Peri-Head Distance Coding in the Mouse Brainstem

SUMMARY

Perceiving the location of nearby objects relative to the body is essential for guiding movement and avoiding danger. Many organisms use touch to estimate object distance, yet how somatosensory circuits extract distance from tactile inputs remains unclear. Here, we investigate how neurons in the mouse whisker brainstem compute peri-head distance. Using extracellular recordings in awake mice during wall-passing stimulation, we find two coding schemes that tile peri-head space: “proximity”, showing monotonically increasing activity as objects approach the face, and “map”, showing peaked tuning at specific distances. The map code enables more precise readout of peri-head distance. Perturbations suggest that long-range SpVi→PrV inhibition is important for generating the map code, subtracting heterogeneous peripheral signals from different whiskers or within the same whisker. These findings reveal an underappreciated computational role for brainstem circuits, where inhibition acts as a neural comparator to transform multiplexed peripheral inputs into a stable representation of peri-head distance.

Graphical Abstract

eTOC Blurb

Xiao & Severson et al. show that the mouse whisker brainstem transforms whisker-centered touch signals into a head-centered map of nearby space. Long-range inhibition within the brainstem converts peripheral inputs into a stable peri-head distance representation, revealing that early brainstem circuits actively compute where objects are relative to the head.

Spatial representations are essential for animals to interact effectively with the world, enabling behaviors such as judging predator distance, guiding reach-to-grasp movements, and navigating complex terrain with obstacles and gaps. To support such behaviors, the brain constructs multiple egocentric representations of space in which sensory signals are transformed into body-centered coordinate systems suitable for action^1–3. These transformations are especially important in peripersonal space (PPS), the space immediately surrounding the body, which serves as a critical interface for object interaction, social engagement, and defensive behavior⁴. Classic primate studies identified multisensory PPS neurons in premotor and posterior parietal cortex that respond both to tactile stimulation on specific body parts and to visual or auditory stimuli presented near those same body surfaces^5,6. Importantly, the visual or auditory receptive fields of these neurons shift with the corresponding body part, indicating that they encode space in body-part-centered coordinates rather than world-centered coordinates.

Recent work has emphasized that PPS is not a fixed geometric zone around the body, but a set of contact-related action fields shaped by interactions between the body and nearby objects, as well as by the actions required to facilitate or avoid contact⁷. Within this framework, body-centered distance is a fundamental component of PPS because it constrains the selection and execution of motor programs. For example, reaching to targets that are near versus far requires distinct patterns of arm and trunk muscle activation^8,9; likewise, crossing narrow versus wide gaps requires different biomechanical and postural adjustments during locomotion¹⁰.

However, how body-centered distance is encoded, and the circuit mechanisms that give rise to such representation, remain unclear. One coding scheme for body-centered distance is a monotonic proximity-based code, in which neural activity increases as external stimuli approach the body^6,11. However, the proximity-based code does not explicitly represent specific distances within PPS. Another possible representation is a map-like code, where each neuron is tuned to a particular body-centered distance. Such non-monotonic, distance-specific tuning could provide a neural substrate that explicitly encodes locations in PPS, potentially supporting rapid motor planning and action selection.

The rodent whisker system provides a natural model to study head-centered distance coding because the grid-like whisker array acts as a built-in spatial sensor, enabling tactile sampling of peri-head space at different distances from the head^12,13 (Figure 1A). Specifically, the whisker array is organized into rows and arcs whose lengths vary systematically along the anterior-posterior axis¹⁴. This organization raises the possibility that signals across the array could be integrated into a head-centered representation of peri-head distances from tactile inputs alone, without requiring concurrent visual or auditory cues¹⁵. Consistent with this idea, neurons in whisker primary somatosensory cortex (wS1) exhibit diverse tuning to wall distance during virtual corridor navigation¹⁶, including non-monotonic responses that peak at specific distances from the head. Such tuning can be interpreted as a map code, in which the peaks of tuning curves map distance from the head, like tick marks on a ruler. At the population level, these preferred distances span the range of whisker reach, effectively forming a head-centered representation of peri-head space.

Figure 1. Acute PrV recordings in a wall-passing paradigm.

(A) Conceptual illustration showing how whiskers map peri-head space in head-centered coordinates.

(B) Left, schematic of a hypothetical brainstem circuit to compute non-monotonic distance code. Two TG neurons (A and B, red) with different distance sensitivities project to the brainstem: neuron A provides excitatory input to a PrV output neuron (C, black), whereas neuron B synapses onto an inhibitory neuron (blue) in the trigeminal brainstem complex. Middle, tuning curves for neurons A and B; the shaded region indicates the difference between their distance-tuning profiles. Right, linear summation yields a non-monotonic tuning curve in neuron C, with a peak at an intermediate peri-head distance.

(C) Left, Schematic of the experimental setup. Head-fixed mouse on a treadmill with a wall moving from anterior to posterior past the mouse’s face. Wall distance from the face was randomly varied between 7–23 mm in 2-mm increments on each pass. Each recording contains 15–20 repetitions per distance. Right, Example video frame showing multiple whiskers contacting the wall (moving left to right). Red arrow indicates measured distance between face and wall.

(D) Example coronal brain section (4.8 mm posterior to Bregma) showing two silicon probe tracks through PrV labeled with DiI (red); cell bodies stained with Nissl (blue).

(E) Example extracellular voltage trace from one recording session. Colored bars indicate wall-passing periods.

By contrast, such a stable population-level peri-head distance code is not present in the periphery. Mechanosensory afferents in the trigeminal ganglion (TG), each of which has a single-whisker receptive field, transduce the mechanical stresses imposed on the follicle-sinus complex during object contact^17–21. Because whiskers are tapered elastic beams, contacts closer to the whisker base generate larger forces than contacts near the tip^22,23. Accordingly, touch-responsive TG neurons encode radial contact position primarily through monotonically increasing firing rate as contact occurs closer to the whisker base^24,25. However, these peripheral responses remain whisker-centered and often multiplexed with whisker kinematics and contact mechanics. Downstream circuits must transform this input into an invariant representation of peri-head distance, such as the map code observed in wS1¹⁶.

We hypothesize that this transformation occurs at the level of second-order neurons in the principal trigeminal nucleus (PrV). PrV receives convergent inputs from multiple TG mechanoreceptor subtypes, including rapidly and slowly adapting afferents^26,27. Its local circuitry includes glycinergic and GABAergic inhibitory interneurons^28,29. PrV also receives topographically organized long-range GABAergic input from the spinal trigeminal nucleus interpolaris (SpVi)^27,30. These circuit features, which support feedforward inhibition, make PrV a plausible site for transforming monotonic, whisker-centered peripheral signals into non-monotonic tuning for peri-head distance. Inspired by theoretical work on representational format transformations in spatial coding³¹, we reasoned that inhibition could enable such a transformation by subtracting proximity-based inputs with distinct mechanical sensitivities to generate non-monotonic distance tuning (Figure 1B). To test this, we performed in vivo electrophysiological recordings from PrV neurons in a wall-passing paradigm.

Results

A wall-passing paradigm to study peri-head distance encoding in the whisker brainstem

To evoke tactile sensation similar to that experienced during naturalistic wall-following behavior, we devised a wall-passing paradigm for awake, head-fixed mice on a linear treadmill²¹ (Figure 1C, left; STAR Methods). During each trial, a wall was positioned at a randomly selected lateral distance from the mouse’s face and moved anterior to posterior at constant velocity (10 mm/s), independent of locomotion. Each wall pass lasted approximately 14 seconds, with wall movement onset at t = 0 and whisker-wall contacts typically occurring between 3–11 seconds. Each distance was presented 15–20 times. Wall distance was measured from the wall to the C2 whisker follicle, with the wall aligned to the head midline (Figure 1C, right).

After habituation, mice exhibited predominantly passive whisker-wall interactions. To assess active retraction, we compared wall-induced whisker bending angles and curvature of a single row of whiskers in awake versus anesthetized conditions (see Supplemental Figure S1). Under anesthesia, the moving wall pushed whiskers in the retraction direction, producing larger absolute bending angles (Supplemental Figure S1D, right) and greater change in whisker curvature (Supplemental Figure S1H, right) at closer wall distances. Awake mice showed similar bending angle (mean difference: 2.28°) and curvature (mean difference: 1.66 × 10⁻⁵ mm⁻¹) changes, with no significant differences from the anesthetized condition across whiskers (Supplemental Figure S1F,J). This observation suggests that mice showed little active whisker retraction to avoid contact and instead allowed their whiskers to passively glide along the passing wall. Thus, wall-induced whisker bending increased approximately monotonically with decreasing wall distance, consistent with a predominantly proximity-based distance encoding in the periphery.

PrV contains both proximity and map codes for peri-head distance

We first examined PrV responses to the passing wall stimulus at varying distances using in vivo extracellular recordings with single tungsten electrodes in mice with only C-row whiskers intact (see Supplemental Figure S2; STAR Methods). Based on our single-unit isolation metrics (see STAR Methods; Supplemental Figure S3J–M), we isolated 46 units from 14 mice (see Table S1 for mouse metadata), including 18 units with single-whisker and 28 units with multi-whisker receptive fields within the C-row. We analyzed single-unit responses using peri-event time histograms (PETHs) by computing the averaged firing rates for wall contact (3–11 s) and baseline (−4.5–0 s) across trials. To determine wall responsiveness, we applied one-tailed Wilcoxon signed-rank tests (see STAR Methods). Of the 46 units, 41/46 (89.1%) were wall-activated and 3/46 (6.5%) were wall-suppressed; one unit showed both activation and suppression at different distances, yielding a total of 43/46 (93.5%) wall-responsive (presumably touch-sensitive) units. To assess distance tuning among the wall-responsive units, we performed two-tailed Wilcoxon signed-rank tests across all distance pairs (p < 0.05, corrected using the Benjamini-Hochberg false discovery rate). Wall-responsive units with significant differences in mean firing rates for at least one distance pair were classified as “distance-tuned” (i.e., distance-selective; 90.7%, 39/43). For each wall-responsive and distance-tuned unit, we computed a z-scored tuning curve (see STAR Methods). The majority of wall-responsive units (55.8%, 24/43) showed a broadly monotonic, inversely proportional relationship with wall distance (e.g., Supplemental Figure S2D). Interestingly, a distinct subset (14.0%, 6/43) showed peak responses at intermediate or far wall distances (e.g., Supplemental Figure S2E), suggesting a heterogeneous representation of peri-head distance within PrV (see Supplemental Figure S2F).

Because high-impedance tungsten electrodes preferentially sample neurons with large-amplitude extracellular spikes, we next used 32-channel single-shank silicon probes to capture broader PrV population activity during the wall-passing paradigm (Figure 1C–E; see STAR Methods). We isolated 1,141 single units across 36 mice (Supplemental Figure S3A–E). Of the 1,141 units, 67.3% (768/1141) were significantly modulated by the passing-wall stimulus and classified as “wall-responsive”. Among these wall-responsive units, 532/768 (69.3%) showed significant activation at one or more distances, 333/768 (43.4%) showed significant suppression, and 97 units met both criteria (activated at some distances and suppressed at others). Given that our probe spanned the entire dorsoventral axis of PrV (Figure 1D), while whisker-responsive neurons are localized to ventral PrV³², the prominence of unresponsive units in dorsal PrV, likely corresponding to unstimulated microvibrissa-, jaw-, or lip-related neurons, is not unexpected (see Supplemental Figure S4I for locations of responsive vs unresponsive units along the silicon probe).

Representative rasters, PETHs, and z-scored tuning curves from six wall-responsive units illustrate diverse responses profiles (Figure 2A–C): Unit 1 showed monotonic increases in firing with decreasing wall distance (“proximity” tuning); Units 2–4 exhibited peak firing at specific intermediate distances (“map” tuning); Unit 5 displayed a complex response profile with both suppression and activation depending on wall distance; and Unit 6 was suppressed during wall contact.

Figure 2. PrV contains multiple neural codes for peri-head distance.

(A) Example spike rasters of six PrV single units during repeated wall passes. Each row is one wall pass, and each dot indicates a 50 ms bin containing at least one spike, colored by wall distance. Wall passing begins at time 0. Wall contact (gray bar) begins between 2–3 s and ends between 9–11 s, depending on wall distance and whisker length. Scale bar: 20 passes.

(B) Peri-event time histograms (PETHs) computed as spike density functions (SDFs) aligned to wall onset time (t = 0 s), for the same example units shown in A. Error shading: SEM.

(C) Z-scored tuning curves (mean ± 95% CI) to wall distance during the wall-passing period (3–11 s) for the same example units shown in A, B.

(D) Heatmap of normalized tuning curves for wall-activated, distance-tuned units (n = 472 from 36 mice), sorted by preferred wall distance from close (top) to far (bottom). Tuning curves were computed during the wall pass period (3–11 s) and linearly interpolated between presented distances; unpresented distances are shown as white space.

(E) Histogram of preferred wall distances, measured as the location of the tuning curve peak, from units in D.

(F) Fraction of wall-responsive units (n = 768) in PrV defined as either untuned (gray), suppressed only (purple), map (yellow, prefer ≥ 8 mm), proximity (pink, prefer < 8 mm), or ambiguous (orange), which could not be statistically defined as either proximity or map (see Figure S4H).

Across the full wall-contact period (8 s), we identified 656 out of the 768 (85.4%) wall-responsive units that were significantly tuned to wall distance. Within this distance-tuned population, 472/656 (72.0%) showed significant activation relative to baseline at one or more distances (hereafter referred to as the activated population), whereas 213/656 (32.5%) showed significant suppression at one or more distances (hereafter referred to as the suppressed population). A small subset of units (29/656; 4.4%) exhibited both activation and suppression across different distances and were therefore included in both populations.

At baseline prior to wall movement onset, units in the activated population showed lower firing rates than untuned units (5.46 ± 0.670 vs. 6.93 ± 0.947 sp/s; two-tailed Kolmogorov-Smirnov (K-S) test: p = 2.47 × 10⁻⁷), whereas units in the suppressed population had baseline firing rates comparable to untuned units (7.31 ± 1.124 vs. 6.93 ± 0.947 sp/s; two-tailed K-S test: p = 0.63; Supplemental Figure S4A). During wall-passing period, the activated population increased firing relative to baseline (16.44 ± 1.640 vs. 6.36 ± 0.704 sp/s; one-tailed K-S test: p = 9.26 × 10⁻²⁸; Supplemental Figure S4B, left), while the suppressed population exhibited a significant decrease in firing (4.47 ± 1.000 vs. 6.19 ± 1.076 sp/s; one-tailed K-S test: p = 3.06 × 10⁻⁴; Supplemental Figure S4B, right). Untuned units showed no significant change between baseline and wall-passing periods (6.93 ± 0.947 vs. 6.32 ± 0.953 sp/s; two-tailed K-S test: p = 0.53; Supplemental Figure S4C), validating our classification criteria.

To visualize distance tuning at the population level, we concatenated normalized tuning curves from the activated and suppressed populations, sorting units by preferred wall distance separately (Figure 2D; Supplemental Figure S4D). In both populations, while a large proportion of units preferred the closest wall position (7 mm), others were best tuned to intermediate or far distances (Figure 2E, Supplemental Figure S4G), revealing coexistence of proximity and map-like coding schemes in PrV.

To evaluate the tuning stability across repeated stimulus presentations, we performed 50% trial cross-validation on the activated and suppressed populations, separately: units were sorted by their preferred wall distance using one randomly selected half of wall passes, and this sorting order was then applied to the remaining half. The resulting distance-tuning heatmaps of the activated population revealed little changes in the preferred distance between cross-validated partitions (Supplemental Figure S4E). In contrast, the suppressed population showed partial reproducibility, where only proximity-preferring responses were stable, while map-like profiles lacked consistent structure (Supplemental Figure S4F). Based on this difference in tuning stability, we focused subsequent categorization of tuning curves on the activated population. Because suppressed responses could still contribute to downstream readout of peri-head distance, as they were observed in wS1¹⁶, we explicitly tested them in the decoding analyses presented later (Figure 3).

Figure 3. The map code enables more accurate decoding of peri-head distance.

(A) Decoding accuracy over time averaged across distances for maximum correlation coefficient-based population decoders trained on subsamples of 118 units from proximity (pink, n = 266), map (yellow, n = 118), map + proximity (green dash, n = 36 map units, 82 proximity units; proportions match the map:proximity ratio in Figure 2F), suppressed-only (purple, n = 119), untuned (black, n = 177), or trial-shuffled control (gray, n = 384) subpopulations. Shaded area: 95% confidence interval (CI). Wall contact occurs around 3–11s (gray bar) after the start of wall movement at t = 0 s.

(B) Population confusion matrices showing proportion of wall distance decoded for each actual distance from proximity, map + proximity, map, and suppressed-only groups. Rows indicate true wall distance and columns indicate predicted distance.

(C) Left, Bar plot of population decoding accuracy across distances (mean ± SEM) for proximity (pink), map + proximity (green), map (yellow), suppressed only (purple), untuned (gray), and shuffled control (white); *, p < 0.05, Wilcoxon rank-sum test with BHFDR correction. Right, Bar plot of decoding error (mm; mean ± SEM). Decoding error was computed by multiplying the confusion matrix by the absolute value of the error matrix (see Figure S5A) and summing across rows. Numerical accuracy/error values and associated p-values for each population group are reported in Tables S2, S3.

To further characterize the activated population, we compared mean firing rates at distances greater than or equal to 9 mm against firing at the closest distance of 7 mm (one-tailed Wilcoxon signed-rank tests, see STAR Methods). Units whose firing peaked at the closest distance (7 mm) were defined as “proximity units” (e.g., Supplemental Figure S4H, left), and units with significantly higher firing at any farther distance (≥ 9mm) were categorized as “map units” (e.g., Supplemental Figure S4H, right). Under this definition, some proximity units could show reduced firing at unsampled distances (<7 mm) with map-like tuning outside the tested range. We therefore use “proximity” and “map” as operational categories that describe the diversity of peri-head distance tuning within the sampled, ethologically relevant peri-head distance range, rather than strict functional classes across all possible distances.

We identified 266 proximity units (34.6%) and 118 map units (15.4%) out of the 768 wall-responsive units (Figure 2F), which were preferentially located in ventral PrV (Supplemental Figure S4J). An additional 88 units (11.5%) were distance-selective but did not meet criteria for either map or proximity and were labeled “ambiguous” (e.g., Supplemental Figure S4H, middle) and excluded from further analysis. Suppressed-only units (i.e., distance-tuned units that show suppression at ≥1 distance and never show activation at any distance) accounted for 119 out of 768 wall-responsive units (15.5%; Figure 2F).

In addition to responses sustained across the entire wall pass, many PrV units showed transient responses at the onset and/or offset of the passing-wall stimulus (e.g., Unit 5 in Figure 2A,B). To quantify these temporal dynamics, we segmented the wall-passing period into onset (2.5–4.5 s), sustained (4.5–9.5 s), and offset (9.5–11.5 s) epochs (Supplemental Figure S4K) and applied our tuning classification to each temporal window (Supplemental Figure S4L, see STAR Methods). Map units were most prevalent during the sustained period among wall-responsive units (15.9%, 122/768), but were also present during the onset epoch (10.8%, 83/768). In contrast, offset responses yielded fewer map units (4.4%, 34/768). A subset of units exhibited shifts in their preferred wall distance across temporal windows, revealing dynamics in peri-head distance tuning between transient and sustained periods of the passing-wall stimulation (Supplemental Figure S4M).

Together, these results illustrate that PrV robustly encodes peri-head distance using both a proximity code, characterized by increasing firing at close distances, and a map code, peaking at intermediate and far distances in peri-head space. Though map cells are more prominent in wS1¹⁶, their presence in PrV revealed representational diversity and pointed to an underappreciated role for the whisker brainstem in active sensory processing beyond simple relay.

The map code in PrV allows more accurate readout of peri-head distance

We next asked whether proximity or map code allows more accurate readout of peri-head distance from neural population activity. We trained maximum correlation coefficient (MCC) decoders (see STAR Methods) on activity from either the proximity or map subpopulations. For comparison, we included the suppressed-only subpopulation (i.e., distance-tuned units that show suppression at ≥1 distance and never show activation at any distance), as well as untuned and shuffled subpopulations as controls. Because downstream circuits likely integrate information from both encoding schemes, we also evaluated a combined “map + proximity” subpopulation at the observed ratio in PrV (proximity:map = 2.3:1; Figure 2F).

As expected, decoding accuracy during the wall-contact period was higher in the map, proximity, suppressed-only, and combined “map + proximity” groups compared to shuffled and untuned controls (Figure 3A). During the sustained wall-passing period, the map group achieved the highest overall decoding accuracy averaged across all nine distances (71.5%), outperforming the proximity (48.9%) and combined “map + proximity” (58.6%) groups. The suppressed-only group had an average overall decoding accuracy of 32.6%. Untuned units performed slightly above chance (15.7%) especially during the transient periods due to residual distance tuning that fell below our stringent statistical criteria for wall-tuning classification.

Confusion matrices revealed complementary distance-specific decoding patterns across groups. The map-only population was most accurate at decoding intermediate and far distances, while the combined “map + proximity” population showed the strongest performance at the two closest distances (7 and 9 mm), followed closely by the proximity population (Figure 3B). For most intermediate distances, the decoding accuracy of the combined group fell between that of the map and proximity populations. To quantify these differences, we compared classification accuracy across nine wall distances for each group (Figure 3C, left; Table S2). At the closest distances (7 and 9 mm), the combined “map + proximity” and proximity populations achieved the highest accuracy. In contrast, the map population was most accurate at decoding more distal wall positions (11–23 mm).

To assess decoding precision in millimeters, we computed distance-classification error (in mm) by multiplying the confusion matrix with a distance error matrix (Supplemental Figure S5A). While classification error generally increased with distance, proximity, map, and the combined “map + proximity” groups showed submillimeter precision when decoding the two closest distances (7 and 9 mm; Figure 3C, right; Table S3). Only the map population achieved submillimeter error at the intermediate distances of 13 and 15 mm. These results are consistent with the performance of single-neuron decoders, although their overall decoding accuracy is lower than that of population decoders (see STAR Methods, Supplemental Figure S5B,C; Table S4). Importantly, these findings were robust to the choice of decoder (see STAR Methods, Supplemental Figure S5D,E, and Tables S2,S3 for SVM classifier).

PrV map tuning can persist in some neurons after reduction to single-whisker input

The presence of a stable map code in PrV suggests that peripheral, whisker-centered inputs are transformed and stabilized into a head-centered representation of peri-head distance. Because the moving wall contacts multiple whiskers of different lengths, and PrV neurons can have multi-whisker receptive fields³³, we asked whether non-monotonic map tuning depended on multi-whisker integration. Within the same recording session, we measured PrV single-unit activity, first with the mouse’s full whisker array, and next after acutely cutting all but one whisker, thereby eliminating multi-whisker inputs and restricting wall-evoked sensory input to the spared whisker (STAR Methods; Figure 4A). Because receptive fields were not systematically mapped for our silicon probe recordings, we could not spare the whisker to match the principal whisker of each recorded neuron. Continuous recordings nonetheless allowed us to follow the same units across conditions and compare each neuron’s distance tuning before and after whisker cutting.

Figure 4. Multi-whisker input supports but is not required for the distance map code.

(A) Top, Example video frame showing multiple whiskers during wall-passing stimulation (wall moving left to right). Middle, same session during whisker cutting. Bottom, same session after cutting to a single B1 whisker. Note that the spared whisker could not be matched to each recorded unit’s principal whisker because receptive fields were not systematically mapped during the silicon probe recording.

(B) Pairwise comparisons of firing rates (log scale) during wall-passing periods for activated units (n = 86 from 18 mice) in the full-array versus single-whisker condition. Lines connect the same unit across conditions. **, Population mean firing rates: 11.58 ± 2.986 sp/s (full array) vs. 8.18 ± 2.602 sp/s (single whisker), p = 4.43 × 10⁻⁸); *, individual neuron, p < 0.05; Wilcoxon signed-rank tests.

(C) Pairwise comparison of preferred wall distances for wall-activated units (n = 33 units from 13 mice) that remained distance-tuned after cutting down to a single whisker. Lines connect the same unit across conditions. Open circles: map units; closed circles: proximity units. Color indicates tuning category: proximity (pink), ambiguous (orange), map (yellow).

(D,E) Two example map units that converted to proximity units after cutting down to a single whisker (B1, top; γ, bottom): (C) SDFs; (D) corresponding tuning curves.

(F,G) Example map unit that retained map tuning but shifted its preferred wall distance closer after cutting down to a single C2 whisker.

(H,I) Two example map units showing gain modulation after cutting down to a single whisker (C1, top; β, bottom): Top, decreased activity; Bottom, increased activity.

(J,K) Two proximity units showing similar gain modulation effects: Top, decreased activity; Bottom, increased activity.

Across the recorded population (n = 195 units from 18 mice), acute whisker cutting did not significantly alter baseline firing rates, measured before wall-movement onset for each wall pass (full array: 7.09 ± 1.034 vs. single whisker: 7.10 ± 1.038 sp/s; Wilcoxon signed-rank test: p = 0.70). In contrast, firing rates of wall-activated, distance-tuned units (activated population; n = 86) during wall-passing periods were significantly reduced following cutting down to a single whisker (full array: 11.58 ± 1.338 vs. single whisker: 8.18 ± 2.602 sp/s, p = 4.43 × 10⁻⁸; Figure 4B). In many cases, wall-evoked responses were nearly abolished after cutting (<1 sp/s), consistent with loss of the neuron’s principal whisker. Among the 86 activated units, 66 showed significant changes in firing rates during wall-passing periods between cutting conditions (two-sided Wilcoxon signed-rank test, p < 0.05). Of these 66 units, 52 exhibited a significant decrease in firing rates, consistent with loss of excitatory input from whiskers within their receptive fields, whereas 14 units displayed increased firing rates, likely reflecting a reduction in inhibition following removal of surround-whisker inputs³⁰.

Of the 21 units categorized as map cells in the full-array condition, 10 remained significantly tuned to wall distance after cutting down to a single whisker. Within this subset, whisker cutting reshaped map-like responses in multiple ways (Figure 4C). Two units shifted their preferred distance to the closest distance (Figure 4D,E), effectively converting from map cells to proximity cells. In contrast, six units shifted their preferred distance while retaining map tuning (e.g., Figure 4F,G), and two additional units preserved their preferred distance but changed in response gain (Figure 4H,I). Thus, reducing sensory input to a single whisker produced heterogeneous effects on map cells, including loss of map tuning in some neurons and preservation of map-like responses in others, suggesting that multi-whisker input is not strictly required for map-like responses. Thus, the heterogeneous inputs in our proposed circuit model (Figure 1B) could, in principle, arise either across whiskers or within a single whisker.

We also examined how proximity cells were affected by the removal of multi-whisker input. Of the 45 units categorized as proximity cells before whisker cutting, only 15 remained distance-selective after cutting down to a single whisker. Of these 15 units, nine remained proximity cells with reduced response gain (e.g., Figure 4J,K, top), four remained proximity cells with increased response gain (e.g., Figure 4J,K, bottom), and two shifted to exhibit map tuning.

We next asked how acute whisker cutting affects population readout of wall distance from proximity and map populations. Using the same population decoder analysis as in Figure 3, we trained MCC decoders on units classified in the full-array condition as map (n = 21) or proximity (n = 21, randomly subsampled from 45 proximity units; see STAR Methods). Whisker cutting reduced overall decoding accuracy in both populations (map: 31.0 ± 0.10% vs. 21.5 ± 0.08%; proximity: 24.1 ± 0.82% vs. 20.2 ± 1.22%; mean ± 95% CI for full-array vs. single-whisker conditions), with the largest decreases occurring at intermediate and far distances in the map population (Supplemental Figure S5F,G). Decoding at the closest distances was comparatively less affected. Because the principal whisker of many recorded units was unavoidably cut, this manipulation reduces both multi-whisker inputs and total excitatory drive. We therefore interpret the decline in post-cut decoding performance as reflecting the combined effects of reduced sensory input and disruption of population-level distance readout.

Inhibition contributes to the PrV map code

We next turned to our proposed circuit model’s defining operation: subtractive integration (Figure 1B). We hypothesized that inhibitory circuitry contributes to this subtraction and tested this prediction using in vivo electrophysiology combined with optogenetic or chemogenetic silencing. The intrinsic circuitry of PrV contains two major sources of inhibition: local glycinergic and GABAergic interneurons within PrV itself, as well as profuse long-range GABAergic projections from the interpolar part of the spinal trigeminal nucleus (SpVi)^29,30. To assess the role of local inhibitory neurons in encoding peri-head distance within PrV, we unilaterally injected AAV-DIO-GtACR^34,35 into PrV of vGat-Cre mice (N=2), enabling Cre-dependent expression of the inhibitory opsin GtACR in PrV inhibitory neurons (Supplemental Figure S6A). We then performed acute optogenetic silencing of these local inhibitory neurons while recording single-unit activity in PrV during the wall-passing paradigm (Supplemental Figure S6B,D). For each wall distance, the “Laser OFF” and “Laser ON” conditions were randomly interleaved. Each distance was repeated 15–20 times in each condition to allow the comparison of distance tuning profiles with and without local inhibition.

Among the 112 well-isolated units, 18 were putative inhibitory neurons expressing GtACR, as they were strongly inhibited during the “Laser ON” period (modulation index ≤ −0.5; see STAR Methods; Supplemental Figure S6C,E). Of distance-tuned units (71/112), 17 were opto-inhibited, exhibiting a decrease in firing rates during laser stimulation (modulation index < 0; see STAR Methods), while preserving their distance tuning profile (i.e., proximity or map, data not shown). Sixteen units showed a reduction in inhibition under the “Laser ON” condition (modulation index > 0; see STAR Methods). Nonetheless, these opto-disinhibited neurons, regardless of whether they were proximity or map cells, retained their preferred distance (e.g., Supplemental Figure S6F,G). These observations suggest that acute suppression of PrV local inhibitory neurons primarily affects response gain, without alerting distance preference of individual PrV neurons.

We next investigated the contribution of long-range inhibitory inputs from SpVi to PrV. Because of the extensive size of SpVi, optogenetic silencing of its inhibitory neurons using a single optical fiber is impractical. Instead, we employed an AAV strategy to achieve sustained silencing by expressing tetanus toxin light chain (TeLC)³⁵ to chronically block synaptic transmission. AAV-DIO-TeLC-EYFP was unilaterally injected into SpVi of vGat-Cre mice (N=11) to target inhibitory neurons known to send long-range projections to PrV (Figure 5A; Furuta et al., 2008). Post-hoc histology confirmed TeLC-EYFP expression in SpVi (Supplemental Figure S6H), as well as labeled axon terminals and boutons in PrV originating from SpVi inhibitory neurons (Figure 5B). We note that the large size of SpVi limited viral access to all PrV-projecting neurons, thereby constraining the extent of silencing.

Figure 5. Silencing internuclear inhibitory projections from SpVi diminishes the PrV map code.

(A) Viral-genetic strategy to express TeLC in inhibitory cells in caudal SpVi, which send long-range projections to PrV.

(B) Example coronal section at 4.6 mm posterior to Bregma, showing axons of TeLC-labeled inhibitory cells (TeLC-EYFP, green) in PrV and a silicon probe recording track labeled with DiI (red). Cell bodies were stained with Nissl (blue).

(C) Heatmap of normalized tuning curves for wall-activated, distance-tuned PrV units in vGat-Cre mice following TeLC-mediated silencing of inhibitory neurons in SpVi known to project to PrV (n = 252 from 11 mice). Units are sorted by preferred wall distance, from near (top) to far (bottom). Tuning curves were linearly interpolated between presented wall distances; unpresented distances appear as white space. The color bar to the right indicates each unit’s tuning category: proximity (pink), ambiguous (orange), map (yellow).

(D) Cumulative distributions of preferred distances for the same units in C, comparing wild-type (black) and SpVi-TeLC (green) populations (one-tailed K-S test: p = 8.02 × 10⁻⁵).

(E) Comparison of tuning categories among wall-responsive units in wild-type population (n = 768 from 36 mice) and in SpVi-TeLC population (n = 395 from 11 mice). TeLC-mediated silencing of PrV-projecting inhibitory neurons in SpVi significantly reduced the prevalence of map units (yellow; 5.8 vs 15.4%; one-tailed Z-test: Z = 4.72, p = 1.20 × 10⁻⁶) and suppressed-only units (purple; 10.9 vs 15.5%; one-tailed Z-test: Z = 2.15, p = 0.02), but significantly increased the prevalence of proximity units (pink; 47.9 vs 34.6%; one-tailed Z-test: Z = −4.33, p = 7.50 × 10⁻⁶). The proportions of untuned (gray; 25.3 vs 23.0%; two-tailed Z-test: Z = −0.86, p = 0.39) and ambiguous (orange; 10.1 vs 11.5%; two-tailed Z-test: Z = 0.62, p = 0.53) units were not significantly different. *, p < 0.05, significant; n.s., not significant.

We recorded activity in PrV 10–15 days after TeLC injection during the wall-passing paradigm and isolated 587 single units (Supplemental Figure S3F–I). Chronic TeLC expression in SpVi did not significantly alter whisker kinematics during wall passes, as measured by whisker array midpoint angle (mean difference across distance = 0.84°; two-tailed Wald’s t-test: t = 1.02, p = 0.31; Supplemental Figure S6I). We then compared distance tuning in SpVi-TeLC and wild-type mice to assess the effect of suppressing long-range inhibitory inputs from SpVi. Comparison of normalized tuning curves of all wall-activated, distance-tuned units revealed a reduction in the proportion of map units in SpVi-TeLC mice (Figure 5C) relative to wild-type mice (Figure 2D), shifting the population to prefer closer distances (Figure 5D). Tuning classification across all wall-responsive units confirmed a significant change in the distribution of tuning types (Figure 5E). In particular, the proportion of map units was nearly threefold lower in SpVi-TeLC mice than in wild-type mice (5.8% vs 15.4%; Z-statistic: 4.721, p = 1.2×10⁻⁶), whereas the proportion of proximity units increased (47.9% vs 34.6%; Z-statistic: −4.33, p = 7.5×10⁻⁶). Our silencing manipulation did not fully abolish map cells, possibly because not all long-range inhibitory internuclear projections were silenced. The proportion of suppressed units decreased modestly (10.9% vs 15.5%; Z-statistic: 2.15, p = 0.02), whereas the proportions of untuned (25.3% vs 23.0%; Z-statistic: −0.86, p = 0.39) and ambiguous units (10.1% vs 11.5%; Z-statistic: 0.62, p = 0.53) remained unchanged.

During both baseline and wall-passing periods, untuned units in SpVi-TeLC mice exhibited significantly higher firing rates than those in the wild-type population, consistent with loss of long-range inhibitory inputs from SpVi (baseline: 10.01 ± 2.431 vs 6.93 ± 0.947 sp/s, p = 3.98 × 10⁻⁵; wall passes: 9.64 ± 2.772 vs. 6.32 ± 0.953 sp/s, p =9.64 × 10⁻⁴; one-tailed K-S tests; Supplemental Figure S6J). In contrast, distance-tuned units in SpVi-TeLC mice showed baseline activity comparable to that of wild-type mice in the absence of wall contact (5.60 ± 0.845 vs. 5.46 ± 0.670 sp/s; one-tailed K-S test: p = 0.14), but exhibited significantly greater wall-evoked firing (9.96 ± 1.209 vs. 9.02 ± 1.005 sp/s; one-tailed K-S test: p = 9.34 × 10⁻³ Supplemental Figure S6K). This enhancement was distance dependent, with the largest increase at the closest wall distance (Supplemental Figure S6L, bottom; Table S5). Thus, SpVi inhibition appears to be strongest at close distances for distance-tuned units, rather than acting as uniform suppression across distances seen in the untuned population (Supplemental Figure S6L, top). These findings suggest that chronic TeLC-mediated silencing of SpVi inputs was associated with a shift in tuning preference toward closer wall distances, while leaving spontaneous activity of distance-tuned units largely unchanged. Thus, long-range inhibition from SpVi appears to contribute critically to the transformation of feedforward proximity-based inputs from TG into a map-like peri-head distance representation in PrV.

Finally, to illustrate how the SpVi-PrV inhibitory circuit could support the proximity-to-map transformation, we implemented a model constrained by known anatomical motifs (Figure 6A; see STAR Methods). In the model, TG afferents send proximity-based inputs to neurons in PrV via two pathways: a direct excitatory pathway and an indirect inhibitory pathway mediated by SpVi inhibitory neurons. TG inputs were modeled as sigmoidal functions with thresholds skewed toward proximal distances (Figure 6B), and each modeled PrV neuron received sparse, random combinations of excitatory TG inputs and inhibitory SpVi inputs. This simple feedforward architecture transformed predominantly monotonic inputs into a heterogeneous population of PrV tuning curves, including both proximity-like and map-like tuning profiles (Figure 6C). Across the modeled population (n = 1225), preferred distances were broadly distributed, reproducing key features of the experimental data (Figure 6D–F, left). When long-range inhibitory input to each PrV neuron was reduced by 50% to simulate TeLC-mediated silencing of SpVi inhibition, the model showed a loss of map units and a corresponding expansion of proximity units (Figure 6D–F, right). This simulated manipulation also recapitulated the experimental population-level changes in tuning curve shape (Figure 6G–I; see STAR Methods): tuning width changed relatively little, tuning amplitude showed a modest rightward shift, and monotonicity exhibited a pronounced leftward shift toward more proximity-like responses. Together, these simulations show that a simple feedforward circuit motif is sufficient to transform largely monotonic sensory signals into a non-monotonic, map-like code capable of supporting an abstract representation of peri-head space, and that inhibition from SpVi is important for generating intermediate- and far -preferring distance tuning in PrV.

Figure 6. Simulation of brainstem circuitry to generate distance-map code.

(A) Schematic of the modeled neural circuit. The input layer consists of a pool of TG neurons whose tuning to wall distance was modeled as sigmoidal functions with varying thresholds, indicated by the outline color, as in B. TG neurons branch and project to PrV neurons via two distinct pathways: a direct excitatory pathway and an indirect inhibitory pathway (blue), mediated by inhibitory neurons in SpVi (blue circles). For illustration, PrV neurons are colored by tuning class: untuned (gray), proximity (pink), and map (yellow).

(B) Top, TG tuning curves (color-coded to match TG neurons outlined in A) are modeled as sigmoidal functions with thresholds spanning the full range of wall distances. Bottom, The sigmoid threshold distribution is exponentially skewed to emphasize the encoding of proximal wall distances.

(C) Example tuning curves of modeled PrV neurons. Y-axis: baseline-subtracted activity predicted by the model (A.U.). Wall distances are normalized between 0 and 1, and nine uniformly spaced distance bins (labeled 0–8) are used for analysis and visualization.

(D) Left, Heatmap of normalized tuning curves for distance-tuned units (n = 590) in the modeled PrV population, sorted by preferred distance bin from close (top) to far (bottom). Right, Same analysis after reducing inhibitory inputs to each PrV neuron by 50%, n = 636 distance-tuned units.

(E) Left, Histogram of preferred distance bins, defined as the peak (max) of the tuning curve, for modeled PrV neurons. Right, Same analysis after reducing inhibitory inputs to each PrV neuron by 50%.

(F) Left, Proportion of modeled PrV neurons classified as proximity (pink, prefer distance bin 0) or map (yellow, prefer > distance bin 0). Right, Same classification following 50% reduction of inhibitory inputs per modeled PrV neuron.

(G) Left, CDFs of tuning curve widths for neuronal populations recorded experimentally in wild-type mice (black) and SpVi-TeLC mice (green). Right, Corresponding CDFs generated by the model. Inset, Illustration of full width (gray arrow) at half maximum (dashed line).

(H,I) Same as G, but for tuning curve amplitude (H) and tuning curve monotonicity (I).

DISCUSSION

Here, we show that the trigeminal brainstem complex transforms multiplexed whisker-centered inputs into a head-centered representation that supports accurate peri-head distance readout. Using electrophysiological recordings from awake, head-fixed mice in a naturalistic wall-passing paradigm, we identified two distance coding schemes in PrV: a monotonic proximity code and a distance-selective map code. Both were robust across repeated presentations of the same wall distance despite trial-to-trial variability in the underlying multiplexed whisker inputs. Mechanistically, we show that long-range SpVi→PrV inhibition is essential for generating the map code, likely by enabling comparisons across heterogeneous TG signals from either different whiskers or within one whisker. While prior research on whisker-based object localization has focused mainly on TG afferents^24,25 and cortex^16,36, our findings revealed an anatomically defined brainstem circuit that abstracts peripheral signals into representations of peri-head space for downstream computation.

While proximity and map cells lie on a continuum of distance tuning rather than forming discrete functional classes, this classification provides a useful framework capturing progressive representational changes along the sensory hierarchy. TG radial distance encoding is mechanically driven and predominantly proximity-like^19,22–25. In PrV, we observed both proximity and map cells, with proximity cells overrepresented. In wS1, the proximity code is reduced, with distance preference more uniformly distributed¹⁶. This gradient suggests a progressive transformation from low-level mechanical signals toward feature-selective representations up the sensory hierarchy.

Our decoder analysis suggests that proximity and map codes make complementary contributions across peri-head distance ranges. The proximity subpopulation disproportionately supported accurate readout of the closest distances, whereas the map subpopulation dominated decoding performance across intermediate and far distances. Combining the two populations at the empirically observed ratio yielded only modest gains confined to the closest distances. However, because this comparison used fixed population size and composition, the absence of synergistic improvement in the combined decoder should not be taken as definitive evidence that proximity signals lack additional information when available to downstream circuits. Downstream circuits could benefit from simultaneous access to both signals, but may integrate them in a distance- and context-dependent manner. Future work is needed to test how downstream regions use these signals to guide behavior.

The map code may be particularly useful for behaviors requiring fine spatial judgments. Its superior decoding of intermediate and far distances could support whisker-guided navigation behaviors, including judging crossable gaps³⁷ and passable apertures^38,39, maintaining distance from walls^16,40, and avoiding obstacles⁴¹. The presence of map code in PrV may also reduce latency and readout complexity by making the peri-head distance encoding available to subcortical sensorimotor circuits. For example, PrV projects directly to the superior colliculus⁴², providing a pathway by which distance signals could modulate premotor circuits, bypassing thalamus and cortex. These potential roles of the map code remain open for future investigation.

SpVi-TeLC experiments and modeling support an E-I-E motif underlying this brainstem sensory transformation. In our model, PrV integrates direct TG→PrV excitation with indirect TG→SpVi→PrV feedforward inhibition, allowing comparison across TG inputs with different distance sensitivity profiles and producing peak responses at specific distances. The SpVi→PrV pathway therefore appears to have a broader role beyond sensory gating^30,43; it actively reshapes peri-head distance tuning, extending earlier evidence that SpVi inhibition sharpens angular tuning in PrV²⁹. This places the SpVi→PrV pathway within a growing class of inhibitory microcircuits that implement feature extraction through comparison or subtraction across sensory modalities. Examples include binaural circuits in the lateral superior olive⁴⁴, inhibitory subnetworks in visual cortex that help sculpt orientation and higher-order feature tuning⁴⁵; and lateral inhibition in olfactory circuits that decorrelates odor representations⁴⁶. Recent connectomic work further suggests that inhibitory neurons participate in stereotyped, feature-targeted motifs⁴⁷, supporting the idea that inhibition is a commonly repeated anatomical substrate for neural code transformation. Our results therefore support a broader view of the E-I-E motif as a computational primitive reused to construct higher-order sensory representations.

Our whisker-cutting experiments suggest that input from one whisker can be sufficient to support the map code. One explanation is that individual whisker follicles contain heterogeneous mechanoreceptor-afferent channels with distinct adaptation properties and mechanical thresholds^24,48,49. Szwed et al.²⁴ reported non-monotonic radial distance-dependent responses in TG neurons. However, these responses are not stable across trials and likely reflect whisker mechanics and receptor-specific sensitivities. In contrast, PrV appears to transform these mechanically derived signals into an invariant, head-centered representation of peri-head distance that supports accurate decoding across near, intermediate, and far distances.

Several limitations of the present study should be acknowledged. First, silicon-probe recordings enabled population-level analysis but not systematic receptive-field mapping for each recorded neuron. Thus, our mechanistic conclusions rely primarily on population-level effects and causal perturbations rather than single-neuron whisker identity. This is particularly relevant for the whisker-cutting experiment, since we could not determine whether the spared whisker was the principal whisker for a given neuron. Map tuning in some neurons might have been preserved when the principal whisker was spared (e.g., Figure 4H,I) and altered when it was cut (e.g., Figure 4D–G). The heterogeneous effects of whisker cutting underscore the need for future work combining precise receptive-field mapping with reversible manipulations of whisker input to distinguish multi-whisker from single-whisker mechanisms. Second, SpVi→PrV inhibition was tested with chronic TeLC perturbation, whereas PrV local inhibition was examined using acute optogenetic suppression. Because these manipulations operate on different timescales, chronic silencing may allow circuit adaptation, limiting direct comparison between long-range and local inhibition. Nonetheless, these two perturbation strategies provide valuable insight into each circuit and set the stage for future studies to compare them directly. Third, our analysis focused on rate-based distance coding, although PrV responses also showed temporal structure. Features such as first-spike latency may carry distance information, but without simultaneous neural recordings and precise whisker-wall contact timing, we could not determine whether temporal features contribute independently of, or alongside, the distance rate code. Finally, our wall-passing paradigm does not capture the full closed-loop sensorimotor dynamics of natural wall-following behavior. We therefore view the sensory code transformation identified here as one computation for peri-head distance estimation whose operation during natural behavior may be shaped by motor strategy and behavioral context. Prior work in wS1 found similar wall-distance tuning in open- and closed-loop conditions and stable preferred distance¹⁶, suggesting that aspects of distance coding may generalize across sensing regimes. An important next step will be to test how the brainstem distance code and inhibitory mechanisms identified here are engaged and potentially refined during natural exploration.

RESOURCE AVAILABILITY

Lead Contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by Fan Wang (fan_wang@mit.edu).

Materials Availability

This study did not generate new unique reagents.

Data And Code Availability

Original MATLAB scripts and analysis code are available on Github (https://github.com/wanglab-neuro/2026_Xiao-Severson_Peri-head-distance, doi: 10.5281/zenodo.20330279), and processed NWB datasets are deposited on DANDI (https://dandiarchive.org/dandiset/001687). Any additional information required to reanalyze the data reported in this work paper is available from the Lead Contact upon request.

STAR* METHODS

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Adult C57BL/6J (JAX stock #000664) and vGat-IRES-Cre (JAX stock #028862; B6J.129S6(FVB)-Slc32a1^tm2(cre)Lowl/MwarJ) mice were obtained from the Jackson Laboratory. During behavioral and electrophysiological experiments, mice were singly housed under a 12-hour light/dark cycle with ad libitum access to food and water. All experiments were conducted during the light phase. Details regarding the sex and age of all animals are provided in Table S1. All procedures were performed in accordance with protocols approved by the Institutional Animal Care and Use Committee at the Massachusetts Institute of Technology.

METHOD DETAILS

Headpost Implantation and Craniotomy

Mice were anesthetized with 3% isoflurane in 0.8 L/min oxygen and positioned in a stereotaxic frame (Model 963, David Kopf Instruments). Anesthesia was maintained at 1.5% isoflurane throughout the surgery. A custom stainless steel headbar (Pololu; 0.060″ ± 0.004″, type 304) and a stainless-steel ground screw were affixed to the skull using C&B Metabond (Parkell, #S380) and reinforced with dental cement (Stoelting). Mice were allowed at least seven days to recover following head-post implantation surgery. On the electrophysiological recording day, a 1 mm × 1 mm craniotomy was created 4.7 mm posterior and 2.0 mm lateral to Bregma using a dental drill (Aseptico). The dura was carefully removed to provide access to the principal trigeminal nucleus (PrV). Craniotomies were acutely sealed with silicone elastomer (Kwik-Cast, WPI).

Viral Vectors

The AAVs used included AAV-EF1a-DIO-GtACR2-EYFP³⁵ and AAV-hSyn-FLEX-TeLC-P2A-EYFP-WPRE (Addgene, #135391).

Viral Injections

For optogenetic silencing experiments, vGat-Cre mice received stereotaxic injections of 300 nl AAV-EF1a-DIO-GtACR2-eYFP into PrV at coordinates 4.7 mm posterior, 2.0 mm lateral to Bregma, and 4.5 mm below the brain surface. For chronic silencing of inhibitory SpVi neurons, 300 nl of AAV-hSyn-FLEX-TeLC-P2A-EYFP-WPRE was stereotaxically injected into caudal SpVi (7.4 mm posterior, 2.0 mm lateral to Bregma, and 5.6 mm below Bregma). Injections were performed using beveled glass micropipettes connected to a microsyringe pump (UMP3, WPI) controlled by a Micro4 controller (WPI) at a rate of 30 nl/min. Following injections, craniotomies were sealed with bone wax (Surgical Specialties, #902), and the scalp was closed using nonabsorbable silk sutures (LOOK, #774B).

Wall-Passing Paradigm

After recovering from headpost implantation, mice were head-fixed using plate clamps (Thorlabs) positioned over a custom linear treadmill²¹, which allowed spontaneous locomotion and whisking. Locomotion was neither trained nor rewarded, and wall movement was delivered in open loop independently of treadmill movement. Mice were habituated to the treadmill for at least two days, followed by at least three days of habituation on a wall-pass apparatus prior to recording. A wall made of clear resin (Formlabs Clear V3; 5 cm × 2.5 cm) was driven along the mediolateral and anteroposterior axes by two linear actuators (Zaber, Models LSM200B-T4 rev 2 and LSM050B-T4 rev 2). During each pass, the wall moved from anterior to posterior across the mouse’s face at a constant velocity of 10 mm/s. In the silicon probe experiments, wall distance from the whisker pad (C2 follicle) was randomly varied across trials between 7 and 23 mm in 2-mm increments (4-mm increments for optogenetic silencing experiments), using a custom MATLAB script. In the tungsten electrode experiments, wall distance was randomly varied between 4 mm and 20 mm in 1 mm increments. All experiments were performed in the dark. After the mediolateral Zaber motor set the wall distance, a randomized delay of 0.1 to 1 s preceded wall movement.

Each recording session in the silicon probe experiments included 15–20 repetitions per wall distance, during which mice freely whisked and interacted with the wall. Tungsten electrode experiments included 10–20 repetitions per distance. Whiskers were illuminated using an infrared LED backlight (Roithner Lasertechnik, #LED660N-66–60) and imaged with a high-speed camera (Basler, Model acA1300–200um) equipped with a telecentric lens (Edmund Optics, #55–349). Prior to each recording, the wall apparatus was aligned to the midline of the mouse’s head, and a calibration image was acquired and converted to physical units (31 pixels/mm). The closest wall distance was defined as the minimum Euclidean distance between the C2 whisker follicle and the wall. Absolute wall distance (in mm) was calculated by adding the calibrated scaling factor to the horizontal position (in 2-mm steps, relative to the closest position) of the stepper motor controlling the wall’s mediolateral position.

High-Speed Videography

High-speed whisker videos (688 × 688 pixels at 500 Hz) were acquired with a machine-vision camera (Basler acA1300–200um) controlled by custom Python software (CamPy⁵⁰). Illumination was provided by a 940 nm LED (Roithner Lasertechnik) collimated with a condenser lens and diffused to form an even backlight. The whiskers appeared in silhouette against this background and were imaged via a front-surface mirror into a 0.25× telecentric lens (Edmund Optics). Frame timing was triggered by a Teensy microcontroller, and TTL pulses were recorded alongside electrophysiology to synchronize spikes and video.

Awake and Anesthetized Wall-Passing Recordings

In a subset of mice (N = 4), high-speed video was continuously acquired during the wall-passing paradigm under both awake and anesthetized conditions. Mice were head-fixed throughout, keeping their position constant relative to the moving wall stimulus. To enable whisker tracking, all whiskers except a single row (C1–C4 and γ or β) were trimmed. While awake, each mouse experienced wall passes with randomized wall distances (7–23 mm in 2mm increments), repeated for 20 passes per distance. Following the awake session, the same animals were anesthetized via intraperitoneal injection of a fentanyl-based cocktail (0.05 mg/kg fentanyl, 5 mg/kg midazolam, 0.5 mg/kg dexmedetomidine). After a 10-minute induction period, the same wall-passing protocol was repeated under anesthesia. Comparisons between awake and anesthetized conditions were performed within-subject across matched wall distances.

Electrophysiology Recording: Silicon Probe

Following habituation to the wall-passing paradigm, mice were head-fixed and allowed to run on the treadmill. The craniotomy was exposed and covered with artificial cerebrospinal fluid (aCSF). For wild-type and TeLC experiments, the tip of a 32-channel silicon probe (NeuroNexus, A1×32-Poly2-10mm-50s-177-A32 or A1×32-Poly3-10mm-50s-177) was coated with the fluorescent dye DiI (Invitrogen^™, #V22885). For optogenetic silencing experiments, a 473 nm laser (Cobolt, 06-MLD 473) was connected to a ferrule attached to a 32-channel silicon optrode (NeuroNexus, A1×32-Poly2-10mm-50s-413, 105 μm/125 μm, 0.22 NA). The probe or optrode was advanced approximately 5 mm into the brain using a micromanipulator (Sutter Instruments, Model MPC-325) until it reached ventral PrV and well-isolated units responsive to manual whisker stimulation were encountered. The brain was allowed to relax for at least 40 minutes to ensure stable recordings. Neural signals were hardware-filtered on the headstage (Blackrock Microsystems, Cereplex M, #10532) between 0.3 Hz and 7.5 kHz for spikes and digitized at 30 kHz using Cereplex Direct software (Blackrock Microsystems, Version 7.0.5). Analog voltage signals from the anterior-posterior stepper motor were simultaneously acquired at 30 kHz. Wall pass start and end times were determined from the rising and falling edges of these voltage signals to synchronize electrophysiological recordings with wall movement. Recording sites were subsequently confirmed by post hoc histological analysis.

Electrophysiology Recording: Single Tungsten Electrode

One day prior to recording, whiskers and facial fur on the left side of the face were trimmed using fine forceps and microdissection scissors (FST, #91500–09) under isoflurane anesthesia (1.5%), sparing the C-row whiskers (γ or β, C1–C5). On the recording day, mice were head-fixed and allowed to run on the treadmill. The craniotomy was exposed and kept covered with aCSF. The tip of a tungsten microelectrode (0.005″ diameter, 5 MΩ, 8° tip angle; A-M Systems, #575300) was coated with the fluorescent dye DiI and lowered approximately 5 mm into the brain using a micromanipulator until it reached ventral PrV. Electrical activity was monitored via audio output from the amplifier (WPI, DAM80). Manual whisker stimulation was performed to identify a well-isolated, whisker-responsive unit. The brain was allowed to relax for at least 20 minutes to achieve stable recordings. Differential signals between the recording and reference electrodes were amplified 10,000× and bandpass filtered between 300 and 3,000 Hz (WPI, DAM80). Following recording, receptive fields and response types (rapidly adapting [RA] or slowly adapting [SA]) were manually classified. Analog voltage signals from the amplifier and the anterior-posterior stepper motor were acquired using Cereplex Direct at a sampling rate of 30 kHz. Wall pass start and end times were extracted from the rising and falling edges of the stepper motor TTL signals to synchronize electrophysiological recordings with wall movement. All recording sites were confirmed by post hoc histological analysis.

Isolation of Single Unit Activity

For silicon probe data, initial single-unit isolation (spike sorting) was performed using Kilosort2.5⁵¹, followed by manual curation in Phy⁵² (github.com/cortex-lab/phy). Spike times and waveforms for isolated clusters were exported from Phy. Most units exhibited narrow spike widths, defined by a trough-to-peak latency of less than 0.4 ms. Duplicate units, identified as those with high Pearson’s correlation coefficients (> 0.1) in spike counts binned at 1 ms resolution, were excluded from further analysis. For tungsten electrode data, spike waveforms were extracted by thresholding high-pass filtered traces (500 Hz) and clustered using MClust-4.4 (A.D. Redish, http://redishlab.neuroscience.umn.edu/MClust/MClust.html).

Isolation quality was assessed using the following metrics: signal-to-noise ratio (SNR), defined as the mean waveform amplitude divided by the standard deviation (SD) of waveform noise; and false alarm (FA) rate, defined as the percentage of spikes with inter-spike intervals (ISI) shorter than 0.8 ms. Units with an SNR less than 5 SDs or an FA rate greater than 2% did not meet single-unit isolation quality criteria and were excluded from subsequent analyses. For silicon probe data, in wild-type mice, 1141 of 1358 units (36 mice), in TeLC experiments, 587 of 665 units (11 mice), and for Tungsten recordings (46 of 46 units (14 mice), passed isolation quality and minimum firing rate (>1 Hz) criteria and were subsequently entered into population analyses. For acute whisker-cutting experiments, only units demonstrating stable waveform properties both before and after cutting were included in analyses.

Acute Whisker-Cutting Manipulation

During sessions involving acute whisker cutting, neural responses were first recorded with the full whisker array intact. Each session included 15 to 20 repetitions of the same wall distance over approximately one hour. The whisker to spare was chosen by online functional screening rather than by systematic receptive field mapping. Candidate whiskers (arcs 0–3; arc 0 corresponds to the Greek whiskers) were manually deflected while monitoring responses across the probe, and the whisker that evoked responses in the largest number of simultaneously recorded units was spared. All whiskers except the designated spared whisker were then cut to the base of the shaft using microdissection scissors (FST, Cat# 91500–09). After cutting, the spared whisker and surrounding whisker pad were brushed with a cotton-tipped applicator to stimulate touch-sensitive units, and the wall-passing procedure was repeated. After the post-cut trials, the spared whisker and whisker pad were brushed again to confirm the continued presence of responsive units. Electrophysiological signals were continuously recorded throughout the pre-cut, cutting, brushing, and post-cut periods.

Optogenetic Silencing of PrV Local Interneurons

Light pulses were generated using a pulse generator (Doric, #OTPG-4), with laser timing synchronized to wall movement via Doric software. Continuous 473 nm light stimulation (2 mW measured at the fiber tip with an optical power meter; Thorlabs) was delivered by a laser (Cobolt, #06-MLD 473) throughout each wall pass, with a 5-second inter-trial interval. Each recording session consisted of 30 to 40 repetitions at each wall distance, ranging from 7 to 23 mm in 4 mm increments, conducted over approximately two hours with randomized “Laser Off” and “Laser On” trials.

Histology

Following recordings, mice were deeply anesthetized and transcardially perfused with phosphate-buffered saline (PBS), followed by 4% paraformaldehyde in PBS. Brains were post-fixed overnight and then cryoprotected in 30% sucrose in PBS at 4°C. Tissue was embedded in OCT compound (Sakura Finetek) and coronally sectioned at 100 μm using a cryostat (Leica Biosystems). For wild-type, optogenetic inhibition, and tungsten recording experiments, sections containing PrV were stained with Neurotrace Blue (1:500; Thermo Fisher Scientific, #N21479) in 0.3% Triton X-100/PBS. For TeLC experiments, sections containing PrV and SpVi underwent immunostaining with goat anti-GFP (1:500; Rockland, #600-101-215) in blocking buffer, followed by either Alexa Fluor 488 (1:500; Jackson ImmunoResearch, #705-545-147) or Alexa Fluor 647 (1:500; Jackson ImmunoResearch, #705-606-147) donkey anti-goat secondary antibodies and Neurotrace Blue (1:500; Thermo Fisher Scientific, #N21479) in blocking buffer. The sections were then briefly rinsed with 0.3% Triton X-100/PBS and mounted on slides with a custom-made mounting medium³⁵. All histological images were acquired with either a Keyence microscope (BZ-X800) or a Zeiss LSM700 confocal microscope.

QUANTIFICATION AND STATISTICAL ANALYSIS

All analyses were performed in MATLAB (MathWorks, versions 2021b and 2023b). Data are reported as mean ± standard error of the mean (SEM) unless otherwise noted. Measures of central tendency, variability, test statistics, p-values, degree of freedom, and figure panel assignments for all analyses are detailed in Table S6. We analyzed single-unit electrophysiological recordings from n = 1,187 units in PrV across 63 mice, as well as high-speed video recordings from four sessions in four mice. The video data comprised 15,679,458 frames, totaling 522.6 minutes. Sample sizes were not predetermined.

Video Analysis

The backbone of each whisker was tracked at subpixel-resolution using the Janelia Whisker Tracker⁵³, yielding a set of “trajectories” (tracked objects in image X-Y coordinates) for each frame.

All subsequent processing of whisker angle (θ) was conducted in MATLAB according to published methods^21,23, with several modifications described below in “Video Analysis” subsections.

Video Analysis: Identifying Tracked Whiskers

To identify the same whiskers across frames (linking), whiskers were ordered by follicle x-position. Trajectories with follicle x- or y-position outside the region of interest (bounding box), with arc length shorter than 150 pixels (4.65 mm), or whisker “score” (Janelia whisker tracker measurements) less than 600 were ignored. Post-hoc median filtering (3-frame window) was applied to smooth over single-frame linking errors.

Video Analysis: Face Masking

Computing time series of whisker angle relies on being able to reliably track the base of the whisker near the follicle location across all video frames. As previously described^21,23, we used a dynamic “mask” to truncate the tracked whisker trajectories near the face to remove “noise” due to occlusion by facial fur. The follicle location was estimated by extrapolating past the intersection of the whisker and the mask using the angle at the whisker base²³. We fit a separate mask for each frame, obtained using a custom algorithm that fitted a smoothing spline to the contour of the face. We blurred the image using a Gaussian filter, binarized the image, and detected the contour by taking a spatial derivative along the y-dimension. The pixel locations of the facial contour were thus detected as the intersection between the white background and black shadow cast by the face. A smoothing spline was then fitted to the facial contour positions and shifted upward by ~10 pixels to yield a 20-point facemask that intersected with the whisker trajectories. Frames in which a tracked whisker did not intersect the facemask were considered missing data.

Data Analysis: Wall-Induced Whisker Retraction, Awake vs Anesthetized

For each tracked whiskers (C1, C2, and Greek), wall-induced retraction was quantified using both whisker angle and curvature. Whisker angle (θ) was defined relative to the mediolateral axis of the face. To slowly varying “centerline” of the whisking trajectory was estimated as the midpoint angle trace (θ_mid). To obtain θ_mid, θ was first smoothed using a Savitsky-Golay filter (21 frame, third order), followed by interpolation of the peaks and valleys to obtain the upper and lower envelope of the θ trace. θ_mid was then calculated as the mean of this envelope. For each wall pass, θ_mid was computed during both the baseline period (−4.5 to 0 s, with t=0 marking wall movement onset) and the wall-passing period (3 to 11 s). Baseline-subtracted θ_mid (Δθ_mid) was then calculated for each pass. For each mouse (N = 4 mice), Δθ_mid was averaged across wall distances and repetitions to obtain a single value per whisker and condition (awake or anesthetized). Data were aggregated across mice into a summary table (mouse × whisker × condition), and paired comparisons were made between awake and anesthetized conditions for each whisker.

Whisker curvature (κ) was computed from the tracked whisker backbone using a local second-order polynomial fit to the x-y coordinates of a continuous 4-mm arclength segment of the tracked whisker. For each frame, the fitted curve was parameterized by a normalized coordinate q, and first and second derivatives (x′, y′, x′′, y′′) were obtained from the polynomial coefficients. Curvature was calculated as the signed planar curvature:

$$\kappa =-\frac{x^{\prime }{y}^{″}-y^{\prime }{x}^{″}}{{\left({x^{\prime }}^2+{y^{\prime }}^2\right)}^{3/2}},$$

with a sign convention chosen so that caudal bending toward the ear produced a negative change in curvature during the wall pass. To quantify wall-evoked bending, change in curvature was computed by subtracting the baseline curvature (κ₀), defined as the mean curvature during the baseline period (−4.5–0 s):

$$\Delta \kappa \left(\text{t}\right)=\kappa \left(\text{t}\right)-\kappa 0.$$

For each pass, Δκ was summarized by averaging Δκ(t) over the wall-passing window (3–11 s), and then (as for Δθ_mid) averaged across wall distances and repetitions to yield a single value per whisker, mouse, and condition.

To compare the effect of wall-induced whisker retraction in awake versus anesthetized conditions, we performed separate two-tailed paired t-tests across mice (within-subjects) for Δθ_mid or ΔK for each whisker. The null hypothesis in each test was that the mean value was equal between awake and anesthetized conditions. Effect size was computed as the mean difference between conditions (awake minus anesthetized). Bonferroni correction was applied to control for multiple comparisons across the three whiskers (α = 0.0167).

Data Analysis: Whisker Kinematics, Wild-Type vs SpVi-TeLC

To assess whether chronic TeLC expression in SpVi altered whisker kinematics during the wall-passing paradigm, we analyzed whisker array-averaged Δθ_mid, which captures both whisker retraction and bending upon wall contact. Because mice had the full whisker array intact, individual whiskers could not be reliably tracked throughout wall passes (due to overlap and occlusion). For each wall pass, we computed mean Δθ_mid across all tracked whiskers in each frame. We compared 15 recording sessions from six wild-type mice to 15 recording sessions from six SpVi-TeLC mice. To test for effects of TeLC expression at each wall distance, we fitted linear mixed-effects models in MATLAB (‘fitlme’, maximum likelihood) of the form: θ_mid ~ TeLC + (1 | subject) + (1 | session), where TeLC is a binary fixed effect (wild-type vs SpVi-TeLC), and random intercepts for subject and session account for repeated measures across animals and recording sessions. To assess TeLC effects across all distances, we fit an omnibus mixed-effects model including distance as a categorical fixed effect to test the main effect of TeLC while adjusting for distance: θ_mid ~ TeLC + distance + (1 | subject) + (1 | session). All hypothesis tests were two-sided.

Data Analysis: Spike Density Function Calculation

Spike density functions (SDFs) were computed by aligning spike times to the onset of wall-pass trials, with time zero defined as the moment when the wall began moving toward the mouse. For all trial conditions, two epochs were defined consistently: a baseline period from −4.5 to 0 s, and a wall-pass period from 3 to 11 s. Spike rasters were generated from the aligned spike times using 1 ms bins. SDF curves were then produced by convolving these spike rasters with a Gaussian filter (5 ms SD, full kernel). SEMs were calculated as the SD of the SDF across trials divided by the square root of the number of trials in each condition. The resulting SDFs and SEMs were averaged across repetitions of the same wall distance and stored for each cell.

Data Analysis: Wall Distance Tuning Curve Calculation

Wall distance tuning curves were generated by uniformly binning distance measurements into nine 2-mm-wide bins with centers at 7, 9, 11, 13, 15, 17, 19, 21, and 23 mm. Distances falling outside the bin edges were excluded. Peri-event time histograms (PETHs) were computed by binning instantaneous firing rates in 200-ms intervals, aligned to the onset of anterior-posterior wall movement as defined by the stepper motor TTL pulse (t=0). Each wall pass trial lasted approximately 14 seconds. Average firing rates during the wall pass period (3–11 s) were computed from the PETH and compared to baseline firing rates calculated from −4.5 to 0 s. For each distance, mean firing rates were averaged across repeated passes to construct the tuning curve. Standard errors were calculated by dividing the SD across repetitions by the square root of the number of repetitions for each distance. CIs were estimated using bootstrap resampling (10,000 iterations) and defined as the 95th percentile range. Z-scored tuning curves were computed by subtracting the mean from firing rate across all baseline periods in the recording session and dividing by the corresponding SD. Uniform binning enabled concatenation of data across sessions for population-level analyses, including decoding analyses described later. Sessions with fewer than nine presented distances were excluded from population analyses. For population heatmaps, each neuron’s tuning curve was normalized to its own minimum and maximum firing rates, and the color scale represents the resulting normalized mean spike rate.

Data Analyses: Characterizing Wall Distance Tuning

Wall responsiveness and distance tuning were assessed independently for each of the 1,187 (46 units from Tungsten dataset, 1141 units from silicon probe dataset) recorded units. To determine responsiveness to wall stimulation, one-tailed Wilcoxon signed-rank tests (p < 0.05) were performed comparing baseline firing rate (−4.5–0 s) to activity during the wall pass period (3–11 s), evaluated separately at each wall distance. Units with significantly higher mean firing rate at any distance were classified as “wall-activated”; those with significantly reduced mean firing rate at any distance were classified as “wall-suppressed”. Units exhibiting significant increases at some distances and decreases at others were classified as both.

Wall distance tuning was assessed independently using two-tailed Wilcoxon signed-rank tests across all pairs of distances. A unit was classified as “distance-tuned” if its mean firing rate at one or more distances differed significantly from that at other distances (p < 0.05, corrected for multiple comparisons using the Benjamini-Hochberg false discovery rate method). A subset of units (n=116; 4 units from Tungsten dataset, 112 units from silicon probe dataset) showed consistent activation or suppression across all wall distances without significant differences in tuning among wall distances. These wall-responsive but not distance-tuned units were excluded from subsequent population-level tuning analyses.

To examine temporal dynamics of distance tuning, the wall pass period was subdivided into three non-overlapping analysis windows: onset (2.5–4.5 s), sustained (4.5–9.5 s), and offset (9.5–11.5 s). These windows were selected based on the presence of distinct transient responses observed at stimulus onset and offset (e.g., Supplemental Figure S4J).

For comparisons of baseline firing rate distributions between distance-tuned and untuned units, the mean and 95% confidence intervals (CI) were calculated using bootstrap resampling (10,000 iterations). The CI was determined by calculating half the range between the upper and lower bounds of the bootstrap interval. Results were reported as mean ± CI. Statistical differences between distributions were assessed using two-sample Kolmogorov-Smirnov (K-S) tests (‘kstest2’ function in MATLAB), performed with one- or two-tailed hypotheses as dictated by the experimental design (see Table S6). Significance was determined at p < 0.05, with Bonferroni correction for multiple comparisons. The same procedures were applied for comparing average firing rates during wall passes and baseline periods between distance-tuned and untuned populations.

To further characterize tuning profile within “wall-activated” and “distance-tuned” units, we performed one-tailed Wilcoxon signed-rank tests comparing mean firing rates at each non-closest distance (≥ 9 mm) to the closest distance (7 mm). Units with significantly higher firing rates at any non-closest distance were classified as “map units,” while those exhibiting maximal firing at the closest distance (7 mm) were classified as “proximity units.” Units that met criteria for “wall-activated” and “distance-tuned” but did not fulfill the criteria for either map or proximity were labeled “ambiguous” and excluded from decoding and population cross-validation analyses.

For analyses of acute whisker-cutting experiments, mean firing rates during baseline and wall-pass periods were compared between full whisker array (control) and single-whisker (cut) conditions. Pairwise comparisons of firing rates were conducted using two-tailed Wilcoxon signed-rank tests. To assess individual neuron responses to whisker cutting, the proportions of units exhibiting significant increases, decreases, or no changes in firing rates were computed. Significance thresholds were corrected for multiple comparisons using Bonferroni adjustment.

To compare the proportions of map, proximity, ambiguous, suppressed, or untuned cells between SpVi-TeLC mice and wild-type mice, a one-tailed Z-test for proportions was performed. The null hypothesis (H₀) stated that the proportion of each cell type in the SpVi-TeLC population (p₂) was less than or equal to that in the wild-type population (p₁), while the alternative hypothesis (Hₐ) posited that p₂ > p₁. Proportions (p₁ & p₂) were calculated as the number of cells of a given type (x₁ & x₁) divided by the total number of recorded units (n₁ and n₂) in each group. A pooled proportion (p_pooled) was used to estimate the standard error (SE) of the difference between groups:

$$p_{\mathit{\text{pooled}}}=\frac{x_1+x_2}{n_1+n_2},\mathit{\text{SE}}=\sqrt{p_{\mathit{\text{pooled}}}\left(1-p_{\mathit{\text{pooled}}}\right)\left(\frac{1}{n_1}+\frac{1}{n_2}\right)}$$

The Z-statistic was computed as:

$$Z=\frac{p_1-p_2}{\text{SE}}$$

P-values were calculated using MATLAB’s cumulative normal distribution function (‘normcdf’ in MATLAB) and evaluated at α = 0.05. A p-value below this threshold indicated a significantly higher proportion of the specified cell type in SpVi-TeLC mice compared to wild-type mice.

To compare the firing rate distributions between SpVi-TeLC and wild-type populations (Group 1 and Group 2), we computed the mean and 95% CI for each group using bootstrap resampling (10,000 iterations). For each group, the results were reported as mean ± CI. To assess the statistical significance of differences between the distributions of the two groups, we performed a two-sample KS test. The test was conducted with a specified directional hypothesis (one- or two-tailed, depending on the experimental design), and the resulting p-value was reported to determine the significance (p < 0.05, with Bonferroni correction for multiple comparisons) of the observed differences.

Data Analysis: Unit Depth Along Silicon Probe

To examine how tuning properties varied with recording location along the silicon probe, we estimated each wall-responsive unit’s position along the shank using the probe channel map. For NeuroNexus A1×32 Poly2 probe, electrode site geometry (inter-site spacing and site order) was defined by the manufacturer layout and implemented in the Kilosort channel map (chanMap.mat), which assigns each recording channel a physical coordinate corresponding to its electrode site on the shank. The active recording sites spanned 0–775 μm along the shank (measured from the channel closest to the probe tip). For each unit, we identified its peak channel from the unit’s mean waveform channel with the largest peak-to-trough amplitude. The unit’s depth from probe tip was then defined as the shank coordinate (μm) of this peak channel, with increasing values corresponding to more dorsal recording sites (further from the tip). Units were grouped into fixed 100 μm-wide depth bins.

Data Analysis: Decoding Wall Distance from Population or Single-Neuron Activity

To decode wall distance from population activity, we applied a maximum correlation coefficient (MCC) decoder implemented in the Neural Decoding Toolbox (www.readout.info)^54,55. All units that passed spike-sorting quality control and were recorded in sessions with all nine wall distances presented (1,141 total units) were concatenated into a pseudo-population. For each unit, spike timestamps were aligned to the onset of anterior-posterior wall movement (t=0) and were binned using a 50 ms bin size with a 150 ms sliding window to create a firing rate matrix, where each row corresponds to a trial and each column corresponds to a time bin. The firing rate matrix spans from time 0 to 14 seconds. To ensure equal trial numbers across conditions (i.e., wall distances), trials were subsampled five times to match 15 trials per condition. The binned data and their corresponding condition labels were randomly split into 10 subsets: 9 used for training the classifier and 1 reserved for validation. The process was repeated 10 times for each unique partition for 10-fold cross-validation to assess decoding accuracy.

The MCC decoder learns a mean population activity vector (template) for each distance class from the training set. During testing, it calculates the Pearson correlation coefficient between a test point and the learned templates. During validation and testing, the class with the highest correlation value is assigned as the predicted label. Decoding accuracy (true positive rate) was averaged over the 3–11 second wall-contact period to obtain the overall decoding accuracy for the sampling epoch. In addition to the decoding accuracy, confusion matrices were computed at each time step. To assess variability, the entire subsampling procedure was repeated 10 times, and SEM was calculated.

To compare subpopulations while controlling for population size, we generated 10 independent random subsamples per group (map, proximity, combined [map + proximity], and untuned), with each subsample matched to the smallest group (n=118 units). For the combined “map + proximity” group, each subsample was constructed by randomly drawing 36 map units and 82 proximity units to preserve the relative proportions observed in the full dataset (see Figure 2F). A “shuffled” control subpopulation was generated by randomly sampling from map and proximity units and shuffling trial distance labels with replacement.

For each subsample and time bin, we computed overall decoding accuracy (percent correct) from the cross-validated confusion matrix. Distance-specific performance was quantified as the true-positive rate (TP/(TP+FN)) for each distance condition. Decoding error (mm) was computed from the confusion matrix as the mean absolute distance deviation between predicted and true labels, weighted by the number of trials assigned to each predicted class. Summary accuracy and error were reported over the wall-contact period (3–11 s) and aggregated across subsamples to obtain mean ± bootstrap 95% confidence intervals.

To test whether relative subpopulation performance depends on the choice of classifier, we repeated the population decoding using a multiclass linear support vector machine (SVM) in place of MCC, while keeping preprocessing, binning, trial matching, and cross-validation identical. SVM decoding used the Neural Decoding Toolbox LIBSVM wrapper (libsvm_CL)⁵⁶ with a linear kernel and default regularization (C=1). Multiclass classification was implemented using a one-versus-one (“all pairs”) strategy.

To assess distance information at the single-neuron level, we implemented single-neuron decoder for wall distance. Spike trains were grouped by the nine wall distances, aligned to trial onset, and converted into spike count vectors using 200 ms non-overlapping bins spanning the 3–11 s wall contact period. For each distance, we computed a mean firing-rate “template” across training trials. Classification was performed with a Euclidean distance decoder: each test trial was assigned to the distance class whose template was closest in Euclidean space. Cross-validation was implemented as a leave-two-out procedure (xval=2), in which two trials per condition were held out as test data, and the remaining trials were used for training. This was repeated exhaustively across held-out pairs, such that each trial pair served as a test set at least once. Single-neuron decoding accuracy was computed as the fraction of correctly classified test trials, averaged across cross-validation iterations. For shuffled controls, distance labels were randomly permuted across trials (1000 iterations) to generate a null distribution.

For both population and single-neuron decoders, differences in mean decoding accuracy and decoding error across wall distances were assessed using Wilcoxon rank-sum tests. Tests were performed separately for each of the nine distances, and p-values were corrected for multiple comparisons using BH-FDR.

Data Analysis: Acute Whisker-Cutting Manipulation

To test for the effect for acute whisker cutting on firing rates, we computed each unit’s mean firing rate during baseline and wall-passing periods on each repetition for the full whisker array (control) or single-whisker condition (manipulation). Pairwise comparison of mean firing rate for neural population and individual units between conditions was performed via two-sided Wilcoxon signed-rank test. To evaluate how individual neurons responded to whisker cutting, we computed the proportion of units that exhibited significant increases, decreases, or no change in firing rate between conditions. P-values were corrected for multiple comparisons across units using the BH-FDR procedure.

To compare the effect of whisker cutting on distance readout from map versus proximity populations, we extended the population decoding analysis to units continuously recorded across full-array and single-whisker conditions within the same session. Units were grouped by their tuning category (map or proximity) in the pre-cut full whisker array condition. To match population size of 21 map units, we used a subset of 21/45 proximity units (randomly subsampled for 10 iterations). For each group (map vs. proximity) and condition (full array vs. single whisker), we trained MCC population decoders using the same time window, distance classes, and cross-validation procedure as previously described, yielding four confusion matrices: pre-cut map, pre-cut proximity, post-cut map, and post-cut proximity.

Data Analysis: Modulation Index in Acute Optogenetic Experiments

We quantified the acute optogenetic effects on firing rates using a modulation index (MI) computed for each unit as:

$$\text{MI}=\frac{\text{F}{\text{R}}_{\mathit{\text{LASER}}\mathit{\text{ON}}}-\text{F}{\text{R}}_{\mathit{\text{LASER}}\mathit{\text{OFF}}}}{\text{F}{\text{R}}_{\mathit{\text{LASER}}\mathit{\text{ON}}}+\text{F}{\text{R}}_{\mathit{\text{LASER}}\mathit{\text{OFF}}}}$$

where FR_{LASER ON} and FR_{LASER OFF} denote the mean firing rates during matched laser-on and laser-off wall-passing periods, respectively. For each unit, we obtained a bootstrap CI on the MI and classified the unit as significantly modulated if the CI did not include zero. Significantly modulated units were further assigned to one of three categories based on their MI: strongly inhibited (MI ≤ −0.5), inhibited (−0.5 < MI < 0), or disinhibited (MI > 0). Units whose MI confidence intervals overlapped zero were classified as no change. All analyses were restricted to units that passed our unit inclusion criteria (see Isolation of Single Unit Activity).

Computational Simulation of Brainstem Circuitry

To model the tuning properties of PrV neurons, we simulated a pool of TG afferents whose activities decreased monotonically with wall distance, consistent with their mechanical sensitivity. These proximity-based TG inputs projected to PrV neurons via a direct excitatory pathway and an indirect inhibitory pathway (SpVi inhibitory projections to PrV). TG tuning curves were modeled as sigmoidal functions with thresholds spanning the distance range. The activity of each PrV neuron $\text{j}\left({\text{v}}_{\text{j}}^{\text{Prv}}\right)$ in response to wall distance $\text{x}\in \left[0,1\right]$ was defined as:

$${\text{v}}_{\text{j}}^{\text{PrV}}=f\left(\sum _{\text{i}\in {\text{C}}_{\text{j}}^{\text{E}}}{\text{w}}_{\text{ij}}{\sigma }_{\text{TG}}\left(\text{x};{\beta }_{\text{i}},{\theta }_{\text{i}}\right)-\sum _{l\in {\text{C}}_{\text{j}}^{\text{I}}}{\text{w}}_{l\text{j}}{\sigma }_{\text{TG}}\left(\text{x};{\beta }_l,{\theta }_l\right)+I_j^{bg}\right)$$

where,

$${\sigma }_{\text{TG}}\left(\text{x};\beta ,\theta \right)=\frac{1}{1+\text{exp}\left(\beta \left(\text{x}-\theta \right)\right)}$$

is the TG neuron’s sigmoidal tuning function. Here, β and θ control sensitivity and selectivity to wall distance, respectively. f(x) = max {x, 0} is the non-linear activation function for PrV neurons and $I_j^{bg}~U(0,1)$ is the background excitatory input to PrV neuron j. ${\text{C}}_{\text{j}}^{\text{E}}$ denotes the set of TG neurons providing direct excitatory input to PrV neuron j, with synaptic weight $\left.{\text{w}}_{\text{ij}}~U\left(0,{\left|C_j^E\right|}^{-\frac{1}{2}}\right]\right)$; ${\text{C}}_{\text{j}}^{\text{I}}$ denotes TG neurons projecting indirectly to PrV neuron j via inhibitory neurons in SpVi, with weight ${\text{w}}_{l\text{j}}~U\left(0,{\left|C_j^I\right|}^{-\frac{1}{2}}\right)$. Subscript i and l denote each TG neuron in set ${\text{C}}_{\text{j}}^{\text{E}}$ and set ${\text{C}}_{\text{j}}^{\text{I}}$. For simplicity, inhibitory neurons were modeled as providing a sign-inverted copy of TG activity, such that ${\text{v}}_{\text{j}}^{\text{PrV}}$ reflect a linear combination of excitatory and inhibitory streams.

Intuitively, the subtraction of two sigmoidal functions can yield a non-monotonic, Gaussian-like function. Here we generalize this by considering a pool of sigmoidal functions (TG neurons). Specifically, we implemented several assumptions:

1. Sparse connectivity: Both ${\text{C}}_{\text{j}}^{\text{E}}$ and ${\text{C}}_{\text{j}}^{\text{I}}$ represented small subsets of TG neurons. Parameter β_i, θ_i, β_l, θ_l, w_ij, w_lj were independently sampled for each PrV neuron. The number of inhibitory inputs $\left|{\text{C}}_{\text{j}}^{\text{I}}\right|$ per PrV neuron j was treated as a random variable sampling from an exponential distribution parametrized by λ_c (truncated to the interval [1,30] to preserve E/I balance):

   $$\text{P}\left(\left|{\text{C}}_{\text{j}}^{\text{I}}\right|=\text{k}\right)=\frac{{\text{e}}^{-{\lambda }_{\text{c}}\text{k}}}{{\sum }_{\text{m}=1}^{30}{\text{e}}^{-{\lambda }_c\text{m}}},\text{k}=1,2,3,\dots 30,{\lambda }_c=0.5$$

   The number of excitatory inputs was fixed at $\left|{\text{C}}_{\text{j}}^{\text{E}}\right|=20$.

2. Skewed distribution of TG tuning: The TG threshold parameter θ was drawn from a truncated exponential distribution within [0,1]:

   $$\theta ~\text{f}\left(\text{x};{\lambda }_{\theta }\right)=\frac{{\lambda }_{\theta }{\text{e}}^{-{\lambda }_{\theta }\text{x}}}{1-{\text{e}}^{{\lambda }_{\theta }}},\text{x}\in \left[0,1\right]$$

   where λ_θ describes the skewness of the distribution. This skewed distribution increases the density of high-threshold TG neurons that fire at proximal but not at distal wall positions, thereby enhancing encoding resolution and reducing overall metabolic cost (efficient coding). We used λ_θ = 2 for the excitatory inputs and λ_θ = 4 for the more skewed inhibitory pathway. Sensitivity parameters were sampled as β_i, β_l~N(20, 5). Untuned PrV neurons were modeled by explicitly weakening all weights for a fraction of neurons to 1%, i.e., p(untuned) = 0.4.

3. Skewed threshold distributions for excitatory and inhibitory TG inputs: The distance tuning of each TG input was modeled as a monotonically decreasing sigmoid,

   $$\sigma =\frac{1}{1+e^{(\beta (x-\theta ))}},$$

   where θ sets the position of the transition along normalized wall distance and β controls its steepness. Larger θ values shift the sigmoid rightward, such that the input remains active over a broader range of distances and falls off only at farther wall positions. To generate distinct excitatory and inhibitory inputs, θ values were sampled from bounded exponential distributions on [0,1], with different skew parameters for the two pathways. Excitatory TG→PrV inputs were sampled with a shallower skew (rate = 2), whereas inhibitory TG →SpVi→PrV inputs were sampled with a more strongly skewed distribution (rate = 4). This produced a higher density of smaller-θ sigmoidal functions in the inhibitory pathway than in the excitatory pathway, biasing the inhibitory stream toward TG inputs with higher mechanical thresholds. Slope parameters were independently sampled for each input as β~20 + 5U (0,1). Together, these pathway-specific threshold distributions defined the relative shapes of excitatory and inhibitory drive, where high E-I ratio leads to monotonic proximity tuning and stronger inhibition leads to non-monotonic map-like tuning in PrV.

The input (distance) range x was normalized to [0,1] for simplicity. We sampled 99 uniformly spaced points corresponding to wall-distance presentations. To evaluate tuning curves, we averaged simulated activity into nine discrete distance bins [0, …, 8]. All neurons with flat tuning curves (max – min < 0.1) were considered untuned. Neurons with max activity below zero were considered suppressed. Among the remainder, proximity cells were defined as neurons with maximal responses at x = 0, while map cells were defined as neurons with peak responses at intermediate or far distances (x > 0). In total, we modeled 1,225 PrV neurons.

To simulate partial silencing of inhibitory projections, TeLC was modeled by scaling long-range inhibitory weights by 0.5, without changing local excitation/inhibition, such that $\left|{\text{C}}_{\text{j}}^{\text{I}}\right|$ for each PrV neuron j is decreased by:

$$\text{P}\left(\left|{\text{C}}_{\text{j}}^{\text{I}}\right|=\text{k}\right)=\frac{{\text{e}}^{-{\lambda }_c\text{k}}}{{\sum }_{\text{m}=1}^{16}{\text{e}}^{-{\lambda }_c\text{m}}},\text{k}=1,2,3,\dots 15.$$

Data Analyses: Distance Tuning Curve Shape

To examine how distance tuning curve shape changes after silencing SpVi→PrV inhibition, we calculated tuning curve width, amplitude, and monotonicity based on metrics described in Watkins & Barbour⁵⁷ across wild-type, SpVi-TeLC, and modeled populations. For empirically recorded neurons, analyses were restricted to units classified as wall-activated and distance-tuned (n = 472 for wild-type mice; n = 252 for SpVi-TeLC mice). For modeled neurons, analyses were restricted to units classified as either “proximity” or “map”.

Tuning curve width.

Tuning width was quantified as the full width at half maximum (FWHM) of each unit’s distance tuning curve, referenced to baseline firing rate. For each unit, the half-maximum response was defined as: half max = baseline + 0.5 × (peak − baseline). The left and right half-maximum crossing points were estimated by linear interpolation between neighboring distance bins, and FWHM was defined as the distance between these crossings, in mm. If the tuning curve remained above half-maximum at one sampled edge, the width was truncated to that edge rather than extrapolated beyond the sampled range. Units with non-positive modulation above baseline or insufficient valid data were assigned undefined width values and excluded from width-based comparisons. For empirically recorded neurons, baseline firing rate was measured during the 4.5s pre-wall baseline. For modeled neurons, the same calculation was applied with the baseline set to zero.

Amplitude.

Response magnitude was defined for each unit as the peak response minus baseline: modulation amplitude = peak response − baseline. This served as the primary response gain summary statistic for population comparisons. In empirically recorded neurons, modulation amplitude is expressed in spikes/s above baseline, whereas in modeled data it is expressed in arbitrary units set by the model output scale.

Monotonicity.

Monotonicity was quantified using a signed edge-preference index derived from the ordered distance tuning curve. To isolate tuning shape from absolute firing-rate offset, each curve was first min-subtracted. A signed monotonicity was then computed as: $\text{monotonicity}=\frac{\text{far}\text{edge}\text{response}-\text{near}\text{edge}\text{response}}{\text{peak}\text{response}}$. After min-subtraction normalization, monotonicity ranges from −1 to 1, with negative values indicating tuning preferring close distance, positive values indicating preferring far distance, and values near 0 indicating either symmetric edge responses or centrally peaked tuning.

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Goat anti-GFP	Rockland	Cat#600-101-215; RRID: AB_2208318
Bacterial and virus strains
AAV-hSyn-FLEX-TeLC-P2A-EYFP-WPRE	Takatoh & Prevosto et al.35	RRID: Addgene_135391
AAV-EF1a-DIO-GtACR2-EYFP	Takatoh & Prevosto et al.35	N/A
Chemicals, peptides, and recombinant proteins
DiI fluorescent dye (Vybrant™ DiI)	Thermo Fisher Scientific	Cat#V22885
Experimental models: Organisms/strains
Mouse: C57BL/6J	The Jackson Laboratory	JAX# 000664; RRID: IMSR_JAX:000664
Mouse: B6J.129S6(FVB)-Slc32a1tm2(cre)Lowl/MwarJ	The Jackson Laboratory	JAX# 028862; RRID: IMSR_JAX:028862
Software and algorithms
MATLAB v.2021b and v.2023b	MathWorks	https://www.mathworks.com/help/docarchives.html
ImageJ	NIH	https://imagej.nih.gov/ij/
CerePlex Direct Central Suite (version v7.0.5)	Blackrock Microsystems	https://blackrockneurotech.com/products/cereplex-direct/
Janelia Whisker Tracker	Clack et al.53	https://www.janelia.org/open-science/whisk-whisker-tracking
Kilosort2.5	Pachitariu et al.51	https://zenodo.org/records/4482749
campy	Dunn & Marshall et al.50	https://github.com/ksseverson57/campy
Neural Decoding Toolbox (NDT)	Meyers54	http://www.readout.info
Original code for brainstem circuitry model	This paper	https://github.com/wanglab-neuro/2026_Xiao-Severson_Peri-head-distance; 10.5281/zenodo.20330279
MClust, v.4.4	A.D. Redish	http://www.redishlab.neuroscience.umn.edu/MClust/MClust.html
Phy	Rossant et al.52	https://github.com/cortex-lab/phy
WaveSurfer	HHMI Janelia Research Campus	http://www.wavesurfer.janelia.org
Doric Neuroscience Studio software (version 6.1.5.0)	Doric Lenses	https://neuro.doriclenses.com/products/doric-neuroscience-studio
Other
Linear actuator, Model LSM200B-T4 rev 2	Zaber Technologies	www.zaber.com
Linear actuator, Model LSM050B-T4 rev 2	Zaber Technologies	www.zaber.com
Controller, Model X-MCB2, firmware v6.30	Zaber Technologies	www.zaber.com
Custom stainless steel head bar	This paper	https://github.com/wanglab-neuro/2026_Xiao-Severson_Peri-head-distance; 10.5281/zenodo.20330279
Student Vannas Spring Scissors, straight, 9 cm	Fine Science Tools	Cat#91500-09
Micromanipulator system	Sutter Instrument	Model MPC-325
Microscope	Keyence	Model BZ-X800
Confocal microscope	Carl Zeiss Microscopy	Model LSM700
High-speed camera	Basler	Model acA1300-200um
High-power LED array, 660 nm	Roithner Lasertechnik	Cat#LED660N-66-60
Condenser lens	Thorlabs	Cat#ACL50832U
Telecentric lens	Edmund Optics	Cat#55-349
32-channel silicon probe, Model A1×32-Poly2-10mm-50s-177-A32 (50 μm thickness)	NeuroNexus	Model A1×32-Poly2-10mm-50s-177-A32
32-channel silicon probe, Model A1×32-Poly3-10mm-50s-177 (50μm thickness)	NeuroNexus	Model A1×32-Poly3-10mm-50s-177
32-channel silicon optrode, Model A1×32-Poly2-10mm-50s-413-	NeuroNexus	Model A1×32-Poly2-10mm-50s-413-OA32LP (105 μm/125 μm, NA 0.22)
32-channel NeuroNexus to omnetics adaptor	NeuroNexus	A32-OM32
Headstage, Cereplex M	Blackrock Microsystems	Cat #10532
Tungsten electrode, 0.005″, 5 MΩ, 8° tip angle	A-M Systems	Cat#575300
AC/DC differential amplifier, Model DAM80	World Precision Instruments	Model DAM80
Microsyringe pump	World Precision Instruments	Model UMP3
Syringe pump controller	World Precision Instruments	Model Micro4
Stereotaxic frame	David Kopf Instruments	Model 963
Laser, 473 nm	Cobolt	06-MLD 473
Pulse generator	Doric Lenses	Model OTPG-4
C&B Metabond® Quick Adhesive Cement System	Parkell	SKU: S380

Highlights.

• PrV neurons encode peri-head distance via proximity and map codes

• Map tuning peaks tile peri-head distance within whisker reach

• PrV map populations support precise distance readout

• Long-range inhibition from SpVi helps generate the PrV map code