A persistent prefrontal reference frame across time and task rules

The stability of neuronal representations underlying consistent behavior over time is a central question in neuroscience. Two contrasting views exist: representational drift, where neuronal responses to identical stimuli gradually change, and stable representations, where neuronal activity remains consistent. Representational drift has been observed in sensory and associational cortices during various tasks, leading to hypotheses about stable readout from population codes, high-dimensional manifolds, or self-correcting assemblies. In contrast, stable neuronal responses have been reported in other brain areas, such as the sensory cortex and hippocampus, often associated with engram formation. The prelimbic region of the mPFC plays a crucial role in spatial working memory and decision-making, but its activity dynamics over extended periods remain largely unexplored. This study aimed to investigate whether prefrontal activity follows a stable or drifting encoding regime over weeks using an olfaction-guided two-choice task in mice, monitoring prelimbic and anterior cingulate area neuronal activity via 1-photon calcium imaging.

Existing research presents conflicting evidence regarding the temporal stability of neuronal representations. Studies have demonstrated representational drift in sensory and associational cortices during perceptual and navigational tasks, suggesting that stable behavior might emerge from population-level dynamics rather than individual neuron stability. These studies proposed mechanisms like stable readout from drifting population codes, high-dimensional population codes, or self-correcting assemblies. Conversely, other studies have shown stable neuronal responses in sensory areas and the hippocampus, particularly in the context of memory engrams. While multiple studies have examined prefrontal subregions during single-day memory and decision-making tasks, the long-term dynamics of mPFC activity remain largely uncharted, raising questions about whether prefrontal encoding is stable or dynamic over extended timescales.

Researchers used 1-photon calcium imaging in Thy1-GCaMP6f mice, a transgenic approach that allows longitudinal imaging of the same deep-layer cortical neurons over extended periods. Retrograde tracing confirmed the presence of both intratelencephalic and pyramidal tract neurons in the recorded population. Mice performed an olfactory-guided two-choice task in a figure-M maze, associating two odors (vanilla or coconut) with reward sites on the left and right arms. Lens implantation in the mPFC resulted in moderate gliosis but didn't affect task learning or electrophysiological profiles of mPFC neurons. Two cohorts of mice were used: cohort 1 began imaging after reaching >70% task accuracy; cohort 2 was imaged during initial task exposure (before learning). Longitudinal registration allowed tracking of the same neurons across days. Behavioral data was recorded using a camera, and calcium signals were analyzed using CalmAn to identify neurons and extract calcium traces. Spatial positions were extracted via EthoVision XT and linearized for analysis. Several analyses were performed, including correlation analysis of spatial tuning functions, calculation of mean activity and side index, quantification of spatial information, and decoder analyses to predict spatial position and trial outcome. Generalized linear modeling (GLM) was used to investigate how different behavioral parameters (position, speed, goal location) influenced neuronal activity. In a subset of mice, the reward rule was reversed (new rule) and then restored (restored rule) to examine the impact on neuronal responses. Additionally, mice were tested in a visually modified arena (novel arena) to assess context generalization. Electrophysiological recordings from a separate cohort of mice were included for comparison.

The study's key findings demonstrate that task-related activity in the mPFC remains remarkably stable over weeks. A substantial proportion (47.8 ± 2.5%) of neurons were repeatedly active throughout the 3-week recording period. These repeatedly active neurons showed stable mean calcium activity and side indices (preference for left or right trials) across days. Behavioral choice could be reliably decoded from neuronal activity across all days using models trained on day 1. Trajectory-specific spatial tuning functions were also highly stable across days, showing a slow representational drift. Linearized position was the most significant predictor of neuronal activity, followed by speed and goal location, and this relationship remained stable across days. The stable representation of task space emerged during learning and was not significantly affected by pauses in task execution. This stable representation also generalized across contexts (familiar vs. novel arena) and persisted across changes in the reward rule, even though some drift was observed after rule inversion. In the before-learning group, spatial consistency of maps between odd and even runs was significantly lower than in the learned group. Consistent with the temporal stability, only a relatively small number of neurons (around 10-15 neurons) were sufficient to achieve high decoding accuracy within a day and across the whole period.

This study's findings provide compelling evidence for the existence of a stable core ensemble of mPFC neurons that maintain a persistent representation of task-relevant space over weeks. This stable reference frame seems to emerge during task learning and is robust against long pauses in task exposure and even changes in the task rules. The high decoding accuracy from a relatively small number of neurons suggests an efficient neural mechanism for driving stereotyped motor responses. The stable prefrontal representation contrasts with observations of significant representational drift in other brain areas like the posterior parietal cortex and hippocampus, highlighting potential differences in the encoding mechanisms underlying different types of memory. The data supports the theoretical concept that reference frames are fundamental for location-feature mapping in neocortical networks. Future research should investigate whether this stability generalizes to non-spatial tasks and other types of rule changes. The study's limitations, including the potential impact of GRIN lens implantation on mPFC circuitry and the focus on deep-layer neurons, also need to be addressed in subsequent research.

This research demonstrates the existence of a remarkably stable neural representation of task space in the mPFC of mice over extended time periods, even across changes in task rules and environmental contexts. This stable representation, likely established during learning, forms a persistent reference frame underpinning consistent behavioral performance. Future studies could explore the role of this stable reference frame in non-spatial tasks and examine the interplay of mPFC and other brain areas in generating stable neural representations.

The study acknowledges limitations stemming from the use of GRIN lens implantation, which may introduce changes in neuronal connectivity and local network dynamics, potentially influencing the findings. The analysis focused primarily on repeatedly active neurons, potentially neglecting potentially important contributions from neurons active on only some of the recording days. The focus on deep layer mPFC neurons also limits the generalization of the findings to other neuronal populations.