Abstract
An economic choice entails computing and comparing the values of individual offers. Offer values are represented in the orbitofrontal cortex (OFC)—an area that participates in value comparison—but it is unknown where offer values are computed in the first place. One possibility is that this computation takes place in OFC. Alternatively, offer values might be computed upstream of OFC. For choices between edible goods, a primary candidate is the gustatory region of the anterior insula (gustatory cortex, GC). Here we recorded from the GC of male rhesus monkeys choosing between different juice types. As a population, neurons in GC represented the flavor, the quantity, and the subjective value of the juice chosen by the animal. These variables were represented by distinct groups of cells and with different time courses. Specifically, chosen value signals emerged shortly after offer presentation, while neurons encoding the chosen juice and the chosen quantity peaked after juice delivery. Surprisingly, neurons in GC did not represent individual offer values in a systematic way. In a computational sense, the variables encoded in GC follow the process of value comparison. Thus our results argue against the hypothesis that offer values are computed in GC. At the same time, signals representing the subjective value of the expected reward indicate that responses in GC are not purely sensory. Thus neuronal responses in GC appear consummatory in nature.
Significance Statement
We recorded from the gustatory cortex (GC) of monkeys choosing between different juice types. This area is interconnected with the orbitofrontal cortex, which plays a primary role in the generation of economic choices. As a population, neurons in GC represented the flavor, the quantity, and the subjective value of the juice chosen by the animal. Conversely, neurons in GC did not systematically represent the value of individual offers. Computationally, the variables encoded in GC follow the process of value comparison. At the same time, GC responses are not purely sensory because they represent the (subjective) chosen value. Thus neuronal responses in GC appear consummatory in nature.
Introduction
Economic choices between goods are thought to rely on the computation of subjective values. While choosing, individuals assign a value to each offer and then make a decision by comparing offer values. Value encoding neurons have been found in numerous cortical and subcortical regions (Roesch and Olson, 2007; Wallis and Kennerley, 2010; Schultz, 2015). However, neurons in the central orbitofrontal cortex (OFC) appear particularly relevant to the decision process. When monkeys choose between juices, different groups of cells in OFC represent the value of individual offers, the binary choice outcome, and the chosen value (Padoa-Schioppa and Assad, 2006; Pastor-Bernier et al., 2019; Gore et al., 2023). Notably, these variables capture both the input and the output of value comparison. The population dynamics recorded in this area is consistent with a decision process (Rich and Wallis, 2016; Balewski et al., 2023). Furthermore, experiments using electrical stimulation demonstrate that offer values encoded in OFC are causally linked to choices (Ballesta et al., 2020) and that this area contributes to value comparison (Ballesta et al., 2022). These observations suggest that economic decisions are formed within OFC. At the same time, previous studies leave open fundamental questions concerning the construction of offer values.
Computing an offer value requires integrating (1) external information coming from sensory stimuli and interpreted based on previous experience and (2) internal variables characterizing the chooser's internal state (Padoa-Schioppa, 2011). Since the orbital cortex receives input from all sensory modalities and from limbic areas (Ongur and Price, 2000), it is often assumed that the integration process takes place within OFC. However, this hypothesis has not been tested. Many studies on the neural mechanisms of economic choices were conducted in monkeys choosing between different juices. In these conditions, neurons in OFC represent options and values in the frame of reference defined by the juice type (Padoa-Schioppa and Assad, 2006; Pastor-Bernier et al., 2019). Thus the most likely alternative to the proposal that offer values are first computed in OFC is that offer values are computed upstream of OFC, in a gustatory representation (Howard et al., 2015; Suzuki et al., 2017). The present study was designed to test this hypothesis. While animals performed a juice-choice task, we recorded the activity of neurons in the gustatory region of the anterior insula (gustatory cortex, GC; Mufson and Mesulam, 1982; Carmichael and Price, 1994; Jezzini et al., 2015). This region provides direct anatomical input to OFC (Ongur and Price, 2000).
In the experiments, monkeys chose between different juices offered in variable amounts. Experimental design and data analysis were similar to those used in previous studies. Over half of GC neurons were modulated by the task in some time window. As a population, neurons in GC encoded three variables related to the choice outcome, namely chosen value, chosen number (i.e., chosen quantity), and chosen juice. These three variables explained the vast majority of task-related responses. Interestingly, while variables chosen value and chosen number were highly correlated, the encoding of these two variables was clearly distinct. Specifically, GC neurons encoded the chosen value early in the trial, long before the animal revealed its choice, whereas the encoding of chosen number became most prominent after juice delivery. Importantly, all three variables identified in GC are computationally downstream of the process of value comparison (post-decision variables). Conversely, and somewhat surprisingly, offer value signals were nearly absent in GC. Thus our findings argue against the hypothesis that offer values are computed in GC. More generally, they challenge the idea that neurons in GC contribute to the ongoing decision process. Importantly, GC responses are not purely sensory because they represent the (subjective) chosen value. Thus neuronal responses in GC appear consummatory in nature.
Materials and Methods
Experimental procedures
All experimental procedures conformed to the NIH Guide for the Care and Use of Laboratory Animals and were approved by the Institutional Animal Care and Use Committee at Washington University.
Surgical protocols, choice task, and procedures for neuronal recordings closely resembled those described in previous studies (Padoa-Schioppa and Assad, 2006; Jezzini and Padoa-Schioppa, 2020). Briefly, two adult male rhesus monkeys (Macaca mulatta; Monkey E, 9.5 kg; Monkey H, 10 kg) participated in the experiments. After familiarization and chair training, we implanted a head post and an oval recording chamber (main axes 30 and 50 mm) under general anesthesia. In Monkey E, the chamber was placed bilaterally, with the longer axis parallel to a coronal plane, and it was centered on 30 mm AP and 0 mm ML. In Monkey H, the chamber was placed in the left hemisphere, with the longer axis parallel to a sagittal plane, and it was centered on 25 mm AP and −10 mm ML. Pre- and post-surgery structural MRIs were used to guide neuronal recordings.
During the experiments, monkeys sat in an electrically insulated enclosure in front of a computer monitor (57 cm distance). The gaze direction was monitored with an infrared video camera (EyeLink, SR Research). The behavioral task was controlled through a custom software (MonkeyLogic) written in MATLAB (MathWorks; Hwang et al., 2019). Monkeys performed a standard juice-choice task (Fig. 1A). In each session, they chose between two juices labeled A and B (with A preferred) offered in variable amounts. At the beginning of the trial, the animal maintained central fixation for 1.5 s. Two offers were then presented simultaneously on the two sides of the fixation point. Each offer was represented by a set of color squares; for each set, the color indicated the juice type, and the number of squares indicated the juice quantity (each square represented one quantum of juice). The offers remained on display for a randomly variable delay (1–2 s), at the end of which the central fixation point was extinguished (go signal). The animal indicated its choice with a saccade. After an additional 0.75 s delay, the chosen juice was delivered. Different offer pairs were presented pseudorandomly. For each offer pair, the spatial configuration (left/right) was counterbalanced across trials. Sessions typically lasted 150–230 trials. Across sessions and across monkeys, we used 20 different juice types and 27 different juice pairings. The juice quantum was set at 70–85 μl in different sessions and was fixed within each session.
Extracellular recordings were conducted using single tungsten electrodes (FHC). Electrodes were guided through a custom-made 1 mm grid. Distances from the top of the grid to the surface of the brain were measured carefully. MRI scans were used to estimate the travel distance necessary to reach GC and to reconstruct the recording coordinates. For each animal, we performed dedicated sessions to map the transition between gray and white matter and to register these measures with those obtained from MRI scans. Neuronal activity was amplified, bandpass filtered (300 Hz–6 kHz; Lynx 8, NeuraLynx), and acquired at 40 kHz [Power 1401, Cambridge Electronic Design (CED)]. Spikes were detected online and saved to disk for off-line sorting (Spike2, CED).
All subsequent analyses were performed in MATLAB (version R2022a; MathWorks). Unless otherwise indicated, all analyses were as described in previous studies (Padoa-Schioppa and Assad, 2006; Jezzini and Padoa-Schioppa, 2020).
Analysis of choice patterns
As a behavioral criterion, we imposed that choice patterns be saturated—for each juice (A, B), the percentage of choices had to exceed 80% for at least one trial type. A total of 333 sessions satisfied this criterion. For each session, choices were analyzed with a probit regression. The quantities of juices A and B offered in any given are indicated as qA and qB, respectively. We used a log-value-ratio probit model (Padoa-Schioppa, 2022) defined as follows:
Analysis of neuronal activity
Neuronal data were analyzed in eight time windows aligned with different behavioral events: pre-offer (0.5 s preceding the offer onset), post-offer (0.5 s after the offer onset), late delay (0.5–1.0 s after the offer onset), pre-go (0.5 s preceding the go cue), reaction time (from the go cue to the saccade start), pre-juice (0.5 s preceding juice delivery), post-juice (0.5 s after juice delivery), and post-juice2 (0.5–1.0 s after juice delivery; Fig. 1B). An “offer type” was defined by two offered quantities; a “trial type” was defined by an offer type and a choice. For each cell and each time window, we averaged spike counts across trials for each trial type. A “neuronal response” was defined as the firing rate of one cell in one time window as a function of the trial type. Trial types with ≤2 trials were discarded. This criterion excluded 1% of trials for the GC dataset and 0.9% of trials for the OFC dataset (see below, OFC dataset).
For each neuron and each time window, we ran a three-way ANOVA (factors, offer type × position of A × movement direction) and we set a significance threshold p < 0.001. The large majority of neurons were not affected by spatial factors. We thus conducted a one-way ANOVA (factor, trial type). Responses that passed the significance criterion (p < 0.001) were included in subsequent analyses. Neurons that passed this criterion in ≥1 time windows are referred to as “task-related.”
We sought to assess what decision variables might be represented in GC during juice choice. Following previous work, we defined a series of variables that GC neurons could conceivably encode. These included variables associated with a single juice (offer value A, offer value B, etc.), variables reflecting the subjective value of the chosen and unchosen juices (chosen value, other value, value difference, etc.), variables defined by the quantity as opposed to the value (chosen number, total number, etc.), and a variable defined by the binary choice outcome (chosen juice). In total, we examined 19 variables (Table 1). For each neuronal response, we performed a linear regression separately on each variable; thus we obtained the regression slope and the R2. If the regression slope differed significantly from zero (p < 0.05), the variable was said to “explain” the response. If the variable did not explain the response, we arbitrarily set R2 = 0. Any given response could be explained by multiple variables. In this case, we identified the variable providing the best explanation (highest R2).
Defined variables
To identify a small number of variables that best accounted for the whole population, we used two procedures for variable selection, namely, stepwise and best-subset. Both procedures were originally designed for multilinear regressions in presence of multicollinearity (Dunn and Clark, 1987; Glantz and Slinker, 2001) and were adapted for the analysis of neuronal populations encoding different variables (Padoa-Schioppa and Assad, 2006). The stepwise procedure is iterative. In the first run, it selects the variable that provides the best explanation (highest R2) for the largest number of responses (pooling time windows). All responses explained by that variable are removed from the pool, and the procedure is repeated on the residual dataset. Iterations continue until when newly selected variables explain less than a minimum percent of responses. In this study, we set the minimum at 3%. Importantly, at each iteration, we verified that all previously selected variables explained at least 3% of responses; variables failing this criterion would be removed from the selected set, although in practice this situation did not occur in the present analyses.
Notably, the stepwise procedure is path dependent and does not guarantee optimality. In contrast, the best-subset procedure is exhaustive. For k = 1, 2, 3…, this procedure examines all the subsets of k variables, computes the number of responses explained by each subset, and identifies the subset that explains the maximum number of responses. Here we used a variant of this analysis, which examined the total R2 explained across responses for each subset (the results obtained with this variant appeared more stable). For each k, the explanatory power of the best subset of variables was higher than that of any other subset. To assess whether this inequality was statistically significant, we conducted a post hoc analysis. In essence, we computed the marginal explanatory power of each selected variable X [namely, ΔR2(X)] with that of other, nonselected variables Y [namely, ΔR2(Y)]. We then performed a binomial test to examine whether the inequality ΔR2(X) > ΔR2(Y) was statistically significant.
We repeated the variable selection analysis separately for each monkey. Because the two datasets were smaller, we examined a reduced list of 13 variables including offer value A, offer value B, chosen value, other value, total value, value difference, value ratio, chosen number, other number, total number, chosen value A, chosen value B, and chosen juice.
OFC dataset
We compared the results obtained for GC with those previously obtained for OFC. For this comparison, we used a previously published dataset collected from two monkeys (H and I) by the same researcher (A.J.; Jezzini and Padoa-Schioppa, 2020). One of the animals (H) was the same that provided GC data. All experimental procedures and methods for data analysis used for OFC were identical to those described here for GC. A total of 608 OFC neurons were recorded in the 130 sessions that satisfied the behavioral criterion (see above, Analysis of choice patterns) and were included in the present analysis.
Neuronal classification
The analysis of neuronal responses indicated that different variables were most prevalent in different time windows. In principle, this observation is consistent with two alternative scenarios. (1) Each neuron might encode a single variable and be untuned in some of the time windows. Alternatively, (2) each neuron might encode different variables in different time windows. To disambiguate between these hypotheses, we divided time windows in early (post-offer, late delay, pre-go) and late (pre-juice, post-juice, post-juice2). We then classified each neuron separately in early and late time windows. We started with early time windows, and we focused on the three selected variables (chosen value, chosen number, chosen juice). For each variable and each time window, we calculated the signed R2, where the sign was that of the regression slope. For each variable, we computed the sum(R2) as the total signed R2 across the three time windows. The encoded variable was that providing the highest ||sum(R2)||, where ||·|| indicates the absolute value. The sign of the encoding was that of sum(R2). Neurons that passed the ANOVA criterion but were not explained by any variable were classified as untuned. We repeated this operation for late time windows. Thus, each cell was classified independently in both early time windows and late time windows. The results of these independent classifications were combined in a contingency table where entries represented cell counts.
We also computed the corresponding table of odds ratios (OR). Indicating with C the table of cell counts, for any entry (i, j), ORi,j was defined as follows:
The final classification of neurons was performed pooling early and late time windows. Again, we computed the sum(R2) across time windows and assigned the cell to the variable providing the highest ||sum(R2)||. The sign of the encoding was that of sum(R2). We thus obtained six distinct groups of neurons.
Analysis of activity profiles
To calculate activity profiles, trials were separately aligned at (1) offer on, (2) go signal, and (3) juice delivery onset. For each trial and each alignment, the spike train was smoothed as follows. Spike times, expressed in 1 ms resolution, were convolved with the kernel as follows:
Results
Two monkeys, E and H, performed a standard juice-choice task (Fig. 1A,B). In each session, the animal chose between two juices labeled A and B, with A preferred. Offers were represented by sets of colored squares and the monkey indicated its choice with a saccade. Offered quantities (qA, qB) and the spatial configuration of the offers on the monitor varied from trial to trial. Choice patterns presented a quality–quantity trade-off (Fig. 1C). For each session, a probit regression provided measures for the relative value of the juices (ρ) and for the sigmoid steepness (η).
Experimental design. A, Choice task. The trial began with the animal fixating a central location. After 1.5 s, two sets of colored squares representing the two offers appeared on the two sides of a fixation point. For each offer, the color indicated the juice type and the number of squares indicated the juice quantity. After a randomly variable delay (1–2 s), two saccade targets appeared by the offers (go signal). The animal indicated its choice with a saccade and maintained fixation on the peripheral target. After 0.75 s, the chosen juice was delivered. B, Time windows defined for neuronal analysis. C, Choice pattern recorded in a representative session. The percentage of B choices (y-axis) is plotted against the log quantity ratio [log(qB/qA); x-axis]. The sigmoid was derived from a probit regression. For this session, the relative value of the juices was ρ = 2.7.
Variables encoded in GC: example cells
We recorded the spiking activity of 1,178 neurons (920 cells from Monkey E, 258 cells from Monkey H). Recording locations were primarily from anterior insula (Fig. 2) and corresponded to the region termed GC (Scott et al., 1986; Rolls et al., 1988; Yaxley et al., 1988, 1990; Smith-Swintosky et al., 1991; Plata-Salaman et al., 1992). We analyzed firing rates in eight time windows aligned with different behavioral events (see Materials and Methods). A “trial type” was defined by two offered quantities and a choice. A “neuronal response” was defined as the mean activity of one cell in one time window, as a function of the trial type.
Anatomical reconstruction. Recording sites are shown on representative coronal sections created based on postsurgery MRI scans. All MRI scans were aligned to a recording grid. In each plot, recording sites were collapsed from different adjacent rostrocaudal locations as indicated for each section. Recording positions were obtained by projecting the dorsoventral and mediolateral coordinates to the aligned coronal sections, separately for each monkey. A–C, Monkey E. D, E, Monkey H. For each coronal section, two magnified images illustrate the number of cells recorded (divided in tuned and untuned; green dashed contour) and the number of cells encoding each of the three variables (chosen value, chosen number, chosen juice; red dashed contour). In each recording location, the radius of the circle indicates the number of neurons. In each location, the radius of the circle indicates the number of neurons recorded in the corresponding site (only tuned cells are included). The mappings between radius and cell numbers are indicated in the most caudal section, separately for each monkey.
Upon visual inspection, a sizeable fraction of responses appeared modulated by the trial type. In many cases, neuronal responses seemed to encode the variable chosen value. Figure 3A1 illustrates one example. In this plot, red diamonds and blue circles represent trial types in which the animal chose juice A and juice B, respectively. In Figure 3A2, the response was plotted against the variable chosen value, which was computed based on the relative value ρ derived from the animal's choices (Table 1). The two regression lines were obtained from linear fits restricted to trial types in which the animal chose juice A or juice B. Notably, the relation between the activity of this neuron and the variable chosen value was close-to-linear and did not depend on the identity of the chosen juice. In Figure 3A3, the same data points were plotted against the variable chosen number (i.e., the chosen quantity). In this case, the two linear regressions have substantially different slopes, consistent with the fact that each slope is proportional to the unit value of the corresponding juice. Plotting the response against the chosen juice (Fig. 3A4) confirmed that the cell activity did not depend on the juice type. Figure 3B illustrates another example of neuronal response encoding the variable chosen value.
Example cells. A, Response encoding the chosen value (time window, post-offer). In panel A1, offer types (x-axis) are ranked according to the quantity ratio (qB/qA). Black dots indicate the choice pattern and color symbols indicate firing rates. Red diamonds and blue circles indicate trials in which the animal chose juice A and juice B, respectively. The relative value derived from the probit fit (ρ = 2.1) is indicated in the panel. In panel A2, the same neuronal response is plotted against the variable chosen value (expressed in units of juice B; uB). The two lines are from linear regressions including only trials in which the animal chose juice A or juice B. Notably, the two regression lines are nearly identical, strongly suggesting that this response encoded the variable chosen value. In panel A3, the response is plotted against variable chosen number. Here the two regression lines are clearly distinguished (and the slope ratio is roughly equal to the relative value ρ). In panel A4, the response is plotted against the variable chosen juice. Here the black line is from a single linear regression including all data points. B, Response encoding the chosen value (time window, late delay). C, Response encoding the chosen number (time window, post-juice). In this case, regression lines in panel C2 (chosen value) are clearly distinguished, while regression lines in panel C3 (chosen number) are nearly identical, strongly suggesting that the response encoded the chosen number. D, Response encoding the chosen number (time window, post-juice2). E, Response encoding the chosen juice (time window, pre-juice). Here, only the linear regression in panel E4 captures well the neuronal response. F, Response encoding the chosen juice (time window, post-juice2). Conventions in panels B–F are as in panel A.
In other cases, neuronal response seemed to encode the variable chosen number. Figure 3C illustrates a representative example. When the response was plotted against the chosen value (Fig. 3C2), the two regression lines were substantially different. Conversely, when the response was plotted against the chosen number (Fig. 3C3), the two regression lines were essentially indistinguishable. Figure 3D illustrates another example of response encoding the chosen number.
Finally, other responses appeared to encode the variable chosen juice. Figure 3E illustrates one example. The activity of this neuron was essentially binary—it was high when the animal chose juice A and low when the animal chose juice B, and it did not depend systematically on the chosen quantity. Similarly, the response illustrated in Figure 3F was low when the animal chose juice A and high when the animal chose juice B and did not depend on the chosen quantity.
As previously noted in other brain regions (Padoa-Schioppa and Assad, 2006; Cai and Padoa-Schioppa, 2012; Jezzini and Padoa-Schioppa, 2020), the encoding of variable chosen value could be positive (increasing firing rates for increasing values) or negative (increasing firing rates for decreasing values). Similarly, the encoding of variable chosen number could be positive or negative. Interestingly, neuronal responses encoding the three variables chosen value, chosen number, and chosen juice appeared most prevalent in different time windows. Specifically, chosen value responses were the most prevalent in early time windows, following the offer and preceding the go signal. Conversely, responses encoding the chosen juice and the chosen number were most prevalent in late time windows, immediately before and after the onset of juice delivery.
Contrary to our initial expectations, we did not find a sizeable group of neurons encoding the value of individual offers (i.e., variables offer value A and offer value B).
Variables encoded in GC: population analysis
For a statistical analysis of the whole population, we proceeded in steps. First, for each cell and each time window, we conducted a three-way ANOVA (factors, offer type × position of A × movement direction), and we imposed a significance threshold of p < 0.001. As detailed in Table 2, 701/1,178 (60%) of neurons were significantly modulated by the offer type in at least one time window. In contrast, fewer cells were modulated by the spatial configuration of the offers (48/1,178 = 4%) or by the movement direction (155/1,178 = 13%; see Discussion). Thus we conducted a one-way ANOVA (factor, trial type). As detailed in Table 2, 709/1,178 (60%) neurons passed the criterion (p < 0.001) in at least one time window. These cells were identified as task related. Unless otherwise indicated, all subsequent analyses focused on this population.
Results of the ANOVAs
Following the procedures adopted in previous studies, we defined 19 candidate variables potentially encoded by neurons in GC (Table 1). These included variables associated with a single juice (offer value A, offer value B, etc.), a variable reflecting the binary choice outcome (chosen juice), variables reflecting the relative value of the two juices (chosen value, other value, etc.), and variables reflecting the numerosity (chosen number, total number, etc.). Of note, these variables were often correlated with each other (Fig. 4). For each response, we performed a linear regression on each variable, and we obtained the slope and the R2. If the slope differed significantly from zero (p < 0.05), the variable was said to “explain” the response. Figure 5 illustrates the results obtained for the whole population. For each time window, Figure 5A indicates the number of cells explained by each variable. Because any response could be explained by >1 variable, each response could contribute to >1 bin in this plot. In contrast, Figure 5B indicates the number of responses for which any given variable provided the best explanation (highest R2). Thus in this plot each response contributes to ≤1 bin. Inspection of Figure 5B reveals that the variable chosen value was the most dominant in early time windows. In contrast, most dominant in late time windows were variables chosen juice and chosen number (and possibly chosen value B, which, however, is highly correlated with both chosen juice and chosen number).
Pairwise correlation between variables. The variables examined in the analysis of neuronal activity (Table 1) were often correlated with each other. For each pair of variables x and y and for each session, we computed the pairwise correlation r(x, y) across trials. The gray scale image shown here illustrates the pairwise correlation averaged across sessions. Pairs of variables for which the mean correlation was >0.75 are highlighted with a while symbol (a circle for pairs including one of the three variables encoded in GC, a cross otherwise).
Linear regressions, population results. A, Explained responses. Each entry indicates the number of responses explained by one variable in one time window. For example, in the post-offer time window, the variable offer value A explained 32 responses. In the same time window, the variable chosen number explained 66 responses. Of note, since variable chosen number was highly correlated with other variables (total number, max number, etc.), many of these responses contributed to multiple bins in this plot. Different shades of gray represent the same results indicated numerically. B, Best fit. Here each entry indicates the number of responses for which the corresponding variable provided highest R2 (best fit). In this panel, each response appears in at most one bin. Notably, the variable chosen value dominates in early time windows, while variables chosen number and chosen juice are most prevalent in late time windows.
The 19 variables included in the analysis collectively explained 1,279/1,476 (87%) of responses passing the ANOVA criterion. To identify a small subset of variables that could best account for the whole population, we conducted a stepwise analysis and a best-subset analysis (see Materials and Methods). In the first two iterations, the stepwise procedure selected variables chosen juice and chosen value, which together explained 967/1,476 (66%) responses (Fig. 6A,B). Once these responses were removed, the residual dataset appeared rather distributed (Fig. 6A). In the third and fourth iterations, the stepwise procedure selected variables chosen value B and offer value B, but the marginal explanatory power of these variables (i.e., the fraction of responses exclusively explained by these variables) did not reach the 3% criterion. Thus these variables were eliminated. Next, the procedure selected variable chosen number, which explained an additional 114/1,476 (8%) of responses. All the variables selected in subsequent iterations failed the 3% criterion. In summary, the stepwise procedure selected variables chosen juice, chosen value, and chosen number, which collectively explained 1,081/1,279 (85%) of the responses explained by all 19 variables.
Variable selection analysis. A, Stepwise procedure. The top panel is as in Figure 5B. In the first two iterations, the procedure selects variables chosen juice and chosen value (red asterisks indicate selected variables). Variables selected in the next two iterations did not reach the 3% criterion (blue dots indicate selected-and-discarded variables). In the fifth iteration, the procedure selects variable chosen number. In subsequent iterations, selected variables never met the 3% criterion. B, Stepwise procedure, explained responses. The percentage of responses explained (y-axis) is plotted against the number of selected variables (x-axis). Here 100% on the y-axis represents the total number of task-related responses (N = 1,476). The dotted line indicates the number of responses collectively explained by the 19 variables defined in Table 1 (N = 1,279). Variables chosen value, chosen number, and chosen juice collectively explained 1,081 responses. C, Best-subset procedure, total R2 explained. For each response, R2max was defined as the highest R2 provided by one of the 19 variables. Thus we expressed the total R2 explained by a subset of variables as a percent of the theoretical maximum, defined as sum(R2max) across the population. In this panel, the percent total R2 (y-axis) is plotted against the number of variables in the best subset. Variables chosen value, chosen number, and chosen juice provided the best subset of three variables and explained 81% of the theoretical maximum R2. D, Final classification of neuronal responses. Each response was assigned to the selected variable that provided the largest R2. If none of the variables explained the response, the response was classified as untuned. Each entry in this table indicates the number of responses assigned to each variable in each time window.
Importantly, the stepwise procedure is path dependent and does not guarantee optimality. Thus we conducted a best-subset analysis, which is exhaustive and identifies the subset of variables providing the highest possible explanatory power. In particular, we selected variables based on the maximum total R2 (see Materials and Methods). Consistent with the results obtained with the stepwise procedure, the best subset of three variables included chosen value, chosen number, and chosen juice (Fig. 6C). This result was replicated separately in each monkey (see Materials and Methods).
To gauge the significance of these results, we conducted a post hoc analysis comparing the marginal explanatory power of the three selected variables with that of other, nonselected variables that were highly correlated with one of the selected variables (see Materials and Methods; Fig. 4). We compared 10 pairs of variables. As detailed in Table 3, all comparisons were statistically significant (p < 0.05), with two exceptions that only tended to significance [chosen value vs (chosen – other) value, p = 0.070; chosen number vs max number, p = 0.078]. We view these two cases as minor caveat on the conclusions drawn from the variable selection analyses.
Post hoc analysis
In summary, neurons in GC encoded variables chosen value, chosen number, and chosen juice. We summarized our results in a table indicating the number of responses encoding each variable in each time window (Fig. 6D).
Comparison with OFC
The variables identified for GC (chosen value, chosen number, chosen juice) partly differ from those previously identified for OFC (offer value A or B, chosen value, chosen juice). We sought to validate this result through a direct comparison of the two brain areas. To do so, we used a dataset of N = 608 cells recorded from OFC in the same experimental conditions used for GC and previously published (Jezzini and Padoa-Schioppa, 2020). For each area, we constructed a reduced table of best-fit data that were equivalent to the table in Figure 5B, except that we only included the relevant variables (offer value A|B, chosen value, chosen number, chosen juice). The collapsed variable offer value A|B was assigned an R2 equal to the highest of the two R2 provided by variables offer value A and offer value B. Visual inspection of Figure 7A,B revealed that the two brain areas differed both for the time course of task relatedness and for the encoded variables. Specifically, OFC was relatively more active in early time windows, while GC was relatively more active in late time windows. Furthermore, neuronal responses best explained by offer value A|B (especially in early time windows) were more prevalent in OFC than in GC. Conversely responses best explained by variable chosen number (especially in late time windows) were more prevalent in GC than in OFC.
Comparing OFC and GC. A, OFC. The panel illustrates a reduced table of best-fit data, analogous to that in Figure 5B, but including only variables offer value A|B, chosen value, chosen number, and chosen juice. Numbers and shades of gray indicate cell counts. B, GC. The same format as in panel A. C, D, Comparing OFC and GC, pooling time windows. For each area, we pooled best-fit data from all time windows. In panel C, numbers and shades of gray indicate cell counts. In panel D, numbers and shades of gray indicate ORs. For each variable (column), Fisher's exact test (two-tailed) was used to assess whether the two brain areas (rows) differed for the proportion of best-fit responses. The resulting p values are indicated below the OR table. E, F, Comparing OFC and GC, early time windows. The same analysis as in panels C and D restricted to early time windows (post-offer, late delay, pre-go). G, H, Comparing OFC and GC, late time windows. The same analysis as in panels C and D restricted to late time windows (pre-juice, post-juice, post-juice2).
To assess the statistical significance of these observations, we pooled responses from different time windows, separately for each brain area. We thus obtained a 2 × 4 contingency table where rows corresponded to brain areas (OFC, GC), columns corresponded to variables, and entries were the number of responses best explained by the corresponding variable (Fig. 7C). Our goal was to assess whether the two areas differed significantly for the distribution across variables. Since the total number of cells recorded in each area was not the same, cell counts in Figure 7C could not be compared directly. Thus we computed the table of ORs (Fig. 7D; see Materials and Methods). For each entry in this table, OR = 1 was the chance level; OR > 1 and OR < 1 indicated that cell counts were higher and lower than expected by chance, respectively. Significant departures from the chance level were assessed using Fisher's exact test. This analysis revealed that responses best explained by offer value A|B were significantly more frequent in OFC than in GC (p = 10−4). Conversely, responses best explained by chosen number were significantly more frequent in GC than in OFC (p = 4 × 10−10). Interestingly, responses best explained by chosen value were significantly more frequent in OFC (p = 4 × 10−6), while responses best explained by chosen juice were significantly more frequent in GC (p = 0.02). We repeated this analysis focusing on early time windows (post-offer, late delay, pre-go; Fig. 7E,F) and, separately, on late time windows (pre-juice, post-juice, post-juice2; Fig. 7G,H) and obtained corroborating results. In early time windows, when decisions took place, the fraction of neuronal responses best explained by offer value A|B was significantly higher in OFC than in GC. Conversely, both in early and in late time windows, the fraction of neuronal responses best explained by chosen number was significantly higher in GC than in OFC.
Importantly, even if one variable were not explicitly encoded in a particular area, we would expect some fraction of responses to be best explained by that variable, because different variables were substantially correlated in the experiments (Fig. 3) and because cortical firing rates are highly stochastic (Conen and Padoa-Schioppa, 2015). This is why chosen number entries in Figure 7A and offer value A|B entries in Figure 7B are nonzero. To confirm this point, we quantified for each area the additional number of responses that would be explained if we included an additional variable in the selected set. For OFC, 475 responses passed the ANOVA criterion (task-related). Variables offer value A|B, chosen value, and chosen juice collectively explained 356 (75%) of them. Adding chosen number to the selected variables would explain an additional four (<1%) responses. For GC, 1,476 responses passed the ANOVA criterion. Variables chosen value, chosen number, and chosen juice collectively explained 1,081 (73%) of them. Adding offer value A|B to the selected variables would explain an additional 77 (5%) responses. Note that adding offer value A|B effectively means adding two variables (offer value A and offer value B), implying that fraction of responses explained by each additional variable is <3%. These results validate the conclusions of the previous section and of earlier studies of OFC (Padoa-Schioppa and Assad, 2006).
Neuronal representation of juice value and juice quantity
Many neuronal responses in GC encoded the variable chosen value or the variable chosen number. For both groups of responses, the activity increased linearly as a function of the chosen quantity of juice, for both juice types. These responses are thus referred as “U-shaped.” As in previous studies of other brain regions (Padoa-Schioppa and Assad, 2006; Cai and Padoa-Schioppa, 2012; Jezzini and Padoa-Schioppa, 2020), we conducted further analyses to assess whether and how these responses reflected the subjective nature of value.
For each response in the dataset, we separated trials in which the animal chose juice A or juice B. For each group of trials, we regressed firing rates on the amount of juice chosen by the animal (as in the third column of Fig. 3). We thus obtained two slopes, βA for juice A and βB for juice B. If both slopes differed significantly from zero (p < 0.01) and had the same sign, the response was said to be “U-shaped”. For U-shaped responses encoding the chosen value, the slope ratio βA/βB should be equal to the relative value of the two juices measured behaviorally from the choice patterns, namely, ρ. Such U-shaped responses would reflect the subjective nature of value. Conversely, for neuronal responses encoding the chosen number, the two slopes βA and βB should be equal (i.e., slope ratio βA/βB = 1), independent of the relative value ρ of the two juices. Thus we examined the relation between the slope ratio βA/βB and the relative value ρ, taking advantage of the fact that, for any given juice pair, the relative value ρ varied from session to session.
We performed an analysis of covariance (ANCOVA) using the log slope ratio [log(βA/βB)] as a dependent variable, the log relative value [log(ρ)] as a predictor, and grouping data by the juice pair. Data recorded from different monkeys and with different juice pairs were grouped separately because relative values generally depend on the subject and on the juice pair. We used the full ANCOVA model, and we included only juice pairs with ≥8 U-shaped responses. This analysis was conducted separately in early time windows preceding the go signal (post-offer, late delay, pre-go) and in late time window following the motor response (pre-juice, post-juice, post-juice2). Our results are illustrated in Figure 8. In early time windows, the whole dataset was distributed on and around the identity line (Fig. 8A), and the same held statistically true for each of 10 juice pairs (Fig. 8B). This is the pattern predicted assuming that neuronal responses encode the chosen value. Conversely, in late time windows, the bulk of the dataset was distributed between the identity line and the abscissa (Fig. 8C), and the same held true for each of six juice pairs (Fig. 8D). This pattern is intermediate between those predicted assuming that neuronal responses encode the chosen value and that predicted assuming that neuronal responses encode the chosen number.
Analysis of U-shaped responses. A, Early time windows. Each color and shape represents one juice pair (see legend below the panel) and each symbol represents one U-shaped response. Data for the two monkeys (E and H) were grouped separately. Lines were obtained from an ANCOVA (full model). The inset illustrates the distributions predicted for responses encoding the chosen value (along the identity line) and for responses encoding the chosen number (along the abscissa). The distribution observed for the whole dataset closely resembles that predicted for chosen value responses. The ANCOVA provided an estimated mean slope, 0.87 ± 0.17 (mean ± SE), and estimated mean intercept, 0.09 ± 0.12 (mean ± SE). B, Regression estimates for individual juice pairs. Lines represent standard deviations. Color symbols along the y-axis indicate different juice pairs according to the legend in panel A. The distribution of individual slopes is roughly centered around the value of 1 (identity). C, D, Late time windows. Same format as in panels A and B. In panel C, the distribution observed for the whole dataset is roughly intermediate between those predicted for chosen value and chosen number. The ANCOVA provided an estimated mean slope, 0.66 ± 0.19 (mean ± SE), and estimated mean intercept, 0.06 ± 0.15 (mean ± SE). In panel D, the distribution of individual slopes is roughly between 0 (abscissa) and 1 (identity).
This finding confirms the observations resulting from the variable selection analysis. Early in the trial, neurons in GC represent the subjective value of the juice expected by the animal (chosen value). Later in the trial, the population of neurons in GC represents both the juice value and the juice quantity (chosen number).
Distinct groups of neurons encoded different variables
At the population level, the representation of variables chosen value, chosen number, and chosen juice in GC varied over the course of the trial (Fig. 6D). In principle, this might be because different subpopulations of neurons were active in different time windows; alternatively, any given cell might encode different variables in different time windows. For example, some neurons might encode the chosen value early in the trial and the chosen number later in the trial. To disambiguate between these hypotheses, we divided time windows in early (post-offer, late delay, pre-go) and late (pre-juice, post-juice, post-juice2; see Materials and Methods). We then classified each cell as chosen value, chosen number, chosen juice, or untuned, separately in early and late time windows. We examined the whole population and constructed a contingency table where each entry (i, j) indicated the number of cells classified as encoding variable i in early time windows and variable j in late time windows (Fig. 9A). We also generated a table of ORs, where each entry (i, j) quantified how the corresponding cell count departed from that expected by chance (Fig. 9B; see Materials and Methods). For each entry in this table, OR = 1 was the chance level and OR > 1 (OR < 1) indicated that the cell count was higher (lower) than expected by chance.
Decision variables encoded in early versus late time windows. A, Contingency table. Rows and columns represent variables encoded in early time windows and late time windows, respectively, and each entry indicates a cell count. Shades of gray are equivalent to the numerical entries. Cells that passed the ANOVA in some time window (either early or late) but were not explained by any variable in the time windows under examination were classified as untuned. B, Table of ORs. For each entry of panel A, we computed the corresponding OR. Asterisks indicate a significant departure from chance level (two-tailed Fisher's exact test; pink for OR > 1, blue for OR < 1; **p < 0.01; ***p < 0.001).
The results obtained from this analysis were noteworthy in several respects. (1) The vast majority (91%) of neurons encoding the chosen juice in late time windows were untuned in early time windows (Fig. 9A). (2) Similarly, the vast majority (87%) of neurons encoding the chosen number in late time windows were either untuned or encoded the chosen number in early time windows (Fig. 9A). Conversely, the incidence of neurons encoding the chosen value in early time windows and the chosen number in late time windows was below chance (OR = 0.68; Fig. 9B). (3) Neurons encoding the chosen value in early time windows either continued to encode the chosen value or became untuned in late time windows (Fig. 9B). Taken together, these observations indicate the presence of three disjoint populations of neurons encoding variables chosen value, chosen number, and chosen juice with different time courses. On this basis, we classified each cell based on both early and late time windows, separating neurons with positive and negative encoding (see Materials and Methods). We thus obtained six groups of cells: 99 cells and 45 cells encoding the chosen value with positive and negative encoding, respectively; 180 cells and 79 cells encoding the chosen number with positive and negative encoding, respectively; and 98 cells and 83 cells encoding the chosen juice A and chosen juice B, respectively.
Evolution of neural signals over the course of a trial
Our final analysis examined the activity profiles of these cell groups over the course of the trial. For each chosen value cells, we divided trials in tertiles according to the chosen value (low, medium, and high), and we computed the activity profile for each tertile. We then averaged the profiles across neurons, separately for cells with positive and negative encoding. Focusing on chosen value + cells (Fig. 10A), their activity became significantly modulated by the chosen value ∼200 ms after offer presentation. This modulation remained sustained throughout the delay, after the go signal, until juice delivery, after which it slowly decayed. The profile of chosen value − cells (Fig. 10B) had a similar time course but a much more modest modulation.
Population activity profiles. A, Chosen value cells with positive encoding (99 cells). To calculate activity profiles, we separately aligned the trials at offer on, go signal, and juice delivery onset. They were divided in three groups according to the chosen value (high, medium, low). For each trial, the spike train was smoothed. Resulting traces were averaged across trials and across cells. B, Chosen value cells with negative encoding (45 cells). C, D, Chosen number cells with positive encoding (C; 180 cells) and negative encoding (D; 79 cells). E, Chosen juice cells (181 cells). Neurons encoding variables chosen juice A (98 cells) and chosen juice B (83 cells) were pooled. The two traces correspond to trials in which the animal chose the juice encoded by the cells (juice E) or the other juice (juice O).
Neurons encoding the chosen number cells were analyzed in a similar way: we divided trials in tertiles according to the chosen number (low, medium, and high), we computed the activity profile for each tertile, and we averaged activity profiles across neurons, separately for positive and negative encoding. Focusing on chosen number + cells (Fig. 10C), their activity became modulated ∼250 ms after the offer, but the modulation remained modest throughout the delay and until juice delivery, after which it increased and reached its peak. A similar trend was observed for chosen number − cells (Fig. 10D), which in fact became modulated only after juice delivery.
For chosen juice cells, we pooled the two populations of chosen juice A and chosen juice B, and we defined juice E as that eliciting higher firing rates. The other juice was defined as juice O. For each neuron, we separated trials depending on the chosen juice (E or O), and we computed the two activity profiles. We then averaged activity profiles across cells. The activity of this cell group (Fig. 10E) became modulated by the chosen juice ∼350 ms after the offer. This modulation was modest, but it persisted throughout the delay. It increased shortly before juice delivery and reached its peak ∼500 ms after the onset of juice delivery.
In summary, different groups of neurons in GC encoded the subjective value, the quantity, and the flavor of the juice chosen, expected, and received by the animal. These three variables were represented by distinct groups of neurons and with different time courses. Specifically, chosen value cells became tuned shortly after the offer and remained tuned throughout the trial. Conversely, chosen number cells and chosen juice cells were predominantly tuned after juice delivery.
Discussion
Economic choices entail representing and comparing subjective values, and numerous results link these mental operations to neuronal activity in OFC. However, previous studies did not address the fundamental question of how subjective values are computed in the first place. In general, constructing the value of an offer requires integrating all the dimensions relevant to the ongoing choice. In broad strokes, one possibility is that disparate signals originating in different brain areas reach OFC and are integrated within OFC. Aptly, OFC receives anatomical input from all sensory modality and from several limbic regions (Ongur and Price, 2000). An alternative possibility is that offer values are computed, at least partly, upstream of OFC, and then transmitted to OFC where they participate in the decision process. Previous studies that let animal choose freely between different types of juices found that the representation of offers and values in OFC is good based and not spatial (Padoa-Schioppa and Assad, 2006). Thus we hypothesized that that the primary input for the computation of offer values would originate from the GC. The results presented in this paper effectively rule out this hypothesis. A large fraction of neurons (60%) in GC was significantly modulated by our choice task. However, the variables represented in GC—namely, chosen value, chosen number, and chosen juice—are computationally postdecision and thus not suitable to construct an offer value or to inform the ongoing choice.
Neuronal signals in GC were primarily nonspatial, although a fraction of neurons reflected the movement direction. Interestingly, this spatial signal was most prevalent immediately before and during the eye movement, at a time when nonspatial signals were least pronounced (Table 2). With this premise, the present study focused primarily on nonspatial signals. Some of these signals preceded the animal's response and reflected the subjective nature of value, and all of them pertained to the good expected and eventually obtained by the animal. In this sense, neuronal responses in GC appear consummatory in nature.
Previous studies in humans (Kuhnen and Knutson, 2005; Preuschoff et al., 2008; Hsu et al., 2009) and nonhuman primates (Mizuhiki et al., 2012; Yang et al., 2022) reported neuronal activity in the anterior insula representing economic decision variables. These experiments specifically focused on decisions under risk. In humans, most prominent signals were associated with risk and risk-prediction errors (Preuschoff et al., 2008). As a caveat, it is not clear whether the region examined here corresponding to GC is homologous to those described in the human papers. Conversely, the region examined here overlaps with that described by Mizuhiki et al. (2012) and Yang et al. (2022). Both these studies documented neuronal signals encoding the value, risk, and expected reward. Critically, in one of the studies, the behavioral paradigm did not include a binary choice (Mizuhiki et al., 2012). In the other study, the data analysis on encoded variables focused exclusively on forced choice trials (Yang et al., 2022). Consequently, the results do not disambiguate between value or risk signals associated with the decision input (offer value, offer risk) and value signals associated with the decision output (chosen value, chosen risk). In this respect, the present study contributes two novel and important notions. First, the presence of choice-related signals in GC is extended to choices that do not involve any stochasticity or risk. Second and most important, choice-related signals in GC are computationally postdecision and thus cannot contribute to the ongoing choice.
Our results shed new light on the neuronal representation of taste, flavor, and value in GC. This region receives input from the gustatory thalamus and is interconnected with OFC and the amygdala (Mufson and Mesulam, 1982; Jezzini et al., 2015). Classic studies in monkeys described labeled-line responses to the five basic tastes (Yaxley et al., 1988; Smith-Swintosky et al., 1991), while other works suggested the presence of a distributed representation of multiple tastes (Katz et al., 2002; Jezzini et al., 2013). Subsequently, it was shown that responses in the rodent GC reflect multiple dimensions (odor, texture, temperature) that contribute to flavor (Small, 2012; Vincis and Fontanini, 2019). The results presented here refine and extend previous notions. They demonstrate the presence in GC of different groups of neurons encoding the flavor (chosen juice), the quantity (chosen number), and the subjective value (chosen juice) of the juices expected and received by the animal. Thus neuronal responses in GC are not purely sensory and reflect the subjective experience of the animal.
To conclude, two main results emerge from the present study. First, contrary to a reasonable hypothesis, GC does not provide input or otherwise contribute to ongoing choices between comestible goods. Second, neurons in GC respond to food intake in a variety of ways. More specifically, distinct groups of cells represent the flavor, the quantity, and the subjective value of the juice animals expect and receive. With respect to the construction of offer values, the present results rule out a direct role for GC, but they leave open multiple hypotheses. In particular, since in our choice task juice offers are represented by visual stimuli, a credible candidate to provide the critical input to OFC is the inferotemporal (IT) cortex. High visual areas in this region send direct projections to central OFC (Ongur and Price, 2000) and represent neural signals relevant to value assignment (Sasikumar et al., 2018). Thus future work shall examine the role of IT in value construction.
Footnotes
We thank M. Barretto-Garcia, Y. Kanazawa, A. Livi, and M. Zhang for their comments on the manuscript. This research was supported by the National Institutes of Health (Grant Number R01-MH104494 to C.P-S.).
The authors declare no competing financial interests.
- Correspondence should be addressed to Camillo Padoa-Schioppa at camillo{at}wustl.edu.
