Hypothalamic interaction with reward-related regions during subjective evaluation of foods

The reward system implemented in the midbrain, ventral striatum, orbitofrontal cortex


Introduction
The reward-related brain regions in humans work as a complex network to calculate the subjective values of presented options and make a choice based on these assigned values and to generate reward prediction error for reinforcement learning ( Delgado et al., 2000 ;Elliott et al., 2000 ;Berns et al., 2001 ;Knutson et al., 2003Knutson et al., , 2005McClure et al., 2003 ;O'Doherty et al., 2003O'Doherty et al., , 2004Haruno et al., 2004 ;Seymour et al., 2004 ;Rolls et al., 2008 ;Lebreton et al., 2009 ;Haber and Knutson, 2010 ;Daw et al., 2011 ;Bartra et al., 2013 ;Clithero and Rangel, 2013 ;Drummond and Niv, 2020 ). Numerous studies have revealed specific functions preformed in each of the brain regions in the socalled reward system. The dopaminergic midbrain region is associated the VMPFC compares multiple reward alternatives ( Elliott et al., 2008 ;Lim et al., 2011 ;Strait et al., 2014 ).
Accumulating evidence has suggested that autonomic factors modulate the reward system via the hypothalamus. In decision-making for rewards, interoceptive states such as hunger and thirst influence decision processing ( Bouret and Richmond, 2010 ;Berrios et al., 2021 ;Eiselt et al., 2021 ). The hypothalamus, a central hub of autonomic functions, which consists of numerous nuclei that perform different functions related to food, is involved in such interoceptive states. While the lateral hypothalamic area (LHA) is a feeding center and the ventromedial nucleus of the hypothalamus (VMH) acts as a satiety center in the classical neuroscientific view ( King, 2006 ;Stuber and Wise, 2016 , for review). Recent animal studies have shown that the arcuate nucleus (ARC) and paraventricular nucleus of the hypothalamus (PVH) are critical for the control of food intake and metabolism ( Balthasar et al., 2005 ;Krashes et al., 2014 ;Li et al., 2019 ). The ARC integrates metabolic signals from the peripheral circulation and projects to both the LHA and PVH ( Atasoy et al., 2012 ;Saper and Lowell, 2014 ). Previous studies of animal anatomy have revealed the anatomical projections from the LHA and PVH to the reward-related regions, such as the ventral tegmental area in the midbrain and the nucleus accumbens in the VS ( Geerling et al., 2010 ;Watabe-Uchida et al., 2012 ;Bonnavion et al., 2016 ). While these animal anatomy studies have suggested that the hypothalamus sends autonomic signals to reward-related regions, it remains unclear how hypothalamic nuclei functionally interact with reward-related brain regions during processing of dietary reward.
To explore the effective interaction between the human hypothalamic nuclei and reward-related regions, in this study, human participants were administered with functional magnetic resonance imaging (fMRI). Higher-resolution images were collected during rest to parcellate the medial hypothalamus into individual nuclei performing different autonomic functions, based on an analytic workflow of areal parcellation in our previous studies ( Osada et al., 2017 ;Ogawa et al., 2020 ). We also measured brain activity when the participants were asked to evaluate food rewards and choose certain or probabilistic options ( Huettel et al., 2006 ;Christopoulos et al., 2009 ;Levy and Glimcher, 2011 ) ( Fig. 1 ). The task involved reward evaluation as a whole, which is thought to activate the reward-related regions. The task also employed probabilistic options to elucidate the effect of risk preference on reward evaluation implemented in the reward-related regions. The brain activity in individual hypothalamic nuclei was examined during evaluation of food rewards, in comparison with monetary rewards, and the neural pathways toward the reward-related regions were then explored.

Participants
Twenty-seven right-handed students participated in this study. Two participants were excluded from the following analyses for their excessive head motion during the task, using a criterion of the mean frame-wise head motion above 0.1 mm. The mean frame-wise head motion was calculated in the human connectome project (HCP) preprocessing pipelines and saved in a file (Movement_RelativeRMS_mean.txt) ( Finn et al., 2015 ;Ogawa, 2021 ). Therefore, the data of twenty-five participants (11 men and 14 women, age: 22.0 ± 3.3 years (mean ± standard deviation) ranging from 20 to 34 years) was included in the analyses. They underwent fMRI scans in the evening, before taking supper. Written informed consent was obtained from all participants in accordance with the Declaration of Helsinki. The Institutional Review Board of the Juntendo University School of Medicine approved the experimental procedures. This study was not preregistered and the analyses and results reported here should be regarded exploratory.

Fig. 1. A reward task used in this study.
Two alternative rewards were compared and chosen. Two options were presented for 2 s, one offering a certain outcome and the other offering a probabilistic outcome. Two pie charts indicate the probability of success at the trial, and the two rectangles indicate the quantity of successful reward. During the presentation of a fixation cross for 1.5 s, the participants were asked to press either the left or right button, corresponding to their choice. The feedback was then presented for 0.5 s that showed the participant's choice. There was an inter-trial interval for a jittered duration of 4, 6, or 8 s. There were two types of rewards: food and money.

Behavioral procedures
The participants performed a reward task, where two alternative rewards were compared and chosen ( Levy and Glimcher, 2011 ) ( Fig. 1 ). In each trial, the two options were presented for 2.0 s, one offering a certain outcome of a fixed amount and the other offering a probabilistic outcome of several amounts, in the certain and probabilistic options, respectively. During the presentation of a fixation cross for 1.5 s, the participants were asked to press a corresponding button (left or right) for their choice. The feedback that showed the participant's choice was then presented for 0.5 s. There was an inter-trial interval for a jittered duration of 4.0, 6.0, or 8.0 s. There were two types of rewards: food and money. For Food condition, the presented food was a set of popular multi-colored buttonshaped chocolates or a set of disc-shaped light-salted crackers. In the chocolate trials, five pieces were presented as the certain options, whereas 10, 20, 40, or 65 pieces were presented as the probabilistic options. In the cracker trials, two pieces were presented as the certain options, whereas 4, 8, 16, or 26 pieces as the probabilistic options. For Money condition, 200 yen was presented as the certain options, whereas 400, 800, 1600, or 2600 yen was presented as the probabilistic options. The probabilities of successful receipt in these probabilistic options were 75%, 50%, 33%, or 15% for both Food and Money conditions, presented in an equal number of trials for all the amount-probability combinations. After the fMRI scans were completed, the participants received actual food and money rewards from one food trial and one money trial, selected randomly from the entire trials. For the probabilistic options, the amount of rewards to be received was also determined after the fMRI scans.
Eight trials were presented in each of four Food blocks and four Money blocks. Four fMRI runs were administered, where Food and Money blocks were alternated. The order of Food/Money blocks was counterbalanced across participants. The participants performed 128 trials in each of Food and Money conditions in total. We used PsychoPy software ( Peirce et al., 2019 ) to present the visual stimuli and record the responses.

Behavioral analysis
Risk preference for each individual participant for each Food/Money condition, was determined as follows. First, the expected subjective value (ESV) of the probabilistic option for t -th trial of the i -th participant was defined as below.
where q t is the reward amount, p t is the reward probability, r i is the i -th participant's risk preference. A value of r i = 1 indicates that the participant is risk-neutral, a value of r i < 1 indicates risk-aversive, and a value of r i > 1 indicates risk-seeking. The probability of choosing the probabilistic option in the t -th trial of the i -th participant was modeled using a softmax function as below.
where c t is the reward amount of the certain option, and the beta controls the balance between the certain and probabilistic options. Individual risk preferences, as well as the beta values, were estimated by fitting P i,t to the participants' choices, using the Broyden-Fletcher-Goldfarb-Shanno (BFGS) method equipped in the optimization toolbox of MAT-LAB (MathWorks, Inc., MA, USA).

MRI procedures
All MRI data were acquired using a 3-T MRI scanner at Juntendo University Hospital (Siemens Prisma, Erlangen, Germany) with a 32ch head coil. T1-and T2-weighted structural images were obtained using 3D magnetization-prepared rapid gradient-echo (MPRAGE) and 3D sampling perfection with application-optimized contrasts using flip angle evolution (SPACE) sequences (resolution = 0.8 × 0.8 × 0.8 mm 3 ). Gradient-echo field map images were also obtained to correct the distortion of the structural images. Functional images with higher resolution of 1.25 mm isotropic voxels were obtained using a multiband gradient-echo echo-planar imaging (EPI) sequence (time of repetition [TR] = 2.3 s, echo time [TE] = 20 ms, flip angle = 73°, in-plane field of view [FOV] = 180 mm × 180 mm, matrix size = 144 × 144, 108 contiguous slices with no gap, phase encoding direction = posterior to anterior (PA), parallel acquisition factor = 2, and multiband factor = 3) ( Feinberg et al., 2010 ;Xu et al., 2013 ). Two images with one AP and one PA encoding direction were acquired using the spin-echo field map sequence before each fMRI run. These images were used to perform the topup distortion correction for functional images ( Andersson et al., 2003 ). A single band image with the same size and resolution as the multiband EPI was acquired at the start of each fMRI session as a reference for motion correction and co-registration to structural images .
On the first day, the participants performed the task for four fMRI runs and then were scanned during the resting state for five fMRI runs. On the second day, the participants were scanned during the resting state for ten fMRI runs. Each task fMRI run contained 240 volumes, and each resting-state fMRI run contained 200 volumes. In total, 960 volumes (37.8 min) and 3,000 volumes (115 min) were acquired for task fMRI and resting-state fMRI, respectively. The days elapsed between the two scans were 5.9 ± 5.5 days (mean ± standard deviation).

Parcellation analysis of the hypothalamus
The HCP preprocessing pipeline  was used, with modifications, to achieve a higher resolution. The functional images were realigned, topup-distortion-corrected, and spatially normalized to the standard space of Montreal Neurological Institute (MNI) coordinates. The resting-state images were used only for the parcellation analyses. For resting-state images, the time series data were projected onto the standard surface (32,492 vertices in each hemisphere) ( Van Essen, 2005 ;Glasser et al., 2013 ). The time series data of the resting state were highpass filtered (cut-off = 2000 s) and cleaned using the ICA + FIX method ( Salimi-Khorshidi et al., 2014 ). MSMAll refined the surface registration using structural images and resting-state images ( Robinson et al., 2014( Robinson et al., , 2018. Global signals were additionally regressed out using the mean of greyordinate time series. Parcellation analyses based on boundary mapping ( Margulies et al., 2007 ;Cohen et al., 2008 ;Buckner et al., 2011 ;Eickhoff et al., 2015 ;Laumann et al., 2015 ;Poldrack et al., 2015 ;Glasser et al., 2016a ;Gordon et al., 2016 ;Osada et al., 2017Osada et al., , 2021Gratton et al., 2018 ;Ogawa et al., 2018Ogawa et al., , 2020Suda et al., 2020 ) were applied to the medial hypothalamus in higher resolution images, similarly to our previous studies ( Osada et al., 2017 ;Ogawa et al., 2020 ). Each voxel in the hypothalamus of each participant, excluding voxels in the mammillary body, was used as a seed to calculate its correlations with the vertices in the cerebral cortex, and the vertex-wise correlation coefficients were transformed to Fisher's z (i.e., correlation map in the cerebral cortex). The similarity of the spatial patterns of the correlation maps for each voxel in the hypothalamus was then evaluated using the transformed correlation coefficients (i.e., similarity map in the hypothalamus). Spatial gradients of the similarity maps were computed for each seed voxel (i.e., gradient map). The gradient maps of the individual participants were averaged to generate a group gradient map. After spatial smoothing (full width at half-maximum [FWHM] = 1.25 mm) of the group gradient maps in the hypothalamus, excluding the third ventricle of each side, a three-dimensional watershed algorithm ( Vincent and Soille, 1991 ) was applied to the gradient maps, and the binary watershed maps were averaged across the seed voxels (i.e., probabilistic boundary map). A high probability indicates that the boundaries of nuclei are likely to exist. A three-dimensional watershed algorithm was again applied to the probabilistic boundary map to parcellate the hypothalamus into subregions. Given the symmetrical localization of the left and right subregions, the names of the hypothalamic nuclei were labeled to each of the subregions based on the atlases of the human hypothalamus ( Baroncini et al., 2012 ;Mai et al., 2015 ;Osada et al., 2017 ;Neudorfer et al., 2020 ).

Regions of interest
Five bilateral ROIs in the hypothalamus were defined for analysis, which are known to be related to the control of food intake ( Morton et al., 2006 ;Hill et al., 2008 ;Simpson et al., 2009 ;Saper and Lowell, 2014 ): the arcuate nucleus of the hypothalamus (ARC), dorsomedial nucleus of the hypothalamus (DMH), lateral hypothalamic area (LHA), paraventricular nucleus of the hypothalamus (PVH), and ventromedial nucleus of the hypothalamus (VMH) (see Table 1 for MNI coordinates).
To determine the ROIs in the brain outside the hypothalamus associated with reward processing, we used the term-based meta-analysis system on Neurosynth (neurosynth.org/analyses/terms/) ( Yarkoni et al., 2011 ). The system automatically generated the association test map for the term "reward" from 922 registered studies. We found the top five peaks with a z-score threshold (z > 7.0): the midbrain, orbitofrontal cortex (OFC), combined left and right ventral striatum (VS) and ventromedial prefrontal cortex (VMPFC). We call these four regions "rewardrelated regions " in this study. Cortical ROIs of the OFC and VMPFC were defined as 6-mm radius spheres. Subcortical ROIs of the midbrain, VS were defined as 4-mm radius spheres. We used different sizes of cortical and subcortical ROIs because of their anatomical volume differences.

Brain activation and effective connectivity
Preprocessing was conducted in a way similar to the restingstate dataset, up to the MNI normalization. After spatial smoothing (FWHM = 4 mm), a general linear model was applied to each voxel using SPM12 ( http://www.fil.ion.ucl.ac.uk/spm/ ). Two events of interest (stimulus presentation in Food and Money conditions), together with nuisance events (choice of a certain option, choice of a probabilistic option and feedback presentation), were coded and modeled into a general linear model as transient events convolved with a canonical hemodynamic response function. The model included six head motion parameters derived from realignment as covariates of no interest. Time-series data were filtered using a high-pass filter (cutoff: 64 s). The magnitude images for individual participants were contrasted between Food and Money conditions. The brain activity related to the subjective expected value of reward was also extracted using two parametric modulation regressors (Food and Money) in a separate general linear model. For a group analysis, contrast images were entered into a two-tailed onesample t-test using a random effect model.
We also performed psychophysiological interaction (PPI) analyses with the seeds of LHA and PVH that showed differential brain activation during Foods vs. Money. The voxel-wise PPI in the whole brain from the two hypothalamic ROIs was calculated using the generalized PPI toolbox ( McLaren et al., 2012 ). Five events (two events of interest, Food and Money and the three nuisance events), timeseries of the ROI (LHA or PVH) and five psychophysiological interactions, together with nuisance regressors (six parameters of head motion), were included into the general linear model. The voxel-wise beta values of the PPI in each of the four ROIs in the reward-related regions were averaged using MarsBaR ( Brett et al., 2002 ) and subjected to group analysis. We also calculated PPIs among the four ROIs using the same procedure.
Based on the PPI results, we further investigated the directionality of the connectivity between the hypothalamic region and the rewardrelated regions using DCM . We considered ROIs comprising the PVH, Midbrain and VMPFC, since we obtained significant PPIs between the PVH and Midbrain/VMPFC (Food > Money) (see Results). We constructed nine models considering the unidirectional and bidirectional modulation of Food on the connectivity between the PVH and Midbrain/VMPFC. All models included driving inputs of Option (combined Food and Money) to the PVH and Midbrain and driving inputs of Food to the PVH; endogenous connectivity between the PVH, Midbrain, and VMPFC; recurrent connectivity within each region. The fixed-effects (FFX) Bayesian model selection (BMS) implemented in SPM was performed to identify the best model for the participants, and Bayesian parameter averaging (BPA) was used to obtain the parameters of the best model ( Stephan et al., 2009 ).

Behavioral results
Participants performed a reward task, in which two options (certain and probabilistic) were evaluated in Food or Money condition ( Levy and Glimcher, 2011 ) ( Fig. 1 ). The participants' reaction times were 460 ± 23 (mean ± standard error of means) ms and 453 ± 22 ms in Food and Money conditions, respectively, and there was no significant difference (t(24) = 1.3, P = 0.21, two-tailed paired t-test). The participants chose the certain options by 72 ± 5 % and 68 ± 6 % in Food and Money conditions, respectively, and there was no significant difference (t(24) = 1.7, P = 0.11, two-tailed paired t-test). The participants' risk preference, the tendency to choose an option that involves high variance in potential outcomes ( Levy and Glimcher, 2011 ), was 0.64 ± 0.06 and 0.69 ± 0.06 in Food and Money conditions, respectively. There was no significant difference (t(24) = -1.5, P = 0.16, two-tailed paired t-test), with significant correlation between them ( r = 0.86. P < 0.001) (Fig. S1), suggesting similar risk-taking strategies in these conditions.

Brain activity related to expected subjective value
The task in the present study was taken from the previous study by Levy and Glimcher (2012) , where the brain activity related to the subjective expected value of reward, expected value with participant's risk preference taken into consideration, was estimated using parametric modulations, with food and money combined. In the previous study, significant brain activity was reported in the VMPFC, but not in the midbrain, OFC or VS. In this study, the brain activity in the VMPFC approached a significance level (t(24) = 1.5, P = 0.16, two-tailed onesample t-test). However, when the ROI is placed in the activation peak reported in the previous study using the same task ( Levy and Glimcher, 2011 ), the brain activity in the VMPFC was significant (t(24) = 3.6, P = 0.0014, two-tailed one-sample t-test). Since very few previous reward-related studies have reported significant brain activity in all of the four regions, to seek the generality of reward-related tasks and the greater statistical power, we employed Neurosynth, which yielded the four regions for subsequent analyses.

Parcellation of the hypothalamus to localize individual nuclei
To localize the hypothalamic nuclei, parcellation analyses based on boundary mapping Gordon et al., 2016 ) were applied to higher-resolution (1.25 × 1.25 × 1.25 mm 3 ) functional images. The analysis yielded a probabilistic boundary map ( Fig. 2 A), where a high probability indicates that the boundaries of nuclei are likely to exist. The hypothalamus was parcellated into subregions based on the probabilistic boundary map. As shown in Fig. 2 B, the subregions were located symmetrically in the left and right hypothalami. Five bilateral ROIs in the hypothalamus were defined as those known to be related to the control of food intake ( Morton et al., 2006 ;Hill et al., 2008 ;Simpson et al., 2009 ;Saper and Lowell, 2014 ): the ARC, DMH, LHA, PVH and VMH.
To confirm the validity of the ROIs representing individual medial hypothalamic nuclei, we compared the locations of the nuclei of the present study and those of Baroncini et al. (2012) . After linear transformation of the MNI coordinates of the medial hypothalamic nuclei reported in Baroncini et al. (2012) (see Fig. S2), the locations of the nuclei were compared ( Fig. 2 C). We further quantified the distance of every combination of the nuclei between the two studies in a sagittal plane to demonstrate the excellent spatial correspondence of the nuclei (Table S1).

Brain activity in the hypothalamic nuclei
We examined differential transient brain activity in the hypothalamus time-locked to presentation of the option stimuli (Food > Money) ( Fig. 3 A). A two-way repeated-measures analysis of variance with laterality (left and right) and nuclei (ARC, DMH, LHA, PVH and VMH) as the main effects was performed. No significant main effect was observed (laterality: F(1, 24) = 0.19, P = 0.66; nuclei: F(4, 96) = 1.2, P = 0.31). Therefore, we explored the differential brain activity in each of the five nuclei in the left-right combined ROIs. Fig. 3 B shows the brain activity in Food and Money conditions in the five ROIs. Overall, the reward evaluation of Food and Money elicited a decrease in brain activity relative to the fixation baseline, as has been commonly observed in previous neuroimaging studies of the hypothalamus ( Smeets et al., 2005 ;Teeuwisse et al., 2012 ;Page et al., 2013 ;Jastreboff et al., 2016 ;Osada et al., 2017 ). Significant differential (Food > Money) brain activity was observed in the PVH (t(24) = -2.9, P = 0.0076, P = 0.038 after correction for five comparisons, two-tailed paired t-test) and LHA (t(24) = -3.0, P = 0.0056, P = 0.028 after correction for five comparisons, two-tailed paired t-test). We conducted additional analyses of the brain activity with (1) a more conservative smoothing kernel of 2.5 mm in FWHM and (2) the inclusion of a nuisance regressor of frame-wise displacement (FD) (Fig. S3). These results showed similar patterns of brain activity, with slightly higher values under the 2.5 mm smoothing kernel or inclusion of the FD regressor. We subsequently highlight the two hypothalamic nuclei in terms of their interactions with the cortical and subcortical reward-related regions.

Effective connectivity between the hypothalamic nuclei and the reward-related regions
The five peaks in the brain outside the hypothalamus associated with reward processing were determined, based on the term-based metaanalysis system on Neurosynth ( Yarkoni et al., 2011 ). We generated the association test map for the term "reward" (Fig. S4): the midbrain, OFC, VMPFC and combined left and right VS (see Table 2 for MNI coordinates). Differential brain activity (Food > Money) was not observed in these four ROIs (all P > 0.42, two-tailed paired t-test). Fig. S5 shows the brain activity during reward evaluation of Food/Money.
It is reasonably assumed that the reward-related regions receive food-specific direct/indirect inputs from the hypothalamus. We Overall, the reward evaluation of Food and Money elicited decreased brain activity relative to the fixation baseline, consistent with previous neuroimaging studies of the hypothalamus. * : P < 0.05, * * : P < 0.01, * * * : P < 0.001. examined effective connectivity by calculating PPIs (i.e., functional connectivity that is differential in Food vs. Money conditions) between the hypothalamic nuclei (LHA and PVH) and the reward-related regions (midbrain, OFC, VMPFC and VS), with the hypothalamic nuclei as seeds. Fig. 4 A and B shows maps of the PPIs with the seeds in the LHA and PVH, respectively. No significant PPI was observed in these ROIs with the seed in the LHA ( Fig. 4 C), whereas significant PPI (Food > Money) with the seed in the PVH was observed in the midbrain (t(24) = 2.6, P = 0.014, two-tailed paired t-test) and VMPFC (t(24) = 3.0, P = 0.0062, P = 0.050 after correction for eight comparisons, two-tailed paired t-test) ( Fig. 4 D).
We conducted an additional gPPI analysis with FD included as a nuisance regressor. As shown in Fig. S6, the results remained almost unchanged. We next examined correlations between functional connectivity and risk preference across participants. In Food condition, significant correlation was observed between PVH-midbrain functional connectivity and risk preference ( r = 0.51, P = 0.0087, P = 0.034 after correction for four comparisons) ( Fig. 4 E). However, no significant correla- Correlation between PVH-midbrain functional connectivity and risk preference in Food condition. One dot represents one participant. F. Correlation between PVH-midbrain functional connectivity and risk preference in Money condition. G. Correlation between PVH-VMPFC functional connectivity and risk preference in Food condition. H. Correlation between PVH-VMPFC functional connectivity and risk preference in Money condition. tion was observed in PVH-midbrain connectivity in Money condition ( r = 0.25, P = 0.23) ( Fig. 4 F), PVH-VMPFC connectivity in Food condition ( r = 0.14, P = 0.51) ( Fig. 4 G), or PVH-VMPFC connectivity in Money condition ( r = 0.34, P = 0.095) ( Fig. 4 H). We conducted a posthoc power analysis for Fig. 4 E to examine the validity of the sample size (N = 25). The power (1 -beta) was 0.81 above the standard criterion (0.8), indicating that the sample size was reasonably acceptable. We further conducted an additional correlation analysis with FD included as a nuisance regressor. As shown in Fig. S7, the results remained almost unchanged.
To analyze the network structure inside the reward-related regions related to food evaluation, the PPIs were further calculated for each pair in the four ROIs in the reward-related regions (midbrain, OFC, VMPFC and VS). No significant PPIs (Food > Money) were found within the reward-related regions, suggesting that the reward-related regions are not food specific. We next investigated the functional connectivity in each pair of the four reward-related regions, common to Food and Money by averaging the functional connectivity related to Food and Money. Significant functional connectivity between the midbrain and VMPFC was found with the midbrain as a seed (t(24) = 2.5, P = 0.021, two-tailed one-sample t-test), and with the VMPFC as a seed (t(24) = 3.9, P = 0.00061, two-tailed one-sample t-test) (Fig. S8).
We further investigated the directionality of connectivity between the PVH and Midbrain/VMPFC. We compared nine dynamic causal modeling (DCM) models using the FFX BMS analysis ( Fig. 5 A). Model 9 was selected as the best model concerning the posterior probability ( Fig. 5 B). Model 9 included the bidirectional modulation of Food on the connectivity between the PVH and Midbrain/VMPFC. Model 3 was the 2nd best with the moderate posterior probability ( Fig. 5 B), including the unidirectional modulation of Food on the connectivity from the PVH to Midbrain/VMPFC. We also performed the random-effects BMS, accommodating differences in the best models of the participants. The results indicate that Model 9 won the selection and Model 3 was the 2nd (Fig.  S9), similarly to the FFX BMS analysis. The results of the BPA analysis of Model 9 showed that the modulation of Food on the connectivity from the PVH to the Midbrain/VMPFC were significant (Bayesian probability p > 0.9), but the modulations in the opposite direction were not significant ( Fig. 5 C). The results of the BPA analysis are summarized in Table 3 . These results indicate causal interaction from the PVH to Midbrain/VMPFC in response to food input. We also performed the BPA of second-place Model 3 that showed moderate posterior probability. The modulation of Food on the connectivity from the PVH to Midbrain/VMPFC were significant (Bayesian probability p > 0.9) ( Fig. 5 C), similarly to Model 9.

Discussion
By analyzing the individual nuclei of the human hypothalamus, we found differential brain activity in the LHA and PVH during subjective evaluation of food relative to money. Effective connectivity analyses further revealed the causal interaction from the PVH to the midbrain region and VMPFC during subjective evaluation of food. The functional connectivity between the PVH and midbrain was correlated with risk preference of the participants. These results suggest dual hypothalamic pathways from the PVH to the midbrain region and to the VMPFC in the human reward-related regions that may modulate reward evaluation for decision making.
The hypothalamic nuclei (parcels) are much smaller than cerebrocortical parcels. As shown in Table 1 , the PVH consisted of only 5 voxels, whereas surrounding voxels consisted of 18 to 21 voxels, demonstrating that such surrounding voxels (boundary voxels) occupied the vast majority of the hypothalamus. Therefore, although they are usually neglected in the cerebrocortical parcels, the boundary voxels seem worth utilizing in calculating the brain activity in the hypothalamic parcels. Specifically, in the PVH proper of 5 voxels, two consecutive voxels constituted the parcel in most the directions. In this case, the smoothing kernel of two voxels (2.5 mm) in FWHM would be appropriate (cf., Glasser et al., 2016b ). On the other hand, in taking the boundary voxels into account, three consecutive voxels (i.e., two proper voxels plus two half boundary voxels on both sides) would constitute the parcel, in which case FWHM of three voxels (approximately 4 mm) would be appropriate. In Fig. S3, we actually compared two smoothing kernels of 2.5 and 4 mm and obtained similar activation results; the larger smoothing kernel yielded more stable results without altering the basic characters of the conservative results. This study contains several limitations to be addressed. First, this study was not preregistered, and there could be several potential analyses and subtests, which may be an issue for multiple comparisons. Multiple comparisons were applied only when there are multiple ROIs/conditions to be tested in Figs. 3 and 4 . Second, we have not collected information about body mass index, hours since last meal, the composition of last meal, and subjective hunger state, which may have influenced reward circuit activity. Third, resting state data was comprised of data collected across two days, under potentially different conditions. For example, in contrast to resting-state scans on day 2, the resting-state scans on day 1 were collected after task, and residual task effects may be included in the resting state data ( Gordon et al. 2014 ). However, that the resting data were used only for the parcellation analyses, and the parcellation results seem reasonably validated in Fig. 2 C and Table S1. In addition, we would also like to acknowledge potential difficulties of fMRI connectivity from the hypothalamus, including small size, signal distortion in this area, and proneness to artifacts from respiration and heartbeat. As for the correlation analyses presented in Fig. 4 E, due to the small sample size (N = 25), the results should be treated as exploratory and require follow-up/replication in the future.
The present study examined the hypothalamic activity based on the contrast of Food vs. rewarding non-Food stimuli (i.e., Money). Another possible contrast would be to use non-rewarding Food stimuli (i.e., unfa- Results of Bayesian parameter averaging of the winning and 2nd place models. Thick lines and arrows indicate the significant effects (Bayesian probability p > 0.9), while dotted lines and arrows indicate non-significant effects. We have summarized the estimated parameters in Table 3 . vorite food) as a control. The former contrast used in the present study was designed to reveal the food-related activity of the hypothalamus as a whole. On the other hand, the latter contrast (favorite Food vs. unfavorite Food) may have revealed a more specific component of hypothalamic functions that promotes feeding behavior ( Saper and Lowell, 2014 ). The PPI in the LHA in the present study is not different between Food and Money but might have been different between favorite and unfavorite Food.
It has been revealed that the ARC integrates the metabolic signals from the peripheral circulation and projects to the LHA, DMH, PVH and VMH ( Balthasar et al., 2005 ;Saper and Lowell, 2014 ), and that the LHA and PVH project to the reward-related extra-hypothalamic regions, such as the ventral tegmental area in the midbrain and the nucleus accumbens in the VS ( Watabe-Uchida et al., 2012 ;Stuber and Wise, 2016 ). Our results of the brain activity decrease observed in the LHA and PVH ( Fig. 3 ) are consistent with the processing stream from the ARC to LHA/PVH in the hypothalamic nuclei ( Balthasar et al., 2005 ;Watabe-Uchida et al., 2012 ;Saper and Lowell, 2014 ;Stuber and Wise, 2016 ), suggesting that the food-related signals are refined in the output nuclei (i.e., the LHA and PVH) of the hypothalamus.
Although differential brain activity during food processing relative to money was observed in the LHA, the connectivity of the LHA with the reward-related regions was not different in this present study ( Fig. 4 ). Animal studies have shown projections from the LHA to various rewardrelated regions such as the ventral tegmental area in the midbrain, nucleus accumbens in the VS, and the VMPFC. It is also known in humans that the functional connectivity of hypothalamus to the reward-related regions may be involved in craving for rewarding materials including foods ( Zhang et al., 2018( Zhang et al., , 2019. One possibility of such non-differential connectivity is that the LHA is a relatively large and functionally heterogeneous region that may comprise multiple subregions and a greater number of types of neurons. It is known that the PVH sends output to neurons in the ventral tegmental area in the midbrain region ( Watabe-Uchida et al., 2012 ;Beier et al., 2015 ;Hung et al., 2017 ). More specifically, the oxytocinergic and vasopressinergic neurons in the PVH project primarily to the GABAergic inhibitory neurons, which then project to the dopaminergic neurons in the ventral tegmental area in the midbrain ( Beier et al., 2015 ). Thus, a decrease in the PVH activity would effect an activity increase in the dopaminergic neurons in the midbrain (Fig. S5), which may release dopamine in the VS in response to reward cues. The effective connectivity between the PVH and midbrain regions in this study ( Fig. 5 and Table 3 ) is consistent with these studies. The dopaminergic neurons in the ventral tegmental area are associated with anticipation of reward amounts ( Schultz et al., 1997 ;O'Doherty et al., 2006 ;D'Ardenne et al., 2008 ), which suggests that the pathway modulates dopaminergic reward processing based on autonomic states such as hunger and thirst ( Minamimoto et al., 2009 ;Bouret and Richmond, 2010 ;Eiselt et al., 2021 ). The modulation may enhance the anticipation of rewards represented in the ventral tegmental area ( Watabe-Uchida et al., 2012 ;Beier et al., 2015 ;Hung et al., 2017 ), which may explain the risk preference of participants ( Levy and Glimcher, 2011 ) ( Fig. 4 E), the tendency to seek seemingly greater rewards irrespective of the probability of being realized.
On the other hand, effective connectivity was also observed between the PVH and VMPFC in the present study. These results are consistent with a rodent study demonstrating that the medial frontal cortex mediates autonomic impact on feeding behavior ( Eiselt et al., 2021 ). Interestingly, however, very few direct projections have been reported between the PVH and VMPFC ( Ongür et al., 1998 ). One possible explanation is that there is a direct pathway from the PVH to VMPFC in a speciesdependent manner, and it has been demonstrated that oxytocin application to the human VMPFC modulates decision-making ( Aoki et al., 2015 ). An alternative possibility is that, due to the polysynaptic nature of the effective connectivity of fMRI data, the PVH sends output to the VMPFC in a polysynaptic manner, presumably via the ventral tegmen-tal area ( Haber and Knutson, 2010 ). The effective interaction between the PVH and VMPFC, whether the projections are direct or not, may modulate the comparison of subjective values based on autonomic states ( Elliott et al., 2008 ;Lim et al., 2011 ;Strait et al., 2014 ). We suggest that the autonomic information from the PVH influences decision-making processes via the midbrain and VMPFC by modulating reward prediction and comparison of subjective values.

Data and code availability
The data and code that support the findings of this study are available at Dryad ( https://doi.org/10.5061/dryad.6q573n620 ), except for raw image data. The raw image data cannot be deposited in a public repository because sharing raw image data was not included in the informed consent. Any additional information required to reanalyze the data reported in this paper is available from the corresponding author upon reasonable request.