How do prefrontal-striatal-hippocampal networks reconfigure during category learning?

A staged handoff, not a uniform increase

The paper asks how three regions interact while a rat learns a category rule, rather than only whether each region carries category information. Nine Long-Evans rats (six male, three female, about three months old) were implanted to record spiking activity and local field potentials simultaneously from the medial prefrontal cortex (mPFC), dorsomedial striatum (DMS), and dorsal hippocampus (dHPC). Across the recorded sessions, 2,391 separable neurons were analyzed: 1,233 in mPFC, 646 in DMS, and 560 in dHPC.

The central result is a change in which pair of regions communicates most strongly through theta oscillations, the 4-10 Hz rhythm the authors focus on. Granger causality, which asks whether past values of one signal improve prediction of another, was used to estimate direction. DMS-mPFC theta causality was strongest in early sessions and ran from DMS to mPFC; mPFC-dHPC causality peaked in middle sessions and ran from mPFC to dHPC; dHPC-DMS causality increased across training and ran from dHPC to DMS. The authors interpret this sequence as early engagement of a frontal-striatal circuit and later engagement of a hippocampal-striatal circuit, with mPFC top-down control as a possible mediator of the transition.

Novel Gabor patches on every trial

Rats categorized black-and-white Gabor patches that varied in spatial frequency and orientation. Two categories were defined as distributions in this two-dimensional stimulus space, aligned so that spatial frequency was the useful dimension and orientation was irrelevant. The distributions were regenerated before each session, so every trial used a novel stimulus and memorizing individual exemplars could not solve the task.

Each session contained 80 trials. A trial began when the rat touched a central star, followed by a 2 s cue period in which the stimulus appeared and had to be touched at least once. Two report keys then appeared for 2 s; the rat chose left or right according to category membership, with the mapping counterbalanced across rats. Correct choices earned a food pellet, and incorrect trials were aborted and repeated without reinforcement. Rats received up to fifteen daily sessions, divided into Early, Middle, and Late phases of five sessions each. Accuracy improved between each phase (linear mixed effects model, t(4.56) = 4.40, p = 0.009) and reaction time became faster (t(4.50) = 3.21, p = 0.027), with no significant male-female difference in accuracy or reaction time.

mPFC selectivity arrives first

Category selectivity here means a reliable firing-rate difference between category A and category B trials. The authors quantified it with a category tuning index (CTI) that compares firing-rate differences for between-category trial pairs with differences for within-category pairs. In all three regions, selectivity appeared early in the cue period and persisted through reward, but its timing and learning trajectory differed.

During the cue period, CTI in DMS and dHPC increased from Early to Middle and Late sessions (DMS: t(623) = 3.52, p = 0.001; dHPC: t(527) = 3.49, p < 0.001). In mPFC, cue-period CTI did not change across the three phases (t(1224) = 1.55, p = 0.121), but it did increase across the five Early sessions (t(452) = 2.90, p = 0.003). Selectivity onset, measured as latency to reach 50% of each neuron's peak normalized CTI, was faster in mPFC than in DMS and dHPC during Early and Middle phases, while onset became faster across training in DMS and dHPC but not mPFC. The authors read this as mPFC category representations developing earlier, with DMS and dHPC representations strengthening after learning.

Choice-period selectivity moved in opposite directions: CTI increased in dHPC (t(454) = 2.69, p = 0.002) and decreased in DMS (t(623) = 2.23, p = 0.007) from Early to later sessions, with no significant change in mPFC. This is a comparison of selectivity magnitude at a defined trial event, not a claim that DMS stops contributing to choices.

Stimulus and choice signals coexist in every region

A linear regression analysis separated stimulus-related from decision-related signals. Regressors for choice, category, and relevant stimulus information each significantly improved model performance in all three regions. Individual neurons were then classified by which regressors mattered: some carried category information, some carried choice information, some carried both, and some neither. The presence of neurons with both stimulus-related and decision-related selectivity, described as mixed selectivity, was therefore not confined to one region.

The authors also aligned activity to the rat's choice touch, using a window from 3 s before to 1 s after the decision and excluding trials in which the choice touch fell within 1 s of reward onset so that feedback-related activity would not be included. In this decision-aligned view, mPFC-dHPC and DMS-dHPC theta causality peaked before the rats' decisions, whereas DMS-mPFC causality was elevated before the decision and declined after it without a clear peak. The authors interpret the first pattern as possibly related to decision-making and the second as engagement in the lead-up to a decision rather than the decision itself.

Direction separates the three theta pathways

Reporting causality strength alone would blur the paper's main distinction, because each pathway had a characteristic direction. DMS-to-mPFC causality was stronger than mPFC-to-DMS throughout training. dHPC-to-DMS causality was stronger than DMS-to-dHPC throughout training. mPFC-dHPC coupling was largely bidirectional, except during Middle sessions, when mPFC-to-dHPC was stronger than dHPC-to-mPFC.

These directional patterns are what support the staged-circuit interpretation. A stronger DMS-to-mPFC signal in early training is consistent with striatal input to prefrontal cortex rather than prefrontal control of striatum at that stage. The middle-session mPFC-to-dHPC dominance is the evidence behind the proposal that top-down mPFC control helps shift the network toward hippocampal-striatal coupling. That proposal is the authors' interpretation of directional Granger estimates, not a direct intervention test of whether mPFC activity causes the shift.

The analysis used only correct trials, a model order of 15, data downsampled to 250 Hz, and detrending; the authors report qualitatively similar results with model orders of 13 and 17. Granger causality describes predictive directionality between signals, so these results should be read as directed statistical influence rather than proven anatomical causation.

What the rat sample can and cannot support

The study's scope is set by its subjects and task. Nine rats learned one rule-based visual categorization task in which spatial frequency was relevant and orientation irrelevant; the authors note they did not run a counterbalanced version with orientation relevant, citing prior work that rats learn both at the same rate and to equivalent levels. Some rats did not complete all fifteen sessions because of lost recording head stages or cable-tether distractions, so the Late phase averages fewer sessions for those animals.

The recording sites also define the network being described: mPFC tetrodes spanned prelimbic and rostral anterior cingulate regions, DMS tetrodes concentrated in dorsal DMS, and dHPC tetrodes concentrated in CA1 with a few extending into dentate gyrus. Firing-rate activity was similar across these subregions and between male and female rats, but the paper does not test whether the theta reconfiguration differs by subregion or sex. The conclusion is therefore about this mPFC-DMS-dHPC network during this task in rats, not a general account of categorization across species or tasks.

The middle phase is the pivot to watch

The most decision-relevant comparison in this paper is not early versus late alone but the middle training phase, where mPFC-to-dHPC theta causality peaks and is directionally dominant. That is the point at which the authors place the handoff from a frontal-striatal to a hippocampal-striatal circuit. Early training is marked by DMS-to-mPFC coupling and early mPFC category selectivity; late training is marked by rising dHPC-to-DMS coupling and stronger dHPC and DMS category selectivity during the cue period.

For readers evaluating the paper's mechanistic claim, the priority is to check whether the directional middle-phase mPFC-to-dHPC signal is treated as a correlate of the transition or as its cause. The paper's own wording is that the transition is potentially mediated by top-down mPFC control, which keeps the causal step open. The strongest supported statement is that theta coupling among these three regions is reconfigured in a staged, directional way as rats learn, with the mPFC-dHPC pathway peaking in the middle of training rather than at the end.

Sources used in this answer