Variable recruitment of distal tuft dendrites shapes new hippocampal place fields

SUMMARY

Hippocampal pyramidal neurons support episodic memory by integrating complementary information streams into new ‘place fields’. Distal tuft dendrites are widely thought to initiate place field formation by locally generating prolonged, globally-spreading Ca²⁺ spikes known as plateau potentials. However, the hitherto experimental inaccessibility of distal tuft dendrites in the hippocampus has rendered their in vivo function entirely unknown. Here we gained direct optical access to this elusive dendritic compartment. We report that distal tuft dendrites do not serve as the point of origin for place field-forming plateau potentials. Instead, the timing and extent of peri-formation distal tuft recruitment is variable and closely predicts multiple properties of resultant place fields. Therefore, distal tuft dendrites play a more powerful role in hippocampal feature selectivity than simply initiating place field formation. Moreover, place field formation is not accompanied by global Ca²⁺ influx as previously thought. In addition to shaping new somatic place fields, distal tuft dendrites possess their own local place fields. Tuft place fields are back-shifted relative to that of their soma and appear to maintain somatic place fields via post-formation plateau potentials. Through direct in vivo observation, we provide a revised dendritic basis for hippocampal feature selectivity during navigational learning.

INTRODUCTION

The central nervous system supports learning and memory by accomplishing two core tasks: (1) integrating sensory information to faithfully represent the external world and (2) updating how it processes this information such that future computations better promote an organism’s overall fitness. Pyramidal neurons (PNs) bear much of this responsibility as they comprise the main source of excitatory drive in the cerebral cortex^{1, 2}. As such, PNs unite multiple levels of brain function to carry out complex computations. A single PN receives many thousands of synaptic inputs across its dendritic arbor^{3, 4}. These inputs typically originate from multiple presynaptic circuits that broadcast complementary streams of information to distinct dendritic compartments^{5, 6}. In turn, dendritic compartments can locally and nonlinearly process synaptic inputs^7–12 according to bespoke integrative rules^13–17 while local plasticity mechanisms update these rules with experience^18–22. This nexus between subcellular and systems levels of brain function endows a single PN with considerable computational capacity^23–27 but is exceedingly difficult to interrogate experimentally. Therefore, one of the most central questions in neuroscience is how PNs integrate their diverse synaptic inputs to drive feature-selective somatic action potential (AP) firing underlying behavioral adaptation^{28, 29}.

PNs in hippocampal area CA1 represent a key cellular substrate for learning and memory and exhibit a striking form of feature selectivity in the form of receptive fields known as ‘place fields’ (PFs). PFs emerge from experience to conjunctively encode context-specific sensory features^{30, 31} and, as a result, are spatially tuned to specific locations within an animal’s environment³². PFs support spatial navigation in rodents by forming flexible ‘cognitive maps’^32–34 and analogous receptive fields are believed to support episodic learning and memory in humans^35–38. New PFs emerge from experience via a recently described synaptic plasticity mechanism known as behavioral timescale synaptic plasticity (BTSP)³⁹. BTSP is unique in its ability to bind together pre- and postsynaptic activity over behaviorally relevant timescales after a single pairing. For this reason, much recent effort has been dedicated to uncovering its still-elusive circuit^{40, 41}, cellular^{20, 41–43}, and molecular^44–46 mechanisms.

Distal tuft dendrites are thought to initiate PF formation. Specifically, they are thought to convert multiplexed sensory and spatial signals from the entorhinal cortex (EC)^47–56 into an instructive signal in the form of a dendritic plateau potential^{39, 42, 57}. According to this model, plateau potentials depolarize the entire PN such that any dendritic spines receiving excitatory input within the seconds-long BTSP association window may undergo plasticity. This model is supported by causal manipulations of EC inputs to CA1^{40, 57, 58}, the re-weighting of more-proximal synapses during PF formation^{20, 41, 43}, and the ability of distal dendrites to drive global depolarization and promote synaptic plasticity^59–62 in vitro. However, distal tuft dendrites have hitherto proven experimentally inaccessible in CA1 PNs due to their location at the deepest reaches of the apical arbor. The inability to directly monitor this elusive dendritic compartment has significantly impaired our ability to test its role in the emergence of hippocampal feature selectivity.

Here we gained direct, simultaneous access to distal tuft dendrites and their soma in single CA1 PNs in vivo. We addressed, for the first time, three broad questions: (1) How does distal tuft dendritic activity inform somatic activity across behavioral states? (2) What experiential features do distal tuft dendrites encode? (3) What roles do distal tuft dendrites play in the emergence of new PFs during spatial navigation? We leveraged single-cell DNA electroporation, multi-plane two-photon Ca²⁺ imaging, and a virtual reality-based serial ‘teleportation’ paradigm to uncover a remarkably broad repertoire of previously unappreciated distal dendritic functions that address each of these questions.

RESULTS

Monitoring somatic and distal tuft dendritic dynamics in single CA1 PNs in vivo

We sought to investigate the role of distal tuft dendrites in shaping somatic activity and forming new PFs. To do so required simultaneous access to the soma and distal tuft dendrites of CA1 PNs. However, these dendrites have not been previously imaged in CA1 PNs in vivo due to their depth and overlap with neighboring dendritic arbors. We first overcame these biological constraints by expressing DNA plasmids encoding the red-shifted Ca²⁺ indicator XCaMP-R⁶³ in individual CA1 PNs using in vivo single-cell electroporation (SCE)^{20, 64–66} (Figure 1A). The use of a red-shifted sensor mitigated depth-dependent light scattering from biological tissue while SCE provided single-neuron sparsity. We then gained simultaneous access to the soma of imaged distal tuft dendrites by using a piezoelectric device to rapidly actuate a 40x objective lens between distant focal planes (Figure 1B). To relate somatic and distal dendritic dynamics to animal behavior and PFs, we imaged electroporated CA1 PNs while mice navigated a 3-m virtual linear track⁶⁷ for randomly located water rewards (Figure 1, C & D; Figure S1). SCE permitted unambiguous allocation of imaged dendrites to their parent soma (Figure 1, E & F).

Figure 1. Simultaneously monitoring somatic and distal tuft dendritic dynamics in isolated CA1 PNs during virtual navigation.

(A) Plasmid DNA encoding the red-shifted Ca²⁺ indicator XCaMP-R was delivered to individual CA1 PNs in anesthetized adult mice. Alexa Fluor 488 was visualized by 2-photon microscopy to guide pipet descent. (B) XCaMP-R was simultaneously imaged in somatic and distal tuft dendritic imaging planes using a 1070 nm fixed-wavelength 2-photon laser and a piezoelectric device that rapidly toggled a 40x water immersion objective lens. Representative motion-corrected, time-averaged images are shown for each focal plane (scale bar: 20 μm). Images were cropped to emphasize regions of interest (white outlines). (C) Mice navigated a 3-m virtual linear track for randomly delivered water rewards (1 / lap). (D) Summary plot depicting somatic and dendritic signals with time-locked behavioral variables. (E) Representative in vivo volumetric scan of same CA1 PN shown in panels (B) and (D). Image stack was processed to maintain consistent contrast across depths. Basal dendrites dorsal to the soma not included. (F) Reconstruction of CA1 PN shown in (E), color-coded by compartment (soma: black, apical trunk and radial oblique dendrites: grey, distal tuft dendrites: red). Numbered arrows indicate imaged dendrites shown in same order as in (D). Dashed lines indicate dendrites that are occluded in 3D reconstruction.

CA1 PN soma distal tuft dendrites are most strongly isolated from their soma during locomotion

The degree of functional autonomy that distal tuft dendrites possess relative to their soma carries major implications regarding their function, including the extent to which they filter EC inputs as well as their role in somatic PF formation. CA1 PN distal tuft dendrites are located > 300 microns from their soma with potentially limited influence on somatic AP firing due to passive^{68, 69} filtering properties of dendrites as well as voltage-dependent leak currents^70–73. On the other hand, CA1 PN tuft dendrites can drive robust somatic AP firing under certain conditions including, but not limited to, rebound spiking from dendritic inhibition^74–76, temporally correlated input to distal and proximal dendrites^{60, 61, 77}, and somatic depolarization⁷³. While these in vitro and computational studies have proven invaluable in appreciating the complexity of dendritic computation in the hippocampus, they have also highlighted the challenges inherent in predicting how tuft dendrites might influence their soma in vivo while sensory experience and animal behavior continuously and variably modulate each of the aforementioned determinants of tuft-soma crosstalk.

To gain insight into the functional properties of distal tuft dendrites, and garner clues as to how they might support higher-level cellular processes such as PF formation, we first examined basic Ca²⁺ transient properties of these dendrites relative to their soma. Tuft dendrites possessed shorter Ca²⁺ transient waveforms than their soma while both compartments displayed strong positive modulation by locomotion (Figure S2, A–C). Tuft dendrites displayed a high degree of autonomous activity, with many Ca²⁺ transients lacking a contemporaneous somatic transient (‘isolated’ transients) or even a contemporaneous tuft transient (Figure 2A). Greater than one third of tuft Ca²⁺ transients were isolated during running periods (med. = 0.37, IQR = 0.18) and, counter to our expectation, this fraction was reduced on average during periods of immobility (med. = 0.29, IQR = 0.31) (Figure 2B).

Figure 2. Bidirectional and state-dependent compartmentalization between CA1 PN soma and distal tuft dendrites.

(A) Soma (black) and tuft (red) traces from an example CA1 PN. Ca²⁺ transients are overlaid in cyan. Asterisks indicate isolated tuft transients. Binary running frames (green) indicate velocity ≥ 1 cm / s. (B) Fractions of tuft Ca²⁺ transients that were isolated, i.e. did not propagate to soma, during running and immobile states. Gaussian kernel density estimates shown on right. (C) Cross-compartment conditional analysis quantifying the fraction of times compartment ‘a’ was recruited when compartment ‘b’ fired (‘a|b’). Values were averaged within-cell. Linear mixed effects model, compartments effect: F_{(1.42, 32.70)} = 6.50, P < 0.01; running effect: F_{(1, 46)} = 0.022, P > 0.05. Post hoc t-test results shown. (D) Top: Mean somatic Ca²⁺ transient waveform triggered on tuft dendrite transient onsets. Vertical lines indicate timestamps for half-maximum amplitudes used for lag calculation. Bottom: Histogram of soma-tuft lags with Gaussian kernel density estimate. (E) Top: Example tuft-based predictions of somatic ΔF/F₀ (red) illustrating range of model performance. True somatic ΔF/F₀ traces are shown in black. Predictions are each based on 5 tuft dendrites for ease of comparison. Bottom: Model performance as a function of the number of dendrites used for model training. Linear mixed effects model: F_{(0.83, 4.46)} = 11.90, P < 0.05. Fisher’s LSD shown for N vs. N – 1 comparisons up to N = 6. Dashed line corresponds to second y-axis and indicates the number of cells for which as least N tuft dendrites (x-axis) were imaged. Violin plots depict full data range with lines at medians and error bars spanning first to third quartiles. *P < 0.05, **P < 0.01. See table S1 for sample sizes and table S2 for additional statistical details.

We next asked whether soma-tuft compartmentalization might be hysteretic, i.e. biased toward propagation in a particular direction. Our temporal resolution did not permit direct observation of event propagation. However, events that occur in tuft dendrites but not their soma can be reasonably interpreted as failures of centripetal (i.e. ‘forward’) propagation and vice-versa. We thus employed a cross-compartment conditional analysis to assess the likelihoods that tuft events accompany a somatic event (‘tuft|soma’), a somatic event accompanies a tuft event (‘soma|tuft’), and tuft events accompany a tuft event (‘tuft|tuft’). This analysis revealed strong overall compartmentalization both between and within compartments. Centripetal propagation from tufts to soma was stronger than propagation between individual tuft segments and non-significantly (P = 0.067) stronger than centrifugal propagation from soma to tufts (Figure 2C). Consistently, somatic Ca²⁺ transients most commonly lagged those of tuft dendrites (median = 0.20 s, IQR = 0.60 s) (Figure 2D). This measurement corresponded to the limit of our temporal resolution set by image acquisition rate across distant focal planes (5–6 Hz). Isolated tuft Ca²⁺ transients displayed higher amplitude, duration, and overall tuft recruitment than those coactive with the soma (Figure S2, D–F). We note that cross-compartmental relationships in Ca²⁺ flux may differ from those of voltage due to a number of reasons, including divergent Ca²⁺ handing mechanisms. Nonetheless, the marked disparities between isolated and global Ca²⁺ events, along with cross-compartment conditional analyses in Figure 2C, indicate that (1) backpropagating somatic APs typically drive voltage-dependent Ca²⁺ influx in only a subset of tuft dendrites and (2) even large, supralinear events recruiting many tuft dendrites can fail to drive detectable somatic Ca²⁺ transients in vivo.

Given the prevalence and prominence of isolated events in tuft dendrites, we next asked how much information is shared between tuft dendrites and their connected soma. To this end, we trained a double-cross-validated Ridge regression model to predict somatic XCaMP-R signals (ΔF/F₀) from random combinations of 1 to 6 connected tuft dendrites (see supplementary materials and methods). Tuft predictive power linearly summed up to the inclusion of 4 dendrites, at which point model performance plateaued at R² = 0.42 ± 0.092 (Figure 2E). We draw two conclusions from this result. First, the modest peak model performance indicates that the overall amount of information shared between these distant cellular compartments is limited. Second, the fact that model performance saturated with only 4 tuft dendrites indicates that the information that is shared between compartments is relatively low-dimensional. In summary, local tuft dynamics, while clearly robust, appear to have limited influence on somatic action potential firing during ordinary foraging behavior in a familiar environment.

Robust detection of somatic place field formation and dendritic plateau potentials in vivo

The high degree of compartmentalization we observed in CA1 PN distal tuft dendrites is amenable to their presumed role in PF formation. While these dendrites exert modest influence on somatic AP firing under basal conditions (Figure 2, B & C), they can generate robust isolated events (Figure S2, D–F) which, if properly timed with input to proximal apical dendrites, may drive dendritic plateau potentials to form new PFs. In the absence of an experimental approach to directly monitor plateau potentials in CA1 PN dendrites, previous in vivo studies have inferred dendritic plateau potentials from their somatic vestige known as the after-depolarizing potential (ADP)^{39, 42, 57, 78}. Therefore, it remains unclear where plateau potentials originate and how globally they spread throughout the dendritic arbor in vivo, i.e. what fraction of eligible dendritic spines will undergo synaptic plasticity during PF formation. These unanswered questions are critical to understanding cellular mechanisms of memory formation. To probe the role of tuft-originating plateau potentials in PF formation, we sought to (1) reliably elicit spontaneous somatic PF formation events, (2) assess tuft Ca²⁺ dynamics at the time of those events, and (3) disambiguate between ordinary tuft Ca²⁺ transients and those driven by dendritic plateau potentials.

While optogenetic stimulation and somatic current injection protocols have recently emerged as a mean to induce PF formation^{20, 39, 41–43, 66, 79} we required an approach that would preserve the in situ role of distal tuft dendrites in PF formation. Exposure to novel contexts has been shown to promote PF formation in area CA1 PNs^80–82 in the absence of neural stimulation. We thus leveraged a virtual context switching task as a causal tool to evoke spontaneous PF formation. Mice were serially ‘teleported’ every 12–15 laps across four novel, 3-m linear virtual tracks. In imaging sessions using fixed-location water reward, mice rapidly learned to slow down and lick as they approach reward zones (Figure S3). Mice then revisited the same contexts (‘familiar’) for the remainder of each 30-minute imaging session (Figure 3A; see supplementary materials and methods).

Figure 3. Detection and characterization of PF formation events and plateau potentials.

(A) Schematic of virtual teleportation paradigm. All contexts were novel upon first exposure. In 6/8 experiments, reward was positioned at a unique fixed location in each context to promote attention to visual cues. (B) Example raster plot of somatic activity, color-coded by context. Tick marks represent deconvolved events from ΔF/F₀ signal which were used for spatial tuning analyses. Each row represents one lap. Shaded regions denote de novo PFs, horizontal dashed lines indicate PF formation laps, and vertical dashed lines indicate animal position at the moment of PF formation. (C) Somatic PF formation events per traversal of a novel or familiar context. (D-F) Hallmarks of BTSP-driven PF formation^{20, 39, 43, 82} including a bias for PFs to form early in novel contexts (D), a correlation between peri-formation running velocity and resultant PF width (E), a tendency for PFs to appear back-shifted in space relative to animal position at time of plateau potential (F, top), and stronger somatic activity during PF formation relative to subsequent PF traversals (F, bottom). (G) Mean event Ca²⁺ waveforms for tuft transients (blue) and plateau potentials (orange). (H) Example somatic (black) and tuft (red) Ca²⁺ traces showing both transients and plateau potentials with context and relative animal position at bottom. Shaded regions indicate PF traversals. (I-K) Comparison of basic Ca²⁺ event properties between ordinary tuft Ca²⁺ transients and plateau potentials including frequency (I), amplitude (J), and duration (K). (L) Somatic responses to ordinary Ca²⁺ transients (blue, left) and plateau potentials (orange, right) detected from mean ΔF/F₀ signal of tuft dendrites imaged for each CA1 PN. (M) Quantification of somatic responses in (L). Violin plots depict full data range with lines at medians and error bars spanning first to third quartiles. **P < 0.01, ***P < 0.001. See table S1 for sample sizes and table S2 for additional statistical details.

We then developed a new approach to detect and precisely timestamp PF formation events based on Ca²⁺ signals (see supplementary materials and methods) and identified a remarkable 39 de novo somatic PF formation events from 30-min recordings of just 9 CA1 PNs (Figure 3, B & C; Figure S4). These PF formation events displayed distinctive hallmarks of BTSP^{20, 39, 82} including a strong association with novel context exposure, a correlation between peri-formation animal running speed and resultant PF width, a tendency for a new PF to emerge backward in space relative to the animal’s location at the time of PF formation, and stronger than usual peri-formation activity levels (Figure 3, C–F). De novo PFs were highly stable across laps as well as repeated traversals of the virtual contexts in which they emerged, despite interleaving exposure to 3 distinct contexts (Figure S5). Therefore, our virtual teleporting task served as a causal tool to reliably evoke PF formation while preserving the in situ role of tuft dendrites in this process.

To gain insight specifically into the role of tuft-driven plateau potentials in PF formation, we devised an unsupervised machine learning approach to factorize and cluster tuft Ca²⁺ transient waveforms into two groups (Figure S6, A–C). Imposing zero assumptions regarding plateau potential Ca²⁺ signals, we identified two distinct activity modes of tuft dendrites (Figure 3G). We then trained a double-cross-validated ensemble of support vector classifiers to correctly predict clustering labels of individual Ca²⁺ events with an accuracy of 99.80% (Figure S6D). Next, we reran our Ca²⁺ transient detection pipeline using template waveforms from both clusters along with the trained classifier to label events as ordinary Ca²⁺ transients or putative plateau potentials (Figure S6E; Figure 3H). Finally, because plateau potentials are considered globally depolarizing cellular events, we applied a winner-take-all rule whereby all imaged tuft dendrites were required to fire an event with at least 50% of those events classified as a plateau. If this condition was met, the classification was applied to all tuft dendrites; otherwise, the event was labeled as an ordinary Ca²⁺ transient. Putative plateau potentials possessed markedly distinct waveforms (Figure S6E; Figure 3, G & H) that were large, long, and rare relative to Ca²⁺ transients (~1/12^th the frequency) (Figure 3, I–K). Consistent with in vitro recordings^{57, 61}, putative tuft plateau potentials elicited significantly stronger somatic responses (Figure 3, L & M) than did Ca²⁺ transients. Therefore, while our identified plateau potentials are necessarily putative in nature, we take their numerous correspondences to in vitro measurements, as well as the unbiased manner in which they were identified, as a strong indicator of external validity.

Distal tuft dendrites are reliably but variably recruited during somatic place field formation

Having developed approaches to promote and detect spontaneous somatic PF formation, and to detect plateau potentials, we first asked a simple question: what do distal tuft dendrites do while a CA1 PN forms a new somatic PF? Based on numerous lines of converging evidence, we hypothesized that distal tuft dendrites would generate plateau potentials to drive global dendritic Ca²⁺ influx. Consistent with this hypothesis, somatic PF formation-triggered tuft Ca²⁺ dynamics revealed that CA1 PN tuft dendrites are robustly activated during PF formation (Figure 4A). Distal tuft dendrites displayed larger than normal Ca²⁺ events during PF formation (Figure 4B) and with peri-formation activation levels that were tightly coupled to those of their soma (Figure 4C). However, we noted two apparent discrepancies between peri-formation tuft activity and their purported role of initiating PF formation via locally generated plateaus. First, the variability of tuft peri-formation activity (Figure 4, B & C) appeared high given the stereotyped depolarizations observed during plateau potential-mediated PF formation in CA1 PNs⁵⁷. Second, whereas novel context exposure robustly promoted somatic PF formation in this study (Figure 3, C & D) and in previous studies^80–82, it did not influence tuft plateau frequency (Figure S7). In fact, novelty was associated with a tonic decoupling of tuft dendrites from their soma and an increased rate of isolated tuft Ca²⁺ transients (Figure S8, A–C). Isolated tuft events were biased toward somatic PFs and predominantly occurred during post-formation laps. Therefore, isolated tuft events did not appear to predict future PF formation events (Figure S8D) as has been observed in CA1 PN basal dendrites⁸⁰. Taken together, these observations affirmed the engagement of tuft dendrites during PF formation under in vivo conditions. However, the variable magnitude and timing of their recruitment raised the intriguing possibility that PF-forming plateau potentials do not typically originate within the distal tuft compartment as previously thought^{40, 57, 61}.

Figure 4. Variable timing and magnitude of distal tuft dendritic activity during spontaneous place field formation.

(A) Somatic (grey) and tuft (red) peri-event time histograms (PETHs) triggered on somatic PF formation events. Shaded regions indicate SEM. Dashed line indicates estimated moment of PF formation (see supplementary materials and methods). (B) Gaussian kernel density estimates (KDEs) showing distributions of all Ca²⁺ transients (blue) compared to those occurring during PF formation for soma (top, grey) and tuft dendrites (bottom, red). (C) Scatterplot showing area under the curve (AUC) of peri-formation PETHs (as in A) for soma (y-axis) and tuft dendrites (x-axis). Black line represents ordinary least squares regression fit. Pearson correlation coefficient is shown. (D) Tuft plateau probability as a joint function of lap relative to PF formation (y-axis) and within-lap distance from somatic PF peak (x-axis). Cyan crosses indicate real data points and heatmap depicts joint Gaussian KDE. (E) Heatmap of tuft PETHs triggered on somatic PF formation events, i.e. data underlying red trace in (A). Each row represents the mean z-scored ΔF/F₀ of all imaged tuft dendrites for a given PF formation event in a single CA1 PN. (F) Histograms and Gaussian KDEs of tuft timing relative to soma (top) and tuft AUC (bottom) during PF formation. (G) Scatterplot showing tuft peri-formation AUC (y-axis) and tuft-soma lag (x-axis). Black line represents Gaussian KDE fit to histogram values (not shown) after binning AUC according to lag. (H) Joint probability distribution of peri-formation tuft AUC and lag relative to soma. Cyan crosses indicate true data points used to fit the heatmapped joint Gaussian KDE. ***P < 0.001. See table S1 for sample sizes and table S2 for additional statistical details.

To specifically probe the role of tuft-originating plateau potentials, as opposed to ordinary Ca²⁺ transients, in PF formation, we assessed their prevalence with respect to lap number and spatial location. This analysis revealed that very few PF formation events were accompanied by tuft plateaus (2 from 39 formation events and 44 plateau potentials within the same virtual context). Instead, tuft plateaus tended to occur after somatic PF formation and slightly forward in space relative to somatic PF peaks (μ = 20.01 ± 10.93 cm, σ = 74.92, N = 48 plateaus from 46 tuft dendrites) (Figure 4D). The occurrence of tuft plateaus after PF-associated somatic events may be explained by a variety of factors including ongoing local excitatory synaptic input, local tuft plasticity induced during somatic PF formation, and/or centrifugally-propagating somatic activity during PF traversals. We note that the forward-shifted nature of post-formation tuft plateaus relative to somatic PFs is reminiscent of the ‘peak shift’ characteristic of the BTSP mechanism (Figure 3F)³⁹, whereby somatic PFs emerge at locations the animal occupied 1–3 s prior to plateau initiation, and consistent with a previously suggested role for plateau potentials in boosting the gain of extant somatic PFs⁵⁷.

Given the surprising lack of association between tuft plateaus and PF formation, we more closely examined peri-formation tuft activity profiles across PF formation events. We observed considerable variability not only in the magnitude of peri-formation tuft activation (μ = 22.96 ± 1.47 area under the curve, σ = 9.064, N = 39 PF formation events), but its timing relative to the moment of peak peri-formation somatic activity (med. = −0.70 sec, IQR = 0.53, N = 39 PF formation events) (Figure 4, E & F). Remarkably, the distribution of tuft-soma timing lags was identical in both shape and time course to the currently unexplained asymmetric plasticity kernel characteristic of BTSP^{39, 42, 43}. The magnitude and timing of peri-formation tuft activity clustered together such that distal tuft dendrites tended to display the strongest activation within 1.5 seconds preceding somatic PF formation (Figure 4, G & H). These data are consistent with prior in vitro experiments indicating that properly timed cortical input is critical to drive plateau potentials in CA1 PN distal tuft dendrites^{57, 61}. However, the diffuseness of this clustering and the lack of tuft-associated plateau potentials during somatic PF formation show that CA1 PNs can accommodate a wide range of tuft activity patterns in generating new PFs.

In summary, while CA1 PN distal tuft dendrites appear to play a prominent role in PF formation, they do not appear to do so via locally generated plateau potentials. From these results, we conclude that (1) PF-forming plateau potentials originate elsewhere in the CA1 PN dendritic arbor (e.g. the apical trunk within stratum radiatum or near the nexus), (2) plateaus do not regeneratively propagate to drive global Ca²⁺ influx throughout CA1 PN dendritic arbor, and (3) given that tuft-associated plateaus nonetheless drive robust somatic responses (Figure 3, L & M), they primarily serve to boost the expression of existing somatic PFs.

The timing and magnitude of distal tuft recruitment jointly determine properties of new place fields

The prominent variability in tuft activity during PF formation prompted us to hypothesize that it may shape aspects of new PFs. We sought to (1) test our hypothesis analytically by devising a model to predict PF properties based on peri-formation tuft recruitment and (2) interrogate this model to understand the logic by which tuft dendrites might shape PF properties. Visually inspecting the timing and magnitude of peri-formation tuft recruitment relative to various properties of PF expression and formation revealed several apparent trends with varying degrees of nonlinearity (Figure S9). These trends presented a challenge as they complicated the use of easy-to-interpret linear regression models. Applying a nonlinear kernel to a linear model was inappropriate as we could not reasonably select and apply one polynomial order to both of our tuft features of interest (Figure S9). We opted for an agnostic approach, transforming both tuft recruitment features into 3^rd order polynomial feature space and allowing a regularized linear model to decide the relative importance of polynomial features corresponding to N^th order coefficients (Figure 5A).

Figure 5. Peri-formation tuft recruitment closely predicts properties of place field expression and formation.

(A) Schematic describing approach to predict somatic PF properties using features of peri-formation tuft recruitment. Tuft-soma lag and area under the curve (AUC) were transformed into polynomial feature space to train a double-cross-validated Ridge regression model (see supplementary materials and methods). Example result for PF width shown at right with predicted values using real (green) and shuffled (purple) data. Dashed line represents equality. (B) Tuft-based model prediction error for properties related to the expression and formation of new somatic PFs. Prediction errors from 1,000 models trained on shuffled data (purple) were normalized and compared to those from real data (green). Violin bodies span full data range with lines at medians and error bars spanning first to third quartiles. (C) Correlation matrix showing Pearson coefficients between target variables in (B). Grey cells indicate non-significant coefficients (P > 0.05). (D, E) Model predictions that significantly outperformed those of models trained on shuffled data, including PF width (D) and spatial information content (E). Top heatmaps in each panel show Ridge model coefficients by tuft feature (rows) and polynomial order (columns). Bottom heatmaps show model predictions across interpolated feature space from Ridge regression models trained on real data. *P < 0.05, ***P < 0.001. See table S1 for sample sizes and table S2 for additional statistical details.

The magnitude and timing of tuft recruitment during PF formation predicted the resultant width of new PFs and, in particular, their spatial information content (Figure 5, A & B). Peri-formation tuft dynamics did not predict hallmarks of the BTSP mechanism that have been previously established and independently reproduced^{20, 39, 43, 82}, including in this study (Figure 3, E & F). Among these hallmarks is a linear relationship between an animal’s running speed during PF formation and the width of the resulting PF (Figure 3E), presumably due to velocity-modulated CA3 inputs recruited a larger proportion of CA1 PN synapses within the seconds-long BTSP plasticity window. Notably, peri-formation tuft activity represents a novel determinant of PF width; neither the magnitude (R = −0.18, P = 0.26) nor the timing (R = 0.28, P = 0.083) covaried with peri-formation animal velocity. Finally, despite varying degrees of multi-collinearity among PF properties (Figure 5C), regression weights indicated that tuft recruitment predicts these properties according to unique rules. PF width was most strongly predicted by the timing of peri-formation tuft recruitment (Figure 5D) whereas PF spatial information was almost entirely predicted by the magnitude of tuft activation (Figure 5E). Exact regression weights are shown in table S3. In summary, this analysis identifies the timing and magnitude of tuft activity, surrounding the moment of somatic PF formation, as previously unappreciated determinants of the hippocampal spatial code.

Local distal tuft spatial tuning is robust and shifted in space relative to somatic tuning

Distal tuft dendrites shape the formation of new PFs while firing plateau potentials during subsequent PF traversals. However, it remains unclear (1) whether tuft dendrites themselves display local PFs and (2) how dendritic PFs hundreds of microns from the soma might relate to somatic PFs. For instance, CA1 PN radial oblique dendrites, located below the distal tuft region within stratum radiatum, display PFs with no apparent relation to their somatic PF²⁰. CA1 PN distal tuft dendrites undergo local, long-term potentiation in vitro⁸³ and receive spatially informative cortical inputs^{40, 47, 48, 55, 56, 58, 84} via clustered synaptic connections rich in plasticity-promoting N-methyl, D-asparate receptors (NMDARs)⁷². We therefore hypothesized that distal tuft dendrites express local PFs. We further hypothesized that, if tuft dendrites do express local PFs, then they may be formed in tandem with somatic PFs and would therefore be tuned to similar locations.

To examine distal tuft spatial tuning relative to somatic tuning, we analyzed data from single-context experiments (Figure 1, C & D) which provided sufficient numbers of laps (μ = 107.25 ± 12.94, σ = 34.24, N = 8 imaging sessions from 7 animals) for robust spatial tuning curve (TC) calculations in all cells (Figure S10). Consistent with our hypothesis, 90 ± 4.8% (σ = 13%, N = 46) of all imaged tuft dendrites displayed PFs (Figure 6, A & B) with similar properties to somatic PFs (Figure S11). Tuft dendritic PFs were systematically shifted backward in space relative to somatic PFs by 21.43 ± 11.77 cm (σ = 68.60 cm, N = 35 tuft-soma PF pairs) (Figure 6, A & C) with a heavy distribution tail at lower absolute distances (Figure 6D). We note that this backward shift is consistent with the tendency for distal tuft dendrites to fire before their soma during somatic PF formation (Figure 4F). Taken together, these result suggests that CA1 PN distal tuft dendrites form local PFs in the process of driving somatic PF formation and their tendency to fire before the soma systematically shifts their spatial tuning preferences backward in space.

Figure 6. Local spatial tuning in distal tuft dendrites of CA1 PNs.

(A) Somatic (black) and distal tuft dendritic (red) TCs. Violet shaded regions indicate PFs. (B) Histogram and Gaussian kernel density estimate (KDE) of fraction of tuft dendrites per cell expressing PFs. Dashed line indicates fraction of soma expressing PFs. (C) Histogram and Gaussian KDE of tuft dendritic PF peak location relative to that of soma. Dashed line indicates equivalent PF peak locations. (D) Vertical histogram of absolute circular distances separating soma and tuft PFs. 150 cm is the maximum possible distance along a 3-m virtual track. (E-F) Vertical histograms showing divergence (Wasserstein distance, W_p, see supplementary materials and methods) between all connected soma-tuft (E, N = 46 pairs from 9 cells) or tuft-tuft (F, N = 160 pairs from 9 cells) pairs. Shaded regions indicate 95% confidence interval (CI) generated from TCs calculated on shuffled data. By-cell Gaussian KDEs at right reflect within-cell variability. (G) Diagram of support vector classifier (SVC) ensemble-based approach to decode animal position from soma or tuft TCs (see supplementary materials and methods). (H) Decoding accuracy plotted against the number of tuft dendrites used for training. Circles represent tufts, Xs separated by dashed lines indicate connected soma, colors indicate cell identity. Black line represents ordinary least squares fit. (I) Tuft decoding accuracy plotted against soma decoding accuracy. Violet symbols indicate cell with somatic PF; grey symbols indicate no PF. (J) Tuft decoding outperformance relative to connected soma (log₂ ratio) plotted against number of tuft dendrites used for training. (K) Tuft spatial information score as a function of the fraction of its Ca²⁺ transients that were isolated. Colors correspond to cell identity. (L) Example spatial tuning heatmaps showing tuft ΔF/F₀ as a function of lap number (y-axis) and animal position (x-axis). Tuft (red) and soma (black) TCs are plotted above. Top: Example consistent with trend shown in (K). Bottom: Counterexample to trend shown in (K). (M-N) Soma spatial information score, as shown in K and L. *P < 0.05, ***P < 0.001. See table S1 for sample sizes and table S2 for additional statistical details.

Since soma and dendrites lacking statistically significant PFs nonetheless displayed subthreshold tuning preferences with apparent cross-compartment relationships (Figure S10), we next compared spatial TCs using pairwise Wasserstein distances. In this setting, Wasserstein distance describes the amount of ‘work’ required to transform one spatial TC into another (e.g. lower values indicate that two TCs are more similar to each other). Consistent with our hypothesis, somatic and distal tuft spatial tuning preferences were generally similar with few tuft TCs significantly differing from those of their soma (Figure 6E). The same was true between individual tuft dendrites of a given cell (Figure 6F), although we noted several exceptions to both of these trends (Figure S10; see also by-cell distributions in Figure 6, E & F). Taken together, these data show that distal tuft dendrites tend to display PFs back-shifted in space relative to somatic PFs, which is accounted for by their activity during somatic PF formation, but they nonetheless have the capacity to express fully autonomous spatial tuning preferences.

Given the non-negligible variability of tuft spatial tuning preferences across and within CA1 PNs, we next asked whether tuft dendrites locally encode spatial information beyond that which might be inherited from backpropagating somatic APs. To this end, we trained a support vector classifier ensemble to predict each animal’s position along the 3-m virtual track based on either tuft or somatic TCs from a single CA1 PN (Figure 6G). To be clear, this model was destined to perform poorly given that it was trained on either 1 soma or 2–11 tuft dendrites (see table S1). Rather than accurately predict animal position, our goal was to assess per-tuft spatial information content relative to the soma. This analysis revealed that tuft dendrites linearly sum to represent an animal’s environment such that the number of dendrites used for model training predicts the extent to which the distal tuft compartment outperformed the soma in decoding animal position (Figure 6, H–J). These results demonstrate that not only are distal tuft dendrites spatially tuned, but their diversity in spatial tuning preferences renders them collectively more spatially informative than the soma.

Finally, although tuft dendrites autonomously encode animal position, this does not appear to be their primary function. To the contrary, a tuft dendrite’s overall fraction of isolated Ca²⁺ transients negatively predicted its spatial information content (Figure 6K, see Figure 6L for example and counterexample). A similar but non-significant trend was observed at the soma (Figure 6, M & N). These data indicate that, despite the high prevalence of PFs in tuft dendrites (Figure 6B), and their cumulatively robust spatial information content (Figure 6, H–I), isolated tuft Ca²⁺ transients primarily encode non-spatial information. To better understand the information content of isolated tuft transients, we attempted to relate isolated tuft activity to a collection of locomotor and reward-related behavioral features using a cross-correlation-based approach (see supplementary materials and methods). Briefly, tuft ΔF/F₀ signals not corresponding to isolated Ca²⁺ transients were masked with zeros. As a control, the same was done for somatic signals not corresponding to a somatic Ca²⁺ transient. Most soma and tuft dendrites were significantly correlated to all features that we assessed: velocity, acceleration, deceleration, rewarded licks, and unrewarded licks (Figure S12, A & B) with variable time lags (Figure S12C). Isolated tuft events were more behaviorally informative than somatic events on the whole, with no one behavioral feature standing out (Figure S12D). In summary, distal tuft dendrites carry out diverse local operations hundreds of microns from their soma. They possess robust, local PFs that tend to be locally similar to each other, and globally back-shifted relative to the somatic PF, while still possessing potentially consequential diversity in their tuning preferences. Finally, while isolated tuft activity can clearly contribute to local spatial tuning, it primarily encodes non-spatial behavioral features.

DISCUSSION

By simultaneously monitoring somatic and distal tuft dendritic Ca²⁺ dynamics in vivo during two virtual spatial navigation tasks, we have uncovered several previously unappreciated aspects of CA1 PN distal tuft dendritic function. Given the lack of prior knowledge regarding distal tuft dynamics in CA1 PNs in vivo, our study is necessarily broad and descriptive. Here we wish to provide additional context to synthesize our findings and to reconcile key observations with the current model for BTSP-driven PF formation.

We find that distal tuft dendrites are strongly compartmentalized from their soma in CA1 PNs and favor centripetal propagation of dendrite-originating events. While strong compartmentalization was largely expected based on past in vitro and computational work^{60, 70–73, 77, 85}, we were surprised to find that compartmentalization broke down during immobility. CA1 PN soma are more active during locomotion and somatic depolarization results in better soma-tuft coupling in vitro due to A-type K⁺ channel inactivation⁷³. At the circuit level, velocity-modulated EC neurons transmit increased excitatory signals to tuft dendrites during running states⁵⁴. Indeed, this is consistent with our observation that tuft dendrites are far more active during running than immobility. Furthermore, EC-driven tuft activation should be boosted by clustered⁸⁶, NMDAR-rich⁷² EC-tuft synapses and depolarization-dependent inactivation of I_h currents^{70, 87–89}. These studies all point to increased soma-tuft coupling during locomotion.

However, it is also the case that somatic AP backpropagation favors single spikes rather than bursts⁹⁰. Inhibition undoubtedly plays a major role in soma-tuft cross talk as well, via either local disinhibition by long-range interneuron-targeting EC inhibitory projections⁹¹, depolarizing I_h tail currents^{75, 76, 92} following the release of velocity-modulated⁹³ dendritic inhibition^94–99, or some combination of both. Therefore, a mixture of inhibitory and cell-intrinsic mechanisms may dynamically modulate distal dendritic excitability as well as its impact on somatic AP firing. We speculate that one function of such gating may be to sparsen the CA1 population code by limiting the rate of PF formation: unchecked plasticity might otherwise bias the hippocampal cognitive map^{32, 33} to disproportionally represent regions first encountered rather than tiling the entirety of an animal’s environment.

Based on our unsupervised plateau detection strategy, which is robust by all measures, tuft dendrites do not form somatic PFs via locally-generated plateau potentials as reasonably postulated based on a substantial amount of converging albeit indirect evidence^{39–42, 57, 61}. To be clear, we do not argue against a role for dendritic plateau potentials in somatic PF formation. ADPs are a somatic vestige of dendritic plateau potentials and their presence during PF formation, even in the absence of experimental perturbation⁵⁷, indirectly indicates that they occur during at least a substantial portion of PF formation events. Instead, we suggest that properly timed input converging onto both distal tuft and radial oblique dendrites leads to plateau potential initiation elsewhere in the apical arbor; likely in the apical trunk if not near the nexus. One implication of this revised model is that plateau potentials may not uniformly depolarize the CA1 PN dendritic arbor such that all dendritic spines receiving presynaptic input within a second-long window around plateau onset will undergo plasticity³⁹. Intuitively, this is not a problem for the BTSP mechanism: one CA1 PN possesses thousands of dendritic spines and depolarizing a plurality of them may well be sufficient to drive robust PF formation.

Perhaps most intriguingly, we found that CA1 PN distal tuft dendrites exert analog, nonlinear control over the expression of new somatic PFs via the timing and magnitude of their peri-formation recruitment. Given recent evidence that BTSP comprises elements of long-term potentiation (LTP) as well as depression (LTD)^{42, 43}, one attractive possibility is that these features of peri-formation tuft recruitment may influence the balance between LTP and LTD to determine PF fidelity¹⁰⁰. We also noted that the distribution of tuft timing lags relative to their soma strikingly resembled the asymmetric, seconds-long temporal association window, or ‘plasticity kernel’^{39, 42, 43}, that renders BTSP particularly relevant to the timescales of action-outcome relationships encountered in the real world. Efforts are already underway to identify its molecular underpinnings^44–46. Here we propose that the BTSP plasticity kernel may be specifically tuned to accommodate variable tuft timing. Alternatively, since this kernel is an estimate based on many PF formation events and CA1 PNs, it may simply reflect the temporal variability of tuft recruitment during PF formation. More broadly, we view our results to strengthen the role of distal tuft dendrites in BTSP-mediated PF formation. By virtue of its variable peri-formation recruitment, this distal neuronal compartment may effectively set the gain, or proverbial importance, of new feature-selective receptive fields based on the nature of concomitant, top-down cortical signals.

Finally, tuft activity during somatic PF formation appears to drive local tuft plasticity that subsequently helps to maintain the new somatic PF. Tuft dendrites express local PFs that are back-shifted relative to their somatic PF. The distance of this backward shift is accounted for by the timing differential between tuft and somatic activation during somatic PF formation, indicating that tuft recruitment produces tuft PFs in parallel via local plasticity. As a result, tuft dendrites will be reactivated to a greater extent during subsequent, post-formation traversals of the CA1 PN’s somatic PF. We therefore propose that concomitantly-formed tuft PFs effectively increase the probability of tuft-associated plateau potentials during somatic PF traversals. Given that we found tuft plateaus to strongly drive somatic activity, this would appear to be an elegant mechanism by which tuft dendrites maintain newly formed somatic PFs. In conclusion, we find that distal tuft dendrites constitute an impressively multifunctional neuronal compartment capable of dynamically modulating somatic activity, representing numerous experiential features, and lastingly determining the relative import of new hippocampal receptive fields.