Sensory and palatability coding of taste stimuli in cortex involves dynamic and asymmetric cortico-amygdalar interactions

....................................................................................................................................... 3 9 Significance Statement ................................................................................................................ 4 10

................................................................................................................................. 5 11

........................................................................................................................................ 8 12 Overview .................................................................................................................................. 8 13 During taste processing, BLA and GC population firing rate transitions happen with zero lag ..... 10 14 BLA-GC interaction dynamics are asymmetric but time-locked ................................................ 13 15 Poisson Generalized Linear Model reveals overlapping groups of GC neurons influencing and 16 influenced by BLA neuron activity ............................................................................................ 20 17 Taste processing in GC is a function of embeddedness in the amygdala-cortical loop ............ 27 18

................................................................................................................................. 30 19 Amygdala-cortical taste activity evolves as a joint attractor network ......................................... 31 20

from research on other sensory systems are consistent with these conclusions ........... 32 21 The function(s) of BLA-GC directional influences ..................................................................... 33 22 The role of projection neurons in sensory processing ............................................................... 34 23

............................................................................................................................ 36 24

.................................................................................................................................... 38 25 Experimental Design and Statistical Analyses .......................................................................... 38 26 Subjects ............................................................................................................................. 38 27 Electrode and intra-oral cannula construction ..................................................................... 38 28 Surgery ............................................................................................................................... 39 29 Acquisition of electrophysiological data ............................................................................... 39 30 Habituation and passive taste administration....................................................................... 39 31 Histology ............................................................................................................................ 40 32 Data and statistical analyses .................................................................................................. 41 33 Local Field Potential processing and analysis ....................................................................... 41 34 Analysis of Spiking Activity .................................................................................................. 44 35 Changepoint Modeling of Population Activity ....................................................................... 46 36 Generalized Linear Modelling .............................................................................................. 47 37 Data Availability ..................................................................................................................... 49 38

................................................................................................................................ 50 39 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 3 40

41 Reciprocal connectivity, which generates nonlinear dynamics within a system, should do the same 42 in the taste circuit, which is rife with between-region feedback. Much evidence supports the 43 existence of specific, characterizable within-region dynamics in taste processing, but scant 44 attention has been paid to dynamic inter-region interactions in taste processing. To fill this gap, we 45 apply investigate simultaneous recordings from rodent Gustatory Cortex (GC) and Basolateral 46 Amygdala (BLA) , testing specific hypotheses about reciprocal amygdala-cortical interactions. We 47 find that initial ly GC and BLA responses are independent, but evolve into synchronous, zero -lag 48 ensemble transitions ( surprising given the long axons connecting the two regions) within a few 49 hundred milliseconds of taste delivery. Spectral Granger Causality (which infers directional 50 influences) revealed that this tight synchrony also characterizes the system’s asymmetric inter-51 region influences wherein the BLA →GC influence is dominant prior to the generation of the 52 behavioral response and the GC→BLA influence bec omes strong at the time that GC has been 53 shown to release a behavior-relevant signal. To better understand the function of single neurons in 54 this process, we then used Poisson Generalized Linear Modeling to categorize GC neurons in terms 55 of their inferred connectivity. This analysis revealed that GC neurons that both influence and receive 56 influence from BLA—the neurons most deeply embedded in the reciprocal circuit—are the ones 57 most strongly involved in tast e processing (quantified as taste-specificity and palatability -58 relatedness). These results, which are consistent with findings in multiple systems and species, 59 support the notion that taste processing is a function of the amygdala-cortical loop. 60 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 4 Significance Statement 61 Consensus is growing that taste (and other sensory) circuitry should be considered as dynamical 62 systems; this theoretical stance requires studying reciprocal interactions. Building on previous work 63 on dynamic cortico-amygdalar interactions, here we show that in response to taste stimuluation, 64 BLA and GC (following an initial transient) coalesce into a single processing network with zero-lag 65 population state transitions; within this unified structure, GC and BLA “take turns” influencing one 66 another. We go on to show, in single -neuron analyses, that taste processing is a function of the 67 amygdala-cortical loop—neurons that are most deeply embedded in that loop are those that code 68 taste most strongly . In sum, this research suggests a unified framework for conceptualizing the 69 evolution of sensorimotor transformation in general. 70 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 5

71 The taste circuit, like those supporting the other senses, is not a feedforward processing stream; 72 rather, most pairs of regions involved in taste processing are reciprocally connected (1–3). Thus, each 73 region receives feedback from its targets at the same time that it influences those targets, a fact that 74 should cause the system to express nonlinear dynamics when activated (4–6). Specifically, several 75 specific, novel predictions can be made about taste processing within pairs of reciprocally-76 connected regions : 1) that each region should “code” tastes dynamically; 2) that these regions 77 should become “coupled” into single processing units when activated by tastes ; 3) that taste 78 processing should involve highly dynamic “conversations” between coupled pairs of brain regions; 79 and 4) that single neurons’ involvement in taste processing should depend upon being a part of that 80 conversation—i.e., being embedded within the task-relevant reciprocal circuit. 81 While the role s/functions of individual brain regions in taste processing ha ve been extensively 82 investigated—investigation which has produced ample evidence for hypothesis 1 above (see 83 below)—the impact of reciprocal circuitry on interactive processing among these regions 84 (hypotheses 2-4) have received far less study. Only a small set of experiments has investigated the 85 impact that perturbations of neural activity in one region have on firing in another (7–12), and a 86 smaller number have analyzed simultaneous electrophysiological recordings from multiple regions 87 (13–16). In general, these studies support the hypothesis that between-region interactions have a 88 role to play in taste and taste learning (as also appears to be the case in other systems). Because 89 they have without exception used simple, non -directional correlation measures, however, these 90 studies were unable to investigate hypotheses related to the “ebb -and-flow” of directed inter-91 regional influences. Asymmetries (17–21) and dynamics of interaction (4,22,23) have been profitably 92 studied in other sensory and decision -making systems, but the kinds of techniques used in these 93 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 6 studies have not yet been applied to taste processing . In sum , previous work on inter -region 94 interactions in taste processing has left largely unaddressed questions regarding which regions and 95 neuronal populations exert specific influence on other regions , when in real time they do so, and 96 what the implications of these interactions might be for our understanding of taste processing. 97 We have recently begun an investigation into such real-time interactions, focusing on the two most 98 prominently discussed nodes of the forebrain taste circuit —basolateral amygdala (BLA) and 99 gustatory cortex (GC , 16). During tasting, neural responses in both structures show evidence of 100 nonlinear dynamics, in that they progress through a sequence of distinct processing states (24,25) 101 separated by sudden, coherent ensemble transitions (16,26; see Fig. 1A); furthermore, activity in the 102 two regions becomes correlated (at the level of single -neuron pairs and population activity) during 103 production of these sequences (16), a result suggesting that BLA and GC may form part of an 104 attractor network (5,27,28). Thus, these responses likely reflect reciprocal BLA-GC interactions (via 105 BLA→GC and GC→BLA axonal projections; 29–32) involved in processing tastes. 106 Here we leverage this current understanding of taste response dynamics to generate novel insights 107 into the systemic nature of taste processing. We first demonstrate that BLA and GC are operating on 108 a single “processing clock” —despite relatively lengthy axons separating the structures , ensemble 109 firing rate transitions occur with zero between-region lag. Spectral Granger Causality brought to bear 110 on simultaneously-recorded local field potentials (LFP) provides converging evidence of this time -111 locking of inter-regional dynamics, but also shows that the influence that BLA exerts on GC during 112 particular epochs of the response differs from the “feedback” influence: in the µ/α frequency band 113 (2-12 Hz) , processing across the period of taste processing is dynamic, with BLA influencing GC 114 during the 300 -1000ms post-stimulus delivery period—the state during which BLA responses are 115 palatability-related (13,16,25), and just before GC responses become so (26,33); GC influence on 116 BLA in this frequency band, meanwhile, peaks only briefly at the end of this period (approx. 1000ms 117 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 7 following stimulus delivery) , around the time of the transition into the state in which GC emits 118 behavior-related signals (34). 119 Finally, we connected these results back to spiking in single neurons using Poisson Generalized 120 Linear Modeling (pGLM), which allows us to divide our GC sample into distinct neural subpopulations 121 based on whether they influence BLA, are influenced by BLA, or both. Analysis of this functional cell-122 type division revealed that GC neurons with taste responses that are influenced by BLA activity show 123 significantly more taste-specific and palatability-related firing than neurons that do not , and that 124 neurons that both influence and receive influence from BLA exhibit the strongest taste specificity and 125 palatability-relatedness. That is, our results demonstrate that the degree to which a GC neuron is 126 embedded in the reciprocal circuit largely determines its involvement in taste processing. 127 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 8

128 Overview 129 Our a nalysis of amygdala -cortical taste responses has 3 parts, beginning with an evaluation of 130 between-region lags in the times of the 3 sudden ensemble firing -rate transitions that characterize 131 the evoked taste-response, and the finding that all but the 1st transition are essentially simultaneous 132 in BLA and GC. The simultaneity of these later transitions shows that BLA and GC become co upled 133 while processing tastes, such that by the onset of palatability-related firing in GC (at the latest), they 134 are essentially showing no sign of the reliable delays that would be expected if these sudden firing 135 rate changes were being “transmitted” from one region to the other via between-region axons. 136 This finding does not , however, imply that BLA and GC necessarily play identical roles in taste 137 processing, any more than the fact that two individuals engaged in a single real -time conversation 138 are necessarily saying the same thing. We therefore go on to use Spectral Granger Causality—a tool 139 that essentially breaks down the degree and timing of each region’s influence on activity in the other 140 (in each frequency band)—to test for “asymmetries” in BLA-GC communication, by which we mean 141 differences between BLA’s influence on GC and GC’s influence on BLA (note that we use the term 142 “influence” to reflect in general terms the fact that one region’s activity influences the other, without 143 claiming additional mechanistic insight such as a distinction between direct and indirect projections 144 or excitatory and inhibitory c onnections). This analysis revealed that the strength of the GC→BLA 145 influence is continuously high in frequency bands above 20 Hz; the oft-described taste processing 146 dynamics, meanwhile, “live” exclusively in asymmetries of influence observed in the µ/ α band (i.e., 147 below 12 Hz; (15,35–37), where BLA influences GC during the 300-1000ms post-stimulus delivery—148 the “middle epoch,” during which BLA firing is palatability -related and GC’s is not (Fig. 1A, and see 149 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 9 11,15,16,24,34,38)—and GC briefly influences BLA just afterward, at the moment that the GC 150 response becomes palatability -related in a transition of firing that has been shown to influence 151 behavior (34). These direction-specific influences, however distinct, change in near -perfect lock -152 step, providing converging evidence that the two regions are behaving as a single processing unit. 153 To better assess the functional implications of these results, we then use Poisson Generalized Linear 154 Modeling to conservatively identify single neurons with inter-regional influence—GC neurons that 155 influence activity in BLA, GC neurons that are influenced by activity in BLA, and neurons that do both. 156 We then evaluated the taste responses of the functionally defined neuron groups; these analyses 157 reveal that involvement in taste processing is determined by the degree to which a neuron is 158 embedded in the reciprocal interaction between the two structures , with GC neurons that both 159 influence and receive influence from BLA showing the largest magnitudes of both taste-specific and 160 palatability-related taste responses. 161 162 163 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 10 During taste processing, BLA and GC population firing rate transitions 164 happen with zero lag 165 Our overarching hypothesis is that BLA and GC, both of which undergo multiple ensemble firing-rate 166 transitions during taste responses ( 16), can be reasonably described as functioning as a single, 167 distributed network (5,27,28,39–41). Such a hypothesis generates the specific prediction that these 168 firing-rate transitions should be simultaneous in the two regions (i.e., there should be zero between-169 region lag), because the transitions in an attractor network are generated by coordinated activity of 170 the regions with no specific “sources” or “epicenters” of the transitions. If BLA and GC cohere into a 171 single network, the oft-described transitions in ensemble firing rates that make up taste responses 172 in both regions (16,34,38,42) aren’t expected to originate in either one, but rather to emerge within 173 the network as a whole (28,40). If, on the other hand, transitions either spread from one region to the 174 other or are generated in a source common to the two , one would likely observe a consistent non-175 zero lag structure (see Discussion). Having shown (16) that transition latencies in GC and BLA are 176 correlated, here we begin by testing the prediction that these transitions are zero-lag. 177 To test this prediction, we apply a model that infers states and state-transitions in stimulus-driven 178 activity to each region independently (16,26,38,42). This analysis shows that such transitions occur 179 in the expected time ranges: in each region and trial, a first transition appears at 300±250ms following 180 taste delivery (but see below) , a second at 850±300ms following taste delivery, and a third at 181 1450±325ms following taste delivery (mean±std, rounded to the closest 25ms for clarity). As shown 182 previously, the precise timing of these transitions varies widely from trial to trial. 183 Fig. 1A, which summarizes the basic taste-processing dynamics that have been identified in GC and 184 BLA (16,25), identifies the transitions as they are used throughout the se figures. Fig. 1B is a 185 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 11 schematic of the analysis comparing the timing of these lags, and Fig. 1C shows the distributions of 186 calculated GC-BLA lags for each transition. Examination of Fig 1Ci, which presents the distributions 187 for each individual session as a heat map , suggests that the between-region lags tend to be highly 188 reliable across the 18 sessions (from 6 rats), appearances that are confirmed by the tightness of the 189 distributions of mean between-region lags for each session and transition (Fig 1Cii). Analysis of these 190 latter distributions reveals that, for the earliest transition, sudden changes in BLA ensemble firing 191 rates tend to lead those in GC by ~100ms (permutation test, 95%CI=[57, 183], p<0.001); we were 192 unable, however, to identify significant between -region lag s for the following two transitions 193 (permutation test, Transition 2: 95%CI=[-74.5, 51.5], p=0.78; Transition 3: 95%CI=[-86, 39.5], 194 p=0.43). When we, a s a further test, compared the distribution of average lags across transitions , 195 transitions 2 and 3 both proved to be significantly different from transition 1 (1 vs. 2: p=0.001, 1 vs. 3: 196 Wilcoxon signed-rank test, p=0.015), the average lags for transitions 2 and 3 , meanwhile, were not 197 significantly different from each other ( Wilcoxon signed-rank test, p=0.747). These data therefore 198 suggest that, after beginning their taste responses at different times, BLA and GC become not only 199 coupled but synchronized, undergoing essentially simultaneous transitions into (and out of) the GC 200 palatability-related state (Fig. 1A; see below for convergent evidence). 201 This pattern of results supports the hypothesis that BLA and GC, rather than being a joint processing 202 network “at rest,” coalesce into a dyad only after producing uncoordinated initial transients in their 203 taste responses. The lag present in the first transition suggests that BLA and GC taste responses are 204 initially driven independently, although there are other possible explanations; regardless, this initially 205 uncoupled activation gives way to the two regions cycling through the later transitions of their taste 206 responses as a single processing unit, such that the relatively long axons connecting the two regions 207 fail to induce delays in the firing-rate transitions that characterize taste processing (see Discussion). 208 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 12 209 Fig 1. BLA initially leads GC in state transition but both regions subsequently show zero-lag 210 coordination. A) Summary schematic of the timing and “coding content” previously demonstrated 211 to characterize state sequences in GC and BLA ensemble spiking activity. Overlain are the numbers 212 used in this study to label these transitions. B) A schematic illustrating how transition lags between 213 GC and BLA were calculated and how we define the sign of the lags . Ci) Distributions of transition 214 lags, presented as heat maps (legend at right shows the # of trials with a specific transition lag ) for 215 18 sessions (y-axis) across 6 animals for each of the 3 transitions. Cii) Summary distributions of lags 216 for each transition , with red arrow s indicating the grand-average lag. The s ignificance of the 217 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 13 difference between that grand-average lag and 0 was evaluated using bootstrapped 95% confidence 218 intervals of the distribution mean (see text). 219 BLA-GC interaction dynamics are asymmetric but time-locked 220 Even seamless zero-lag coupling of firing-rate transitions in BLA and GC does not necessarily imply 221 that each has identical influence on the other, any more than a deduced general function of a single 222 brain region negates the possibility of differences between the action of distinct subpopulations of 223 neurons. Rather, it implies only that the degree to which one region influences the other should be 224 dynamic, and that any such dynamics of this influence should occur in concert with dynamics in the 225 feedback influence (see Discussion for consideration of other coupled systems in which asymmetric 226 patterns of influence nonetheless evolve in concert). With that in mind, any complete understanding 227 of the function of this amygdala -cortical dyad will necessarily involve an examination of the 228 dynamics of directional influence—of the moment-to-moment degree to which activity in each area 229 impacts activity in the other. 230 We performed this examination by applying Spectral Granger Causality (GrCa) to LFPs recorded from 231 BLA and GC . Rather than simply assessing similarity between timeseries as a measure of 232 coordination (as done in a cross-coherence or correlational analysis), GrCa allows us to disentangle 233 directional dependencies by making use of time-lagged regressions: in essence, GrCa assumes that 234 if the future of timeseries A is better predicted using both the histories of timeseries A and B 235 (compared to just using the history of timeseries A), then B can be thought to “causally” influence 236 the future of A (43,44) broken down by frequency (45–47). Use of GrCa allows us to obtain a time-237 resolved view (a “big-picture” view, owing to the use of LFPs as a signal) of the dynamics of interaction 238 between BLA and GC. Significance of influence in GrCa can then be assessed for each time and 239 frequency bin. 240 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 14 Fig. 2 shows the output of this analysis brought to bear on simultaneously recorded GC and BLA local 241 field potentials. Collapsed across time, the influences of the regions on one another are almost 242 entirely contained within the 1-60Hz range (Fig. 2A; note that the low-pass filter used when collecting 243 the LFPs allows for the detection of influence well above 60Hz, see Methods), and quite distinct from 244 one another within that range (Fig. 2B): GC→BLA influence, while above zero across this entire range 245 of frequencies, is highest (and dominant) from 20 to 60Hz (the β and γ frequency ranges; see Buzsáki 246 and Draguhn, 2004 ); BLA →GC influence, meanwhile, which is essentially non -existent (and non -247 significant) across most of the β/γ range, is dominant in frequencies (2-12Hz, the θ and µ range, the 248 latter of which is sometimes referred to as α) which are known to be both particularly prominent in 249 GC evoked activity and modulated in a manner related to the dynamics of GC taste responses (35–250 37). 251 We note that it is unlikely that this relatively clean separation in directional frequencies is an artifact 252 of the methodology, as GrCa can appropriately recognize influences in identical or overlapping 253 frequency bands (18,19,43,49). Furthermore, the distinctness between the frequency bands of the 254 BLA→GC and GC →BLA influences aligns with previous work show ing that pharmacological 255 inhibition of taste thalamus reduces GC LFP power in the 20 -60Hz band, but increases it in the 2-256 12Hz band (15)—a finding that suggests that the 2-12Hz band may be the “channel” for amygalo -257 cortical interaction, whereas the 20 -60Hz band may be the “channel” for the thalamo -cortical 258 interaction (19,20, see Discussion). 259 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 15 260 Fig. 2) GC-BLA interaction shows processing-related asymmetries and dynamics. A) Mean (bold 261 lines) and individual sessions’ (thin lines) significance of Spectral Granger Causality for 0 -2000ms 262 post-stimulus activity across datasets for BLA →GC (black) and GC →BLA (red) influences. B) 263 Directional differences in mean (bold red line) and individual subjects’ (thin black lines) influence 264 across frequency bands for data shown in A). Above -zero differences mean stronger BLA →GC 265 influence, and below -zero differences mean stronger GC →BLA influence. BLA →GC is stronger 266 below 12Hz, and GC→BLA dominates between 20 and 60Hz; ** = p<0.05, *** = p<0.005 C) The data 267 from Fig. 2A unpacked into significant time-frequency bins for i) BLA→GC and ii) GC→BLA influence. 268 Brighter color indicates that a larger fraction of the recording sessions displayed significant Granger 269 Causality at that time-frequency bin (heatmap legend to right). In both panels the difference between 270 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 16 higher and lower frequency bands can be seen, and across panels the asymmetry in influence can 271 be seen; furthermore, dynamics of influence are apparent in the lower frequency band. D) The data 272 shown in C collapsed across frequencies within each identified bandwidth (frequency band limits 273 are given above each panel). Dynamics of influence are restricted to the <12Hz bandwidth; shaded 274 regions indicate periods of strongest significance determined using cluster-permutation testing. See 275 text and Methods for more details. 276 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 17 This difference between low and high frequencies is easily seen in plot s of the significance of 277 direction-specific influence through time (Fig. 2C; compare the upper two thirds to the lower third in 278 each panel), as is the difference between GC →BLA and BLA →GC influence (compare the left and 279 right panels). This presentation further suggests that, in general, β/γ influence may be slightly higher 280 spontaneously than during taste processing, and that the differences between BLA→GC (Fig. 2Ci) 281 and GC→BLA (Fig. 2C ii) influence in θ/µ frequencies change repeatedly across taste processing 282 time—that they reflect the oft-described GC dynamics. 283 To rigorously assess these seeming phenomena, we plotted the average proportion of sessions in 284 which power in each frequency-time bin is significantly above zero (the “significant fraction”) in each 285 direction of influence, collapsed across the frequencies with in the two dominant bands . This 286 analysis reveal s near-tonic influence s in the higher frequency band (Fig. 2 Di and ii; the average 287 GC→BLA magnitude of influence is of course higher than that of BLA→GC influence), although there 288 is a slight trend toward a reduction of influence following taste delivery . The µ/α band results , 289 meanwhile, accord well with our understanding of the known GC taste-response dynamics (Fig. 2Diii 290 and iv). In each panel, the shaded regions represent the result of cluster-based permutation testing, 291 identifying the longest continuous period s in which this result is significantly higher than that of 292 randomized data (there are no such periods in Fig. 2Di/ii). With regard to BLA→GC influence in the 293 lower frequency band ( Fig. 2Diii ), this period is a good match to the average GC identity epoch 294 (approx. 300-1000ms post-stimulus, see Fig. 1A). Given that BLA responses are palatability-related 295 during this period, and that GC responses become palatability-related only at the end of this period, 296 this result also accords well with classic studies on the role of amygdala in stimulus valuation, which 297 have led both human and rodent researchers to theorize that amygdala “passes” emotional valence 298 to cortical regions (7,11,51,52, see also Discussion). GC→BLA influence in this frequency range , 299 meanwhile (Fig. 2Div ), peaks only briefly at the end of th is BLA→GC push . This finding further 300 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 18 corroborates the above interpretation, suggesting that GC, after being influenced by BLA, emits 301 decision-related activity capable of influencing behavior, and as such it is also consistent with work 302 demonstrating that the transition into palatability-related firing in GC (which typically occurs around 303 this time) is causally linked to the driving of behavior (26,34). 304 Again, the fact that this amygdala-cortical interaction dynamic is asymmetric, in that BLA strongly 305 influences GC during a period of post -stimulus time in which GC does not influence BLA (and vice 306 versa), was not unexpected. If the two regions truly form a recurrently connected system, however, 307 one would further expect that even these asymmetries of influence should run on the same “clock”—308 that is, that changes in influence should occur simultaneously. To test this hypothesis, we performed 309 Principal Component Regression on each time-series shown in Fig. 2C. Fig. 3A illustrates the 310 trajectories for the data in the two Fig. 2C panels obtained from this regression overlain on one 311 another. T he dynamics of the two timeseries , which are visualized as changes in the trajectories’ 312 directions through space, are clearly very similar. We quantified this seeming similarity by assessing 313 the mean -squared error (MSE) of linearly predicting one time -series from the other (Fig. 3B , red 314 dashed line) compared to that produced by the same linear regression of the two time -series after 315 shuffling them temporally (linear, inter, shuffled), and also by linear and non-linear (i.e., using multi-316 layered artificial neural network) auto-regression (i.e., predictions of each time series from itself; see 317 Methods) to give us a lower-bound for prediction error. The results of these analyses make it clear 318 that the linear prediction of the BLA→GC time-series from the GC→BLA time-series was almost as 319 accurate as the non-linear auto-regression and better than the linear auto-regression (this was not 320 surprising given the nonlinearities in the BLA-GC interaction dynamics). Simply put, the inter-series 321 prediction nears the limit of prediction accuracy , a fact that serves as converging evidence for our 322 hypothesis that BLA and GC form, upon receiving taste input, a single distributed dyad with time-323 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 19 locked, coupled dynamics—albeit dynamics within which GC→BLA and BLA →GC influences are 324 asymmetric. 325 326 327 Fig. 3) Dynamics of BLA→GC and GC →BLA influences are time -locked. A ) Aligned principal 328 components of spectra shown in Fig. 2C, indicating that the dynamics of BLA →GC and GC →BLA 329 influence are strongly coordinated , in that the trajectories change direction together through time . 330 Elapsing post-stimulus time is shown with changing colors (legend to left), and taste delivery time is 331 shown as a red dot. B) Quantification of alignment shown in A). Prediction of one time -series 332 (GC→BLA) from the other (BLA →GC) using linear regression (Linear, Inter) is better than both 333 shuffled linear regression (Linear, Inter, Shuffled), and linear autoregression (Linear, Auto) of 334 timeseries, and almost as good as non-linear autoregression (Non-linear, Auto). 335 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 20 Poisson Generalized Linear Model reveals overlapping groups of GC 336 neurons influencing and influenced by BLA neuron activity 337 As noted above, the finding that different patterns of influence appear to “live” in different frequency 338 bands is not without precedent—bottom-up and top-down influences in visual pathways have also 339 been suggested to segregate into separate frequency bands (18,19,53). Meanwhile, the fact that 340 BLA→GC influence is focused in the 2-12Hz range is intriguing, given previous studies showing that 341 GC taste-processing dynamics are also prominently visible in GC µ activity (15,35–37): these facts 342 add to the evidence that BLA and GC become a single dynamical object while processing tastes, 343 further motivating our hypothesis that the involvement of GC neurons in taste processing (which can 344 be quantified in terms of the magnitude of GC responses’ taste -specificity and palatability -345 relatedness) might be a function of their coupling with BLA. 346 To test th is hypothesis, we trained Poisson Generalized Linear Models (pGLM, see Methods for 347 important differences between pGLM and the more commonly-used General Linear Model) to infer 348 connectivity and influence between individual GC-BLA neuron pairs —an aim that cannot be 349 achieved with GrCa (see Methods) , and th at allows us to categorize individual GC neurons as 350 putatively “sending” and/or “receiving” inter-regional influence (or both). We fit pGLMs to single 351 neurons in our dataset, using spike trains from all other simultaneously recorded neurons as input 352 (54–56). An advantage of this analytic approach to inferring connectivity is that, while it cannot 353 capture the dynamics of influence, it is inherently more conservative than correlational methods 354 such as spike-train cross-coherence (57). This is because stimulus input and the history of spiking in 355 the current neuron “compete” against one another, and with the recent spiking of other neurons (Fig. 356 4A), for significance during pGLM inference. Evoked activity of a neuron can potentially be explained 357 as a function of any of these variables, and therefore only explanatory value above and beyond what 358 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 21 is predicted by all other variables is deemed evidence of a functional inter -regional influence —359 “influence” that can be attributed to other sources is essentially “partialled out .” Standard 360 correlational approaches are significantly more prone to false positives due to the shared latent 361 dynamics across neurons within regions and across regions (55). 362 We fit models to 151 GC and 160 BLA well-isolated single neuron waveforms (mean±SD per session: 363 9.4±4.14, GC; 10.0±4.61, BLA). Before we could test our central hypothesis (that taste processing is 364 a function of embeddedness in the amydala-cortical loop), it was first necessary to test whether the 365 models successfully fit the data. Toward that end, Fig. 4B shows 2 representative examples of actual 366 PSTHs (left) and PSTHs predicted by the fit pGLMs (right). The apparent similarity between the two is 367 quantified in Fig. 4C, which shows that the model consistently produces significantly better fits to 368 the data than models fit on spike-trains shuffled between trials (Wilcoxon signed -rank test, p < 369 0.001), a control that preserves the average statistics of the spike -trains used, eliminating only the 370 within-trial coordination of spiking between neurons . This result highlights the fact that , despite 371 significant trial-to-trial variability in spiking, the pGLM learns relationships that are meaningful within 372 trials. 373 As an additional test of the pGLM fits, we also compared correlations between actual PSTHs and: 1) 374 trial-matched predictions; 2) cross-unit predictions; and 3) predictions on circularly shuffled data 375 (Fig. 4D, see Methods for details ). We find, as expected , that trial -matched predictions (blue 376 histogram and summary box plot) perform best, followed by unit-shuffled/cross-unit predictions (in 377 orange); we suspect that these control correlations were significantly above zero because many GC 378 neuron pairs tend to exhibit broadly similar neural dynamics. Worst were correlations with circularly 379 shuffled (temporally garbled) data , which were essentially zero (in green). Statistical comparisons 380 confirm that t rial-matched predictions are significantly better than either control comparison 381 (Wilcoxon signed-rank test, p<0.001 for each comparison). Together, these tests allow us to conclude 382 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 22 that the poisson pGLM generates reasonable models of the mechanics ( functional circuitry) 383 underlying the observed activity. 384 385 Fig. 4) Generalized Linear Modelling of single-neuron interactions between BLA and GC. A) pGLM 386 structure. Every neuron model depends on: 1) stimulus timing, 2) the neuron’s spiking history, and 3) 387 input from (or spiking history of) all other simultaneously recorded neurons. B) Actual and Predicted 388 PSTHs of two representative units across 4 delivered tastants (different colors), showing that fitted 389 models (right column) accurately recapitulate single -neuron activity (left column). C) The vast 390 majority of models built on real data showed higher cross-validated likelihood than models trained 391 on shuffled data (i.e., the bulk of the distribution is to the right of the vertical red dashed line at equal 392 predictability). This confirms that the models are learning single -trial relevant features. The white 393 arrow indicates the median of the distribution. D) The distributions of correlations (Spearman’s Rho) 394 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 23 between: 1) actual and predicted PSTHs (blue); 2) unit-shuffled data and prediction (orange); and 3) 395 temporally scrambled data and prediction (Circ shuffle: green). 396 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 24 Having established that we can successfully model the amygdala-cortical dataset using pGLM, we 397 moved on to using the model to identify any significant influence of the activity of one simultaneously 398 recorded neuron on that of another (i.e., identifying significant coupling filters). In what follows, we 399 refer to neurons with responses that are concluded to influence activity in the other region as 400 “Output” neurons, and neurons with responses that are concluded to be influenced by activity in the 401 other region as “Input” neurons. Fig. 5A shows the distribution of neurons so identified: clearly, our 402 BLA sample contains more Output neurons (40%) than Input neurons (6.3%), with an additional 9.4% 403 categorized as both; the Output and Input populations in GC were more balanced (17.2% and 22.5%, 404 respectively), with an additional 11.9% characterized as doing both. Note that no claim is made 405 about whether these neurons might output to or take input from any structures from which we didn’t 406 record, and therefore no further analysis is done on the (probably highly diverse set of) neurons that 407 were categorized as neither Input nor Output. 408 Given the relatively small overlap of Input and Output designations, it was important to ask whether 409 the two groups were structurally distinct—that is, whether a neuron (or neuron type) was specifically 410 likely to be either an Input or Output neuron, but not both. Two separate analyses were performed on 411 this front. First, we performed a direct, Bayesian assessment of the overlap between Output and 412 Input sub-groups in both regions , comparing the data to a null distribution in which labels for 413 connection types were assigned randomly so that we could ask whether, given the likelihood of a 414 neuron being categorized as Input or Output, the conditional probability (i.e., the likelihood of it being 415 categorized as Input given that it was categorized as Output, and vice-versa) was lower or higher than 416 chance. Fig. 5B shows that the overlap we observe in the real data falls squarely within the chance 417 probability of overlap for this null -distribution for both regions; we therefore conclude that there is 418 no evidence that Output and Input populations should be thought of as segregated groups. 419 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 25 We next assessed the likelihood that the separation of neurons into Output and Input groups aligns 420 with anatomically-defined neuron categories. Specifically, we asked whether putative pyramidal 421 cells, identified on the basis of spike width, were more likely to be Output neurons than Input 422 neurons. Fig. 5C demonstrates that our sample of isolated single neurons broke, as expected, neatly 423 into two groups on this basis, with a 0.45ms trough-to-peak time providing a clean cutoff between 424 fast-spiking putative interneurons (Int) and slower-spiking putative pyramidal neurons (Pyr ; see 425 (11,58–60). Slightly under 10% of our sample fell into the Int category , a number that is not wildly 426 inconsistent with expectation based on an extensive previous literature using chronic extracellular 427 electrophysiology. Given the small fraction of interneurons, it can be concluded that putative 428 pyramidal neurons are approximately equally likely of being Input and Output neurons. In conclusion, 429 we could find no evidence suggesting that Output and Input neurons are distinct anatomical types 430 (at least at the level of gross subdivision). 431 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 26 432 Fig. 5) pGLM brought to bear on amygdala -cortical ensembles identifies functionally (but not 433 structurally) distinct subpopulations of neurons. A) Inference of Input and Output population in 434 the recorded BLA-GC populations (number outside colored circles indicates percentage of neurons 435 with no identified between-region connections). B) The amount of overlap between Input and Output 436 populations is not significantly different from that expected by chance in either brain region . C) 437 Inferred neuron-types based on trough-to-peak times of waveforms (Inset: solid line = GC, crosses = 438 BLA). Most neurons in the dataset (expectedly) are putative pyramidal neurons , which means that 439 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 27 the vast majority (and similar percentage) of each functional neural subtype as shown in panel A is 440 pyramidal. 441 Taste processing in GC is a function of embeddedness in the amygdala-442 cortical loop 443 While Output and Input neurons do not appear to be separate neuron types, it is nonetheless 444 possible that their positioning within the network causes them to differ with regard to their 445 involvement in taste processing. Next, we tested this possibility; specifically, we tested whether we 446 can discern a relationship between pGLM label and the average magnitude of the neuron’s taste-447 specificity (more distinctive responsive ness to different tastes) and palatability -relatedness 448 (correlation with the hedonic value of the tastants)—the basic variables that describe involvement in 449 taste processing (11,13,16,24–26,33,34,38,61). 450 Initial analyses did in fact reveal significant differences in both taste-coding variables as a function 451 of subpopulation (Fig. 6A): GC Input neurons (those shown by pGLM to be influenced by BLA) showed 452 higher magnitudes of both taste specificity and palatability relatedness than did Output neurons. 453 Despite the status of GC as primary gustatory cortex, recipient of ascending taste input from 454 thalamus (2,15,30,62), involvement in taste processing (as measured by taste specificity and 455 palatability relatedness of taste responses) appears to depend on receiving “feedback” input from 456 BLA. 457 As taste discriminability and palatability are (necessarily) semi-redundant and correlated (Fig. 6B, 458 Spearman’s Rho=0.73, p<0.001), it was useful to aggregate and summarize the results shown in Fig. 459 6A into a single measure . We used for this summary t he vector norm (the magnitude of an arrow 460 drawn to a point in 2D -space, see Methods) of normalized discriminability and palatability coding, 461 an aggregate measure of taste -processing involvement that we here refer to as “tastiness .” An 462 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 28 analysis of neuron tastiness (Fig. 6C) confirms the findings of Fig. 6A, in that GC Input neurons show 463 significantly higher tastiness than GC Output neurons. Note that these results were not simply an 464 artifact of lower firing rates in the Output neurons, in that we observed no significant link between 465 firing rates and the metrics tested above (Spearman’s Rho: Firing Rate vs. Discriminability = 0.1, Firing 466 Rate vs. Palatability = 0.0). 467 The division of responses into Input and Output, which was made without reference to taste coding, 468 effectively reveals the importance of BLA feedback in taste responsiveness. But given that our central 469 hypothesis is that involvement in taste processing depends upon being “embedded” in the 470 amygdala-cortical loop, the most rigorous test involves an analysis of the specific set of neurons that 471 are most deeply so embedded—that is, the GC neurons that are categorized as both influencing and 472 receiving influence from BLA ( i.e., “Input/Output” neurons). If our hypothesis is correct , we should 473 expect that these neurons will exhibit the strongest tastiness—even stronger than that of GC Input 474 neurons. Indeed, this proved to be the case: the tastiness of GC Input/Output neurons is significantly 475 higher than that of GC Input neurons (Student’s t-test p=0.01), confirming our prediction that the 476 degree of “embedding” in the recursive circuit determines its degree of involvement in taste 477 processing (Fig. 6C), a result that is particularly striking given the near total lack of “tastiness” of GC 478 Output neuron responses . The processing of tastes appears to be largely performed by the 479 amygdala-cortical loop as a single dynamical object (see Discussion). 480 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 29 481 Fig. 6) GC neurons that are embedded in the amygdala -cortical loop are the ones that most 482 strongly process tastes. A) Both taste discriminability (i) and palatability-relatedness (ii) of GC Input 483 neurons (those influenced by BLA activity), measured in the standard manner (see Methods), are 484 significantly stronger than that in GC Output neurons. ** = p<0.005 . B) Discriminability and 485 Palatability are significantly correlated for GC neurons ( Spearman’s Rho=0.73, p<0.001) and hence 486 can be combined into a measure ( “Tastiness”) that represents a single measure of involvement in 487 taste processing. C) Comparison of tastiness among subpopulations confirms the results of A 488 (higher tastiness in Input neurons than Output neurons), but also shows that the small population of 489 GC Input/Output neurons produces responses with significantly higher tastiness than even the Input 490 population (2-sample Student’s T-test, ** = p<0.005, * = p<0.05) 491 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 30

492 This study reveals taste processing to centrally involve dynamical activity of the amygdala -cortical 493 loop. Following a n initial transient in which they fire independently of one another , BLA and GC 494 respond to taste input with zero-lag coupling of ensemble transitions; within this tight coupling, the 495 flow of influence between the two structures is asymmetric (as assessed using spectral granger 496 causality), even as those differences in influence change synchronously. Put simply, BLA and GC 497 become a single dynamical system , such that any individual GC neuron’s involvement in taste 498 processing—measured in terms of the taste -specificity and palatability -relatedness of its taste 499 responses—is a function of its embeddedness in the reciprocal circuit . Collectively, these results 500 support our argument that the BLA/GC dyad behaves, probably because of the direct recurrent 501 connectivity between the two regions, as a joint attractor network, processing tastes by collectively 502 transitioning through quasi-stable states. 503 These results highlight general principles that , we argue, are probably at work in other sensory 504 processing circuits. Specifically, it is reasonable to propose that, within any nominally sensory 505 network, activity evolves in a distributed yet coordinated fashion that is not present “by default” but 506 rather established in response to stimulus input, and that neurons involved in inter -region 507 communication reliably play outsized role s in processing. The following sections discuss this 508 proposal in further detail , specifically considering : 1) h ow stimulus-evoked metastable dynamics 509 seen in the amygdala -cortical circuit are well -described as a distributed attractor network ; 2) 510 whether this conceptualization can be generalized to describe work from other sensory systems; 3) 511 the fact that the “flow of influence” in tightly connected dynamical systems is likely to be 512 asymmetric, with distinct dynamics of influence acting in concert with one an other; and 4) 513 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 31 implications of the finding that neurons most deeply “embedded” in the task-relevant circuit show 514 the strongest task-related encoding. 515 Amygdala-cortical taste activity evolves as a joint attractor network 516 The metastable population dynamics that make up GC taste responses have already been 517 successfully modeled as an attractor network (28,40,63,64). Noise-driven attractor hopping (28,40) 518 has recapitulated many GC response characteristics—the exponential distribution of state 519 durations, activity in partially overlapping population pools, and distinct neural encoding in different 520 states—and has been shown to perform noisy decision-making better than the same network in an 521 “integration regime. ” Even a very simple instantiation of such a model, comprised of mutually 522 inhibitory neural populations, effectively captures the changing levels of LFP coherence observed 523 between population states (see Ref. #16, Fig. 10). 524 Our current results reveal that this attractor network is made up not just of GC, but of the dyad of GC 525 and BLA. As predicted by such a model, all but the initial transitions in taste-evoked activity happen, 526 on average, with lags that are distributed normally around zero, despite the fact that the regions are 527 connected by relatively long axons—a result that is difficult to explain using hierarchical/feedforward 528 models which, by construction, contain sources and sinks of influence (65–67). In a joint attractor 529 network, there is no “source” of transitions, which are triggered by coordination of the noise in the 530 system. And while there was some (inevitable) variability in the inference of the state transitions, this 531 did not preclude us from identifying the non-zero between-region lag that exists in the 1st transitions 532 expressed between GC and BLA (which is also the only transition uncorrelated between BLA and GC, 533 see 16). Thus, we argue that taste input causes the BLA -GC system to switch from resembling a 534 feedforward system to resembling a recurrent attractor network as processing unfolds. 535 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 32

from research on other sensory systems are consistent with these 536

537 Evidence supporting our proposal that nominally separate regions may “coalesce” into a single 538 perceptual processing unit abounds in research on other sensory systems. Work from Gaillard and 539 Dehaene (68,69), for instance, reveals properties of the processing of visual stimuli that are 540 qualitatively similar to the state transitions we see in taste processing. According to the model that 541 these authors use to explain their data, Global Workplace Theory (GWT), sensory information is first 542 transmitted in a feedforward manner to specific brain regions ; activation of these regions then 543 engages a distributed network via reciprocal projections , which essentially binds the regions 544 together. 545 In Gaillard and Dehaene’s GWT analysis, as in our analyses, different states reflect different stages 546 of perceptual processing and transitions to different stages are determined by different 547 mechanisms. Of particular relevance, their 2nd state is driven by recurrent interactions, is internally 548 driven (as opposed to being stimulus -bound), and reflects “conscious” stimulus processing. This 549 mirrors work show ing that the late taste-processing state “codes” the psychological variable of 550 palatability—the variable that is used to make consumption decisions (26,64,70–73). It is also 551 consistent with our demonstration that perturbation of the transition into this state most strongly 552 impacts decision making with regard to consumption (Mukherjee 2019). 553 In work that even more concretely parallels our own (23), researchers recorded intracranial EEG from 554 an ensemble of six cortical regions (early visual cortex, middle fusiform gyrus, pars triangularis, pars 555 opercularis, ventral sensorimotor cortex, and auditory cortex) . Using Hidden Markov Model ing on 556 data collected as human subjects viewed and verbally identifi ed pictures, the researchers 557 demonstrated that these regions transition collectively through a sequence of 3 states: 1) a stimulus-558 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 33 locked state lasting approximately 250ms, within which activity flows in a feedforward manner from 559 Early Visual Cortex to the other regions ; 2) a variable -length pre-output state characterized by 560 distributed connectivity (i.e., with no single source or sink region; see also 74); and 3) an output state, 561 centered on the period 1000-1500ms post-stimulus delivery, during which behavioral responses are 562 released. This 3-state dynamic mirrors several features of responses (see Fig. 1A) observed during 563 awake tasting in rats: 1) the timings of state transitions, 2) the fact that these transitions occur across 564 a distributed set of brain regions behaving as a collective whole, and 3) the evolution of processing 565 from feedforward to strongly recurrent. 566 Supp et al. (75) provide converging evidence from still another system, in the form of analysis of EEG 567 responses to small wrist shocks in human subjects under different levels of anesthesia. These 568 responses were, in conscious subjects, characterized by an initial period of roughly 200ms in which 569 activity is primarily seen in the somatosensory cortex, followed by a second period of roughly 200ms 570 characterized by distributed activation of multiple cortical regions (a result suggesting long-range 571 interactions, although the coupling between regions was not explicitly tested). Anesthesia, which 572 diminishes the behavioral response to this painful stimulus, eliminates distributed activation of 573 regions in the second state without impacting activity in the first (feedforward) state, suggesting that 574 in this system as well the 2nd state is related to conscious processing. 575 The function(s) of BLA-GC directional influences 576 While BLA and GC progress through taste processing in lock-step, as is reasonable to expect given 577 the reciprocal connections between them, the manner with which each region influences the other 578 is asymmetric. It is actually not unusual (quite the opposite) for recurrently interacting systems to 579 contain such asymmetric influences. Examples of dynamical systems in which feedforward and 580 feedback interactions have different strengths include half-center oscillators (76), Lotka -Volterra 581 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 34 (prey-predator) models (77), and reservoir networks (78). Furthermore, researchers have reported 582 asymmetric interactions analogous to our own in data collected from visual system recordings 583 (19,21,50). 584 By Granger Causality (GrCa) analysis, the dynamics of the amygdala-cortical interaction accord well 585 with classic stimulus-valuation studies, including those performed by Damasio (Bechara 1999) and 586 Schoenbaum (Schoenbaum 1999), which have been interpreted as suggesting that amygdala, after 587 playing a primary role in determining the value of a stimulus, provides this information to cortex 588 which then uses the information to generate a behavioral decision. This interpretation is also 589 consistent with : 1) the fact that GC neurons primarily encode taste quality (discriminability, 590 (24,26,33) during the ~200 -800msec middle epoch, at a time that BLA neurons are already 591 responding in a palatability-related manner (13,25); 2) the appearance of a peak in the GC→BLA 592 influence aligned with the average time of the transition into palatability-related firing in GC (just after 593 1000ms post-stimulus delivery); and 3) the role of this GC transition in driving behavioral responses 594 to the taste (26,34). Our GrCa analysis provides the first representation of this “signal” being emitted 595 by GC. In this context, however, it is worth noting that Forseth et al. (23) observed a similar transient 596 “signal” emanating from somatomotor regions approximately 250ms prior to the onset of response 597 articulation in the above-discussed visual-naming task. 598 Again, this analysis is consistent with our central suggestion that BLA and GC act not as separate 599 structures but rather as a single processing unit. Taste processing is not actually feedforward, but 600 rather emerges out of reciprocal interaction (see 9,69). 601 The role of projection neurons in sensory processing 602 The results of our Poisson General Linear Modeling (pGLM) make it clear that the GC neurons most 603 deeply involved in taste processing —quantified in terms of the magnitude of taste -specificity and 604 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 35 palatability-relatedness of their responses, basic response parameters that are summarized in our 605 “tastiness” metric —are those that both receiv e influence from and send influence to BLA . 606 Furthermore, the responses of GC neurons that are “simply” influenced by BLA (i.e., that don’t also 607 influence BLA) are still more deeply taste- and palatability-specific than those that don’t. 608 These results suggest that a breakdown of neuron types based on embeddedness in the reciprocal 609 circuit may be more meaningful for understanding function than more classically considered 610 subdivisions. Previous work has shown that while afferents from thalamus and BLA broadly overlap 611 in GC, the “hotspots” of these projections are visibly distinct (79). While our data do not directly test 612 this hypothesis, it is likely that some thalamocortical afferents, which are those carrying ascending 613 taste information, synapse on neurons that are not directly influenced by BLA; this consideration 614 carries with it the somewhat surprising realization that this ascending influence does not uniquely 615 determine the magnitude of taste responsiveness , but ( in at least in a subset of neuron s) taste 616 responsiveness is generated by a rec iprocal interaction with amygdala (see (14) for a similar result 617 using cue-triggered responses). Furthermore, the pattern that we have observed is not well explained 618 by the unit’s putative cell-type, suggesting that the “embedding” of the neuron in the circuit is more 619 significantly linked to function than morphology and molecular identity. 620 It is tempting to speculate that GC itself powers the processing attributed to BLA, perhaps by sending 621 taste quality information to BLA while BLA returns palatability information, with the two regions co -622 evolving in this manner to converge to a decision (80). While GC neurons that influence BLA activity 623 without themselves being influenced by BLA activity (i.e., GC Output neurons) showed little or no 624 taste coding, t he above conclusion can nonetheless be easily reconciled with previous work (8) 625 showing that a subset of GC→BLA projection neurons have more discriminatory taste-responses 626 compared to the average of non-projecting neurons, in response to aversive tastants vs palatable 627 tastants. Th ese researchers showed that a subset of projection neurons more distinctly encode 628 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 36 valence than the average of non-projecting neurons. We predict that this subset of strongly 629 discriminatory GC→BLA projection neurons observed by Lavi et al. (8) were those deeply embedded 630 in the reciprocal circuit, that is, Input/Output neurons (although it is also possible that the fact that 631 Lavi et al. were examining responses in the context of CTA learning, which involves both GC and BLA, 632 may have changed functional connectivity). 633 This interpretation of BLA-GC “co-determination” is further supported by combining our GrCa results 634 with previous work (15) showing that pharmacological inhibition of taste thalamus suppressed 20 -635 60Hz LFP power in GC – exactly the band that dominates GC→BLA influence. Hence, and as noted 636 above, GC→BLA influence in this range may indicate simultaneous participation of GC in a parallel 637 circuit, likely the thalamocortical circuit, implicating the larger taste circuit in BLA-GC dynamics. In 638 future investigations we will directly test whether activity in GC →BLA axons, perhaps even activity 639 specifically confined to the γ-frequency range in which the GC →BLA influence is strongest, is 640 necessary for the evolution of amygdala-cortical taste processing (81). 641 Finally, there is evidence from studies examining other circuits (82–84) suggesting enhanced coding 642 in projection populations compared to non -projecting neurons. This may be rationalized from the 643 perspective that task -specific computations invoke different circuit configurations, therefore 644 different projections populations will be involved in specific tasks and neurons involved in the task-645 relevant circuit will have the strongest encoding, as we observe in our results. 646

647 Our study reveals that cortico-amygdalar interactions subserve taste processing (and thereby taste 648 decision-making), showing that mechanisms previously theorized to underly state transitions in 649 cortical stimulus-evoked activity (attractor networks) are in fact part and parcel of the dynamics of 650 the GC-BLA dyad. By extension, we speculate that these dynamics are potentially a function of the 651 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 37 lager taste circuit including the brainstem (85). This taste network is initialized in a feedforward state 652 but coalesces into a reciprocally interacting network that evolves toward a point shown previously to 653 generate the behavioral output. Finally, our investigation of the makeup of GC Input and Output 654 neurons suggests that network interactions are stronger determinants of neural encoding than are 655 specific cell -types. All results presented herein align with previous studies from other sensory 656 systems as well as investigations of GC -BLA interactions during learning , allowing us to 657 conceptualize these previous results in a unified theoretical framework. 658 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 38

659 Experimental Design and Statistical Analyses 660 Subjects 661 Adult, female Long -Evans rats (n = 6; 300 –350g at time of electrode implantation, Charles River 662 Laboratories) served as subjects in our study (we have observed no sex differences in the basic 663 cortical dynamics of taste responses between male and female rats, and therefore use female Long-664 Evans rats because they are, in our hands, calmer than males). The rats were housed in individual 665 cages in a temperature- and humidity-controlled environment under a 12:12 hr light:dark cycle, given 666 ad libitum access to food and water prior to the start of experimentation, and weighed daily following 667 surgery to ensure that they never dropped below 80% of their pre -surgery weight. All experimental 668

were in compliance with National Institutes of Health guidelines and were approved in 669 advance by the Brandeis University Institutional Animal Care and Use Committee. 670 Electrode and intra-oral cannula construction 671 Custom microwire bundle drives , optimized for chronic (i.e., multi -day and multi -week) recording 672 quality, were made with either 16 or 32 electrodes per recording site (design and construction details 673 available at https://katzlab.squarespace.com/technology). Intra-oral cannulae—flexible tubing with 674 a flanged tip and washer to ensure stability, connected to a plastic top complete with a locking 675 mechanism—were built to allow the delivery of tastants directly onto the tongue (86). 676 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 39 Surgery 677 Rats were anaesthetized with an intraperitoneal injection of ketamine/xylazine cocktail (100mg/kg 678 and 5.2 mg/kg respectively) and mounted in a stereotaxic instrument (David Kopf Instruments; 679 Tujunga, CA) with blunt (atraumatic) ear bars. A midline incision exposed the skull and trephine holes 680 (~2 mm diameter) were drilled above BLA and GC. Microwire bundles were implanted 0.5mm above 681 GC (coordinates: AP +1.4 mm, ML -5.0 mm, DV −4.4mm from dura) and BLA (coordinates : AP -682 3.0mm, ML -5.0mm, DV -6.8mm from dura). Once in place, electrode bundles were cemented to the 683 skull. Once electrode bundles were secured, an intra -oral cannula (IOC) was threaded through the 684 masseter muscle (inside the zygomatic arch) to the space between the lip and gums, and the top of 685 the cannula was cemented to the rest of the assembly with dental acrylic (Fontanini and Katz, 2006). 686 The rat’s body temperature was monitored and maintained at ~37°C by a heating pad throughout the 687 duration of the surgery. 688 Acquisition of electrophysiological data 689 Electrophysiological signals from the micro -electrodes were sampled at 30 kHz using 32 -channel 690 analog-to-digital converter chips (RHD2132) from Intan Technologies, digitized online at the head 691 stage and sampled jointly, along with signals from actuators marking tastant delivery, using an Intan 692 RHD USB interface board (Part #C3100), which routed records to the hard drive of a PC for saving. 693 The experimental chamber was ensconced in a Faraday cage that shielded recordings from external 694 electromagnetic influences. 695 Habituation and passive taste administration 696 Following their recovery from surgery, we habituated rats to the experimental chamber for 2 days, to 697 the IOC/electrode harness for the next 2 days, and to passive water deliveries for the following 2 days, 698 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 40 before beginning data collection. Starting with the second habituation day, we also placed rats on a 699 mild water restriction schedule—20mL of water (not including the ~4mL delivered during habituation 700 sessions themselves) per day. This water restriction schedule was maintained till the end of the 701 experiment. For the 2 final habituation sessions, we attached the rats to the taste delivery apparatus, 702 and infused 120 pulses of distilled water (∼30μL per pulse; 20s inter-pulse interval) into the animal’s 703 oral cavity through the IOC, and drove electrode bundles deeper (by 250 μm) into target structures. 704 By the end of this procedure, the tips of the electrodes lay within GC and BLA. We then recorded taste 705 responses during 3-4 days of taste delivery sessions, between each of which the microwire bundle 706 was driven down approximately 60 μm. During these sessions, Sucrose (0.3M), Sodium Chloride 707 (0.1M), Citric Acid (0.1M), and Quinine (1mM), dissolved in ultra -pure water (∼30μL per pulse; 20s 708 inter-pulse interval, 30 trials/tastant) were delivered to passive rats (i.e., no behavior was required to 709 elicit delivery). These concentrations were chosen to represent a range of hedonic values, and 710 because they are known to evoke robust responses in both GC and BLA (25,33). 711 Histology 712 In preparation for histology, rats were deeply anesthetized with an overdose of the ketamine/xylazine 713 mixture. We perfused the rats through the heart with 0.9% saline followed by 10% formalin and 714 harvested the brain. The brain tissue was incubated in a fixing mixture of 30% sucrose and 10% 715 formalin for 7 days before being sectioned into 50 μm coronal slices on a sliding microtome (Leica 716 SM2010R, Leica Microsystems). Sections containing the electrode implant sites around GC and BLA 717 were imaged at 2x. 718 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 41 Data and statistical analyses 719 The analysis of data and statistical tests were performed using custom written software in Python as 720 described below. 721 Local Field Potential processing and analysis 722 Filtering / power and phase extractions 723 LFPs were extracted from broadband digitized signals using a 2nd order bandpass Butterworth filter 724 (1-300Hz), to de-emphasize spiking and emphasize frequencies typically of interest in such data. In 725 order to avoid contamination from noise/artifacts on noisy/broken channels, only channels 726 containing isolable single neurons (see below) were used for analyses. Signal was down sampled to 727 1000Hz to reduce computational load for downstream analyses. 728 We note that this filtering process is sufficient for analysis up to the maximum frequency shown in 729 the Results section (that is, 100Hz), and is unlikely to be the reason we don’t observe effects above 730 60Hz. First, a maximum frequency of 300Hz is sufficient for a Nyquist frequency for accurately 731 detecting signals up to 150Hz. Secondly, we use a Butterworth filter, which is “maximally flat” (see 732 scipy.signal.butter, 87) within the pass band and is even permissive of frequencies higher than 733 300Hz. Hence, signals with frequencies 100Hz or lower are unlikely to be lost due to filtering. 734 Spectral Granger Causality 735 Spectral Granger Causality was assessed between pairs of electrodes in GC and BLA using the 736 Spectral Connectivity package (47). We selected for this analysis the electrodes from which activity 737 was most similar to the mean activity for the region ( smallest mean-squared error relative to the 738 mean phase across all channels for each region; this selection of channels was constant for all trials 739 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 42 in a single analyzed session, and the same set of channels was used for all frequencies), to ensure 740 reliable representation of each. 741 To satisfy the normality assumption for use of Granger Causality, signals from both electrodes were 742 first preprocessed according to previously established criteria to remove outliers and “detrend” the 743 data (88). Briefly: 744 1. Any trials with data points exceeding 3 Median Absolute Deviations calculated using all the 745 trials, were removed 746 2. Linear detrending was performed on a single-trial basis (scipy.signal.detrend) 747 3. The temporal (single -trial) mean was subtracted from the data, and all data divided by the 748 temporal standard deviation 749 4. Trial-averaged mean was subtracted, and all data divided by the trial -averaged standard 750 deviation 751 Following preprocessing, the Augmented Dickey -Fuller test was run on each channel to confirm 752 stationarity (normality) of the signal. All preprocessed data passed the test and hence was included 753 in further analyses. 754 Using the preprocessed data, spectral granger causality was calculated using multitaper methods 755 implemented in the Spectral Connectivity package. The default set of tapers (Slepian) was used. 756 Specific parameters included: 757 • “sampling_frequency” = 1000 (Hz), 758 • “time_halfbandwitch_product” = 1 759 • “time_window_duration” = 0.3 (sec) 760 • “time_window_step” = 0.05 (sec) 761 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 43 Significance of the granger causality for each frequency -time bin on a single -session basis was 762 obtained using a shuffling procedure. Granger Causality was recalculated for 500 sets of data where 763 trial labels were shuffled (mismatched) between BLA and GC to generate a null distribution of “non-764 specific” Granger Causality between the two regions. Deviation of the value in each frequency-time 765 bin from its respective null distribution was taken to indicate a significant interaction in that bin . 766 Multiple comparisons were addressed using the Bonferroni correction with a base alpha of 0.05 with 767 each frequency-time bin treated as a comparison. We further supplemented this “bin-level” testing 768 as below. 769 Cluster Permutation Testing 770 To perform statistical testing on the Granger Causality results while minimizing issues of multiple 771 comparisons, we opted to use a cluster-based permutation test (89,90) to detect the largest 772 statistically-significant deviation in the time-series for the 0-12Hz and 20-60Hz frequency bands. 773 Briefly, the procedure involves: 774 1. Z-scoring the time-series on which testing needs to be done (in our case, this is the mean 775 significance of Granger Causality for a frequency band) 776 2. Assessing the summed value of the largest cluster (continuous chunk) above a specific 777 threshold for the absolute of the z-scored time-series (we used a threshold of Z = 2). This 778 value is used as the test statistic. 779 3. This summed value of the largest cluster is then recalculated from shuffled data to obtain a 780 null-distribution for the test statistic. 781 4. If the actual value of the statistic significantly deviates from the null-distribution, the null 782 hypothesis is rejected (alpha=0.05). 783 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 44 Granger Causality Similarity Analysis 784 To analyze similarities in the dynamics of the directional Granger Causality timeseries, we performed 785 Principal Component Analysis on the Granger Causality significance timeseries for 0 -100Hz, 786 reducing each multivariate timeseries down to 3 components to mitigate overfitting in the following 787 regression analyses. All comparisons were then performed on these dimensionally reduced 788 timeseries. We next performed 1) Linear Regression between the two timeseries, 2) Linear regression 789 between circularly (temporally) shuffled versions of the two timeseries (as a negative control / 790 context for poor prediction), 3) Linear Autoregression, and 4) non -Linear Autoregression 791 (Multilayered perceptron regressor, (91). For both forms of autoregression, we used the previous 3 792 timepoints to predict the current one. To obtain an estimate of the uncertainty of prediction, we 793 performed bootstrapping of the data with 500 reruns of model fitting and prediction. 794 Analysis of Spiking Activity 795 Single Unit Isolation 796 Spikes from electrophysiological recordings were sorted and analyzed off -line using an in-house 797 Python-based pipeline (92). Putative single -neuron waveforms with > 5:1 signal -to-noise ratio 798 (median absolute deviation) were sorted using a semi-supervised algorithm: recorded voltage data 799 were filtered between 300 -3000Hz, grouped into potential clusters by a Gaussian Mixture Models 800 (GMM) fit to multiple waveform features; clusters were then labeled and/or refined manually (to 801 increase conservatism) by the experimenters. 802 Single Unit Evoked Response Characterization 803 Evaluating single-neuron response taste specificity. To statistically determine the degree to which a 804 single neuron’s response contained taste -specific information, firing rates were estimated by 805 convolving spike-trains with a rectangular window (length: 2 00ms). A one -way ANOVA (between 806 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 45 tastes, across trials) was then run on each window to identify if the response of a single neuron to 807 one taste was different from its responses to any other tastes at that timepoint. The (taste) specificity 808 of a single neuron was defined as the average of the ANOVA effect sizes for 2000ms of post-stimulus 809 activity. 810 Evaluating single-neuron response palatability-relatedness. To statistically determine the degree to 811 which a single neuron’s response reflected the hedonic value of the stimuli delivered, we estimated 812 firing rates as described above, and calculated the Pearson’s correlation coefficient between the 813 evoked firing rates and the palatability ranks of the tastants at each timepoint. Palatability ranks — 814 sucrose (1) > NaCl (2) > citric acid (3) > quinine (4) — directly reflected consumption of (earlier-run 815 squads of) rats in a Brief Access Task (see 33). This ordering is canonical, and has been replicated in 816 many studies and with multiple measures of stimulus appreciation (86,93,94). The palatability of a 817 single neuron was defined as the average of the absolute correlation value (as the correlation can 818 also be strongly negative) for 2000ms of post-stimulus activity. 819 Calculating “Tastiness” 820 To combine taste specificity and palatability into a single metric, we first log-transformed specificity 821 to make the scales of specific and palatability more similar. This is because palatability as an 822 absolute value of a correlation is bounded between [0,1] whereas specificity as an effect size has the 823 range [0, ∞). Log-Specificity and Palatability were then Min-Max normalized following which the L2-824 norm of the vector generated by the normalized metrics (normalized log-specificity, normalized 825 palatability) was calculated on a single-neuron basis and defined as “tastiness”. 826 Determination of putative pyramidal vs inter-neurons 827 From the average of filtered spikes for each putative single-unit, we determined the time taken from 828 the trough to the next peak. As established previously, inter-neurons and pyramidal neurons can be 829 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 46 reliably separated using this “rise time” (11,58–60). The specific threshold varies depending on the 830 filtering frequencies used. We used a value of 0.45ms based on previous work using the same 831 bandpass filter frequencies as we have used here (11,60) 832 Changepoint Modeling of Population Activity 833 Model Fitting 834 Changepoint modelling of neural activity was performed as described previously (95) using the pytau 835 library (https://github.com/abuzarmahmood/pytau) written in Python . Briefly, spike -trains were 836 binned into 50ms bins. Changepoint models with independent Poisson emissions were fit to these 837 timeseries of binned spike -counts for BLA and GC populations separately using automatic 838 differentiation variational inference (ADVI) functionalities implemented in the pymc3 probabilistic 839 programming library (96). Once fit, the procedure returns: 1) the estimate of the posterior distribution 840 for single-trial population changepoint locations, and 2) the emission (firing) rates for each neuron 841 during every state. Since ADVI utilizes spherical Gaussians as the surrogate distribution, we use the 842 mode of the distribution for individual parameters as the “best estimate” of the parameter for 843 downstream analyses. Models with 4 states were fit to 2000ms of post-stimulus activity. Please refer 844 to (95) for more details on the process of model/state-number selection. 845 Calculation of Transition Lags 846 Lags between BLA and GC transitions were calculated on a single-trial basis by subtracting the “best 847 estimate” (see above) of the matched transition (that is, transition 1 in GC was compared to transition 848 1 in BLA for any given trial) between models fit to each region. This resulted in a distribution of lags 849 for each recording session (120 values, 1 for each trial in a session) . The means of each session’s 850 distribution were used to generate a distribution of average lags for each transition across all the 851 recording sessions. Finally, a grand average of each transition’s distribution of average lag s was 852 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 47 calculated. The significance of the deviation of this grand -average lag from zero was calculated by 853 bootstrapping at the level of trials for each session with a total of 5000 permutations. The 95% 854 confidence interval of each grand average was calculated using these bootstrapped distributions, 855 and deviation of zero from this 95% interval was taken as a significant difference. 856 Generalized Linear Modelling 857 As has been shown previously, Granger Causality methods perform poorly on spike-train data due to 858 the strong violation of the stationarity assumption used by the autoregressive processes on which 859 Granger Causality is based (97,98). This issue is addressed by Generalized Linear Models (GLM ; 860 distinct from General Linear Models which also follow the stationarity/normality assumption) which 861 can model Poisson-distributed data, and furthermore can model temporal dependencies explicitly 862 over a large range of timescales (54,66,99,100). 863 Model Specification and Model Fitting 864 The GLM included 3 sets of filters for each neuron modelled: 865 • Spike History Filter: Captures how a neuron's past spiking activity influences its current firing 866 rate. This can model refractoriness (reduced probability of firing immediately after a spike) 867 and bursting behavior. 868 • Stimulus Filter: Models how external stimuli drive neural activity. 869 • Coupling Filter: Captures how the activity of other neurons influences the target neuron, 870 modeling functional connectivity. 871 The duration of the spike history and coupling filters was set to 100ms, while the duration of the 872 stimulus filter was set to 300ms, based on previous modelling work (54). To reduce number of fit 873 parameters and mitigate over -fitting of the models, raised cosine -basis functions were used to 874 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 48 parameterize the filters (54,55,99): a basis of 10 functions was used for every filter, with peaks 875 distributed logarithmically to model fast and slow timescales with appropriate temporal resolutions. 876 The conditional intensity of the neural activity was hence defined as: 877 λ(t) = exp(β₀ + β_hist * h(t) + β_stim * s(t) + β_coup * c(t)) 878 • λ(t) is the firing rate at time t 879 • β₀ is the baseline firing rate 880 • h(t) represents the neuron's own spike history 881 • s(t) represents the stimulus 882 • c(t) represents the coupling with other neurons 883 • β terms are the corresponding weights 884 The Python “statsmodels” library was used for model fitting (101). 885 Quality of fit was assessed in two ways. First, by comparing the log-likelihood of models fit to actual 886 data with models fit to data where the trial labels were shuffled for neurons in the ensemble being 887 fit, which would only allow the model to learn an “average” coupling filter between the neurons in the 888 model being fit. Second, we compared Spearman’s Rho between predicted peri -stimulus time 889 histograms (PSTH) of a given neuron with PSTHs predicted for 1) the neuron the model was fit to 890 (actual data), 2) a different neuron (Unit Shuffled), and 3) temporally scrambled data for the same 891 neuron (Circular Shuffle). 892 Once models were fit, putative inter-neuron connections were determined using the coupling filters 893 between neurons. The fitting process returns p -values for each coefficient in the coupling filter 894 (which, in our case due to the parametrization of the filter using basis functions corresponded to the 895 scaling factor of each function in the basis). We asserted that a coupling filter indicated a statistically 896 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 49 significant interaction if a p-value in the filter was below alpha=0.05 following a Bonferroni correction 897 (corrected p-value = 0.05 / 10 = 0.005). 898 Assessment of overlap between input and output populations 899 We assess whether the overlap observed between subpopulations labelled as “Input” and “Output” 900 in GC and BLA was different from that expected by chance using Monte Carlo simulations. 901 Essentially, keeping the number of Input, Output, and Total number of neurons in each region 902 constant, we label neurons as Input or Output randomly and assess the number of neurons with both 903 labels. We perform 10,000 simulations for each region to generate a null distribution for this overlap 904 as expected by random assignment. Deviation of the observed value of overlap from the 2.5 -97.5 905 percentile range of this distribution would indicate that the observed value is significantly different 906 from that expected by chance with an alpha=0.05. 907 Data Availability 908 We have structured our electrophysiology datasets in a hierarchical data format (HDF5) and are 909 hosting the files on a university -wide network share managed by Library and Technology Services 910 (LTS) at Brandeis University. These HDF5 files contain our electrophysiology recordings, sorted 911 spikes, single-neuron and population-level analyses (and associated plots and results). These files 912 are prohibitively large to be hosted on a general -purpose fileshare platform - we request anyone 913 interested in our datasets to contact the corresponding author, Donald Katz ([email protected]) 914 who can put them in touch with LTS in order to create a guest account at Brandeis through which they 915 can securely access our datasets (hosted on the katz-lab share at files.brandeis.edu). 916 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 50

917 1. Yarmolinsky DA, Zuker CS, Ryba NJP . Common Sense about Taste: From Mammals to Insects. 918 Cell. 2009 Oct 16;139(2):234–44. 919 2. Carleton A, Accolla R, Simon SA. Coding in the mammalian gustatory system. Trends in 920 Neurosciences. 2010 Jul 1;33(7):326–34. 921 3. Rolls ET. Taste and smell processing in the brain. In: Handbook of Clinical Neurology [Internet]. 922 Elsevier; 2019 [cited 2025 Jan 19]. p. 97–118. Available from: 923 https://linkinghub.elsevier.com/retrieve/pii/B9780444638557000071 924 4. Kietzmann TC, Spoerer CJ, Sörensen LKA, Cichy RM, Hauk O, Kriegeskorte N. Recurrence is 925 required to capture the representational dynamics of the human visual system. Proceedings of 926 the National Academy of Sciences. 2019 Oct 22;116(43):21854–63. 927 5. van Bergen RS, Kriegeskorte N. Going in circles is the way forward: the role of recurrence in 928 visual inference. Current Opinion in Neurobiology. 2020 Dec 1;65:176–93. 929 6. Kriegeskorte N. Deep Neural Networks: A New Framework for Modeling Biological Vision and 930 Brain Information Processing. Annual Review of Vision Science. 2015;1(1):417–46. 931 7. Piette CE, Baez-Santiago MA, Reid EE, Katz DB, Moran A. Inactivation of Basolateral Amygdala 932 Specifically Eliminates Palatability-Related Information in Cortical Sensory Responses. J 933 Neurosci. 2012 Jul 18;32(29):9981–91. 934 8. Lavi K, Jacobson GA, Rosenblum K, Lüthi A. Encoding of Conditioned Taste Aversion in Cortico-935 Amygdala Circuits. Cell Reports. 2018 Jul 10;24(2):278–83. 936 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 51 9. Kayyal H, Yiannakas A, Chandran SK, Khamaisy M, Sharma V , Rosenblum K. Activity of Insula to 937 Basolateral Amygdala Projecting Neurons is Necessary and Sufficient for Taste Valence 938 Representation. J Neurosci. 2019 Nov 20;39(47):9369–82. 939 10. Arieli E, Gerbi R, Shein-Idelson M, Moran A. Temporally-precise basolateral amygdala activation 940 is required for the formation of taste memories in gustatory cortex. The Journal of Physiology. 941 2020;598(23):5505–22. 942 11. Lin JY , Mukherjee N, Bernstein MJ, Katz DB. Perturbation of amygdala-cortical projections 943 reduces ensemble coherence of palatability coding in gustatory cortex. Schoenbaum G, 944 Wassum KM, editors. eLife. 2021 May 21;10:e65766. 945 12. Yiannakas A, Kolatt Chandran S, Kayyal H, Gould N, Khamaisy M, Rosenblum K. Parvalbumin 946 interneuron inhibition onto anterior insula neurons projecting to the basolateral amygdala 947 drives aversive taste memory retrieval. Current Biology. 2021 Jul 12;31(13):2770-2784.e6. 948 13. Grossman SE, Fontanini A, Wieskopf JS, Katz DB. Learning-Related Plasticity of Temporal 949 Coding in Simultaneously Recorded Amygdala–Cortical Ensembles. J Neurosci. 2008 Mar 950 12;28(11):2864–73. 951 14. Samuelsen CL, Gardner MPH, Fontanini A. Effects of Cue-Triggered Expectation on Cortical 952 Processing of Taste. Neuron. 2012 Apr 26;74(2):410–22. 953 15. Samuelsen CL, Gardner MPH, Fontanini A. Thalamic Contribution to Cortical Processing of 954 Taste and Expectation. J Neurosci. 2013 Jan 30;33(5):1815–27. 955 16. Mahmood A, Steindler J, Germaine H, Miller P , Katz DB. Coupled Dynamics of Stimulus-Evoked 956 Gustatory Cortical and Basolateral Amygdalar Activity. J Neurosci. 2023 Jan 18;43(3):386–404. 957 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 52 17. Bressler SL, Coppola R, Nakamura R. Episodic multiregional cortical coherence at multiple 958 frequencies during visual task performance. Nature. 1993 Nov;366(6451):153. 959 18. van Kerkoerle T, Self MW, Dagnino B, Gariel-Mathis MA, Poort J, van der Togt C, et al. Alpha and 960 gamma oscillations characterize feedback and feedforward processing in monkey visual 961 cortex. Proc Natl Acad Sci U S A. 2014 Oct 7;111(40):14332–41. 962 19. Bastos AM, Vezoli J, Bosman CA, Schoffelen JM, Oostenveld R, Dowdall JR, et al. Visual Areas 963 Exert Feedforward and Feedback Influences through Distinct Frequency Channels. Neuron. 964 2015 Jan 21;85(2):390–401. 965 20. Arce-McShane FI, Ross CF , Takahashi K, Sessle BJ, Hatsopoulos NG. Primary motor and 966 sensory cortical areas communicate via spatiotemporally coordinated networks at multiple 967 frequencies. Proceedings of the National Academy of Sciences. 2016 May 3;113(18):5083–8. 968 21. Mejias JF , Murray JD, Kennedy H, Wang XJ. Feedforward and feedback frequency-dependent 969 interactions in a large-scale laminar network of the primate cortex. Science Advances. 2016 970 Nov 1;2(11):e1601335. 971 22. Michaels JA, Dann B, Scherberger H. Neural Population Dynamics during Reaching Are Better 972 Explained by a Dynamical System than Representational Tuning. PLOS Computational Biology. 973 2016 Nov 4;12(11):e1005175. 974 23. Saravani AG, Forseth KJ, Tandon N, Pitkow X. Dynamic Brain Interactions during Picture 975 Naming. eNeuro [Internet]. 2019 Jul 1 [cited 2020 Feb 16];6(4). Available from: 976 https://www.eneuro.org/content/6/4/ENEURO.0472-18.2019 977 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 53 24. Katz DB, Simon SA, Nicolelis MAL. Dynamic and Multimodal Responses of Gustatory Cortical 978 Neurons in Awake Rats. J Neurosci. 2001 Jun 15;21(12):4478–89. 979 25. Fontanini A, Grossman SE, Figueroa JA, Katz DB. Distinct Subtypes of Basolateral Amygdala 980 Taste Neurons Reflect Palatability and Reward. J Neurosci. 2009 Feb 25;29(8):2486–95. 981 26. Sadacca BF , Mukherjee N, Vladusich T, Li JX, Katz DB, Miller P . The Behavioral Relevance of 982 Cortical Neural Ensemble Responses Emerges Suddenly. J Neurosci. 2016 Jan 20;36(3):655–69. 983 27. Hopfield JJ. Neural networks and physical systems with emergent collective computational 984 abilities. PNAS. 1982 Apr 1;79(8):2554–8. 985 28. Miller P , Katz DB. Stochastic Transitions between Neural States in Taste Processing and 986 Decision-Making. J Neurosci. 2010 Feb 17;30(7):2559–70. 987 29. McDonald AJ. Cortical pathways to the mammalian amygdala. Progress in Neurobiology. 1998 988 Jun 27;55(3):257–332. 989 30. Shi CJ, Cassell MD. Cortical, thalamic, and amygdaloid connections of the anterior and 990 posterior insular cortices. The Journal of Comparative Neurology. 1998 Oct 5;399(4):440–68. 991 31. Stone ME, Maffei A, Fontanini A. Amygdala Stimulation Evokes Time-Varying Synaptic 992 Responses in the Gustatory Cortex of Anesthetized Rats. Front Integr Neurosci [Internet]. 2011 993 [cited 2019 May 8];5. Available from: 994 https://www.frontiersin.org/articles/10.3389/fnint.2011.00003/full 995 32. Haley MS, Fontanini A, Maffei A. Laminar- and Target-Specific Amygdalar Inputs in Rat Primary 996 Gustatory Cortex. J Neurosci. 2016 Mar 2;36(9):2623–37. 997 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 54 33. Sadacca BF , Rothwax JT, Katz DB. Sodium Concentration Coding Gives Way to Evaluative 998 Coding in Cortex and Amygdala. J Neurosci. 2012 Jul 18;32(29):9999–10011. 999 34. Mukherjee N, Wachutka J, Katz DB. Impact of precisely-timed inhibition of gustatory cortex on 1000 taste behavior depends on single-trial ensemble dynamics. Colgin LL, Maffei A, Maffei A, 1001 Ghazanfar AA, Verhagen J, Mazzucato L, editors. eLife. 2019 Jun 24;8:e45968. 1002 35. Fontanini A, Katz DB. 7 to 12 Hz Activity in Rat Gustatory Cortex Reflects Disengagement From a 1003 Fluid Self-Administration Task. Journal of Neurophysiology. 2005 May 1;93(5):2832–40. 1004 36. Tort ABL, Fontanini A, Kramer MA, Jones-Lush LM, Kopell NJ, Katz DB. Cortical Networks 1005 Produce Three Distinct 7–12 Hz Rhythms during Single Sensory Responses in the Awake Rat. J 1006 Neurosci. 2010 Mar 24;30(12):4315–24. 1007 37. Pavão R, Piette CE, Lopes-dos-Santos V , Katz DB, Tort ABL. Local Field Potentials in the 1008 Gustatory Cortex Carry Taste Information. J Neurosci. 2014 Jun 25;34(26):8778–87. 1009 38. Moran A, Katz DB. Sensory Cortical Population Dynamics Uniquely Track Behavior across 1010 Learning and Extinction. J Neurosci. 2014 Jan 22;34(4):1248–57. 1011 39. Bell T. Attractor transitions: A basis for sequential processing in neural networks? Neural 1012 Networks. 1988;1:163. 1013 40. Miller P , Katz DB. Accuracy and response-time distributions for decision-making: linear perfect 1014 integrators versus nonlinear attractor-based neural circuits. J Comput Neurosci. 2013 Dec 1015 1;35(3):261–94. 1016 41. Brennan C, Proekt A. Attractor dynamics with activity-dependent plasticity capture human 1017 working memory across time scales. Commun Psychol. 2023 Oct 25;1(1):1–15. 1018 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 55 42. Jones LM, Fontanini A, Sadacca BF , Miller P , Katz DB. Natural stimuli evoke dynamic sequences 1019 of states in sensory cortical ensembles. PNAS. 2007 Nov 20;104(47):18772–7. 1020 43. Ding M, Chen Y , Bressler SL. Granger Causality: Basic Theory and Application to Neuroscience 1021 [Internet]. arXiv; 2006 [cited 2023 Apr 26]. Available from: http://arxiv.org/abs/q-bio/0608035 1022 44. Seth AK, Barrett AB, Barnett L. Granger causality analysis in neuroscience and neuroimaging. J 1023 Neurosci. 2015 Feb 25;35(8):3293–7. 1024 45. Geweke J. Measurement of Linear Dependence and Feedback Between Multiple Time Series. 1025 Journal of the American Statistical Association. 1982;77(378):304–13. 1026 46. Chicharro D. On the spectral formulation of Granger causality. Biol Cybern. 2011 Dec;105(5–1027 6):331–47. 1028 47. Denovellis EL, Myroshnychenko M, Sarmashghi M, Stephen EP . Spectral Connectivity: a python 1029 package for computing multitaper spectral estimates and frequency-domain brain connectivity 1030 measures on the CPU and GPU. Journal of Open Source Software. 2022 Dec 13;7(80):4840. 1031 48. Buzsáki G, Draguhn A. Neuronal Oscillations in Cortical Networks. Science. 2004 Jun 1032 25;304(5679):1926–9. 1033 49. Baccalá LA, Sameshima K. Partial directed coherence: a new concept in neural structure 1034 determination. Biol Cybern. 2001 May 1;84(6):463–74. 1035 50. Arce-McShane FI, Ross CF , Takahashi K, Sessle BJ, Hatsopoulos NG. Primary motor and 1036 sensory cortical areas communicate via spatiotemporally coordinated networks at multiple 1037 frequencies. PNAS. 2016 Apr 15;201600788. 1038 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 56 51. Bechara A, Damasio H, Damasio AR, Lee GP . Different Contributions of the Human Amygdala 1039 and Ventromedial Prefrontal Cortex to Decision-Making. J Neurosci. 1999 Jul 1;19(13):5473–81. 1040 52. Schoenbaum G, Chiba AA, Gallagher M. Neural Encoding in Orbitofrontal Cortex and 1041 Basolateral Amygdala during Olfactory Discrimination Learning. J Neurosci. 1999 Mar 1042 1;19(5):1876–84. 1043 53. Michalareas G, Vezoli J, van Pelt S, Schoffelen JM, Kennedy H, Fries P . Alpha-Beta and Gamma 1044 Rhythms Subserve Feedback and Feedforward Influences among Human Visual Cortical Areas. 1045 Neuron. 2016 Jan 20;89(2):384–97. 1046 54. Pillow JW, Shlens J, Paninski L, Sher A, Litke AM, Chichilnisky EJ, et al. Spatio-temporal 1047 correlations and visual signalling in a complete neuronal population. Nature. 2008 1048 Aug;454(7207):995–9. 1049 55. Stevenson IH, Rebesco JM, Miller LE, Körding KP . Inferring functional connections between 1050 neurons. Curr Opin Neurobiol. 2008 Dec;18(6):582–8. 1051 56. Magrans de Abril I, Yoshimoto J, Doya K. Connectivity inference from neural recording data: 1052 Challenges, mathematical bases and research directions. Neural Networks. 2018 Jun 1053 1;102:120–37. 1054 57. Kobayashi R, Kurita S, Kurth A, Kitano K, Mizuseki K, Diesmann M, et al. Reconstructing 1055 neuronal circuitry from parallel spike trains. Nat Commun. 2019 Oct 2;10(1):4468. 1056 58. Sirota A, Montgomery S, Fujisawa S, Isomura Y , Zugaro M, Buzsáki G. Entrainment of 1057 Neocortical Neurons and Gamma Oscillations by the Hippocampal Theta Rhythm. Neuron. 1058 2008 Nov 26;60(4):683–97. 1059 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 57 59. Quirk MC, Sosulski DL, Feierstein CE, Uchida N, Mainen ZF . A defined network of fast-spiking 1060 interneurons in orbitofrontal cortex: responses to behavioral contingencies and ketamine 1061 administration. Front Syst Neurosci [Internet]. 2009 Nov 3 [cited 2025 May 8];3. Available from: 1062 https://www.frontiersin.orghttps://www.frontiersin.org/journals/systems-1063 neuroscience/articles/10.3389/neuro.06.013.2009/full 1064 60. Herzog LE, Pascual LM, Scott SJ, Mathieson ER, Katz DB, Jadhav SP . Interaction of Taste and 1065 Place Coding in the Hippocampus. J Neurosci. 2019 Apr 17;39(16):3057–69. 1066 61. Li JX, Yoshida T, Monk KJ, Katz DB. Lateral Hypothalamus Contains Two Types of Palatability-1067 Related Taste Responses with Distinct Dynamics. J Neurosci. 2013 May 29;33(22):9462–73. 1068 62. Kosar E, Grill HJ, Norgren R. Gustatory cortex in the rat. II. Thalamocortical projections. Brain 1069 Research. 1986 Aug 6;379(2):342–52. 1070 63. Mazzucato L, Fontanini A, Camera GL. Dynamics of Multistable States during Ongoing and 1071 Evoked Cortical Activity. J Neurosci. 2015 May 27;35(21):8214–31. 1072 64. Lang L, Camera GL, Fontanini A. Temporal progression along discrete coding states during 1073 decision-making in the mouse gustatory cortex. PLOS Computational Biology. 2023 Feb 1074 7;19(2):e1010865. 1075 65. Parras GG, Nieto-Diego J, Carbajal GV , Valdés-Baizabal C, Escera C, Malmierca MS. Neurons 1076 along the auditory pathway exhibit a hierarchical organization of prediction error. Nat Commun. 1077 2017 Dec 15;8(1):2148. 1078 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 58 66. Glaser JI, Perich MG, Ramkumar P , Miller LE, Kording KP . Population coding of conditional 1079 probability distributions in dorsal premotor cortex. Nature Communications. 2018 May 1080 3;9(1):1–14. 1081 67. Heidari-Gorji H, Ebrahimpour R, Zabbah S. A temporal hierarchical feedforward model explains 1082 both the time and the accuracy of object recognition. Sci Rep. 2021 Mar 11;11(1):5640. 1083 68. Dehaene S, Sergent C, Changeux JP . A neuronal network model linking subjective reports and 1084

physiological data during conscious perception. Proceedings of the National 1085 Academy of Sciences. 2003 Jul 8;100(14):8520–5. 1086 69. Gaillard R, Dehaene S, Adam C, Clémenceau S, Hasboun D, Baulac M, et al. Converging 1087 Intracranial Markers of Conscious Access. PLOS Biology. 2009 Mar 17;7(3):e1000061. 1088 70. Rogers PJ. Why a palatability construct is needed. Appetite. 1990 Jun;14(3):167–70; discussion 1089 180. 1090 71. Araujo JA, Studzinski CM, Larson BT, Milgram NW. Comparison of the cognitive palatability 1091 assessment protocol and the two-pan test for use in assessing palatability of two similar foods 1092 in dogs. 2004 Nov 1 [cited 2025 May 6]; Available from: 1093 https://avmajournals.avma.org/view/journals/ajvr/65/11/ajvr.65.11.1490.xml 1094 72. Breslin PAS. The Sense of Taste Encompasses Two Roles: Conscious Taste Perception and 1095 Subconscious Metabolic Responses. In: Think Tank [Internet]. Yale University Press; 2018 [cited 1096 2025 May 6]. p. 110–8. Available from: 1097 https://www.degruyterbrill.com/document/doi/10.12987/9780300235470-1098 017/html?srsltid=AfmBOopo1-t48Heuq2k_xlyWiTnwcA6QP4aYMaFChOt_KAeaiSt3OwPX 1099 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 59 73. Vincis R, Chen K, Czarnecki L, Chen J, Fontanini A. Dynamic Representation of Taste-Related 1100 Decisions in the Gustatory Insular Cortex of Mice. Current Biology. 2020 May 18;30(10):1834-1101 1844.e5. 1102 74. Forseth KJ, Pitkow X, Fischer-Baum S, Tandon N. What The Brain Does As We Speak [Internet]. 1103 bioRxiv; 2021 [cited 2025 May 6]. p. 2021.02.05.429841. Available from: 1104 https://www.biorxiv.org/content/10.1101/2021.02.05.429841v1 1105 75. Supp GG, Siegel M, Hipp JF , Engel AK. Cortical Hypersynchrony Predicts Breakdown of Sensory 1106 Processing during Loss of Consciousness. Current Biology. 2011 Dec 6;21(23):1988–93. 1107 76. Skinner FK, Gramoll S, Calabrese RL, Kopell N, Marder E. Frequency Control in Biological Half-1108 Center Oscillators. In: Eeckman FH, editor. Computation in Neurons and Neural Systems 1109 [Internet]. Boston, MA: Springer US; 1994 [cited 2025 May 6]. p. 223–8. Available from: 1110 https://doi.org/10.1007/978-1-4615-2714-5_36 1111 77. Diz-Pita É, Otero-Espinar MV . Predator–Prey Models: A Review of Some Recent Advances. 1112 Mathematics. 2021 Jan;9(15):1783. 1113 78. Vrugt M te. An introduction to reservoir computing [Internet]. arXiv; 2024 [cited 2025 May 6]. 1114 Available from: http://arxiv.org/abs/2412.13212 1115 79. Stone ME, Fontanini A, Maffei A. Synaptic Integration of Thalamic and Limbic Inputs in Rodent 1116 Gustatory Cortex. eNeuro [Internet]. 2020 Jan 1 [cited 2025 May 8];7(1). Available from: 1117 https://www.eneuro.org/content/7/1/ENEURO.0199-19.2019 1118 80. Ahissar E, Assa E. Perception as a closed-loop convergence process. Kleinfeld D, editor. eLife. 1119 2016 May 9;5:e12830. 1120 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 60 81. Jones SR, Kerr CE, Wan Q, Pritchett DL, Hämäläinen M, Moore CI. Cued Spatial Attention Drives 1121 Functionally Relevant Modulation of the Mu Rhythm in Primary Somatosensory Cortex. J 1122 Neurosci. 2010 Oct 13;30(41):13760–5. 1123 82. Beyeler A, Chang CJ, Silvestre M, Lévêque C, Namburi P , Wildes CP , et al. Organization of 1124 Valence-Encoding and Projection-Defined Neurons in the Basolateral Amygdala. Cell Reports. 1125 2018 Jan 23;22(4):905–18. 1126 83. Lui JH, Nguyen ND, Grutzner SM, Darmanis S, Peixoto D, Wagner MJ, et al. Differential encoding 1127 in prefrontal cortex projection neuron classes across cognitive tasks. Cell. 2021 Jan 1128 21;184(2):489-506.e26. 1129 84. Gongwer MW, Klune CB, Couto J, Jin B, Enos AS, Chen R, et al. Brain-Wide Projections and 1130 Differential Encoding of Prefrontal Neuronal Classes Underlying Learned and Innate Threat 1131 Avoidance. J Neurosci. 2023 Aug 9;43(32):5810–30. 1132 85. Baez-Santiago MA, Reid EE, Moran A, Maier JX, Marrero-Garcia Y , Katz DB. Dynamic taste 1133 responses of parabrachial pontine neurons in awake rats. Journal of Neurophysiology. 2016 Jan 1134 20;115(3):1314–23. 1135 86. Fontanini A, Katz DB. State-Dependent Modulation of Time-Varying Gustatory Responses. 1136 Journal of Neurophysiology. 2006 Dec 1;96(6):3183–93. 1137 87. Virtanen P , Gommers R, Oliphant TE, Haberland M, Reddy T, Cournapeau D, et al. SciPy 1.0: 1138 fundamental algorithms for scientific computing in Python. Nat Methods. 2020 Mar;17(3):261–1139 72. 1140 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 61 88. Ding M, Bressler SL, Yang W, Liang H. Short-window spectral analysis of cortical event-related 1141 potentials by adaptive multivariate autoregressive modeling: data preprocessing, model 1142 validation, and variability assessment. Biol Cybern. 2000 Jun 1;83(1):35–45. 1143 89. Maris E, Oostenveld R. Nonparametric statistical testing of EEG- and MEG-data. Journal of 1144 Neuroscience Methods. 2007 Aug 15;164(1):177–90. 1145 90. Candia-Rivera D, Valenza G. Cluster permutation analysis for EEG series based on non-1146 parametric Wilcoxon–Mann–Whitney statistical tests. SoftwareX. 2022 Jul 1;19:101170. 1147 91. Pedregosa F , Varoquaux G, Gramfort A, Michel V , Thirion B, Grisel O, et al. Scikit-learn: Machine 1148 Learning in Python. J Mach Learn Res. 2011 Nov 1;12(null):2825–30. 1149 92. Mahmood A, Mukherjee N, Stone B, Raymond M, Germaine H, Lin JY , et al. 1150 katzlabbrandeis/blech_clust: v1.1.0 [Internet]. Zenodo; 2025 [cited 2025 May 15]. Available 1151 from: https://zenodo.org/records/15175273 1152 93. Clarke SNDA, Ossenkopp KP . Taste reactivity responses in rats: influence of sex and the 1153 estrous cycle. American Journal of Physiology-Regulatory, Integrative and Comparative 1154 Physiology. 1998 Mar;274(3):R718–24. 1155 94. Travers JB, Norgren R. Electromyographic analysis of the ingestion and rejection of sapid stimuli 1156 in the rat. Behavioral Neuroscience. 1986;100:544–55. 1157 95. Mahmood A, Steindler J, Germaine H, Miller P , Katz DB. Coupled Dynamics of Stimulus-Evoked 1158 Gustatory Cortical and Basolateral Amygdalar Activity. J Neurosci. 2023 Jan 18;43(3):386–404. 1159 96. Salvatier J, Wiecki TV , Fonnesbeck C. Probabilistic programming in Python using PyMC3. PeerJ 1160 Comput Sci. 2016 Apr 6;2:e55. 1161 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint 62 97. Kispersky T, Gutierrez GJ, Marder E. Functional connectivity in a rhythmic inhibitory circuit using 1162 Granger causality. Neural Syst Circuits. 2011 May 25;1:9. 1163 98. Gerhard F , Kispersky T, Gutierrez GJ, Marder E, Kramer M, Eden U. Successful Reconstruction of 1164 a Physiological Circuit with Known Connectivity from Spiking Activity Alone. PLOS 1165 Computational Biology. 2013 Jul 11;9(7):e1003138. 1166 99. Stevenson IH, Rebesco JM, Hatsopoulos NG, Haga Z, Miller LE, Kording KP . Bayesian Inference 1167 of Functional Connectivity and Network Structure From Spikes. IEEE Transactions on Neural 1168 Systems and Rehabilitation Engineering. 2009 Jun;17(3):203–13. 1169 100. Keeley SL, Zoltowski DM, Aoi MC, Pillow JW. Modeling statistical dependencies in multi-1170 region spike train data. Current Opinion in Neurobiology. 2020 Dec 1;65:194–202. 1171 101. Seabold S, Perktold J. Statsmodels: Econometric and Statistical Modeling with Python. In 1172 Austin, Texas; 2010 [cited 2025 May 16]. p. 92–6. Available from: 1173 https://doi.curvenote.com/10.25080/Majora-92bf1922-011 1174 1175 .CC-BY-NC-ND 4.0 International licenseavailable under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (whichthis version posted July 4, 2025. ; https://doi.org/10.1101/2025.07.01.662567doi: bioRxiv preprint