Letter
The prefrontal cortex is a critical neuroanatomical hub for controlling motivated
behaviors across mammalian species
1–3
. In addition to intra-cortical connectivity, prefrontal projection neurons innervate
subcortical structures that contribute to reward seeking behaviors, such as the ventral
striatum and midline thalamus
4
. While connectivity among these structures contributes to appetitive behaviors
5–13
, how projection-specific prefrontal neurons encode reward-relevant information to
guide reward seeking is unknown. Here we use in vivo two-photon calcium imaging to
monitor the activity of dorsomedial prefrontal neurons during an appetitive Pavlovian
conditioning task. At the population level, these neurons display diverse activity
patterns during the presentation of reward-predictive cues. Furthermore, recordings
from prefrontal neurons with resolved projection targets reveal that individual corticostriatal
neurons show response tuning to reward-predictive cues, such that excitatory cue responses
are amplified across learning. In contrast, corticothalamic neurons gradually develop
new, primarily inhibitory responses to reward-predictive cues across learning. Finally,
bidirectional optogenetic manipulation of these neurons reveals that stimulation of
corticostriatal neurons promotes conditioned reward seeking after learning, while
activity in corticothalamic neurons suppresses both the acquisition and expression
of conditioned reward seeking. These data show how prefrontal circuitry can dynamically
control reward-seeking behavior through the opposing activities of projection-specific
cell populations.
Neurons in the prefrontal cortex (PFC) respond diversely to reward-predictive cues
14–17
, although how this cue encoding fits into a broader circuitry to guide reward seeking
is unknown. To address this, we designed a Pavlovian conditioning task that allows
two-photon imaging of deep cortical tissue during behavior. Head-fixed mice were trained
to associate one conditioned stimulus (CS+), but not another (CS−), with sucrose (Figure
1a,b). Following multiple training sessions, mice behaviorally discriminated between
the cues by displaying anticipatory licks to the CS+ but not CS− (Figure 1c), confirming
that the cue-reward contingencies had been established by the late sessions (Figure
1d,e; Extended Data Fig. 1). To monitor neural activity during this task, we injected
a virus into dorsomedial PFC for delivery of a calcium indicator
18
under the control of the calcium/calmodulin dependent protein kinase II alpha (CAMK2A)
promoter (AAVdj-CaMKii-GCaMP6s; Figure 1f), which putatively targets cortical excitatory
neurons
19
. Ex vivo brain slice recordings revealed that fluorescent deflections of GCaMP6s-expressing
PFC neurons reliably tracked elevations and reductions in action potential frequency,
whereas hyperpolarization from rest alone did not influence GCaMP6s-mediated fluorescence
(Extended Data Fig 2). Next, we implanted optical cannulae ~2.2mm beneath the surface
of the brain, allowing chronic optical access to hundreds of dorsomedial PFC neurons
in each awake, behaving mouse (Figure 1g–i; Supplementary Video 1). Collectively,
we recorded from GCaMP6s-expressing PFC neurons before learning (n=1,473) and after
learning (n=1,571), and found that while many of these neurons displayed increased
activity in response to reward-predictive cues, other neurons exhibited inhibitory
cue responses (Figure 1j,k,n,o). These responses were most prevalent during presentation
of the CS+, but not the CS−, after learning (Figure 1l,m,p,q; Extended Data Fig. 3a,b;
Extended Data Fig. 4a–c). Thus, the responses of many individual PFC neurons could
be used to decode whether the CS+ or CS− was presented on any given trial after learning
(Extended Data Fig. 4d).
We found that many dorsomedial PFC neurons encoded cues after learning; however, both
inhibitory and excitatory responses were common. Considering this, it is possible
that distinct populations of neurons in the PFC have opposing cue-encoding properties.
Two major projection targets of the PFC are the nucleus accumbens (PFC-NAc) and paraventricular
nucleus of the thalamus (PFC-PVT)
4
. Furthermore, studies examining reward seeking in rodent addiction models have found
that PFC-NAc neurons express immediate early genes following cue-induced relapse
21,22
, and inhibition of these cells can prevent drug seeking
6–8,12
. Additionally, reward-predictive cues may influence the activity of PVT neurons
23
, and such activity is thought to modulate cue salience
24
and conditioned behaviors
25,26
. Thus, we hypothesized that PFC-NAc and PFC-PVT neurons encode reward-predictive
stimuli to orchestrate appetitive learning. To test this hypothesis, we first determined
whether PFC-NAc and PFC-PVT neurons are distinct or overlapping populations of neurons.
We injected retrograde cholera toxins conjugated to different fluorophores into NAc
and PVT of the same mice. Electrophysiological recordings and histological analysis
revealed that PFC-NAc and PFC-PVT neurons are physiologically distinct and anatomically
segregated within dorsomedial PFC (Extended Data Fig. 5), such that PFC-NAc neurons
are in layers II/III and V (46% layer II/III, 51% layer V, 4% layer VI) whereas PFC-PVT
neurons are in layer VI (0% layer II/III, 20% layer V, 79% layer VI). These data are
consistent with anatomical studies, as corticostriatal projection neurons reside in
layers II/III and V in rats, whereas layer VI is specific for corticothalamic neurons
27,28
. Furthermore, using a retrograde rabies-tracing strategy to sparsely label PFC-NAc
or PFC-PVT neurons as well as their afferent inputs, we found that these cell populations
have both shared and unshared afferent connectivity (Extended Data Fig 6). Finally,
PFC-NAc and PFC-PVT neurons express CaMKii (Extended Data Fig. 7a–d), collectively
revealing that these neurons make up two non-overlapping subpopulations of CaMKii-expressing
projection neurons.
We next targeted PFC-NAc and PFC-PVT neurons for in vivo two-photon calcium imaging.
Cre-inducible GCaMP6s (AAVdj-DIO-GCaMP6s) was injected into dorsomedial PFC, and in
the same surgery a retrogradely transported virus, canine adenovirus-2 encoding cre-recombinase
(Cav2-cre), was injected into either the NAc or PVT (Figure 2a,g). This resulted in
projection-specific GCaMP6s expression in PFC-NAc and PFC-PVT neurons (Figure 2b,h;
Extended Data Fig. 7e–j). Next, mice underwent Pavlovian conditioning with simultaneous
head-fixed two-photon calcium imaging. Data revealed that after learning, but not
before learning, PFC-NAc neurons primarily displayed excitation to the CS+, whereas
fewer neurons responded to the CS− (Figure 2c–e; Extended Data Fig. 3c,d). In contrast,
PFC-PVT neurons primarily displayed inhibition to the CS+ after learning, whereas
fewer neurons responded to the CS− (Figure 2i–k; Extended Data Fig. 3e,f). Finally,
we found that activity in either PFC-NAc neurons (Figure 2f) or PFC-PVT neurons (Figure
2l) could be used to decode whether the CS+ or CS− was presented on any given trial
after learning.
We found that PFC-NAc neurons and PFC-PVT neurons show distinct cue encoding profiles
after learning, but how these response properties emerge across learning was unclear.
To examine this, we identified individual PFC-NAc and PFC-PVT neurons whose activity
could reliably be monitored throughout training (Figure 3a,d). We found that many
PFC-NAc neurons initially displayed low amplitude, phasic calcium transients to both
the CS− and CS+, such that there was little cue discrimination (Figure 3b). However,
across learning the responses of individual PFC-NAc neurons became selective, such
that responses to the CS+ were amplified, whereas responses to the CS− were diminished
(Figure 3b,c). In contrast to this response tuning in PFC-NAc neurons, PFC-PVT neurons
showed little response to either the CS+ or CS− before learning, but instead many
neurons acquired new inhibitory responses after learning (Figure 3e,f). Thus, responses
of PFC-NAc, but not PFC-PVT, neurons after learning could be predicted based on responses
before learning (Figure 3c,f, correlations). Collectively, corticostriatal and corticothalamic
neurons showed distinct functional plasticity across appetitive learning, such that
PFC-NAc neurons adjusted their activity to allow cue discrimination, whereas PFC-PVT
neurons gradually acquired responses for cue discrimination.
Our data reveal that projection-specific PFC neurons show distinct cue encoding properties
across appetitive learning. However, whether this activity controls the acquisition
of conditioned reward seeking is unclear. To test this, we targeted PFC-NAc or PFC-PVT
neurons for optogenetic manipulation by injecting Cav2-cre into the NAc or PVT, and
cre-inducible channelrhodopsin-2 (AAV5-DIO-ChR2-eYFP), halorhodopsin (AAV5-DIO-eNpHR3.0-eYFP),
or control (AAV5-DIO-eYFP) into dorsomedial PFC (Figure 4a–c, j–l). Next, we optogenetically
manipulated these cells during cue delivery throughout Pavlovian conditioning (sessions
1–8; Figure 4d–f, m–o), and tested the effects on acquisition of this task during
a subsequent test without laser. Data reveal that activation (in PFC-NAc::ChR2 mice)
or inactivation (in PFC-NAc::eNpHR3.0 mice) of PFC-NAc neurons did not influence CS+-evoked
anticipatory licking during the no-laser test (Figure 4g–i), implying that these cells
do not control the acquisition of anticipatory licking. In contrast, optogenetic activation
of PFC-PVT neurons (in PFC-PVT::ChR2 mice) reduced CS+-evoked anticipatory licking
during the no-laser test, whereas optogenetic inactivation of PFC-PVT neurons (in
PFC-PVT::eNpHR mice) increased anticipatory licking during the no-laser test (Figure
4p–r). Thus, inhibitory cue encoding in PFC-PVT neurons, but not PFC-NAc neurons,
contributes to the acquisition of conditioned reward seeking.
Although activity in PFC-NAc neurons did not control the acquisition of conditioned
reward seeking, differences in licking behavior were apparent between groups during
the laser conditioning sessions (sessions 1–8; see Figure 4g,h). Thus, one possibility
is that PFC-NAc neurons control the expression of conditioned reward seeking, rather
than acquisition. To test this directly, we next performed the optogenetic manipulations
after conditioning in separate groups of mice. We found that optogenetic activation
of PFC-NAc neurons during the CS+ increased anticipatory licking (Figure 5a,d; Figure
5c,f), whereas inactivation of PFC-NAc neurons reduced anticipatory licking (Figure
5b,e; Figure 5c,f). In contrast, optogenetic activation of PFC-PVT neurons during
the CS+ reduced anticipatory licking (Figure 5g,j; Figure 5i,l) whereas inactivation
of PFC-PVT neurons had no effect (Figure 5h,k; Figure 5i,l). The lack of effect for
PFC-PVT inactivation is likely due to CS+-evoked inhibition of these cells (see Figures
2 and 3), as neuronal hyperpolarization from sub-threshold potentials should not result
in further adjustment of neuronal output. Collectively, activation of PFC-NAc neurons
and inhibition of PFC-PVT neurons supports the expression of conditioned reward seeking.
In addition to encoding the reward-predictive stimulus after learning, some PFC-NAc
and PFC-PVT neurons displayed CS− encoding (see Figure 2). Considering this, we also
evaluated how projection-specific cells influenced licking during the CS−. We found
that optogenetic perturbations of PFC-NAc and PFC-PVT neurons during the CS− did not
influence the acquisition of CS− licking (Extended Data Fig. 8a–f). In contrast, inactivation
but not activation of PFC-NAc neurons reduced the expression of CS− licking, whereas
optogenetic manipulations of PFC-PVT neurons had no effect on the expression of CS−
licking (Extended Data Fig. 8g–l). Moreover, optogenetic manipulation of PFC-NAc or
PFC-PVT neurons during random time epochs outside of cue delivery had no effect on
licking (Extended Data Fig. 9), nor did optogenetic manipulations produce a real-time
place preference or aversion (Extended Data Fig. 10). Thus, activity in PFC-NAc and
PFC-PVT neurons specifically during the cue period is critical for conditioned reward
seeking.
Here we found that the opposing population dynamics of prefrontal output circuits
orchestrates conditioned reward seeking. In addition to divergent population encoding,
subsets of cells in both corticostriatal and corticothalamic circuits have response
profiles that are opposite to the population overall (i.e., some PFC-NAc neurons have
inhibitory CS+ responses, whereas some PFC-PVT cells have excitatory CS+ responses).
Although the cause and function of such heterogeneity is unclear, it is possible that
subpopulations of cells within each group may have distinct cell-type specific targets,
afferent inputs, and spatial locations. Thus, while it remains unclear how all cells
in PFC function to orchestrate reward seeking, our data suggest that projection-specific
and anatomically segregated prefrontal neurons can have opposing activity dynamics,
plasticity profiles, and functional control of conditioned reward seeking.
Methods
Subjects and surgery
Adult male C57BL/6J mice (Jackson Laboratories) or Ai9 reporter mice (Cg-Gt(ROSA)26Sortm9(CAG-tdTomato)Hze
IJ; Jackson Laboratories) were group housed (25–35g; 6–8 weeks old) with littermates
until surgery. For all experiments, mice underwent surgery during which they were
anesthetized with 0.8–1.5% isoflurane vaporized in pure oxygen (1L/min) and placed
within a stereotactic frame (David Kopf Instruments). Ophthalmic ointment (Akorn)
and a topical anesthetic (2% Lidocaine; Akorn) were applied during surgeries, and
subcutaneous injections of sterile saline (0.9% NaCl in water) were administered to
prevent dehydration. During surgeries, virus injections were administered unilaterally
(for two-photon microscopy experiments) or bilaterally (for optogenetics or anatomical
experiments) targeting dorsal medial PFC (specifically prelimbic cortex; 500nl/side;
relative to bregma: AP, +1.85mm; ML, ±0.60mm; DV, −2.50mm), bilaterally targeting
NAc (500nl/side; relative to bregma: AP, +1.42mm; ML ±0.73mm; DV, −4.80mm), and/or
on the midline targeting PVT (300nl; relative to bregma: AP, −1.46mm; ML −1.13mm;
DV, −3.30mm; 20° angle). The UNC Vector Core packaged all viruses except canine adenovirus-2
encoding cre (Cav2-cre; Institut de Génétique Moléculaire de Montpellier, France).
For two-photon imaging experiments, an optical cannula (Inscopix, CA) was implanted
above the PFC injection site (relative to bregma: AP, +1.85mm; ML −0.8mm; DV, −2.2mm;
see Resendez et al., 2016 for details using similar surgical protocols for imaging
experiments)
29
. For optogenetic experiments, custom-made optical fibers
30
were implanted bilaterally ~0.5mm above the PFC injection sites (relative to bregma:
AP, +1.85mm; ML ±0.83mm; DV, −1.93mm; 10° angle). For experiments involving head-fixed
behavior, a custom-made ring (stainless steel; 5mm ID, 11mm OD) was attached to the
skull during surgery to allow head fixation (see Figure 1a). Following surgeries,
mice received acetaminophen in their drinking water for two days, and were allowed
to recover with access to food and water ad libitum for at least 21 days. After recovery,
mice were water restricted (water bottles taken out of the cage), and 0.6mL of water
was delivered every day to a dish placed within each home cage. Behavioral experiments
began when mice weighed less than 90% of free drinking weight (~10 days for all experiments).
To ensure good health and weight maintenance, mice were weighed and handled daily.
This protocol resulted in weight stabilization between 85–90% of free-drinking weight
during each experiment. No mouse was given more or less than 0.6mL of water for weight
concerns during water restriction procedures, nor did any health problems related
to dehydration arise at any point from these protocols. All experimental procedures
were performed in accordance with the Guide for the Care and Use of Laboratory Animals
(National Institutes of Health), and were approved by the Institutional Animal Care
and Use Committee at the University of North Carolina a priori.
Head-fixed behavior
Following recovery from surgery, mice were habituated to head fixation for 3 days,
during which unpredictable drops of sucrose (10% sucrose in water; 2.0–2.5µL) were
delivered intermittently for one hour (~60 drops/hour) through a gravity-driven, solenoid-controlled
lick tube. Once the mice displayed sufficient licking (>1000 licks per session), they
underwent Pavlovian conditioning. During each conditioning session, two cues (3kHz
pulsing or 12kHz constant tones, 2 seconds, 70dB) were randomly presented 50 times
before the delivery of sucrose (CS+, 10% sucrose in water; 2.0–2.5µL) or no sucrose
(CS−), such that there was a one second trace interval between delivery of the CS+
and sucrose (see Figure 1b). The cue contingencies were counterbalanced across cohorts
of mice to ensure that mice acquired conditioned licking in response to either tone
when paired with sucrose. The inter-trial interval between the previous reward delivery
(CS+) or withholding time (CS−) and the next cue was chosen as a random sample from
a uniform distribution bounded by 40s and 80s. Cue discrimination was quantified using
the area under a Receiver Operating Curve (auROC) formed by the number of licks during
the CS+ versus CS− trace intervals. For both two-photon and optogenetic behavioral
experiments, we classified sessions as ‘Early’ or ‘Late’ in learning, defined by both
behavioral performance (Early, auROC < 0.65; Late, auROC > 0.66) and session number
(Early, sessions 1–5; Late, sessions 7 or later). These criteria were used as post-hoc
analysis revealed that an auROC > 0.66 approximates high performance in a phase space
formed by behavioral performance across sessions. Finally, behavioral data is displayed
and analyzed throughout the manuscript as the change in lick rate (“ΔLick Rate”) between
each 3-second cue period and 1-second baseline period (baseline period is immediately
before each cue). In addition, we show raw lick rates during both the cue and baseline
periods for all imaging experiments (see Extended Data Fig 1). Baseline lick rates
remained relatively low across all experiments, and therefore for optogenetics studies
only “ΔLick Rate” is shown and analyzed (see Figures 4, Figure 5; Extended Data Fig.
8, Extended Data Fig. 9).
Two-photon microscopy
Experimental design
Two-photon microscopy was used to visualize activity dynamics of PFC neurons in vivo.
A virus encoding the calcium indicator GCaMP6s
18
(AAVdj-CaMKiiα -GCaMP6s; 5.3×1012 infectious units/mL) was injected into PFC (see
subjects and surgery). For imaging projection-specific neurons, a virus encoding the
cre-dependent calcium indicator GCaMP6s (AAVdj-ef1α-DIO-GCaMP6s; 3.1×1012 infectious
units/mL; from Karl Deisseroth) was injected into PFC, and the retrogradely transported
canine adenovirus encoding cre-recombinase
31,32
was injected into either NAc or PVT (Cav2-cre; 4.2×1012 infectious units/mL). After
a minimum of 8 weeks to allow virus transport and infection, mice underwent Pavlovian
conditioning during which GCaMP6s-expressing neurons were visualized using two-photon
microscopy.
Data acquisition, signal extraction, and analysis
A two photon microscope (FVMPE-RS) was equipped with the following to allow imaging
of PFC in vivo: a hybrid scanning core set with galvanometers and fast resonant scanners
(allows up to 30Hz frame-rate acquisition; set to 2.5 Hz), multi-alkali PMT and GaAsP-PMT
photo detectors with adjustable voltage, gain, and offset features, a single green/red
NDD filter cube, a long working distance 20× air objective designed for optical transmission
at infrared wavelengths (Olympus, LCPLN20XIR, 0.45NA, 8.3mm WD), a software-controlled
modular XY stage loaded on a manual z-deck, and a tunable Mai-Tai Deep See laser system
(Spectra Physics, laser set to 955nm, ~100fs pulse width) with automated four-axis
alignment. Before each conditioning session, a particular field of view (FOV) was
selected by adjusting the imaging plane (z-axis), and each FOV was spaced at least
50µm from one another to prevent visualization of the same cells across multiple FOVs.
During each conditioning session, two-photon scanning was triggered for each trial
7s before cue delivery, and a 20s video was then collected for each trial. Data were
both acquired and processed using a computer equipped with FluoView (Olympus, FV1200)
and cellSens (Olympus) software packages. Following data acquisition, videos were
motion corrected using a planar hidden Markov model (SIMA v1.3)
33
and regions of interest (ROIs) were hand drawn around each cell using the standard
deviation projection of the motion-corrected video using ImageJ. Next, calcium transient
time series data were extracted with SIMA and analyzed using custom Python data analysis
pipelines written in the lab (by VMN). For analysis, data were split into two groups
(Early and Late) that were defined based on behavioral performance and the day of
conditioning (see head-fixed behavior). Next, each recorded neuron was defined as
having an excitatory response, inhibitory response, or no response. Significant responses
represent significant two-tailed auROC comparing average fluorescence (Δf/f) of the
trace interval (1s after CS offset) versus baseline (1s before CS onset) where p<0.05
after Benjamini-Hochberg false discovery rate correction. Each p-value for auROC was
defined by calculating the p-values for the corresponding Mann-Whitney U statistic.
Chi-squared tests were then used to compare the number of CS+ responders to CS− responders
for each group. For additional decoding analysis (e.g. Figure 2f,l), we tested whether
the identity of the cue on any given trial could be decoded from the mean trace interval
response on that trial using support vector machines. To this end, we used the Python
module, scikitlearn, with GridSearchCV and a support vector classification (SVC) estimator
with a radial basis function kernel, optimizing across the following parameters: γ:
{10−2, 10−1, 100, 101, 102}, C: {10−2, 10−1, 100, 101, 102}. Quantification of performance
was done using 10-fold validation
34
For each neuron, the highest accuracy score across these parameters was used as the
metric of accuracy. In order to determine whether the population of accuracy scores
across all neurons was significantly different from that expected by chance, we performed
a single shuffle per neuron by randomizing the cue identity on every trial. The population
of shuffled accuracy scores across one shuffle was then compared to the population
of unshuffled accuracy scores using a two-tailed Welch's t-test. Note that since the
metric of accuracy was optimized across parameters, the mean accuracy score expected
by chance is not 0.5, but is instead closer to 0.55 (Figure 2f,l and Extended Data
Fig. 4d). We also further tested whether the mean activity during the trace interval
on a given trial for one neuron could be used to decode the number of licks in the
trace interval. This was performed using support vector regression (SVR) in scikitlearn
with GridSearchCV with a radial basis function kernel, optimizing across the following
parameters: C: 5 logarithmically equidistant points between 10−3 and 103 {10−3, 3.16×10−2,
100, 3.16×102, 103}, ε: 5 logarithmically equidistant points between 10−3 and 103
{10−3, 3.16×10−2, 100, 3.16×102, 103}, γ: 10 logarithmically equidistant points between
10−6 and 106 {10−6, 2.15×10−5, 4.64×10−4, 10−2, 2.15×10−1, 4.64, 102, 2.15×103, 4.64×104,
106}. Quantification of performance was done using 10-fold validation of the R2 metric
(note that this metric can be infinitely negative, indicating arbitrarily poor performance,
but is bounded on the positive end at 1, indicating perfect decoding). We found that
as a population, the number of anticipatory licks during the trace interval could
not be decoded in the late sessions in CaMKii-expressing neurons (mean R2=−1.21),
PFC-NAc neurons (mean R2=−0.92) or PFC-PVT neurons (mean R2=−0.39). These negative
numbers reflect the absence of a relationship between licking and calcium activity
in each cell population.
Optogenetics
Behavioral optogenetics were performed as described in detail (see Sparta et al.,
2011)
30
. Briefly, during surgery a virus encoding cre-inducible channelrhodopsin-2 (AAV5-EF1α-DIO-hChR2(H134R)-eYFP;
5.0×1012 infectious units/mL), halorhodopsin (AAV5-EF1α-eNpHR3.0-eYFP; 8.0×1012 infectious
units/mL), or control (AAV5-EF1α-eYFP; 6.0×1012 infectious units/mL) was injected
into PFC; and the retrogradely transported canine adenovirus encoding cre-recombinase
31,32
was injected into either NAc or PVT (Cav2-cre; 4.2×1012 infectious units/mL). After
a minimum of 8 weeks to allow sufficient virus transport and infection, mice underwent
Pavlovian conditioning.
For acquisition experiments (e.g., Figure 4), mice underwent 8 daily conditioning
sessions with laser followed by a test session (no laser). For photoactivation manipulations
in ChR2 or control mice, the laser (473nm; 8–10mW) was turned on for 5ms pulses (20Hz)
during 80% of the cue trials, starting at the cue onset and ending at the reward delivery.
For photoinhibition manipulations in eNpHR3.0 or control mice, the laser (532nm; 8–10mW)
did not pulse. Because there was no effect of laser in the control mice, these data
were collapsed across PFC-NAc and PFC-PVT groups. For expression experiments (e.g.,
Figure 5), after mice reached high performance criterion (‘Late’, auROC>0.66), they
underwent 6 daily conditioning sessions. Furthermore, every other session was selected
for optogenetic manipulations, during which the laser was presented for 3s during
either the cue and trace interval or at random time epochs outside of cue or reward
delivery. Because there was no effect of laser in the ChR2 or eNpHR3.0 control mice,
these data were collapsed for PFC-NAc groups and PFC-PVT groups. In addition, for
expression experiments subsets of control mice were used twice, once as ChR2 controls
(blue light), and again as eNpHR3.0 controls (green light). Following experiments,
histological verification of fluorescence and optical fiber placements were performed
as described previously
35
.
Behavioral data (ΔLick Rate, see above) was analyzed based on a priori comparisons
of interest (effect of laser on ChR2/eNpHR animals versus effect of laser in eYFP
animals). For acquisition experiments (Figure 4; Extended Data Fig. 8a–f), we analyzed
data from the no laser test day only, and specifically compared the ΔLick Rate between
the ChR2 or eNpHR groups versus the eYFP group. To correct for the double comparison
(ChR2 or eNpHR versus eYFP), we performed a Benjamini-Hochberg multiple comparisons
correction. For expression experiments (Figure 5; Extended Data Fig. 8 g–l; Extended
Data Fig. 9), in each pair of sessions (no laser, laser) we calculated the difference
in mean lick rate between the two in order to obtain a statistical measure of the
“effect of laser” per session pair. Next, we compared the effects of laser from the
ChR2 or eNpHR groups versus the corresponding effect of laser in the eYFP group. To
correct for the double comparison (ChR2 or eNpHR versus eYFP), we again performed
a Benjamini-Hochberg multiple comparisons correction. Considering this, for optogenetics
experiments all p-values (which are two-tailed throughout the manuscript) have been
corrected for multiple comparisons.
Retrograde tracing
The anatomy and electrophysiological properties of PFC-NAc and PFC-PVT neurons were
evaluated through retrograde tracing
36
. Specifically, during surgeries the retrograde tracer cholera toxin subunit B conjugated
to Alexa Fluor (CtB-488, CtB-594; Molecular Probes) was injected bilaterally into
NAc (500nl/side) and on the midline in PVT (300nl; color counterbalanced across mice).
Ten days following surgery, animals were sacrificed for histology (n=3 mice) or slice
electrophysiology (n=3 mice). For anatomical experiments, a student blind to all experiments
(E.P.M.) and conditions counted the number of CtB-488 positive, CtB-594 positive,
and double-positive neurons in prelimbic medial prefrontal cortex (a subregion of
dorsal medial PFC). The distance of each cell from the midline and the layer specificity
of each cell were then measured using ImageJ. For electrophysiological experiments,
mice were euthanized ten days following surgeries for patch-clamp recordings ex vivo
(see below for details).
Rabies tracing
The monosynaptic afferents to PFC-NAc and PFC-PVT neurons were identified using a
glycoprotein-deleted rabies strategy
37
in combination with Cav2-cre targeting of projection-specific neuron populations.
Specifically, during the first surgery a cocktail containing the cre-dependent starter
viruses encoding the G-protein and TVA were injected into PFC (3:1 of AAV5-FLEX-RG
and AAV5-FLEX-TVA-mCherry; 300nl/side), and Cav2-cre was injected into either NAc
(500nl/side) or PVT (300nl). Five weeks later, mice were given a second surgery in
which the G-deleted rabies virus was injected into PFC (1:5 diluted EnvA-Rabies-GFP).
Finally, 8 days after the rabies injection each mouse (n=3/group) was sacrificed for
histology and cell quantification. Our rabies protocol led to sparse labeling of PFC
projection neurons, allowing quantification of individual cells in each brain section
(40µm thick). Each ROI was selected based on previous PFC tracing experiments
38
, as well as the fluorescence intensity observed in our experiments. Next, out of
all tissue collected for each ROI in each mouse, we selected the 3 sections containing
the most cells per region, and used confocal microscopy to get cellular-resolution
images of all cells in each of those sections. For each section, we quantified all
individual input neurons (GFP+) and starter cells (both GFP+ and mCherry+). Considering
that the anterior cingulate cortex (ACC) was close to the PFC injection site, some
sections containing ACC also had starter cell labeling. Thus, because we were interested
in long-range inputs from ACC only, only sections that did not have mCherry labeling
were used for ACC input quantification. Finally, rabies-tracing data were analyzed
by comparing the number of cells in each section across groups (raw neuron count),
and by comparing the percentage of input neurons per starter cell for each particular
mouse.
Patch-clamp electrophysiology
Mice were anesthetized with pentobarbital (50mg/kg) before transcardial perfusion
with ice-cold sucrose cutting solution containing the following (in mM): 225 sucrose,
119 NaCl, 1.0 NaH2P04, 4.9 MgCl2, 0.1 CaCl2, 26.2 NaHCO3, 1.25 glucose, 305 mOsm.
Brains were then rapidly removed, and coronal sections 300µm thick were taken using
a vibratome (Leica, VT 1200). Sections were then incubated in aCSF (32°C) containing
the following (in mM): 119 NaCl, 2.5 KCl, 1.0 NaH2P04, 1.3 MgCl, 2.5 CaCl2, 26.2 NaHCO3,
15 glucose, ~306 mOsm. After an hour of recovery, slices were constantly perfused
with aCSF (32°C) and visualized using differential interference contrast through a
40× water-immersion objective mounted on an upright microscope (Olympus BX51WI). Whole-cell
recordings were obtained using borosilicate pipettes (3–5 ΜΩ) back-filled with internal
solution containing the following (in mM): 130 K gluconate, 10 KCl, 10 HEPES, 10 EGTA,
2 MgCl2, 2 ATP, 0.2 GTP (pH 7.35, 270–285 mOsm).
Current-clamp recordings were obtained from GCaMP6s-expressing neurons to identify
how action potential frequency correlated with GCaMP6s fluorescence. Specifically,
to determine how elevations in action potential frequency influence GCaMP6s fluorescence,
a 1 second train of depolarizing pulses (2nA, 2ms) was applied at a frequency of 1,
2, 5, 10, or 20Hz. To determine how attenuations in action potential frequency influence
GCaMP6s fluorescence, a 3 second pause was applied after a 10 second baseline train
of depolarizing pulses (2nA, 2ms; 1, 2, 5, 10, or 20Hz). Finally, to determine if
hyperpolarization influences GCaMP6s fluorescence in the absence of action potential
frequency modulation, a 3 second hyperpolarizing step (150pA) was applied in neurons
that were held either below or above resting membrane potential. During electrophysiological
recordings, GCaMP6s fluorescence dynamics were visualized using a mercury lamp (Olympus,
U-RFL-T) and microscope-mounted camera (QImaging, optiMOS). Imaging data were acquired
using Micro-Manager, and extracted through hand-drawn ROIs for each recorded neuron
using ImageJ.
Current-clamp recordings were also obtained to identify the intrinsic properties of
PFC-NAc and PFC-PVT neurons in retrograde tracing experiments, as previously described
39
. First, action potential firing was examined by applying a series of long depolarizing
sweeps (800ms) at +25pA steps (0–450pA). Next, rheobase (the minimum amount of current
required for an action potential to fire) was measured by applying a series of short
depolarizing sweeps (50ms) at +10pA steps (starting at 0pA) until the recorded neuron
fired an action potential. For all patch-clamp experiments, data acquisition occurred
at 1 kHz sampling rate through a MultiClamp 700B amplifier connected to a Digidata
1440A digitizer (Molecular Devices). Data were analyzed using Clampfit 10.3 (Molecular
Devices).
Data Collection
The nature of all imaging and behavioral experiments yields high-power data sets,
as we can test responses to reward-predictive cues hundreds of times within a single
session. Thus, although the experiments themselves require rigorous experimentation,
the number of mice that are required for each experiment is generally 3–6 per group,
depending on the effect size (which was not predetermined for these experiments).
Mice were randomly picked for each group in each experiment, by alternating the surgery
for each mouse in a cage. During data collection, investigators were only blind to
the conditions for rabies tracing cell counting and CtB cell counting. The only mice
excluded from final analysis were those that died before or during the experiments
(n=3). For optogenetics experiments, mice were excluded if histology confirmed ectopic
virus expression outside of PFC (n=1), or if cannula placements were not in dorsomedial
PFC (n=0). For data analysis, equal variance was not assumed for behavioral optogenetics
or imaging datasets. Equal variance was deemed equivalent and was therefore not assumed
for cell counting experiments and electrophysiological experiments.
Code and Data Availability
We used Python (codes written by VMN) to analyze imaging and optogenetics datasets
included in this manuscript (see Figures 1–5). That data, as well as the codes used
for analysis, are openly available online: <https://github.com/stuberlab>. All other
data are available upon request from the corresponding author.
Extended Data
Extended Data Figure 1
Mice used for imaging experiments acquired cue-specific anticipatory licking across
conditioning
a, Average lick rate during the 1-second baseline period (immediately before each
cue delivery) for all imaging experiments (Early, n=30; Late n=30). b, Average lick
rate during each cue (rather than the change in lick rate presented in the main figures)
for all imaging experiments (Early, n=30; Late n=30). c, Individual behavioral discrimination
(licking during CS+ versus CS−; auROC-0.5) scores during early and late conditioning
sessions for all imaging sessions used in this manuscript (Early, n=30; Late, n=30;
t(58)=43.0, p<0.001). Line graphs represent the mean±SEM. These data are presented
in a summarized form in Figure 1d and 1e.
Extended Data Figure 2
Elevations and reductions in GCaMP6s fluorescence track action potential frequency
but not voltage per se
a, Virus injections of AAVdj-CaMKii-GCaMP6s into dorsomedial PFC allowed subsequent
whole-cell patch-clamp slice recordings of GCaMP6s-expressing neurons. Coronal cartoon
redrawn based on Paxinos and Watson, 2007
40
. b,c, Example traces showing GCaMP6s fluorescence (b) during current-clamp recordings
(c), in the absence (left) and presence (right) of action potentials (n=8 neurons;
n=2 mice). d, Grouped data revealing that hyperpolarization resulted in negative GCaMP6s
fluorescence deflections in the presence of baseline action potentials, but not in
the absence of baseline action potentials (interaction: F(1,14)=20.0; p<0.001; post-hoc
tests: baseline APs, p>0.4; no baseline APs, p<0.001). e, Example traces showing a
series of depolarizing pulses (1–20Hz) applied in current clamp mode to drive trains
of action potentials (bottom), during which GCaMP6s fluorescence was tracked in recorded
neurons (top; n=12 neurons; n=2 mice). f, Action potential generation resulted in
linear elevations in GCaMP6s fluorescence (r=0.776, p<0.001), such that a single action
potential was detectable (red waveform; peak=12.6±4.0% Δf/f). g, A series of baseline
depolarizing pulses (1–20Hz) was applied before and after a 3 second pause (n=7 neurons;
n=2 mice). h, The pause in action potentials resulted in linear reductions in GCaMP6s
fluorescence (r=−0.656, p<0.001), such that a 1Hz reduction in firing frequency was
detectable (red waveform; peak=−8.4±2.0% Δf/f). Line graphs represent the mean±SEM.
AP, action potential; ILc, infralimbic cortex; PLc, prelimbic cortex.
Extended Data Figure 3
Cue responses in PFC CaMKii-expressing neurons, PFC-NAc neurons, and PFC-PVT neurons
before appetitive learning
a,b, Population heat plots showing average responses for all individual PFC CaMKii-expressing
neurons (n=1473 neurons; n=8 mice) across all CS− trials before learning (a) and all
CS+ trials before learning (b). c,d, Population heat plots showing average responses
for all individual PFC-NAc neurons (n=84 neurons; n=4 mice) across all CS− trials
before learning (c) and all CS+ trials before learning (d). e,f, Population heat plots
showing average responses for all individual PFC-PVT neurons (n=92 neurons; n=3 mice)
across all CS− trials before learning (e) and all CS+ trials before learning (f).
Data shown here are from neurons presented in Figures 1 & 2. Vertical dotted lines
refer to the time of sucrose delivery (CS+ trials) or no sucrose delivery (CS− trials).
CS, conditioned stimulus.
Extended Data Figure 4
Cue discrimination in PFC CaMKii-expressing neurons before and after appetitive learning.
a–c
Histograms for all recorded CaMKii-expressing PFC neurons (Early, n=1473 neurons;
Late, n=1571 neurons; n=8 mice), showing CS+ responses (a), CS− responses (b), and
CS+/CS− discrimination (c) during both early and late Pavlovian conditioning sessions.
d, CDF plot showing that the dynamics of individual CaMKii-expressing neurons could
be used to accurately decode whether the CS+ or CS− was presented in early conditioning
sessions (compared to early shuffled: Welch’s t(2925.61)=7.30, p<0.001), as well as
in late conditioning sessions (compared to late shuffled: Welch’s t(2727.06)=24.84,
p<0.001). Data shown here are from neurons presented in Figure 1. CDF, cumulative
distribution frequency; CS, conditioned stimulus; EarlySh, early shuffled; LateSh,
late shuffled.
Extended Data Figure 5
Corticostriatal and corticothalamic projection neurons are anatomically and electrophysiologically
distinct
a,b, CtB-488 and CtB-594 were injected (n=3 mice) into either NAc or PVT (scale bars=250µm)
(a), resulting in retrograde labeling of NAc-projecting and PVT-projecting neurons
in dorsomedial PFC (scale bars=50µm) (b). c–e, Coronal sections of anterior (c), middle
(d), and posterior (e) dorsomedial PFC revealing spatial separation of PFC-NAc and
PFC-PVT neurons (scale bars=250µm). f–h, Population histograms from all mice (n=3)
showing cell counts from anterior (f), middle (g), and posterior (h) dorsomedial PFC
reveal that NAc-projecting neurons (red) are in more superficial layers as compared
with PVT-projecting neurons (blue). Black refers to overlap (medial-lateral axis)
between red and blue bars, and purple refers to double-labeled neurons (i.e., both
NAc and PVT projection neurons). i, Current clamp recordings from dorsomedial PFC
CtB-labeled neurons projecting to either NAc (top; n=9 neurons; n=3 mice; scale bars=25µm)
or PVT (bottom; n=10 neurons; n=3 mice; scale bars=25µm). j,k, Representative waveforms
(j) and averaged data (k) showing that PFC-NAc neurons fired fewer action potentials
(spikes) as compared to PFC-PVT neurons during somatic depolarization (interaction:
F(16,272)=16.6, p<0.001). l,m, Representative waveforms (l) and averaged data (m)
revealing no differences in rheobase (the minimum current required to evoke an action
potential) between PFC-NAc and PFC-PVT neurons (t(17)=1.22, p>0.2). n,o, Representative
waveforms (n) and averaged data (o) showing that PFC-NAc neurons had larger peak AHPs
as compared with PFC-PVT neurons (t(17)=4.67, p<0.001). p, The maximum number of action
potentials (spikes) in each neuron was correlated with the peak AHP (r=0.80, p<0.001).
Line and bar graphs represent the mean±SEM. AHP, afterhyperpolarization, cc, corpus
callosum; ILc, Infralimbic cortex; MO, medial orbitofrontal cortex; PLc, prelimbic
cortex; VO ventral orbitofrontal cortex.
Extended Data Figure 6
Corticostriatal and corticothalamic projection neurons have distinct monosynaptic
inputs
a–d, Viral strategy for rabies tracing experiments in which the monosynaptic inputs
to (a,b) PFC-NAc and (c,d) PFC-PVT neurons were evaluated (n=3 mice/group). e–g, Example
images showing (e) mCherry+ cells (TVA expression), (f) RV-GFP+ cells (local interneurons),
and (g) overlap revealing mCherry+/RV-GFP+ cells (starter cells) or only GFP+ cells
(local interneurons). h, The number of local inputs neurons (nonstarter; only GFP+
cells per section) to each projection population, as quantified by raw neuron count
and by the percent of starter cells for each mouse, was equivalent for PFC-NAc and
PFC-PVT neurons (raw neuron count: t(16)=0.59, p=0.56; % starter cells: t(16)=0.13,
p=0.90). i, Representative image showing RV-GFP but not mCherry expression in the
ACC. j, The number of input neurons from ACC was higher for PFC-PVT neurons as compared
with PFC-NAc neurons (raw neuron count: t(16)=3.51; p=0.003; % starter cells: t(16)=3.31,
p=0.004). k, Representative image showing RV-GFP but not mCherry expression in the
LPO. l, The number of input neurons from the LPO was equivalent for PFC-NAc and PFC-PVT
cells (raw neuron count: t(16)=1.77; p=0.01; % starter cells: t(16)=0.20, p=0.84).
m, Representative image showing RV-GFP but not mCherry expression in the vHipp. n,
The number of input neurons from vHipp was higher for PFC-NAc neurons as compared
with PFC-PVT neurons (raw neuron count: t(16)=4.44; p<0.001; % starter cells: t(16)=4.00,
p=0.001). o, Representative image showing RV-GFP but not mCherry expression in the
VTA. p, The number of input neurons from the VTA was equivalent for PFC-NAc and PFC-PVT
cells (raw neuron count: t(16)=0.56; p=0.59; % starter cells: t(16)=0.09, p=0.93).
Bar graphs represent the mean±SEM. ACC, anterior cingulate cortex; RV-GFP, rabies
virus encoding green fluorescent protein; LPO, lateral preoptic area; vHipp, ventral
hippocampus; VTA, ventral tegmental area. *Note: no RV-GFP+ neurons were detected
in any nucleus of the amygdala for either projection group.
Extended Data Figure 7
Corticostriatal and corticothalamic projection neurons express CaMKii and have distinct
basal activity dynamics
a–d, Injections of AAV5-CaMKii-eYFP into dorsomedial PFC and the retrograde tracer
CtB-594 into NAc (a) or PVT (c) resulted in expression of eYFP in CtB-labeled PFC-NAc
neurons (b) and PFC-PVT neurons (d). These data reveal that PFC-NAc and PFC-PVT are
subpopulations of CaMKii-expressing neurons (n=2 mice/group). e,f, In ai9 reporter
mice, (e) DIO-GCaMP6s injections in dorsomedial PFC and Cav2-cre injections into PVT
(f) result in expression of GCaMP6s and tdTomato (marker for cre-recombinase), which
have spatial overlap in PFC (n=2 mice). These data reveal that GCaMP6s expression
is specific to the projection cells of interest. g, Example traces revealing spontaneous
calcium dynamics from in vivo two-photon imaging in GCaMP6s-expressing PFC-NAc neurons
(top; n=69 neurons; n=4 mice) and PFC-PVT neurons (bottom; n=61; n=3 mice) in awake,
head-fixed mice. Red and blue dots refer to detected events. h–j, Averaged data reveal
no differences in event amplitude (h) or event duration (i); however, PFC-NAc neurons
had significantly shorter inter-event intervals (j) as compared to PFC-PVT neurons
(amplitude: t(130)=1.10, p>0.2; duration: t(130)=0.68, p>0.4; interval: t(130)=2.30,
p<0.05). Bar graphs represent the mean±SEM. CtB, cholera toxin subunit B; tdT, tdTomato.
Extended Data Figure 8
Effects of corticostriatal and corticothalamic optogenetic manipulations on acquisition
and expression of CS− licking
Acquisition: a, Line graph showing average CS− lick rate during conditioning (with
laser) and test (no laser) from PFC-NAc::ChR2 (n=5), PFC-NAc::eNpHR (n=6), and PFC-NAc::eYFP
mice (n=10). b,c, CDF plots and bar graphs showing CS− lick rate during conditioning
(b) and test (c). No differences were observed between PFC-NAc groups during the no-laser
test (ChR2 vs. eYFP: auROC=0.53, BHC p=0.43; eNpHR vs. eYFP: auROC=0.45, p=0.43).
d, Line graph showing average CS− lick rate during conditioning (with laser) and test
(no laser) from PFC-PVT::ChR2 (n=6), PFC-NAc::PVT (n=5), and PFC-PVT::eYFP mice (n=10).
e,f, CDF plots and bar graphs showing CS− lick rate during conditioning (e) and test
(f). No differences were observed between PFC-PVT groups during the no-laser test
(ChR2 vs. eYFP: auROC=0.48, BHC p=0.48; eNpHR vs. eYFP: auROC=0.32, p=0.30). Expression:
g–i, CDF plots and bar graphs showing CS− lick rates for PFC-NAc::ChR2 (n=5), PFC-NAc::eNpHR
(n=5), and PFC-NAc::eYFP mice (n=8). There were no significant differences in CS−
lick rate for PFC-NAc::ChR2 mice (vs. PFC-NAc::eYFP: auROC=0.43, p=0.26), although
there was an effect of laser for PFC-NAc eNpHR mice (vs. PFC-NAc::eYFP: auROC=0.23,
p=0.006). j–l, CDF plots and bar graphs showing CS− lick rates for PFC-PVT::ChR2 (n=5),
PFC-PVT::eNpHR (n=5), and PFC-PVT::eYFP mice (n=6). There were no significant differences
in CS− lick rate for PFC-PVT::ChR2 mice (vs. PFC-PVT::eYFP: auROC=0.35, p=0.15) or
PFC-PVT::eNpHR mice (vs. PFC-PVT::eYFP: auROC=0.55, p=0.31). Line and bar graphs represent
the mean±SEM. CDF, cumulative distribution frequency; NL, no laser test.
Extended Data Figure 9
Effects of corticostriatal and corticothalamic optogenetic manipulations are timing
dependent
a–c, CDF plots (top) and bar graphs (bottom) show anticipatory licking rates for PFC-NAc::ChR2
(n=5) or PFC-NAc::eNpHR3.0 (n=5) versus PFC-NAc::eYFP mice (n=6) during sessions in
which the laser was randomly presented outside of cue delivery. There were no significant
differences in anticipatory lick rate for PFC-NAc::ChR2 mice (vs. PFC-NAc::eYFP: auROC=0.56,
BHC p=0.30) or PFC-NAc::eNpHR mice (vs. PFC-NAc::eYFP: auROC=0.63, p=0.23). d–f, CDF
plots (top) and bar graphs (bottom) show anticipatory licking rates for PFC-PVT::ChR2
(n=5) or PFC-PVT::eNpHR3.0 (n=5) versus PFC-PVT::eYFP (n=8) mice during sessions in
which the laser was randomly presented outside of cue delivery. There were no significant
differences in anticipatory lick rate for PFC-PVT::ChR2 mice (vs. PFC-PVT::eYFP: auROC=0.42,
p=0.21) or PFC-PVT::eNpHR mice (vs. PFC-PVT::eYFP: auROC=0.36, BHC p=0.14). g–i, CDF
plots (top) and bar graphs (bottom) show CS− lick rates for PFC-NAc::ChR2 (n=5) or
PFC-NAc::eNpHR3.0 (n=5) versus PFC-NAc::eYFP mice (n=6) during sessions in which the
laser was randomly presented outside of cue delivery. There were no significant differences
in CS− lick rate for PFC-NAc::ChR2 mice (vs. PFC-NAc::eYFP: auROC=0.41, p=0.19) or
PFC-NAc::eNpHR mice (vs. PFC-NAc::eYFP: auROC=0.40, p=0.19). j–l, CDF plots (top)
and bar graphs (bottom) show CS− lick rates for PFC-PVT::ChR2 (n=5) or PFC-PVT::eNpHR3.0
(n=5) versus PFC-PVT::eYFP (n=8) mice during sessions in which the laser was randomly
presented outside of cue delivery. There were no significant differences in CS− lick
rate for PFC-PVT::ChR2 mice (vs. PFC-PVT::eYFP: auROC=0.39, p=0.12) or PFC-PVT::eNpHR
mice (vs. PFC-PVT::eYFP: auROC=0.36, p=0.12). Bar graphs represent the mean±SEM. CDF,
cumulative distribution frequency.
Extended Data Figure 10
Optogenetic manipulations of corticostriatal and corticothalamic neurons are not appetitive,
aversive, and do not affect movement
a, Tracking data from single example mice during real time place preference experiments
showing that PFC-NAc::ChR2 (left; 5) and PFC-NAc::eNpHR3.0 (right; n=5) mice spent
equivalent time in chambers that were paired with laser (PFC-NAc::eYFP mice, n=8).
b, Grouped data show that laser stimulation in PFC-NAc mice did not lead to a real-time
place preference (interaction: F(2,30)=2.15, p>0.13). c, Grouped data show that optogenetic
manipulations in PFC-NAc mice did not influence velocity of movement (interaction:
F(2,30)=0.12, p>0.88). d, Tracking data from single example mice during real time
place preference experiments showing that PFC-PVT::ChR2 (left; n=5) and PFC-PVT::eNpHR3.0
(right; n=5) mice spent equivalent time in chambers that were paired with laser (PFC-PVT::eYFP
mice, n=5). e, Grouped data show that laser stimulation in PFC-PVT mice did not lead
to a real-time place preference (interaction: F(2,24)=0.15, p>0.86). f, Grouped data
show that optogenetic stimulation in PFC-PVT did not influence velocity of movement
(interaction: F(2,24)=0.10, p>0.90). g,h, Coronal plates (redrawn based on Paxinos
and Watson, 2007
40
) located 1.98, 1.78, and 1.54 mm anterior to bregma illustrate the placements of
optical fiber tips in PFC-NAc mice (g) and PFC-PVT mice (h). Bar graphs represent
the mean±SEM. NoStim, no laser stimulation; Stim, laser stimulation.
Supplementary Material
video
video caption