Figure 8 odor sequence task, cocaine self-administration, and behavior

(A) Schematic representation of the trial events in the odor-sequence task. The initiation of each trial was indicated by the illumination of two overhead houselights. After poking into the central odor port and sampling the presented odor, rats had the option to respond with a ‘go’ to receive a sucrose reward or a ‘no-go’ to avoid a prolonged inter-trial interval. (B) The six odors were grouped into two sequences, S1 and S2, each consisting of four odor positions (P1–P4). The sequences alternated in a “figure eight” pattern. The numbers at each position represent odor identities, and the blue/red symbols (+/−) indicate rewarded and non-rewarded trials. (C) Timeline of the experimental procedure. (D) The number of active (solid lines) and inactive (dashed lines) lever presses during sucrose self-administration sessions (upper; n = 3) and cocaine self-administration sessions (lower; n = 3). (E) Reconstruction of recording sites, with red squares indicating the locations of electrodes. (F) Percentage of correct responses (% correct) are shown for each trial type during single-unit recording sessions. A significant group difference emerged specifically at position P3 between the Sucrose and Cocaine groups (F (1,288) = 13.1, p = 3.4 × 10-4, η p 2 = 0.043; One-way ANOVA). Blue denotes rewarded trial types, while red denotes non-rewarded trial types. (G) Reaction time was defined as the interval between odor port exit and water well entry. For correct no-go trials, where no movement toward the water well was made, a fixed reaction time of 2 seconds—the full response window—was assigned. Reaction time differed significantly between the Sucrose and Cocaine groups overall (F (1,1158) = 6.6, p = 0.01, ηp2 = 0.006; One-way ANOVA), with position-specific analyses revealing significant differences at P1, P3, and P4 (P1: F (1,288) = 152.9, p = 1.9 × 10-28, ηp2 = 0.35; P3: F (1,288) = 9.3, p = 2.5 × 10-3, ηp2 = 0.03; P4: F(1, 288) 139.6, p = 1.6 × 10-26, ηp2 = 0.33; One-way ANOVA). (H) Poke latency, measured as the time from light onset to odor port entry. Significant group differences were found across positions (F (1,1158) = 4.0, p = 0.04, ηp2 = 0.004; One-way ANOVA), with specific differences emerging at positions P2, P3, and P4 (P2: F (1,288) = 10.0, p = 1.7 × 10-3, ηp2 = 0.034; P3: F (1,288) = 12.4, p = 5.0 × 10-4, ηp2 = 0.04; P4: F (1,288) = 5.7, p = 0.018, ηp2 = 0.02; One-way ANOVA). (I) For the absolute S2–S1 difference in reaction time across positions between Sucrose and Cocaine groups, two-way ANOVA revealed significant main effects of group (F (1,572) = 6.4, p = 0.012, ηp2 = 0.01) and position (F (3,572) = 3.3, p = 0.021, ηp2 = 0.017). Black circles denote sucrose data, with shading that deepens progressively from P1 through P4. Cocaine data are indicated by red circles, with coloration shifting from light to dark red across positions P1 to P4. (J) Two-way ANOVA revealed significant main effects of group (F (1,572) = 5.6, p = 0.018, ηp2 = 0.01) and position (F (3,572) = 37.7, p = 3.2 × 10-22, ηp2 = 0.16) on the absolute poke latency difference between S1 and S2 in Sucrose and Cocaine groups. Sucrose data are plotted as graded black circles, with shading progressing from light to dark across positions P1 to P4. Cocaine data are plotted as graded red circles, with shading progressing from light red to dark red across positions P1 to P4. Error bars represent standard errors of the mean (SEMs). *p < 0.05; **p < 0.01; ***p < 0.001.

Discriminating sequences S1 versus S2 at single-unit level

(A-D) Heatmaps showing the activity of individual OFC neurons. Each panel illustrates the firing pattern of a single neuron, capturing its response dynamics across task epochs. Each small grid represents one epoch at a specific position, arranged in the direction of the arrow following the sequence: [“ITI-a”, “Light”, “Poke”, “Odor”, “Unpoke”, “Choice”, “Outcome”, “Post-Out”, “ITI-b”], repeated for all four positions of each sequence. The left columns represent single OFC neurons from the Sucrose group, where the intensity of left-side epochs closely matches that of the right-side epochs, indicating strong generalization. In contrast, the Cocaine group shows comparatively weaker generalization across sides. S1, sequence 1; S2, sequence 2. (E) Percentage of neurons selective at each position in S1 versus S2 in the Sucrose (upper panel) and Cocaine (lower panel) groups. The selectivity was calculated for different task epochs [“ITI-a”, “Light”, “Poke”, “Odor”, “Unpoke”, “Choice”, “Outcome”, “Post-Out”, “ITI-b”] for all four positions (one-way ANOVA). Epochs within the red dashed boxes indicate selectivity exceeds a 95% confidence interval (CI) (Sucrose: χ2’s > 5.4; p’s < 0.02; Cocaine: χ2’s > 4.0; p’s < 0.05; Chi-squared test). The p-values were not corrected, and the dotted lines represent the chance level of selectivity according to this criterion. Notably, rats could discriminate between the two sequences (S1 vs. S2) based solely on current sensory information at two task epochs [“Odor” at P3 and P4; black bars]. At all other task epochs, indicated by gray bars, the discrimination relied on an internal memory of events. (F) Line plot showing neuronal selectivity at each position in S1 versus S2 for each neuron in both the Sucrose (black line) and Cocaine (red line) groups. The selectivity was calculated at different task epochs for all four positions. Each asterisk indicates that there are significant differences between the two groups (χ2’s > 4.8; p’s < 0.03; Chi-squared test). *p < 0.05; error bars are standard deviations (SDs).

Cocaine use reduces the ability of OFC to generalize task-irrelevant odor-overlapping sequences.

(A) Decoding accuracy of S1 versus S2 was evaluated within each of the nine task epochs for positions P1 to P4. Error bars represent SDs, and each asterisk indicates that the mean decoding accuracy exceeds a 95% confidence interval calculated using the same decoding process with label-shuffled data. The dotted lines represent the chance level of decoding. The meaning of the black bars is consistent with Figure 2E. (B) Decoding accuracy of S1 versus S2 at each position was assessed using varying ensemble sizes for the Sucrose (black) and Cocaine (red) groups across all epochs. The Cocaine group demonstrated higher decoding accuracy and reduced generalization of task-irrelevant, overlapping sequences compared to the Sucrose group (F (1,118) = 16.8, p = 7.8 × 10-5, ηp2 = 0.12; One-way ANOVA). P1, P2, P3, and P4 represent positions 1, 2, 3, and 4, respectively. (C) Decoding accuracy of S1 versus S2 at each position was assessed using varying ensemble sizes for the Sucrose group (black, same data as in Figure 3B) and the Less Training group (blue, Zhou et al., 2021) across all epochs. The decoding accuracy in the Less Training group was comparable to that observed in the Cocaine group (Figure 3B, red) (F (1,118) = 0.01, p = 0.92, ηp2 = 9.0 × 10-5; One-way ANOVA). Additionally, the Less Training group exhibited significantly higher decoding accuracy than the Sucrose group (F (1,118) = 18.1, p = 4.2 × 10−5, ηp2 = 0.13; One-way ANOVA). P1, P2, P3, and P4 represent positions 1, 2, 3, and 4, respectively. *p < 0.05; error bars indicate SDs. (D) Confusion matrices showing within-sequence position decoding from OFC ensemble activity at four task positions for the Sucrose and Cocaine groups. The y-axis denotes the rats’ actual position, and the x-axis indicates the predicted position. Brighter colors reflect higher decoding probabilities. (E) Quantification of within-sequence decoding accuracy across positions revealed a significant difference between groups (p = 0.027; W = 1118; two-sided Wilcoxon rank-sum test), with Sucrose shown in black and Cocaine in red. (F) Confusion matrices showing across-sequence position decoding from OFC ensemble activity at the same four task positions. Axes are as in (D), with brighter colors indicating higher decoding probabilities. (G) Quantification of across-sequence decoding accuracy also revealed a significant group difference (p = 0.046; W = 1144; two-sided Wilcoxon rank-sum test), with Sucrose shown in black and Cocaine in red.

Cocaine use impairs the OFC’s function in temporal cognition and decreases the generalization of task-irrelevant odor-overlapping sequences

(A-B) Ten components of TCA applied to Sucrose(A) and Cocaine(B) neuron activity are displayed. Each component consists of a neuron factor (left column), a temporal factor spanning nine events with eight time points per event (middle column), and a trial factor (right column). Temporal factors are grouped as follows: 1–8 (preTrial), 9–16 (Light), 17–24 (Poke), 25–32 (Odor), 33–40 (Unpoke), 41–48 (Choice), 49–56 (Outcome), 57–64 (postTrial1), and 65–72 (postTrial2). The dataset includes eight trial types, each with 40 trials. Trial factors 1–80 and 241–320 represent rewarded (positive) trials, while 81–240 correspond to unrewarded (negative) trials. Trials are further categorized as follows: 1–80 (P1), 81–160 (P2), 161–240 (P3), and 241–320 (P4). These low-dimensional components were extracted from a 10-component model for both the Sucrose(A) and Cocaine groups(B). In the Sucrose group, components 1, 6, 7, and 10 encode positive trials, while components 2, 3, and 8 encode negative trials. Similarly, in the Cocaine group, components 1, 4, 6, 7, and 10 encode positive trials, whereas components 2, 3, 5, 8, and 9 encode negative trials. Notably, in the Sucrose group, beyond reward-related components, additional components encode aspects of temporal cognition (components 5 and 9), along with early components broadly active across most trials (component 4). (C) Plot illustrating mutual information (MI) between across-trial factors and trial types. The Sucrose group exhibits lower MI, particularly at components 5 and 9. A two-way ANOVA confirmed a significant reduction in MI for the Sucrose group (F (1,1360) = 571.6, p = 9.7 × 10−106, ηp2 = 0.3), especially at these components, which showed consistent patterns across all trial types, aligning with the plots in (A) of the Sucrose group, indicating a weaker dependency between across-trial factors and trial types. A significant difference was also observed across components (F (9,1360) = 266.3, p = 1.2 × 10-292, ηp2 = 0.64). (D) Plot illustrating mutual information (MI) between across-trial factors and time. The Sucrose group demonstrates greater MI, particularly at components 5 and 9, suggesting a stronger dependency between across-trial factors and time. A two-way ANOVA confirmed a significant increase in MI for the Sucrose group (F (1,1360) = 850.1, p = 1.4 × 10-145, ηp2 = 0.38). Additionally, a significant difference was observed across components (F (9,1360) = 176.8, p = 1.4 × 10-221, ηp2 = 0.54). Error bars are SEM.

The ability of the OFC to generalize task-irrelevant odor-overlapping sequences decreases as learning progresses

(A) Organization of sixteen odors into two pairs of sequences (S1a-S1b and S2a-S2b), with each sequence consisting of six trials or positions (P1-P6). The blue ‘+’ symbol represents rewarded trials, while the red ‘-’ symbol represents non-rewarded trials. Odor identities are represented by numbers 0-15, and arrows indicate transitions between sequences. (B-C) The percentage of decoding accuracy was analyzed across different groups with varying training days and ensemble sizes. Significant differences were observed between S1a and S1b across all time periods (days 1–5, days 6–10, and days 11–15) for different numbers of selected neurons (F (2,42) = 11.2, p = 1.3 × 10-4, ηp2 = 0.35). In contrast, no significant differences were found between S2a and S2b (F (2,42) = 0.17, p = 0.84, ηp2 = 8.2 × 10−3). Additionally, significant differences were identified between S1 and S2 (F (1,88) = 8.69, p = 4.1 × 10-3, ηp2 = 0.09). The error bars indicate SDs. 24-Position Odor Sequence Taska Rats were trained in 18” x 18” aluminum behavioral chambers equipped with an odor port and a fluid well for odor and sucrose delivery, respectively. Odor and fluid delivery were controlled by solenoids via a custom C++ program. Infrared beam sensors monitored entries into both ports. Each trial was cued by the illumination of two house lights in the front wall above the odor panel. Rats initiated a trial by nose-poking into the odor port within 5 seconds of light onset. Upon successful initiation, an odor was delivered after a 500-ms delay and continued until the rat withdrew. Trials were aborted if the rat exited the port before 500 ms. Otherwise, after withdrawal, the rat had 2 seconds to respond at the fluid well. On rewarded trials, a correct response at the well triggered sucrose delivery (10% w/v; 50 μL) after a variable delay (400–1500 ms). The trial ended when the rat left the well, at which point the house lights turned off, initiating the inter-trial interval (ITI). If no response was made within the 2-s window, or if the rat responded on a non-rewarded trial (a rare occurrence), the trial also ended without reward. Correct responses were followed by a 4-s ITI; incorrect responses led to an 8-s ITI. Odors were drawn from a pool of 16 and organized into four sequences: S1a, S1b, S2a, and S2b. Rewarded (+) and non-rewarded (–) trials were defined by the following sequence structures:

Training proceeded in stages. Rats were first shaped to poke into the odor port and then respond at the fluid well. They were initially trained on a single odor pair (one rewarded, one not) from sequence S1a or S1b. Sessions contained up to 480 trials. Once rats achieved >90% correct performance, additional odor pairs were gradually introduced until they performed reliably on full S1a and S1b sequences. Next, rats were trained to discriminate odors 13 and 14 from sequence S2, including multiple reversals of their reward valence. Once they achieved >75% accuracy across trial types and completed three reversals, the remaining odors from S2a and S2b were added. In the final training phase, rats performed sessions with all four sequences (S1a, S1b, S2a, S2b). Each sequence was presented 20 times per session (480 trials total), with transitions between S1 and S2 sequences occurring in a pseudo-random order. S1 sequences were followed by S2 with ∼55%/45% probability (2a/2b), and S2 sequences were followed by S1 with ∼67%/37% probability (1a/1b). This sequence structure was repeated throughout each session. Prior to electrode implantation, rats were trained on the full task for at least three weeks and had to reach >75% accuracy on all trial types.

Two-way ANOVA of sequence and position effects

(A–B) Percentage of neurons showing significant effects of sequence or interaction in the Sucrose (A) and Cocaine (B) groups. No significant sequence or interaction effects were detected across any epoch in the Sucrose group (χ2 < 6.5, p > 0.06; Chi-squared test). In contrast, the Cocaine group showed a markedly higher proportion of neurons exhibiting interaction effects than sequence effects during the Light, Poke, Odor, Post-Out, and ITIb epochs (χ2 > 5.5, p < 0.034; Chi-squared test). (C) Percentage of neurons showing sequence effects. A significant group difference was observed only at the ITIb epoch (χ2 = 11, p = 8.3 × 10−3; Chi-squared test). (D) Percentage of neurons showing position effects. Significant group differences were found at the epochs of Poke, Odor, Unpoke, Choice, Outcome, and ITIb (χ2 > 13.8, p < 3.0 × 10−4; Chi-squared test). (E) Percentage of neurons showing interaction effects. The Cocaine group showed significantly more interaction effects than the Sucrose group at epochs of Poke, Odor, Post-Out, and ITIb (χ2 > 14.5, p < 4.2 × 10−4; Chi-squared test).

Correlation of preferred positions

Plots display regression lines for the Sucrose and Cocaine groups. The black line represents the Sucrose group, while the dark red line corresponds to the Cocaine group. A statistical comparison using Fisher’s r-to-z transformation revealed a significant difference between the two groups (z = 4.92, p = 8.6 × 10−7). P1, P2, P3, and P4 denote different positions.

Error and similarity plots for TCA

The optimization landscape of TCA may contain suboptimal solutions (local minima), necessitating iterative optimization to minimize a cost function. To evaluate the stability of this process, we generated error and similarity plots. Specifically, we ran the TCA optimization algorithm 100 times for 10 components, each initialized with random conditions, and plotted the normalized reconstruction error across all runs. This approach enabled us to assess whether certain runs converged to local minima with high reconstruction errors. As shown in Figure 4—figure supplement 1A-B (left panels), the error plots indicate that all 100 runs at a fixed number of components produced nearly identical reconstruction errors. Additionally, models with 9 and 10 components exhibited very similar errors. The similarity plot (Figure 4—figure supplement 1A-B, right panel) further quantifies the consistency of TCA models across different component numbers (horizontal axis). For each model, similarity scores were computed relative to the best-fit model with the same number of components, with the lines representing the mean similarity as a function of the number of components. Across all tested component numbers, the 100 repeated runs per component showed substantial overlap and consistently yielded similarity scores above 0.8, indicating high quantitative consistency. Our tests indicate that adding more than ten components reduces model reliability, making them less identifiable. (A-B) Error plots (left panel) and similarity plots (right panel) for Sucrose (A) and Cocaine group(B). Error plots illustrate the normalized reconstruction error across TCA models with varying component numbers. Each black dot corresponds to a model fit using different initial parameters, all of which produced nearly identical performance. Notably, reconstruction error showed almost no further improvement beyond ten components for both groups. The similarity plot displays the similarity score for TCA models with different component numbers. Each black dot represents the similarity of a given model to the best-fit model with the same number of components, demonstrating high repeatability for both groups.

Four component model, cocaine impairs OFC temporal cognition and sequence generalization with irrelevant overlapping odors

(A–B) Four components from TCA applied to Sucrose (A) and Cocaine (B) neuron activity are shown. Each component includes a neuron factor (left), a temporal factor spanning 72 time points across nine events (middle), and a trial factor (right). Temporal events are grouped as: 1–8 (preTrial), 9–16 (Light), 17–24 (Poke), 25–32 (Odor), 33–40 (Unpoke), 41–48 (Choice), 49–56 (Outcome), 57–64 (postTrial1), and 65–72 (postTrial2). The dataset contains eight trial types (40 trials each): trials 1–80 and 241–320 are rewarded (P1, P4), and 81–240 are unrewarded (P2, P3). In Sucrose, components 1, and 2. In Cocaine, components 1, 2, and 4 encode positive trials; 3 encode negative trials. Additionally, Sucrose components 3 and 4 relate to temporal cognition. (C) Mutual information (MI) between across-trial factors and trial types is lower in the Sucrose group, especially at components 1, 3 and 4, indicating weaker dependency. Two-way ANOVA confirmed a significant group difference (F (1,544) = 1074.5, p = 6.7 × 10-131, ηp2 = 0.66) and component effect (F (3,544) = 17.7, p = 5.2 × 10-11, ηp2 = 0.089). (D) MI between across-trial factors and time is higher in Sucrose, particularly at components 1, 2 and 4, reflecting stronger temporal encoding. ANOVA showed a significant group effect (F (1,792) = 36.5, p = 2.3 × 10-9, ηp2 = 0.044) and component effect (F (3,792) = 6.5, p = 2.4 × 10-4, ηp2 = 0.024). Error bars denote SEM.

Fourteen component model, cocaine use impairs OFC function in temporal cognition and hinders generalization across odor-overlapping, task-irrelevant sequences

(A–B) Fourteen components from TCA applied to Sucrose (A) and Cocaine (B) neural activity are shown. Each component includes a neuron factor (left), a temporal factor spanning nine events with eight time points each (middle), and a trial factor (right). Temporal segments are grouped as: preTrial (1–8), Light (9–16), Poke (17–24), Odor (25–32), Unpoke (33–40), Choice (41–48), Outcome (49–56), postTrial1 (57–64), and postTrial2 (65–72). The dataset comprises eight trial types (40 trials each): rewarded trials (1–80, 241–320; P1, P4) and unrewarded trials (81–240; P2, P3). Low-dimensional components were extracted from a 14-component model for each group. In Sucrose, components 2, 7, 8, 12 and 13 encode positive trials; components 1, 4, 10, 11 and 14 encode negative trials. In Cocaine, positive trial components are 2, 3, 5,7, 8, 11, 12, and 13; negative trials are encoded by components 1, 4, 6, 9, 10, and 14. Additionally, Sucrose components 3, 5 and 6 encode temporal structure, and component 9 is broadly active across trials. (C) Mutual information (MI) between across-trial factors and trial types is lower in the Sucrose group, particularly at components 3, 5 and 6, indicating weaker encoding of trial identity. A two-way ANOVA confirmed a significant group difference (F (1,308) = 571.6, p = 4.1 × 10-25, ηp2 = 0.29) and component effect (F (13,308) 33.0, p = 1.0 × 10-50, ηp2 = 0.58). (D) MI between across-trial factors and time is higher in Sucrose, especially at components 5 and 9, reflecting stronger temporal encoding. A two-way ANOVA revealed a significant group effect (F (1,2772) = 3.8, p = 0.046, ηp2 = 0.0014) and component effect (F (13,2772) = 5.5, p = 5.6 × 10-10, ηp2 = 0.025). Error bars represent SEM.

Ten-component model for individual Sucrose rats

(A, C, E) Ten-component TCA models are presented for neuronal activity recorded from individual sucrose-trained rats. Each component comprises three factors: a neuron factor (left column), a temporal factor spanning nine task events with eight time points per event (middle column), and a trial factor (right column). Temporal factors are organized as follows: 1–8 (preTrial), 9–16 (Light), 17–24 (Poke), 25–32 (Odor), 33–40 (Unpoke), 41–48 (Choice), 49–56 (Outcome), 57–64 (postTrial1), and 65–72 (postTrial2). The dataset includes eight trial types with 40 trials each. Trial factors 1–80 and 241–320 correspond to rewarded (positive) trials, while 81–240 correspond to unrewarded (negative) trials. Trials are further categorized by position: 1–80 (P1), 81–160 (P2), 161–240 (P3), and 241–320 (P4). These low-dimensional components reflect the latent structure of population activity. Across Sucrose rats #J505, #J506, and #J507, several TCA components consistently distinguished trial valence. Components 3, 4, and 9 were associated with positive trials in all three rats, with components 6 and 7 additionally contributing in #J505 and #J506. Negative trials were reliably encoded by components 5 and 8 across all rats, with component 6 also contributing in #J507. Furthermore, each rat exhibited temporal components (typically 2 and 10, with 7 also involved in #J507) and a broadly active early component (1) that was consistent across trial types. (B, D, F) Error plots (left panels) and similarity plots (right panels) are shown for each Sucrose rat. The error plots demonstrate that all 100 runs with a fixed number of components yielded nearly identical reconstruction errors, indicating stable model convergence. Models with 9 and 10 components produced similarly low errors, with minimal improvement beyond 10 components. The similarity plots quantify consistency across model runs. For each component number, similarity was computed relative to the best-fit model, and mean similarity is plotted as a function of component number. Across all tested component counts, repeated runs produced high similarity scores (generally > 0.8), indicating that the decomposition was highly stable and reproducible for all Sucrose rats.

Ten-component TCA model applied to individual rats from the Cocaine group

(A, C, E) TCA-derived ten-component models are shown for individual Cocaine rats. Each component includes a neuron factor (left), a temporal factor spanning nine task events (middle), and a trial factor (right). Temporal factors cover preTrial (1–8), Light (9–16), Poke (17–24), Odor (25–32), Unpoke (33–40), Choice (41–48), Outcome (49–56), postTrial1 (57–64), and postTrial2 (65–72). The dataset includes eight trial types (40 trials each): trials 1–80 and 241–320 are rewarded (positive), and 81–240 are unrewarded (negative), grouped by position (P1–P4). In the Cocaine rats, positive trials were consistently encoded by components 1, 3, 4, 6, 9, and 10. Negative trials were represented by components 2, 5, and 8 across all Cocaine rats, with component 7 additionally contributing to rats #J509 and #510. (B, D, F) Error plots (left) and similarity plots (right) are shown for each Cocaine rat. The error plots indicate that across 100 runs with a fixed number of components, reconstruction errors were nearly identical, demonstrating stable model convergence. Models with 9 or 10 components achieved comparably low errors, with little improvement beyond 10 components. The similarity plots quantify consistency across runs: for each component count, similarity was calculated relative to the best-fit model, and mean similarity is plotted as a function of component number. Across all component counts, repeated runs yielded high similarity scores (generally >0.8), confirming that the decomposition was stable and reproducible in all sucrose rats.