organoid development. Each overlaid trace represents a single event during the
same recording session.
(F) The number of subpeaks during an event is maximized and stereotypical at 6 months, developing nonlinearly and following an inverted-U trajectory.
(G) Network variability, measured as the coefficient of variation of the inter-event interval, increases monotonically throughout development.
(legend continued on next page)

network activity. Starting at a single cellular level, we used temporal correlation between spontaneous network events.
whole-cell patch-clamp recording from 6-month cortical organo- The inter-event interval CV consistently increased over
ids (Figures S2A–S2E). The action potential firing activity and the 10 months of differentiation (Figure 2G), from extremely regular
voltage-dependent Na+ current were tetrodotoxin (TTX) sensi- latencies (CV y 0) at 2 months to irregular, Poisson-like
tive. Application of glutamate receptor antagonists (NBQX and (CV y 1) at 10 months. This indicates increased variability be-
AP5) fully inhibited the spontaneous excitatory postsynaptic tween consecutive network events initiation. Additionally, spatial
currents recorded at 60 mV, confirming the presence of func- and temporal irregularity on a shorter timescale (within event)
tional excitatory neurons. also increased with development, suggesting a breakdown of
To further evaluate the cortical organoid functionality in a deterministic population dynamics from the onset of network
mesoscopic level, we performed weekly extracellular recordings events.
of spontaneous electrical activity using multi-electrode arrays Periodic oscillatory activity is often defined as a ‘‘bump’’ over
(MEAs). Cortical organoids were plated per well in 8 wells of a the characteristic 1/f background in the power spectral density
MEA plate containing 64 low-impedance (0.04 MU) platinum (PSD) of extracellular signals above and beyond the aperiodic
microelectrodes with 30 mm of diameter spaced by 200 mm, 1/f signal (Buzsáki et al., 2013; Gao et al., 2017). In organoid
yielding a total of 512 channels. We separately analyzed sin- LFPs, we observed both prominent oscillatory peaks in the
gle-channel and population firing characteristics derived from low-frequency range (1–4 Hz) and in the aperiodic signal charac-
channel-wise spike times and the local field potential (LFP), a teristic of neural recordings (Ben-Ari, 2001; Voytek et al., 2015).
measure of aggregate synaptic currents and other slow ionic The development of oscillatory activity in cortical organoids over
exchanges (Buzsáki et al., 2012; Figure 2A). The spikes were time was quantified by computing the PSD for each LFP
defined by the event unit waveforms standard structure with recording (Figure 2H, inset). Oscillatory power in the delta range
typical refractory periods and by pharmacological intervention. (1–4 Hz) increased for up to 24 weeks in culture, tapering off
These spikes from each channel do not represent putative sin- slightly in subsequent recordings and plateauing during the
gle-unit action potentials but represent multi-unit activity last 10 weeks. This inverted-U trajectory reflects the network’s
(MUA). Because both the spatial and temporal resolution of initial acquisition of oscillatory modes at steady frequencies
MEA sampling is sparse, single-unit spike trains were not iso- and the dispersion of this regularity at later time points. The
lated, instead submitting channel-wise and whole-well activity LFP results reveal the development of the cortical organoid cul-
for further analysis rather than individual spike trains. Over the tures across different network states: from sparse activity with
course of 10 months, cortical organoids exhibited consistent in- extreme rigidity and regularity to one that acquires repetitive
creases in electrical activity, as parametrized by channel-wise and regular oscillatory patterns (Voytek and Knight, 2015), until
firing rate, burst frequency, and synchrony (Figures 2B–2D and it finally reaches a stage of higher spatiotemporal complexity
S2F–S2I), which indicates a continually evolving neural network. and variability that is reminiscent of self-organized networks
Additionally, the variability between replicates over 40 weeks of (Tetzlaff et al., 2010; Figures 2I, 2J, and S3C–S3F).
differentiation was significantly lower compared to iPSC-derived
neurons in monolayer cultures (Figures 2C, inset, and S2J). Oscillatory Coordination of Neural Ensembles and Its
During individual recordings, cultures displayed a robust Synaptic Mechanisms
pattern of activity, switching between long periods of quies- Oscillatory dynamics have been postulated to coordinate spiking
cence and short bursts of spontaneous network-synchronized across neural ensembles. In the LFP and other mesoscopic brain
spiking (hereafter referred to as ‘‘network events’’). These signals, this manifests as a phenomenon known as cross-fre-
network events were periodic (0.05 Hz) but infrequent early in quency phase-amplitude coupling (PAC) (Voytek and Knight,
development (2 months), occurring roughly every 20 s and 2015), wherein the high-frequency content of the LFP is
decayed monotonically after the initial onset (Figure 2E). From entrained to the phase of slow oscillations (Manning et al.,
4 months onward, a secondary peak emerged 300–500 ms after 2009; Miller et al., 2007; Mukamel et al., 2005). In the cortical
the initial network activation, leading to the presence of a nested organoids, we observed greater PAC between oscillatory delta
faster oscillatory (2 or 3 Hz) pattern up to 6 months in culture (Fig- (1–4 Hz) and broadband gamma activity (100–400 Hz; see
ures 2F and S3A–S3F). Notably, this robust fast timescale nested STAR Methods) during network events compared to quiescent
oscillation was not observed in 3D neurospheres, suggesting periods (Figures 3A–3C). This broadband gamma is non-oscilla-
that the spherical arrangement of neurons is insufficient for the tory but has been shown to be an LFP surrogate of population
emergence of nested oscillations (Figures S3G–S3J). The regular spiking (Manning et al., 2009; Miller et al., 2007).
oscillatory activity during network events transitioned to stron- We further evaluated the role of glutamatergic and GABAergic
ger, yet more variable, oscillations over time. To quantify this synaptic transmission in forming oscillations by pharmacological
network complexity, we tracked the regularity (coefficient of intervention. Organoid neural networks were susceptible to both
variation of inter-event intervals [CV]) and the spatial and glutamate receptor antagonists (AP5 and CNQX; NMDA and
(H) 1- to 4-Hz oscillatory power in the LFP increases up to the 25th week in culture and plateaus at 30 weeks. Inset: oscillatory power is calculated by fitting a
straight line (dashed) over the aperiodic portion of the PSD and taken as the height of narrow peaks rising above the linear fit.
(I) Pairwise correlation of LFP across all electrodes (coherence) within a well during network events initially increases and then decreases after 30 weeks.
(J) An example of sequential frames during a network event shows the spatial propagation of wave spreading and then disappearing again after 100 ms.
The data shown in (C), (D), (F), (G), (H), and (I) are presented as mean ± SEM. *p < 0.05; **p < 0.01; ***p < 0.001; unpaired Student’s t test (C), quadratic (F, H, and I)
and linear (G) regression.

Figure 3. Cortical Organoid Serves as a
Model of Functional Oscillations and Their
Synaptic Mechanisms
(A–C) Phase-amplitude coupling is observed in
organoid LFP during network events.
(A) Example of raw LFP during a network event
decomposed into its low-frequency component
(1- to 4-Hz delta) and the amplitude envelope of
the high-frequency, broadband gamma compo-
nent (200–400 Hz). Analysis was repeated for
100–200 Hz with near identical effect size and
significance.
(B) Normalized gamma amplitude binned by delta
phase during network events (black) shows
greater modulation depth by low-frequency delta
than during non-event periods (red).
(C) Phase-amplitude coupling during network
events is significantly greater than non-event pe-
riods in all batches.
(D) Effect of selective drug treatments on neuronal
electrical activity in 6-month organoids. Repre-
sentative raster plots and burst measurements of
untreated and treated organoids are shown. The
pharmacological manipulation was performed
using cortical organoid plated on 4 MEA wells
(n = 4; cortical organoid culture for each treat-
ment). Scale bar, 20 s. Exposure to AP5 + CNQX,
baclofen, and muscimol reversibly extinguish the
network bursts (synchrony), and no changes were
promoted by bicuculline.
(E and F) Pharmacological perturbation of network
events (E) and oscillatory activity (F) during
network events in 6-month organoids. Pre and
post refer to before treatment administration and
after administration, respectively. Application of
bicuculline and picrotoxin increases the number of
network events, and CNQX + AP5 and baclofen
completely abolish synchronized network events.
Bicuculline blocks oscillatory network activity, but
not the network event itself. Data are shown as
mean ± SEM; unpaired Student’s t test.
AMPA/kainate, respectively) and GABA receptor agonists discontinu (Tolonen et al., 2007), quiescent periods are
(muscimol, GABAA; baclofen, GABAB) by significantly reducing punctuated by high-amplitude oscillations (spontaneous activity
the number of spike events and bursts, with a subsequent transients [SATs]) lasting a few seconds. Intervals of complete
extinction of synchronous activity. The electrical activity was quiescence disappear as infants become of term, and the EEG
abolished in the presence of TTX (Figures 3D and 3E). Blockade is dominated by continuous and low-amplitude desynchronized
of GABAergic transmission by bicuculline increased the number activity in adult brains (Figures 4A, S4A, and S4B).
of network-synchronized events and did not affect peak popula- Due to the inability to interrogate the electrophysiology of
tion firing rates but abolished nested 2-Hz oscillatory activity by intact human embryonic brains, we attempted to quantitatively
erasing subsequent reverberant peaks (Figure 3F). compare network activity in cortical organoids to preterm human
EEG. We analyzed a publicly available dataset of 101 serial EEG
Cortical Organoid Network Development Resembles recordings from 39 preterm infants ranging from 24 to 38 weeks
Some Preterm EEG Features post-menstrual age (PMA): 567 data points total (Stevenson
Despite emergence of complex oscillatory network activity in et al., 2017). The dataset contains 23 precomputed features for
organoids, it is unclear whether the spontaneous developmental each EEG record, ranging from timing, rate, and variability of
trajectory observed is representative of programmed early SATs (or bursts), as well as spectral power in canonical oscilla-
neurodevelopment. Although network activity from cortical orga- tory bands (delta, theta, etc.; see Table S2 for full list of features).
noids does not exhibit the full temporal complexity seen in We computed analogous features from each organoid LFP
adults, the pattern of alternating periods of quiescence and recording when appropriate. It is important to note that the
network-synchronized events resembles electrophysiological biophysics of scalp EEG is drastically different from extracellular
signatures present in preterm human infant EEG. During trace field potential in the organoid, due to factors such as spatial

Figure 4. Cortical Organoid Network Dynamics Mimic Those of Premature Neonates after 28 Weeks of Maturation
(A) Representative LFP trace from cortical organoid, highlighting instances of network events (yellow). Comparable events between periods of quiescence
(discontinuous network dynamics) are shown in human preterm neonate EEG at 35 weeks gestational age, and a different pattern of continuous activity is
observed in adult EEG. SAT, spontaneous activity transient.
(B) Examples of analogous features in preterm neonate EEG and organoid LFP show various levels of similarity throughout development. RMS, root-mean-
square; 50% and 95% refer to 50th and 95th percentile of feature distribution within a recording.
(C) Pearson’s correlation coefficients between age and electrophysiological features (mean ± SD of bootstrapped distribution) for both organoid and premature
neonates show different degrees of developmental similarity for individual features (12 total selected). For example, SATs (events) per hour shows remarkable
similarity over time between organoid and neonates.
(D) Schematic of machine learning procedure for age prediction: EEG features from 39 premature neonates (n = 567 recordings) between 25 and 38 weeks PMA
(post-menstrual age) were used to train and cross-validate a regularized regression model (ElasticNet), optimizing for preterm neonate age prediction based on
their EEG features only (top). The model was clamped after training and applied directly on organoid LFP features and control datasets, including held-out preterm
neonate data, mouse primary culture, 2D iPSC culture, and human fetal brain culture.
(E) Model-predicted developmental time (y axis, age in weeks) follows actual weeks in culture (x axis) for organoids (orange and blue), as well as true age of held-
out preterm neonate data points (black). Dashed line represents unity, signifying perfect prediction. Large circles on solid lines and shaded regions denote mean ±
SD of prediction, respectively, and dots indicate per sample prediction (n = 8 for organoids at all time points).
(F) Pearson’s correlation coefficient between predicted and actual developmental time for organoid and control datasets. Significant positive correlations indicate
the model’s ability to capture developmental trajectory in a particular dataset.

filtering by the scalp and orientation of neuronal populations in observed for organoid and held-out EEG datasets (Figure 4F).
relation to the recording electrode. Therefore, we selected a sub- Although the developmental trajectory of cortical organoids is
set of 12 features to compare in the organoid LFP (highlighted in not identical to, and more variable than, that of the fetal human
Table S2); the majority of those correspond to duration and brain, the two populations share similarities in how their network
timing of SATs. Although features like EEG SD (or root-mean- electrophysiological properties change over time, suggesting
square [RMS]) and interhemispheric synchrony are likely altered genetically programmed developmental timelines that can be
by skull thickening during early development, the large-ampli- detected by a simple machine learning algorithm.
tude network synchronous events are reliably detected in both
EEG and LFP. Timing features derived from SAT times (duration,
DISCUSSION
inter-SAT interval, etc.) were compared between cortical orga-
noid and preterm neonates.
Although brain organoids have been shown to mimic early
By comparing specific timing features between cortical orga-
human neurodevelopment at the cellular and molecular levels,
noids and preterm infants, we found a range of correlations in the
evidence of network activity maturation and the corresponding
developmental trajectory of features with age, as well as similar-
cellular basis have not been previously explored. Here, we report
ities in development between the two datasets (Figures 4B and
the formation of small-scale functional electrophysiological
C). For instance, ‘‘SATs per hour’’ (‘‘events per hour’’ in organo-
networks in human cortical organoids while tracking their gene
ids; Figure 4B) and 95th percentile of inter-SAT duration distribu-
expression profile and cellular composition over time. Single-
tion showed high similarity both in absolute value and their
cell RNA sequencing (RNA-seq) at multiple time points spanning
developmental trajectory (correlation with age), and ‘‘root-
10 months show development of various cellular subclusters,
mean-square SAT duration’’ and median (50%) SAT duration
transitioning from progenitor cells to neuronal and glial
showed different trends and absolute value, from 25 to 38 weeks
populations.
in both datasets (all features presented in Figures S4C and S4D).
Cortical organoids begin to exhibit highly synchronous and
To compare the similarity of developmental trajectory quantita-
stereotypical network activity at 2 months, which transition into
tively, we computed the average resampled correlation between
2- or 3-Hz rhythmic activity by 4–6 months. Subsequently,
each feature and developmental time in both datasets (Figures
network activity becomes more variable spatiotemporally, coin-
4C and S4D; see STAR Methods for details). These results sum-
ciding with the development of inhibitory populations. Oscillatory
marize what is shown in Figure 4B: SATs per hour consistently
activity at 6 months exhibits cross-frequency coupling, a poten-
increase during development in both organoids and preterm in-
tial signature of functional neuronal network communication;
fants, while the variability of SAT duration (TSAT RMS) consis-
pharmacological intervention demonstrates the causal involve-
tently decreases. Other features showed inconsistent develop-
ment of glutamate and GABA in generating and sustaining oscil-
mental trajectories over time between organoid LFPs and
lations. Finally, we observe similarities in the developmental tra-
preterm EEGs.
jectory of some electrophysiological features between
Taking into consideration the wide range of similarities
organoids and human preterm infants, where a machine learning
observed across the two datasets, we asked which features’
model trained to predict neonatal age from their EEG features
developmental trajectory was most informative of the develop-
can predict organoid developmental timeline. Taken altogether,
mental time and whether those were conserved between orga-
these results demonstrate the utility of human-stem-cell-derived
noids and preterm infants. To accomplish this in an objective
brain organoids as a viable neuroscience research model, not
fashion, we trained a regularized regression model with cross-
only for the shifting landscape of molecular and cellular compo-
validation (ElasticNet; L1 and L2 regularized) to predict preterm
sition but also for the maturation of functional activity in brain net-
infant age from their EEG features. In other words, the regression
works during early neurodevelopment.
model was only optimized to predict preterm infant age based on
their EEG and was blind to the organoid data. Following training
and hyperparameter selection, the regression model was Diversity of Excitatory and Inhibitory Populations
‘‘locked’’ while we directly applied it on the organoid LFP dataset We used longitudinal single-cell transcriptional profiling followed
and various control datasets to obtain their predicted develop- by immunostaining and functional validation to demonstrate the
mental time (Figure 4D). cellular dynamics of cortical organoids during long-term devel-
Although the regression model predicted organoid develop- opment, revealing an unprecedented diversity of cell types.
mental time poorly before 25 weeks (Figure 4E, orange) and Notably, GABAergic neurons were mainly restricted to 6- to
with high variability, mean predicted developmental time fol- 10-month organoids, reaching around 10%–15% of the total
lowed culture time with much higher fidelity after 25 weeks neural population after 10 months, consistent with its presence
(blue). A subset of the preterm EEG data held out during training later in the in vivo development (Uylings et al., 2002). A metabo-
was used to further validate the performance of the model lomic identification of GABA released in the culture media was
(black), in addition to other control datasets, including mouse pri- further used to validate the presence and functionality of the
mary culture, iPSC monolayer culture, and human fetal brain cul- GABAergic system. Although our aim was not to investigate
ture (Figure S4E; details in STAR Methods). To quantify how well the origin of GABAergic neurons in the human neocortex, we
developmental trend over time was captured by the regression cannot exclude the possibility of aberrant cellular differentiation,
model, we computed the Pearson correlation coefficient be- warranting further dissection of the cortical organoid model for
tween the model-predicted age and the true age of the various novel neurodevelopmental pathways. The dynamic cell popula-
datasets. Note that a significant positive correlation was only tion and the presence of neurotransmitter systems suggest the

activity of the basic components for the emergence of a neural may arise from the EEG acquisition system and electrodes posi-
network in a developing in vitro model. tioning. Second, clinical confounds due to potential neurological
condition and medications may also impact the dataset. Lastly, it
Synchronous Oscillations as a Signature of Functional is important to highlight that the regression model cannot be
Network Activity extended to neurotypical adult, as adult EEGs do not display
The presence and changes of oscillations at fast timescales the observed bursting patterns under normal conditions; thus,
(>1 Hz) is a hallmark of the in vivo brain, and coupling across the relevant features (e.g., SAT timing features) cannot be
different frequencies has been proposed to coordinate the flow computed. Nonetheless, although we do not claim functional
of information across regions (Buzsáki, 2004; Fries, 2005; Voytek equivalence between the organoids and a full cortex—neonatal
et al., 2015). With the cellular components for the generation of a or adult—the current results represent the first step toward an
functional neural network in place, we tested whether the cortical in vitro model that captures some of the complex and oscillatory
organoids display activity typically found in organized cortical spatiotemporal dynamics of the human brain.
networks. Robust extracellular electrical activity was observed
at earlier stages and progressively developed into an organized Conclusions
oscillatory network. Cortical organoids initially exhibited periodic Given the potential roles of synchronized and oscillatory network
and highly regular nested oscillatory network events that were dynamics in coordinating information flow between developed
dependent on glutamatergic and GABAergic signaling. Our brain regions (Uhlhaas et al., 2010), these results highlight the
data suggest that GABA transmission is crucial for the mainte- potential for cortical organoids to advance our understanding
nance, but not the initiation, of faster oscillatory activity. This is of functional electrophysiology. Additionally, by applying spiking
consistent with accounts of inhibition rhythmically coordinating and LFP analysis that is traditional to animal models, our findings
pyramidal populations’ activity during early development (Opitz offer a link between microscale organoid physiology and sys-
et al., 2002). Additionally, during periods of high network activity, tems neuroscience. Finally, considering the diversity and matu-
the power of high-frequency (>100 Hz) activity is coupled to the ration of cell types generated, the robustness of the neuronal
phase of the 3-Hz oscillation in the LFP. Without positing its networks, the presence of structural traits of mature neurons,
functional role, this observation suggests that more complex and the possibility of using sensory experience to modulate
oscillatory activity can indeed manifest and be studied in this neuronal activity collectively, cortical organoids may be used
in vitro system. to model cellular interactions and neural circuit dysfunctions
It is also unclear the biological basis of the increased variability related to neurodevelopmental and neuropsychiatric pathol-
in the number of spontaneous events in organoid cultures, espe- ogies. Importantly, this organoid model is small, approximately
cially after 28 weeks in culture. We believe that different factors one million times smaller than the human brain, but ethical
could increasingly introduce variability or diversity into the neural implications cannot be ignored about the future possibility of
population during maturation. In this context, because we larger and more complex organoids (Farahany et al., 2018).
started with single cells that aggregate to form organoids, small Nevertheless, our findings illuminate a link between microscale
population differences at early stages of organoid formation organoid physiology and systems neuroscience. This offers a
could lead to changes in activity. Differences in the organoid promising, small-scale experimental model of human neocortex
positioning on the MEA and manipulation could also affect the to help address neurodevelopmental pathologies that affect
signal acquisition. Additionally, we do not exclude the possibility millions of people but otherwise lack an existing animal model.
of the formation of independent network profiles based on
intrinsic activity and retro-feedback properties. STAR+METHODS
Comparing to the Early Developing Brain: Insights and Detailed methods are provided in the online version of this paper
Limitations and include the following:
Some features of early network dynamics in humans (e.g., SATs)
can be recapitulated by the in vitro model, with no additional d KEY RESOURCES TABLE
constraints other than structural and genetic similarities. The d LEAD CONTACT AND MATERIALS AVAILABILITY
regularized regression model presented here was built on d EXPERIMENTAL MODEL AND SUBJECT DETAILS
preterm EEG data only, following an internal cross-validation B Human cell source
procedure to estimate the hyperparameters. It was then directly B Rodent cell source
applied to organoid LFP data—previously unseen by the classi- d METHOD DETAILS
fier—to produce a ‘‘predicted developmental time,’’ in addition B Generation of cortical organoids
to data from several other cellular models for validation. Simulta- B Neurosphere generation
neous MEA and EEG seizure recordings, in human subjects, B Mycoplasma testing
share common features in the EEG frequency range (Schevon B Immunofluorescence staining
et al., 2012). However, when comparing the in vitro MEA and B Electron microscopy (EM)
neonatal EEG features, it is crucial to remove any comparison B 10X genomics single-cell and analysis
of features affected by the spatial filtering properties of the skull. B Mass spectrometry
Moreover, there are a few factors that might challenge the inter- B Whole-cell patch-clamp
pretation of the regression model results. First, it is difficult to B Multi-electrode array (MEA) recording
control external variation in infant EEG due to differences that B Custom MEA analysis

B Network event analysis Received: September 4, 2018
B Oscillatory spectral power analysis Revised: May 3, 2019
B Phase Amplitude Coupling (PAC) Accepted: August 6, 2019
Published: August 29, 2019
B Pharmacology
B Preterm neonatal EEG REFERENCES
B Resampled feature-age correlation
B Neonate-organoid development time regres- Allène, C., Cattani, A., Ackman, J.B., Bonifazi, P., Aniksztejn, L., Ben-Ari, Y.,
sion model and Cossart, R. (2008). Sequential generation of two distinct synapse-driven
d QUANTIFICATION AND STATISTICAL ANALYSIS network patterns in developing neocortex. J. Neurosci. 28, 12851–12863.
B Statistical analysis Ben-Ari, Y. (2001). Developing networks play a similar melody. Trends
Neurosci. 24, 353–360.
B Statistics and Regression for custom MEA analysis
d DATA AND CODE AVAILABILITY Birey, F., Andersen, J., Makinson, C.D., Islam, S., Wei, W., Huber, N., Fan,
H.C., Metzler, K.R.C., Panagiotakos, G., Thom, N., et al. (2017). Assembly of
B Single-cell RNA sequencing data
functionally integrated human forebrain spheroids. Nature 545, 54–59.
B The unnormalized feature weights
Blankenship, A.G., and Feller, M.B. (2010). Mechanisms underlying sponta-
neous patterned activity in developing neural circuits. Nat. Rev. Neurosci.
SUPPLEMENTAL INFORMATION 11, 18–29.
Butler, A., Hoffman, P., Smibert, P., Papalexi, E., and Satija, R. (2018).
Supplemental Information can be found online at https://doi.org/10.1016/j. Integrating single-cell transcriptomic data across different conditions, technol-
stem.2019.08.002. ogies, and species. Nat. Biotechnol. 36, 411–420.
Buzsáki, G. (2004). Large-scale recording of neuronal ensembles. Nat.
Neurosci. 7, 446–451.
ACKNOWLEDGMENTS
Buzsáki, G., and Draguhn, A. (2004). Neuronal oscillations in cortical networks.
This work was supported by grants from the California Institute for Science 304, 1926–1929.
Regenerative Medicine (CIRM) DISC1-08825 and DISC2-09649; the NIH Buzsáki, G., Anastassiou, C.A., and Koch, C. (2012). The origin of extracellular
through R01MH108528, R01MH094753, R01MH109885, R01MH100175, fields and currents–EEG, ECoG, LFP and spikes. Nat. Rev. Neurosci. 13,
and R56MH109587; a SFARI grant no. 345469; and a NARSAD Independent 407–420.
Investigator Grant to A.R.M. B.V. is supported by a Sloan Research Fellow- Buzsáki, G., Logothetis, N., and Singer, W. (2013). Scaling brain size, keeping
ship, the Whitehall Foundation (2017-12-73), and the National Science Foun- timing: evolutionary preservation of brain rhythms. Neuron 80, 751–764.
dation (1736028). I.A.C. is a San Diego IRACDA Fellow supported by National Camp, J.G., Badsha, F., Florio, M., Kanton, S., Gerber, T., Wilsch-Bra €uninger,
Institutes of Health NIH/NIGMS K12 GM068524 Award. A.R.M. is supported M., Lewitus, E., Sykes, A., Hevers, W., Lancaster, M., et al. (2015). Human ce-
by U19MH1073671, part of the National Cooperative Reprogrammed Cell rebral organoids recapitulate gene expression programs of fetal neocortex
Research Groups (NCRCRG) to Study Mental Illness. R.G. is supported by development. Proc. Natl. Acad. Sci. USA 112, 15672–15677.
the Natural Sciences and Engineering Research Council of Canada (NSERC
Cederquist, G.Y., Asciolla, J.J., Tchieu, J., Walsh, R.M., Cornacchia, D., Resh,
PGS-D), UCSD Kavli Innovative Research Grant (IRG), Frontiers for Innovation
M.D., and Studer, L. (2019). Specification of positional identity in forebrain or-
Scholars Program, and Katzin Prize. We thank Patrick S. Cooper for his help on
ganoids. Nat. Biotechnol. 37, 436–444.
the Australian EEG Database. We thank Dr. V. Taupin for her expert assistance
with electron microscopy and Dr. Bruce Barshop and Dr. Jon Gangoiti (UCSD) de Hemptinne, C., Swann, N.C., Ostrem, J.L., Ryapolova-Webb, E.S., San
for their expertise in mass spectrometry. We also thank the Epigenomics Cen- Luciano, M., Galifianakis, N.B., and Starr, P.A. (2015). Therapeutic deep brain
ter (UCSD). This work was supported by the UC San Diego School of Medicine. stimulation reduces cortical phase-amplitude coupling in Parkinson’s disease.
Nat. Neurosci. 18, 779–786.
Farahany, N.A., Greely, H.T., Hyman, S., Koch, C., Grady, C., Pașca, S.P.,
AUTHOR CONTRIBUTIONS Sestan, N., Arlotta, P., Bernat, J.L., Ting, J., et al. (2018). The ethics of exper-
imenting with human brain tissue. Nature 556, 429–432.
C.A.T., R.G., and P.D.N. should be considered co-first authors, as each de-
signed the experiments and conducted the analyses with input from A.R.M. Fries, P. (2005). A mechanism for cognitive dynamics: neuronal communica-
and B.V.; C.A.T. and P.D.N. generated and characterized the cortical organo- tion through neuronal coherence. Trends Cogn. Sci. 9, 474–480.
ids and performed the MEA recordings. C.A.T., J.G., J.B., S.P., A.W., I.A.C., A. Gao, R., Peterson, E.J., and Voytek, B. (2017). Inferring synaptic excitation/in-
Domissy, and G.W.Y. performed and analyzed 103 Genomics single-cell ex- hibition balance from field potentials. Neuroimage 158, 70–78.
periments. C.A.T., M.V., and A. Devor performed functional experiments. W.W. Gertsman, I., Gangoiti, J.A., and Barshop, B.A. (2014). Validation of a dual
and G.G.H. performed whole-cell patch-clamp electrophysiology analysis. LC-HRMS platform for clinical metabolic diagnosis in serum, bridging quanti-
C.A.T. and P.D.N. analyzed the MEA data, and R.G. performed the custom tative analysis and untargeted metabolomics. Metabolomics 10, 312–323.
MEA and EEG analyses. A.R.M. and B.V. should be considered co-senior au-
Giandomenico, S.L., Mierau, S.B., Gibbons, G.M., Wenger, L.M.D., Masullo,
thors, as they contributed equally to directing the overall study design, with
L., Sit, T., Sutcliffe, M., Boulanger, J., Tripodi, M., Derivery, E., et al. (2019).
A.R.M. leading the cortical organoid development and analyses and B.V. lead-
Cerebral organoids at the air-liquid interface generate diverse nerve tracts
ing the electrophysiological analyses. C.A.T., P.D.N., R.G., B.V., and A.R.M.
with functional output. Nat. Neurosci. 22, 669–679.
wrote the manuscript. All authors reviewed the manuscript for publication.
Haller, M., Donoghue, T., Peterson, E., Varma, P., Sebastian, P., Gao, R., Noto,
T., Knight, R.T., Shestyuk, A., and Voytek, B. (2018). Parameterizing neural po-
DECLARATION OF INTERESTS wer spectra. bioRxiv. https://doi.org/10.1101/299859.
Henriques, J.B., and Davidson, R.J. (1991). Left frontal hypoactivation in
A.R.M. is a co-founder and has equity interest in TISMOO, a company dedi-
depression. J. Abnorm. Psychol. 100, 535–545.
cated to genetic analysis focusing on therapeutic applications customized
for autism spectrum disorder and other neurological disorders with genetic or- Johnson, M.H. (2001). Functional brain development in humans. Nat. Rev.
igins. The terms of this arrangement have been reviewed and approved by the Neurosci. 2, 475–483.
University of California, San Diego in accordance with its conflict of interest Kelava, I., and Lancaster, M.A. (2016). Stem cell models of human brain devel-
policies. opment. Cell Stem Cell 18, 736–748.

Khan, S., Gramfort, A., Shetty, N.R., Kitzbichler, M.G., Ganesan, S., Moran, Quiroga, R.Q., Reddy, L., Kreiman, G., Koch, C., and Fried, I. (2005). Invariant
J.M., Lee, S.M., Gabrieli, J.D., Tager-Flusberg, H.B., Joseph, R.M., et al. visual representation by single neurons in the human brain. Nature 435,
(2013). Local and long-range functional connectivity is reduced in concert in 1102–1107.
autism spectrum disorders. Proc. Natl. Acad. Sci. USA 110, 3107–3112. Renner, M., Lancaster, M.A., Bian, S., Choi, H., Ku, T., Peer, A., Chung, K., and
Khazipov, R., and Luhmann, H.J. (2006). Early patterns of electrical activity in Knoblich, J.A. (2017). Self-organized developmental patterning and differenti-
the developing cerebral cortex of humans and rodents. Trends Neurosci. 29, ation in cerebral organoids. EMBO J. 36, 1316–1329.
414–418. Schevon, C.A., Weiss, S.A., McKhann, G., Jr., Goodman, R.R., Yuste, R.,
Lancaster, M.A., and Knoblich, J.A. (2014). Generation of cerebral organoids Emerson, R.G., and Trevelyan, A.J. (2012). Evidence of an inhibitory restraint
from human pluripotent stem cells. Nat. Protoc. 9, 2329–2340. of seizure activity in humans. Nat. Commun. 3, 1060.
Lancaster, M.A., Renner, M., Martin, C.A., Wenzel, D., Bicknell, L.S., Hurles, Stevenson, N.J., Oberdorfer, L., Koolen, N., O’Toole, J.M., Werther, T.,
M.E., Homfray, T., Penninger, J.M., Jackson, A.P., and Knoblich, J.A. (2013). Klebermass-Schrehof, K., and Vanhatalo, S. (2017). Functional maturation in
Cerebral organoids model human brain development and microcephaly. preterm infants measured by serial recording of cortical activity. Sci. Rep.
Nature 501, 373–379. 7, 12969.
Lisman, J.E. (1997). Bursts as a unit of neural information: making unreliable Stuart, T., Butler, A., Hoffman, P., Hafemeister, C., Papalexi, E., Mauck, W.M.,
synapses reliable. Trends Neurosci. 20, 38–43. 3rd, Hao, Y., Stoeckius, M., Smibert, P., and Satija, R. (2019). Comprehensive
Luo, C., Lancaster, M.A., Castanon, R., Nery, J.R., Knoblich, J.A., and Ecker, integration of single-cell data. Cell 177, 1888–1902.e21.
J.R. (2016). Cerebral organoids recapitulate epigenomic signatures of the hu- Tang, X., Kim, J., Zhou, L., Wengert, E., Zhang, L., Wu, Z., Carromeu, C.,
man fetal brain. Cell Rep. 17, 3369–3384. Muotri, A.R., Marchetto, M.C.N., Gage, F.H., and Chen, G. (2016). KCC2
Manning, J.R., Jacobs, J., Fried, I., and Kahana, M.J. (2009). Broadband shifts rescues functional deficits in human neurons derived from patients with Rett
in local field potential power spectra are correlated with single-neuron spiking syndrome. Proc. Natl. Acad. Sci. USA 113, 751–756.
in humans. J. Neurosci. 29, 13613–13620. Tetzlaff, C., Okujeni, S., Egert, U., Wörgötter, F., and Butz, M. (2010). Self-
Mariani, J., Simonini, M.V., Palejev, D., Tomasini, L., Coppola, G., Szekely, organized criticality in developing neuronal networks. PLoS Comput. Biol. 6,
A.M., Horvath, T.L., and Vaccarino, F.M. (2012). Modeling human cortical e1001013.
development in vitro using induced pluripotent stem cells. Proc. Natl. Acad. Thomas, C.A., Tejwani, L., Trujillo, C.A., Negraes, P.D., Herai, R.H., Mesci, P.,
Sci. USA 109, 12770–12775. Macia, A., Crow, Y.J., and Muotri, A.R. (2017). Modeling of TREX1-dependent
McInnes, L., Healy, J., and Melville, J. (2018). UMAP: uniform manifold approx- autoimmune disease using human stem cells highlights L1 accumulation as a
imation and projection for dimension reduction. arXiv arXiv:1802.03426v2. source of neuroinflammation. Cell Stem Cell 21, 319–331.e8.
Miller, K.J., Leuthardt, E.C., Schalk, G., Rao, R.P., Anderson, N.R., Moran, Tolonen, M., Palva, J.M., Andersson, S., and Vanhatalo, S. (2007).
D.W., Miller, J.W., and Ojemann, J.G. (2007). Spectral changes in cortical sur- Development of the spontaneous activity transients and ongoing cortical activ-
face potentials during motor movement. J. Neurosci. 27, 2424–2432. ity in human preterm babies. Neuroscience 145, 997–1006.
Moore, S.M., Seidman, J.S., Ellegood, J., Gao, R., Savchenko, A., Troutman, Tort, A.B.L., Komorowski, R., Eichenbaum, H., and Kopell, N. (2010).
T.D., Abe, Y., Stender, J., Lee, D., Wang, S., et al. (2019). Setd5 haploinsuffi- Measuring phase-amplitude coupling between neuronal oscillations of
ciency alters neuronal network connectivity and leads to autistic-like behaviors different frequencies. J. Neurophysiol. 104, 1195–1210.
in mice. Transl. Psychiatry 9, 24. Uhlhaas, P.J., and Singer, W. (2010). Abnormal neural oscillations and syn-
Mukamel, R., Gelbard, H., Arieli, A., Hasson, U., Fried, I., and Malach, R. chrony in schizophrenia. Nat. Rev. Neurosci. 11, 100–113.
(2005). Coupling between neuronal firing, field potentials, and FMRI in human Uhlhaas, P.J., Roux, F., Rodriguez, E., Rotarska-Jagiela, A., and Singer, W.
auditory cortex. Science 309, 951–954. (2010). Neural synchrony and the development of cortical networks. Trends
Murrell, P. (2005). R Graphics (Chapman & Hall/CRC Press). Cogn. Sci. 14, 72–80.
Nageshappa, S., Carromeu, C., Trujillo, C.A., Mesci, P., Espuny-Camacho, I., Uylings, H.B.M., Delalle, I., Petanjek, Z., and Koenderink, M.J.T. (2002).
Pasciuto, E., Vanderhaeghen, P., Verfaillie, C.M., Raitano, S., Kumar, A., et al. Structural and immunocytochemical differentiation of neurons in prenatal
(2016). Altered neuronal network and rescue in a human MECP2 duplication and postnatal human prefrontal cortex. Neuroembryology 1, 176–186.
model. Mol. Psychiatry 21, 178–188. van de Leemput, J., Boles, N.C., Kiehl, T.R., Corneo, B., Lederman, P.,
Opitz, T., De Lima, A.D., and Voigt, T. (2002). Spontaneous development of Menon, V., Lee, C., Martinez, R.A., Levi, B.P., Thompson, C.L., et al.
synchronous oscillatory activity during maturation of cortical networks (2014). CORTECON: a temporal transcriptome analysis of in vitro human
in vitro. J. Neurophysiol. 88, 2196–2206. cerebral cortex development from human embryonic stem cells. Neuron
Pașca, S.P. (2018). The rise of three-dimensional human brain cultures. Nature 83, 51–68.
553, 437–445. Van Hove, J.L.K., and Thomas, J.A. (2014). Disorders of glycine, serine,
Paşca, A.M., Sloan, S.A., Clarke, L.E., Tian, Y., Makinson, C.D., Huber, N., GABA, and proline metabolism. In Physician’s Guide to the Diagnosis,
Kim, C.H., Park, J.Y., O’Rourke, N.A., Nguyen, K.D., et al. (2015). Functional Treatment, and Follow-Up of Inherited Metabolic Diseases, N. Blau,
cortical neurons and astrocytes from human pluripotent stem cells in 3D cul- M. Duran, K.M. Gibson, and C. Dionisi Vici, eds. (Springer Berlin
ture. Nat. Methods 12, 671–678. Heidelberg), pp. 63–83.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Voytek, B., and Knight, R.T. (2015). Dynamic network communication as a
Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., et al. (2011). Scikit-learn: unifying neural basis for cognition, development, aging, and disease. Biol.
machine learning in Python. J. Mach. Learn. Res. 12, 2825–2830. Psychiatry 77, 1089–1097.
Power, J.D., Fair, D.A., Schlaggar, B.L., and Petersen, S.E. (2010). The devel- Voytek, B., Kayser, A.S., Badre, D., Fegen, D., Chang, E.F., Crone, N.E.,
opment of human functional brain networks. Neuron 67, 735–748. Parvizi, J., Knight, R.T., and D’Esposito, M. (2015). Oscillatory dynamics coor-
Qian, X., Nguyen, H.N., Song, M.M., Hadiono, C., Ogden, S.C., Hammack, C., dinating human frontal networks in support of goal maintenance. Nat.
Yao, B., Hamersky, G.R., Jacob, F., Zhong, C., et al. (2016). Brain-region-spe- Neurosci. 18, 1318–1324.
cific organoids using mini-bioreactors for modeling ZIKV exposure. Cell 165, Wickham, H. (2016). ggplot2 (Springer).
1238–1254. Xiang, Y., Tanaka, Y., Patterson, B., Kang, Y.J., Govindaiah, G.,
Quadrato, G., Nguyen, T., Macosko, E.Z., Sherwood, J.L., Min Yang, S., Roselaar, N., Cakir, B., Kim, K.Y., Lombroso, A.P., Hwang, S.M., et al.
Berger, D.R., Maria, N., Scholvin, J., Goldman, M., Kinney, J.P., et al. (2017). (2017). Fusion of regionally specified hPSC-derived organoids models
Cell diversity and network dynamics in photosensitive human brain organoids. human brain development and interneuron migration. Cell Stem Cell
Nature 545, 48–53. 21, 383–398.e7.

Xiang, Y., Tanaka, Y., Cakir, B., Patterson, B., Kim, K.-Y., Sun, P., Kang, Y.-J., Zerbino, D.R., Achuthan, P., Akanni, W., Amode, M.R., Barrell, D., Bhai, J.,
Zhong, M., Liu, X., Patra, P., et al. (2019). hESC-derived thalamic organoids Billis, K., Cummins, C., Gall, A., Girón, C.G., et al. (2018). Ensembl 2018.
form reciprocal projections when fused with cortical organoids. Cell Stem Nucleic Acids Res. 46 (D1), D754–D761.
Cell 24, 487–497.e7. Zheng, G.X., Terry, J.M., Belgrader, P., Ryvkin, P., Bent, Z.W., Wilson, R.,
Yoon, S.-J., Elahi, L.S., Pașca, A.M., Marton, R.M., Gordon, A., Revah, O., Ziraldo, S.B., Wheeler, T.D., McDermott, G.P., Zhu, J., et al. (2017).
Miura, Y., Walczak, E.M., Holdgate, G.M., Fan, H.C., et al. (2019). Reliability Massively parallel digital transcriptional profiling of single cells. Nat.
of human cortical organoid generation. Nat. Methods 16, 75–78. Commun. 8, 14049.

STAR+METHODS
KEY RESOURCES TABLE
REAGENT or RESOURCE SOURCE IDENTIFIER

Antibodies
Mouse anti-Nestin Abcam ab22035
Rat anti-CTIP2 Abcam ab18465
Rabbit anti-SATB2 Abcam ab34735
Chicken anti-MAP2 Abcam ab5392
Rabbit anti-Ki67 Abcam ab15580
Rabbit anti-TBR1 Abcam ab31940
Rabbit anti-TBR2 Abcam ab23345
Rabbit anti-beta-catenin Abcam E247
Mouse anti-GABA Abcam ab86186
Mouse anti-GABA B Receptor 1 Abcam ab55051
Rabbit anti-Calretinin Abcam ab92341
Rabbit anti-TTF1 (NKX2.1) Abcam ab76013
Rabbit anti-Synapsin1 Millipore AB1543P
Mouse anti-NeuN Millipore MAB377
Mouse anti-Parvalbumin Millipore MAB1572
Rat anti-Somatostatin Millipore MAB354
Rabbit anti-SOX2 Cell Signaling 2748
Rabbit anti-GFAP DAKO Z033429
Donkey anti-Mouse IgG- Alexa Fluor 488 Thermo Fisher R37114
Donkey anti-Rabbit IgG- Alexa Fluor 488 Thermo Fisher R37118
Donkey anti-Rat IgG- Alexa Fluor 488 Thermo Fisher A21208
Goat anti-Mouse IgM- Alexa Fluor 488 Thermo Fisher A21042
Goat anti-Chicken IgY- Alexa Fluor 488 Thermo Fisher A11039
Donkey anti-Rabbit IgG- Alexa Fluor 555 Thermo Fisher A31572
Donkey anti-Mouse IgG- Alexa Fluor 555 Thermo Fisher A31570
Donkey anti-Mouse IgG- Alexa Fluor 647 Thermo Fisher A31571
Donkey anti-Rabbit IgG- Alexa Fluor 647 Thermo Fisher A31573
Goat anti-Mouse IgM- Alexa Fluor 647 Thermo Fisher A21238
Goat anti-Chicken IgY- Alexa Fluor 647 Thermo Fisher A21449
Chemicals, Peptides, and Recombinant Proteins
Dorsomorphin R&D Systems 3093
Stemolecule SB431542 StemGent 04-0010-10
ROCK inhibitor (Ri) Y-27632 dihydrochloride Sigma-Aldrich 688000
FGF-Basic (AA 1-155) Recombinant Human Protein Life Technologies PHG0263
Animal-Free Recombinant Human EGF Peprotech AF-100-15
Recombinant Human BDNF Peprotech 450-02
Recombinant Human GDNF Peprotech 450-10
Recombinant Human NT-3 Peprotech 450-03
L-Ascorbic acid Sigma-Aldrich A4403
N6,20 -O-Dibutyryladenosine 30 ,50 -cyclic Sigma-Aldrich D0627
monophosphate sodium salt
Bicuculline methiodide Tocris 2503
Picrotoxin Tocris 1128
Muscimol Tocris 0289/1
(Continued on next page)
e1 Cell Stem Cell 25, 558–569.e1–e7, October 3, 2019

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
CNQX disodium salt Tocris 1045
D-AP5 Tocris 0106
(R)-Baclofen Tocris 0796/10
Tetrodotoxin citrate Tocris 1069
Critical Commercial Assays
Chromium Single Cell 30 v2 Library kit 10X Genomics PN-120237
MyOne Silane Beads Thermo Fisher 37002D
SPRIselect Reagent Kit Beckman Coulter B23317
Qubit dsDNA HS Assay Kit Thermo Fisher Q32854
Deposited Data
Single-cell RNA sequencing data This paper NCBI GEO: GSE130238
Experimental Models: Cell Lines
Control iPSCs N/A N/A
Experimental Models: Organisms/Strains
C57 black 6 The Jackson Laboratory C57BL6
Software and Algorithms
GraphPad Prism GraphPad Software N/A
AxIS Software Axion Biosystems N/A
Neural Metrics Tool Axion Biosystems N/A
MATLAB MathWorks N/A
Cell Ranger software version 2.1.1 10X Genomics N/A
LEAD CONTACT AND MATERIALS AVAILABILITY
This study did not generate new unique reagents.

Further information and requests should be addressed to and will be fulfilled by the Lead Contact, Alysson R. Muotri (muotri@
ucsd.edu).
EXPERIMENTAL MODEL AND SUBJECT DETAILS
Human cell source

iPSC lines derived from control individuals have been previously characterized elsewhere (Nageshappa et al., 2016; Tang et al.,
2016). iPSC colonies were expanded on Matrigel-coated dishes (BD Biosciences, San Jose, CA, USA) with mTeSR1 medium
(StemCell Technologies, Vancouver, Canada). The cells were routinely checked by karyotype and CNV arrays to avoid genomic
alterations in the culture. Embryonic samples were obtained from fetus brains and cultured in Neurobasal (Life Technologies,
Carlsbad, CA, USA) supplemented with GlutaMAX (Life Technologies), 1% Gem21 NeuroPlex (Gemini Bio-Products, West Sacra-
mento, CA, USA), 1% MEM nonessential amino acids (NEAA; Life Technologies), 1% penicillin/streptomycin (PS; Life Technologies).
The study was approved by the University of California San Diego IRB/ESCRO committee (protocol 141223ZF).
Rodent cell source

Newborn mouse primary culture was performed as described elsewhere (Moore et al., 2019). The cells were maintained in
Neurobasal medium with GlutaMAX, 1% Gem21 NeuroPlex, 1% NEAA and 1% PS. The study was approved by the University of
California San Diego IACUC committee (protocol S09005).
METHOD DETAILS
Generation of cortical organoids

Feeder-free iPSCs were fed daily with mTeSR1 for 7 days. Colonies were dissociated using Accutase (Life Technologies) in PBS (1:1)
for 10 minutes at 37 C and centrifuged for 3 minutes at 150 x g. The cell pellet was resuspended in mTeSR1 supplemented with 10 mM
SB431542 (SB; Stemgent, Cambridge, MA, USA) and 1 mM Dorsomorphin (Dorso; R&D Systems, Minneapolis, MN, USA).
Approximately 4 3 106 cells were transferred to one well of a 6-well plate and kept in suspension under rotation (95 rpm) in the
Cell Stem Cell 25, 558–569.e1–e7, October 3, 2019 e2

presence of 5 mM ROCK inhibitor (Y-27632; Calbiochem, Sigma-Aldrich, St. Louis, MO, USA) for 24 hours to form free-floating
spheres. After 3 days, mTeSR1 was substituted by Media1 [Neurobasal (Life Technologies) supplemented with GlutaMAX, 1%
Gem21 NeuroPlex (Gemini Bio-Products), 1% N2 NeuroPlex (Gemini Bio-Products), 1% NEAA (Life Technologies), 1% PS (Life
Technologies), 10 mM SB and 1 mM Dorso] for 7 days. Then, the cells were maintained in Media2 [Neurobasal with GlutaMAX, 1%
Gem21 NeuroPlex, 1% NEAA and 1% PS] supplemented with 20 ng/mL FGF2 (Life Technologies) for 7 days, followed by 7 additional
days in Media2 supplemented with 20 ng/mL of FGF2 and 20 ng/mL EGF (PeproTech, Rocky Hill, NJ, USA). Next, cells were
transferred to Media3 [Media2 supplemented with 10 ng/mL of BDNF, 10 ng/mL of GDNF, 10 ng/mL of NT-3 (all from PeproTech),
200 mM L-ascorbic acid and 1 mM dibutyryl-cAMP (Sigma-Aldrich) to promote maturation, gliogenesis and activity]. After 7 days,
cortical organoids were maintained in Media2 for as long as needed, with media changes every 3-4 days.
Neurosphere generation
The neurosphere generation protocol was published elsewhere (Nageshappa et al., 2016). Briefly, iPSC were dissociated using
Accutase (Life Technologies), centrifuged and resuspended in medium (IMDM medium, 15% fetal bovine serum, 2 mM L-glutamine,
1% NEAA, 1 mM sodium pyruvate, 100 U PS, 200 mg/mL iron-saturated transferrin, 10 mM b-mercaptoethanol, 50 mg/mL ascorbic
acid; supplemented with 10 mM SB (Stemgent) and 1 mM Dorso (R&D Systems) on a ‘‘low-attachment’’ plate for embryoid body (EB)
formation. After 8 days, the EBs were plated for rosette formation and expansion of neural progenitors in the presence of defined
medium DMEM/F-12 supplemented with Gem21 NeuroPlex (Gemini Bio-Products) and 20 ng/mL of FGF2. For neurosphere
generation, 4,000 neural progenitors were seeded on ‘‘low-attachment’’ plate under rotation with no FGF2. The neurospheres
were developed for around 8 weeks prior to MEA plating.
Mycoplasma testing
All cellular cultures were routinely tested for mycoplasma by PCR. Media supernatants (with no antibiotics) were collected,
centrifuged, and resuspended in saline buffer. Ten microliters of each sample were used for a PCR with the following primers:
Forward: GGCGAATGGGTGAGTAAC; Reverse: CGGATAACGCTTGCGACCT. Only negative samples were used in the study.
Immunofluorescence staining
Cortical organoids were fixed with 4% paraformaldehyde overnight at 4 C and then transferred to 30% sucrose. After the 3D struc-
tures sink, they were embedded in O.C.T. (Sakura, Tokyo, Japan) and sliced in a cryostat (20 mm slices). Following air dry, the slides
containing the sliced samples were permeabilized/blocked with 0.1% Triton X-100 and 3% FBS in PBS for 2 hours at room temper-
ature, and incubated with primary antibodies overnight at 4 C. Primary antibodies used in this study were: mouse anti-Nestin, Abcam
(Cambridge, UK) ab22035, 1:250; rat anti-CTIP2, Abcam ab18465, 1:500; rabbit anti-SATB2, Abcam ab34735, 1:200; chicken anti-
MAP2, Abcam ab5392, 1:2000; rabbit anti-Synapsin1, EMD-Millipore AB1543P, 1:500; mouse anti-NeuN, EMD-Millipore MAB377,
1:500; rabbit anti-Ki67, Abcam ab15580, 1:1000; rabbit anti-SOX2, Cell Signaling Technology 2748, 1:500; rabbit anti-GFAP, DAKO
Z033429, 1:1000; rabbit anti-TBR1, Abcam ab31940, 1:500; rabbit anti-TBR2, Abcam ab23345, 1:500; rabbit anti-beta-catenin,
Abcam E247, 1:200; mouse anti-GABA, Abcam ab86186, 1:200; mouse anti-GABA B Receptor 1, Abcam ab55051, 1:100; mouse
anti-Parvalbumin, Millipore MAB1572, 1:500; rabbit anti-Calretinin, Abcam ab92341, 1:200; rat anti-Somatostatin, Millipore
MAB354, 1:100; rabbit anti-TTF1 (NKX2.1), Abcam ab76013, 1:200. Next, the slices were washed with PBS and incubated with
secondary antibodies (Alexa Fluor 488-, 555- and 647-conjugated antibodies, Life Technologies, 1:1000) for 2 hours at room
temperature. The nuclei were stained using DAPI solution (1 mg/mL). The slides were mounted using ProLong Gold antifade reagent
and analyzed under a fluorescence microscope (Axio Observer Apotome, Zeiss).
Electron microscopy (EM)

EM was performed at the CMM Electron Microscopy Facility at University of California San Diego. Four-month organoids were
immersed in modified Karnovsky’s fixative (2.5% glutaraldehyde and 2% paraformaldehyde in 0.15 M sodium cacodylate buffer,
pH 7.4) for at least 4 hours, post fixed in 1% osmium tetroxide in 0.15 M cacodylate buffer for 1 hour and stained in 2% uranyl acetate
for 1 hour. Samples were dehydrated in ethanol, embedded in Durcupan epoxy resin (Sigma-Aldrich), sectioned at 50 to 60 nm on a
Leica Ultracut UCT (Leica, Bannockburn, IL), and transfered onto Formvar and carbon-coated copper grids. Sections were stained
with 2% uranyl acetate for 5 minutes and Sato’s lead stain for 1 minute. Grids were analyzed using a JEOL 1200EX II (JEOL, Peabody,
MA) transmission electron microscope equipped with a Gatan digital camera (Gatan, Pleasanton, CA).
10X genomics single-cell and analysis

Cortical organoids were manually dissociated and sorted for single-cell RNA-seq analysis on the same day. Dissociated cells were
first placed on ice, diluted in 2-5 mL of cell media and treated with flavopiridol (5 mM, Sigma-Aldrich) to arrest transcriptional activity.
Cells were stained with the DNA-dyes DAPI (300 nM, Invitrogen) and DRAQ5 (2.5 mM, Thermo-Fischer), and incubated on ice for
10 minutes prior to sorting. Cells were sorted using a SH800 sorter (Sony) into 50 mL of cell media using a gating strategy that first
isolated large, cell-sized particles and then sorted based on viability. Sorted cells were pelleted (3 min, 100 x g, 4 C) and resuspended
in 40 mL of fresh cell media. Cell concentration was determined and the minimum population viability threshold for downstream single
cell RNA-seq processing was set at 80%.

Single cell RNA-seq libraries were constructed using the Chromium Single Cell 30 v2 Library kit (10x Genomics, (Zheng et al., 2017)
according to manufacturer descriptions; approximately 12,000 cells were loaded per sample. Reverse transcription and other ampli-
fication steps were carried out on a T100 thermal cycler (Bio-Rad). After reverse transcription, GEMs (Gel beads in emulsion) were
lysed and cDNA was cleaned up with MyOne Silane Beads (Thermo Fisher Scientific). Single stranded cDNA was PCR-amplified for
12 cycles and purified using SPRIselect Reagent Kit (Beckman Coulter). Next, cDNA was enzymatically fragmented followed by dou-
ble size selection with SPRIselect Reagent Kit (B23317, Beckman Coulter). Subsequently, adaptors were ligated and libraries were
constructed by PCR. Another round of double size selection was performed using SPRIselect Reagent Kit to generate final libraries
with a size of 200-700bp. Final libraries were quantified using Qubit dsDNA HS Assay Kit (Thermo Fisher Scientific) and size
distribution was measured using Tapestation (High Sensitivity D1000, Agilent). Average fragment size of successful libraries was
500 bp. The libraries were loaded at a concentration of 13 pM and sequenced on a Hiseq 4000 sequencer (Illumina) with the following
parameters (Read1 26 cycles; Index 1 8 cycles; Read 2 98 cycles).
Raw sequencing data from 1, 3, 6, 10-month organoids were preprocessed with Cell Ranger software (version 2.1.1, 10X Geno-
mics, Pleasanton). Reads were aligned to hg38 human reference genome (Zerbino et al., 2018) and the feature-barcode matrix was
generated. The secondary analysis performed on the feature-barcode matrix was processed via the Seurat v2.0 package (Butler
et al., 2018). For the analysis of individual time points, all genes that were not detected in at least 5 cells and cells with less than
200 genes detected were discarded. The additional filtering was based on gene-UMI distribution and percentage of mitochondrial
reads. The filtered matrix was log-normalized and scaled to 10,000 transcripts per cell. Variable genes across the single cells were
identified with the FindVariableGenes function and unwanted sources of variation, such as UMI counts per cell, percent of mitochon-
drial reads, were regressed out with the ScaleData function. Dimension reduction of the pre-processed matrix was performed by
principal component analysis (PCA). The number of principal components was identified based on a strong enrichment of genes
with low p values, which were computed by a resampling test. This procedure was implemented with the JackStraw function in
Seurat. With the selected dimensions, cellular distance matrix was first organized into a K-nearest neighbor (KNN) graph and then
partitioned into clusters with Louvain algorithm via the FindClusters function. Finally, cells within the graph-based clusters were
co-localized on the UMAP plot (McInnes et al., 2018) of two dimensions by the RunUMAP function. Identifying top differentially
expressed genes for each cluster was performed using the FindAllMarkers function.
Datasets from the four time points were merged with the MergeSeurat function and then the merged matrix was used as an input to
the Seurat v3 anchoring procedure, which assembles datasets into an integrated reference by identifying cell pairwise correspon-
dences for single cells across different datasets. Further analysis was processed with Seurat v3.0 package (Stuart et al., 2019).
Default parameters including a dimensionality of 30 were set to run the FindIntegrationAnchors and IntegrateData function. On
the integrated datasets, clustering was performed with a resolution parameter set to be 1.0 and a dimensionality of 30 by
FindNeighbors and FindClusters. With the graph-based clustering, a total of 14 clusters were generated, which were further merged
into seven main clusters based on expression of marker genes. UMAP plots displayed by the DimPlot function were used to visualize
and explore the integrated datasets. Dot plots and UMAP plots for transcript abundance of marker genes were made using ggplot2
package (Wickham, 2016), while the barplot was created with graphics v3.5.3 package (Murrell, 2005). The dot plots show the
percentage of cells that express more than one transcript for each gene and its log-normalized expression level across main cell
clusters. Violin plots for marker gene expression across all clusters were produced with the VlnPlot function.
Mass spectrometry
Samples were assayed using an adaptation of published protocol (Gertsman et al., 2014). Cortical organoid media (100 mL) was
mixed with 2 mM 13C4-4-aminobutyric acid, as internal standard. Metabolites were extracted using 80% ice-cold methanol. After
incubation for 30 minutes at 20 C, samples were deproteinized at 4 C by centrifugation at 17,136 x g for 10 minutes. Supernatants
were evaporated to dryness in a centrifugal evaporator at 36 C (Savant SPD121P Speed-Vac concentrator. Thermo Fisher, Asheville,
NC) and reconstituted in 100 mL of 10% methanol in water + 0.1% formic acid, by means of consecutive vortexing, orbital shaking and
sonication. 5 uL of which were injected into a Sciex 4500 triple quadrupole mass spectrometer (Sciex, Foster City, California, USA) to
determine 4-aminobutyric acid (GABA) concentration. Chromatographic separations were conducted in a 3 mm ACE C18-PFP
reversed-phase HPLC column (Mac-mod analytical, Chadds Ford, PA, USA) using an Acquity binary pump (Waters, Milford, MA,
USA) equipped with an in-line degasser at 0.3 mL/min flow-rate and at 13 C, to enhance the retention of the low-retained com-
pounds, by means of a simple binary gradient of acetonitrile partitioning in 3% acetonitrile in water, both containing 0.1% formic
acid. Compounds were eluted during the first 3 minutes, then it ramped to 100%. Total run time was 45 minutes. Positive electrospray
ionization multiple reaction monitoring transitions were optimized for GABA (and 13C4-GABA), m/z 104.2 > 87 (108.2 > 90.9) and m/z
104.2 > 68.9 (108.2 > 73), using collision energies of 15 and 23, respectively, and unit mass resolution. GABA concentrations were
calculated by interpolation using an 8-point calibration curve, spanning 0.01 to 0.2 mM, constructed by supplementing medium with
the appropriate amounts of GABA. Quantification was conducted using MultiQuant 2.1 software (Sciex, Foster City, CA, USA).
Whole-cell patch-clamp
Whole-cell patch-clamp recordings were performed from cells of cortical organoids in a similar condition as for MEA recordings: 6- to
8-week cortical organoids were plated on 35 mm dishes that were previously coated with 100 mg/mL poly-L-ornithine and 10 mg/ml
laminin. Cells were fed twice a week and were maintained for 24 weeks. The extracellular solution for patch-clamp experiments con-
tained (in mM) the following: 130 NaCl, 3 KCl, 1 CaCl2, 1 MgCl2, 10 HEPES, and 10 glucose; pH 7.4 with 1 M NaOH (4 mM Na+

added). The internal solution for patch electrodes contained (in mM) the following: 138 K-gluconate, 4 KCl, 10 Na2-phosphocreatine,
0.2 CaCl2, 10 HEPES (Na+ salt), 1 EGTA, 4 Mg-ATP, 0.3 Na-GTP; pH 7.4 with 1 M KOH (3 mM K+ added). The osmolarity of all
solutions was adjusted to 290 mOsm. Electrodes for electrophysiological recording were pulled on a Flaming/Brown micropipette
puller (Model P-87, Sutter Instrument, CA, USA) from filamented borosilicate capillary glass (1.2 mm OD, 0.69 mm ID, World Precision
Instruments, FL, USA). The electrode resistances were 3–8 MU. Patch-clamp experiments were performed with an Axon CV-4 head-
stage and Axopatch 200A amplifier (Molecular Devices, CA, USA) at room temperature. Liquid junction potential correction (10 mV)
was not applied. Electrophysiology data were analyzed offline using pCLAMP 10 software (Molecular Devices, CA, USA).
Multi-electrode array (MEA) recording

6-week cortical organoids were plated per well in 12-well MEA plates (Axion Biosystems, Atlanta, GA, USA). Each well contains 64
low-impedance (0.04 MU) platinum microelectrodes with 30 mm of diameter spaced by 200 mm, yielding a total of 512 channels (8
wells containing organoids and 4 internal control). The plate was previously coated with 100 mg/mL poly-L-ornithine and 10 mg/
mL laminin, and we performed four independent experiments in duplicates. Cells were fed twice a week and measurements were
collected 24 hours after the medium was changed, once a week, starting at two weeks after plating (8 weeks of organoid
differentiation). Recordings were performed using a Maestro MEA system and AxIS Software Spontaneous Neural Configuration
(Axion Biosystems) with a customized script for band-pass filter (0.1-Hz and 5-kHz cutoff frequencies). Spikes were detected with
AxIS software using an adaptive threshold crossing set to 5.5 times the standard deviation of the estimated noise for each electrode
(channel). The plate was first allowed to rest for 3 minutes in the Maestro device, and then 4 minutes of data were recorded. For the
MEA analysis, the electrodes that detected at least 5 spikes/min were classified as active electrodes using Axion Biosystems’ Neural
Metrics Tool. Bursts were identified in the data recorded from each individual electrode using an inter-spike interval (ISI) threshold
requiring a minimum number of 5 spikes with a maximum ISI of 100 ms. A minimum of 10 spikes under the same ISI with a minimum of
25% active electrodes were required for network bursts in the well. The synchrony index was calculated using a cross-correlogram
synchrony window of 20 ms. Bright-field images were captured to assess for cell density and electrode coverage.
Custom MEA analysis

Custom MEA analysis and developmental time regression model can be found in: https://github.com/voytekresearch/
OscillatoryOrganoids. Raw MEA recordings were converted to .mat files using Axion-provided functions and analyzed offline using
custom MATLAB functions and scripts. Local field potential signals (LFP) from each of the 64 electrodes were generated by low-pass
filtering (FIR filter) and downsampling raw signals from 12,500 Hz to 1,000 Hz (resample.m). Multi-unit spikes were detected as
follows: each channel was first referenced to the well median for every time point, similar to a common average reference (64 chan-
nels). The median was used instead of the mean to avoid biasing the reference during high firing rate periods. Next, the re-referenced
signal was bandpass filtered for 300-3,000 Hz with a 3rd-order Butterworth filter (butter.m). The adaptative spike threshold was set to
be 5.5 standard deviations, where the standard deviation was estimated from the median as previously described (Quiroga et al.,
2005) to avoid biasing the threshold for channels with high firing rates (thus an artificially high threshold). Spike timestamps were
taken as the peak time after the absolute value of the signal crossed the threshold, but at least 1 ms from another spike (findpeaks.m).
Spike timestamps were then converted into binary vectors (1 ms bin size), summed across 64 channels, and smoothed (conv.m) with
a normalized 100-point (0.1 s) Gaussian window (gausswin.m) to create a population spiking vector for each MEA well. Note that
spikes from each channel do not represent putative single-unit action potentials, as the spatial resolution of MEA electrodes were
too sparse. Multi-unit spiking was not sorted since total population spiking (of well) was submitted for further analysis, rather than
individual spike trains.
Network event analysis

A network event was detected when population spiking was i) greater than 80% of the maximum spiking value over the length of the
recording; ii) at least 1 spike/s; and iii) 1 s away from any other network events. The first peak after all 3 criteria was satisfied was
marked as t = 0, and the window of data from 0.5 s before to 2.5 s after the peak was collected as the network event, as events almost
always subsided 2.5 s after onset by both algorithmic detection and visual inspection. Nearly all spiking channels experienced a
significant firing rate increase during network events. LFP data from all 64 channels from the same time frame were also collected
for analysis. All events from different MEA wells obtained on the same recording day were aggregated for statistical analysis and
plotting. Subpeaks within an event were identified using findpeaks.m, where a subpeak must satisfy the following: i) peak height
of at least 25% of the first peak; ii) peak width of at least 50 ms; iii) at least 200 ms away from the previous peak; and iv) peak prom-
inence of 1 over Peak 1 height. Subpeak time and the height relative to the initial peak were recorded. The inter-event interval coef-
ficient of variation (IEI CV) was calculated as the standard deviation of the inter-event interval divided by its mean, where IEI is the time
between consecutive network events within the same MEA well. Event temporal correlation was calculated as the mean Pearson
correlation coefficient of population spiking vector between each pair of network event in the same MEA well across a single
recording session. Event spatial correlation was calculated as the mean Pearson correlation coefficient between all pairs of 64
LFP channels during each 3 s network event.

Oscillatory spectral power analysis
Power spectral density (PSD) estimates were computed using Welch’s method (pwelch.m), with a window length of 2 s and overlap of
1 s. Oscillatory power was defined as peaks in the PSD above the aperiodic 1/f power law decay. Thus, for each channel, a straight
line was fit over the PSD in double-log space between 0.5-20 Hz using robust fit (robustfit.m), and oscillatory power was computed as
the difference between the mean log PSD power and the mean log fitted power (baseline), over 2.5-4.5 Hz. This method accounts for
non-oscillatory changes, such as slow transients or the aperiodic 1/f background component, whereas standard wavelet filtering
methods will confound the two (Haller et al., 2018).
Phase Amplitude Coupling (PAC)

LFP data from all 64 channels of each well was first lowpass/bandpass filtered (eegfilt.m, EEGLAB) for delta (0-4 Hz) and high-fre-
quency, broadband (100-400 Hz) activity, sometimes referred to as high gamma. Delta phase was extracted by taking the phase
angle of the bandpassed delta signal Hilbert transform (hilbert.m, angle.m), while gamma power was extracted by taking the squared
magnitude of the filtered gamma. Gamma power was smoothed with the same delta-band filter for display purposes, but not for
subsequent analysis. Note that analysis was performed for 100-200 Hz and 200-400 Hz separately, as LFP spectrum follows an in-
verse power law (1/f), and grouping a wide frequency band (100-400 Hz) together would bias power estimates toward lower
frequency limits (100 Hz). To compute PAC, instantaneous delta phase was binned into 20 equidistant bins between -p and p,
and gamma power was sorted based on the corresponding delta phase at the same sample time and averaged across the same
phase bin. This procedure was performed separately for event and non-event indices, where event indices are the same 3 s windows
as described above in Network Event Analysis, while all other times are considered as non-event time points. Modulation Index was
computed as the Kullback-Leibler divergence between the sum-normalized distribution of gamma power across phase bins and a
uniform distribution (Tort et al., 2010). Figure 3C presents well-averaged MI across all 64 channels. For visualization in Figure 3B, the
binned gamma power vector for each channel was circularly shifted such that the phase of maximum gamma power was -p.
Pharmacology
The pharmacological manipulation was performed with cortical organoids plated on 4 MEA wells (n = 4, cortical organoid culture)
using the following drugs: 10 mM bicuculline, 100 mM picrotoxin, 50 mM muscimol, 20 mM CNQX, 20 mM AP5, 25 mM baclofen and
1 mM TTX. In this assessment, baseline recordings were obtained immediately before and 15 minutes after the addition of the com-
pound. Three washes with PBS for total removal of the drug were performed in washout experiments; fresh media was added and
another recording was conducted after 2 hours.
Preterm neonatal EEG

A preterm neonatal EEG dataset was obtained from a publicly available dataset (Stevenson et al., 2017). Raw recordings were not
available due to patient confidentiality concerns. The dataset includes 567 recordings from 39 preterm neonates (24-38 weeks PMA),
consisting of 23 EEG features computed from the entirety of each recording (Table S2). See cited publication for details of features.
Briefly, we chose to include features derived from duration and timing of (interval between subsequent) network events in neonates
and organoids, as these are least affected by anatomical differences between the two model systems (i.e., skull filtering), as well as
spectral features (delta, theta, and alpha power). 5%/50%/95% refer to percentile of the feature distribution from a recording.
Resampled feature-age correlation

We computed Pearson’s correlation coefficient between neonate age and each of the 12 EEG features, after a leave-K-groups-out
resampling procedure N times, where K is the number of neonates from whom all recordings were left out in computing the correlation
(50% of all neonates, resampling N = 100). An identical procedure was performed to compute the correlation between organoid cul-
ture development time and LFP features (K = 4 out of 8, 50%, N = 100). Mean and standard deviation were then computed over all
resampled draws in order to compare between organoid LFP and neonatal EEG to produce Figures 4C and S4D.
Neonate-organoid development time regression model

To compare the similarity of developmental trajectory of cortical organoids and the preterm human brain, we trained an Elastic Net
(L1- and L2- regularized) regression model on only the preterm neonatal EEG features and used that model (with all parameters held
the same) to predict an equivalent organoid development time for each recording time point over 40 weeks in culture. Specifically, the
training dataset consisted of a subset of the preterm EEG data; we discarded all ‘‘low-activity-period’’ features (Lisman, 1997) since
there was no equivalent period for organoid recordings, as well as features for which we could not sensibly compute from organoid
LFPs, such as interhemispheric synchrony. This selection was done a priori, and 12 features remained, including 3 features for rela-
tive spectral power in distinct frequency bands. The features corresponding to aspects of spontaneous activity transient (SAT) timing,
such as SATs per hour and SAT duration, were similarly computed on organoid LFPs after network event detection described earlier
(see Table S2 for a full list of included and rejected features). This latter organoid LFP test dataset was never seen by the regression
model until prediction time. Training was performed using scikit-learn linear model module [ElasticNetCV (Pedregosa et al., 2011)],
with K-Group shuffle split cross-validation on regularization hyperparameters, where K = 25% of groups, N = 200 shuffles. In other
words, we found the best regularized linear model possible for predicting the age of preterm neonates using those precomputed EEG
features. This model was directly applied on organoid LFP features to determine the corresponding development time of the

organoids during 40 weeks in culture. Control datasets were also submitted for prediction, including held-out preterm EEG (positive
control), and mouse primary culture, 2D iPSC culture, and human fetal culture (negative controls). To quantify the model’s ability to
predict the developmental trend of the out of sample datasets, we compute the Pearson correlation coefficient between the pre-
dicted and actual age of each dataset. To eliminate the potential confound of a difference in frequency-dependent filtering properties
of the skull and difference in spatial integration of currents in macroscopic EEG electrodes compared to microscopic planar MEA
electrodes, the same analysis was performed after discarding the spectral features (leaving 9 features total). This result is presented
in Figure S4E, in addition to the prediction for the control datasets.
QUANTIFICATION AND STATISTICAL ANALYSIS
Statistical analysis
Data are presented as mean ± s.e.m., unless otherwise indicated, and it was obtained from different samples. No statistical method
was used to predetermine the sample size. The statistical analyses were performed using Prism software (GraphPad, San Diego,
CA, USA). Student’s t test, Mann–Whitney-test, or ANOVA with post hoc tests were used as indicated. Significance was defined
as p < 0.05(*), p < 0.01(**), or p < 0.001(***).
Statistics and Regression for custom MEA analysis

To fit linear or quadratic models in Figures 2F, 2G, and 2I, we used organoid developmental time (in days) as input and electrophys-
iological features as output (LinearModel.fit, MATLAB). Reported R2 and p values are model statistics over the entire dataset. All
events from different MEA wells on the same recording day were aggregated as samples drawn from the same distribution.
DATA AND CODE AVAILABILITY
Single-cell RNA sequencing data

All datasets and/or analyses generated during the current study are available from the Lead Contact upon reasonable request.
Single-cell RNA sequencing data that support the findings of this study have been deposited at NCBI GEO: GSE130238.
The unnormalized feature weights

The code can be found online:
https://github.com/voytekresearch/OscillatoryOrganoids/blob/master/organoid_EEG_age_regression.ipynb.

Complex Oscillatory Waves Emerging From Cortical Organoids Model Early Human Brain Network Development

Uploaded by

Copyright:

Available Formats

Complex Oscillatory Waves Emerging From Cortical Organoids Model Early Human Brain Network Development

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Complex Oscillatory Waves Emerging From Cortical Organoids Model Early Human Brain Network Development

Uploaded by

Copyright:

Available Formats

Short Article

Complex Oscillatory Waves Emerging from Cortical

d Cortical organoids exhibit phase-amplitude coupling during

d Differential role of glutamate and GABA in initiating and

d Network-level events are similar to the human preterm

Trujillo et al., 2019, Cell Stem Cell 25, 558–569

Complex Oscillatory Waves Emerging from Cortical

La Jolla, CA 92093, USA

SUMMARY human brain. Neural oscillations, a prominent, rhythmic brain

(C) Organoid growth during different developmental stages.

560 Cell Stem Cell 25, 558–569, October 3, 2019

Cell Stem Cell 25, 558–569, October 3, 2019 561

562 Cell Stem Cell 25, 558–569, October 3, 2019

Cell Stem Cell 25, 558–569, October 3, 2019 563

564 Cell Stem Cell 25, 558–569, October 3, 2019

Cell Stem Cell 25, 558–569, October 3, 2019 565

566 Cell Stem Cell 25, 558–569, October 3, 2019

Cell Stem Cell 25, 558–569, October 3, 2019 567

568 Cell Stem Cell 25, 558–569, October 3, 2019

Cell Stem Cell 25, 558–569, October 3, 2019 569

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

e1 Cell Stem Cell 25, 558–569.e1–e7, October 3, 2019

LEAD CONTACT AND MATERIALS AVAILABILITY

This study did not generate new unique reagents.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Human cell source

Rodent cell source

Generation of cortical organoids

Cell Stem Cell 25, 558–569.e1–e7, October 3, 2019 e2

Electron microscopy (EM)

10X genomics single-cell and analysis

e3 Cell Stem Cell 25, 558–569.e1–e7, October 3, 2019

Cell Stem Cell 25, 558–569.e1–e7, October 3, 2019 e4

Multi-electrode array (MEA) recording

Custom MEA analysis

Network event analysis

e5 Cell Stem Cell 25, 558–569.e1–e7, October 3, 2019

Phase Amplitude Coupling (PAC)

Preterm neonatal EEG

Resampled feature-age correlation

Neonate-organoid development time regression model

Cell Stem Cell 25, 558–569.e1–e7, October 3, 2019 e6

QUANTIFICATION AND STATISTICAL ANALYSIS

Statistics and Regression for custom MEA analysis

DATA AND CODE AVAILABILITY

Single-cell RNA sequencing data

The unnormalized feature weights

e7 Cell Stem Cell 25, 558–569.e1–e7, October 3, 2019

You might also like