Dendrite-targeting OLM interneurons regulate the formation of learning-related CA1 place cell representations

Abstract

Spatial learning depends on the rapid formation of hippocampal CA1 place cell representations through behavioral timescale synaptic plasticity (BTSP)^1–7. BTSP is driven by dendritic Ca²⁺ plateau potentials in the distal apical dendrites of CA1 pyramidal neurons and is thought to arise from the interaction of an excitatory target signal from entorhinal cortex layer 3 (EC3) with inhibitory feedback reflecting the current CA1 population state8. However, the cellular source of this feedback remains unknown. To identify this circuit element, we combined in vivo two-photon calcium imaging with bidirectional optogenetic manipulation to examine the role of dendrite-targeting oriens lacunosum-moleculare (OLM) interneurons^9,10. We found that axonal and somatic activity of OLM interneurons increased with learning and was spatially biased toward behaviorally salient locations, closely matching the evolving CA1 population representation and the environment-specific EC3 target signal3. Causal manipulations revealed that optogenetic silencing of Chrna2α-positive OLM interneurons, a genetically defined OLM subset, late in learning increased dendritic plateaus and promoted place cell formation at stimulated locations, whereas activation of the same population early in learning suppressed plateau initiation and place field formation. Together, these findings identify OLM interneurons as the key inhibitory feedback element that dynamically regulates BTSP and stabilizes hippocampal representations during learning.

OLM axonal activity increases with spatial learning.

Spatial learning engages behavioral timescale synaptic plasticity (BTSP), which is triggered by Ca²⁺ plateau potentials (“plateaus”) in the distal apical tuft dendrites of CA1 pyramidal neurons^11–13 and generates CA1 place cell representations^1–7. In previous work, we showed that entorhinal cortex layer 3 (EC3) provides a stable, target-like excitatory input enhanced at behaviorally salient locations, which promotes dendritic plateau initiation and drives BTSP, shaping CA1 population activity toward an overrepresentation of salient locations required for successful spatial learning^14–18. We hypothesize that plateau initiation reflects a mismatch between the EC3-derived excitatory target signal and an inhibitory feedback signal that reports the current state of the CA1 population, leading to plateau activity when excitatory drive exceeds feedback inhibition (Fig. 1a). Oriens–lacunosum moleculare (OLM) interneurons are well positioned to provide such feedback, as they receive excitatory input from CA1 pyramidal neurons and selectively target the tuft dendritic compartment^19–21.

Figure 1: OLM population activity during learning.

a, Preliminary BTSP working model. b, Top: experimental setup, in which mice learn the location of a water reward. Bottom: schematic of the CA1 microcircuit and imaging plane (dashed line) c, Left: representative, two-photon, time-averaged images of axons from CA1 SST-OLMs expressing GCaMP6f and tdTomato. Right: example traces of distance, licking, velocity, and GCaMP6f Δf/f (mean of the field). d, Top: Task design of day 0 (the last day of habituation), and day 1 (the first day of learning). Middle: mean running profile across space. Bottom: mean lick probability across space. e, Representative day 1 running, and Δf/f corrected for dwell time (•s). Top: along laps. Bottom: spatial profiles. f, Days 0 and 1 mean z-scored Δf/f•s (day 0: n=15, day 1: n=15). Top: along the percentage of laps run. Bottom: spatial profiles. The peak-zone is 90 cm beginning at the spatial bin with reward, the away-zone is the same but 90 cm away. g, Days 0 and 1 mean z-scored Δf/f•s within peak-zone and away-zone (paired two-tailed t-test, day 1: P=2.8•10⁻³). h, Mean z-scored Δf/f•s vs. lap quintiles (two-way RM ANOVA with post-hoc FDR correction, group × quintile: P=9.7•10⁻³, day 0: P=1.9•10⁻², day 1: P<1•10⁻⁴). i, Quintiles 1 and 5 mean z-scored Δf/f•s vs. position (bin=18 cm, two-way RM ANOVA with post-hoc FDR correction, quintile 1/group × bin: P=7.0•10⁻⁴, quintile 5/group × bin: P<1•10⁻⁴, quintile 5/bin 5: P<1•10⁻⁴). In d-i, n=15 mice. In g, dots represent mice. Data are shown as mean ± s.e.m.

To test whether OLM interneurons constitute the inhibitory feedback signal, we performed two-photon Ca²⁺ imaging of OLM axons in dorsal CA1 of somatostatin (SST)-IRES-Cre mice²² expressing the Ca²⁺ indicator GCaMP6f and tdTomato while animals ran on a linear treadmill (with three somatosensory cues) and learned a fixed reward location (Fig. 1b). OLM axonal Ca²⁺ signals were recorded in stratum lacunosum-moleculare (SLM), the CA1 sublayer where entorhinal cortex layer 3 (EC3) inputs converge with OLM-mediated inhibition and averaged within each field of view to yield a population-level measure of OLM output alongside behavioral variables (Fig. 1c, Extended Data Fig. 1a). As in prior work^3,14, animals restricted their licking to the region surrounding the reward and reduced their running speed when approaching the reward delivery site during the day 1 session (belt with tactile cues, fixed reward) when comparing early (quintile 1) to later laps (quintile 5; Fig. 1d). These behavioral changes were not observed during the day 0 sessions (last day of habituation; no cues, random reward), when licking and running speed remained relatively uniform across space. On day 1 and subsequent recording sessions, after correcting for dwell time to account for learning-related changes in running speed, OLM axonal Ca²⁺ activity developed spatial structure, increasing modestly across the track but exhibiting a pronounced peak at the reward location during later laps (Fig. 1e–i, Extended Data Fig. 1b–g). A comparable learning-dependent increase in OLM activity at salient locations was observed in a separate group of animals recorded in a distinct environment, in which animals ran on a blank belt and a visual cue was positioned 50 cm before the reward (Extended Data Fig. 2). In this environment, OLM axonal activity likewise increased with learning but instead peaked near the cue location.

Because our prior work has shown that both EC3 target patterns and CA1 representations differ across these two environments³, the corresponding shifts in OLM activity we observe here indicate that these interneurons track the evolving CA1 population state and align with the environment-specific EC3 target structure. Together, our results identify OLM interneurons as the likely source of a learning-dependent inhibitory feedback signal that interacts with the entorhinal excitatory input at distal apical dendrites to regulate plateau initiation during learning. Notably, we observed the OLM activity development despite averaging across all axons within the imaging plane, potentially suggesting a population-level modulation of OLM output rather than the emergence of highly selective individual axons.

Spatial learning induces population-wide changes in OLM interneuron activity.

Therefore, to determine whether the learning-related OLM activity change observed at the population axonal level reflected coordinated, experience-dependent changes in a large population of OLM interneurons rather than being driven by a small subset of strongly modulated cells, we examined Ca²⁺ dynamics in the somata of identified OLM interneurons during the same behavioral paradigm. For these experiments, we used a Chrna2α-Cre mouse line, which provides selective and reliable genetic access to OLM interneurons^20,23, and performed two-photon Ca²⁺ imaging of OLM somata located in stratum oriens, where OLM cell bodies reside (Fig. 2a). Two-photon imaging revealed sparsely labeled Chrna2α-positive OLM somata within each field of view, allowing for the analysis of single-cell neuron activity. Example traces from three OLM neurons illustrate heterogeneous but coordinated activity across laps (Fig. 2b). We first quantified the population-averaged OLM somatic activity (Fig. 2c–g) and found that it closely matched the spatial and temporal profiles observed for SST axonal activity (Fig. 1, Extended Data Fig. 1b–g), indicating that the axonal signal reflects changes in OLM activity rather than additional axonal processing. Next, we examined how learning shaped activity patterns at the individual neuron level (Fig. 2h–n). Spatially binned, dwell-time–corrected activity maps from three example neurons illustrate an increase in activity primarily at the reward location over the course of the session in all three cells (Fig. 2h). Consistent with this, heatmaps summarizing the activity of all recorded Chrna2α-positive OLM neurons (n = 57) during early (quintile 1) and late (quintile 5) phases of the session reveal a population-wide increase in activity at the reward location (Fig. 2i). Consistent with this observation, quantitative analyses showed modest but significant increases in both spatial and reward selectivity in quintile 5 compared to quintile 1 across the OLM population (Fig. 2j–l). In addition, the peak activity location of individual neurons shifted toward the reward location over the course of the session (on average 13.3 +/− 3.1 cm), with a wide distribution of shift magnitudes but with 73.7% of the population shifting closer to reward (Fig. 2m). Notably, neurons whose activity peaks were initially located farther from the reward exhibited larger shifts by the end of learning (Fig. 2n). Together, these results demonstrate that learning-related changes in OLM activity are distributed across the population, with most neurons contributing to the evolving increase in inhibitory feedback signal to varying degrees.

Figure 2: Single-cell OLM activity during learning.

a, Left: Schematic of the CA1 microcircuit with imaging plane (dashed line). Right: representative, two-photon, time-averaged image of Chrna2α-OLMs expressing GCaMP6f. b, Example traces of distance, velocity, and Δf/f for three OLM neurons and the mean of all neurons in the field. c, Representative day 1 running and mean Δf/f•s. Top: along laps. Bottom: spatial profiles. d, Top: Task design of day 0 and day 1. Middle: mean z-scored Δf/f•s along the percentage of laps run. Bottom: spatial profiles. The peak-zone is 90 cm beginning at the spatial bin with reward, the away-zone is the same but 90 cm away. e, Days 0 and 1 mean z-score Δf/f•s within the peak-zone and away-zone (paired two-tailed t-test, day 1: P=8.8•10⁻³). f, Mean z-scored Δf/f•s vs. lap quintiles (Two-way RM ANOVA with post-hoc FDR correction, day 0: P=3.4•10⁻², day 1: P=1.1•10⁻²). g, Quintiles 1 and 5 mean z-score Δf/f•s vs. position (Two-way RM ANOVA with post-hoc FDR correction, quintile 5/group × bin: P=7•10⁻⁴, quintile 5/bin 5: P=1.8•10⁻³). h, Representative day 1 Δf/f•s for three OLM neurons. Top: along laps. Bottom: spatial profiles. i, Min/max scaled mean Δf/f•s for all neurons (n=55 neurons, scaling from quintiles 1 and 5, sorted by quintile 5 peak). The reward zone is 39.6 cm centered at the reward. j, Mouse reward selectivity (the fraction of neurons with mean Δf/f•s peaks within the reward-zone (paired two-tailed t-test, P=2.0•10²). k, OLM Spatial selectivity (max/mean Δf/f•s, paired two-tailed t-test, P=4.4•10⁻²). l, OLM reward selectivity (the fraction of mean Δf/f•s within the reward-zone, paired two-tailed t-test, P=1.1•10⁻²). “In” and “out” refer to activity inside or outside the reward zone. m, Mean distance that Δf/f•s peaks shifted toward reward from quintile 1 to 5. n, The absolute distance Δf/f•s peaks shifted from quintile 1 to 5, vs. the absolute peak-to-reward distance in quintile 1. In d-n day 0: n=12 mice, day 1: n=11 mice. In e,j dots represent mice, and in k-n they represent neurons. Data are shown as mean ± s.e.m.

Silencing OLM interneurons late in learning permits BTSP-mediated place cell formation.

If OLM activity indeed constitutes the inhibitory feedback predicted to regulate BTSP, then manipulating OLM activity should bidirectionally constrain plateau initiation and place field formation. To test this hypothesis, we first optogenetically suppressed Chrna2α-positive OLM interneurons during spatial learning using a combined transgenic and viral strategy that enabled simultaneous Ca²⁺ imaging and optogenetic manipulation, with GCaMP6f expressed in CA1 pyramidal neurons and ArchT selectively expressed in OLM interneurons (Fig. 3a–b). Optogenetic manipulation was restricted to an “opto-zone” centered around the fixed reward location (Fig. 3a). Each session was divided into two blocks, with no light delivered during block 1 (laps 1–50) and OLM silencing during block 2 (laps 51–100) via 594 nm-light activation of archaerhodopsin-T (ArchT)²⁴ (Fig. 3a). This design allowed us to perturb OLM activity late in learning around the reward, when and where OLM activity is highest (Figs. 1–2). Mice that did not express ArchT but were otherwise treated identically served as controls; in a subset of animals, we further confirmed that 594 nm-light effectively suppressed the activity of ArchT-expressing Chrna2α-positive interneurons (Extended Data Fig. 3a–f). Importantly, this optogenetic manipulation did not alter running or licking behavior, indicating that ArchT activation did not disrupt task engagement or reward anticipation (Fig. 3c). Consistent with the absence of optogenetic manipulation during block 1, CA1 place cell activity was similar in control and ArchT-expressing animals during this initial phase of the session (Fig. 3d–e, Extended Data Fig. 4a–c). In contrast, during block 2, silencing OLM neurons resulted in an increase in place cell density within the opto-zone around the reward in ArchT-expressing animals (Fig. 3d–e). To quantify the temporal dynamics of this effect, we analyzed the emergence of new place fields, distinguishing fields whose peaks fell within the opto-zone from those located in an “away-zone” approximately 90 cm from the opto-zone (Fig. 3f). In ArchT-expressing animals, new place field formation increased selectively within the opto-zone during block 2, with no corresponding change in the away-zone. As a consequence, the fraction of CA1 place cells was significantly higher in the opto-zone during block 2 in ArchT-expressing animals compared to controls, while no differences were observed in block 1 or outside the opto-zone (Fig. 3g). Together, these data indicate that OLM suppression selectively facilitates place cell formation at the manipulated location and late stage of learning.

Figure 3: Silencing Chrna2α-OLMs late in learning.

a, Top: day 1 manipulation paradigm wherein ArchT in Chrna2α-OLMs is activated during block 2 (50-lap blocks) within 30 cm of the reward (the opto-zone is 60 cm, centered at the reward, the away-zone is the same, but 90 cm away from the opto-zone). Bottom: Imaging and optogenetics apparatus. b, Left: Schematic of the CA1 microcircuit with two imaging planes (dashed lines). Right: representative, two-photon, time-averaged images of Chrna2α-OLMs expressing Arch-tdTomato and pyramidal neurons expressing GCaMP6f. c, Blocks 1 and 2 behavior. Top: mean running profile across space. Bottom: mean lick probability across space. d, Blocks 1 and 2 peak-normalized mean Δf/f vs. position for all place cells. Analyzed independently within blocks, sorted by place field peak location. Top: control (block 1: n=1188, block 2: n=989). Bottom: ArchT (block 1: n=762, block 2: n=748). e, Blocks 1 and 2 fraction of CA1 place cells vs. place field peak location (unpaired two-tailed t-test: block 2/bin 5 P=4•10⁻²). f, Blocks 1 and 2 CA1 place cell onsets. Top: place cells with place field peaks within the opto-zone (two-tailed t-test, block 2: P=1•10⁻²). Bottom: same for away-zone. g, Place cell distributions in the opto-zone and away-zones. Left: block 1, Right: block 2 (unpaired two-tailed t-test, opto-zone: P=1•10⁻²). h, Schematic of criteria for putative plateaus. i, Example place cell from an ArchT mouse. Mean Δf/f along laps with putative plateau marked by the white arrow. j-k. Blocks 1 (j) and 2 (k) mean plateau probability for all cells vs. position (two-way RM ANOVA with post-hoc FDR correction, block 2/group × bin: P=7•10⁻⁴, block 2/bin 5: P=0.012). In c-k, control: n=14 mice, ArchT=9 mice. In g dots indicate mice. Data are shown as mean ± s.e.m.

To determine whether this increase in place field formation in the block 2 opto-zone was associated with an increase in plateau activity, we next identified putative dendritic plateaus using established Ca²⁺ imaging signatures with GCaMP6f, derived from prior whole-cell and imaging studies^1,3,25,26. Plateau events were defined as a large-amplitude Ca²⁺ transient followed by sustained increases in activity across at least three subsequent laps relative to pre-event activity levels, with only the first qualifying event per cell included in the analysis^1,3,25,26 (Fig. 3h–k, Extended Data Fig. 5a–b, see Method section). We found that the CA1 population in control animals exhibited a marked reduction in plateau probability from block 1 to block 2 (Fig. 3j–k; 46% (± 4) of CA1 neurons exhibited a plateau in block 1 vs. 10% ± 0.8 in block 2, Full belt, mean ±SEM paired two-tailed t-test, P<1•10⁻⁴), with plateaus preferentially concentrated near the reward location. A similar global reduction in plateau probability was observed in ArchT-expressing animals (51% ± 4 vs. 17% ± 1, block 1 vs. block 2 full belt; paired two-tailed t-test, P<1•10⁻⁴). However, compared with controls, the probability of observing a plateau event during later laps in block 2 was significantly elevated in ArchT animals (unpaired two-tailed t-test, P<1•10⁻⁴). This effect was driven by an approximately 2.5-fold increase in plateau probability within the center of the opto-zone relative to controls (Fig. 3j–k), indicating a selective enhancement of plateau activity at the manipulated location. Notably, the amplitudes of the lower 80% of Ca²⁺ events (used as a proxy for non-plateau-related neuronal activity) measured in the opto-zone were not significantly affected by ArchT activation (Extended Data Fig. 5c).

Together, these results indicate that OLM interneuron activity constrains BTSP-mediated place cell formation by suppressing dendritic plateau initiation in CA1 pyramidal neurons. Silencing OLM neurons during the later phase of learning selectively releases this constraint, allowing additional place fields to form at the reward location, driven by the excitatory input pattern from EC3.

Activating OLM interneurons early in learning suppresses BTSP-mediated place cell formation.

To test whether OLM interneuron activity is sufficient to suppress place cell formation, we next performed optogenetic activation of Chrna2α-positive OLM neurons during the early phase of spatial learning (Fig. 4, Extended Data Fig. 3g–l). OLM neurons expressing ReaChR²⁷ (Chrna2α:ReaChR mice) were selectively activated within the opto-zone during block 1, a phase when place cell formation and BTSP induction are generally high due to low OLM activity (Figs. 1–2). CA1 pyramidal neurons were imaged following AAV-mediated expression of GCaMP6f in Chrna2α:ReaChR mice. Activation of OLM interneurons during block 1 significantly inhibited the development of CA1 spatial representations (Fig. 4a–e; Extended Data Fig. 4d–f). Heatmaps of CA1 population activity revealed a marked reduction in the emergence of place fields within the opto-zone surrounding the reward in ReaChR-expressing animals compared to controls. This suppression was evident both in the overall number of place cells and in their spatial distribution (Fig. 4a–e), indicating that increased OLM-mediated inhibition during early learning constrains the formation of new place fields. Consistent with the block-specific manipulation, place cell development during block 2 was comparable between groups once the optogenetic manipulation was discontinued. These results demonstrate that increasing OLM activity is sufficient to suppress place cell formation during early learning when the probability of BTSP induction is high.

Figure 4: Activating Chrna2α-OLMs early in learning.

a, Top: manipulation paradigm wherein ReaChR in Chrna2α-OLMs is activated during block 1 (the opto-zone is 60 cm centered at the reward, the away-zone is the same, but 90 cm away from the opto-zone). Middle: Blocks 1 and 2 peak-normalized mean Δf/f vs. position for all place cells from control mice. Bottom: histogram showing place cell distribution (block 1: n=1031 cells, block 2: n=901 cells). Place cells were analyzed independently within blocks, sorted by place field peak location in the heatmaps. b, same as a for ReaChR mice (block 1: n=688 cells, block 2: n=702 cells). c, Blocks 1 and 2 fraction of CA1 place cells vs. place field peak location (bin=18 cm, two-way RM ANOVA with post-hoc FDR correction, block 1/group × bin: P=1.0•10⁻³, block 1/bin 5: P=7.1•10⁻³, block 1/bin 6: P=4.0•10⁻³). d, Blocks 1 and 2 CA1 place cell onsets. Top: place cells with place field peaks within the opto-zone (unpaired two-tailed t-test, block 1: P=1.0•10⁻³). Bottom: same as top for away zone. e, Place cell distributions in the opto-zone and away-zone. Left: block 1, (unpaired two-tailed t-test, opto-zone: P=6.1•10⁻³, away-zone: P=2.8•10⁻²). Right: block 2. f-g. Blocks 1 (f) and 2 (g) mean plateau probability for all cells vs. position (block 1 two-way RM ANOVA, block 1/group × quintile: P=2.7•10⁻³, block1/bin 5: P<1•10⁻⁴, block 1/bin 6: P=7.2•10⁻³). h, BTSP working model with OLM feedback added. i, Spatial profiles of the proposed signals in early and late learning that regulate BTSP and CA1 spatial representation. Left: typical learning in control animals. Right: manipulation of typical learning by either activation or inhibition of OLMs. In a-g, control: n=8 mice, ReaChR=8 mice. In e, the dots indicate mice. Data are shown as mean ± s.e.m.

We next asked whether this suppression was due to altered plateau probability. Using the same Ca²⁺ imaging-based criteria as before (Fig. 3h), we identified putative dendritic plateau events (Fig. 4f–g, Extended Data Fig. 5d–e, see Method section). In ReaChR-expressing animals, plateau events were significantly reduced during block 1 in the opto-zone (Fig. 4f), consistent with OLM activation suppressing plateau initiation in CA1 pyramidal neurons. We further found that, while control animals showed an approximate 2-fold increase in plateau probability at the reward, this enrichment was lost in ReaChR animals (1.90±0.09 fold vs. 1.07±0.33 fold for controls vs. ReaChR animals, bin 5/mean of bins 1–3, unpaired two-tailed t-test P=2.97•10⁻²). Following the end of OLM activation, plateau occurrence returned to control levels, paralleling the normalization of place field formation. Importantly, this effect was not accompanied by a global change in neuronal activity, as the amplitudes of the lower 80% of Ca²⁺ events were not significantly altered by ReaChR activation (Extended Data Fig. 5f). Together, these experiments demonstrate that increasing OLM interneuron activity early in learning is sufficient to suppress BTSP-mediated place field formation.

Summary and Conclusions

Here, we address a central question left open by our prior work establishing BTSP as a mechanism for experience-dependent hippocampal learning: why BTSP-mediated place field formation declines with learning despite a stable EC3 target signal. We identify a learning-dependent, circuit-level inhibitory feedback signal that dynamically regulates when and where BTSP can occur (Fig. 4h–i). OLM interneurons provide this feedback, consistent with their known circuit position as recipients of excitatory input from CA1 pyramidal neurons and inhibitors of distal apical tuft dendrites. Early in learning, low OLM activity permits EC3-driven excitation to trigger dendritic plateaus and robust BTSP-mediated place field formation; as learning proceeds and CA1 population activity increases, recruitment of OLM-mediated inhibition progressively suppresses further plateau initiation. Rather than simply suppressing plasticity, this feedback enables the development of hippocampal representations to terminate once they have converged on their target pattern. This model is supported by our optogenetic manipulations: activating OLM interneurons early suppresses plateau initiation and place field formation, whereas silencing them later reinstates plateau firing and BTSP and enables additional place fields to emerge. Together, these findings establish OLM interneurons as a core control element that stabilizes hippocampal representations by constraining plasticity once learning has occurred, while preserving the capacity for rapid adaptation when representations remain incomplete or inaccurate.

Our results provide additional experimental support for the idea that dendritic plateaus act as locally computed error signals. Rather than requiring the explicit broadcasting of a global error across the circuit²⁸, plateau initiation in individual CA1 neurons depends on whether excitation exceeds inhibition within their tuft compartment, providing a cell-specific signal that additional synaptic modification is warranted. This computation may be well-suited for guiding learning in complex neuronal networks because it implicitly accounts for the many parameters that link activity across brain circuits to behavior^28–30. OLM interneurons are uniquely positioned to implement this form of inhibitory feedback, and our data indicate that learning-dependent changes in their activity are distributed across the population, consistent with them broadly sampling the evolving CA1 representation. Consistent with this interpretation, OLM activity on subsequent training days developed the reward peak more rapidly than on day 1 (Extended Data Fig. 1f–g), reflecting the accelerated stabilization of experience-dependent CA1 representations of salient locations²⁵. An additional feature of this mechanism is its generality across behavioral contexts. We found OLM axonal activity was biased toward behaviorally salient locations across environments with distinct EC3 target patterns, including one in which salience was defined by a visual cue rather than by reward delivery. Thus, OLM-mediated feedback appears not to be tied to reward per se but rather to reflect task-relevant structure. Such flexibility may be essential for spatial learning in natural environments, where the nature of salient features varies across contexts and over time.

Our results identify OLM interneurons (and specifically the Chrna2α-positive subtype) as a central control element of BTSP. However, a growing body of work suggests that additional interneuron classes play an important role in regulating hippocampal plasticity. Interneurons targeting the soma and distinct dendritic compartments exhibit task-related activity changes during goal-directed learning, interact bidirectionally with OLM cells in a state- and context-dependent manner, and influence place field activity when manipulated^{9,20,23,31–37}. In hippocampal slices, dendrite-targeting interneurons control dendritic plateau initiation^38,39, and recent work has shown that a distinct OLM subtype (i.e., NDNF/Nkx2–1-expressing) regulates plateau termination in vitro⁴⁰. Disinhibitory circuits involving VIP interneurons provide an additional layer of control. VIP interneurons, which are activated by novelty⁴¹, influence the development of hippocampal spatial representations and are thought to act indirectly through OLM-mediated disinhibition of distal apical dendrites^14,42. This mechanism has recently been linked to BTSP⁴². Beyond local inhibitory circuits, neuromodulatory systems are also likely to shape BTSP-mediated place cell formation⁴³, hippocampal plasticity⁴⁴, and, more broadly, the balance between dendritic excitation and inhibition during learning⁴⁵. For example, cholinergic subcortical nuclei directly innervate hippocampal OLM interneurons, while VIP interneurons also receive neuromodulatory input, suggesting that neuromodulation may scale OLM activity at multiple circuit levels during reward-based learning⁴⁶. Future work will be required to determine how these inhibitory and neuromodulatory pathways interact, either through OLM interneurons or in parallel, to regulate BTSP during behavior.

Together, our findings demonstrate that learning in hippocampal circuits is governed by a dendritic feedback computation that provides a biologically grounded mechanism for controlling synaptic plasticity at behavioral timescales and across hierarchically organized circuits and may generalize to other brain regions that require balancing flexibility with stability.

Online methods

All experiments were performed according to methods approved by the Institutional Animal Care and Use Committees at Baylor College of Medicine (Protocol AN-7734) and Brandeis University (Protocols 22022 & 25001).

Surgery

Experiments were conducted on either sex of adult (older than two months postnatal) SST-IRES-Cre²² (n = 23, JAX, #018973, also kindly provided by Andreas Tolias), Chrna2α-Cre²³ (n=21, kindly provided by Simon Chamberland, Richard Tsien, Jun Wu, and Gina Poe), and GP5.17⁴⁷ (n=4, JAX, #025393) mice, as well as double transgenic Chrna2α-Cre:GP5.17 (n=14), and Chrna2α-Cre:ReaChR²⁷ (JAX, #024846) mice (n=16). Mice were housed in a reverse light cycle facility (12 hours dark/12 hours light) maintained at 30–50% humidity and 21 °C. Surgeries were performed using stereotactic methods while mice were kept under deep anesthesia using isoflurane. Following local antisepsis and the application of a topical anesthetic, the scalp was removed and the skull cleaned. Then, the skull was leveled, and two craniotomy locations for virus injections in dorsal CA1 were marked (2.15 mm | 2.35 mm posterior from bregma, 1.8 mm | 2.2 mm lateral from the midline) as well as the center location of a cranial window (2.2 mm posterior from bregma and 2.0 mm lateral from the midline). Following the virus injection, a 3 mm diameter craniotomy was made above the hippocampus, and the cortical tissue above the hippocampus was aspirated while rinsed with 0.9% sterile saline. Once the hippocampal capsule was visible, the window implant (a 3 mm-diameter, 1.7 mm-high metal cannula (McMaster) with a window on the bottom and 0.5 mm of shrink tubing around the top) was implanted and cemented in place. Lastly, a titanium head bar was secured to the skull using dental acrylic (Ortho-Jet, Lang Dental). Virus injections were carried out using a microinjector (Drummond, Nanoject II) loaded with a pulled glass pipette that was trimmed and beveled to 25–35 μm and backfilled with mineral oil (Sigma). The virus was loaded by retracting the injector’s plunger. The microinjector was positioned over the craniotomies using a micro-manipulator (MP-285A, Sutter Instruments). For each injection site, a 0.5 mm round craniotomy was drilled, and 46 nl of virus dilution was injected at two depths per site (1200–1300 μm and 1000–1050 μm). Virus injections varied across experimental paradigms. For imaging SST-OLM axons: AAV2/1.Syn.Flex.GCaMP6f.WPRE.SV40 (Addgene, #100833-AAV1, titer: 8×10¹¹-2×10¹²) with AAV2/1-syn-Flex-TdTomato (UNC Neurotools, titer: 3–9×10¹⁰). For imaging Chrna2α-OLM somata: AAV2/1.Syn.Flex.GCaMP6f.WPRE.SV40 (titer:8×10¹¹-2×10¹²). For inhibition of Chrna2α-OLMs: AAV2/1.hsyn.Flex.ArchT.tdTomato (UNC Neurotools, titer: 3.3×10¹¹), or AAV2/1.hsyn.Flex.ArchT.tdtomato (UNC Neurotools, titer: 3.3×10¹¹) with AAV2/1.Syn.GCaMP6f.WPRE.SV40 (Addgene, #100837, titer: 1–3×10¹²). For inhibition controls: AAV2/1-syn-Flex-tdTomato (UNC Neurotools, titer: 3×10⁹-9×10¹⁰), or AAV2/1-syn-Flex-tdTomato (UNC Neurotools, titer: 3×10⁹-9×10¹⁰) with AAV2/1.Syn.GCaMP6f.WPRE.SV40 (Addgene: #100837-AAV1, titer: 1–3×10¹²). For activation of Chrna2α-OLMs and controls: AAV2/1.Syn.GCaMP6f.WPRE.SV40 (Janelia Viral Core, titer:1×10¹²).

Behavioral training and task on the linear track treadmill

The treadmill comprised a fabric belt (180 cm; McMaster Carr) and a custom-built lick port that dispensed sucrose water, controlled by a solenoid valve (quiet operation; The Lee Company). The belt was self-propelled by the animal, and its speed was tracked using an encoder at the front wheel axle. An Arduino-based behavioral state machine and corresponding MATLAB GUI (Bpod, Sanworks) were used to control the reward value and sensors, and to track the encoder. A custom-written MATLAB GUI controlled the light for optogenetic manipulations. The behavioral data, stimulations, and frame times were monitored and recorded using a NIDAQ-based acquisition system (National Instruments, PCIe-6341) and the MATLAB-based Wavesurfer software (version 0.982, Janelia).

5 to 7 days following surgery, the mice were placed on a water restriction protocol (receiving 1.5ml per day). Three or four days into water restriction, animals were habituated to the experimenter for 5–6 days, followed by 3–5 days of head-fixed treadmill training. Training days were conducted during the animal’s dark cycle on a belt with no cues and with a sucrose reward (5–10% solution) dispensed at a random location each lap. During experimental days, the mice learned to navigate a cue-rich belt (three cue sections: Velcro tape patches, thin hot glue sticks, and correction fluid dots) to a single fixed location for a sucrose reward. In the light-cue learning experiments (Extended Data Fig. 2), the animals were on a blank belt (no tactile cues) and received a bilateral visual stimulus of blue light (10 Hz for 500 ms) 50 cm before the fixed reward location. For OLM activity experiments, recordings lasted 30–60 minutes. Mice were excluded if they ran <75 laps or >350 laps in 60 minutes. Mice recorded past day 1 (Extended Data Fig. 1) were recorded for a maximum of five more days. For the optogenetic experiments, analyses were limited to 88–120 laps (mean±SEM: 101±1 laps) per session. Animals were recorded at most once per day.

In vivo two-photon Ca²⁺ imaging

All experiments were conducted in a dark box, and Ca²⁺ images were captured using custom-built two-photon microscopes (detailed descriptions below). Experiments recording OLM interneuron activity were performed on both microscopes. Optogenetic experiments activating ArchT or ReachR were performed on the microscopes at Brandeis or Baylor, respectively. The images were recorded using ScanImage at 30 Hz and 512 × 512 pixels. Fields of view were 280 × 280 um and for animals with multiple recording days, we attempted to return to the same field of view (Extended Data Fig. 1a). GCaMP6f, and in some cases tdTomato or Citrine, were imaged through Nikon 16x, 0.8-numerical-aperture objectives (both microscopes) and excited at 920 nm (typically 30–75 mW measured under the objective)

For experiments conducted at Baylor recording the activity of CA1 OLM interneurons, the microscope (Janelia MIMMS 2.0 design) used a Ti:Sapphire laser (Coherent, Chameleon Ultra II), passed through a FF705-Di01 (Semrock) primary dichroic filter. Images were collected with ScanImage R2022 premium software (Vidrio). Emission light passed through a 565 DCXR dichroic filter (Chroma) and either stacked 505/19 nm (GCaMP6f channel; Semrock) and 514/44 nm (GCaMP6f channel; Semrock), or a 612/69 nm (tdTomato channel, Semrock) bandpass filters. Light was detected by two GaAsP photomultiplier tubes (11706P-40SEL, Hamamatsu). For optogenetic ReaChR manipulation experiments, in which only the GCaMP6f channel was recorded, a 594 nm laser (Obis, Coherent) was used. Diffuse 594 nm light was reflected into the beam path by a DMLP805R (Semrock) dichroic filter, and the primary dichroic was substituted with a Di02-R561 (Semrock) dichroic filter.

For experiments conducted at Brandeis, the microscope at Brandeis (INSS, UK) included a Spectra-Physics Insight X3 laser. Images were collected with ScanImage R2021 software (Vidrio). Emission light passed through a T565lpxr (Chroma) dichroic filter and either an ET 510/80 nm (GCaMP6f channel, Chroma) or a 630/75 nm (tdTomato channel, Chroma) bandpass filter, then was detected by respective GaAsP photomultiplier tubes (11706P-40, Hamamatsu). For optogenetic ArchT manipulation experiments, in which only the GCaMP6f channel was recorded, diffuse 594 nm light from an LED (Thorlabs) passed through an ET570LP (Chroma) low-pass filter, and the filter in front of the PMT in the GCaMP6f channel was replaced with an ET 510/80 (Chroma) bandpass filter stacked between two BG39 (Chroma) filters.

Optogenetic perturbation of CA1 OLM interneurons

To examine the effect of OLM inhibition (performed on the microscope at Brandeis), we used the following groups: 1) experimental mice (Chrna2α-Cre:GP5.17 injected with AAV2/1.hsyn.Flex.ArchT.tdtomato, n=6; Chrna2α-Cre injected with AAV2/1.hsyn.Flex.ArchT.tdtomato and AAV2/1.Syn.GCaMP6f.WPRE.SV40, n=3); 2) control mice (Chrna2α-Cre:GP5.17 injected with AAV2/1.hsyn.Flex.tdtomato, n=8; Chrna2α-Cre injected with AAV2/1.hsyn.Flex.tdtomato and AAV2/1.Syn.GCaMP6f.WPRE.SV40, n=2; non-injected GP5.17, n=4). The experimental (ArchT-expressing) and control mice underwent the same procedures and paradigm: the window was implanted in the same surgery as the injections. Imaging and optogenetic experiments were conducted 18 to 28 days post-surgery. Recordings were obtained on days 1–3 under the otherwise same conditions. Continuous diffuse light of 594 nm from an LED (Thorlabs) at 4 mW measured under the objective was delivered through the objective. In all optogenetic experiments, the region of the belt where 594 nm light was applied (the “opto-zone”) was 60 cm centered at the reward location. Light was applied only in block 2 (approximately 50 laps) when the animals were around the fixed reward location (10 – 70 cm); there was no time limit. To validate the effect of ArchT on OLM activity (Extended Data Fig. 3a–f), a subset of Chrna2α-Cre mice that expressed both GCaMP6f and ArchT (n=3) or tdT (n=2) were used. A field of view was located that contained OLM neurons expressing GCAMP6f, and these neurons were or were not expressing ArchT or tdTomato, respectively. The Ca²⁺ signals of these cells were recorded while the LED was on for approximately 2.5–3 seconds, or, in some laps, for 60 cm of the animal’s travel on the belt.

To examine the effect of OLM excitation (performed on the microscope at Baylor), we used the following groups: 1) experimental mice (Chrna2α-Cre:ReaChR injected with AAV2/1.Syn.GCaMP6f.WPRE.SV40, n=8); 2) littermate control mice (Chrna2a-Cre:ReaChR litter mates that were confirmed by genotype to lack at least one allele required for ReaChR expression, injected with AAV2/1.SynGCaMP6f.WPRE.SV40, n=8). The experimental ReaChR-expressing mice and littermate control mice underwent the same procedures and paradigm as ArchT mice in OLM inhibition experiments. Optogenetic manipulation in ReaChR-expression animals was restricted to day 1 on the cue-enriched, fixed-reward belt. Diffuse light of 594 nm was delivered from a laser (Obis, Coherent) at 40 Hz (square pattern) at 0.8–0.95 mW measured under the objective. 594 nm light was only applied during block 1 (50 laps) while the mouse was within the opto-zone. The laser was turned off if the mouse’s velocity dropped below 2.5 cm/s, or if it had been applied for more than 5s continuously. To validate the effect of ReaChR activation on OLM activity in mice used for the data presented in Fig. 4, stratum oriens interneurons were found directly above the field of view used for pyramidal neuron recordings, on the last day of experimentation (Extended Data Fig. 3g–l). The Ca²⁺ signals of these interneurons were recorded while the laser was either activated for 5 seconds (“fixed-duration”, n=8 mice) or within the opto-zone (“fixed-location”, n=7 mice). Each mouse was validated to have responsive OLMs, which showed> 2-fold increases in mean Δf/f while the laser was on (for fixed-duration: a 3s window before stimulation was compared with a 4s window during; for fixed-location: the opto-zone was compared with surrounding locations).

Data analysis

Ca²⁺ signal extraction and activity map generation

To extract axonal Ca²⁺ signals from the population of CA1 SST-expressing OLM interneurons, motion correction was performed using the Suite2p pipeline⁴⁸, with both green (GCaMP6f) and red (tdTomato) channels aligned, and the red channel used for image registration. GCaMP6f fluorescence was then averaged across the field of view using custom MATLAB functions (version 2024b), excluding a narrow border (2% of frame width) to minimize motion-related artifacts. For somatic Ca²⁺ imaging, motion correction and ROI detection were performed using Suite2p on GCaMP6f fluorescence. Functional ROIs were detected for Chrna2α-expressing OLM somata. For pyramidal neurons in ArchT mice, ROIs were identified using a custom-pretrained Cellpose⁴⁹ model. In ReaChR mice, which exhibited dense pyramidal neuron labeling, functional ROI detection was used and cross-validated with the same Cellpose model. All ROIs were manually curated, and those with insufficient signal quality were excluded. In recordings from the stratum pyramidale in ReaChR mice, interneurons were identified based on event amplitude, duration, and frequency, and manually removed. All datasets in which motion correction was unsuccessful were excluded.

Further analyses were performed using custom MATLAB code. First, raw fluorescence signals were converted to Δf/f, calculated as (f − f₀)/f₀. For pyramidal neurons, f₀ was defined as the mode of the fluorescence histogram, whereas for OLM interneurons, f₀ was defined as the 8th percentile of fluorescence values. For pyramidal cell recordings, this was followed by neuropil correction (subtraction of 70% of the neuropil Δf/f signal obtained via Suite2p) and the generation of spatial activity maps by dividåing the 180-cm belt (one lap) into 50 spatial bins (3.6 cm each). For each lap, mean Δf/f values were computed within each spatial bin during periods when running speed exceeded 2.5 cm/s. These spatially binned activity maps were used to calculate noise correlations between neighboring ROIs (within three soma diameters). Pearson’s correlation coefficient thresholds of 0.35–0.5 were applied to identify potential signal cross-contamination, and contaminated ROIs were excluded from further analysis. For interneuron recordings, spatial Δf/f maps were multiplied by the animal’s dwell time within each spatial bin to account for learning-related changes in running speed and isolate spatial patterns of inhibitory activity. Dwell time was computed as d_bin/V_bin, where d_bin is the spatial bin width (3.6 cm), and V_bin is the mean velocity within that bin. Significant Ca²⁺ events were identified as transients exceeding three standard deviations of the baseline noise (mode of the histogram). For visualization, spatial maps were smoothed using a three-point boxcar filter and centered such that either the reward location (cue-belt environment) or the visual cue location (visual cue environment) corresponded to spatial bin 25.

CA1 place cell identification

Place cells were determined as previously described^3,18. In brief, from each CA1 pyramidal neuron’s spatial map of mean Δf/f, we first identified potential onset laps for a place field (induction lap). Place fields were defined as those spatial bins with contiguous activity >20% of the peak mean Δf/f. Induction occurred at a lap where significant Ca²⁺ activity was within the neuron’s eventual place field, and two out of the following five laps also had significant activity in the eventual place field. If multiple laps fit these criteria per neuron, the first instance was the induction unless the place field generated was weak and disappeared for more than 20 laps at some point. Only the laps following place field induction were used to determine place cell identity. CA1 place cells were defined as neurons exhibiting (1) significant spatial information about the linear track position (outside the 95% confidence interval of values obtained from shuffled data) and (2) high place field reliability, defined as significant activity on more than 33% of laps following place field induction. The spatial information was calculated as described previously¹⁸ and compared to 100 shuffles of the activity. For shuffling, Δf/f from periods when the mouse’s velocity exceeded 2.5 cm/s was circularly rotated 500 frames and divided evenly into six sections that were permuted in pseudo-random order. Place cells were determined independently across the laps of blocks 1 and 2, within which only one place field was used per neuron. If the spatial bin containing the peak of a place cell’s place field moved by >20 cm between the laps of blocks 1 and 2, its block 2 place field was considered a new place field for that cell.

Plateau detection

Only cells with at least 20 significant events (amplitude > three times the standard deviation of baseline noise) were included in this analysis. Putative plateau (Figure 3–4) events were identified using criteria adapted from previous whole-cell recordings and descriptions of GCaMP6f imaging of Ca²⁺ events^1,3,25,26. Putative plateaus were defined as events that (1) were large Ca²⁺ transients, reaching amplitudes within the top 10% of all significant events for each neuron, and > 1Δf/f; (2) were followed by significant activity within 45 cm of the event peak in at least three of the subsequent five laps; (3) were followed by a forward shift (relative to the event peak) in the mean Δf/f center of mass across the subsequent five laps; (4) Only one event per cell, i.e., the first event meeting criteria (1)-(3), was included in subsequent analyses (the exception is Extended Data Fig. 5 a, b, d, e; see below). Plateau identification was performed across all cells, regardless of spatial location. Spatial maps of plateaus were generated for each cell, wherein every spatial bin × lap was scored as having a plateau (1) or not (0). Plateau probability per animal for blocks 1 and 2 was calculated as the total number of detected plateau events within each spatial bin (10 bins, 18 cm per bin), normalized by the number of regions of interest (ROIs) in the animal. For Extended Data Fig. 5a, b, d, e, only plateau events from place cells that occurred within their place field at or following the place field induction were included. For correlation between velocity and place field width, the mean velocity (cm/s) across the time points of every Δf/f value averaged within the spatial bin containing the peak of the plateau event was used versus the place field computed over the ten following laps.

Behavioral data quantification

As with the analysis of Ca²⁺ data, spatial maps for behavioral data were constructed by first dividing the belt length (also one lap, 180 cm) into 50 spatial bins (3.6 cm each). Along all laps, the mean velocity was averaged within each spatial bin. Lick probability was computed by scoring each spatial bin as containing (1) or lacking (0) at least one lick by the mouse, then by taking the mean along this binary matrix.

Statistical methods.

The exact sample size (n) for each experimental group is indicated in the figure legend or in the main text. No statistical methods were used to predetermine sample sizes, but our sample sizes are similar to those reported in previous publications. In some cases, when the data distribution was assumed but not formally tested to be normal, data were analyzed using two-tailed paired or unpaired t-tests, as stated in the text or figure legends. Data were analyzed automatically without consideration of trial conditions or experimental groups. Experiments and data analyses were not randomized and not performed blind to the experimental conditions. If not otherwise indicated in the figure, data are shown as mean ± SEM.

Data Availability

The data that support the findings of this study are available from the corresponding author upon request.

Code Availability

The code that supports the findings of this study is available from the corresponding author upon request.