MRI data and preprocessing. MRI acquisition parameters are summarized in Table S2. We included MRI data from three different cohorts: (1) normative diffusion MRI dataset from 178 unrelated subjects from the HCP who underwent a ultra-high-resolution acquisition on a ''Magnetom'' 7T MRI scanner (Siemens Medical Systems, Erlangen, Germany) were obtained from the publicly available S1200 WashU-Minn-Ox HCP dataset1–3; (2) functional resting-state, and diffusion MRI data from 34 binge-prone females recruited by the Stanford Eating Disorders Program on a 3T MRI scanner (Discovery MR750, GE Healthcare, Milwaukee, Wisconsin).
Resting-state fMRI scans from the binge-prone cohort were preprocessed using fMRIPrep 1.2.34. In brief, the preprocessing of the functional image involved skull-stripping, co-registration to the T1 reference image, and head motion and susceptibility distortion corrections. After removal of non-steady state volumes and spatial smoothing with a 6mm FWHM isotropic Gaussian kernel, ICA-AROMA was used to identify motion-related noise components in the BOLD signal5. Framewise displacement (FD) and root mean square variance over voxels of the temporal derivative of time courses (DVARS) were calculated6,7. Global signals were extracted within the cerebrospinal fluid (CSF), white matter (WM), gray matter (GM), and whole-brain masks. XCP Engine 1.0 was used to perform denoising of the preprocessed BOLD output from fMRIPrep, utilizing the estimated confound parameters8,9. This included demeaning and removal of any linear or quadratic trends and temporal filtering using a first-order Butterworth bandpass filter (0.01-0.08 Hz). These preliminary preprocessing steps were then followed by confound regression of ICA-AROMA noise components, together with mean WM, CSF, and global signal regressors. All regressors were bandpass filtered to retain the same frequency range as the data to avoid frequency-dependent mismatch8. Whereas preprocessing was performed on the diffusion MRI data from the binge-prone cohort to prepare the images for probabilistic tractography using the FSL suite10,11, the normative HCP diffusion MRI data had already been preprocessed (with the minimal preprocessing pipeline). The diffusion-weighted images were corrected for motion and geometric distortions using the ‘topup’ and ‘eddy’ functions, similar to that applied in HCP’s preprocessing pipeline. For each subject, diffusion and T1-weighted images were co-registered using boundary-based registration.
Probabilistic tractography. Probabilistic tractography was used to evaluate the connections between LH and hippocampus. The LH mask was defined on the standard T1 MNI152 09c template adapted from CIT168 Subcortical In Vivo Probabilistic Atlas12, while the hippocampi masks were defined using the Harvard-Oxford brain atlas. This registration was performed using Advanced Normalization Tools (ANTs), which consists of two successive steps of linear and nonlinear registration between the subject’s brain and the MNI brain. In a third step, the MNI-defined ROIs were registered to the subject's space. FSL’s Bayesian Estimation of Diffusion Parameters Obtained using Sampling Techniques (BEDPOSTX) was used to conduct Monte Carlo sampling of probability distribution of diffusion parameters at each voxel, accounting for up to three crossing fibers directions within a voxel13. Fiber tracking was performed with FSL’s Probtrackx2, using distance correction and each hippocampal voxel as seed and the LH as target14. A total of 5000 seed points were used to generate streamlines from each seed voxel, and only the streamlines that reached the target were retained for further analysis. The results of Probtrackx are summarized in a map of “streamline probability” and “waytotal”, representing the probability for each seed voxel to reach the target and the total number of streamlines from a given seed that reached the target, respectively. The strength of the connections between seed and target was calculated as a tractography connectivity index (tractography-CI), as defined in a previous study by the formula: log(waytotal)/log(5000 x Vseed)15. The waytotal resulting from the tractography was log-transformed and divided by the log-transformed product of the generated sample streamlines in each seed voxel (5000) and the number of voxels in the respective seed mask (Vseed). The log-transformation increased the likelihood of reaching normality, which was tested using the Shapiro-Wilk test16.
Hippocampal segmentation. Tractography was used to generate a probabilistic map each with voxel’s streamline probability to the LH for all 178 subjects from the normative HCP dataset. Each subjects’ streamline probability map to the LH was transformed to standard MNI 09c space so that they could be averaged and concatenated into a normative weighted average group map of streamline probability between the hippocampal area and the LH across the 178 HCP subjects. We performed this analysis to define the hippocampal subregions in the normative HCP data and then applied these subregions to the binge-prone cohort. We then used k-means to segment group average hippocampus streamline probability maps. This hypothesis-free method uses successive iterations to assign each voxel to one of two clusters, without the application of external spatial constraints. For the case of large inter-voxel similarities in streamline count, the algorithm fails to identify two distinct clusters. Resulting clusters represented normative hippocampal subregions based on its connectivity to the LH in standard MNI 09c space. Finally, we co-registered the normative clusters to the MRI images from our subjects implanted with depth electrodes as well as the binge eating cohort.
Sweet-fat incentive paradigm. Participants (n=9, Table S1) underwent surgical implantation of depth electrodes for neurosurgical epilepsy monitoring. The location of electrode implantation was determined solely based on clinical needs and thus varied across participants. The inclusion criteria for this study were the presence of at least one hippocampal depth electrode. Consent to participate in this study was obtained according to the Declaration of Helsinki and approved by the institutional ethical committee. We have adapted the sweet-fat incentive paradigm, also known as the Milkshake task17 for intracranial electrographic recordings during cued anticipation and consumption of a sweet-fat and a taste-neutral solution, as previously published and described in detail by our group18,19. Data acquisition and experimental task were previously described18. Briefly, neural activity was sampled at 1024 Hz from AdTech electrodes while subjects engaged in a sweet-fat incentive computer-based paradigm (Fig. S1). Each trial in this paradigm began with a 2-sec fixation cross presented on a computer screen - this period is referred to as the pre-stimulus period. This was followed by a 1-sec presentation of an image of a glass of either water or of milkshake, which served as a cue for the solution to be subsequently delivered through a mouthpiece to the subject for consumption. Before the solution was delivered, a 2-sec image of a fixation cross was viewed. The 1-sec presentation of the solution to be delivered and this 2-sec fixation cross period are referred to as an anticipatory period (3-sec). Following the anticipatory period is a 5-sec receipt/consummatory period, consisting of a 3-sec solution delivery period followed by a 2-sec consumption of solution period. Sweet-fat and taste-neutral trials were presented in a randomized order, with a total of 80 to 100 trials evenly split between sweet-fat and taste-neutral conditions. Upon task completion, the participants were asked to rate on a 1-10 Likert scale the quality of the sweet-fat solution (Likert scale, 1–10) and which solution (sweet-fat vs. taste-neutral) they preferred.
Electrode localization. Pre-surgical MRI scan was co-registered to post-surgical CT scan for electrode visualization and localization as described previously20. Locations of depth electrodes within the medial temporal lobe were then examined by one rater with expertise in medial temporal lobe anatomy in neuroimaging (D.A.N.B). Electrodes in direct contact with the hippocampal area were selected for further evaluation. Then, we co-registered the normative hippocampal clusters (i.e., dlHPC and non-dlHPC) that we had previously defined in standard MNI09c template brain to each subject’s native space (Fig. 1D). All hippocampal electrodes (n=54) were labelled according to whether they were in direct contact with dlHPC or not (i.e., non-dlHPC). This localization was performed prior to time-frequency analysis.
Task data preprocessing and analyses. Electrophysiological data was downsampled to 1000 Hz, notch filtered for 60 Hz and 2nd-3rd harmonics, and Laplacian re-referenced in FieldTrip as previously described18,19. Artifact timepoints were defined as voltage values greater or less than the mean signal of all 10-sec trials concatenated, recorded from the same channel plus 4 multiples of its standard deviation. Any trial with at least 1 detected artifact timepoint was excluded. Time-Frequency analysis was implemented using the wavelet toolbox in Matlab. There are three input parameters: 1) minimum frequency, set to 3, 2) maximum frequency, set to 250, and 3) NumVoices, set to 32. The toolbox generates ‘scales’ based on the desired frequency range (defined by min to max frequencies), which then get mapped into frequencies. The trial vector, scales vector, and ‘morl’ are inputs to the cwtft function in matlab, which generates the wavelets and power extraction. Wavelets were first tested on ground truth data with known spectral properties before use on experimental data.
Trial instantaneous power values were normalized to power at the same frequency and channel during the 1-sec prestimulus period across all trials in the same condition (condition specific pre-stimulus normalization). The prestim duration of any trial with at least 1 detected artifact timepoint was excluded from the normalizing distribution (see previous section). Condition specific pre-stimulus normalization was used to account for possible differences in baseline power before stimulus presentation. Results were reproduced using an alternative normalization method whereby power values were normalized relative to the distribution of power in the same frequency and channel, during the entire recording.
Spectral analyses were primarily focused on the anticipation period (1-sec cue, and 2-sec post-cue fixation). Statistical differences in time-frequency power between conditions were calculated using cluster-based permutation testing21. Briefly, this involved calculating a t-statistic in each time-frequency voxel, between the two conditions (sweet fat vs. neutral), thereby generating the observed t-map. The distribution for each voxel was generated by pooling the time-frequency maps from all channels and subjects (trials were averaged to generate a single map per channel). The observed t-map was then compared to a null distribution (shuffled condition labels) of t-maps generated over 1000 paired permutations. A p-value for each voxel was obtained by comparing the observed to the null t-value at the same time-frequency voxel, thereby generating a p-map. Clusters of contiguous voxels with a p < .05 were identified and compared to the null-distribution cluster size. Observed clusters with sizes larger than the 95th percentile of those from the null distribution were considered significant after correction for multiple comparisons.
Evoked potentials. As previously described, we performed single-pulse stimulations at rest using an intracranial electrical waveform generator and switchbox22,23 (MS-120BK-EEG and PE-210AK, Nihon Kohden, Tokyo, Japan). Electrical stimulation was delivered through adjacent pairs of electrodes in biphasic pulses (6mA; 200 μs/phase, 49 trials) at a frequency of 0.5 Hz for a total of 120 seconds. We measured electrical potentials in response to stimulation with a video EEG monitoring system using a sampling rate of 2000 Hz (version WEE-1200, Nihon Kohden). We analyzed the single-pulse stimulation data using custom Matlab scripts. We first applied a high-pass butterworth filter (1 Hz) to exclude slow varying effects and segmented evoked responses time series from recording channels were into 2500 ms epochs time-locked to stimulus onset (500 ms pre-stimulus and 2000 ms post-stimulus). Then, we re-referenced the data to the common average signal, excluding stimulated channels, channels with artifacts, and channels with large, evoked responses, as previously described24. Finally, to exclude potential effects of pre-stimulus signal fluctuations, we applied a baseline correction by subtracting the average signal between 200 ms and 20 ms prior to stimulus onset. To ensure that these preprocessing steps did not introduce a bias, we also provided line traces of bipolar re-referenced and the single-trials raw signal from the recording electrodes (Fig. S5).
Human hippocampal sample. In accordance with the local IRB, a post-mortem sample of the left hippocampal area (Fig. 2C, left panel) was obtained from a whole brain with no known pathologies that had been extracted 24 hours after death and placed in 10% formalin for one day. The sample was perfused and stored in PFA 4%. For the brain-clearing procedure, we extracted a representative dorsolateral hippocampal section (Fig. 2C, mid panel) and transferred the sample to 1M phosphate-buffered saline (PBS). The sample was stored at 4°C until the iDISCO protocol was performed.
Antibody validation protocol. A validation protocol was performed to be sure that the anti-MCH (Phoenix Pharmaceuticals, Inc- USA, H070-47, Lot No. 01629-10) was compatible with the reagents used in the immunolabeling-enabled 3D imaging of solvent cleared organs (iDISCO) protocol25. Slices were obtained on a Vibratome at 60 µM in 1M phosphate-buffered saline (PBS) solution. The free-floating sections were permeabilized for 3 hours with methanol at room temperature and after were rinsed two times with PBS for 20 min and then rinsed with PBS with 2% TritonX-100. The sections were then incubated with permeabilization solution, PBS with 0.2% Triton, during 30 min and blocking solution, PBS with 0.2% TritonX-100, 10% DMSO, and 6% Donkey Serum for 1 hour. The anti-MCH was incubated 1:500 in PTwH (PBS with 0.2% Tween-20, 1% Heparin (10mg/ml), 0.2% Sodium Azide), overnight at -4°C. The samples were rinsed 3 times for 5 minutes and the secondary antibody, Alexa Fluor 647 anti-Rabbit (ThermoFischer Scientific-USA, A-31573) were incubated 1:250 in PTwH and 3% Donkey serum and 0.2% Sodium azide during 1 hour at RT, covered by the light. Finally, the samples were rinsed in PTwH 3 times for 5 minutes and the slices were mounted with DAPI (Vectashiels-VECTOR, USA). Images were acquired with confocal microscopy (data not shown).
iDISCO brain-clearing 3D histology. After antibody validation, we selected a representative dorsolateral section for the iDISCO protocol to confirm our hippocampal area segmentation (Fig. 1A). The section was approximately 1.0 x 0.8 x 0.4 cm and was pretreated with methanol following the iDISCO protocol25 using a modified immunostaining protocol. The sample was rinsed with PBS with 2% TritonX-100 two times for 1 hour. We incubated the sample in permeabilization solution, PBS with 0.2% Triton, for 30 min, and blocking solution, PBS with 0.2% TritonX-100, 10% DMSO, and 6% Donkey Serum, for 1 hour. The sample was incubated with 1:500 anti-MCH in PTwH (PBS, 0.2% Tween-20, 1% Heparin 10mg/ml, 0.2% Sodium Azide) for 10 days, nutating at 37° C. After 10 days, the samples were rinsed 3 times for 5 minutes and then every few hours and left nutating at room temperature overnight. On the next day, the sample was incubated in the secondary antibody, Alexa Fluor 647 anti-Rabbit (ThermoFischer scientific-USA, A-31573), 1:250 in PTwH and 3% Donkey serum and 0.2% Sodium azide at 37°C, nutating for 10 days, covered from light. Following secondary incubation, the sample was again rinsed in PTwH for 2 days, and the iDISCO clearing protocol was followed (https://idisco.info/idisco-protocol/update-history/)
Histological assessment of lateral hypothalamic connections in the dorsolateral hippocampus. The iDISCO brain-clearing 3D histology results were used to confirm the hippocampal area segmentation. The anti-MCH was used to identify orexigenic projections within our hippocampal sample. We assessed whether our sample from the dorsolateral hippocampal subregion, defined based on the higher number of probabilistic tractography streamlines, contained projections expressing MCH, an orexigenic neuropeptide primarily produced in the LH area. First, we manually identified the location of our whole sample in a corresponding coronal slice in the high-resolution MNI 09c template brain (Fig. S7) and then extracted a representative dorsolateral section that overlapped with the dlHPC (Fig. 2C). Following immunolabeling and clearing, the final MCH-stained sample was again manually overlaid to the corresponding coronal slice in the high-resolution MNI template brain and the tractography-based hippocampal probability map of LH area streamlines (Fig. 2D).
Demographics, clinical and behavioral data of binge-prone cohorts. Subjects’ consent was obtained according to the Declaration of Helsinki and approved by the institutional ethical committee (IRB-35204). We analyzed the available clinical and behavioral data from 34 binge-prone females, defined by at least one weekly episode of eating large amounts of food in short periods accompanied by the feeling of loss of control eating over the prior 6 months (i.e., binge-prone cohort; mean age = 26 ± 5.6 years; BMI = 27.9 ± 8.5; binge frequency = 2.7 ± 1.4 episodes/week)26. The number of binge-eating episodes per week was assessed with the Eating Disorder Examination, a standardized diagnostic interview27. The Beck Depression Inventory (BDI) and the Beck Anxiety Inventory (BAI) were used to screen for depression and anxiety, respectively28,29. The Difficulties in Emotion Regulation Scale was used to assess impairment in emotion regulation30. The binge-eating-cohort was divided into two subgroups: (1) lean (n=17): BMI < 25 (referred to as lean group); (2) overweight/obese (n=17): BMI > 25 (referred to as obese/overweight group).
Resting-state functional connectivity analysis. Resting-state functional connectivity analysis was performed on the binge-prone cohort’s preprocessed resting-state fMRI data using DPABI 4.3/DPARSF which is based on Statistical Parametric Mapping (SPM12, https://www.fil.ion.ucl.ac.uk/spm)31. Three subjects were excluded due to excessive movement as measured by 1) mean FD > 0.2mm, 2) more than 20% of FD over 0.2mm, or 3) any FD > 5mm9. A seed-based approach was performed to examine rsFC in the binge-prone cohort (n=34) by calculating the rsFC between the LH mask as defined above and each tractography-identified hippocampal subregion. Functional connectivity values were extracted for all subjects and utilized in further correlational analyses.
Statistical analyses. Statistical analyses were performed using the RStudio Version 1.2.5042 (RStudio, Inc.). Given the sensitivity of metrics derived from resting-state fMRI and diffusion MRI proneness to numerical distortions related to data acquisition or analytical pipeline, we employed the Tukey’s method to remove outliers for each connectivity metric. After checking for normality, we then used the Student’s T-test to compare rsFC as well as tractography-CI between hippocampal subregions and LH in overweight/obese and lean group. One outlier was identified and removed from the obese binge eating group and two outliers were identified and removed from the lean group in both the rsFC and structural CI analyses in Figures 3C and 3D (left panel), respectively. Additionally, we identified and removed one outlier from the lean group in the CI analysis in Figure 3D (right panel). Mann–Whitney U Test was used to compare the corrected number of streamlines between the LH and hippocampal subregions in the binge-prone cohort. To assess potential effect of confounders in the identified connectivity differences between the obese and lean group, we fit a multivariate logistic regression model to predict whether a subject belongs to the overweight/obese or lean group including the available demographic and behavioral variables in addition to the LH-dlHPC connectivity measurements. The comprehensive list of variables included: age, depression (BDI), anxiety (BAI), binge frequency, restrained eating, emotional eating and externally driven eating scores (from DEBQ), LH-dlHPC-LH rsFC, LH-non-dlHPC rsFC, LH-dlHPC structural connectivity, and LH-non-dlHPC structural connectivity. We then used backwards elimination to identify which combination of variables provided highest predictive power with the lowest total number of explanatory variables to avoid over-fitting (Akaike Information Criterion). Finally, a variance inflation factor was calculated to assess potential correlations between explanatory variables32. Significance was defined by p < 0.05 for all tests.
Data availability. Anonymized data that support the findings of this study as well as the code and materials used for analyses are available from the corresponding author upon reasonable request.
1. Van Essen, D. C. et al. The Human Connectome Project: a data acquisition perspective. NeuroImage 62, 2222–2231 (2012).
2. Glasser, M. F. et al. The minimal preprocessing pipelines for the Human Connectome Project. NeuroImage 80, 105–124 (2013).
3. Sotiropoulos, S. N. et al. Advances in diffusion MRI acquisition and processing in the Human Connectome Project. NeuroImage 80, 125–143 (2013).
4. Esteban, O. et al. fMRIPrep: a robust preprocessing pipeline for functional MRI. Nat. Methods 16, 111–116 (2019).
5. Pruim, R. H. R. et al. ICA-AROMA: A robust ICA-based strategy for removing motion artifacts from fMRI data. NeuroImage 112, 267–277 (2015).
6. Power, J. D., Barnes, K. A., Snyder, A. Z., Schlaggar, B. L. & Petersen, S. E. Spurious but systematic correlations in functional connectivity MRI networks arise from subject motion. NeuroImage 59, 2142–2154 (2012).
7. Power, J. D. et al. Methods to detect, characterize, and remove motion artifact in resting state fMRI. NeuroImage 84, 320–341 (2014).
8. Ciric, R. et al. Benchmarking of participant-level confound regression strategies for the control of motion artifact in studies of functional connectivity. NeuroImage 154, 174–187 (2017).
9. Parkes, L., Fulcher, B., Yücel, M. & Fornito, A. An evaluation of the efficacy, reliability, and sensitivity of motion correction strategies for resting-state functional MRI. NeuroImage 171, 415–436 (2018).
10. Smith, S. M. et al. Advances in functional and structural MR image analysis and implementation as FSL. NeuroImage 23 Suppl 1, S208-219 (2004).
11. Andersson, J. L. R., Skare, S. & Ashburner, J. How to correct susceptibility distortions in spin-echo echo-planar images: application to diffusion tensor imaging. NeuroImage 20, 870–888 (2003).
12. Pauli, W. M., Nili, A. N. & Tyszka, J. M. A high-resolution probabilistic in vivo atlas of human subcortical brain nuclei. Sci. Data 5, 180063 (2018).
13. Behrens, T. E. J., Berg, H. J., Jbabdi, S., Rushworth, M. F. S. & Woolrich, M. W. Probabilistic diffusion tractography with multiple fibre orientations: What can we gain? NeuroImage 34, 144–155 (2007).
14. Jenkinson, M., Beckmann, C. F., Behrens, T. E. J., Woolrich, M. W. & Smith, S. M. FSL. NeuroImage 62, 782–790 (2012).
15. Tschentscher, N., Ruisinger, A., Blank, H., Díaz, B. & Kriegstein, K. von. Reduced Structural Connectivity Between Left Auditory Thalamus and the Motion-Sensitive Planum Temporale in Developmental Dyslexia. J. Neurosci. 39, 1720–1732 (2019).
16. Royston, P. Approximating the Shapiro-Wilk W-test for non-normality. Stat. Comput. 2, 117–119 (1992).
17. Stice, E., Spoor, S., Bohon, C. & Small, D. M. Relation between obesity and blunted striatal response to food is moderated by TaqIA A1 allele. Science 322, 449–452 (2008).
18. Huang, Y. et al. The insulo-opercular cortex encodes food-specific content under controlled and naturalistic conditions. Nat. Commun. 12, 3609 (2021).
19. Kakusa, B. et al. Anticipatory human subthalamic area beta-band power responses to dissociable tastes correlate with weight gain. Neurobiol. Dis. 154, 105348 (2021).
20. Liu, S. & Parvizi, J. Cognitive refractory state caused by spontaneous epileptic high-frequency oscillations in the human brain. Sci. Transl. Med. 11, eaax7830 (2019).
21. Cohen, M. X. Analyzing Neural Time Series Data: Theory and Practice. (The MIT Press, 2019).
22. Keller, C. J. et al. Mapping human brain networks with cortico-cortical evoked potentials. Philos. Trans. R. Soc. B Biol. Sci. 369, 20130528 (2014).
23. Shine, J. M. et al. Distinct Patterns of Temporal and Directional Connectivity among Intrinsic Networks in the Human Brain. J. Neurosci. 37, 9667–9674 (2017).
24. Miller, K. J., Müller, K.-R. & Hermes, D. Basis profile curve identification to understand electrical stimulation effects in human brain networks. PLoS Comput. Biol. 17, e1008710 (2021).
25. Renier, N. et al. iDISCO: a simple, rapid method to immunolabel large tissue samples for volume imaging. Cell 159, 896–910 (2014).
26. American Psychiatric Association. Diagnostic and Statistical Manual of Mental Disorders. (American Psychiatric Association, 2013). doi:10.1176/appi.books.9780890425596.
27. Fairburn, C. G. & Cooper, Z. The Eating Disorder Examination (12th edition). in Binge eating: Nature, assessment, and treatment 317–360 (Guilford Press, 1993).
28. Beck, A. T., Ward, C. H., Mendelson, M., Mock, J. & Erbaugh, J. An Inventory for Measuring Depression. Arch. Gen. Psychiatry 4, 561–571 (1961).
29. Beck, A. T., Epstein, N., Brown, G. & Steer, R. A. An inventory for measuring clinical anxiety: psychometric properties. J. Consult. Clin. Psychol. 56, 893–897 (1988).
30. Gratz, K. L. & Roemer, L. Multidimensional Assessment of Emotion Regulation and Dysregulation: Development, Factor Structure, and Initial Validation of the Difficulties in Emotion Regulation Scale. J. Psychopathol. Behav. Assess. 26, 41–54 (2004).
31. Yan, C.-G., Wang, X.-D., Zuo, X.-N. & Zang, Y.-F. DPABI: Data Processing & Analysis for (Resting-State) Brain Imaging. Neuroinformatics 14, 339–351 (2016).
32. Johnston, R., Jones, K. & Manley, D. Confounding and collinearity in regression analysis: a cautionary tale and an alternative procedure, illustrated by studies of British voting behaviour. Qual. Quant. 52, 1957–1976 (2018).