Neural Processing of Taste-Related Signals in the Mediodorsal Thalamus of Mice

MD reciprocal connection with the gustatory insular cortex

To investigate the type of taste-related information processed by MD neurons in mice, we first evaluated the neural connection between the primary taste cortex (i.e. the gustatory insula, GC) and MD using anatomical tracing approaches. The rationale for these experiments was twofold: (1) to confirm and expand on the existing evidence of the connection between the gustatory portion of the insular cortex (GC) and the MD in rodents (Allen et al., 1991; Gehrlach et al., 2020) and (2) to locate the specific parts of the mouse MD that could potentially be involved in processing taste and/or taste-related information. Given the small size of this thalamic nucleus in mice, the latter step was crucial for us to ensure accurate targeting for silicon probe implantation for neural recording, the focus of the remainder of this study. For these experiments, we first injected a retrograde neural tracer (CTB-555; Fig. 1A,B) into a region of the GC known to process gustatory information (Chen et al., 2011; Kusumoto-Yoshida et al., 2015; Levitan et al., 2019; Bouaichi and Vincis, 2020; Bouaichi et al., 2023) and receive direct axonal projections from the gustatory thalamus (VPMpc; Chen et al., 2011, 2021; Fletcher et al., 2017). This approach allowed us to label the somata of neurons with terminal fields in the GC (Fig. 1A,B). The injections were unilateral and targeted all subdivisions of the GC (granular, dysgranular, and agranular; Fig. 1A, left panel). As anticipated, we observed a substantial number of thalamocortical neurons labeled in the ipsilateral VPMpc (Fig. 1A, middle panel). Furthermore, beyond the VPMpc, we identified neurons labeled in the MD (Fig. 1A, right panel). To evaluate the relative distribution of GC-projecting neurons, we compared the number of CTB⁺ neurons in the two thalamic nuclei. Analysis of the relative number of retrogradely labeled neurons revealed significant differences in the distribution of thalamo-GC–projecting neurons across the two nuclei (Fig. 1B; Wilcoxon rank sum; W = 16; p = 0.02857). While the majority of labeled neurons originated from the gustatory thalamus (VPMpc), the presence of a substantial population in MD suggests that this higher-order thalamic nucleus may also play a role in influencing GC activity. Figure 1C shows the qualitative evaluation of the location of the MD neurons that projected to the GC obtained from six mice, with the largest overlap present in the medial subregion of the MD. To further explore this neural connection, we then used anterograde viral tracers (AAV9-hSyn-mCherry) in the GC to qualitatively map the projections from the GC to the MD (Fig. 1C). Similar to what was observed for retrograde tracer experiments, the medial subregion of the MD appeared to be the recipient of most of the GC–MD projections (Fig. 1D). Although MD is not part of the classical taste sensory pathway, its reciprocal connection with the GC suggests a potential role in processing gustatory information even in the absence of a previously established association with another oral sensory modality, such as retronasal odor (Fredericksen and Samuelsen, 2022). For these reasons, we next investigated the response profile of MD neurons in representing chemosensory gustatory information, focusing on taste quality and taste concentration, in active licking mice.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1.

A, Leftmost panel, Coronal section of the mouse brain showing the CTB injection site (in red) in the GC counterstained with Hoechst (in cyan). Central panel, On the top, a coronal section showing the location of CTB⁺ neurons in the gustatory thalamus (VPMpc); on the bottom, high-magnification image of CTB⁺ neurons in the VPMpc. Rightmost panel, A coronal section showing the location of CTB⁺ neurons in the MD with nuclear Hoechst counterstaining in cyan. On the right, high-magnification image of CTB⁺ neurons in the MD. B, Individual percentages (filled circles) and averaged (horizontal bars; n = 4 mice) of CTB⁺ neurons in the two thalamic nuclei. C, Qualitative reconstruction of the MD regions showing the highest amount of CTB⁺ neurons (n = 6 mice). D, Leftmost panel, Coronal section of the mouse brain showing the injection site of the adeno-associated viral construct AAV9-hSyn-mCherry (in orange) in the GC counterstained with Hoechst (in cyan). Central panel, A coronal section showing the location of mCherry⁺ axonal projections in the midline thalamic nuclei. The white-dotted lines highlight the MD borders. Rightmost panel, A qualitative reconstruction of the MD regions showing the highest amount of mCherry⁺ axonal projections (n = 5 mice). Note that the qualitative labeling schematics shown in C and D refer only to MD and not to adjacent areas. MD, mediodorsal nucleus of the thalamus; MDc, mediodorsal nucleus of the thalamus, central division; MDl, mediodorsal nucleus of the thalamus, lateral division; MDm, mediodorsal nucleus of the thalamus, medial division; PV, paraventricular nucleus of the thalamus; IMD, intermediodorsal nucleus of the thalamus; VPMpc, parvicellular portion of the ventroposteromedial nucleus, also known as the gustatory thalamus.

Taste quality processing in the mouse MD

To investigate how individual MD neurons represent taste quality information in ad libitum licking mice, we recorded single units using movable silicon probes (Cambridge NeuroTech) mounted on a nanodrive shuttle (Cambridge NeuroTech) implanted unilaterally in the MD (Fig. 2A,B). After habituation to head restraint, water-deprived mice were engaged in a behavioral task in which they had to lick six times on a dry spout to obtain a 3 μl drop of one out of four gustatory stimuli at a fixed concentration (0.1 M sucrose, 0.05 M NaCl, 0.01 M citric acid, 0.001 M quinine) or deionized water, all presented at room temperature (Fig. 2C). To distinguish neural activity evoked by gustatory stimuli from electrophysiological correlates of sensory and oromotor/palatability aspects of licking, we initiated neural recording sessions only after confirming that the licking patterns evoked by each gustatory stimulus were statistically similar within a 1.5 s poststimulus temporal window, as assessed using a Kruskal–Wallis test (Bouaichi et al., 2023). The taste quality-independent licking ensured that differences in neural responses between gustatory stimuli within 1.5 s after taste delivery could not be attributed to overt licking/palatability behavioral correlates. Figure 2D shows the raster plots and PSTHs of three representative MD neurons. Visual inspection of the graphs indicated that each of these neurons responded to more than one taste stimulus with time-varying and multiphasic changes in activity similar to those reported by Fredericksen and Samuelsen (2022) when oral stimuli were delivered via intraoral cannulae. As a first step, we wanted to understand how many MD neurons, such as the three shown in Figure 2D, were selectively modulated by gustatory stimuli. Neurons were defined as taste-selective if they exhibited a significant change in activity compared with the baseline for at least one taste stimulus and had significantly different responses to the five stimuli (see Materials and Methods). This analysis revealed that more than half of the MD neurons recorded (66.5%, 141/212) were taste-selective (Fig. 2E; χ² test for given probabilities, χ² = 23.113; df = 1; p = 1.527 × 10⁻⁰⁶).

• Download figure
• Open in new tab
• Download powerpoint

Figure 2.

A, Schematic showing the recording setup and a head-restrained mouse licking a spout to obtain tastants. B, Left, Example of histological sections showing the tracks (red) of one shank of the CN-P1 probe in the MD. Red arrow points to the tip of the probe. Right, Schematic of the summary of probe tracks from the mice used in this study. Red shanks are implanted along the caudal–rostral axis (one shank visible), while yellow shanks are perpendicular, showing both shanks. C, Top, A diagram of the taste delivery paradigm: taste stimuli are delivered after six consecutive dry licks to the spout; bottom, a raster plot of licking activity during one experimental session: each line represents an individual lick. Taste delivery occurs at time 0 s (highlighted in red). D, Raster plots and PSTHs of three MD neurons showing taste responses. Trials pertaining to different tastants are grouped together (in the raster plots) and color-coded (both in the raster plots and PSTHs), with sucrose (S; “brown”), quinine (Q; “purple”), NaCl (N; “red”), citric acid (CA; “green”), and water (W; “orange”). Smaller text within the PSTHs highlights the sharpness (SI; blue) and entropy (H; red) values for each of the neurons. E, Top, A pie chart showing the proportion of MD taste-selective neurons. Bottom left, fraction of taste-selective neurons responding to one or more gustatory stimuli. Bottom right, The distribution of the breadth of tuning [expressed as SI (“blue”) and H (“red”)] of the taste-selective neurons. High SI or low H values indicate narrowly responsive neurons, whereas low SI or high H values imply that the same neuron is modulated by multiple tastants.

Next, we evaluated the tuning profile of the taste-selective neurons. We aimed to understand whether the taste-selective MD neurons preferentially responded to only one single taste stimulus (i.e., narrow tuning) or, as suggested by the example neurons shown in Figure 2D, if they were capable of encoding information pertaining to multiple tastes (i.e., broad tuning). This analysis served only to estimate how many taste-selective neurons were modulated by more than one tastant independent of the quality/identity of the taste. Our analysis revealed that, while only 25.5% (36/141) of MD taste-selective neurons responded to one taste, the vast majority (74.5%; 105/141) were modulated by more than one gustatory stimulus (Fig. 2E; χ² test for given probabilities, χ² = 33.766; df = 1; p = 6.216 × 10⁻⁰⁹). To further investigate differences in the tuning profiles of MD neurons, we calculated the response SI and the response H for each taste-selective neuron. These two analyses are standard techniques used to evaluate the breadth of tuning of single neurons (Smith and Travers, 1979; Rainer et al., 1998; Yoshida and Katz, 2011; Bouaichi and Vincis, 2020). Figure 2E shows the distribution of the response SI (Yoshida and Katz, 2011). An SI of 1 describes a neuron that responds to only one taste, and an SI of 0 describes a neuron that responds to all five stimuli. The distribution of SI values strongly implied that most taste-selective neurons were broadly tuned, suggesting that the majority of taste-selective neurons in MD are modulated by more than one taste. Similar results were obtained by analyzing the response H. Low H values are evidence of narrowly tuned neurons, whereas high H values indicate broadly tuned neurons. It is important to note that when interpreting entropy values, the use of absolute neural response values can bias the results toward higher H values. However, despite this potential bias, the results from the three different analyses of the responsiveness profile of taste-selective MD neurons, as shown in Figure 2E, consistently indicate that MD neurons are broadly tuned.

After characterizing the profiles of chemosensory stimuli in individual MD neurons, we shifted our focus to the neural activity at the population level. While the activity of single neurons can represent important features of sensory stimuli, it is the information encoded in populations or ensembles (networks) of neurons that informs behavioral choices. As a result, we computed the taste quality decoding performances in our taste-selective neuron population (n = 141; gold trace in Fig. 3A). The onset of taste decoding occurred in the third temporal bin, ∼200  ms after stimulus delivery, and reached its peak later than 500 ms after taste delivery (Fig. 3A). Additionally, while the overall classification value started decreasing after 500 ms, decoding performances remained significantly above chance and shuffled control (gray trace in Fig. 3A; permutation test, p < 0.05) until the end of the temporal window analyzed (Fig. 3A). As a control, the same analysis was performed using licking instead of spiking data (Fig. 3B). As expected, the taste classification accuracy using licking data never exceeded the shuffled control level. The lack of taste classification with licking data, as opposed to spiking data, further indicated that the MD neural activity encodes specific information about the taste quality that could not be attributed to overt oromotor/palatability behavioral correlates. Figure 3C shows how classification during a 1.5 s poststimulus temporal window changed as the decoder gained access to progressively more MD neurons. While the classifiers that were trained with a small number of neurons were less accurate at identifying taste quality information, increasing the number of neurons in the population drastically increased the decoding performance, reaching up to 50% accuracy with <20 neurons. Next, we wanted to provide detailed information on the temporal dynamics of the taste classification performance with respect to each gustatory stimulus. To do this, we constructed confusion matrices and characterized the classification performance for each tastant in 150 ms time bins with 50 ms moving window around taste delivery (Fig. 3D). Figure 3A shows that the average decoding performance in taste-selective neurons was well above chance and shuffled control during the first 500 ms (between ∼200 and 500 ms). However, inspection of the confusion matrices in Figure 3D revealed that, in this first phase of taste processing, not all taste qualities were equally classified. In this plot, the main diagonal represents the fraction of trials where the classifier accurately matched the predicted taste stimulus to its actual category. Comparison of the fraction of trials correctly classified in the first 500 ms revealed that all stimuli, with the exception of quinine and deionized water, were predicted by the decoding algorithm (Marascuilo's test; Tables 1, 2). As time progressed, all stimuli were decoded with accuracy above chance (20%) within 1 s from fluid delivery. However, as represented in the last row of Figure 3D, the classifier is less likely to decode water compared with the classical taste stimuli (Marascuilo's test; Tables 1, 2) and is more likely to decode sucrose and quinine compared with NaCl and citric acid (Marascuilo's test; Tables 1, 2). Altogether, these data indicate that ensembles of MD neurons recorded from active licking mice dynamically encode gustatory information up to 2 s after stimulus delivery. Considering that the neural activity in response to taste stimuli was recorded in the absence of overt differences in oromotor behaviors related to palatability (licking; Figs. 2C, 3B), the encoded information likely reflects the chemosensory discriminative aspects of the stimulus. In addition, taste coding showed a relatively “late” onset—at least compared with the taste cortex (Bouaichi and Vincis, 2020)—with the time course of the averaged population decoding performance significantly rising above shuffled control within ∼200 ms.

• Download figure
• Open in new tab
• Download powerpoint

Figure 3.

A, Time course of decoding performance (white line) considering the population of taste-selective neurons. Gold-shaded area indicates the 1–99 percentile range of the 20 times the decoder was run, each time using different training and testing splits (n = 20). Gray-shaded areas indicate the 1–99 percentile range of the decoding performance over time after shuffling (20 times) stimulus labels for all trials. The horizontal black bar above the trace denotes bins when the classification accuracy significantly differed from the shuffled (permutation test, p < 0.05). B, Time course of decoding performance (white line) considering the licking data extracted from the behavioral sessions from which the taste-selective neurons where recorded. The gold-shaded area indicates the 1–99 percentile range of the 20 times the decoder was run, each time using different training and testing splits (n = 20). Gray-shaded areas indicate the 1–99 percentile range of the decoding performance over time after shuffling (20 times) stimulus labels for all trials. C, The mean accuracy of the decoder trained to discriminate the different gustatory stimuli (using a temporal window of 1.5 s after taste delivery) as the decoder gained access to progressively more neurons (gold dots). D, Confusion matrices showing decoding performance for each gustatory stimulus in different 150 ms bins with 50 ms moving window temporal epochs around taste delivery (0 s). Color codes the classification accuracy, with bright hues indicating a higher fraction of correct trials. The main diagonal highlights the number of trials in which the classifier correctly assigned the taste stimulus (predicted taste) to its real category (true taste). On the rightmost panels, the average decoding accuracy for three 500 ms temporal windows (−500 to 0 ms, top panels; 0–500 ms, middle panels; 500–1,000 ms, bottom panels; 0 s taste delivery) shown both as confusion matrices and bar plots. In the bar plots, red dotted lines represent the chance level.

It is important to note that these results are obtained by administering taste stimuli at fixed and relatively low concentrations. This approach was primarily chosen to facilitate comparison with previous neural recordings in the mouse primary taste cortex (Levitan et al., 2019; Bouaichi and Vincis, 2020; Neese et al., 2022). Although the concentrations selected for each taste quality are suprathreshold, they are not equi-detectable. Therefore, differences in the MD neural activity may also result from changes in stimulus concentration. Concentration is a critical aspect of taste stimuli, and many studies have shown that perceptual/behavioral and neural responses to taste stimuli often correlate with concentration (Breza and Contreras, 2012; Sadacca et al., 2012; Wilson and Lemon, 2014; Spector et al., 2015; Fonseca et al., 2018). For these reasons, our next step was to explore if and how MD neurons integrate taste concentration information, focusing specifically on the taste qualities of sweet (sucrose) and salty (NaCl).

Taste concentration processing in the mouse MD

To begin investigating how individual MD neurons represent sucrose concentration, we recorded individual units as described above. However, after habituation to head restraint, water-deprived mice were engaged in a task in which, this time, they had to lick six times on a spout to obtain a 3 μl drop of one of the five peri- and suprathreshold sucrose concentrations (0.02, 0.04, 0.06, 0.1, and 0.3 M; Treesukosol et al., 2009; Spector, 2015), all presented at room temperature. Figure 4A shows the raster plots and PSTHs of two MD neurons that produced distinct responses to different sucrose concentrations. The responses of both neurons differed from the baseline spiking activity and showed a trend of increasing (Neuron #1) and decreasing (Neuron #2) firing rates with increasing stimulus concentration. Interestingly, monotonic positive and negative responses to changes in taste concentrations have been reported in other brain regions along the taste pathway (Yamamoto et al., 1984; Breza and Contreras, 2012; Sadacca et al., 2012; Wilson and Lemon, 2014; Fonseca et al., 2018). To quantify how many MD neurons respond selectively to sucrose concentrations, we used an approach similar to the one described above for taste quality. MD neurons were defined as concentration-selective if they exhibited a significant change in activity compared with the baseline for at least one stimulus and responded differently to multiple stimuli. Our analysis revealed that, of the 154 MD neurons recorded, more than half (56%, 87/154) were sucrose concentration-selective. To further evaluate their responsiveness as a function of concentration, we normalized the evoked firing rate in a 1.5 s poststimulus interval to the one evoked by the lowest concentration in our stimulus panel (0.02 M; Fig. 4B). As suggested by the example neurons in Figure 4A, the average spiking activity of these neurons was correlated with the increase in sucrose concentrations in a positive or negative monotonic manner (positive monotonic, R² = 0.94; p = 0.0058; negative monotonic, R² = 0.93; p = 0.00749; Fig. 4B). Importantly, this correlation was found to be independent of oromotor licking activity (Fig. 4B). When comparing the different types of neural responses, we did not observe significant differences in the proportion of neurons showing positive or negative monotonic responses (χ² test for given probabilities, χ² = 0.10345; df = 1; p = 0.7477; Fig. 4B). However, it is important to note that while the correlation between sucrose concentration and the single-neuron firing rate appears strong, this may be influenced primarily by the highest suprathreshold concentration of 0.3 M. Indeed, when we reanalyzed the data excluding the highest concentration, the linear correlation was lost (positive monotonic, R² = 0.64; p = 0.19; negative monotonic, R² = 0.2; p = 0.49), suggesting that for the concentrations tested here, the main difference at the level of single-neuron response is driven by the highest sucrose concentration rather than a continuous gradient across all concentrations.

• Download figure
• Open in new tab
• Download powerpoint

Figure 4.

A, Raster plots and PSTHs of two MD neurons showing sucrose concentration responses. Trials pertaining to different sucrose concentrations are grouped together (in the raster plots) and color-coded (both in the raster plots and PSTHs), with 0.02 M in gray, 0.04 M in yellow, 0.06 M in orange, 0.1 M in magenta, and 0.3 M in brown. B, On the left panel, the average (circles) and standard error of the mean (vertical lines) of the firing rate of all neurons with sucrose concentration-related information that either increased (green) or decreased (purple) their activity as the sucrose concentration increased. Firing rate values are normalized to 0.02 M. The average Pearson's R² coefficients and associated p values are also shown for increases and decreases. On the top-right panel, the sucrose-independent lick rate (red circles) in the same time window used for the quantification of spiking activity shown in the left panel. On the bottom-right panel, bar plot showing the fraction of neurons with sucrose concentration-related information that either increased (green) or decreased (purple) their activity as the sucrose concentration increased. C, Left panel, Time course of decoding performance (white line) considering the population of sucrose concentration-selective neurons. The gold-shaded area indicates the 1–99 percentile range of the 20 times the decoder was run, each time using different training and testing splits (n = 20). Gray-shaded areas indicate the 1–99 percentile range of the decoding performance over time after shuffling (20 times) stimulus labels for all trials. The horizontal black bars above the trace denote bins when the classification accuracy significantly differed from the shuffled (permutation test, p < 0.05). Right panel, Confusion matrix showing decoding performance for each sucrose concentration in a 500 ms temporal window (500–1,000 ms after stimulus delivery). Color codes the classification accuracy, with bright hues indicating a higher fraction of correct trials. The main diagonal highlights the number of trials in which the classifier correctly assigned the predicted sucrose concentration to its real category. D, In black, the distribution of max decoding accuracy of taste quality obtained by randomly selecting 87 neurons from the taste quality dataset (n = 141, shown in Figs. 2, 3); the dotted vertical red line highlights the max decoding accuracy obtained from the sucrose concentration-selective neurons (n = 87).

We then wanted to investigate how well sucrose concentration was encoded at the level of population of MD neurons. To this end, we calculated the decoding performances using the neural population of sucrose concentration-selective neurons (Fig. 4C). The classification accuracy of the sucrose concentration-selective population exceeded the chance level and was significantly different from that of the shuffled control beginning >250 ms after stimulus delivery (permutation test, p < 0.05). The delayed onset of significant decoding appeared to be in accordance with the timing described for the onset of taste quality decoding (Fig. 3A). However, the decoding performance appeared to be qualitatively lower for sucrose concentration compared with taste quality. Although this comparison involves different populations of MD neurons, which could partly explain the observed difference, we can speculate as to why the average classification performance for sucrose concentration is lower than the one for taste quality. One reason could be that MD-evoked neural activity contains less information, particularly at lower sucrose concentrations, leading to lower average classification accuracy. This is supported by the confusion matrix shown in the right panel of Figure 4C; the classification performance for each concentration in a time window of 500 ms indicates that higher sucrose concentrations were better classified than lower ones. Another possibility is that the population of neurons that were selective for sucrose concentration was smaller than the group of neurons selective for taste, 87 and 141 neurons, respectively, which may explain the reduced classification. However, when we compared the maximum decoding accuracy for sucrose concentration with a distribution of maximum decoding accuracy for taste quality using subsets of 87 neurons, we found that the sucrose concentration performance lay below the fifth percentile of the maximum decoding performance distribution for taste quality (Fig. 4D).

Next, we explored the extent to which MD neurons represent the concentration differences of another taste quality, salt. This time after mice had become accustomed to head restraint, water-restricted animals were trained on a licking task where they needed to lick six times on a spout to receive a 3 μl drop of one of the five peri- and suprathreshold NaCl concentrations (0.02, 0.04, 0.07, 0.15, and 0.3 M; Treesukosol and Spector, 2012; Spector, 2015). Figure 5A shows the raster plots and PSTHs of two MD neurons that produced NaCl concentration-selective responses. In this case, the responses of both neurons differed from the baseline spiking activity and showed a trend of increasing firing rates with increasing NaCl concentration. Our analysis revealed that, of the 84 MD neurons recorded, 46% were NaCl concentration-selective. Similarly to what was described for sucrose, we found that stimulus-evoked spiking activity was correlated with the increase in concentrations in a positive and negative monotonic manner (positive monotonic, R² = 0.98; p = 0.0005; negative monotonic, R² = 0.78; p = 0.0466; Fig. 5B) and was independent of oromotor licking. Similarly to what we described above for sucrose taste, we reanalyzed the data excluding the highest suprathreshold NaCl concentration. In this case, removing the highest concentration did not eliminate the correlation (positive monotonic, R² = 0.96; p = 0.0196; negative monotonic, R² = 0.93; p = 0.0333), indicating that for the NaCl concentration tested here, the firing properties of individual MD neurons remain correlated to the stimulus gradient even in the absence of the highest concentration. In addition, we found a higher proportion of neurons showing positive monotonic responses (χ² test for given probabilities, χ² = 5.7692; df = 1; p = 0.01631; Fig. 5B). The time course of the NaCl classification average across concentrations (Fig. 5C, left panel) showed the onset and accuracy similar to what was described for sucrose, with higher NaCl concentrations classified better than lower ones (Fig. 5C, right panel). Finally, when comparing decoding accuracy using subsets of neurons, NaCl concentration performance was also significantly lower, below the first percentile of the taste quality performance distribution (Fig. 5D).

• Download figure
• Open in new tab
• Download powerpoint

Figure 5.

A, Raster plots and PSTHs of two MD neurons showing NaCl concentration responses. Trials pertaining to different NaCl concentrations are grouped together (in the raster plots) and color-coded (both in the raster plots and PSTHs), with 0.02 M in gray, 0.04 M in yellow, 0.07 M in orange, 0.15 M in red, and 0.3 M in brown. B, On the left panel, the average (circles) and standard error of the mean (vertical lines) of the firing rate of all neurons with NaCl concentration-related information that either increased (green) or decreased (purple) their activity as the NaCl concentration increased. Firing rate values are normalized to 0.02 M. The average Pearson's R² coefficients and associated p values are also shown for increases and decreases. On the top-right panel, the NaCl-independent lick rate (red circles) in the same time window used for the quantification of spiking activity shown in the left panel. On the bottom-right panel, a bar plot shows the fraction of neurons with NaCl concentration-related information that either increased (green) or decreased (purple) their activity as the NaCl concentration increased. C, Left panel, Time course of decoding performance (white line) considering the population of NaCl concentration-selective neurons. The gold-shaded area indicates the 1–99 percentile range of the 20 times the decoder was run, each time using different training and testing splits (n = 20). Gray-shaded areas indicate the 1–99 percentile range of the decoding performance over time after shuffling (20 times) stimulus labels for all trials. The horizontal black bars above the trace denote bins when the classification accuracy significantly differed from the shuffled (permutation test, p < 0.05). Right panel, Confusion matrix showing decoding performance for each NaCl concentration in a 500 ms temporal window (500–1,000 ms after stimulus delivery). The color codes the classification accuracy, with bright hues indicating a higher fraction of correct trials. The main diagonal highlights the number of trials in which the classifier correctly assigned the predicted NaCl concentration to its real category. D, In black, the distribution of max decoding accuracy of taste quality obtained by randomly selecting 39 neurons from the taste quality dataset (n = 141, shown in Figs. 2, 3); the dotted vertical red line highlights the max decoding accuracy obtained from the NaCl concentration-selective neurons (n = 39).

Together, these data revealed that a substantial fraction of neurons in the mediodorsal thalamus represents sucrose and NaCl concentrations (and/or the “concentration” of water), with spiking activity that was correlated with the increase in stimulus concentrations in a positive and negative monotonic manner and independent of behavioral correlates of palatability (licking). However, our single-neuron analyses suggest that, for sucrose, this correlation is primarily driven by the highest concentration, whereas for NaCl, the relationship persists even after its removal. Furthermore, decoding analysis indicates that while individual MD neurons exhibit some degree of concentration selectivity, the discrimination of taste stimulus concentrations appears to be more robust at the population level, with higher concentrations being classified more accurately than lower ones. These findings support the idea that MD may contribute to the processing of sensory-discriminative features of gustatory chemosensory stimuli. However, it is important to note that although the concentrations tested here span a relatively broad range of perithreshold and suprathreshold values based on previous studies (Treesukosol and Spector, 2012; Spector, 2015), our mice were “passively” experiencing these stimuli without using this information to perform a behavioral task, such as detection or discrimination. Given this, the concentration-selective activity observed here, both at the single-neuron level and, more prominently, at the population level, likely represents only the tip of the iceberg in terms of the potential role of MD in concentration-selective gustatory processing. It is plausible that when concentration information becomes behaviorally relevant—especially in the perithreshold range—MD neurons would exhibit even stronger and more selective responses, reflecting their role in task-driven sensory processing.

MD processing of taste-specific expectation

Gustatory experiences are often perceived against the background of previous expectation. Cues of all sensory modalities frequently provide information predictive of taste outcomes and play a crucial role in shaping taste-driven behaviors (Samuelsen et al., 2012; Gardner and Fontanini, 2014; Kusumoto-Yoshida et al., 2015; Vincis and Fontanini, 2016; Livneh et al., 2017). Experiments in rats have highlighted MD as a brain region that responds to sensory cues associated with a specific behavioral outcome (Oyoshi et al., 1996; Kawagoe et al., 2007) and to the association of intraorally sourced odor and taste stimuli relevant to flavor (Fredericksen and Samuelsen, 2022). To determine how neurons in the mouse MD process information related to a taste-specific expectation, we relied on a two-cue–taste association task. We trained mice (n = 5) to associate one auditory stimulus with the delivery of 0.3 M sucrose and another with the delivery of 0.03 M quinine (Fig. 6A). Two-second-long auditory cues (2 and 12 kHz) were followed by a 0.5 s delay and by taste delivery (Fig. 6A). Cue contingencies were counterbalanced across mice. To track associative learning, we analyzed the licking rate in anticipation of the two gustatory stimuli (Fig. 6B) by computing a “selectivity index” using a ROC analysis (see Materials and Methods). The association was established for each individual animal when the “selectivity index” was negative (indicating a higher lick rate for cue–sucrose association) and significant (p < 0.05) compared with the distribution of “pseudoselectivity indices” obtained after randomizing the trials. All mice reached the criterion within 2 weeks of training. After specific cue–taste associations were established, we performed neural recordings. Figure 6C illustrates four representative examples of MD neurons that were modulated by the presentation of auditory cues. Visual inspection of raster plots and PSTHs (Fig. 6C) indicated that, from a qualitative point of view, the activity of the MD neuron could be enhanced or suppressed by both auditory cues, irrespective of the valence of their associated outcome. It should be noted that cue responses consistently preceded the onset of licking. To quantify the fraction of thalamic neurons whose activity was modulated by different auditory stimuli, we analyzed changes in the firing rate that occurred before and after the onset of the cue (see Materials and Methods). Of all MD neurons recorded, an overwhelming majority (90.7%; 78/86; χ² test for given probabilities, χ² = 56.977; df = 1; p = 4.41 × 10⁻¹⁴) were modulated by at least one auditory signal (Fig. 6D). Next, we analyzed whether and to what extent the activity of cue-responsive neurons provided information predictive of specific gustatory outcomes (as suggested by the example neurons shown in Fig. 6C) or represented general anticipatory signals. We addressed this issue using the following analysis. For each cue-responsive neuron, we compared the spiking activity evoked by the two different auditory stimuli. Briefly, the poststimulus firing rate (0–2.5 s) was divided into five 500 ms bins, and a two-way ANOVA was used with “cue” and “time” as variables (Jezzini et al., 2013; Levitan et al., 2019; Bouaichi and Vincis, 2020). Cue-responsive neurons that had significantly different responses to the two tones in either the main effect “cue” or the “cue” × “time” interaction were defined as cue-selective (Fig. 6D). This method revealed that the majority of cue-responsive neurons (73%, 57/78) responded selectively to cues predicting different outcomes, while the remaining 27% likely responded to general anticipatory signals (Fig. 6D; χ² test for given probabilities, χ² = 16.615; df = 1; p = 4.578 × 10⁻⁰⁵). In addition, the vast majority (80%, 46/57; Fig. 6E) of cue-selective neurons responded to cues predicting both outcomes with only a small fraction responding to only one auditory stimulus (20%, 11/57; χ² test for given probabilities, χ² = 21.491; df = 1; p = 3.555 × 10⁻⁰⁶). Among the latter, a similar number of neurons were modulated only by the sucrose-predicting or the quinine-predicting cue (54%, 6/11 for cue–quinine; 45%, 5/11 for cue–sucrose; χ² test for given probabilities, χ² = 0.090909; df = 1; p = 0.763). It is important to note that, according to the cue selectivity training paradigm, it was not possible to determine whether a given neuron responded to sucrose and/or quinine prior to training. While this does not invalidate the observed cue-selective responses, it is a factor to consider when interpreting cue selectivity. Future studies will further investigate how pre-existing taste responsiveness may influence cue-driven MD neural activity.

• Download figure
• Open in new tab
• Download powerpoint

Figure 6.

A, Schematic representation of the behavioral task. Mice learned to associate specific auditory stimuli, Tone 1 (gold) and Tone 2 (green), with the delivery of two specific taste outcomes (sucrose and quinine, counterbalanced). B, Raster plots and PSTHs of licking activity for a representative experimental session. In the raster plot, the colored areas indicate blocks of cue–sucrose (cue-S; gold) and cue–quinine (cue-Q; green) trials. The PSTHs show the average licking for cue-S and cue-Q trials in gold and green, respectively. The black dashed vertical lines mark the onset and offset of the auditory cue (0 and 2 s, respectively), while the red dashed vertical line indicates taste delivery (2.5 s). C, Raster plots and PSTHs for four MD cue-selective neurons. Raster plots and PSTHs are color-coded, with activity evoked by cue-S and cue-Q shown in gold and green, respectively. Black dashed vertical lines indicate the onset and offset of the auditory cue (0 and 2 s, respectively), while the red dashed vertical line indicates taste presentation (2.5 s). D, Pie charts showing the proportion of MD cue-responsive (“leftmost chart”) and cue-selective (“rightmost chart”) neurons. E, Fraction of cue-selective neurons responding to one or both auditory cues. F, Fraction of cue-selective neurons modulated by cue-S and sucrose taste that exhibited matching responses (enhanced firing to both cue-S and sucrose taste; i.e., Neuron #3 in panel C) and nonmatching responses (enhanced firing to cue-S but suppressed to sucrose taste or vice versa; i.e., Neuron #2 in panel C).

Next, we performed a series of analyses to first determine whether cue-selective neurons also respond to the taste outcome and, second, to evaluate if the responsiveness of a neuron to a specific cue is related to its response to the taste. An important caveat must be recognized here. The successful learning of the cue–taste association task led to reduced anticipatory licking for cue-Q, resulting in very few trials with quinine consumption. Therefore, modulation in the firing rate after “quinine” could result from movements in the mouth and snout associated with avoiding licking to avoid consumption. For this reason, we restricted these analyses to sucrose trials. Among sucrose cue-selective neurons, 56% (29/51) were also selectively modulated by the sucrose taste. We then performed an additional analysis to examine the relationship between their response to cue-S and their response to sucrose taste, specifically assessing whether the “sign” of neural modulation (enhancement or suppression) was consistent between both conditions (Fig. 6F). The results revealed that among MD neurons selectively modulated by cue-S, similar proportions exhibited matching responses (enhanced firing to both cue-S and sucrose taste) and nonmatching responses (enhanced firing to cue-S but suppressed to sucrose taste or vice versa; χ² test for given probabilities, χ² = 0.31034; df = 1; p = 0.5775; Fig. 6F).

Together, these results show that individual neurons in the mouse MD can respond selectively to cues anticipating taste outcomes. In addition, cue-selective neurons appear to preferentially represent information from cues predicting both rewarding and aversive taste, and some cue-selective neurons have different response patterns to the cues and the predicted outcomes.

Population decoding of cue–taste expectation by the MD

After characterizing the profile of cue-related responses in single neurons, we focused our attention on neural activity at the population level. Thus, to further evaluate how the MD network encodes specific expectations relevant to guiding consummatory behavior, we computed the decoding performances in the population of cue-responsive neurons. The time course of the classification average is shown in Figure 7A. The decoding performance of the cue-selective population (Fig. 7A) showed an early onset, with a classification accuracy above the chance level and significantly different from that of the control (i.e., shuffled data) from the first bin after cue onset (0–150 ms; permutation test, p < 0.05). As an additional control, the same analysis was performed using licking instead of spiking data (Fig. 7A, blue dotted line). In particular, the decoding with neural data showed an onset that occurred seconds before the decoding with licking data, indicating that neural responses, especially those in the first 1.5 s after the onset of the auditory stimulus, precede the licking activity. Furthermore, the classification accuracy of the cue-selective population significantly differed from that of the shuffled control for the entire 5 s window including cue and taste–outcome (Fig. 7A). This suggests that cue-selective MD neurons are capable of encoding both anticipatory and taste–outcome information.

• Download figure
• Open in new tab
• Download powerpoint

Figure 7.

A, Time course of decoding performance (white line) considering the population of cue-selective neurons. The gold-shaded area indicates the 1–99 percentile range of the 20 times the decoder was run, each time using different training and testing splits (n = 20). Gray-shaded areas indicate the 1–99percentile range of the decoding performance over time after shuffling (20 times) stimulus labels for all trials. The horizontal black bar above the trace denotes bins when the classification accuracy significantly differed from the shuffled (permutation test, p < 0.05). Dotted blue line represents the time course of decoding performance when licking data were used. Yellow box indicates time of auditory cue presentation. The dotted horizontal red line indicates the chance level (50%); the dotted vertical red line indicates time of taste outcome delivery. B, Time course of decoding performance (white line) considering the population of cue-nonselective neurons. The purple-shaded area indicates the 1–99 percentile range of the 20 times the decoder was run, each time using different training and testing splits (n = 20). Gray-shaded areas indicate the 1–99 percentile range of the decoding performance over time after shuffling (20 times) stimulus labels for all trials. The horizontal black bars above the trace denote bins when the classification accuracy significantly differed from the shuffled (permutation test, p < 0.05). The dotted blue line represents the time course of decoding performance when licking data were used. The yellow box indicates time of auditory cue presentation. The dotted horizontal red line indicates the chance level (50%); the dotted vertical red line indicates time of taste outcome delivery. C, Time course of population decoding performance over time by the cue-selective (gold) and cue-nonselective neurons. The horizontal black bars above the trace denote bins when the classification accuracy significantly differed between the two populations (permutation test, p < 0.05). The dotted gold line represents the time course of decoding performance with a population of cue-selective neurons composed of 21 randomly selected cells. The yellow box indicates time of auditory cue presentation. The dotted horizontal red line indicates the chance level (50%); the dotted vertical red line indicates time of taste outcome delivery. D, Confusion matrices showing decoding performance during different 500 ms temporal epochs around the cue onset (0 s) and taste–outcome delivery (2.5 s) for the cue-selective (gold background) and cue-nonselective (purple background) populations. The color codes the classification accuracy, with bright hues indicating a higher fraction of correct trials. The main diagonal highlights the number of trials in which the classifier correctly assigned the stimulus (predicted stimulus) to its real category (true stimulus). E, In red and blue, the distribution of max decoding accuracy during cue presentation (cue epoch, 0–2.5 s; red) and after taste outcome (taste epoch, 2.5–4.5 s; blue) obtained by randomly selecting 21 neurons from the cue-selective neuron population; the dotted vertical red and blue lines highlight the max decoding accuracy during cue presentation (cue epoch, 0–2.5 s; red) and after taste outcome (Taste epoch, 2.5–4.5 s; blue) obtained from cue-nonselective neuron population (n = 21).

Next, we switched our focus to the population of cue-nonselective neurons; those MD neurons whose firing activity evoked by auditory cues differed from the baseline but was similar between the two taste-predicting stimuli. As expected, the decoding performance of the cue-nonselective neurons showed that its population activity did not represent information about the auditory cues. However, the classification accuracy of the cue-nonselective neuron population exceeded the chance level and differed significantly from that of the control (i.e., shuffled data) beginning hundreds of milliseconds after taste outcome delivery (Fig. 7B; permutation test, p < 0.05). Thus, the cue-nonselective population of MD neurons may primarily process information related to actual taste outcomes rather than anticipatory auditory cues.

We then compared the classification accuracy of the two populations (Fig. 7C). As expected, the decoding performance of the cue-selective population differed significantly from that of the cue-nonselective population from the first bin after the cue onset and continuing until taste outcome delivery (permutation test, p < 0.05). This can also be observed from a qualitative analysis of the confusion matrices shown in Figure 7D, where the classification performances for each stimulus are represented. The significant difference in classification accuracy between the two populations was briefly disrupted after taste outcome delivery (Fig. 7C,D), where both populations of MD neurons encoded taste outcome information with similar accuracy. The brief disruption in the significant difference after taste delivery could reflect a transient period in which both populations of neurons are equally capable of encoding the presence or absence of fluid. However, the significant difference in decoding performance reemerged for the remaining temporal window, highlighting that cue-selective neurons better encode sucrose taste information compared with cue-nonselective neurons (Fig. 7C,D). This is consistent with the time course of taste quality decoding shown in Figure 3 where the taste classification average showed a delayed onset ∼200 ms after taste delivery. It is important to note that from a quantitative point of view, the differences in the time course of the two populations do not appear to depend on the number of neurons forming each population (Fig. 7C, dotted gold line). However, to further evaluate whether the decoding differences between the two populations during the cue period and, more importantly, during taste outcomes were the result of differences in population sizes, we performed an additional analysis. We compared the maximum decoding accuracy for the cues and taste outcomes of the cue-nonselective neurons with a distribution of the maximum decoding accuracy using subsets of 21 cue-selective neurons (Fig. 7E). The decoding accuracies of the cue-nonselective populations lay below the fifth percentile of the decoding performance distribution for cue-selective neurons (Fig. 7E), suggesting that differences in encoding capabilities cannot be exclusively explained by the different population sizes.

Together, these results suggest that cue-selective neurons in the MD network are specifically tuned to auditory cues that predict different outcomes, allowing early and precise encoding of these signals. Furthermore, cue-selective neurons also reliably encoded taste outcomes, demonstrating their dual role in both anticipatory and outcome processing. In contrast, cue-nonselective neurons responded more generally and only showed decoding performance—though not as accurately as cue-selective neurons—after the actual taste outcome was delivered, indicating their involvement in processing the outcome rather than predicting it. In summary, these results highlight the role of MD neurons in anticipatory encoding, which is crucial to guide behavior based on expected outcomes.