Computational mechanisms for temporal integration in the anterior claustrum
- School of Biological Sciences, College of Natural Sciences, Seoul National University, Republic of Korea
Figures
Figure 1 with 3 supplements
RNN simulation capable of performing the delayed escape test.
(A) Schematic diagram of the delayed escape task. On day 1, animals were conditioned with a light cue paired with an electric footshock. On day 2, they were adapted to the test context with only a single crossing permitted. On day 3, animals were placed in one compartment of the test context and presented with a 20-s CS, followed by a 5-s delay, after which the outlet door was opened to allow escape. The test was conducted only once. (B) Schematic diagram of the task structure that the RNN model was designed to simulate. (C) Graphs of crossing latency, measured as the interval from door opening to crossing in rats and as derived from RNN simulations. (Rats: CS, n = 17; Neutral, n = 16; *p = 0.0335, unpaired t-test; RNN: CS, n = 147; Open, n = 108; ****p < 0.0001, Mann–Whitney). (D) Clustering of all recorded claustral neurons in rats (n = 203). (E) Mean Z-scored firing rates of non-exploratory Cluster 1 neurons in the CS group (n = 25, rat data). Non-exploratory Cluster 1 denotes the cluster showing increased persistent activity both during CS presentation and thereafter. The CS group corresponds to the CS +door-opening condition in the RNN model. (F) Average Z-score for each behavioral period shown in E. Bars: Friedman test, ****p < 0.0001; Dunn’s multiple comparisons—Base vs Cue, ****p < 0.0001; Base vs Interval, **p = 0.0085; Base vs Early Open, **p = 0.0075; Base vs Late Open, *p = 0.0455; Base vs Post-cross, p > 0.9999. (G) Correlation between crossing latency and Cluster 1 activity in the CS group (n = 12; Spearman r = 0.6503, *p = 0.0257). (H) Clustering for simulated neurons (n = 100 neurons). (I) Mean Z-scored firing rates of Cluster 1 neurons in the CS +door-opening condition (n = 24 neurons, RNN) across all simulated trials. (J) Average Z-score for each behavioral period shown in I. Bars: Friedman ****p < 0.0001; Dunn’s—Base vs Cue, ****p < 0.0001; Base vs Interval, ***p = 0.0001; Base vs Early Open, ****p < 0.0001; Base vs Late Open, **p = 0.0034; Base vs Post-cross, p = 0.7134. (K) Correlation between crossing latency and Cluster 1 activity in the CS + door-opening condition (147 trials; Spearman r = –0.1628, *p = 0.0488). (L) Heatmaps of single-neuron activity corresponding to panels E and I. Left: rat data (firing rate, Hz); Right: RNN data (Z-score). Neurons are ordered by overall activity. (M) Crossing latency in rats with anterior claustrum inhibition during the 5-s period between CS offset and door opening, compared with control virus-expressing animals (left, inhibition, n = 8; control, n = 9; **p = 0.0079, Mann–Whitney). Corresponding results in the RNN are shown for selective inhibition of Cluster 1 neurons during the same 5-s window simulation (right, inhibition, n = 28; control, n = 37; ****p < 0.0001, Mann–Whitney). Bars show mean ± SEM. Panels A, D–G, and L (left) are adapted from Figures 1A and 4B, D, E, and S4E of Han et al., 2024, Cell Reports, Elsevier.
Figure 1—figure supplement 1
RNN activity of Clusters 1–3 under CS and door-opening only conditions.
(A) Cluster 1, door-opening only. Left, Z-scored firing rate of Cluster 1 neurons (n = 24) in a representative trial. Middle, average Z-score for each behavioral period across all trials; Friedman test, ****p < 0.0001; Dunn’s post hoc—Base vs Cue, ****p < 0.0001; Base vs Interval, ***p = 0.0001; Base vs Early Open, ****p < 0.0001; Base vs Late Open, **p = 0.0034; Base vs Post-escape, p = 0.7134. Right, Spearman correlation between crossing latency and Cluster 1 activity (n = 147; r = 0.1143, p = 0.2388). (B) Cluster 2, CS. Left, Z-scored firing rate of Cluster 2 neurons (n = 38) in a representative trial. Middle, average Z-score for each behavioral period across all trials; Friedman ****p < 0.0001; Dunn’s—Base vs Cue, ***p < 0.003; Base vs Interval, ****p < 0.0001; Base vs Early Open, **p = 0.0072; Base vs Late Open, ***p = 0.0002; Base vs Post-escape, ****p < 0.0001. Right, correlation between crossing latency and Cluster 2 activity (n = 147; r = 0.0618, p = 0.4573). (C) Cluster 2, door-opening only. Left, Z-scored firing rate of Cluster 2 neurons (n = 38) in a representative trial. Middle, average Z-score for each behavioral period across all trials; Friedman ****p < 0.0001; Dunn’s—Base vs Cue, p > 0.9999; Base vs Interval, p > 0.9999; Base vs Early Open, ***p = 0.0009; Base vs Late Open, ****p < 0.0001; Base vs Post-escape, ****p < 0.0001. Right, correlation between crossing latency and Cluster 2 activity (n = 147; r = –0.0717, p = 0.4607). (D) Cluster 3, CS. Left, Z-scored firing rate of Cluster 3 neurons (n = 38) in a representative trial. Middle, average Z-score for each behavioral period across all trials; Friedman ****p < 0.0001; Dunn’s—Base vs Cue, p > 0.9999; Base vs Interval, p > 0.9999; Base vs Early Open, ****p < 0.0001; Base vs Late Open, ****p < 0.0001; Base vs Post-escape, ****p < 0.0001. Right, correlation between crossing latency and Cluster 3 activity (n = 147; r = 0.3335, ****p < 0.0001). (E) Cluster 3, door-opening only. Left, Z-scored firing rate of Cluster 3 neurons (n = 38) in a representative trial. Middle, average Z-score for each behavioral period across all trials; Friedman ****p < 0.0001; Dunn’s—Base vs Cue, p = 0.2867; Base vs Interval, p > 0.9999; Base vs Early Open, p > 0.9999; Base vs Late Open, ****p < 0.0001; Base vs Post-escape, ****p < 0.0001. Right, correlation between crossing latency and Cluster 3 activity (n = 147; r = 0.1861, p = 0.0539). Bars show mean ± SEM.
Figure 1—figure supplement 2
Contribution of each cluster to the output during the delayed escape test.
Relative contributions of Clusters 1–3 to the network output across behavioral periods in the delayed escape test.
Figure 1—figure supplement 3
RNN activity of Clusters 2 and 3 under inhibition and escape latency under 180-s delay condition.
(A) Schematic of the inhibition-simulation protocol. In the RNN model, inhibition was applied for 5 s during the delay interval after CS offset and before the door-opening signal. (B) Bar graphs showing crossing latency under inhibition of Cluster 2 or 3 neurons.(Cluster 2 : inhibition, n = 130; control, n = 37; **p = 0.0023; Cluster 3 : inhibition, n = 127; control, n = 37; p = 0.1511 (both Mann–Whitney)). (C) Cross-latency under a prolonged 180-s CS–door-opening interval. Bar graphs show crossing latencies in rats (left; CS, n = 14; No CS, n = 14; p = 0.9720, unpaired t-test) and in the RNN model (right; CS, n = 128; no CS, n = 113; p = 0.3574, unpaired t-test). Bars show mean ± SEM.
Figure 2 with 1 supplement
Comparison of recurrent connectivity in the claustrum and in the RNN simulation.
(A) Mean absolute inter-cluster synaptic weight in the RNN. (B) Schematic representation of whole-cell recordings of ChrimsonR-non-expressing neurons during optogenetic stimulation using LEDs to stimulate ChrimsonR-expressing neurons across the entire optical field. (C) Confocal image showing the expression pattern of the ChrimsonR virus in sagittal claustrum slices. (D) Representative EPSCs evoked by 2 ms LED pulses before and after pharmacological treatments (red bar indicates stimulation). (E) Pooled data of EPSC amplitudes (n = 7) with statistical analysis using the Wilcoxon matched-pairs signed rank test (*p = 0.0156). (F) Schematic depiction of local stimulation using a digital mirror device (DMD). (G) Optical stimulation of each divided part produces variable amplitudes of EPSCs in whole-cell patched neurons expressing no ChrimsonR. Heatmap showing EPSC amplitudes upon stimulation of the designated part of optical field. Please note that when stimulating closer to the recorded neuron, larger amplitude EPSCs were induced. (H) The graph displays EPSC amplitudes and latencies in the region marked with dashed lines on the heatmap shown in G. (I) Left: schematic representation showing an experimental configuration in which brief electrical stimulation produces a persistently enhanced activity in rsCla slices. Right: representative images of GCaMP6f fluorescence changes immediately before and 2 s after stimulation. (J) Heatmap of fluorescence changes for individual puncta in the slice shown in I. (K) Population-averaged fluorescence trace of all puncta in J (n=89; mean ± SEM). (L) Schematic representation showing an experimental configuration in which effects of the blockers for AMPA/NMDA receptors on the persistent response were examined. (M) Calcium imaging before and after an aCSF puff (top panels; n = 12) and an NBQX + D-AP5 puff (bottom panels; n = 16): heatmaps (left) and population-averaged traces (right). (N) RNN analogue: heatmaps (left) and averaged Z-scores (right) for Cluster 1 neurons following brief excitation (top panels; n = 24) and after the addition of recurrent inhibition (bottom panels; n = 24). Shaded areas represent SEM.
Figure 2—figure supplement 1
Slice physiology and pharmacological manipulation of claustral persistent activity.
(A) Membrane-potential decay τ during whole-cell recordings. (n = 11) (B) Top: voltage trace showing persistent spiking after 20 Hz, 200 µA, 1-s stimulation; Bottom: corresponding GCaMP6f fluorescence changes recorded from the same cell. (C) Relationship between the distance from the stimulating electrode to the recorded cell and the persistence ratio (peak fluorescence/fluorescence at 10 s), calculated from the dataset shown in Figure 2J. (n = 89) (D) Schematic of pressure injection of drugs onto whole-cell patch-clamped neurons in horizontal claustral slices. (E) Representative EPSCs recorded under control (aCSF, no pressure injection) and after NBQX + D-AP5 pressure injection with the drug pipette positioned 30 µm or 70 µm from the recorded cell. EPSCs were evoked by electrical stimulation at 0.5 and 3.5 s after injection. (F) Decay kinetics measured from the graphs in Figure 2M, N under NBQX + D-AP5 or aCSF treatment. Left, decay τ after application of NBQX + D-AP5 (n = 16) or aCSF (n = 12) 10 s after electrical stimulation; **p = 0.0012, Mann–Whitney. Right, decay τ with vs without recurrent inhibition in the RNN (n = 24 each); ****p < 0.0001, Mann–Whitney. Bars show mean ± SEM.
Figure 3 with 2 supplements
RNN PCA trajectories.
(A–C) Time-normalized, trial-averaged PCA trajectories for each simulation condition. All trajectories begin at the Start event (black circle), proceed through CS onset and offset (CSon and CSoff; green circle) in the CS-containing and inhibition simulations, then through the Open event (orange circle) in the Open-containing and inhibition simulations, and terminate either at the Cross event (red circle) for crossing conditions or at the End event (black circle) for the CS-only condition. (A) PCA trajectory for the CS + door-opening condition. The mean latency from door opening to crossing was 44.16 ± 0.9583 s. (B) PCA trajectory for the door-opening only condition. (C) PCA trajectory for the CS only condition. (D) Simulated inhibition applied during a 5-s interval between CSoff and door-opening. Top: schematic of the simulated task. Bottom: time-normalized, trial-averaged PCA trajectory. (E) Simulation with a 180-s interval between CSoff and door-opening (no inhibition applied). Top: schematic of the simulated task. Bottom: time-normalized, trial-averaged PCA trajectory. (F) Schematic of the trajectory-combination model: the predicted CS + door-opening trajectory (purple line)—obtained by model of which input are the CS only and door-opening only trajectories—is plotted against the actual CS +door-opening trajectory (orange line). (G) Model-fit comparison for Cluster 1: difference in residual sum of squares (ΔRSS) between the linear regression and multilayer perceptron (MLP) models, normalized to the mean RSS of the linear model. Bar colors denote condition (CS + door-opening = green, inhibition = red, 180-s interval = gray). One-way ANOVA with Holm–Sidak’s multiple comparisons test (CS + door-opening, n = 147; inhibition, n = 119; 180-s interval, n = 128): CS + door-opening vs inhibition, **p = 0.0063; CS + Open vs 180-s interval, ****p < 0.0001. Bars show mean ± SEM. Solid lines represent mean PCA trajectories; shaded areas denote SEM.
Figure 3—figure supplement 1
PCA trajectories of Clusters 2 and 3 under various simulation conditions.
Cluster 2: time-normalized, trial-averaged PCA trajectories for (A) CS + door-opening, (B) door opening only, (C) CS only, (D) overlay of A–C, (E) inhibition simulation, and (F) 180-s delay between CS offset and door-opening. (G–L) Cluster 3 trajectories arranged as in panels A–F. Trial numbers: CS + door-opening, n = 147; door opening only, n = 108; CS only, n = 121; inhibition, n = 119. Solid lines represent mean PCA trajectories; shaded areas denote SEM.
Figure 3—figure supplement 2
Model-fit comparison across clusters.
(A) Residual sum of squares (RSS) from a linear regression model and a multilayer perceptron (MLP) across normalized time bins. Data are taken from the Open event to the Cross event (for the CS only case, from the open event to the mean cross latency of the CS + door-opening condition, n = 147). Green line: RSS of the linear regression model; blue line: RSS of the MLP. (B–D) Difference in residual sum of squares (ΔRSS) between linear regression and MLP fits, normalized by the mean RSS of the linear model, across conditions.; bars, mean ± SEM. (B) Cluster 1 (CS + door-opening, n = 147; inhibition, n = 119; 180-s interval, n = 128): CS + door-opening vs inhibition , **p = 0.0063; CS + door-opening vs 180-s delay, ****p < 0.0001. (C) Cluster 2 (CS + door-opening, n = 147; inhibition, n = 130; 180-s interval, n = 128): CS + door-opening vs inhibition, p = 0.6735; CS + door-opening vs 180-s delay, **p = 0.0021. (D) Cluster 3 (CS + door-opening, n = 147; inhibition, n = 127; 180-s interval, n = 128): CS + door-opening vs inhibition, p = 0.4067; CS + door-opening vs 180-s delay, ****p < 0.0001 (one-way ANOVA with Holm–Šidák).
Figure 4
Biological neural GPFA trajectories and quantification of local geometry.
(A) GPFA trajectories of in vivo single-unit recording data: (left) CS rats, (right) neutral rats (n = 25 neurons, 16 rats). The mean latency from door opening to crossing was 56.363 ± 8.22 s. (B) Validation of GPFA trajectories in CS group recordings. Neurons were ranked in descending order of mean z-scored firing rates (baseline to 5 s post-crossing) and split into odd- and even-indexed subsets. The trajectory from subset 1 (left; n = 13 neurons, 6 rats) served as the reference, while the trajectory from subset 2 (right; n = 12 neurons, 10 rats) was aligned accordingly. For clarity, only the segment from door opening to crossing is displayed. The mean latency from door opening to crossing for subset 1 was 53.306 ± 14.857 s. The mean latency from door opening to crossing for subset 2 was 50.969 ± 7.417 s. (C) Third eigenvalue divided by the sum of all eigenvalues (i.e., the proportion of total three-dimensional variance explained by PC3), computed using sliding windows from the post-open RNN trajectory, comparing the CS + Open (green; n = 20 windows) and Open-only (gray; n = 20 windows) conditions. **** Unpaired t-test, p < 0.0001. (D) λ₃/Σλ of the biological neural trajectory within the sliding window, comparing the CS (green) and Neutral (gray) group pseudopopulations. **** Mann–Whitney test, p < 0.0001. (E) Neuron-wise contribution to the mean λ₃/Σλ, defined using a leave-one-neuron-out analysis of the biological neural trajectory, comparing CS (green; n = 25 neurons) and Neutral (gray; n = 17 neurons). ** Mann–Whitney test, p = 0.0012. Bars show mean ± SEM. Raw data from Figure 4 of Han et al., 2024, Cell Reports, were reanalyzed using GPFA. Elsevier.
Figure 5 with 2 supplements
Decoder accuracy, and Partial Information Decomposition of Cluster 1 RNN.
Neurons under different simulation conditions. (A) Left: schematic of the CS decoder. A classifier was trained to distinguish CS + door-opening trials from door-opening only trials using Cluster 1 neuron activity. Right: CS decoding accuracy (%) across time bins. (B) Left: schematic of the door-opening decoder. A classifier was trained to distinguish CS + door-opening trials from CS only trials using Cluster 1 neuron activity. Right: door-opening decoding accuracy (%) across time bins. (C) Conceptual diagram of Partial Information Decomposition. Mutual information about CS and door-opening is decomposed into CS-unique information, door-opening-unique information, redundancy, and synergy. (D) Schematic of the PID analysis. CS decoding accuracy and door-opening decoding accuracy were computed for each trial type (CS + door-opening, CS only, door-opening only, and None), and used as input variables for neuron-wise PID. PID terms were then averaged across neurons. (E) Synergy information of Cluster 1 neurons. (F) Comparison of synergy between inhibition and no-inhibition simulations. (G) Comparison of synergy between the 180-s interval and the 5-s interval simulations. Green line: 5-s interval; green shaded box indicates CS presentation window. Gray line: 180-s interval. Solid lines represent means; shaded areas indicate SEM. (H) Model fit comparison between Cluster 1 RNN neurons with high synergy and those with low synergy. The difference in residual sum of squares (ΔRSS) between linear regression and multilayer perceptron (MLP) models is shown, normalized to the mean RSS of the linear model. Bar colors indicate condition (high synergy = green, low synergy = gray). (High synergy: n = 147 trials; Low synergy: n = 147.) Mann–Whitney test: high vs low synergy, ****p < 0.0001. Bars show mean ± SEM. In this figure, Open denotes door-opening.
Figure 5—figure supplement 1
Decoding performance and PID analysis across clusters.
(A) Decoding accuracy for classifying CS-present vs CS-absent trials using cluster-specific decoders. (B) Decoding accuracy for classifying Open-present vs Open-absent trials using cluster-specific decoders. (C–E) PID metrics (CS-unique, Open-unique, synergy) for neurons in each cluster. (F–J) Same analyses as in panels A–E under the inhibition condition. (K–O) Same analyses as in panels A–E under a 180-s CS–Open interval. Cluster sizes: Cluster 1, n = 24 neurons; Cluster 2, n = 38 neurons; Cluster 3, n = 38 neurons. Solid lines represent means; shaded areas indicate SEM.
Figure 5—figure supplement 2
PCA trajectories of high- and low-synergy neurons.
Time-normalized, trial-averaged PCA trajectories for (A) high-synergy and (B) low-synergy Cluster 1 RNN neurons. Solid lines represent mean PCA trajectories; shaded areas denote SEM.
Figure 6
Trajectory coding hypothesis and biological neural trajectories.
(A) Cross-temporal decoding of integration-specific information from Cluster 1 RNN neurons. (B) Cross-temporal decoding of integration-specific information from Cluster 1 RNN neurons with high synergy. (C) Cross-temporal decoding of integration-specific information from Cluster 1 RNN neurons with low synergy. (D) Left: heatmaps of raw firing rates for RNN neurons from a representative CS + door-opening trial. Right: heatmaps for biological Cluster 1 non-exploratory neurons in the CS group during the delayed escape task (right). Neurons in the right heatmap are ordered by overall activity. (E) Hypotheses on dynamic coding. Left: attractor hypothesis: multiple or continuous stable converging points exist, and changes in input cause neural dynamics to move from one point to another, thereby encoding varying states. Right: trajectory hypothesis: neural dynamics are not, or are only minimally, related to attractors (which may not exist); instead, changes in input cause changes in trajectory, which encode varying states. (F) Fixed-point analysis of a representative trial in an RNN. Green dots represent stable fixed points (CS period), while black and yellow X-marks denote unstable fixed points (interval and open periods). Panel D (right) is adapted from Fig. S4E of Han et al., 2024, Cell Reports, Elsevier.
Additional files
Download links
A two-part list of links to download the article, or parts of the article, in various formats.
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Computational mechanisms for temporal integration in the anterior claustrum
eLife 15:RP109539.
https://doi.org/10.7554/eLife.109539.3
