Summary of rat subjects and experimental conditions.

Annotation of EMG waveforms using scored video labels.

A) A schematic of the data recording and annotation setup. Rats implanted with electromyography (EMG) electrodes and an intraoral cannula were given a battery of tastants while EMG and video recording were captured.B) Video scoring similarity between the two raters was significantly higher than chance for all labels (*** p < 0.001, one-tailed t-test). x, + = accuracy score with scorer 1,2 as ground-truth respectively. C-D) Examples of video-scored behaviors (top), and corresponding EMG signals (bottom) for trials of aversive quinine (C) and palatable sucrose (D). Note different y-scales for EMG envelopes between C and D. E) Overlain waveforms for each group of labelled movements (Lateral Tongue Movements [LTMs], which made up <5% of behaviors, are not shown). F) Breakdown of the labelled dataset by behavior.

Optimization space of classifier hyperparameters.

The XGBoost model accurately classifies EMG data.

A) The XGBoost architecture represented as a set of tree-based models. B) Interrogation of feature importance using SHAP values shows that all sets of features show stronger contribution to predictions than expected by chance (Null SHAP values; ***: p < 0.001, Wilcoxon signed-rank). C) Classifier model performance is robust to hyperparameter specification—models trained without hyperparameter optimization exhibit high cross-validation accuracy and hyperparameter optimization results in only small accuracy gains. Inset) Distributions of final sets of hyperparameters (normalized to search range) following optimization. Parameters span almost the entire search range but some show tighter clustering.

XGB outperforms previously used classifiers.

A) Example predictions by QDA, BSA, and XGB classifiers for four tastants. B) Session-averaged confusion matrices of class-specific predictions for all three models (mean±STD). C) Weighted accuracy for all three models. XGB outperforms both QDA and BSA (Tukey post-hoc). D) Prediction precision and recall for all three models across label class. XGB performs on par with or better than QDA and BSA for all classes on both metrics. E) Weighted F1-score for all 3 models. XGB outperforms both QDA and BSA (Tukey post-hoc). F) D’ values for binary classifications. XGB consistently performs best for all class labels. G) Comparison of temporal specificity of rejection/ingestion-related responses (paired t-tests). * p < 0.05, ** p < 0.01, *** p < 0.001, for all panels

MTMs are comprised of three distinct behavioral subtypes.

A) Example MTM waveforms in UMAP space, showing individual waveforms falling into 3 clusters. Inset) BIC values for 1-14 clusters indicating that three clusters are optimal for this dataset, based on the elbow method. B) Distribution of optimal cluster number across 50 iterations for each test session (mode = 3). C) Comparison of intra-cluster versus inter-cluster Mahalanobis distances indicates high separability of clusters for each session (*** p < 0.001, Kolmogorov-Smirnov test). D) Schematic depicting cluster index alignment across test sessions using cosine similarity. E) MTM waveforms across all test sessions projected into UMAP space, colored by cluster index label. (Left) Two-dimensional view (“the croissant”) highlighting the spread of the MTM distribution. (Right) Surface plot showing cluster densities with density peaks (red dots). F) Confusion matrix from a SVM (support vector machine) classifier showing robust cluster prediction accuracy for leave-one-animal-out cross-validation. G) Representative waveforms for each cluster, selected as those closest to the density peak. H) The first principal component (PC 1) values, which are derived from waveform shape, are significantly different across clusters (*** p < 0.001, Mann-Whitney U-Test). I) Counts of video-annotated lateral tongue movements classified into each behavioral category.

Summary of Kruskal-Wallis test results for feature differences between clusters.

MTM subtype frequencies shift across the taste response.

A) Example contour plot of UMAP-reduced MTM feature distributions categorized by their timing relative to average palatability activity onset (“before” [blue] versus “after” [green] 800 ms post-stimulus delivery).B) Session-wise p-values quantifying the shift in MTM feature distributions before versus after average palatability activity, calculated using bootstrap resampling. The median p-value across multiple repeats for all sessions is <0.05. Colors denote sessions from the same animal. C) Bottom: Raster plot, and Top: mean rate of occurrence of MTM cluster #2 across 30 trials of sucrose in an example test session, aligned to the average palatability activity onset (800 ms post-stimulus delivery). D) Distribution of Poisson means test p-values across all tastant-behavior comparisons for every test session (gray) compared to a null distribution (yellow). E) Tastant-behavior matrix, with brightness of color indicating the proportion of significant test sessions showing broad changes in behavior frequency.

Emergence of palatability encoding in GC neural activity precedes changepoints in MTM occurrence.

A) Comparison of changepoint models with 2-8 states applied to timeseries of oral behaviors reveals that 3-state models best describe behavior dynamics in 2000 ms post-stimulus delivery (** p < 0.01, pairwise paired t-test with Bonferroni correction). B) Representative example of inference performed by changepoint model. i) Raster plot of behaviors across tastes; ii) State sequences inferred by changepoint model; iii) Histograms of probabilities for each behavior per state and taste inferred by changepoint model. N = no movement, G = gape, 1, 2, 3 = MTM cluster 1, 2, and 3, respectively. C) Comparison of normalized magnitudes of difference between inferred behavior probability vectors for subsequent states, including either all categories or MTMs only. Only quinine (QHCl) is affected when gapes and no movements are removed (*** p < 0.001, t-test with Bonferroni correction). D) Comparison of changepoint models with 2-8 states applied to timeseries of electrophysiological responses; 3-state models best describe neural dynamics in 2000ms post-stimulus delivery (** p < 0.01, pairwise paired t-test with Bonferroni correction). E) Only activity in the 3rd neural state (i.e., the state that follows the 2nd changepoint) is significantly correlated with palatability. F) The neural and behavioral latencies of both changepoints 1 & 2 are significantly correlated in a higher number of sessions than expected by chance G) Comparison of lags between neural and behavioral changepoints latencies shows a significant neural-to-behavioral lag for the 2nd changepoint but not the 1st. Black dots and connecting lines show session-average latencies. H) Same data as in G) plotted as lags of behavioral changepoints against mean latency of neural changepoint to highlight difference of lags between both changepoints.