Whole-body central processing of lateral line inputs encodes flow direction relative to the center-of-mass

Significance

Spatial and directional cues are essential to select appropriate actions in response to changes in the environment. How information from broadly distributed sensors across the entire body is organized and contributes to motor selection remains elusive. By directionally stimulating in the horizontal plane, each and virtually all sensors distributed along the flow sensitive lateral line of larval zebrafish, our study uncovers that spatial and directional inputs from each flow sensor are encoded in the brain relative to the animal’s center-of-mass to subsequently recruit neurons that control motor commands driving forward and turning propulsion. These results establish a framework for understanding how broadly distributed inputs get integrated to recruit motor command neurons responsible for producing diverse behaviors.

Abstract

From shifting visual scenes to tactile deformations and fluid motion, animals interpret patterns of sensory flow around their body to construct internal models and produce adaptive behavior. Understanding how such transformations are encoded in the brain remains incomplete. To tackle this question, we leverage the lateral line of larval zebrafish as a tractable system sensitive to fluid motion and used to steer navigation, feed, and avoid predators. By presenting stimuli of either direction to neuromasts along the body, we mapped hindbrain responses via high-resolution calcium imaging. Our findings challenge the notion that central lateral line processing lacks topographic structure by revealing a simple, yet powerful principle centered on an egocentric spatial framework: the direction and location of local flow motions are encoded in reference to the animal’s center-of-mass. Brainstem neurons that encode flow toward the center-of-mass broadly project to form bilateral connections onto reticulospinal neurons that coordinate forward locomotion while those that encode flow away from the center-of-mass displayed a more selective and unilateral projection profile to command neurons eliciting turns. This simple representation enables the brain to register complex flow patterns and provides a robust basis for behavioral action selection. Our finding represents a shift from purely somatotopic encoding toward an integrative representation of axial position and directionality along the horizontal plane combined, revealing a central principle for encoding spatiodirectional cues. This study advances our understanding of how complex mechanosensory inputs select appropriate motor outputs via simple egocentric neural maps in the hindbrain.

Animals must continuously transform diverse sensory inputs into coherent internal representations of the environment to generate appropriate behavioral responses. To guide survival behaviors such as navigation, feeding, and escape, sensory systems must accurately encode both the location and direction of external stimuli. At the population level, spatially distributed sensors are able to disambiguate the location and direction of a stimulus, yet individual sensory afferent fibers can be easily confounded or remain agnostic to the location and direction of the sensory cue. Integration centers must be efficiently organized within the brain to ease sensory processing and subsequent motor decisions. To integrate the key features relevant for guiding behavior, we must resolve how sensory systems encoding dynamic inputs from sensors distributed across the body are compactly represented within the brain.

One common strategy employed by a diversity of sensory systems is the formation of topographic maps (1). Such maps are well characterized in systems with compact sensory epithelia. In the visual system, retinotopic maps preserve the continuous spatial relationships of photoreceptors in the retina (2). Similarly, in the auditory system, tonotopic maps organize cochlear hair cell positions along the continuum of sound frequencies (3). Other sensory systems integrate discrete stimulus features, such as interaural time and intensity differences, into spatially distinct maps (4, 5). The rodent somatosensory cortex incorporates both continuous and discrete mapping providing precise spatial encoding of whisker stimulation (6, 7). Despite these insights for organization in compact sensory organs, how spatial and directional cues from body-wide sensor arrays are represented in the brain remains incompletely understood. One technical challenge lies in selectively stimulating individual sensors, often overlapping in space and receptive field, at the scale of the entire body plan while comprehensively recording the sensory inputs and interneuron responses in the brain.

To tackle this question, we leverage the lateral line system (LL) of larval zebrafish (Danio rerio), which offers a tractable sensory organ distributed along the entire body platform that is sensitive to fluid motion to mediate behaviors such as navigation (8–10), prey detection (11–13), and predator avoidance (14, 15). Spatial and directional cues from the fluid environment are essential to action selection of appropriate motor responses (10, 15, 16). To sample the location and orientation of the flow, the LL is composed of distributed neuromasts (NMs), comprised of polarized mechanoreceptive hair cells, which are embedded on the head (anterior lateral line; aLL) and body (posterior lateral line; pLL) of the animal to detect and transmit information about changes in flow direction to the hindbrain (17–21). Individual sensory afferents preserve directional cues at the level of the periphery by only innervating hair cells sensitive to the same flow direction ostensibly leading to preservation of directional cues at the primary projection locus (22–24). By revealing how spatial and directional information across the entire body is processed in the brain, we can gain fundamental insights into how multidimensional inputs are integrated to subsequently modulate the activity of motor commands driving locomotion. The primary projection site of the LL is the medial octavolateralis nuclei (MON) located in the rostral-dorsal hindbrain. Recording the MON offers a unique opportunity to dissect causal mechanisms that inform central processing in the brain during sensory perception. Anatomical studies have shown that the aLL and pLL nerves terminate onto distinct MON regions, preserving an incipient somatotopic organization (25, 26), but at the physiological level we have yet to uncover any topographical representation of sensor position or directional cues (18, 27, 28).

To address this gap, we presented stimuli of opposing polarities onto nearly all and each NM innervating the head, body, and tail of larval zebrafish. This approach provided a comprehensive map of MON responses to spatially distributed stimuli across body axes and directions. Precise bidirectional stimulation combined with high-resolution MON calcium imaging revealed a simple yet powerful organizing principle challenging the notion that central LL processing lacks topographic structure. We showed that directional inputs are encoded in reference to the animal’s center-of-mass, establishing an egocentric spatial framework. This representation facilitated the registration of complex flow patterns as motion toward or away from the center-of-mass, simplifying subsequent action selection. By elucidating how spatial and directional cues from a distributed sensory array are mapped in the brain, our work uncovers how vertebrate systems integrate multidimensional inputs to drive context-appropriate motor outputs.

Results

Functional Identification of MON in Larval Zebrafish.

To functionally define the MON loci in 4 to 6 d post fertilization (dpf) larval zebrafish, we deployed controlled flow along the anterior–posterior axis to stimulate individual NMs while simultaneously recording neuronal responses in the hindbrain (Fig. 1A). The anatomical perimeter of the larval zebrafish MON has been originally defined by the primary projection patterns of aLL and pLL afferents (17, 29) whereas the functional space has been probed only via stimulation of tail NMs (18, 27, 28). To access NMs from head to tail, we secured paralyzed larvae to the base of a silicone semicircle using tungsten pins, avoiding the encapsulation of head neuromasts associated with traditional agarose head-restraint (Fig. 1A). We stimulated individual NMs three times for 5 s (18, 30) with 40 s intertrial intervals using a gentle unilateral single-pulse flow ejected from a large-tip (~10 µm) micropipette. In this protocol, we monitored for cupulae and kinocilia deflection along the intended axis (Fig. 1B), measured kinocilia height (18.8 ± 3.6 µm), tip-displacement (13.1 ± 4.7 µm), and deflection angle (47.8 ± 9.3°; Fig. 1C), which overlapped with displacement distances (~8 to 14 µm) sufficient to elicit the entire range of evoked afferent firing rates (30–32). Hindbrain responses were recorded over 10 planes (z = 100 µm) to capture the entire MON calcium signals using the panneuronal, nuclear-tagged calcium indicator Tg(elavl3:H2B-GCaMP6f)^jf7Tg transgenic line (33) (Fig. 1A). As the MON can integrate auditory and vestibular inputs (34) along LL inputs, we verified that our stimuli selectively targeted the LL: we recorded responses to NM stimulation before and after 5 mM bath application of the calcium chelator BAPTA that disrupts tip-links between stereocilia necessary for mechanotransduction in superficial hair cells (35) (Fig. 1 D, i–iii). Flow-evoked responses (Fig. 1E) were significantly reduced following this chemical ablation, confirming that the observed hindbrain activity originated from LL input (Shapiro–Wilk test: P = 0.372; paired t test: t₍₃₎ = 2.75, P = 0.035; Fig. 1F). Across all 572 trials with consistent recruitment, we identified over 3,900 flow responsive neurons across unique NMs (n = 100) and individuals (n = 36). To enable comparisons of response profiles across different stimulus parameters and individuals, we registered flow responsive neurons to a reference atlas (Fig. 1G). Stimulation of a single NM led to consistent activation of ~32 neurons on average across the entire imaging volume. The number of flow responsive neurons upon aLL and pLL stimulation was identical for anterior-to-posterior (tailward) flow or posterior-to-anterior (headward) flow (P = 0.79, P = 0.80, respectively; Fig. 1H), indicating no bias in sensitivity to LL subtype or flow direction. To map the structure of left MON, we generated a 3D mask by applying kernel density estimation (KDE) at the 95th percentile threshold on neurons responsive to ipsilateral NM stimulation (Fig. 1I). This mask reveals functional subdivisions within the MON, highlighting regions of concentrated flow responsive neurons forming localized clusters. Altogether, our map provides a comprehensive representation of the MON in larval zebrafish, integrating both aLL and pLL input location with directional specificity and enabling a detailed exploration of the somatotopic organization and central processing mechanisms of whole-body sensory integration.

Fig. 1.

Hindbrain response properties to single neuromast stimulation with unidirectional flow. (A) Schematic of experimental preparation of a paralyzed larval zebrafish (4 to 6 dpf) to enable stimulation of individual neuromasts (NM, green) along the head and tail. Black box outlines the field of view of hindbrain imaging (see D). (B) Brightfield imaging of a single NM before (i) and during (ii) flow stimulus. Dashed line outlines the perimeter of the cupula and black arrows highlight the angle of the kinocilia deflection. (C) Average kinocilia height (n = 53, 18.8 ± 3.6 µm), tip displacement (13.1 ± 4.7 µm), and deflection angle (47.8 ± 9.3°) across NMs. (D) Hindbrain imaging of panneuronal, nuclear-targeted genetically encoded calcium indicator in Tg(elavl3:H2B-GCaMP6f) transgenic larvae before stimulation (i), during stimulation with an intact LL (ii) and after chemical ablation of the LL (5 mM BAPTA, iii). (E) Calcium transients from D-ii of hindbrain neurons that reliably respond during all three stimulus periods (blue bars represent stimulus period). (F) Box-and-whisker plot representing flow responsive neuron counts before and after BAPTA exposure demonstrates a significant reduction in flow responsive neurons after LL ablation (n = 4 fish/4 NMs, P = 0.03). (G) Reference atlas of all flow responsive hindbrain neurons (n = 3,918) across fish (n = 36) and NMs (n = 100) depicted in dorsal (Left), sagittal (Top-Right), and coronal views (Bottom-Right) color-coded according to depth-of-view. (Scale bars, 50 µm.) (H) Box-and-whisker plots of responsive neuron counts normalized by number of fish does not significantly differ when stimulating aLL (39.97 ± 48.89 neurons) vs. pLL NMs (51.39 ± 67.43 neurons; Mann–Whitney U stat = 649.5, P = 0.72; Top) nor when stimulating with tailward (30.29 ± 34.62 neurons) vs. headward flow (28.30 ± 32.33 neurons; Mann–Whitney U stat = 977.0, P = 0.80; Bottom). (I) Three-dimensional mask of the left medial octavolateralis nucleus (MON) generated from contour maps of the 95th percentile from kernel density estimates (KDE) of coordinates from ipsilateral NM connected neurons in the atlas.

Sensory Somatotopy within MON Represents Sensor Position Across Body Axes.

Lateral line afferents lack contralateral projections maintaining lateralized separation of left vs right prior to projecting to the hindbrain (17), which impacts how the MON processes left/right asymmetries in flow-evoked signals. To verify our method and investigate the recruitment of MON neurons during single-NM stimulation across the left–right axis, we bootstrapped random samples of left vs right NMs on the head and calculated the relative density of tailward flow-evoked responses in the hindbrain (Fig. 2). In the prepontine, pontine, and retropontine (r1–r6) regions, most flow-evoked responses were ipsilateral to the stimulated NMs (P < 0.001; Fig. 2 A–C). A subset (~18%) exhibited contralateral activation (Fig. 2 B, iii and C, iii), verifying that flow-responsive neurons in the MON predominantly encode lateralized inputs and preserve the lateralized segregation of sensory signals that facilitates rheotaxis (10).

Fig. 2.

Flow responsive MON neurons demonstrate ipsilateral somatotopy and reliably predict neuromast position along the rostral–caudal axis. (A) An illustration of a larval zebrafish indicating the position of individually stimulated neuromasts (NMs) located on the left (cyan) or right (magenta) side of the head. (B-i) Contour heat maps in dorsal, sagittal, and coronal views represent the relative kernel density estimates (KDE) of flow responsive hindbrain neurons that responded when stimulating individual NMs (randomly sampled n = 10, bootstrapped n = 1 k) on the left-side of the head. (ii) Histogram visualizes distribution of responsive neurons along the left–right axis. (iii) Quantification of significantly more responsive neurons ipsilateral to NMs stimulated on the left side of the body represented in a box-and-whisker plot (N = 29 fish, n = 54 NMs, n = 1,799 neurons; Mann–Whitney U-stat = 99 × 106, P < 0.001; iii). (C, i–iii) A mirrored pattern of ipsilateral activation along the left–right axis is observed when stimulating NMs on the right-side of the head (N = 6 fish, n = 15 NMs, n = 760 neurons; Mann–Whitney U-stat = 1.5 × 105, P < 0.001). (D) An illustration of a larval zebrafish indicating the position of individually stimulated NMs innervated by either the anterior lateral line (aLL; cyan) or posterior lateral line (pLL; magenta) on the left-side of the body. (E and F) Contour heat maps in dorsal, sagittal, and coronal views represent the KDE of flow responsive hindbrain neurons (n = 2,504) that responded when stimulating individual NMs (N = 33 fish, n = 84 NMs; randomly sampled n = 10 NMs, bootstrapped n = 1 k) from either the aLL or pLL on left-side of the body. (G) Histogram and box-and-whisker plot of responsive neurons normalized by the number of NMs sampled in either the aLL (cyan) or pLL (magenta) along the left–right (F-stat = 0.12, P = 0.73, i), anterior–posterior (F-stat = 6.15, P = 0.01, ii), and dorsal–ventral (F-stat = 12.07, P < 0.001, iii) axes. (H) Scatter plots with line of best fit of average coordinate position of flow responsive neurons along anterior–posterior (left; R² = 0.01, F-stat = 0.11, P = 0.75) and dorsal–ventral (right; R² = 0.45, F-stat = 8.93, P = 0.01) axes with respect to the relative position of each NM from head to tail. (I) Heatmap representing KS-statistics for pairwise comparisons between aLL and pLL NMs coarse-grained distributions along the MON anterior–posterior axis (Top) and dorsal–ventral axis (Bottom). Lower KS values (blue) represent greater similarity, while higher values (red) represent more distinct distributions. (J) Confusion matrices generated from a random forest model representing the moderate probability (71.3 ± 6.6%) of successfully predicting the corresponding LL subtype (aLL = 72%; pLL = 67%; Top) based on coarse-grained coordinates of the responsive neurons increased overall accuracy (79.1 ± 4.9%) when predicting whether the stimulated NM was embedded on the head (81%) or tail (74%; Bottom). (K) A confusion matrix representing a single exemplar iteration of the bootstrapped random forest classifier (SI Appendix, Fig. S4) with an overall accuracy of 55% when predicting NM identity based on the coarse-grained coordinates of the responsive neurons from randomly selected NMs indicated a lack of spatial resolution to reliably differentiate responses across all sampled NMs. Scale bars represent 50 µm and error bars represent SD of the mean.

The LL system extends well beyond the head, raising the question of whether a similar somatotopic organization exists along the anterior–posterior axis, where spatial encoding could support directional flow processing and sensorimotor integration. The spatial distribution of NMs along the anterior–posterior axis suggests that central processing of fluid motion may integrate sensory input according to the detection site. At the periphery, incipient somatotopy is evident in the primary projections of lateralis afferents in the hindbrain (17) between the anterior and posterior LL ganglia. To determine whether the MON shows somatotopic organization of inputs distributed across the entire body, we stimulated individual NMs along the head and tail (Fig. 2D). Stimulation of aLL NMs recruited ventral MON neurons in the prepontine and pontine areas (r1–r3; Fig. 2E) while stimulation of pLL NMs recruited dorsal and posterior MON in the retropontine area (r4–r6; Fig. 2F).

To test whether LL subtype (aLL/pLL) and hindbrain area influenced the spatial distribution of MON recruitment, we analyzed normalized neuron counts along the lateral–medial, anterior–posterior, and ventral–dorsal axes. Similar receptive field distributions occur for both aLL and pLL inputs along the medial–lateral axis, (Fig. 2 G, i) while along the anterior–posterior axis, aLL stimulation preferentially recruited neurons in the anterior MON (P = 0.035; Fig. 2 G, ii). Similarly, along the dorsal–ventral axis, aLL stimulation led to enhanced recruitment in the ventral MON (P = 0.007; Fig. 2 G, iii). Altogether, aLL stimulation preferentially recruits neurons within the anterior-ventral MON whereas pLL receptive fields are more broadly distributed and overlapping with aLL receptive fields.

Examining the response patterns of individual NMs (SI Appendix, Fig. S1) revealed no evidence of a continuous somatotopic map along the anterior–posterior axis of the hindbrain (R² = 0.01, P = 0.75), but a significant gradient along the dorsal–ventral axis (R² = 0.45, P = 0.01; Fig. 2H). Stimulation of anteriormost NMs recruited the ventralmost MON neurons, while the posterior-most NMs preferentially recruited the dorsalmost MON neurons.

To assess whether receptive field distributions differed across NMs, we compared coarse-grained spatial coordinates of neuronal responses and found significant variation (KS-stat = 0.15, P < 0.001). Pairwise comparisons revealed that receptive field distributions for pLL NMs on the head (e.g., MI1, MI2, D1) were indistinguishable from neighboring NMs, while pLL NMs on the tail showed distinct, nonoverlapping patterns (Fig. 2I). The overlap between the LL subtypes thus arises from pLL NMs on the head whose receptive fields resemble more aLL NMs than trunk and tail pLL NMs, supporting evidence for a positional gradient in pLL NM receptive field topography (head vs. tail). We therefore pool all NMs located on the head, irrespective of their ganglionic origin, while treating NMs on the trunk and tail as a separate population (Fig. 2J).

To study receptive field overlap at the single cell level, we identified MON neurons responding to stimulation of multiple NMs. Using coarse-grained neuronal coordinates (10 μm³) to filter for voxels that contained responses to more than one NM within individual fish, most neurons responded to a single NM (1,212 voxels) and few to multiple NMs (108 voxels). Multiresponsive neurons typically responded to stimulation of 2 to 4 NMs among proximal NMs (SI Appendix, Fig. S2A) and rarely located in head and tail (22 voxels; SI Appendix, Fig. S2B). Discrete somatotopic representation of head and tail was tested using a bootstrapped random forest model, trained on ensembles of coarse-grained neuronal coordinates representing responses to each NM. The model achieved high accuracy in classifying whether the stimulated NM belonged to the aLL (72%) or pLL (67%) (overall accuracy of 71.3 ± 6.6%, Fig. 2J). Model performance improved when NMs were grouped by head or tail location (81% and 74% accuracy respectively), suggesting a somatotopic organization based on axial position rather than LL subtype (79.1 ± 4.9% overall accuracy). When predicting individual NMs based on coarse-grained neuronal coordinates, the model achieved low accuracy (51.1 ± 2.6%; Fig. 2K and SI Appendix, Fig. S3A). Mixed-effects modeling indicated that NM identity influenced recall (P < 0.001) where several NMs (i.e., SO1, SO3, MI1, L4, and L6) had distinguishably higher recall (+2%) than baseline (~50%; P < 0.001). However, the overall accuracy to predict each NM remains relatively low (SI Appendix, Fig. S3B) and off-target misclassification errors were more likely to occur between spatially adjacent neuromasts according to head or tail (SI Appendix, Fig. S3C). Therefore, while a general somatotopic organization exists between head and tail, significant overlap of receptive fields among neighboring NMs reduce localization precision across the entire population.

Altogether the receptive fields of head vs tail NMs are functionally distinct. Subsequently, NMs were categorized as belonging to either head or tail based on position relative to the animal’s center-of-mass, ~1 mm from the rostrum within the swim bladder (36, 37). The somatotopic organization in the MON tuned to spatial location along the body axis relative to the center-of-mass sets the stage for exploring how directional flow cues from the head and tail are distinctly represented.

Sensory Somatotopy within MON Represents Flow Direction Relative to Center-of-Mass.

Although few larval zebrafish NMs are sensitive to dorsal–ventral (21, 38), we restricted stimulation in the horizontal plane. While previous analyses had focused on responses to anterior-to-posterior or “tailward” flow, we examined also responses to the reversed flow, posterior-to-anterior or “headward” to investigate how directional cues shape somatotopy in the hindbrain (Fig. 3). Hair cells are sensitive to flow in a single direction which depends on the orientation of the ciliary bundle. Within each NM, hair cells of opposite polarities are antagonistically organized relative to one another and equally represented within each NM (39). Afferents can innervate multiple NMs but only synapse onto hair cells with matching polarity (22–24). Yet, it is unknown whether opposing directional inputs project to distinct regions within the brain. Resolving this question is critical for understanding if the brain preserves directional specificity at early stages of processing.

Fig. 3.

Receptive fields of head and tail change with respect to flow direction. (A) An illustration of a larval zebrafish indicating the position of individually stimulated anterior lateral line (aLL) neuromasts (NMs) with both tailward (cyan) and headward (magenta) flow. Contour heatmaps indicate the relative kernel density estimates (KDE) of flow responsive neurons within the hindbrain in dorsal, sagittal, and coronal views when stimulating individual aLL NMs in the tailward direction (B) or headward direction (C). (D) Histograms of normalized counts of hindbrain neurons that respond to stimulation of NMs located on the head of the animal with either tailward (cyan) or headward (magenta) flow (N = 11 fish, n = 16 NMs, n = 575 neurons). (E) An illustration of a larval zebrafish indicating the position of posterior lateral line (pLL) NMs individually stimulated with both tailward (cyan) and headward (magenta) flow. Contour heat maps indicate the KDE of flow responsive neurons when stimulating individual pLL NMs in the tailward direction (F) or headward direction (G). (H) Histograms of normalized counts of hindbrain neurons that respond to stimulation of NMs located on the body of the animal with either tailward (cyan) or headward (magenta) flow (N = 4 fish, n = 7 NMs, n = 663 neurons). (I) Heatmaps of pairwise comparisons of MON activity using a three-dimensional KDE to determine the proportion of overlap of receptive fields encoding different flow directions along the head and body. Tailward flow demonstrated moderate overlap across the entire MON (42 ± 6%) with a maximum probability density function (pdf) overlap of 4.2 × 10⁻⁶ (i) and headward flow demonstrated low overall overlap (38 ± 6%) with a maximum pdf of 4.2 × 10⁻⁶ (ii). Flow toward or away from the center-of-mass demonstrated moderate overall overlap (44 ± 9% and 41 ± 6%, respectively) particularly in either the anterior-ventral (iii) or the posterior-dorsal (iv) MON with the maximum pdfs at 5.0 × 10⁻⁶ and 6.1 × 10⁻⁶, respectively (indicated by black arrows). (Scale bars, 50 µm.)

To map the recruitment of MON neurons according to stimulus direction, we imaged 4 to 6 dpf Tg(elavl3:H2B-GCaMP6f) transgenic larvae during stimulation of NMs with bidirectional flow (Fig. 3A). To monitor hindbrain responses to both directions on the same NM and fish, we positioned stimulus pipettes on either side of the target NM along the anterior–posterior axis enabled tailward flow followed by headward flow to volley an individual cupula back-and-forth. Upon tailward flow on individual head NMs, recruitment occurred in the prepontine, pontine, and retropontine (rhombomeres r1 to r6) (Fig. 3 A and B), consistent with previous patterns (Fig. 2E). Headward flow on the same head NMs recruited more caudal and dorsal MON neurons in the retropontine and pontine region (r4–r6) (Fig. 3C), resembling recruitment for tailward stimulation of tail NMs (Fig. 2F). Pairwise comparisons of distributions of tailward and headward flow responses on head NMs indicate distinct receptive fields along the anterior–posterior and dorsal–ventral axes within the MON (A–P axis: KS-stat = 0.25, P < 0.001; V-D axis: KS-stat = 0.29, P < 0.001; Fig. 3D). For tail NMs, tailward flow led to dorsal-caudal activation (Fig. 3E) while headward flow recruited a discrete rostral-ventral cluster (A–P axis: KS-stat = 0.35, P < 0.001; V-D axis: KS-stat = 0.32, P < 0.001; Fig. 3 F–H). Altogether, responses to headward flow along the body resemble those of tailward flow along the head.

This unexpected observation motivated closer examination of overlapping receptive fields of flow direction relative to body position. Analyzing neuronal ensembles across different flow directions revealed overlapping neural activity patterns. To quantify this effect, we fit a 3D KDE to the neural coordinates and computed the overlap coefficient as well as the probability density function (pdf). Across the MON, overlap between head and tail responses was generally low, reflecting largely segregated receptive fields (tailward flow: 42%, Fig. 3 I, i; headward flow: 38%, Fig. 3 I, ii), with similar low pdf maxima (~4.2 × 10⁻⁶). The low overlap within tailward and headward flow receptive fields across the entire body further supports a discrete somatotopic organization of stimulus direction across the head and tail. Interestingly, when flow was analyzed relative to the center-of-mass, the overall degree of overlap remained comparable (toward center-of-mass: 44%, Fig. 3 I, iii; away from center-of-mass: 41%, Fig. 3 I, iv), but the locations of overlapping neurons were more concentrated, as reflected in elevated pdf maxima (toward center-of-mass: 5.0 × 10⁻⁶, Fig. 3 I, iii; away from center-of-mass: 6.1 × 10⁻⁶, Fig. 3 I, iv). Considering both metrics together allows us to distinguish global overlap from localized convergence of the responsive neurons which revealed a concentration of responses in either the rostral (rMON) or caudal (cMON) MON.

Pooling responses according to flow direction relative to center-of-mass revealed hotspots of activity within the MON (Fig. 4A). We analyzed high-density regions surrounding the center of distribution of neurons to test for somatotopic organization of head- and tail-responsive neurons within the newly defined subregions. We normalized the NM positions relative to the center-of-mass to allow symmetric scaling of anterior–posterior NM positions regardless of their absolute distance. For neurons that responded to flow toward the center-of-mass (rMON; Fig. 4 B, i) and away from the center-of-mass (cMON; Fig. 4 C, i), we observed a clear somatotopic organization along the mediolateral and dorsoventral axes. Within both subnuclei, neurons driven by head neuromasts were systematically positioned more medially and dorsally than those driven by tail neuromasts (rMON: mediolateral P = 0.02, δ = 0.15; dorsoventral P < 0.001, δ = 0.61, Fig. 4 B, ii; cMON: mediolateral P < 0.001, δ = 0.40; dorsoventral P < 0.001, δ = 0.24; Fig. 4 C, ii). Along the anteroposterior axis, this organization was specific to the rMON, where head- and tail-responsive neurons were strongly segregated (P < 0.001, δ = – 0.48), whereas the cMON showed no separation (P = 0.86, δ = 0.01). Together, head- and tail-responsive populations occupy systematically offset positions within the rMON and cMON providing evidence for somatotopy within each subnuclei that retains positional information about the NM relative to the center-of-mass.

Fig. 4.

Flow responsive MON neurons discretely organize according to flow direction relative to the animal’s center-of-mass. (A) Contour heatmaps indicate the relative kernel density estimates (KDE) of flow responsive neurons within the hindbrain in dorsal, sagittal, and coronal views reveal anatomically discrete receptive fields when stimulating individual NMs toward the center-of-mass (rMON blue, N = 12 fish, n = 20 NMs, n = 650 neurons) or away from the center-of-mass (cMON; red, N = 15 fish, n = 23 NMs, n = 588 neurons). (B) Coordinates of individual responsive neurons to flow toward the center-of-mass color-coded to a normalized distance metric that symmetrically scaled values, enabling a direct comparison of anterior–posterior positioning regardless of magnitude. Neurons outside the high-density region are depicted with higher transparency. Dot diameter represents average response amplitude (0.1 − 1.0 ΔF/F) (i). Histograms of normalized responsive neurons counted across the left–right, anterior–posterior, and dorsal–ventral axes of the rMON (ii). (C) Coordinates of responsive neurons to flow away from the center-of-mass (i). Histograms of normalized neuron counts across cMON axes (ii). (D) Coarse-grained coordinates of responsive neurons to both flow toward and away from the center-of-mass in the same treatment revealing bidirectionally tuned neurons specific to head (green) or tail (magenta; i). Pie-chart demonstrates the low proportion of bidirectionally tuned neurons (6%) relative to those selective to flow toward (48%) or away from (46%) the center-of-mass (ii). (E) A confusion matrix generated from a bootstrapped (n = 100 iterations) random forest model representing the high accuracy (80.3 ± 1.5%) of successfully predicting whether the sensor experienced flow toward (88%) or away from (63%) the center-of-mass based on ensembles of coordinates of the responsive neurons during bidirectional stimulation of individual NMs (i). Predicting flow direction relative to the center-of-mass as well as whether it occurred on the head or tail modestly decreased predictability (69.2 ± 1.9%; ii). (Scale bars, 50 µm.)

We next identified neurons that responded to flow toward and away from the center-of-mass to assess directionally nonselective responses. A small subset of neurons (6%) was bidirectionally tuned cells encoding the location of the stimulus (i.e. head or tail) but activated under both flow conditions (Fig. 4 D, i), compared to the majority that were directionally selective for flow either toward (48%) or away from (46%) the center-of-mass (Fig. 4 D, ii). The small population of bidirectional neurons emphasizes the importance of the egocentric framework but may encode location-based computations independent of flow direction.

To further test whether MON subregions encode both relative flow direction and axial position, we trained a random forest classifier on ensembles of coarse-grained neuronal coordinates representing responses to each NM under both flow directions. Model accuracy substantially exceeded random chance (1/18 classes; 5.5%) when classifying individual NMs and flow direction relative to the center-of-mass (47.9 ± 2.4%). Despite positional information being stored in some capacity within the MON the overall accuracy indicates low spatial resolution to reliably predict stimulus origin at the sensor level (SI Appendix, Fig. S5). High accuracy was achieved when predicting whether the stimulated NM experienced flow toward (88%) or away (63%) from the center-of-mass, with an overall accuracy of 80.3 ± 1.5% (Fig. 4 E, i). Predictive accuracy moderately decreased when classifying head or tail during flow directed either toward (83% for head; 42% for tail) or away (59% for head; 43% for tail) from the center-of-mass, yielding an overall classification accuracy of 69.6 ± 1.9% (Fig. 4 E, i). Rather than encoding axial location, the MON appears to prioritize egocentric directional information, effectively parsing ebbs and flows from the center-of-mass, thereby simplifying the neural representation of distributed sensory input for downstream motor commands.

MON Neurons Project to Reticulospinal Neurons and Higher-Order Processing Centers.

To promptly select the motor output for either forward or ipsi-/contralateral directed turns, it is not known whether MON neurons project onto reticulospinal neurons that integrate inputs from different sensory modalities and higher brain areas and in turn, instruct the spinal circuits producing locomotion (40). We therefore investigated projections from MON to reticulospinal neurons (Fig. 5). Numerous reticulospinal neurons specified by the vsx2 transcription factor (41) demonstrate functional specialization according to birthdate (42). Distinct subsets of vsx2⁺ reticulospinal neurons elicit forward vs. turn bouts and receive direct inputs from mesencephalic locomotor region neurons (43–45).

Fig. 5.

Anatomical connectivity of secondary projections from MON to motor commands referred to as reticulospinal neurons in the hindbrain, vestibular nuclei, interstitial nucleus of Cajal (INC) and torus semicircularis (inferior colliculus). (A) Optical backfills of rostral MON soma (green) reveal extensive projections contralaterally to form putative connections (arrowheads) with vsx2⁺ (magenta) reticulospinal neuron soma such as (i) RoM3, (ii) MiD2i as well as putative axo-dendritic connections to (iii) retropontine neurons ipsilaterally and (iv) rostrally medullary contralaterally (iv). Scale bars represent 10 µm and the asterisk marks the photoactivation site. (B) Dorsal view of a whole brain with secondary projections from the MON color coded according to receptive fields encoding flow toward the center-of-mass (n = 43 neurons; blue) or away from the center-of-mass (n = 61 neurons; magenta) taken from the single-cell atlas MapZeBrain (https://mapzebrain.org/home). (C) Stacked bar charts comparing the proportion of neurons within the rMON and cMON that project ipsilaterally, contralaterally, and bilaterally (i) as well as local, ascending, and descending project profiles (ii). (D) Illustrated schematic of representative projections patterns (SI Appendix, Fig. S6) from the rMON (blue) and cMON (magenta) to different regions within the brain. (E) Dorsal view of rMON projections overlaid with reticulospinal neurons with 3 µm projections of planes that revealed extensive projections to reticulospinal neurons (arrows; i), the interstitial nucleus of Cajal (INC, ii), and bilateral projections to vestibular nuclei (iii). (F) Dorsal view of cMON with unilateral projections to canonical vsx2⁺ reticulospinal neurons (i) and bilateral projections to the vestibular nuclei (ii). (Scale bar, 100 µm.) (G) Single cell traces of rMON (i) and cMON (ii) projections demonstrate unilateral axosomatic overlap with reticulospinal neurons engaged during forward swimming and turns on either the ipsilateral or contralateral side.

To reveal the extent of rostral MON (rMON) projections onto vsx2⁺ reticulospinal neurons, we performed optical backfills of rMON soma in the double transgenic Tg(Cau.Tuba1:c3paGFP; vsx2:LOXP-DsRed-LOXP-GFP)^{a7437Tg;nns3Tg} larvae by leveraging targeted illumination of photoactivatable GFP expressed in all neurons in combination with a red fluorescent protein in vsx2⁺neurons (Fig. 5A). The intense brightness at the photoactivation site hindered ipsilateral projections but uncovered contralaterally projecting axons of rMON neurons overlapping with the soma of vsx2⁺ reticulospinal neurons (Fig. 5A). We found evidence for putative axosomatic connections onto contralateral RoM3 cells driving forward locomotion (Fig. 5 A, i) and MiD2i neurons driving forward locomotion and steering (46) (Fig. 5 A, ii). Posteriorly projecting secondary fibers ran on ipsilateral and contralateral sides suggesting putative axo-dendritic synapses onto medullary vsx2⁺ reticulospinal neurons (Fig. 5 A, iii and iv), known to be recruited during forward locomotion (42, 45).

To study how MON neurons project onto more reticulospinal neurons, we leveraged the virtual single-cell tracing of the publicly available brain atlas mapZebrain (47) (https://mapzebrain.org/home) at the cellular resolution onto the canonical large caliber reticulospinal neurons (48) (Fig. 5B). rMON neurons (n = 43) exhibit an extensive connectivity profile forming robust, bilateral overlap with large-caliber reticulospinal neurons and higher-order processing centers (Fig. 5C).

rMON neurons project both ipsilaterally (30%) and contralaterally (30%), with a majority projecting bilaterally (40%; Fig. 5 C, i and SI Appendix, Fig. S5). Across this broad projection profile, a majority project locally (49%), with a subset ascending to the midbrain and forebrain (40%), and a few descending along the reticulospinal complex (12%; Fig. 5 C, ii). Local projections overlap with the ipsilateral and contralateral MON while ascending projections terminate in the midbrain within the torus semicircularis, the interstitial nucleus of Cajal (INC or also referred to as nMLF (49, 50), and the thalamic region (29) (Fig. 5D and SI Appendix, Fig. S5). The torus semicircularis is an interface for high-order feature detector circuits homologous to the inferior colliculus in Amniotes. The INC serves as a critical integration center for the coordination of eye position, posture, and forward locomotion (51–53) while the thalamus integrates cues across sensory modalities (54, 55). rMON projected on reticulospinal neurons responsible for both forward locomotion and steering (ROM1, MiM1, RoV3, and MiV1-2; Fig. 5 E, i and G, i; arrowheads) (46). The bilateral pattern of rMON projections suggests activation that primarily drives symmetrical left–right recruitment of spinal circuits essential for forward locomotion as well as postural stability (40).

In contrast, the caudal MON (cMON) neurons (n = 61) displayed a more selective and unilateral projection profile (Fig. 5B). The dorsal commissure across the reticulospinal formation primarily projected to a contralateral receptive field within the hindbrain (88%) with a majority of those projections terminating locally (60.7%) and few projecting ipsilaterally and bilaterally (5% and 7%, respectively; Fig. 5C) (29). This pattern demonstrates a substantial bias for contralateral projections relative to the rMON’s bilateral projections. Of the contralaterally projecting neurons, no neurons extended beyond the contralateral MON and torus semicircularis (Fig. 5D and SI Appendix, Fig. S5). The most prominent putative contralateral connections occur onto reticulospinal neurons eliciting turns (RoM1-2, RoV3, and the MiVs; Fig. 5 F, i) (42, 45, 53, 56, 57) with some encoding forward and steering bouts during the optomotor response (MiM1, RoL1, and ROM2; Fig. 5 F, ii and G, ii) (46). Unlike the broader rMON projection pattern, cMON projects contralaterally to contact command neurons relevant for steering.

Notably, rMON and cMON share the characteristic of bilateral projections to the vestibular nucleus, indicating that both MON subdivisions modulate the vestibular system (Fig. 5 E, iii and F, ii). The bilateral innervation pattern suggests a role in coordinating bilateral vestibular processing for maintaining postural adjustments during locomotor maneuvers with direct relevance to the center-of-mass of the animal (58).

Discussion

Integration from Single Sensors Spread Over the Entire Body.

Natural hydrodynamic environments are highly dynamic, shaped by currents and eddies that interact with the body at the fine scale of individual sensory loci. For larval zebrafish, these localized perturbations provide critical spatial and directional cues essential for survival. While flow may engulf the entire body, the lateral line (LL) system does not encode the absolute flow but the local changes in flow at the level of individual neuromasts (59) (NMs). A fundamental question in sensory neuroscience is how the activation of spatially distributed sensors is compactly and reliably encoded in the brain. Inputs of single-sensor is well established in other sensory systems. Single photoreceptors in the mammalian retina can detect individual photons (60), while single olfactory receptors can drive behavioral responses (61). Single sensory units, despite operating within a noisy, redundant system, can carry salient information that shapes perception and behavior. In the LL, single NM deflection is sufficient to elicit escape responses (30), emphasizing its behavioral importance and sensitivity. However, the downstream integration in the brainstem as a function of NM location on the body and directionality of the flow had not been described. By stimulating individual NMs, our results provide a comprehensive anatomical and functional characterization of the MON in larval zebrafish. While it has been hypothesized that MON receptive fields could be organized by both somatotopic position and flow direction to enhance signal detection and noise robustness (62–64), such organization has remained elusive due to limited sampling and the confounding effects of bidirectional or bulk stimulation. By incorporating both aLL and pLL inputs, and systematically probing responses to horizontal flow in opposite directions, we reveal a spatially structured and directionally tuned MON map. This approach clarified how individual sensor responses are represented in the hindbrain and uncovers how MON activity can inform how whole-body, multisensor inputs are processed during natural behaviors.

Previous studies, including brain-wide activity mapping (27), characterized responses to bulk flow across trunk and tail lacking the resolution to detect fine-scale topographic organization within the MON. Single NM stimulation had yielded mixed conclusions: bidirectional mechanical deflection identified directionally selective MON neurons (18), whereas optogenetic activation of individual NMs found no evidence of somatotopy (28). Optogenetic methods however stimulate hair cells of both polarities simultaneously and thereby mask direction-selective responses. Moreover, when investigating central processing in larval zebrafish, the posterior ramus of the pLL had previously been exclusively probed despite representing less than half of all NMs. Contrary to the view of MON as a homogeneous processing center, we reveal anatomically and functionally distinct subregions that differentially encode flow direction as a function of the sensor axial position. By sampling most NMs from aLL and pLL and parsing direction-specific responses, we demonstrate a structured somatotopic and directionally tuned organization. The MON topography provides insights into how the brain integrates distributed sensor inputs into a coherent, egocentric representation of flow to generate adequate motor output.

Topographic Flow Mapping.

Topographic maps are a hallmark of sensory systems, enabling precise localization and discrimination of stimuli based on spatial origin (1). In the LL, such somatotopy has been observed in the midbrain of Teleostean and amphibians (25, 26, 64–66). We find a similar organization in the hindbrain, where MON reveals discrete neural maps that integrate spatial and directional cues into topographically organized receptive fields. A small population of contralaterally responsive neurons is consistent with cross-midline projections thought to mediate lateral inhibition, helping to prevent bilateral coactivation and enhancing directional discrimination (67–69). This lateralized organization is relevant for integrative reflexes such as rheotaxis, which require detection of gradients across the body midline to orient appropriately in flow (10).

The somatotopy we observed along the MON dorsoventral axis during unidirectional tailward stimulation aligns with previous anatomical studies showing that pLL afferents project more dorsally than aLL afferents, which terminate ventrally (17, 18, 25, 26, 70). However, resolving a clear somatotopic representation of the body’s anterior–posterior axis within the MON is more difficult due to overlapping termination fields of aLL and pLL along this axis. Our single-sensor approach revealed substantial functional overlap between aLL and anterior pLL inputs, similar to observations in adult goldfish where some MON neurons responded to both head and rostral body stimulation, while other neurons responded selectively to caudal trunk and tail stimulation (70). These results suggest that somatotopic organization in the MON reflects the positional origin of a sensor (e.g. head vs. tail) more than strict adherence to LL type (aLL vs. pLL). MON somatotopy likely emerges by primary projections from the periphery during development. In the pLL, early-born afferents from the first primordium innervate tail NMs and project dorsally onto the MON, while late-born pLL afferents from the second primordium target head-NMs and terminate more ventrally (24, 71, 72). Similarly, the aLL undergoes a mirrored developmental sequence along the anterior–posterior axis, where anterior NMs are innervated earlier, and posterior aLL NMs later (73). However, aLL development occurs over a narrower window, potentially resulting in less stratification among the ventrally projecting aLL afferents innervating NMs sampled in this study. The anterior–posterior MON topography we identified is thus consistent with the temporal order of LL neurogenesis. Prior to our findings, there was no evidence that polarized lateralis afferents transmitting opposing directional information terminated onto different groups of MON neurons. Determination of the developmental origin of the directional neural map remains challenging given no genetic tools exist to distinguish polarized afferents and their anatomical projections.

Differential Projections Onto Motor Command Neurons and Sensorimotor Integration Centers.

Our findings support two differential patterns of secondary projections along the antero–posterior axis of the MON suggesting complementary yet distinct functional roles in motor control and postural regulation (29). rMON neurons project more extensively to higher-order regions such as the tectum and the thalamus, supporting longer timescale processing that likely integrates LL input with broader sensory and motor contexts. In contrast, the cMON neurons preferentially project directly to the contralateral MON and reticulospinal circuits providing substrate for short-latency recruitment and rapid reflexive adjustments to changes in flow. The organization of these connections provides insight into how MON neurons orchestrate complex locomotor strategies across timescales via modulation of reticulospinal and higher-order motor pathways. The bilateral projection pattern characteristic of rMON neurons likely facilitates forward swimming by providing a symmetrical excitatory drive to high brain centers and reticulospinal neurons to produce coordinated movements typical of forward locomotion (74, 75). Conversely, the predominantly unilateral projections of cMON neurons provide the anatomical framework for generating turns. Asymmetrical activation of reticulospinal pathways through unilateral cMON activation would result in steering maneuvers. The uniquely rMON projections to the interstitial nucleus of Cajal (INC) further expand the postural regulatory capacity of inputs encoding flow toward the center-of-mass. Not only is the INC involved in generating forward locomotion (52, 76), but it also controls head positioning, gaze stabilization, and postural maintenance through its connections with oculomotor and neck motor systems (50–53). Through this diencephalic connection, rMON neurons may influence both forward locomotion and higher-order postural control mechanisms that coordinate head–body relationships and maintain appropriate spatial orientation during locomotor activities in stabilizing flow. This connection suggests that rMON may serve as a critical link between locomotor circuits and postural control systems, enabling integrated control of whole-body movement and posture.

The center-of-mass serves as a universal reference that allows the integration of sensory inputs from different modalities. Vestibular inputs provide information about body tilt, and the LL encodes spatial and directional information about flow. If both are encoded relative to the center-of-mass, the brain can more easily combine these signals to generate appropriate movement responses. In fact, the MON functions as the sensorimotor integration center for both systems and we found bilateral projections to the vestibular nucleus from both cMON and rMON. MON neurons likely modulate vestibular processing and, consequently, postural reflexes and spatial orientation mechanisms. This vestibular influence would be particularly important during dynamic navigation where continuous postural adjustments are required to maintain stability and orientation.

It is important to note that the differences in projection patterns and reticulospinal neuron targets across MON subnuclei are not mutually exclusive. While rMON neurons tend to exhibit a bias toward bilateral projections and cMON neurons toward contralateral projections, both subnuclei also contain neurons with opposing projection patterns (e.g., unilateral–contralateral projections within the cMON, and vice versa). This anatomical overlap suggests that both subnuclei may contribute to varying extents, to forward and turning locomotion. Future studies should aim to characterize the recruitment patterns of the reticulospinal neurons at the population level depending on stimulus direction, sensor position, and stimulus amplitude. Uncovering the stimulus parameters that lead to transitions between forward locomotion and steering or ipsilateral and contralateral recruitment will reveal fundamental principles about sensory mediated action selection.

An Egocentric Directional System Coupled to Motor Commands.

The center-of-mass provides a stable, intrinsic frame of reference that is invariant to body movement and is well suited for sensorimotor processing that supports navigation and interaction with the environment. A key insight from our study is that directional inputs in the horizontal plane to the MON are organized relative to this egocentric reference frame, reflecting a transformation from raw peripheral input into a body-centered representational scheme. Center-of-mass–referenced tuning can serve as an organizing principle that assigns information to local directional signals. Although individual lateral line afferents do not independently encode flow direction (55), the prevalence of monodirectionally tuned hindbrain neurons indicates the MON as the site of early directional processing (27). This organization may reduce reliance on complex computational mechanisms like intrinsic coincidence detectors that compare input across multiple afferents (59) or inhibitory–excitatory ensembles tuned to opposing directions (10, 18). By aligning directional tuning on an egocentric map of sensor position, the MON can extract behaviorally relevant information about fluid motion relative to the body, rather than its absolute spatial source. Encoding flow direction relative to the center-of-mass provides a rapid, local cue to functionally distinct motor command neurons responsible for initiating forward bouts or turn bouts. cMON and rMON may be differentially recruited for respectively reorientation versus postural control phases of behavior. Egocentric encoding of flow direction can act as a complementary signal to bias transitions between navigation strategies, operating alongside established circulation-based and bilateral comparison mechanisms (10, 18, 27).

The integration of egocentric flow direction in MON combined with projections on command neurons in the reticular formation enables to act as a fast, low-dimensional proxy for the complex body–flow interactions, making rapid motor decisions without detailed hydrodynamic inference. Future work using combinatorial, antagonistic cues across rostral–caudal and left–right axes will be critical to test how these subnuclei cooperate in action selection.

Materials and Methods

Detailed methods are provided in SI Appendix. Briefly, single neuromast stimulation was provided via a stimulus pipette connected to a pneumatic microinjector while simultaneously recording change in fluorescence of a calcium indicator in a panneuronal promoter (Tg(elavl3:H2B-GCaMP6f). Response topography was quantified via spatial density estimate analyses alongside machine-learning algorithms to identify predictive classifiers of stimulus source. Last, we leveraged optical backfills and a publicly available brain atlas to characterize MON neuron projection patterns. Code and data can be found here https://github.com/wyartlab/Lunsford-PNAS-2026.

Supplementary Material

Appendix 01 (PDF)

Movie S1.

Rotational-3D video of a whole brain with neuronal projections from the MON color coded according to receptive fields encoding flow toward the center-of-mass (n = 43 neurons; blue) or away from the center-of-mass (n = 61 neurons, magenta) taken from the reconstructions of single-cell morphology from the MapZeBrain atlas (https://mapzebrain.org/home).

Data, Materials, and Software Availability

Code and data can be found here https://github.com/wyartlab/Lunsford-PNAS-2026 (77). All other study data are included in the article and/or supporting information.