A persistent prefrontal reference frame across time and task rules

The study addresses how stable behavior over long timescales is supported by neuronal activity: do cortical representations drift over time or remain stable? Prior findings in sensory, associative cortices and hippocampus have shown representational drift to identical stimuli and during navigation, raising questions about how stable behavior arises from changing codes. Alternative frameworks propose stable readouts from drifting populations, high-dimensional stable manifolds, or self-correcting assemblies. Conversely, temporally stable neuronal responses (engram-like) have been reported in certain systems. The prelimbic mPFC supports spatial working memory and decision-making, yet its dynamics across weeks are unclear. The authors hypothesize that a stable, task-specific reference frame may emerge in mPFC during learning and persist across time, pauses, and changing task rules, enabling consistent behavior.

Evidence for representational drift has been reported in visual cortex, primary olfactory cortex, associational (parietal) cortex, and hippocampal CA1 during perceptual and navigational tasks. Mechanistic solutions for stable behavior despite drift include stable readouts from non-random drift, stable population-level manifolds, and self-healing assembly codes. Conversely, temporally stable responses have been observed in sensory cortex during deprivation, dentate gyrus, and neocortical engrams for fear memory. The mPFC is necessary across phases of working memory and shows spatially tuned activity during cognitive tasks in single-day studies. Stable responses over weeks have been noted in orbitofrontal cortex, while pronounced drift was reported in posterior parietal cortex. Theoretical work suggests reference frames could underlie learning and memory mapping in neocortex. This work situates mPFC within these competing views by testing long-term stability of trajectory-specific tuning.

Experimental model and task: Adult Thy1-GCaMP6f mice were implanted with a GRIN lens over mPFC (prelimbic/anterior cingulate) for head-mounted 1-photon calcium imaging of predominantly layer 5 pyramidal neurons. Mice performed an olfaction-guided two-choice spatial task on an M-shaped maze: vanilla or coconut odor sampled in the center stem predicted reward at the left or right arm. In each imaging session (~15 min), mice completed ~40 trials. Cohorts and longitudinal design: Cohort 1 (n=8) was imaged after reaching criterion performance (≥70% correct) across 24 days (13 recording days). A subset (n=6) underwent rule reversal (odor–reward mapping inversion) and later restoration of the original rule (up to 68 days after initial learning). Cohort 2 (n=5 before learning; n=4 after learning) was imaged during initial exposure before learning (∼47% correct) and after learning. Additional manipulations compared daily continuous exposure vs 7-day pauses without task exposure, and familiar vs visually modified (novel) arenas. Imaging and signal extraction: Imaging used Inscopix nVoke/nVista microscopes at 20 Hz. Motion correction and source extraction employed CalmAn/CNMF-E; components were curated via a custom GUI. Calcium traces were baseline-corrected and z-scored; significant transients were defined (>3 SD, ≥0.2 s). Longitudinal cell registration across days used CellReg, aligning projection images and assigning same-cell probabilities; analysis focused primarily on repeatedly active neurons detected on each day. Behavioral/positional processing: Animal trajectories were tracked and linearized to a 1D skeleton from center entry to side-arm reward. Spatial tuning functions were computed (20 bins), normalized by occupancy, separately for left and right trajectories. A side index quantified trial-direction preference. Spatial information (SI) was computed from transient rates and occupancy. Decoding analyses: Trial outcome decoding used logistic regression (also SVC and an ANN as controls), trained on day 1 center-arm activity and tested across days. Spatial position decoding used LinearSVR trained on day 1 and tested across days; performance was mean absolute error of linearized position. Single-neuron decoding and neuron-count subsampling analyses quantified contributions to decoding. GLM encoding: A generalized linear model with categorical predictors assessed contributions of linearized position (8 bins plus ends), speed (quartiles), and goal side to calcium activity, quantifying explained variance (r^2), decreases upon time-shifting predictors, and per-cell significance. Controls: Electrophysiological recordings in Ai32 mice performing the same task assessed comparability to calcium imaging. Histology assessed gliosis after GRIN implantation. Statistical analyses used paired/unpaired tests and repeated-measures ANOVAs with corrections as appropriate.

• Stable behavior and cell ensemble across weeks: Behavioral performance remained stable across weeks (F=0.93, p=0.527). Longitudinal registration showed 82.9±0.5% of L5 neurons registered across consecutive days, and 47.8±2.5% were active on every imaging day over 24 days (mean 63±16 neurons/mouse, n=8), indicating a repeatedly active ensemble. • Stable choice-related signals: Mean center-arm calcium activity patterns were stable (Spearman r first vs last 0.635, p=2×10^-16). Side index remained highly correlated across days (r=0.813, p=7×10^-12; n=133 neurons). A decoder trained on day 1 accurately predicted left/right choice up to day 24 with no accuracy-time interaction (F=1.38, p=0.276) and remained significant vs shuffled (F=177.17, p=3×10^-40). Single-neuron within-day decoding accuracies were correlated over weeks (r=0.573, p=6×10^-10; n=133). As few as ~10 randomly selected cells achieved ~90% within-day choice decoding; ~15 cells gave ~80% accuracy across 24 days. • Persistent trajectory-specific spatial tuning: Task-active L5 neurons tiled the maze; tuning was more similar for same-direction trajectories than opposite. Spatial information remained correlated across days (left r≈0.654; right r≈0.648; n=502 neurons). Trajectory-specific tuning correlations decayed slowly at ~−0.006/day yet stayed well above shuffled controls (≈0.5 average correlation by day 24). Position decoders trained on day 1 generalized across weeks with significant accuracy vs shuffled. • Encoding dominated by position: On day 1, the full GLM (position, speed, goal) explained 13.2±1.2% variance. Position alone captured the most variance; time-shifting position induced the largest drop in explained variance. Proportions of cells significantly modulated: position 67.6±3.8%, speed 41.6±5.8%, goal 26.2±2.9%. These contributions remained stable across days. • Learning stabilizes the reference frame: Before learning (~47% correct), trajectory-specific consistency within day and stability across 4 days were lower than after learning (consistency: t=−5.93, p=3×10^-8; stability: t=−3.98, p=0.001). Decoders trained on day 1 had larger spatial errors across 4 days before learning (t=2.23, p=0.047). • Drift depends on time, not experience: Seven-day pauses in task exposure did not reduce correlation of trajectory-specific tuning compared with continuous daily exposure (t=−2.01, p=0.084). Position and choice decoding generalization from the first to the seventh day were comparable between pause and continuous conditions (p>0.39 and p=0.600, respectively). • Context generalization: Tuning stability was similar between familiar–familiar and familiar–novel arenas (t=0.94, p=0.379). Position decoding from familiar-trained models remained accurate in the novel arena with mildly increased error (p=0.039); choice decoding accuracy was comparable (p=0.138). • Robustness to rule reversals: After inverting odor–reward associations and later restoring the original rule (up to 68 days after original learning), trajectory-specific tuning retained significant correlation to the original state despite slow decay. Position and choice decoders trained on the original rule generalized to new and restored rules with significant accuracy vs shuffled.

The findings show that a subset of mPFC layer 5 neurons forms a core ensemble that provides a temporally stable, trajectory-specific reference frame for task space over weeks. This stability persists despite pauses in task exposure, changes in visual context, and even reversal and restoration of cue–reward contingencies, indicating only slow representational drift that is driven primarily by time rather than repeated task experience. The dominance of spatial position in GLM encoding and the successful decoding of both position and choice across weeks suggest that downstream circuits could reliably read out behaviorally relevant information from relatively few, consistently active neurons. The results reconcile stable behavior with neural dynamics by revealing a stable scaffold within mPFC, contrasting with stronger drift in hippocampus and posterior parietal cortex. Such a reference frame may facilitate consistent, learned motor responses and contribute to long-term memory-guided performance, potentially arising from robust recurrent weight changes within mPFC during learning or from stable inputs from other regions.

This work identifies a temporally stable, trajectory-specific reference frame in mouse mPFC that emerges during learning and persists across weeks, pauses, context changes, and rule reversals, with only mild time-dependent drift. Position is the principal driver of prefrontal activity during the task, and a small, repeatedly active core ensemble suffices to decode choice reliably across time. These results provide a neural substrate for consistent behavior over extended periods. Future studies should test the causal necessity of the stable ensemble for task execution, determine whether similar reference frames arise in non-spatial or extra-dimensional rule-shift tasks, examine contributions of superficial layers, and dissect circuit mechanisms transforming dynamic hippocampal codes into stable prefrontal representations.

GRIN lens implantation produces a cortical lesion; although behavior and electrophysiological profiles were comparable to controls, altered connectivity cannot be fully excluded. Day-to-day detection of active cells may reflect biological variability or technical factors (e.g., focus shifts); analyses emphasized repeatedly active cells, potentially underestimating changes in unregistered neurons. Thy1-GCaMP6f expression largely restricted recordings to deep layers; stability in superficial mPFC layers remains to be tested. The necessity of the identified stable reference frame for behavior was not directly probed.