Striatal cell-type specific encoding stability and reorganization for goal-directed behavior and habit
Year: 2026
Type: software
Abstract: Goal-directed behavior is guided by learned action-outcome associations, allowing actions to adapt when outcomes change. With repetition, behavioral control shifts toward habit, enabling more automatic action execution, but reducing flexibility to outcome changes. Here, we used longitudinal, cellular-resolution calcium imaging, chemogenetics, instrumental conditioning, and the outcome devaluation test to ask how dorsomedial striatum D1+ and D2/A2A+ neurons contribute to the formation of goal-directed behavior and habit. DMS D1+ neurons stably encode instrumental actions, carry information about behavioral engagement and organization, and develop and maintain action-reward representational similarity and conjunctive activation with learning. Correspondingly, these neurons are critical for action-outcome learning and goal-directed decision making. In contrast, one ensemble of DMS A2A+ neurons transiently encodes actions during early action-outcome learning and another slowly starts to encode actions as habits form. Similarly, DMS A2A+ neurons transiently mediate action-outcome learning and become dispensable for goal-directed decision making. Thus, action-outcome associations are learned through DMS D1+ and A2A+ neuronal activity but only DMS D1+ neurons maintain support of adaptive, goal-directed decision making. Habit formation is associated with reorganization of A2A+ neuronal encoding. These data reveal a cell-type-specific dissociation between stable v. reorganizing striatal action representations paralleled by persistent v. transient contributions to goal-directed control. Methods Miniscope system. Calcium imaging videos were recorded using an open-source, head-mounted UCLA Miniscope imaging device, which interfaces with the UCLA open-source Miniscope DAQ hardware and software (http://miniscope.org/) to capture the neuronal calcium dynamics with corresponding frame timestamps. A webcam (Logitech, San Jose, CA) recorded the Med Associates interface LEDs, which display light pulses coincident with task events (lever presses, food-port entry, pellet delivery). This video was simultaneously saved on the same computer alongside the calcium imaging video via the miniscope software with synchronized frame timestamps. Calcium trace extraction and spike inference. Videos were initially processed using custom Python code124 to concatenate all the videos from a session into one tiff stack, spatially downsample by a factor of 2, and temporally downsample to an effective frame rate of ~7.5 frames/s to accelerate image processing. The Python implementation of the Calcium Imaging Analysis package, ‘CaImAn’125, was then used to process the videos. We applied the nonrigid motion correction algorithm to remove tissue motion artifacts and then applied constrained non-negative matrix factorization for microendoscopic data to extract individual spatial contours and demix temporal fluorescence traces for each detected neuron. Individual fluorescent signals were deconvolved to estimate discrete spikes for each neuron126. The deconvolved traces, the estimation of discrete spikes that generate the calcium traces, were derived from the denoised fluorescence traces using CaImAn’s ‘deconvolveCa’ function. Neurons were coregistered across sessions using the open source package CellReg127. Activity for each neuron was Z-scored to the whole session to allow comparison of baseline and phasic activity across cells and subjects with different levels of GCaMP expression and/or lens proximity. Behavior extraction. All behavior-event videos from a session were concatenated and downsampled as above. Using custom MATLAB code, the brightness of each LED on the interface was measured for each frame. A task event was considered to have occurred if the LED brightness was >2x the background brightness threshold (determined individually per recording), thus binarizing task events in each frame of the video. The extracted frame-by-frame event information was then interpolated to the timestamps of the calcium imaging video. Cell classification. We used an analysis based on the area under the receiver operating characteristic curve to identify neurons tuned to key behavioral events: ‘action initiation’, first press in session, first press after earned reward, and first press after a non-reinforced food-port check; ‘action termination’, last press before earned reward collection and last press before a non-reinforced food-port check; ‘food-port check’, entries into the food delivery port, for those in which there was no reward present (a.k.a. non-reinforced entries); ‘reward consumption’, consumption of the earned food pellet occurring at the time of exit from the food-delivery port. The auROC provides a non-parametric measure of the separability between two distributions of data. To increase the temporal precision in linking neural activity to behavior, we used the deconvolved traces to distinguish between a neuronal activity preceding v. in response to an event. For each neuron, we defined two distributions of activity: one from the activity aligned to the ±2.9 s window around repeated occurrences of a specific behavioral event and one from equated baseline periods in which subjects were not engaged in the task. We compared these two distributions of neuronal activity using the MATLAB perfcurve function to generate receiver operating characteristic curves of the true positive rate and the false positive rate and measured the area under the curve (auROC). To determine if a neuron significantly encoded an event, we generated a null distribution of auROC values by circularly shifting the activity trace for each neuron and recalculating the auROC 1000 times. Neurons with true auROC values >97.5% of the auROCs in the null distribution were considered significantly excited by an event and those with true auROC values <2.5% of the auROCs in the null distribution were considered significantly suppressed. The 2.9-s window before and after the event of interest was selected to ensure that the cell classification was not biased toward neurons modulated exclusively before or after the event, thereby capturing the range of event-related dynamics, including both anticipatory and reactive responses. To validate the robustness of the classification and characterize the temporal dynamics of these responses, the time at which neurons were significantly modulated within the ±2.9 s window, the auROC was further calculated at each individual frame within the ±2.9 s window around each event. To test whether action-initiation neuron identity is preserved across sessions above chance, we ran a 10000-fold Monte Carlo permutation simulation. For each individual subject, we simulated the null distribution of overlapping neurons based on that subject's specific daily proportions of neurons modulated by action initiation and total tracked neurons. To test if daily neuronal overlap between action initiation and reward was above chance, we ran a Monte Carlo permutation simulation as described above. In this case, chance overlap was simulated using the single-session proportions of action- and reward-modulated neurons. In all simulations, differences in total neuron yield across animals was controlled by normalizing the observed overlap and the simulated chance overlap as a percentage of the total tracked neuron pool for that subject. Behavior decoding. We used support vector machines (SVM) to decode key behavioral events (action initiation, action termination, food-port check, and reward consumption) using the activity of specific neuron populations. We used the temporal fluorescence trace because it is a continuous signal that can reflect recent neural activity that contains a more integrated signal important for changes between behaviors. To predict behavior events prior to onset, observation data used to train and test the classifier consisted of the 0.4-s epoch preceding each of the 4 distinct event labels. Since there were large variations in the number of each type of behavioral event, we controlled for the number of each event label. First, for the behavior event labels that were overrepresented (usually food-port entries), we found the median number for all event labels, and limited those event labels to that value. Then, for the behavior event labels that were underrepresented, we used the synthetic minority oversampling technique (SMOTE)128 prior to the test/train split to generate synthetic activity samples to match the event label with the highest number of events. For each session, the classifier was trained (MATLAB function fitcecoc) on ¾ of the data, randomly selected, to generate a model and behavior was predicted (MATLAB function predict) with the generated model on the remaining data. Decoding performance was averaged across 20 different random data splits. Decoding performance was assessed by computing the F-score, a standard measure of binary classification accuracy that weighs recall (proportion of actual positives that are correctly predicted) against precision (the proportion of positive predictions that are actually positive) of the classifier. It is calculated as the harmonic mean of precision and recall. We performed a 1000-fold shuffling procedure to confirm that the average decoding of randomized data is at the theoretical chance level of ¼ correct. To examine whether the ongoing lever-press rate, switch rate, and interpress-interval fluctuation could be explained by the neuronal activity, we fit a linear model (MATLAB function: fitlm) using the temporal fluorescence traces as predictors and behavior as the response. The discrete lever presses were transformed into an ongoing behavior rate by convolving the lever presses with a half-gaussian-post event kernel width of 30 s and a standard deviation of the distribution of 6 (see Extended Data Figure 1-5f for example). Switch points were identified as a shift from a lever press to a foo
Author: Malvaez, Melissa
Language: English
Subfield: Cellular and Molecular Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7203665892