The Crunchometer System and Workflow for Acoustic Feeding Analysis.

A) Schematic of the Crunchometer setup for behavioral feeding studies. This diagram illustrates the system’s integrated components for the acquisition and analysis of acoustic feeding data in mice. The inset details the microphone’s optimized placement, within the outer wall of the 11-hole box, and the precise locations of the food pellets. B) Exploded view of Crunchometer components. The acrylic box (1) in which a High-fat diet (HFD) (2) and Chow (3) pellets were positioned on the inner wall of the behavioral box (See Video 1 for pellet preparation), at the same height as the condenser microphone (4). The microphone is mounted on the adjacent outer wall. A bottom-view camera (5) is installed beneath the box to record mouse locomotor activity (an optional top-view camera is not shown). A contact lickometer (6), controlled by an Arduino device or MedAssociate system (7), provides 1 µL of 10% sucrose per lick. C) Sound recording and processing workflow. Audio was sampled at 44.1 kHz, and video was recorded at 30 frames per second (fps). The spectrogram (1 s resolution) is also shown, with a color bar indicating power in dB. D) Power spectrum filtering. The power spectrum of the recorded audio was averaged over 500-950 Hz band to isolate potential bite sounds. The average dB in this band is plotted. A fixed threshold of -85 dB (determined by trial and error) was used in our setup. Every 1 s bin exceeding the threshold was assigned a value of 1; otherwise, it was assigned a value of 0. E) Identification of bite frames and feeding bouts. Black lines indicate putative bites, obtained through signal binarization and labeled as bite frames. Asterisks mark representative examples of feeding bouts, with red lines indicating bout onset and green lines indicating bout offset. Gray boxes represent 1-second time bins. Each feeding bout consists of one or more bites, separated by pauses of less than 5 seconds. Feeding bouts were detected within two defined Regions of Interest (ROIs) on the video recording, where the area corresponding to each pellet was outlined (yellow box for Chow, pink box for HFD). F) After identifying feeding bouts, the software automatically classified each video snippet by sorting them into four folders: “Chow,” “HFD,” “Gnawing,” or “Artifact.” The primary classification was determined by motion energy within two regions of interest (ROIs) drawn over the food: a yellow square for the Chow pellet and a pink square for the HFD pellet. For example, if motion energy was greater in the Chow ROI, the snippet was classified as ’Chow.’ Snippets were labeled ’Gnawing’ if visual inspection showed mouth movements without food consumption, such as biting the plastic cap that holds the pellet (or any other non-edible object). Finally, a snippet was labeled as ‘Artifact’ if a noise occurred while the mouse was outside both food ROIs. G) Human validation is a critical step to correcting automated classification errors. For example, the system may misclassify a snippet as “feeding” simply because the mouse is positioned within a feeding region of interest (ROI), even if it isn’t eating. The examples highlighted by the red arrows show such cases. During the manual review, a user corrects these errors by moving the misclassified snippets from the diet folders (Chow or HFD) to their proper category, such as “artifact” or “gnawing.” H) The final ethogram displays feeding bouts identified by two methods: the supervised Threshold method and a Support Vector Machine (SVM). In the graph, colored lines represent feeding on Chow pellets (yellow), HFD pellets (pink), and gnawing behavior (black). While the SVM method is more efficient at distinguishing bite-like sounds and less prone to artifacts, this precision comes at the cost of underestimating gnawing behavior. In Supplementary Fig. 1-1, the bill of materials is provided. Supplementary Fig. 1-2 shows the robustness of Crunchometer bite detection to additive white noise. The Threshold method is sensitive to loud sounds overlapping the bite-related frequency band (500–950 Hz) but remains more reliable than the SVM method across the tested signal-to-noise ratio (SNR) range.

The Threshold and SVM methods were more precise and reliable than humans at detecting the start and end of a feeding bout; however, these methods tend to fragment feeding bouts more than humans do, as they estimate more events and longer feeding periods.

A) Representative ethograms illustrate the feeding behavior of two fasted mice. Red and yellow marks denote feeding bouts related to HFD and Chow pellets, respectively. A horizontal dashed black line separates human-annotated feeding bouts from method-detected feeding bouts. The experiment in the top panel (M1-Saline) demonstrates greater consistency between human observations and Crunchometer detections, whereas the black arrows in the bottom panel (M6-Saline) highlight two feeding bouts that were uniquely detected by either mathematical methods or by human observers. Also, the asterisk (*) indicates periods in which a human failed to detect a feeding bout correctly. B) To evaluate the similarity between the automated methods and human detections, Normalized Mutual Information (NMI) was used, with varying shades of orange indicating different NMI values. While both human observers and the methods showed high within-group correlation, there was notably reduced similarity between human and method detections (highlighted by the blue rectangle). C) To assess the overlap probability between the Crunchometer methods and human observers, we first concatenated all feeding bouts detected across the six experiments. Subsequently, we quantified the total number of feeding bouts identified by human observers and by each Crunchometer method. Solid bars represent the overlap probability that an SVM and seven human observers agreed on the detection of a single feeding bout, relative to the 267 total bouts detected by the Threshold method across all six mice. Empty bars display the overlap probability between Threshold and SVM detections relative to each of the seven human observers (with total feeding bouts detected by each observer being 251, 286, 228, 248, 583, 215, and 444 for human 1 through 7, respectively). Horizontal black lines indicate the mean ± SEM. D) Time delay in detections by SVM and Human observers relative to Threshold-based onsets. The probability distributions of detection times by the SVM (top panels) and Human observers (bottom panels) are shown relative to the start (left panels) and end (right panels) of Threshold-based detections. The solid black line indicates the Threshold onset, while the red dashed line depicts the mean detection delay. Delay values are presented as mean ± std. E) Scatter plots illustrate the number and size of feeding bouts detected by human observers and the Crunchometer methods. Points falling on the diagonal dashed line indicate similar detection rates between human and automated methods. Deviations below the line (lower right) suggest greater detection by the Crunchometer methods, while deviations above the line (upper left) signify more detections by human observers. The left panels compare human detections against the threshold-based method, and the right panels compare them against the SVM-based method. Errors indicate the mean ± SEM. F) Total number (top) and size (bottom) of feeding bouts detected by human observers and the Crunchometer methods. Bars represent the mean for each detection source, with error bars indicating the mean ± SEM. G) Pearson coefficient of determination, R-squared, of mouse intake with bout size and bout number from automated methods and human observers. These scatter plots display the correlation between food intake and either feeding size or the number of bouts across all experiments. The black dashed line represents the linear regression excluding one subject (mouse number 5), whereas the green solid line indicates the regression including all subjects. For human detections, error bars are the mean ± SEM.

The Crunchometer captures detailed differences in feeding microstructure between satiated and fasted states.

A) Representative feeding ethograms from a satiated (Fed) and a food-deprived (Fasted) mouse. Lines indicate bouts of HFD consumption (pink), standard Chow consumption (yellow), and non-edible gnawing (black). The lower panel shows the cumulative feeding time (only for Chow and HFD pellets, gnawing is excluded) over a two-hour session for the fed (green) and fasted (purple) groups (n = 6 mice). Shaded areas indicate the standard error of the mean (± SEM). B) The top panel shows the mean Inter-Bout Interval (IBI) in 10-minute bins, illustrating the progression toward satiety. The bottom panel shows the corresponding feeding rate; red dots indicate a significant difference between fed and fasted states. The scatter plot (right) shows a significant correlation between total feeding time and food intake (R² = 0.49, p = 0.011). C) Quantification of feeding parameters, including total feeding time, number of bouts, bout size, latency to first bite, and mean IBI. Data are expressed as mean ± SEM. *p < 0.05, **p < 0.01, ***p < 0.001.

Semaglutide suppresses appetite and reduces the preference for a high-fat diet.

A) Feeding bouts and cumulative intake in fasted mice following subcutaneous administration of saline (control group, Ctrl). B) The same parameters were measured in the same mice 24 hours later, under fasting conditions, following subcutaneous semaglutide administration (Sem group). C) Feeding behavior recorded 24 hours after semaglutide administration under fed conditions (Post-Sem group). In the ethograms, pink lines represent HFD pellet consumption, yellow lines indicate pellet consumption, and black lines denote gnawing events. In cumulative intake plots, solid lines show group averages (n = 6 mice), and shaded areas represent the ± SEM. All three experimental conditions (Ctrl, Sem, and Post-Sem) were tested sequentially in the same animals, with 24-hour intervals between sessions, as indicated by the arrow in the experimental timeline. D) Percentage change in the body weight of mice. The mice were weighed prior to the onset of the Crunchometer test. The black arrow indicates the administration of semaglutide. The percentage change was calculated by subtracting the body weight recorded on the saline administration day (Ctrl) from the weights on subsequent days (semaglutide and post-semaglutide), then dividing by the baseline (Ctrl) weight and multiplying by 100. Asterisks indicate statistically significant differences between Ctrl vs D1, D4 and D5 Post-sem. Data are expressed as mean ± SEM. *p < 0.05, **p < 0.01, ***p < 0.001. E) Bar plots of food intake, feeding time, and number of bouts for Chow pellet (yellow bars) and HFD pellet (pink bars) across the three experimental groups. The sum of the yellow and pink bars represents the total food intake. Asterisks indicate statistically significant differences between experimental groups (Ctrl (Sal/Fasted), Sem (Sem/Fasted), and Post sem (Post-semaglutide/Fed)), while the hash symbol (#) denotes significant differences between Chow and HFD pellets within the same group. F) Quantitative analysis of feeding variables measured using the Crunchometer: latency to the first bite, bout size, IBI, and intake of a 10% sucrose solution. Data are expressed as mean ± SEM. *p < 0.05, **p < 0.01, ***p < 0.001. G) Average IBIs and feeding rate calculated in 10-minute bins across the session. For the feeding rate, red dots indicate a significant difference between groups (two-way ANOVA, time effect: F(11,165) = 1.386, p = 0.1837, Treatment effect: F(2,15) = 7.310, p = 0.0061, and interaction: F(22,165) = 0.8404, p = 0.6721; post hoc analysis: Ctrl vs Post sem time bins = 0.17 h, 0.33 h, 1 h (p < 0.05); Ctrl vs Sem time bins = 0.33 h, 1 h (p < 0.05). Asterisks indicate statistically significant differences between Ctrl vs Sem (Sal/Fasted vs Sem/Fasted), while a hash symbol denotes significant differences between Ctrl vs Post sem (Sal/Fasted vs Post-semaglutide/Fed). The scatter plot in the right panel shows the correlation between feeding time and total food intake (R² = 0.57, p = 0.00031). A group analysis, as shown here, can be performed using the “short.mat” files generated by the Crunchometer software, via Tab 6 (Group analysis). The “short.mat” files need to be in a folder named after each experimental group.

Bilateral chemogenetic activation of GABAergic neurons in the LH promotes spillage and gnawing behavior.

A) Schematic of viral infection and representative immunofluorescence images (right panels) showing expression of red fluorescent protein (mCherry, red) and nuclear labeling with 4ʹ,6-diamidino-2ʹ-phenylindole (DAPI, blue) in neuronal somata within LH of Vgat-cre mice. The white scale bar in the lower right corner represents 10 μm. Cumulative feeding time in B) fed (control group, Ctrl) and C) fasted (Sal) mice following intraperitoneal administration of saline. D) Feeding behavior in fed mice following intraperitoneal injection of Clozapine-N-oxide (CNO), a ligand for hM3D(Gq). In the feeding ethograms, pink lines represent HFD pellet consumption, yellow lines indicate Chow pellet consumption, and black lines denote gnawing events. In cumulative feeding time plots, solid lines show group averages (n = 4 mice), and shaded areas represent the ± SEM. E) Average IBIs and feeding rate were calculated in 10-minute bins across the session (top and middle panels, respectively) (two-way ANOVA, time effect: F(11,108) = 4.942, p < 0.0001, Treatment effect: F(2,108) = 15.83, p < 0.0001, and interaction: F(22,108) = 2.262, p = 0.0030; post hoc analysis: Fed Saline vs. Fed CNO time bins = 0.17 h, 0.33 h, 0.5 h, All p’s < 0.05; Fasted Saline vs. Fed CNO time bins = 0.17 h, 0.33 h, 0. 5 h, 0.83 h; All p’s < 0.05). In the feeding rate measure, red dots indicate a significant difference between groups. * Indicate statistically significant differences between Ctrl vs CNO (Saline/Fed vs CNO/Fed), and a # symbol denotes significant differences between Sal vs CNO (Saline/Fasted vs CNO/Fed). Bottom panel: Scatter plots display correlations between feeding time and food intake with or without spillage (left and right panels, respectively). Red lines indicate the correlation considering all groups (No spillage R² = 0.32, p = 0.5782; Spillage R² = 0.79, p = 0.0001), while blue lines denote the correlation excluding the CNO group (No-Spillage R² = 0.55, p = 0.0354; Spillage R² = 0.51, p = 0.0462). F) Quantitative analysis of feeding behavior, including total food intake, spillage, gnawing bout size, number of feeding bouts, bout size, latency to the first bite, IBIs, and intake of a 10% sucrose solution. Data are shown as mean ± SEM. *p < 0.05, **p < 0.01, ***p < 0.001. Unilateral chemogenetic DREADD activation of LH GABAergic neurons (n=3 mice) produced similar results (Supplementary Fig. 5-1).

The Cruchometer and electrophysiological responses of LH neurons in freely behaving mice.

A) To record electrophysiological responses with the Crunchometer, a pulse generator system synchronizes the Crunchometer data with neuronal recordings. Video and audio signals captured by the Crunchometer are aligned with the electrophysiological recordings using a pulse generator, whose output is simultaneously recorded through the electrophysiology analog input (red line) and used to blink an LED (blue line), which is captured frame by frame by the camera during the session. B) Example of LED blinking used to synchronize video and audio signals with electrophysiological recordings. Frames showing the LED off (top) and on (bottom) were subtracted from the video to generate a signal resembling the analog input; these frames correspond in time to the horizontal black lines in panel A. In Supplementary Fig. 6-1, there are more details of how to build a 1 Hz pulse generator to blink an LED. C) Representative activity of LH neurons during feeding behavior. The firing rates of individual LH neurons were calculated (top plots) and aligned to feeding bouts (bottom plots). The top panel exhibits a neuron that increased its firing rate during feeding bouts (red bars), whereas the bottom panel displays a neuron that decreased its firing rate during the feeding bouts (blue bars). Feeding bouts for Chow and HFD pellets are marked by yellow and pink squares, respectively; licking behavior is indicated by blue squares. The inset depicts the action potential waveform in two representative neurons. D) Population activity from four LH neuronal recording sessions. Top plots display feeding bouts in yellow and pink for Chow and HFD pellets, respectively; bottom plots exhibit the z-scored neuronal population activity of the LH across sessions. Red and blue arrows highlight the representative neurons in panel C that were selectively activated or inhibited during food intake, respectively.

Microendoscope calcium imaging of LH GABAergic neurons during liquid and solid food intake.

A) Two pilot experiments are shown (left and right panels). Left: GCaMP7s fluorescence images displaying ROIs of individual GABAergic neurons in the LH. Scale bar = 20 µm. Right: Mouse trajectories over a 30-minute session with access to Chow pellets, HFD pellets, and a licking chamber delivering water or sucrose solutions. The color gradient indicates time progression (blue to red). Dashed boxes indicate the locations of Chow, HFD, and licking zones. Scale bar = 5 cm. B) Top: z-scored calcium traces from all detected neurons for each experiment, sorted by their similarity in activity patterns (Coss et al., 2022; Pérez-Ortega et al., 2024). Bottom: Feeding and licking bouts automatically detected by the Crunchometer, color-coded by food type Chow (yellow), HFD (pink), and licking (blue). Running speed over time, estimated from video tracking. C) Calcium traces from representative GABAergic neurons from each experiment. Neurons display diverse activity profiles, including activation and suppression during feeding bouts. In the left panel, neuron 7 exhibited more activity during feeding but no activity during licking sucrose; in contrast, neurons 16, 18, and 19 mainly responded to liquid sucrose but not to solid food. Neuron 23 exhibited activity during both licking and solid food intake, suggesting that it may participate in processing both types of consummatory behaviors. In the panel on the right, during feeding periods with solid food pellets, neurons 6, 7, and 8 become active, while neuron 3 is inhibited, exhibiting a mirror-like pattern, and neuron 14 shows rebound disinhibition. In this session, the mouse did not lick for sucrose. D) Receiver operating characteristic (ROC) curves showing the ability of individual neurons to predict feeding (left) or licking (right) events. Each curve corresponds to a neuron whose decoding performance is significantly different from chance (p < 0.05, shuffle test based on temporal shifting of neuronal activity). Only significant neurons are shown. Curves with AUC > 0.5 are shown in red and indicate neurons whose activity increases during the behavioral event, whereas curves with AUC < 0.5 are shown in blue and indicate neurons whose activity decreases during the event. In LH GABAergic neurons, a similar percentage of neurons exhibited significant AUC predictive activity associated with solid-food feeding (n=32/79; 41%) and licking (n = 24/52; 46%) (chi-square (with Yates correction): χ²(1) = 0.21, p = 0.646). E) Boxplots showing the distribution of AUC values for neurons significantly predicting feeding (left) or licking (right). Only neurons that reached statistical significance in the shuffle test (same neurons shown in D) are included. F) Scatter plot comparing feeding AUC versus licking AUC for neurons recorded in sessions where both behaviors occurred. Each dot represents one neuron. Dashed lines indicate chance-level decoding (AUC = 0.5) and divide the plot into quadrants corresponding to neurons that significantly predict feeding, licking, or both behaviors. Icons indicate the type of significant response exhibited by each neuron (see legend). The number of significant neurons responsive to liquid and/or solid food relative to the total recorded population is shown in the plot (n=28 out of 52; ∼54%), and the percentage of neurons in each response category is summarized in the legend. A subset of LH GABAergic neurons exhibited significant ROC-based discriminability between solid-food feeding and licking, either through selective responses or through opposite-sign modulation of activity across the two behaviors.

LH glutamatergic neurons differentially encode solid and liquid ingestive behaviors, with a bias toward solid food consumption.

A) Two representative experiments are shown (left and right panels). Left: GCaMP7s fluorescence image showing ROIs of individual glutamatergic neurons in the lateral hypothalamus. The experiment on the right was performed by recording 3 planes, which allowed us to obtain a larger number of neurons. Scale bar = 20 µm. Right: Mouse trajectories during 30-minute sessions with access to Chow pellets, HFD pellets, and a licking sipper (with a metal floor for a contact lickometer) delivering water or sucrose solutions. Colors represent time progression (blue to red). Dashed boxes indicate the locations of Chow (yellow ticks), HFD (pink), and licking (blue) zones. Scale bar = 5 cm. B) Top: z-scored calcium traces from all detected glutamatergic neurons in each experiment. Bottom: Automatically detected feeding and licking bouts as in Figure 7. Mouse’s running speed traces from both experiments. C) Example calcium activity traces from glutamatergic neurons illustrating different response profiles across solid and liquid ingestive behaviors. In the experiment on the left, only two neurons (17 and 18) exhibited activity tightly coupled to licking for liquid sucrose but remained unresponsive during the intake of solid food.