Fig 1.
Schematic showing the steps in analyzing decision-making dynamics within the cortical-basal ganglia-thalamic (CBGT) network.
(A) CBGT network, with connections color-coded as follows: classical direct pathway in green, classical indirect pathway in blue, and pallidostriatal pathway in gold. Arrows ending in dots indicate postsynaptic sites of inhibitory connections, while arrows ending in triangles indicate excitatory connections. Cx, cortical neurons; CxI, inhibitory cortical interneurons; FSI, fast spiking interneurons; dSPN, direct spiny projection neurons; iSPN, indirect spiny projection neurons; GPe, external globus pallidus; GPe_P, prototypical neurons; GPe_A, arkypallidal neurons; GPi, internal globus pallidus; STN, subthalamic nucleus; Th, thalamus. (B) Simulated behavioral data, including firing rates, choices, and decision times, generated by a spiking model for the CBGT network. (C) A chain of state transitions derived from the processed data. (D) States grouped into zones, with associated transition probabilities. (E) Levels of activity within zones mapped to DDM parameters through control ensemble analysis, showing how dynamic decision policies emerge from flow of network activity between zones.
Fig 2.
Example of firing rate time course and binarization in a two-choice task.
Each red (blue) trace corresponds to activity in the left (right) action channel in a CBGT region, and the grey traces correspond to the populations (i.e., CxI and FSI) common to both action channels. Pink regions represent the decision-making phase, occurring before the thalamus (Th) of one of the action channels reaches the decision threshold of 30 Hz (dashed black line in Th panel). Grey regions represent the consolidation phase, where partial cortical input to the selected channel is sustained [55,56]. The unshaded regions represent the inter-trial interval. In each panel the right -axis corresponds to binarized firing rates (dots at 0 or 1), where the horizontal green line indicates the binarization threshold (see Fig 10 for details on how thresholds were determined).
Fig 3.
CLAW (Circuit Logic Assessed via Walks) diagram for CBGT network dynamics.
(A) CLAW diagram. Bold numbers in boxes indicate the network states, and the transition probability from a current state to a subsequent state is indicated by the number near the arrow pointing from the current state. The numbers below certain states (e.g., “0.82 end” below state 663) represent the probability that if these states are reached, then they are the final states prior to decisions. (B) Overall decision time distribution across 300 networks, categorized into three equal-count tertiles defining fast (left), intermediate (middle), and slow (right) networks, demarcated by vertical dashed black lines. The coloring of each CLAW state in panel A corresponds to the mean DT for all trials that visit this state, following the same color-coding scheme as in panel B. (C) Details of the states. A complete explanation of the full set of state properties, including those related to the right choice, is presented in the Supporting Information S1 Table. From left to right, after state labels: binarized firing rates of dSPN, iSPN, GPi, GPeP, and Th for left (-L) and right (-R) channels; probability of activation (binarized firing rate = 1) for STN and GPeA for left and right channels; mean DT over the trials that visit each state; the ratio of trials that chose left/right for those that visit each state; Kullback–Leibler (KL) divergence between left and right trials’ DT distributions. The grey rows correspond to the initial states that occur early in each trial and never lead directly to a decision, and the green and blue rows correspond to outer CLAW and inner CLAW states, respectively. The states in the lower half of the CLAW are not shown; these are symmetric – up to the swap of certain L and R channel binary values – with the states that lie in corresponding positions in the upper half.
Fig 4.
(A) Zone CLAW diagram. Zone I contains the pre-stimulated state 60 as well as states 61, 62, and 63, where only the binarized firing rates of thalamic populations may cross threshold. Zones II and V correspond to the initial and second deliberation phases within the inner CLAW, respectively. The left and right arms of the outer CLAW are represented by zones III and IV, respectively, each of which has a low probability of transitioning into zone VI, which consists of a single neutral state 783. The transition probability from one zone to another is indicated by the number near the arrow pointing from the source zone. The loop arrows represent the probability of staying in a zone or reaching the decision threshold from that zone. (B) Details of the zones. From left to right, after zone labels: probability of activation for dSPN, iSPN, GPi, GPeP, Th, GPeA, and STN for left (-L) and right (-R) channels, when networks are in each zone. The grey rows correspond to the launching zone, and the green and blue rows correspond to the zones containing the outer CLAW and inner CLAW states, respectively.
Fig 5.
Temporal activation of CBGT pathways for different decision trajectories along the CLAW.
The direct, indirect, and pallidostriatal pathways across both left and right channels are enclosed in boxes, with their activation strength colored in blue, green, and gold, respectively (color bars at bottom). Darker shades indicate stronger activation within each pathway. Nuclei that play a key role for some phase transitions are indicated near the corresponding arrows. (A) Launching region Left commitment. (B) Launching region
Right commitment. (C) Launching region
Initial deliberation
Launching region. (D) Launching region
Initial deliberation
Second deliberation. The Roman numeral on the upper right corner of each pattern corresponds to the zone labels in Fig 4.
Fig 6.
Canonical loading matrices obtained by applying the CCA.
(A) Loadings of , corresponding to CBGT firing rates. The subscript ‘sum’ refers to the sum of rates in a region across both action channels, while the subscript ‘diff’ refers to the rate in the left channel minus that in the right channel. (B) Loading of
, corresponding to DDM parameters. These three components are referred to as choice, responsiveness, and pliancy control ensembles, respectively.
Fig 7.
Time evolution of control ensemble drives.
The percentage changes in each drive are averaged across all trials within each of four decision groups—(A) fast left, (B) slow left, (C) fast right, and (D) slow right, from the stimulus onset time to the decision time (set to be 0). Since each trial had a different decision time, we performed the averaging by aligning trials on their decision times, with the averages represented by dots connected by lines (light dashed for fast and dark solid for slow). Shadowed areas represent the standard deviation for each group. The vertical dotted lines with colors matching the plotted curves mark the timing at which each control ensemble’s drive shows a notable response. For fast and slow trials to the same choice, each inset at the right aligns the drives, zoomed in on the time range with pronounced changes (~75 ms), on the same scale for comparison.
Fig 8.
Modulation of decision policy parameters by CBGT activity along the CLAW.
The arrows indicate variations of the individual DDM parameters (drift rate , panel A; boundary height
, panel B; onset time
, panel C) associated with transitions between zones of the CLAW, computed based on the activity changes from each zone to the one immediately downstream from it. The magnitude of each variation is indicated by the coloring (see color bar) and the number at the start of each arrow, representing the percentage change in parameter values.
Fig 9.
Mapping changes in CBGT activity across CLAW zones to variations in DDM parameters.
(A) Firing rate changes during zone transitions are mapped to the DDM space through control ensembles. The vectors and
represent the difference in the firing rate measurements from zone I to zone III (blue) and zone IV (red), respectively. These changes update the static DDM parameter set from
to
or
. (B) The corresponding DDM behavior that fits the decision outcomes, assuming that CBGT activity stays in one zone (here, either zone I, zone III, or zone IV) over the entire decision process. (C) Prediction about the actual dynamic DDM behavior (magenta) as trials travel from zone I to zone III, the DT distribution of which is constrained by those corresponding to the static cases.
Fig 10.
Firing rate histograms for all trials up to decision times.
In each panel the vertical green line represents the binarization threshold for the firing rates. In unimodal histograms, the binarization threshold for GPeP and GPi was defined as the firing rate at 10% of the counts, with values above this threshold binarized to 1 and values below it binarized to 0, while for GPeA, dSPN, and iSPN the threshold was set at 90% of the counts. In bimodal histograms, the midpoint between the two peaks was defined as the binarization threshold.
Fig 11.
Procedures for processing firing rate data.
(A) Populations of interest are specified, with the simulated temporal activity. (B) Firing rates within each population in each time bin (10 ms) are binarized into 1 (above threshold) or 0 (below threshold). (C) Transition probabilities are computed using the full set of binarized activity sequences.