Reward salience but not spatial attention dominates the value representation in the orbitofrontal cortex

Abstract

The orbitofrontal cortex (OFC) encodes value and plays a key role in value-based decision-making. However, the attentional modulation of the OFC’s value encoding is poorly understood. We trained two monkeys to detect a luminance change at a cued location between a pair of visual stimuli, which were over-trained pictures associated with different amounts of juice reward and, thus, different reward salience. Both the monkeys’ behavior and the dorsolateral prefrontal cortex neuronal activities indicated that the monkeys actively directed their spatial attention toward the cued stimulus during the task. However, the OFC’s neuronal responses were dominated by the stimulus with higher reward salience and encoded its value. The value of the less salient stimulus was only weakly represented regardless of spatial attention. The results demonstrate that reward and spatial attention are distinctly represented in the prefrontal cortex and the OFC maintains a stable representation of reward salience minimally affected by attention.

Full text

Article https://doi.org/10.1038/s41467-022-34084-0
Reward salience but not spatial attention
dominates the value representation
in the orbitofrontal cortex
Wenyi Zhang
1,2,3
,YangXie
1,2,3
& Tianming Yang
1
The orbitofrontal cortex (OFC) encodes value and plays a key role in value-
based decision-making. However, the attentional modulation of the OFC’s
value encoding is poorly understood. We trained two monkeys to detect a
luminance change at a cued location between a pair of visual stimuli, which
were over-trained pictures associated with different amounts of juice reward
and, thus, different reward salience. Both the monkeys’behavior and the
dorsolateral prefrontal cortex neuronal activities indicated that the monkeys
actively directed their spatial attention toward the cued stimulus during the
task. However, the OFC’s neuronal responses were dominated by the stimulus
with higher reward salience and encoded its value. The value of the less salient
stimulus was only weakly represented regardless of spatial attention. The
results demonstrate that reward and spatial attention are distinctly repre-
sented in the prefrontal cortex and the OFC maintains a stable representation
of reward salience minimally affected by attention.
Imagine that you are in a bookstore trying to find an ideal gift for your
young child. While you direct your search in the children’s book sec-
tion, an eye-catching poster of your favorite writer in the bestseller
section would divert your attention and lead you to check out the
books over there. While task demands or behavior context often lead
to internally generated attention, commonly referred to as top-down
attention, salient stimuli, either based on physical or reward salience,
may capture attention via a bottom-up mechanism1–6. Attention from
top-down and bottom-up sources may be integrated or compete
against each other in the brain and affect decision making.
Previous studies suggested that attention may affect value-based
decision making by modifying choice options’subjective value7,8.
Moreover, attention, both in the overt form with gaze shifts and in the
covert form without eye movements, has been shown to modulate the
representation of value in the orbitofrontal cortex (OFC)9–11, which is a
key brain area involved in value-based decision making and adaptive
behavior12–16.Inaddition,theOFCneuronalactivitiesencodedeach
choice option’s value alternately during value-based decision making,
which might be related to attentional shifts17,18. Based on these studies,
it has been proposed that the OFC activity reflects the value of the
attended item while multiple items are presented.
However, in these studies, the researchers inferred the location of
attention either from reward or visual salience11 or from gaze
location9,10. Without clear behavior or neurophysiology markers of
attention, these studies only provided indirect evidence for the pro-
posed theory. Especially when the behavior task and the stimulus or
reward salience direct attention to distinct locations, an experiment
that allows one to measure the attention location directly is necessary
for investigating how attention from different sources modulates the
OFC’s value encoding.
To this end, we trained two macaque monkeys to perform a visual
detection task with a Posner cueing paradigm19. A pair of visual stimuli
were presented simultaneously, and a cue indicated which of the sti-
muli would change its luminance. The monkeys had to direct their
attention toward the cued stimulus to detect the luminance change
better. In addition, the stimuli were well-trained shapes that were
associated with different amounts of reward. They acquired different
levels of salience through over-training and couldcapture attention via
Received: 11 August 2021
Accepted: 11 October 2022
Check for updates
1
Institute of Neuroscience, Key Laboratory of Primate Neurobiology, Center for Excellence in Brain Science and Intelligence Technology, Chinese Academy of
Sciences, Shanghai 200031, China.
2
University of Chinese Academy of Sciences, Beijing 100049, China.
3
These authors contributed equally: Wenyi Zhang,
Yang Xie. e-mail: tyang@ion.ac.cn
Nature Communications | (2022) 13:6306 1
1234567890():,;
1234567890():,;
Content courtesy of Springer Nature, terms of use apply. Rights reserved

a bottom-up mechanism. We were able to verify that the monkeys
directed their attention according to the cue based on both the
monkeys’behavior and the dorsolateral prefrontal cortex (DLPFC)
neuronal activities. However, contrary to the proposed theory, we
observed that the value encoding in the OFC was dominated by reward
salience, regardless of where the cue and the attention were. The
results indicate that the OFC maintains a stable salience-based repre-
sentation of value that is only weakly modulated by spatial attention.
Results
Behavioral task and subject performance
We trained two monkeys to perform a visual detection task (Fig. 1a, d).
In this task, the monkeys were instructed to covertly pay attention to
one of the two stimuli to detect a luminance change at a random delay.
The stimulus with the luminance change was cued by a frame that was
presented on the side opposite to the change location 200 ms before
the onset of the stimulus pair. The monkeys were required to report
the change by making a saccade toward an eye-movement target
located above the fixation spot to get a juice reward. For monkey D, the
cue was valid in 90% of the trials. In the remaining trials (invalid-cue
trials), the change location was on the same side of the frame (un-cued
location). Monkey G was also trained to detect the luminance change
at the cued location (target trials: 80%). In addition, to encourage
monkey G to direct its attention appropriately, we required it to ignore
the luminance change at the un-cued location and not to make a
response (distractor trials: 20%). In half of the distractor trials, another
luminance change would happen at the cued location after the dis-
tractor(distractor+target trials: 10%), and monkey G was rewarded for
detecting the change. In the other half of the distractor trials (dis-
tractor-only trials: 10%), there was no luminance change at the target
stimulus, and the monkey had to hold its fixation till the endof the trial
to receive a reward.
The stimulus set included 5 pictures for each animal. Each picture
was associated with a juice reward of a different size (Fig. 1a, d, inset).
The reward for correct responses was randomly chosen between the
rewards associated with the two stimuli. Therefore, the reward out-
comes were only 50% certain. The initial frame cue, as well as the
stimulus where the luminance occurred, did not convey information
about which stimulus’s reward would be delivered. We initially trained
the monkeys withsingle-stimulus trials, in which only one stimulus was
presented, and the frame always appeared on its opposite side. Its
associated reward would be delivered for a correct response, and there
was no ambiguity about the reward outcome.
For the ease of the discussion, between the two stimuli presented
in each trial, we refer to the stimulus that is cued to have a luminance
change as the cued stimulus, its value as cued value (CV). The other
stimulus is referred to as the un-cued stimulus and its value as un-cued
value (UCV). In addition, as the monkeys were over-trained with this
task and were familiar with the stimuli, the stimuli acquired different
levels of salience due to their reward associations: high salience for
stimuli associated with larger rewards and low salience for stimuli
associated with smaller rewards. Accordingly, we refer to the stimulus
a
Monkey D
Fixation
Attention cue on
Visual stimulus on
Attention cue off
Luminance change
1000 ms
200 ms
200 ms
400-1800 ms
Valid-cue
Invalid-cue
~0 1 2 4 8
Stimulus set:
Juice (drops):
90%
10%
Cued Salient
d
Monkey G
Fixation
Attention cue on
Visual stimulus on
Attention cue off
Luminance change
1000 ms
200 ms
200 ms
400-2300 ms
Target
Distractor-only
~0 1 2 4 8
Stimulus set:
Juice (drops):
Distractor
+ target
80%
10%
10%
Cued Salient
p=0.081
Valid Invalid
150
200
250
Reaction time (ms)
**
Valid Invalid
0.5
0.6
0.7
0.8
0.9
1
Hit rate
Lum change on salient stim
Lum change on non-salient stim
** p=0.25
**
bc
Target (hit)
Distractor-only/distractor+target
(corr rej)
Distractor+target (hit)
0.5
0.6
0.7
0.8
0.9
1
Hit/correct rejection rate
**
**
**
150
200
250
300
350
Reaction time (ms)
Target (hit)
Distractor-only/distractor+target
(false alarm)
Distractor+target (hit)
**
*
**
** **
ef
p=0.37
Fig. 1 | Behavioral tasks and monkeys’performance. a Monkey D had to detect
when any of the two stimuli changed its luminance. The location of the luminance
changewas indicated bya cue (90% valid) thatappeared on theside oppositeto the
luminance change. When the cue was invalid (10% of trials), the luminance change
occurred on the same side of the cue. The monkey reported the luminance change
by makingan eye movementto the eye-movement targetlocated abovethe fixation
point. The stimuli were associated with different amounts of juice reward (inset).
Random pairings of stimuli were selected for each trial. The stimulus with a larger
associated reward is referred to as the salient stimulus, and the stimulus where
attention shouldbe direct to is thecued stimulus.The computer randomly selected
one stimulus from the pair and delivered its associated reward for a correct
response. bThe proportion of correct responses whenthe cue was validand invalid
and when the luminance change occurred at the salient and the non-salient sti-
mulus. cThe reaction times when the cue was valid and invalid and when the
luminance changeoccurred at the salient and thenon-salient stimulus. dMonkey G
was requir ed to detect the luminance change at the ta rget location, ind icated with
the cue frame (target trials, 80%), and ignore the change at the un-cued location
(distractor trials, 20%). In half of the distractor trials, another luminance change
happened at the cued location after the distractor changed its luminance (dis-
tractor+target trials, 10%). The stimuli were associated with different amounts of
juice reward (inset). eThe proportion of correct responses in the three trial types
and when the luminance change occurred at the salient and the non-salient sti-
mulus. fThe reaction time of detecting the target and the distractor luminance
change and when the luminancechange occurred atthe salient and the non-salient
stimulus. The error bars are SEM across sessions (monkey D: n= 71; monkey G:
n= 100), * denotes p< 0.01, ** denotes p< 0.001, two-tailed Wilcoxon signed-
rank test.
Article https://doi.org/10.1038/s41467-022-34084-0
Nature Communications | (2022) 13:6306 2
Content courtesy of Springer Nature, terms of use apply. Rights reserved

associated with a larger reward as the salient stimulus, and the other
one as the non-salient stimulus. Their values are termed salient value
(SV) and non-salient value (NSV). SV and NSV are the same in trials with
a pair of identical stimuli.
Both monkeys learned the task and used the cue appropriately to
direct their attention. When the cue was valid, monkey D were more
accurate (valid: mean hit rate = 94.2%, SEM = 0.2%; invalid: mean hit
rate = 90.3%, SEM = 0.4%. p≪0.001, two-tailed Wilcoxon signed-rank
test. SEMs are across sessions) and responded faster (valid: mean
reaction time (RT) = 224.5 ms, SEM = 2.9 ms; invalid: mean RT = 234.1
ms, SEM = 3.6 ms. p≪0.001, two-tailed Wilcoxon signed-rank test).
Similarly, Monkey G accurately detected the luminance change at the
target (target: mean hit rate = 92.2%, SEM = 0.3%; distractor+target:
mean hit rate = 90.0%, SEM = 0.6%) and ignored the distractor’slumi-
nance change (distractor trials: mean correct rejection rate = 81.4%,
SEM =0.6%). The response latency of the target was shorter than that
of the distractor in the false alarm trials (target responses in target
trials: mean RT = 278.2 ms, SEM = 0.9ms; target responses indistractor
+target trials: mean RT = 271.2 ms, SEM = 1.1 ms; distractor responses:
mean RT = 297.8 ms, SEM = 1.0 ms. p« 0.001 for the difference
between the target response in either target or distractor+ target trials
and the distractor responses, two-tailed Wilcoxon signed-rank test).
The stimulus salience also affected the monkeys’performance,
although to a much lesser degree than the cue. The monkeys were
more likely to detect a luminance change of a salient stimulusthan that
of a non-salient stimulus (monkey D: valid salient: mean hit rate =
94.8%, SEM = 0.2%; valid non-salient: mean hit rate = 93.9%, SEM =
0.2%; monkey G: target salient: mean hit rate = 94.0%, SEM = 0.3%;
target non-salient: mean hit rate = 91.2%, SEM = 0.4%. p≪0.001 for
both monkeys, two-tailed Wilcoxon signed-rank test, Fig. 1b, e). Mon-
key G also responded faster when the target stimulus was the salient
stimulus (target salient: 273.8 ms, SEM=1.0ms; target non-salient:
282.2 ms, SEM = 1.0 ms. p≪0.001, two-tailed Wilcoxon signed-rank
test, Fig. 1f), while such an improvement was not seen in monkey D
(Fig. 1c). These behavior improvements caused by reward salience
were however much smaller than those caused by the cue (monkey D:
valid non-salient: mean RT = 225.4 ms, SEM = 2.9 ms; invalid salient
mean RT = 236.3 ms, SEM = 3.1 ms; monkey G: target non-salient: mean
RT = 282.2 ms, SEM = 1.0 ms; distractor salient: mean RT = 294.4 ms,
SEM = 1.2 ms. p≪0.001, two-tailed Wilcoxon signed-rank test,
Fig. 1c, f). As the reward for correct responses was randomly chosen
between the two stimuli, paying attention to the salient stimulus did
not bring any behavior benefits to the monkeys or provide information
about the expected reward. Correspondingly, the behavior of the
monkeys indicated that their attention was dominated by the cue.
DLPFC population encodes spatial attention shifts
In addition to the evidence from the monkeys’behavior, we recorded
single-unit activity from the DLPFC neurons to further confirm that the
monkeys’attention was directed to the cued location. A total of 406
DLPFC single units (monkey D: 240; monkey G: 166) from Walker’s
areas 8a, 8b, 46d, and 46v (Supplementary Fig. 1) were recorded, and
their activities were used to decode the location of attention20,21.
One important feature of our task is that the frame serving as the
cue was displayed on the side opposite to where the attention needed
to be directed. The onset ofthe frame might first capture the monkeys’
attention to a wrong location via a bottom-up process. Afterward, the
monkeys had to shift their attention to the opposite side where the
luminance change was expected. This attention shift was captured by
many DLPFC neurons that encoded spatial attention location. In Fig. 2a
we show such an example neuron. The neuron had larger responses
initially when the frame appeared to the left of the fixation point (blue
trace). After the cue offset, its response became greater in trials that
had the cue frame on the right but the cued stimulus on the left (red
trace). The neuron’s responses reflected the attention shift from the
side of the frame to the opposite side where the luminance change was
expected.
Population analyses confirmed that the DLPFC neurons encoded
spatial attention location and reflected the attentional shift. To pool
neurons with different spatial selectivity together, we considered the
whole population of recorded neurons as a high-dimensional repre-
sentation of task-related variables17,22,23 and trained a decoder based
on the linear discriminant analysis (LDA) algorithm with a pseudo
neuronal ensemble to decode the frame location. The pseudo neuro-
nal ensemble was composed in such a way that each neuron had the
same number of left and right frame trials (see Methods). Thedecoders
based on the LDA provided the posterior probabilities of the frame
location.
We trained and tested the decoders with DLPFC neurons’
responses using a sliding window (size: 25 ms, step: 10 ms) and plotted
the decoders’performance as a heat map (Fig. 2b). When the decoders
were trained and tested with the responses at the same time point,
indicated by the diagonal in Fig. 2b, the cue location can be reliably
decoded shortlyafter the onset of the attention cue (Fig. 2c, left) until
after the luminance change and the animal’sresponse(Fig.2c, right).
During the period from ~200 ms after the stimulus onset until after the
luminance change, the encoding of the attentionlocation was stable so
that the cross-temporal decoding performance was maintained
(Fig. 2b). Furthermore, the cross-temporal decoder’s performance
revealed the attention shift after the shape onset. We trained a decoder
with the neuronal responses 50–200 ms before the stimuli onset when
only the frame but not the stimuli were on the screen. The decoder’s
performance was then tested with the responses at the other time
points in a trial (Fig. 2d). The decoder’s performance quickly dropped
below the shuffled level after the stimulus onset and reached its
minimum after the luminance change. Lower-than-shuffled perfor-
mance indicates that the attention location shifted to the side opposite
to where it was when the decoder was trained. This attention shift can
be also observed in Fig. 2b in the dark blue regions—an indication of
the performance below the chance level—where the decoders were
trained and tested at different time points when the frame was pre-
sented and when the stimuli were presented.
The encoding of attention location in the DLPFC correlated with
the monkeys’performance. We divided the trials by the monkeys’RT
into two halves and tested the decoder’s performance in each half of
the trials. The decoder performed significantly better before the onset
of the luminance in trials when the monkeys reacted faster to the
luminance change, suggesting that the attention was more likely to be
allocated correctly in those trials (Fig. 2e, Supplementary Fig. 2). In
addition, we were able to decode attention in the correct trials better
than in the error trials (Supplementary Fig. 3).
OFC responses were dominated by the salient value
After verifying that the monkeys directed their attention to the proper
location, we investigated how the OFC neurons’value encoding was
modulated by attention.
We recorded the activities of 349 OFC neurons (monkey D: 212;
monkey G: 137) from Walker’s areas 11 and 13 (Supplementary Fig. 1).
Many of them encoded stimulus values. An example neuron is shown
in Fig. 3a, d, g, j. We first used trials with two identical stimuli to
measure the neuron’s value selectivity. In these trials, the monkeys
should be certain of the reward they would get for a correct response,
and our previous study indicated that the OFC neurons responded
similarly to a pair of identical stimuli and to the same stimulus pre-
sented alone11. The example neuron encoded the value of the stimuli;
its responses were larger when the value was larger (Fig. 3a). To study
the neuron’s responses to a pair of distinct stimuli, we first groupedthe
trials by SV. The neuron’sresponsesreflected SV stably during
the stimulus period well until after the luminance change (Fig. 3d).
We calculated the neuron’sfiring rates between the stimulus onset and
Article https://doi.org/10.1038/s41467-022-34084-0
Nature Communications | (2022) 13:6306 3
Content courtesy of Springer Nature, terms of use apply. Rights reserved

the luminance change and plotted them against SV (Fig. 3g). The
responses to SV were similar to the responses when a pair of the same
stimuli were presented. The non-salient stimulus was largely ignored.
In comparison, the neuron exhibited similar responses when the trials
were grouped by CV and when the trials were grouped by the UCV
(Fig. 3j). Both CV and UCV contributed to the neuron’s responses
similarly. The results suggested that the reward salience but not the
cue dominated the example neuron’s responses.
The OFC population exhibited the same trend. Among the 349 OFC
neurons from which we recorded, 126 neurons were selective to value
(Supplementary Table 1). We divided these neurons into positively
tuned neurons and negatively tuned neurons based on their value
tuning (see Methods). The positively tuned neurons had larger
responses when the values were higher (Fig. 3b, 73 out of 126 value-
tuned neurons), and the negatively tuned ones had the opposite tuning
(Fig. 3c, 53 out of 126 value-tuned neurons). Both groups of neurons
encoded SV stably during the stimulus period (Fig. 3e, f). Their firing
rates sorted by SV were similar to when a pair of the stimuli with the
same value as SV were presented (Fig. 3h, i and Supplementary Fig. 4),
and their responses did not distinguish between CV and NCV (Fig. 3k, l
and Supplementary Fig. 5). The results were consistent across the two
individual monkeys (Supplementary Fig. 6). Finally, to study the inter-
action between cue and reward salience, we grouped the trials by cued-
SV, un-cued-SV, cued-NSV, and un-cued-NSV and plotted the OFC neu-
rons’responses to each group condition. The responses were mostly
divided between the salient and non-salient groups, while the cued and
the un-cued groups exhibited similar responses, indicating little inter-
action between cue and reward salience. (Supplementary Fig. 7).
b
d
-400 -200 0 200 400 600 -400 -200 0 200 400
0
0.2
0.4
0.6
0.8
1
Posterior probability
Time from stim on (ms)
-400
-200
0
200
400
600
Time from stim on (ms) Time from lum change (ms)
-400 -200 0 200 400
Time from lum change (ms)
Training
Testing
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Posterior
probability
-400 -200 0 200 400 600
-400
-200
0
200
400 Cue on Cue off
Cue on Cue off
Stim on Luminance change
Luminance change
Stim on
c
-400 -200 0 200 400 600 -400 -200 0 200 400
0
0.2
0.4
0.6
0.8
1
Posterior probability
Attention cue loc
Control
Cue on Cue offStim on Luminance change
Cue on Cue offStim on Luminance change
e
-400 -200 0 200 400 600 -400 -200 0 200 400
Time from stim on (ms)
0
0.2
0.4
0.6
0.8
1
Posterior probability
Quick
Slow
Time from lum change (ms)
Cue on Cue offStim on Luminance change
-400 -200 0 200 400 600 -400 -200 0 200 400
0
5
10
15
20
25
30
35
Firing rates (spikes/s)
Right attention cue
Left attention cue
Time from stim on (ms)
Time from lum change (ms)
Cue on Cue offStim on Luminance change
a
Time from stim on (ms)
Time from lum change (ms)
Time from stim on (ms)
Time from lum change (ms)
Attention cue loc
Control
Fig. 2 | DLPFC encoded spatial attention. a The responses of an example DLPFC
neuron.The red line indicates trials when thecue was on the right side and the blue
line indicates trials when thecue was on the left side.The shaded areasaround each
line denote SEM across trials. bThe posterior probability of the attention location
decodedfrom DLPFC pseudo-population ensemble activities.Both the training and
the testing of the LDA decoder used sliding windows of 25 ms at 10ms steps. cThe
posterior probability of the attention location from the decoder that was trained
and testedwith the DLPFCneuronal responses atthe same time (thediagonal linein
b). Significance was assessed with two-tailed paired t-tests (p<0.01 with FDR cor-
rections for multiple comparisons), tested against the data with shuffled cue
location labels. The black solid trace indicates the posterior probability calculated
from the actual data. The black dashed trace indicates the posterior probability
calculated from the shuffled data.The black segments at the top indicate when the
actual data performed significantly better than the shuffled data. Thin gray lines
represent SEM across trials. d.Sameasbexcept that the decoder was trained with
the mean activities at 50–200 ms before the stimulus onset. The black segments at
the bottomindicate whenthe performance was significantly lower thanthe control,
indicating the flip of attention to the opposite side. Significance was assessed with
two-tailed paired t-tests (p<0.01 with FDR corrections for multiple comparisons).
Thin gray lines represent SEM across trials. eThe posterior probability of the
attention location decoded from the fast (black line) and the slow trials (gray line).
The trialswere divided by the median reaction time in each neuron. Thin gray lines
represent SEM across trials.Significance was assessed with two-tailed paired t-tests
(fast versus slow, at p< 0.01 with FDR corrections for multiple comparisons).
Article https://doi.org/10.1038/s41467-022-34084-0
Nature Communications | (2022) 13:6306 4
Content courtesy of Springer Nature, terms of use apply. Rights reserved

We further quantified the contributions of SV and NSV to the OFC
neuronal responses with a linear regression model that contained both
value variables as well as a binary term that indicated the frame loca-
tion. We calculated the coefficient of partial determination (CPD) to
measure how much variance was explained by each variable (see
Methods). The average CPD of SV rose significantly above the baseline
shortly after the stimulus onset (Fig. 4a). In contrast, the CPD of NSV
remained low and was not significantly different from the baseline in
most time bins. The paired t-tests that compared the CPDs between SV
and NSV indicated that the OFC neurons encoded the SV much better
in almost all the time bins after the stimulus onset (p< 0.005 with FDR
correction for multiple comparisons). The frame location was tran-
siently encoded when the frame was presented and after the lumi-
nance change but was not significantly encoded during most of the
stimulus presentation period.
This SV dominance is also observed at the level of individual
neurons. We calculated each neuron’s average responses during the
stimulus period and used them to determine the CPD for SV and NSV
a
-1
-0.5
0
0.5
1
1.5
2
Normalized firing rates (spikes/s)
~0 drop
1 drop
2 drops
4 drops
8 drops
-400 -200 0 200 400 600
Time from stim on (ms)
-400-200 0 200 400
Time from lum change (ms)
Cue on Cue offStim on Luminance change
(n=73)
-0.4
-0.2
0
0.2
0.4
0.6
Normalized firing rates (spikes/s)
-400 -200 0 200 400 600
Time from stim on (ms)
-400 -200 0 200 400
Time from lum change (ms)
Cue on Cue offStim on Luminance change
b
(n=53)
-0.4
-0.2
0
0.2
0.4
0.6
Normalized firing rates (spikes/s)
-400 -200 0 200 400 600
Time from stim on (ms)
-400 -200 0 200 400
Time from lum change (ms)
Cue on Cue offStim on Luminance change
c
Same
value
-1
-0.5
0
0.5
1.5
2
Normalized firing rates (spikes/s)
-400 -200 0 200 400 600
Time from stim on (ms)
~0 drop
1 drop
2 drops
4 drops
8 drops
-400-200 0 200 400
Time from lum change (ms)
d
Cue on Cue offStim on Luminance change
-0.4
-0.2
0
0.2
0.4
0.6
Normalized firing rates (spikes/s)
-400 -200 0 200 400 600
Time from stim on (ms)
-400 -200 0 200 400
Time from lum change (ms)
Cue on Cue off
Stim on Luminance change
e
-0.4
-0.2
0
0.2
0.4
0.6
Normalized firing rates (spikes/s)
-400 -200 0 200 400 600
Time from stim on (ms)
-400 -200 0 200 400
Time from lum change (ms)
Cue on Cue offStim on Luminance change
f
Value size
~0 248
-1.5
-1
-0.5
0
0.5
1
1.5
2
Normalized firing rates (spikes/s)
Salient value
Same value
1
g
~0 2 4 8
Value size
-1.5
-1
-0.5
0
0.5
1
1.5
2
Normalized firing rates (spikes/s)
Cued value
Uncued value
1
j
~0 1 2 4 8
Value size
-0.4
-0.2
0
0.2
0.4
0.6
Normalized firing rates (spikes/s)
h
~0 1 2 4 8
Value size
-0.4
-0.2
0
0.2
0.4
0.6
Normalized firing rates (spikes/s)
k
~0 1 2 4 8
Val ue si ze
-0.4
-0.2
0
0.2
0.4
0.6
Normalized firing rates (spikes/s)
i
~0 1 2 4 8
Value size
-0.4
-0.2
0
0.2
0.4
0.6
Normalized firing rates (spikes/s)
l
Salient
value
Same value
versus
salient value
Cued value
versus
uncued value
Example neuron
Positive-tuned
neurons
Negative-tuned
neurons
(n=73)
(n=73)
(n=73)
(n=53)
(n=53)
(n=53)
1
-1.5
-1.5
Fig. 3 | Attentional modulation of the OFC responses. a The responses of an
example OFC neuron intrials when the same rewards wereassociated with the two
stimuli, aligned to the stimulus onset (left) and the luminance change (right). The
reward sizes are indicated by different shades of gray. Shading areas indicate SEM
across trials. bThe population response of the positively tuned OFC neurons
(n= 73) in trials when the same rewards were associated with the two stimuli.
Shading areas indicate SEM across neurons. cSame as b, but for the negatively
tuned OFC neurons (n=53).dThe responses of the example neuron with the trials
grouped by SV. The reward sizes are indicated by different shades of yellow.
Shading areas indicate SEM across trials. eThe responses of the positively tuned
OFC neurons with the trials grouped by SV. The reward sizes are indicated by
different shades of yellow. Shading areas indicate SEM across neurons. fSame as
(d), but for the negatively tuned OFC neurons. gThe average responses of the
example neuron during the period between the stimulus onset and the luminance
change. The black trace is based on the trials with the same-reward stimulus pairs.
The yellow trace is based on the trials grouped by SV. The error bars indicate SEM
across trials. Two-way ANOVA (group: F
1413
=0.79,p=0.37;value: F
4413
= 40.40,
p≪0.001). hThe average responses of the positively tuned OFC neurons during
the period betweenthe stimulusonset and the luminance change. The blacktrace is
based on the trials with the same-reward stimulus pairs. The yellow trace is based
on the trials grouped by SV. The error bars indicate SEM across neurons. Two-way
ANOVA (group: F
1724
=3.22, p=0.07;value:F
4724
= 77.71, p≪0.001). iSame as (h),
but for the negatively tuned OFC neurons. Two-way ANOVA (group: F
1524
=1.97,
p=0.16; value: F
4524
=46.03,p≪0.001). jSame as g, but trials are grouped by CV
(red) and NCV (purple). Two-way ANOVA (group: F
1690
=0, p= 0.99; value:
F
4690
=21.04
,
p≪0.001). kSame as h, but trials are grouped by CV (red) and NCV
(purple). Two-way ANOVA (group: F
1724
=0.16, p= 0.68; value: F
4724
=104.56,
p≪0.001). lSame as i, but trials are grouped by CV (red) and NCV (purple). Two-
way ANOVA (group: F
1524
=0.01, p=0.92;value:F
4524
=88.54,p≪0.001).
Article https://doi.org/10.1038/s41467-022-34084-0
Nature Communications | (2022) 13:6306 5
Content courtesy of Springer Nature, terms of use apply. Rights reserved

for each value-selective neuron (Fig. 4b). The data points of the neu-
rons spread along the axis of the CPD for SV (SV: 105/349 significant,
p< 0.05, linear regression). Their CPDs for NSV, on the other hand,
clustered within a small range (NSV: 34/349 significant, p< 0.05, linear
regression). The distribution of the neurons’CPD for SV was much
more extended in the range than that for NSV, and the mean of the
CPDs for the SV was larger than that of UCV (t= 6.95, p=1.8e−11, two-
tailed paired t-test) (Fig. 4c).
By contrast, when we regressed the OFC neuronal responses
against CV and UCV, we obtained overall similar CPDs, suggesting that
CV and UCV contributed similarly to the responses (Fig. 4d). Other
than a short period after the stimulus onset and another one after the
luminancechange,thedifferencebetweenCV’sandUCV’sCPDswas
not significant during most of the stimulus period. During the period
when the CPD of CV was larger, there were extra visual inputs, the
frame initially and the luminance change in the end. When we plotted
the neurons’CPDs for CV and those for UCV, they tended to lie closely
to the diagonal (Fig. 4e). Nevertheless, the CPDs for CV were slightly
but significantly higher than those for NCV (t=3.09, p=2.2e−3, two-
tailed paired t-test). These results indicate that the cue and the spatial
attention weakly and transiently modulated the neurons’value tuning.
These results were again consistent across the two individualmonkeys
(Supplementary Fig. 8). Further regression analyses based on the four
combinations of SV, NSV, CV, and NSV also point to the same con-
clusion (Supplementary Fig. 9).
Finally, to confirm that the weak attentional modulation was not
due to a washout effect by the many trials in which CV and UCV had
similar values, we studied the trial conditions in which we would
expect the largest modulation on the neurons’responses by spatial
attention. According to our previous study11, the cue-induced mod-
ulation might be most evident when the cue directed the attention
toward the non-salient picture when the SV was the greatest (8 drops
of juice). In this condition, the cue shifted the attention away from the
most salient picture. This would provide us an opportunity to observe
any potential modulation of the neurons’responses by the cue.
Therefore, in these trials, we compared the neurons’responses when
the cue directed the attention toward the 8-drops-of-juice cue (CV = 8)
and when the cue directed the attention away from it (CV =~0, 1, 2, 4,
or 8 and UCV=8). Again, we looked at the positively and negatively
tuned OFC neurons separately. When the attention was on the 8-drops-
of-juice stimulus, the cue and the reward salience were consistent, and
the OFC neurons largely ignored the value of the less salient stimulus
(linear regression, H
0
:slope=0:p= 0.44, positively tuned neurons;
p= 0.30, negatively tuned neurons). The responses were highest for
the positively tuned neurons and lowest for the negativelytuned ones,
both similar to their responses when a pair of 8-drops-of-juice stimuli
were presented (Fig. 5a, b). When the cue directed the attention away
from the 8-drops-of-juice stimulus, it lowered the responses of posi-
tively tuned neurons very slightly (F
1723
=3.74, p= 0.05) and failed to
increase the responses of the negatively tuned neurons (F
1524
= 1.18,
p= 0.28). A complete shift of responses toward the less salient sti-
mulus would have produced responses similar to those when a pair of
the less salient stimuli were presented (black trace in Fig. 5a, b).
Attending away from the most salient stimulus did not abolish the
-400 -200 0 200 400 600 -400 -200 0 200 400
Time from lum change (ms)
0
0.01
0.02
0.03
0.04
CPD
Attention cue loc
Salient value
Non-salient value
-400 -200 0 200 400 600 -400 -200 0 200 400
Time from lum change (ms)
0
0.01
0.02
0.03
0.04
CPD
Attention cue loc
Cued value
Uncued value
0 0.1 0.2 0.3 0.4
CPD of salient value
0
0.1
0.2
0.3
0.4
CPD of non-salient value
(p = 1.8e-11)
0 0.1 0.2 0.3 0.4
CPD
0
10
20
30
40
50
Number of units
0.074
Sig salient value
Non-sig salient value
0
20
40
60
80
100
Number of units
0.014
Sig non-salient value
Non-sig non-salient value
0 0.1 0.2 0.3 0.4
CPD of cued value
0
0.1
0.2
0.3
0.4
CPD of uncued value
(p = 2.2e-3)
0
10
20
30
40
50
Number of units
0.054
Sig cued value
Non-sig cued value
0 0.1 0.2 0.3 0.4
CPD
0
20
40
60
80
100
Number of units
0.043
Sig uncued value
Non-sig uncued value
Time from stim on (ms)
Cue on Cue offStim on Luminance change
Time from stim on (ms)
Cue on Cue offStim on Luminance change
a
d
c
f
b
e
（n=349） （n=118）（n=118）
（n=127）
（n=349）（n=127）
Fig. 4 | OFC neuronal responses dominantly encoded the salient value. a OFC
neuron (n= 349) firing rates were regressed against attention cue location, SV, and
NSV. Plotted is the time course of the population average coefficients of partial
determination (CPD). Significance was assessed with two-tailed paired t-tests
(p< 0.005,with FDR corrections for multiple comparisons) compared to a baseline
computed with the CPD between 0 and 200ms before the cue onset averaged
across different regressors. The blue, yellow,and green bars at the top indicate the
significant CPDs of cue location, SV, and NSV, respectively, and the black bar
indicates significant differences between SV and NSV. The error bars indicate SEM
across neurons. bThe CPDs of SV against the CPDs of NSV for individual value-
selective OFC neurons (n= 118). The yellow data points are the neurons with non-
zero coefficients for SV only, the green data points are the neurons with non-zero
coefficients for NSV only, and the black data points are the neurons that encoded
both SV and NSV (one-samplet-test, p< 0.05, without multiple comparisons). Two-
tailed paired t-test was conducted to compare the mean CPDs of SV and NSV of all
the OFC neurons (n=349).cTop: the distribution of the CPDs for SV of the value-
selective OFC neurons (n= 118).Bottom: the distribution of the OFC neurons’CPDs
for NSV. Vertical dashed lines indicate the mean. Filled bars indicate significant
neurons. dSame as a, but for CV and UCV (n=349).eSameas b, but for CV and UCV
(n= 127). fTop: the distribution of the CPDs for CV of the value-selective OFC
neurons (n= 127). B ottom: the distr ibution of the OFC neurons’CPDs for UCV.
Vertical dashed lines indicate the means. Filled bars indicate significant neurons.
Article https://doi.org/10.1038/s41467-022-34084-0
Nature Communications | (2022) 13:6306 6
Content courtesy of Springer Nature, terms of use apply. Rights reserved

dominant representation of SV by the OFC, and the overall neuronal
responses were very similar regardless of whether the cue was on the
most salient stimulus or not. The results were confirmed in the two
individual monkeys (Supplementary Fig. 10).
DLPFC’s value encoding was also dominated by reward salience
It has been suggested that the DLPFC sits downstream of the OFC in
the processing of value information. We wonder whether the dom-
inance of the value encoding in OFC by reward salience may be
inherited by the DLPFC. On the other hand, the DLPFC has a strong
spatial attention signal that is lacking in the OFC (Supplementary
Fig. 11). One may expect that spatial cue and attention may play a more
substantial role in DLPFC’s value encoding.
We carried out the analyses on DLPFC neuronal responses parallel
totheanalysesontheOFC.Theresults,whicharesummarizedinFig.6
and Supplementary Figs. 12–14, indicated that the DLPFC’svalue
encoding was similarly dominated by reward salience, and spatial
attention did not affect DLPFC’s value encoding significantly. The
regression analyses similar to Fig. 4but done to the DLPFC neuronal
responses reveal that the population average CPD of SV was sig-
nificantly above the baseline right after the stimulus onset and lasted
until the end of the trials (Fig. 6a). The CPD of NSV was low, although
initially significant, and did not last until the luminance change.
Importantly, the CPDs of SV were consistently larger than the CPDs of
NSV during the whole stimulus period (Fig. 6a). The CPDs of individual
units for SV were also significantly greater than those for NSV (Fig. 6b,
c, t=7.27, p=1.9e−12, two-tailed paired t-test). The dominance of SV
was strong in monkey G (Supplementary Fig.13g–i). Monkey D’s DLPFC
neurons encoded SV relatively weakly, yet the encoding was still
stronger than that of NSV (Supplementary Fig. 13a–c). In contrast, the
encodings of CV and UCV in the DLPFC were similar, both when in the
combined data (Fig. 6d–f, t=1.56,p= 0.12, two-tailed paired t-test) and
in the individual monkeys (Supplementary Fig. 13d–f, j–l). Therefore,
despite that a strong spatial attention signal is present in the DLPFC
(blue traces in Fig. 6a, d), its modulation of the value encoding
was weak.
Comparison between OFC and DLPFC
Finally, to compare the representations of spatial attention and value
in the OFC and the DLPFC, we plotted the CPD for attention cue
location against the CPD for SV for each neuron from the two areas
(Fig. 7). The CPDs were calculated from regression models similar to
those used in Figs. 4b, 6b, but the average neuronal responses were
computed with the period between the frame cue offset and the
luminance change, during which spatialattention wasmostly drivenby
the top-down mechanism and on the opposite side of the cue. The
CPDs of OFC neurons are largely distributed along the SV axis, while
the CPDs of DLPFC neurons were more evenly distributed along both
axes. Furthermore, for the OFC neurons, the distribution of the neu-
rons’difference between the two CPDs strongly skewed toward SV
(Fig. 7insert, mean ΔCPD = −0.072, p=1.7e−9, two-tailed paired t-test).
In contrast, the DLPFC neurons did not exhibit such a difference (mean
ΔCPD = −0.0031, p= 0.43, two-tailed paired t-test). The analyses con-
firmed that SV dominatedthe OFC responses, while bothSV and spatial
attention were well represented in the DLPFC. Similar results were
observed in individual monkeys (Supplementary Fig. 15).
Discussion
Here, we demonstrated that while both the behavior and the DLPFC
neuronal responses indicated that the monkeys directed their atten-
tion to the cued location, the OFC neuronal responses were never-
theless dominated by stimulus reward salience and only weakly
affected by spatial attention. The results argue against the previously
proposed theory that attention serves a s a selection mechanism for the
OFC’s value encoding when multiple items with different value asso-
ciations are presented11.
The dominance of reward salience that we observed cannot be
attributed to spatial attention, decision making, or eye movements.
First, unlike reward salience, which was constant in a trial, the cue that
indicated the luminance change was on the opposite side of the
luminance change, and the spatial attention needed to switch sides.
Therefore, spatial attention could be clearly identified with the switch
of location and be distinguished. Second, the monkeys had to report
their decision by making a saccade to a target above the fixation point.
The motor preparation signals were thereby dissociated from the
attention and salience signals. Third, the reward that the monkey
received was randomlychosen between the two stimuli. Neither the CV
nor the SVwas the expected rewardoutcome. Lastly, no decisions were
guided by the values of the stimuli, so potential attention shifts
accompanying the value-based decision-making process were mini-
mized with the current behavior paradigm.
Spatial attention in our paradigm was dictated by the attention
cue via a top-down mechanism, while reward salience may create
attention from a bottom-up subcortical circuitry, which includes
~0 12 4
8
Value size
-0.6
-0.4
-0.2
0
0.2
0.4
Normalized firing rates (spikes/s)
(n =73)
Cued value = 8
Uncued value = 8
Same value
~0 12 4
8
Value size
-0.6
-0.4
-0.2
0
0.2
0.4
Normalized firing rates (spikes/s)
(n =53)
ab
Fig. 5 | Spatial attention failed to switch the value encoding in the OFC to the
non-salient stimulus. a The positively tuned OFC neurons’(n= 73) r esponses to an
8-drops-of-juice stimulus (SV) paired witha stimulus associated with~0, 1, 2, 4, or 8
drops of juice (NSV). The red line denotes trials when the attention was toward the
8-drops-of-juice stimulus, and thepurple line denotes trialswhen the attention was
away from the 8-drops-of-juice stimulus. A two-way ANOVA (attention location:
F
1723
=3.74,p=0.05;value:F
4723
=3.45,p=8.4e−3) indicatesa subtle but significant
attentional modulation. A complete switch of the value encoding to NSV would
produce responses close to the black line, which are the neurons’responses to the
trials of a pair of stimuli with the same reward. All error bars indicate SEM across
neurons. bSame as a, but for the negatively tuned OFC neurons (n= 53). Two-way
ANOVA with group (CV = 8/UCV = 8) and value was performed (attention location:
F
1524
=1.18, p=0.28;value:F
4524
=3.68, p=5.8e–3).
Article https://doi.org/10.1038/s41467-022-34084-0
Nature Communications | (2022) 13:6306 7
Content courtesy of Springer Nature, terms of use apply. Rights reserved

superior colliculus, pulvinar, ventral midbrain, and amygdala24,25.
Although reward salience is distinct from the kind of bottom-up
attention based on physical salience, interestingly, we also observed
bottom-up physical salience signals in the OFC. First, the OFC neurons
robustly, although only briefly, encoded the onset of the frame (Fig. 4a,
d, blue trace), which was a salient bottom-up visual signal. Second,
there was a slightly stronger representation of the cued stimulus than
that of the un-cued after the luminance change (Fig. 4d), which,
although subtle, might also capture the attention via a bottom-up
mechanism. Therefore, while the OFC activities were minimally affec-
ted by spatial attention from the top-down source, influences from the
bottom-up sources, no matter whether they were based on physical or
reward salience, were readily reflected in the OFC.
The dominance of reward salience, which was almost all-or-
none, is in stark contrast to how neurons in the visual cortex are
modulated by spatial attention26. The responses of the visual neurons
to multiple stimuli presented in their receptive field can be well
described with normalization models that perform a weighted aver-
age of the responses to each stimulus presented alone. Attention
modulates how the responses are weighted by adding bias toward
the attended stimulus, and the neurons’stimulus preference plays an
important role. Whether attended or not, a preferred stimulus pre-
sented in a neuron’s receptive field has a significant influence on its
responses27. OFC neurons, however, only encode the value of the
most salient stimulus. That is true even for the negatively tuned
neurons, which prefer the non-salient stimulus. Less salient stimuli
are largely ignored by OFC neurons, and spatial attention, when
directed toward a less salient stimulus, does not enhance its encod-
ing very much in the OFC.
One might argue that the weak representation of NSV in the OFC
was because of the task design, in which the reward for a correct
response was randomly chosen between the two stimuli. Yet, the value
that was relevant to the monkeys in this task—the mean value of the
two stimuli—was also not represented by the OFC. For that to be true,
one would expect that SV and NSV be similarly represented. Yet, the
findings of the dominance of SV encoding and the weak attentional
modulation may not generalize to other species. Notably, the OFC in
rodents was reported to be modulated by task context28. Given the fact
that the OFCs in rodents and primates are not evolutionarily homo-
logues, such a discrepancy may highlight an interesting difference
previously unknown between the two species.
Finally, our results suggest that the representations of spatial
attention and reward expectancy are dissociable in the brain. The
modulation effects of attention and reward on neuronal activities are
similar, and many previous studies have not clearly distinguished
attention and reward in the experiment design2. Here, we demonstrated
dissociable representations of reward salience and spatial attention in
the prefrontal cortex. Spatial attention was well represented only in the
DLPFC. The reward salience signal remains stable in both the OFC and
the DLPFC even when spatial attention shifted sides via the top-down
mechanism. Such a representation of the salience signal independent of
top-down control may be necessary for the brain to evaluate the
potential targets toward which spatial attention could be directed.
Future experiments could be designed to investigate this hypothesis.
Methods
Subjects
Two male rhesus monkeys (Macaca mulatta) were used. They weighed
9.3 kg (subject D) and 6.8 kg (subject G) at the beginning of the
training. All procedures followed the protocol approved by the Animal
Care Committee of Shanghai Institutes for Biological Sciences, Chinese
Academy of Sciences (CEBSIT-2021004).
0 0.1 0.2 0.3
CPD of salient value
0
0.05
0.1
0.15
0.2
0.25
0.3
CPD of non-salient value
(p = 1.9e-12)
0
10
20
30
40
50
Number of units
0.045
Sig salient value
Non-sig salient value
0 0.1 0.2 0.3
CPD
0
20
40
60
80
100
120
Number of units
0.011 Sig non-salient value
Non-sig non-salient value
0 0.1 0.2 0.3
CPD of cued value
0
0.05
0.1
0.15
0.2
0.25
0.3
CPD of uncued value
(p = 0.12)
0
10
20
30
40
50
60
Number of units
0.039
Sig cued value
Non-sig cued value
0 0.1 0.2 0.3
CPD
0
10
20
30
40
50
60
Number of units
0.032
Sig uncued value
Non-sig uncued value
a
d
c
f
b
e
(n=141) (n=141)
(n=143)
(n=143)
-400 -200 0 200 400 600
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
CPD
Time from lum change (ms)Time from stim on (ms)
-400 -200 0 200 400
Attention cue loc
Salient value
Non-salient value
Cue on Stim on Cue off Luminance change
(n=406)
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
CPD
Attention cue loc
Cued value
Uncued value
-400 -200 0 200 400 600
Time from lum change (ms)Time from stim on (ms)
-400 -200 0 200 400
Cue on Stim on Cue off Luminance change
(n=406)
Fig. 6 | DLPFC neuronal responses were dominated by the salient value.
aDLPFC neuron (n= 406) firingrates were regressedagainst attention cue location,
SV, and NSV. Plotted is the time course of the population average coefficients of
partialdetermination(CPD). Conventions as in Fig. 4a.bThe CPDs of SV against the
CPDs of NSVfor individual value-selective DLPFC neurons (n= 141).Conventions as
in Fig. 4b. cTop: the distribution of the CPDs for SV of the value-selective DLPFC
neurons (n= 141). B ottom: the distr ibution of the DLPFC neurons’CPDs for NSV.
Vertical dashed lines indicate the mean. Filled bars indicate significant neurons.
dSame as (a), but for CV and UCV (n=406).eSame as b, but for CV and UCV
(n= 143). fTop: the distribution of the CPDs for CV of the value-selective DLPFC
neurons (n= 143). Bottom: the distribution of the DLPFC neurons’CPDs for UCV.
Vertical dashed lines indicate the means. Filled bars indicate significant neurons.
Article https://doi.org/10.1038/s41467-022-34084-0
Nature Communications | (2022) 13:6306 8
Content courtesy of Springer Nature, terms of use apply. Rights reserved

Behavioral task and materials
Head-restrained monkeys were seated in a primate chair facing a
23.6 inch computer monitor at 60cm away from their eyes. The center
of the screen was adjusted to align to the midpoint of the two eyes.
Behavioral tasks were run with the MATLAB based software
MonkeyLogic29. Subjects’eye position and pupil dilation were tracked
with an infrared oculomotor system at a sampling rate of 500 Hz
(EyeLink 1000). Juice was delivered by a computer-controlled
solenoid.
We trained two monkeys to perform a visual detection task
(Fig. 1a). The monkeys had to hold their gaze ata central fixation point
on the screen within a 3° wide window. After maintaining fixation for
0.5 s, a yellow saccade target was presented 6.0° (monkey D) or 7.6°
(monkey G) above the fixation point. The saccade target remained on
the screen until the end of the trial. After 0.5 s, a square frame
(attention cue) 2.4° wide was presented on the left or right (monkey D:
7.0°; monkey G: 7.6°) of the fixation point. After another 0.2 s, two
visual stimuli of 2.4° size appeared on the left and right sides (monkey
D: 7.0°; monkey G: 7.6°) of the fixation point. The stimulus locations
were rotated slightly (23° counter-clockwise) for monkey G to mini-
mize its left-right bias. The attention cue disappeared 0.2 s after the
stimulus onset. After a random variable delay, one of the two stimuli
would change its luminance.
For monkey D, there were 90% of trials (valid-cue trials) the
luminancechangewasontheoppositesideoftheattentioncue(cued
location). In the remaining trials (invalid-cue trials), the change wason
the same side of the cue (un-cued location). In both cases, the monkey
was required toreport the change within0.1–0.4 s by making a saccade
toward the eye movement target to receive a juice reward. The onset
latency of the luminance change was subject to an exponential decay
function (tau = 2.5 s, cut off at 1.4 s) plus 0.4 s.
To encourage monkey G to direct the attention appropriately
according to the cue, it was trained to report only the luminance
change at the cued location (target) but ignore the change at the un-
cued location (distractor). In 80% of trials, the luminance change
occurred at the cued location. Among the rest of 20% of trials, half of
them had a luminance change at the un-cuedlocation only (distractor-
only trials, 10%), and the other half had first a luminance change at the
un-cued location and then another at the cued location (distractor +
target trials, 10%). The latencies of the luminance change at the target
and the distractor locations were subject to their respective expo-
nential decay functions (target: tau = 2.5 s, cut off at 1.9 s; distractor:
tau =1.25 s, cut off at 1.9 s) plus 0.4s. The monkey was rewarded when
they made a timely response to the target in both the target and the
distractor+target trials, and when they held their fixation till the end of
the trial in the distractor-only trials.
The visual stimuli were associated with different amounts of juice
that the monkeys might get for a correct response. There were five
stimuli for each monkey, each was associated with 1 small drop
(0.033ml), 1, 2, 4, and 8 standard drops (0.10, 0.16, 0.29, and 0.55ml)
of juice, respectively. For convenience, we label the stimulus with a
small drop of juice as ~0 standard drops in the figures and use 0 in the
regression models. The stimuli in each trial were randomly selected
from the stimulus set, and their locations (left or right) were coun-
terbalanced. For correct responses, one of the stimuli was randomly
selected, and the monkey would receive the reward associated with
that stimulus. Therefore, neither the cue nor the reward salience
provided information on which stimulus’s associated reward would
be given.
During the initial training, only one stimulus was presented on the
screen, and the monkeys were rewarded with its associated juice
amount if they detected the luminance change correctly. The attention
cue always appeared on the opposite side of the fixation point (always
valid). During the recording sessions, the single-stimulus and double-
stimulus trials were interleaved. There were 7.1% and 10.1% of single-
stimulus trials for monkey D and monkey G, and the rest were double-
stimuli trials.
Surgery and MRI
Before the behavioral training, both monkeys received a chronic
implant of a titanium headpost. After a 2-month recovery, they were
trained to perform the behavioral task until they achieved satisfactory
performance. Then, the monkeys received structural Magnetic Reso-
nance Imaging (MRI) scans for us todetermine the recording chamber
location. After the chamber implant surgery, a manganese-enhanced
MRI scan was conducted to verify the chamber placement. The scans
were carried out in a Siemens 3 T scanner.
During the surgery, the monkeys were sedated with ketamine
hydrochloride (10 mg/kg), and anesthesia was then induced and
maintained with isoflurane gas (1.5–2%, to effect). Body temperature,
heart rate, blood-oxygen concentration, and expired CO
2
were mon-
itored throughout the surgical procedures.
Neuronal recordings
A 2 cm * 1.5 cm chamber was implanted on the surface of the left
(subject G) and the right (subject D) prefrontal cortex, centered 31.5
(subject G) and 27.5 (subject D) mm anterior to the interaural line
(Supplementary Fig. 1). We recorded extracellular single-unit activities
with tungsten microelectrodes (FHC: 0.3–2MΩ;AlphaOmega:
0.5–3MΩ). Each electrode was driven by an independent microdrive
(AlphaOmega EPS) through a stainless-steel guide tub e. The guide tube
was placed within a grid with holes 1 mm apart. The depth of
the penetration was confirmed by the transitions between gray and
0 0.1 0.2 0.3 0.4
CPD of attention cue location
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
CPD of salient value
OFC
DLPFC
-0.2 0.2
Δ CPD
-0.0031
-0.072
0
0.02
0.04
0.06
Proportion of units
0
(n=225)
Fig. 7 | Comparison of the spatial attention and value signals between OFC and
DLPFC neurons. The CPDs of the attention cue location are plotted against the
CPDs of SV for individual OFC (orange) and DLPFC (magenta) neurons. The dotted
line is the diagonal. Only the neurons with non-zero coefficients for the attention
cue location or SV were plotted (n=225,p< 0.01, one-sample t-test without mul-
tiple comparisons). The insert shows the distribution of the CPD differences
between the two CPDs. Vertical dashed lines indicate the mean.
Article https://doi.org/10.1038/s41467-022-34084-0
Nature Communications | (2022) 13:6306 9
Content courtesy of Springer Nature, terms of use apply. Rights reserved

white matter. At mostfour electrodes were lowered at a time. Neuronal
signals were recorded with an AlphaOmega SnR system at a sampling
rate of 44 kHz. We recorded neurons from the OFCand the DLPFC. The
OFC recording locations were between the lateral and medial orbital
sulci in Walker’s areas 11 and 13. The DLPFC neurons were recorded
from the area rostral to the arcuate sulcus (8a and 8b), including both
the dorsal and the ventral bank of the principal sulcus (46d and 46 v).
There were 171 recording sessions overall, 71 from monkey D and
100 from monkey G. Each recording session contained on average
1381 trials (monkey D: 1668 trials; monkey G: 1177 trials), of which 885
were correct trials (monkey D: 905 trials; monkey G: 871 trials).
Neurons with <100 valid-cue (target) trials were excluded from the
analyses.
Voltage signals of putative single neurons were isolated offline
manually with Plexon Offline Sorter (Plexon, Dallas, TX). Neurons with
poor isolation or with a lower than 1Hz response rate were excluded.
There were no additional selection criteria for neurons.
Behavioral analyses
All the behavior analyses were based on the monkeys’performance
during the recording sessions.
Accuracy and reaction time
For monkey D, an eyemovement to the saccade target within 0.1–0.4 s
after the luminance change is considered a hit. Otherwise, it would be
considered as a miss. For monkey G, hits and misses are similarly
defined, except that any eye movements to the saccade target within
the 0.1–0.4 s time window on the distractor’s luminance change are
counted as false alarms. The RT is defined as the time from the lumi-
nance change to the eye-movement initiation in the hit and false alarm
trials.
Neuronal selectivity
To determine how the value information was encoded by the neurons,
linear regressions (fitlm function in Matlab Statistical Toolbox) were
performed for each task variable for each individual unit in three task
epochs:
Rnr,tðÞ=β0,ntðÞ+β1,ntðÞ*variable rðÞ+ε,ð1Þ
where R
n
(r, t) was the average neural response of unit nfor a given trial
rwithin time window t. In the single-stimulus trials, the only relevant
task variable is stimulus value (V
sin
). In the double-stimuli trials, the
task variablesinclude left stimulusvalue (LV), right stimulus value (RV),
difference between LV and RV (LV-RV), CV, UCV, difference between
CV and UCV (CV-UCV), SV, NSV, difference between SV and NSV (SV-
NSV), total value (TV) and attention cue’s location (att cue loc). εwas
independent Gaussian noise. The task epochs were the cue-stimulus
epoch (0.0–0.2 s after the stimulus onset), the early stimulus epoch
(0.2–0.6 s after the stimulus onset), and the late stimulus epoch
(0.0–0.4 s before the luminance change). A neuron is selective to a
variable if β
1,n
is significantly different than 0 (p< 0.005).
A neuron is considered as value selective if it is selective to any
value variables above during any task epochs (cue-stimulus, early sti-
mulus,andlatestimulus).Wefurtherdividevalueselectiveneurons
into positively tuned and negatively tuned neurons. Neurons with
contradictory tuning are not included in Figs. 3,5, and Supplementary
Table1(e.g.,positivelytunedtoonevaluevariablebutnegatively
tuned to another value variable or positively tuned during one epoch
but negatively tuned during another epoch).
Decoding spatial attention
Linear discriminant analyses (LDA, classify function in Matlab Statis-
tical Toolbox) were used to decode spatial attention location from the
neuronal responses of DLPFC neurons. LDA aims to classify an
observation into one of the K classes through modeling the posterior
probability pðB=b∣A=aÞ,whereAis the observation (neural respon-
ses), and Bis the predictor (left attention cue or right attention cue).
The posterior probability is calculated through Bayes’theorem which
requires prior probability p(B) and conditional probability density
functions pðA=a∣B=bÞ. LDA assumes that the probability density
function given each condition is a normal distribution. Because the left
attention cue and the right attention cue were equally likely, the prior
was set to be 0.5.
Before conducting the LDA, we z-scored the responses of each
neuron across all the trials and all time points (−400–600 ms around
stimulus onset and −400–400 ms around luminance change). The
normalized responses were further smoothed in 100ms windows with
a Gaussian kernel (sigma = 50 ms). The LDA was performed using a
sliding window of 25 ms with 10 ms steps.
We constructed pseudo-neuron ensembles as follows. To balance
the contribution from each neuron, we first selected the neurons with
more than 100 trials in both the left attention cue condition and the
right attention cue condition. We randomly chose 100 trials without
replacements from each condition for each neuron and constructed
the confusion matrix X2RM×T×Nwith the sampled trials, where Mis
the number of trials (M= 200, 100 trials in each cue condition), Tis the
number of time bins, and Nis the number of neurons (subject D:
N= 102; subject G: N=93).Toreducenoise,werantheprincipal
component analysis (PCA) and kept the first Pcomponents that cap-
tured at least 70% of the variance for the LDA.
To estimate how spatial attention signals fluctuated with time
(Fig. 2c, e), PCAs were run on the neuron dimension of Xat each time
bin. Only the first Pprincipalcomponentsthatcapturedatleast70%of
the total variance were used to reconstruct the subspace
(Y2RM×T×P). The decoder was trained and tested with the recon-
structed subspace at each time bin. The posterior probability of
attention cue location given the neuronal responses was used to
quantify the decoder performance. The results were based on 200
independent leave-one-out cross-validations. Significance was estab-
lishedby two-sided paired t-tests comparingthe actual data against the
posterior probabilities calculated in the same procedure but with the
shuffled data.
In Fig. 2b, d, a PCA was run on the neuron dimension of the
confusion matrix that combined all time bins (X02RMT ×N). This
ensured that the neural representation at each time bin shared the
same subspace after PCA. Only the first Qprincipal components that
captured at least 70% of thetotal variance were used to reconstructthe
subspace (Y02RMT ×Q). Y0was reshaped into a RM×T×Qmatrix to
enable cross-temporal decoding. The decoder was trained with the
neural responses in one time bin (Fig. 2b: 25 ms time bin, stepped by
10 ms; Fig. 2d: 50–200 ms before stimulus onset), and tested with the
responses in all time bins.
Linear regression models
In Figs. 4a, d and 6a, d, linear regression models were fittoaverage
firing rates calculated with a 25 ms time window, aligned to the sti-
mulus onset and the luminance change, respectively. We constructed
two linear regression models:
Rnr,tðÞ=β0,ntðÞ+β1,ntðÞattention cue loc rðÞ+β2,ntðÞSV rðÞ+β3,ntðÞNSV rðÞ+ε,
ð2Þ
Rnr,tðÞ=β0,ntðÞ+β1,ntðÞattention cue loc rðÞ+β2,ntðÞCV rðÞ+β3,ntðÞUCV rðÞ+ε,
ð3Þ
where Rnr,tðÞis the average firing rate of neuron nin trial rat time t,
and the predictors were attention cue location(coded as −1 or 1), SV (0,
1,2,4,or8),NSV(0,1,2,4,or8),CV(0,1,2,4,or8),UCV(0,1,2,4,or8).
Article https://doi.org/10.1038/s41467-022-34084-0
Nature Communications | (2022) 13:6306 10
Content courtesy of Springer Nature, terms of use apply. Rights reserved

Each predictor was z-scored. The CPD was calculated as:
CPDi,ntðÞ=SSEi,nðtÞSSEall,nðtÞ
SSEi,nðtÞ,ð4Þ
where CPDi,ntðÞis the CPD of variable iin neuron n, SSEi,nðtÞis the
residual sum of squares of the regression model without variable i,
SSEall,nðtÞis the residual sum of squares of the full model. The sig-
nificance of CPDi,ntðÞ is tested against the baseline (paired t-test,
p< 0.005 with FDR correction for multiple comparisons), which is the
CPD calculated with the firing rate in a 200 ms time window before the
attention cue onset averaged across different regressors:
CPDbaseline,n=1
3X
3
i=1
CPDi,nðt0Þ,ð5Þ
TheCPDsinFigs.4b, d, e, f and 6b, d, e, f were calculated similarly
but with the average firing rate between the stimulus onset and the
luminance change. To specifically study the attention from the top-
down source, CPDs in Fig. 7were calculated with the average firing rate
between the frame cue offset and the luminance change.
Individual monkey analyses
All results presented in the main text are based on the combined data
from both monkeys. Analyses based on each monkey individually can
be found in Supplementary Figs. 2, 6, 8, 10, 13, 15 corresponding to
Figs. 2–7in the main text. The results are consistent.
Reporting summary
Further information on research design is available in the Nature
Research Reporting Summary linked to this article.
Data availability
The data used in this study are available at https://doi.org/10.5281/
zenodo.7090240.
Code availability
The custom codes supporting the findings of this study are availableat
https://github.com/tmyang-lab/reward_salience_in_OFC.
References
1. Connor, C. E., Egeth, H. E. & Yantis, S. Visual attention: Bottom-up
versus top-down. Curr. Biol. 14,850–852 (2004).
2. Maunsell, J. H. R. Neuronal representations of cognitive state:
Reward or attention? Trends Cogn. Sci. 8,261–265 (2004).
3. Awh, E., Belopolsky, A. V. & Theeuwes, J. Top-down versus bottom-
up attentional control: a failed theoretical dichotomy. Trends Cogn.
Sci. 16,437–443 (2012).
4. Chelazzi, L., Perlato, A., Santandrea, E. & Della Libera, C. Rewards
teach visual selective attention. Vis. Res. 85,58–72 (2013).
5. Gottlieb,J.,Hayhoe,M.,Hikosaka,O.&Rangel,A.Attention,reward,
and information seeking. J. Neurosci. 34,15497–15504 (2014).
6. Anderson, B. A., Laurent, P. A. & Yantis, S. Value-driven attentional
capture. Proc. Natl Acad. Sci. U. S. A. 108,10367–10371 (2011).
7. Krajbich, I., Armel, C. & Rangel, A. Visual fixations and the compu-
tation and comparison of value in simple choice. Nat. Neurosci. 13,
1292–1298 (2010).
8. Towal, R. B., Mormann, M. & Koch, C. Simultaneous modeling of
visual saliency and value computation improves predictions of
economic choice. Proc. Natl Acad. Sci. U. S. A. 110,
E3858–E3867 (2013).
9. McGinty, V. B., Rangel, A. & Newsome, W. T. Orbitofrontal Cortex
Value Signals Depend on Fixation Location during Free Viewing.
Neuron 90,1–14 (2016).
10. McGinty, V. B. Overt attention toward appetitive cues enhances
their subjective value, independent of orbitofrontal cortex activity.
eNeuro 6,1–19 (2019).
11. Xie, Y., Nie, C. & Yang, T. Covert shift of attention modulates the
value encoding in the orbitofrontal cortex. Elife 7,e31507(2018).
12. Wallis, J. D. & Miller, E. K. Neuronal activity in primate dorsolateral
and orbital prefrontal cortex during performance of a reward pre-
ference task. Eur. J. Neurosci. 18,2069–2081 (2003).
13. Padoa-Schioppa, C. & Assad, J. A. Neurons in the orbitofrontal
cortex encode economic value. Nature 441, 223–226 (2006).
14. Ballesta, S., Shi, W., Conen, K. E. & Padoa-Schioppa, C. Values
encoded in orbitofrontal cortex are causally related to economic
choices. Nature 588,450–453 (2020).
15. Hunt, L. T. et al. Triple dissociation of attention and decision com-
putations across prefrontal cortex. Nat. Neurosci. 21,
1471–1481 (2018).
16. Murray, E. A. & Rudebeck, P. H. Specializations for reward-guided
decision-making in the primate ventral prefrontal cortex. Nat. Rev.
Neurosci. 19,404–417 (2018).
17. Wallis, J. D. & Rich, E. L. Decoding subjective decisions from orbi-
tofrontal cortex. Nat. Neurosci. 7,973–980 (2016).
18. Hayden, B. Y. & Moreno-Bote, R. A neuronal theory of sequential
economic choice. Brain Neurosci. Adv. 2,1–15 (2018).
19. Posner, M. I. Orienting of attention. Q. J. Exp. Psychol. 32,
3–25 (1980).
20. Tremblay, S., Pieper, F., Sachs, A. & Martinez-Trujillo, J. Attentional
Filtering of Visual Information by Neuronal Ensembles in the Pri-
mate Lateral Prefrontal Cortex. Neuron 85,202–215 (2015).
21. Tremblay,S.,Doucet,G.,Pieper,F.,Sachs,A.&Martinez-Trujillo,J.
Single-Trial Decoding of Visual Attention from Local Field Potentials
in the Primate Lateral Prefrontal Cortex Is Frequency-Dependent. J.
Neurosci. 35,9038–9049 (2015).
22. Rigotti, M. et al. The importance of mixed selectivity in complex
cognitive tasks. Nature 497,585–590 (2013).
23. Mante,V.,Sussillo,D.,Shenoy,K.V.&Newsome,W.T.Context-
dependent computation by recurrent dynamics in prefrontal cor-
tex. Nature 503,78–84 (2013).
24. Takakuwa, N., Kato, R., Redgrave, P. & Isa, T. Emergence of visually-
evoked reward expectation signals in dopamine neurons via the
superior colliculus in V1 lesioned monkeys. Elife 6, e24459 (2017).
25. McFadyen, J., Mattingley, J. B. & Garrido, M. I. An afferent white
matter pathway from the pulvinar to the amygdala facilitates fear
recognition. Elife 8,e40766(2019).
26. Maunsell, J. H. R. Neuronal mechanisms of visual attention. Annu.
Rev. Vis. Sci. 1,373–391 (2015).
27. Ni,A.M.,Ray,S.&Maunsell,J.H.R.Tunednormalizationexplains
the size of attention modulations. Neuron 73,803–813 (2012).
28. Nogueira, R. et al. Lateral orbitofrontal cortex anticipates choices
and integrates prior with current information. Nat. Commun. 8,
1–13 (2017).
29. Asaad,W.F.,Santhanam,N.,Mcclellan,S.&Freedman,D.J.High-
performance execution of psychophysical tasks with complex
visual stimuli in MATLAB. J. Neurophysiol. 109,249–260 (2013).
Acknowledgements
We thank John Maunsell for his comments during the preparation of
the paper. We also thank Ruixin Su, Wei Kong, and Lu Yu for their help
in all phases of the study. This work was supported by the National
Science and Technology Innovation 2030 Major Program (Grant No.
2021ZD0203701), National Key R&D Program of China (Grant No.
2019YFA0709504), Shanghai Municipal Science and Technology
Major Project (Grant No. 2018SHZDZX05), and the Strategic Priority
Research Program of Chinese Academy of Science (Grant No.
XDB32070100).
Article https://doi.org/10.1038/s41467-022-34084-0
Nature Communications | (2022) 13:6306 11
Content courtesy of Springer Nature, terms of use apply. Rights reserved

Author contributions
T.Y. conceived the original idea of the study. W.Z. and Y.X. collected the
behavioral and neurophysiological data and performed the analysis.
W.Z. and T.Y. wrote the paper. All authors designed the experiments,
discussed the results and provided the feedback on the paper.
Competing interests
The authors declare no competing interests.
Additional information
Supplementary information The online version contains
supplementary material available at
https://doi.org/10.1038/s41467-022-34084-0.
Correspondence and requests for materials should be addressed to
Tianming Yang.
Peer review information Nature Communications thanks Pieter Roelf-
sema and the other, anonymous, reviewer(s) for their contribution to the
peer review of this work. Peer reviewer reports are available.
Reprints and permissions information is available at
http://www.nature.com/reprints
Publisher’s note Springer Nature remains neutral with regard to jur-
isdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons
Attribution 4.0 International License, which permits use, sharing,
adaptation, distribution and reproduction in any medium or format, as
long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons license, and indicate if
changes were made. The images or other third party material in this
article are included in the article’s Creative Commons license, unless
indicated otherwise in a credit line to the material. If material is not
included in the article’s Creative Commons license and your intended
use is not permitted by statutory regulation or exceeds the permitted
use, you will need to obtain permission directly from the copyright
holder. To view a copy of this license, visit http://creativecommons.org/
licenses/by/4.0/.
© The Author(s) 2022
Article https://doi.org/10.1038/s41467-022-34084-0
Nature Communications | (2022) 13:6306 12
Content courtesy of Springer Nature, terms of use apply. Rights reserved

1.
2.
3.
4.
5.
6.
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:
use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
use bots or other automated methods to access the content or redirect messages
override any security feature or exclusionary protocol; or
share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at
onlineservice@springernature.com