The two-chambered heart of the zebrafish embryo starts beating 24 hours post-fertilization, with atrial contraction followed by ventricular contraction, driven by its intrinsic pacemaker [16, 17]. To investigate the emergence of a minimal circuit for neurocardiac control in larval zebrafish, we developed a system, as shown in Figure 1A, that enables the quantification of the heart’s physiological function while simultaneously exposing the animal to visual environmental stimuli. In our setup, the larval zebrafish head is immobilized using agarose, while the tail is free to move (Supp. Fig. 1G). We quantified individual atrial or ventricular contractions based on the reflection of near-infrared light, as shown in Figure 1B. Using this approach, we examined the spontaneous heart rate of fish during development and found, consistent with prior work [18, 19], that basal heart rate first decreases from 4 to 5 days post fertilization (dpf) and increases from 5 to 7 dpf (Supp. Fig. 1B).

To induce transient changes in heart rate, we used a brief UV flash (Fig. 1C), a flash of darkness (Fig. 1D), or looming stimuli (Supp. Fig. 1C-F), which are well-studied visual threats. We subjected larval zebrafish to these stimuli across developmental stages ranging from 4 to 12 dpf. We found that while at 4 dpf, fish are capable of behavioral escapes as marked by tail flicks (Supp. Fig. 1G), their heart rate remained constant (Fig. 1C, D, Supp. Fig. 1C, H). At 5dpf, larval fish start displaying increased heart rate responses to UV flash, dark flash, and looming (Fig. 1C, D, Supp. Fig. 1D, H). These responses change their temporal dynamics and intensity as the fish matures. By 12 dpf, the response is larger, more rapid in onset and adaptation, and more stereotypical than at earlier stages (Fig. 1C, D, Supp. Fig. 1F, H).

To define the minimal neural structure required for heart rate modulation, we examined the spatiotemporal development of cardiac innervation using immunostaining for acetylated alpha-tubulin to label neurites and Myl7 to label the myocardium (Fig. 1E). We observed that cardiac innervation begins at 5 days post-fertilization (dpf), when the first neurites reach the entry of the atrium, known as the sinoatrial plexus (SAP), a region that includes the pacemaker cells [20,21]. As development proceeds, innervation expands in a stereotyped sequence: the atrium is innervated by 7 dpf, followed by the atrioventricular plexus (AVP) near the valve region, and the ventricle by 12 dpf.

Next, we asked whether intracardiac neurons are required for the initiation of heart rate control. To track their development, we used transgenic zebrafish expressing GFP under the control of the phox2bb promoter, which labels all intracardiac neurons (ICNs) [22]. Consistent with our acetylated α-tubulin staining, we observed no cardiac innervation at 4 dpf (Fig. 1F). By 5 dpf, neurites begin to reach the heart, but no intracardiac neurons are yet present (Fig. 1F; Supp. Vid. 1). The first intracardiac neuron migrates into the heart by 7 dpf (Fig. 1F; Supp. Vid. 2), and by 12 dpf, multiple intracardiac neurons are evident (Fig. 1F; Supp. Vid. 3). To visualize neuronal somata more clearly, we performed immunostaining for Phox2bb alongside Myl7 to label the myocardium (Fig. 1G). At 5 dpf, no neuronal cell bodies were observed within the heart. By 12 dpf, we identified 6 neurons near the SAP, 6 near the AVP, and a cluster of 16 neurons in the dorsal-posterior atrium, which we designate the Intracardiac Ganglion (ICG).

To summarize the timing and variability of cardiac innervation and neuronal migration, we quantified developmental progression across animals (Fig. 1H, I): at 4 dpf, 0/12 fish showed cardiac innervation; at 5 dpf, 12/15 fish showed innervation; at 7 dpf, 14/20 fish had at least one intracardiac neuron; and by 12 dpf, all 12/12 fish displayed robust innervation of both cardiac chambers with multiple intracardiac neurons. Together, these results show that functional cardiac innervation precedes the arrival of intracardiac neurons, establishing that early modulation of heart rate occurs through direct descending motor pathways alone.

We next sought to determine which branches of the autonomic nervous system contribute to progressive innervation of the heart between 5 and 12 dpf. Because the motor vagus nerve is cholinergic [10], we began by labeling it with antibodies against choline acetyltransferase (ChAT), the enzyme responsible for synthesizing acetylcholine, and labeling the myocardium with antibodies against Myl7. This approach allowed us to visualize cholinergic innervation of the heart beginning at 5 dpf (Fig. 2A).

We also detected ChAT-positive intracardiac neurons (ICNs), including the first neuron to arrive at 7 dpf, suggesting that this cell is a postganglionic neuron of the motor vagus nerve (Fig. 2A–C). Motor vagus postganglionic neurons are known to be cholinergic and receive input from vagal preganglionic fibers and form direct connections with cardiac tissue [23]. By 12 dpf, we observed six ChAT-positive ICNs near the sinoatrial plexus (SAP). We confirmed the cholinergic identity of the first ICN using fluorescent in situ hybridization chain reaction (HCR) for ChAT, together with GFP boosters to confirm overlap with ChaTA-Gal4, UAS-GFP and phox2bb:GFP (Supp. Fig. 2). We did not detect neurons by the atrioventricular plexus (AVP) or elsewhere, suggesting that the other ICNs identified in Fig. 1G may use alternative neurotransmitters.

We then used ChaTA-Gal4, UAS-GFP transgenic fish to label motor vagal fibers selectively. Because red fluorophores in the heart interfered with the detection of thin axons, we instead imaged the myocardium in transmitted light using a photodiode-based detector, while capturing green fluorescence in a separate channel. This strategy enabled reliable identification of even fine cholinergic neurites, based on their anatomical position and rhythmic motion during the heartbeat. Consistent with our immunostaining and phox2bb:GFP imaging, we observed no innervation at 4dpf (Fig. 2D) (Supp. Vid. 4), emerging innervation at 5 dpf (Fig. 2D, Supp. Vid. 5), emergence of the first ICN at 7 dpf (Fig. 2D, Supp. Vid. 6), and further extension to the AVP by 12 dpf (Fig. 2D, Supp. Vid. 7). We did not detect cholinergic innervation of the ventricle at any stage up to 12 dpf.

To quantify the developmental progression of cholinergic cardiac innervation across individual animals, we scored each fish for the presence of cardiac innervation and the number of cholinergic intracardiac neurons (ICNs) observed at 4, 5, 7, and 12 dpf (Fig. 2E). At 4 dpf, the vast majority of fish (14/15) lacked any evidence of cholinergic cardiac innervation. By 5 dpf, most fish (12/14) displayed innervation of the sinoatrial region, but none had ICNs. At 7 dpf, 11 of 17 fish showed a single cholinergic ICN, while the remaining 6 had innervation but no ICNs. By 12 dpf, all fish exhibited robust innervation of the atrium and SAP, with 11 of 14 fish containing between two and six ICNs. These data highlight a gradual and stage-specific acquisition of cholinergic innervation and ICNs, with clear stereotypy in the timing of each step in the maturation of the vagal motor circuit.

To investigate if the first ICN already receives neural inputs, we performed serial-section electron microscopy (ssEM) of the SAP in a 7 dpf fish (Fig. 2G). We identified five cell types—cardiomyocytes, epithelial cells, blood cells, endothelial cells, and one neuron, consistent with our light microscopy observations (Fig. 2G). Vesiclerich synaptic contacts were found both onto the ICN (Fig. 2H) and from the ICN onto cardiomyocytes (Fig. 2I, J), supporting its integration as a functional node in the developing cardiac motor circuit.

To explore how this first anatomical evidence of innervation correlates with the onset of physiological control, we used calcium imaging to measure the activity of motor vagus neurons in the brain. Previous anatomical studies have shown that motor vagus neurons in larval zebrafish are cholinergic and located in the hindbrain [24]; thus, we focused on that population. In order to characterize the dynamics of those neurons, we customized a 2-photon imaging rig with an optical path for simultaneous quantification of cardiac function using near-infrared light and brain functional imaging at single-cell resolution using the 2-photon laser (Fig. 3A). Using ChaTA-Gal4 to drive UAS-GCaMP6s expression, we selectively imaged cholinergic hindbrain neurons while concurrently monitoring heart rate. Dark flash stimulation elicited robust tachycardic responses in 5, 7, and 12 dpf fish (Fig. 3C-E) but not in 4dpf fish (Fig.3B), consistent with our results in Fig. 1. Neural responses were diverse and even when large fractions of neurons displayed synchronous increases in activity, those were not consistently accompanied by changes in heart rate (Fig. 3B-E).

We asked whether motor-vagus activity carries a time-resolved signature of cardiac control. To do so, we fit a generalized linear model (GLM) that decodes heart-rate (HR) dynamics from the population activity, assigning each neuron a causal finite-impulse-response kernel that quantifies its influence across time lags. A static sigmoid nonlinearity captured saturation in HR output. To visualize the temporal motifs, we pooled predictive neurons—those exceeding an Otsu threshold on cross-validated predictive power (Methods, Decoding GLMs)—and plotted their kernels as a shape-sorted heatmap (Fig. 3F). This revealed a single canonical waveform that recurred across cells with opposite polarities: some neurons expressed the positive version and others the negative of the same kernel (Fig. 3F-H).

Model performance was evaluated with blocked 5-fold cross-validation (80% train, 20% test per fold; Methods). Across 31 fish (5, 7, and 12 dpf), only 2 fish at 5 dpf and 2 at 7 dpf fell below CV R² = 0.20; in all others, the model predicted HR trends with CV R² ranging from 0.25 to 0.60 (Fig. 3I). Even for animals near the cohort median (CV R² = 0.35 and 0.43), the predicted traces closely tracked stimulus-evoked HR responses (Fig. 3J). Although predictive neurons (defined by the Otsu threshold on CV performance) comprised <5% of recorded cells at each age, their prevalence tripled at 12 dpf relative to 5 dpf (Fig. 3K). Among encoders, the fraction with negative kernels—the HR-decreasing sign—rose from ~40% at 5 dpf to ~90% at 12 dpf, consistent with the cardioinhibitory role of the motor vagus (Fig. 3K). Spatially, encoders were concentrated medially across ages; positive-kernel encoders were more anterior and dispersed, whereas negative-kernel encoders formed a reproducible medial cluster (Fig. 3L–O).

Given the stereotyped, medial clustering of negative-kernel encoders and the established cardio-inhibitory role of the motor vagus, we hypothesized that activating this population would reduce heart rate. We expressed the light-gated channel CoChR under ChaTA-Gal4 in cholinergic hindbrain neurons and delivered patterned photostimulation to the negative-kernel locus using structured illumination steered by dual galvos (Fig. 3P,Q). At 4 dpf, before anatomical heart innervation is detectable, photoactivation of this locus did not measurably alter heart rate (Fig. 3R). By 7 dpf, once innervation is present, the same stimulation decreased heart rate (Fig. 3S). As a control, optogenetic activation of the anterior spinal cord did not result in changes in heart rate (Supp. Fig. 3). These data indicate that a small, medial subset of cholinergic hindbrain neurons functions as a vagal premotor nucleus for cardiac control.

In zebrafish, the sympathetic nervous system consists of a chain of paravertebral ganglia that extends from the visceral cavity to the tail (Fig. 4A). As in other vertebrates, these ganglia can be visualized using antibodies against tyrosine hydroxylase (TH), the rate-limiting enzyme in catecholamine biosynthesis [26]. To determine when the heart first receives sympathetic input, we performed TH immunostaining alongside Myl7 to label myocardial tissue. At 5 dpf, no TH-positive fibers were detected in the heart (Fig. 4C). By 7 dpf, TH-positive projections reached the sinoatrial plexus (SAP), indicating the onset of sympathetic cardiac innervation (Fig. 4C). These fibers originated from the most anterior sympathetic ganglia, which we refer to as the anterior paravertebral ganglia (APG) (Fig. 4C). Over development, the APG expand and adopt a distinctive C-shaped morphology (Fig. 4B). Despite this morphological maturation, TH-positive projections remain restricted to the SAP through 12 dpf, with no evidence of innervation to the atrioventricular plexus or ventricle.

To quantify the developmental dynamics and interindividual variability of sympathetic innervation, we assessed individual animals at 5, 7, and 12 dpf and recorded the presence or absence of TH-positive fibers at the heart. At 5 dpf, none of the nine animals examined showed sympathetic innervation. By 7 dpf, 6 of 13 animals exhibited TH-positive fibers at the SAP, indicating a variable onset of innervation across individuals. By 12 dpf, all 10 animals displayed consistent TH-positive projections to the SAP (Fig. 4D, E). These findings reveal a gradual and variable onset of sympathetic connectivity to the heart, with full anatomical incorporation of the APG circuit established by 12 dpf.

To investigate how the anterior paravertebral ganglia (APG) acquire the ability to regulate cardiac function, we expressed GCaMP6s under the TH1-Gal4 driver and we simultaneously imaged sympathetic neuron activity and heart rate dynamics in response to brief UV flashes (Fig. 5A). At 5 dpf, the APG neurons respond to the UV stimulus with transient increases in activity (Fig. 5B), suggesting early engagement of the sympathetic system in fight-or-flight visceral responses even before its anatomical integration with the heart. At 7 dpf, sympathetic neurons respond with apparent faster dynamics (Fig. 5C). By 12 dpf, as the ganglion increased in size, we observed emergent heterogeneity: newly appearing neurons decreased their activity following the same stimulus (Fig. 5D), hinting at the maturation of diverse sympathetic subtypes.

To further investigate the diversity of neural dynamics and their temporal relationship with heart rate, we trained a generalized linear model (GLM) to predict heart rate from the activity of all recorded APG neurons. This model estimates a temporal kernel for each neuron that captures how its activity at different time lags influences heart rate. Pooling across animals revealed a striking diversity of kernel shapes (Fig. 5E). Unsupervised clustering of kernels identified six reproducible motifs (Fig. 5F, Supp. Fig. 4). Two motifs were biphasic with opposite polarity (C1, C4), resembling derivative filters that emphasize the effect of changes in neural activity. Two were monophasic bumps (C2, C5), consistent with proportional decoding of magnitude. The remaining two were monotonic ramps (C3, C6), consistent with integrator functions that accumulate activity over time. Together, these motifs map onto canonical motor control operations—derivative, proportional, and integrator—that may form the building blocks of cardiac control.

In our decoding model, neural activity is passed through its assigned kernel, summed across neurons, and then transformed by a static sigmoid nonlinearity to predict heart rate (Fig. 5G). Cross-validated performance improved significantly with age, with CV-R² increasing from 0.2 at 5 dpf to ~0.5–0.6 at 12 dpf (Fig. 5H), indicating that APG neurons acquire predictive power as circuits mature. To illustrate the predictive power of the model even with modest CV-R², we display the heart rate over time (in blue) and the heart rate predicted by the model (in orange) (Fig. 5I). Cluster composition also shifted across development: derivative motifs were absent early (except for one neuron) and emerged at 7 and 12 dpf (Fig. 5J), suggesting that the APG progressively establishes the computational repertoire needed for cardiac control.

To investigate whether specific neuron types are spatially organized in the ganglion, we plotted the imaged APGs coloring each neuron by cluster. We found that locations are variable per fish and per side, without a stereotypic location (Fig. 5K-M).

Last, to directly test the causal role of sympathetic neurons in modulating heart rate, we optogenetically activated APG neurons expressing CoChR under the TH1-Gal4 promoter. Using galvo mirrors, we delivered spatially targeted stimulation to the APG while simultaneously recording heart rate (Fig. 5O). At 5 dpf, stimulation failed to evoke any change in cardiac output (Fig. 5P), consistent with the absence of sympathetic innervation. At 7 dpf, however, optogenetic activation induced modest heart rate increases (Fig. 5Q), which became progressively larger by 9 and 12 dpf (Fig. 4R). Importantly, this response was abolished by propranolol, a β-adrenergic receptor antagonist (Supp. Fig. 5), confirming that sympathetic neurons modulate cardiac function via β-adrenergic signaling.

The somata of the sensory vagus are organized into four ganglia per side. The largest and most posterior is the nodose ganglion. It is located below the lateral line ganglion and innervates the heart and most of the viscera. The three most anterior ganglia are the epibranchial ganglia, which innervate each of the gill arches and are believed to track oxygen and other ventilatory parameters [28] (Fig. 6A). To visualize vagal ganglia, we expressed GFP under the control of the vglut promoter, which labels the cranial sensory neurons [29]. To assess the developmental timing and variability of cardiac innervation by the sensory vagus nerve, we examined the presence of sensory axonal projections to the heart at 5, 7, and 12 dpf. At 5 dpf, all fish (8/8) had formed vagal sensory ganglia but lacked projections to the heart, indicating that ganglion specification precedes target innervation. By 7 dpf, only 2 of 10 animals exhibited cardiac sensory innervation, revealing substantial interindividual variability in the onset of sensory connectivity. By 12 dpf, all fish (11/11) showed robust vagal sensory projections to the heart, demonstrating that sensory innervation becomes fully established at this stage.

In order to investigate how vagal sensory neurons may encode heart rate dynamics, we expressed GCaMP6s under the control of the vglut2-Gal4 driver. We selected this driver because in zebrafish, all cranial sensory neurons, including vagal sensory neurons (VSNs), are glutamatergic [29]. We did not restrict our experiments to 12dpf – when vagal sensory innervation of the heart is established – because heart rate information could be available to VSNs from innervation of vasculature, gills, or other indirectly coupled structures. We stimulated fish with UV-flashes while recording calcium activity in the vagal ganglia concomitantly with heart rate (Fig. 7A). We found robust heart rate responses across development, from 5 to 12 dpf, and diverse neural dynamics, including subpopulations of VSNs that displayed temporally locked responses (Fig. 7B-D). Using generalized linear models (GLMs), we asked how well heart-rate fluctuations predict neural activity, assigning each neuron a causal temporal kernel. Neurons exceeding a cross-validated performance threshold (CV R² > 0.2) were classified as cardiac-encoding. The proportion of encoding neurons was comparable at 5, 7, and 12 dpf (Fig. 7E), indicating that subsets of VSNs track heart-rate trajectories even before direct sensory innervation of the heart is fully established.

Kernel heatmaps aligned by shape revealed that the diversity of response filters could be described as a single temporal motif expressed at systematically shifted phases (Fig. 7F). In other words, rather than multiple distinct kernel families, each encoding neuron appeared to express the same biphasic kernel advanced or delayed in time. Principal-component analysis confirmed this interpretation: the first two PCs accounted for nearly all variance (Supp. Fig. 6) and arranged the kernels along a circular trajectory in phase space (Fig. 7G). This geometry is consistent with a basis set of sinusoidal quadrature components, where neurons are distributed around the circle such that each one samples a different phase of the same underlying waveform (Fig. 7H).

The quality of this representation improved across development. Model performance increased, with crossvalidated (CV) R² values rising from ~0.3 at 5 dpf to >0.5 by 12 dpf (Fig. 7I), indicating that HR became a progressively stronger predictor of neural activity. To illustrate the predictive power of the model even with CVR² values near the median, we display the ΔF/F₀ of a neuron with phase π (top) and a neuron with phase 0 (bottom) over time (in black) and the ΔF/F₀ predicted by the model (in green) (Fig. 7J). In both cases the signal is tracked accurately, with every large fluctuation in measured ΔF/F₀ matched by a large fluctuation in the predicted ΔF/F₀. This is also important to illustrate that neurons with phase π decrease activity when heart rate increases following the dynamics of the canonical kernel, and neurons with phase 0 increase activity when heart rate increases, also following the dynamics established by the canonical kernel.

The distribution of kernel phases also changes systematically during development: at 5 dpf, neurons did not occupy phases uniformly, but instead showed a bias toward 0 and π (Fig. 7K, blue radial histogram). By 7 dpf, these biases strengthened (Fig. 7K, green radial histogram), and by 12 dpf neurons clustered into two modes of representation: a positive biphasic kernel (at phase 0) and a negative biphasic kernel (at phase π) (Fig. 7K red radial histogram). We confirmed this observation using a pairwise permutation test on the kernel phases, which showed that 5dpf fish have a different distribution than fish at 7 and 12 dpf (Fig. 7K inset heatmap), consistent with emergent bimodal encoders with opposite phases.

Anatomical mapping showed no spatial segregation of phase groups: neurons from different ensembles were intermixed within vagal ganglia and across both sides (Fig. 7L–N). Thus, the phase structure of the code is functional rather than anatomical, arising from temporal offsets in kernel phase rather than physical clustering.

Together, these findings demonstrate that vagal sensory neurons implement a distributed, phase-shifted population code for heart rate. By expressing a single canonical kernel at different phase offsets, the ensemble provides a robust and redundant representation of cardiac dynamics, well before direct cardiac sensory innervation is fully stabilized, and, as the fish matures, the encoding shifts to a bimodal representation with kernels of opposite phases.

The observation that a single kernel shape, expressed at shifted phases, accounts for heart rate encoding across fish and developmental stages suggests that this is a core computational property of vagal afferents. To test this hypothesis, we asked whether the same phase-shifted kernel emerges under distinct modes of heartrate modulation: (1) spontaneous fluctuations, which occur without environmental perturbations (Fig. 8A); (2) fluctuations elicited by mosaic optogenetic activation of small subsets of cardiomyocytes (Fig. 8B); and (3) acute tachycardia triggered by brief optogenetic activation of the entire myocardium (Fig. 8C). Spontaneous fluctuations were chosen because, in contrast to fight-or-flight responses, they occur while the external sensory environment is constant and therefore span a broad range of amplitudes, durations, and temporal profiles (Fig. 8D). Optogenetic manipulations were included because they perturb the heart directly and independently of external stimuli. For these experiments, we generated transgenic zebrafish expressing the light-gated cation channel CoChR under the cmlc2 promoter, which labels cardiomyocytes [30]. Mosaic activation produced local perturbations that did not evoke immediate tachycardia but instead generated delayed fluctuations, sometimes tens of seconds after stimulation (Fig. 8E). In contrast, global heart activation produced immediate and robust tachycardia (Fig. 8F). Together, these three conditions provided diverse trajectories of heart rate and neural activity (Fig. 8D–F) without altering the external environment, allowing us to probe encoding mechanisms under controlled and distinct perturbations.

We next trained GLMs, as in Fig. 7, to identify causal kernels that predict neural activity from heart-rate dynamics. Across all conditions, kernel heatmaps revealed the same smooth green↔magenta gradient (Fig. 8G–I) seen previously under UV stress (Fig. 7F), indicating that most encoding neurons are well described by a single canonical biphasic filter, differing primarily in phase (latency/polarity). PCA of kernel shapes confirmed this structure: in all three conditions, kernels distributed along a circular trajectory in PC space, with neurons clustering near phase 0 and π (Fig. 8J–L), as observed in 12 dpf fish during UV stimulation (Fig. 7K). Encoding strength was similar across conditions, with median CV R² values in the 0.4–0.55 range (Fig. 8M). Moreover, the canonical kernel identified in each case converged to the same shape (Fig. 8N).

As in the neural recordings of responses to UV flashes, anatomical mapping showed no spatial segregation of phase groups: neurons from different ensembles were intermixed within vagal ganglia and across both sides (Fig. 8O–Q). Across fish, we observe a higher likelihood of neurons being located on the ventral side of the nodose ganglion, however, there are also neurons in more dorsal parts of the ganglia (Fig. 7L-N, Fig. 8O-Q).

These results demonstrate that vagal sensory neurons implement a stimulus-invariant, phase-shifted (latency/polarity) population code for heart rate. Whether fluctuations arise from environmental stressors, spontaneously, from localized perturbations, or from optogenetically-induced tachycardia, the dynamics of VSN encoding are well described by a linear time-invariant filter set, with phase (latency/polarity) offset as the primary degree of freedom.

Our findings demonstrate that the first stage of functional cardiac control in zebrafish is mediated exclusively by descending vagal motor neurons (Fig. 1 and 2). By 5 dpf, cholinergic fibers reach the sinoatrial plexus, and vagal activity becomes capable of altering heart rate even in the absence of intracardiac neurons. Optogenetic activation of hindbrain cholinergic neurons confirmed that these descending projections are sufficient to slow the heart, revealing a minimal efferent circuit. This stage is particularly striking because it separates the earliest onset of neural control from the later incorporation of intracardiac ganglia. In mammals, parasympathetic ganglia embedded in the cardiac plexus are thought to be essential for modulation of pacemaker activity [31, 32], but our results suggest that brain-to-heart projections alone can transiently carry this function. Such a configuration may provide a rapid developmental solution for ensuring that cardiac activity is already under neural control during the critical window when larvae first engage with the external environment.

In parallel with the onset of vagal motor projections, we identified a small medial cluster of hindbrain neurons whose activity reliably predicts cardiac inhibition. Encoding models revealed a canonical kernel with negative polarity, and targeted optogenetic activation of this locus elicited bradycardia in animals as early as 7 dpf (Fig. 3R). These data indicate that a premotor nucleus in the hindbrain provides a dedicated node for descending cardiac control. The temporal precision and polarity of these kernels suggest that the premotor nucleus exerts a gain-limited inhibitory drive on the heart, consistent with the well-established cardioinhibitory role of the motor vagus. Although the molecular identity of these premotor neurons remains to be fully resolved, their reproducible medial location and robust functional impact indicate that they are a conserved substrate for parasympathetic regulation.

The second stage of circuit assembly involves the arrival of sympathetic projections from the anterior paravertebral ganglia (Fig. 4). Anatomical innervation emerges and optogenetics show that sympathetic neurons develop the capacity to accelerate the heart at 7dpf. Importantly, encoding analyses of neural activity uncovered kernels with proportional, integrative, and derivative-like temporal motifs. This computational repertoire mirrors classical control architectures, suggesting that the sympathetic system contributes not only to the magnitude but also to the dynamics of cardiac responses. Such motifs may allow the system to respond proportionally to changes in heart rate, integrate sustained deviations, or accentuate rapid fluctuations, thereby shaping the gain and stability of autonomic control. Mammalian baroreflex studies indicate modulation of gain and sensitivity that is consistent with proportional control, and dynamic changes in reflex latency under different physiological conditions, suggesting that components of the control motifs we observe may be conserved [33, 34]. Our findings are the first evidence for heart rate integral- or derivative-like temporal dynamics in sympathetic neurons.

Although our primary focus was on the staged emergence of autonomic control, our results also provide new insight into the developmental progression of the intracardiac nervous system (ICNS). By tracking phox2bb+ intracardiac neurons alongside functional assays of heart rate control, we found that cardiac innervation is first detectable at 5 dpf, but intracardiac neurons are absent at this stage. The earliest intracardiac neuron appears by 7 dpf, and by 12 dpf multiple clusters are evident, including a reproducible population in the dorsal–posterior atrium that we designate the intracardiac ganglion. Thus, the arrival of ICNs lags behind the onset of functional vagal efferent control, which is already functional by 5 dpf. This indicates that the initial establishment of parasympathetic influence over heart rate occurs independently of intracardiac ganglia, with these neurons contributing only after the basic circuit is in place.

In mammals, intrinsic cardiac ganglia are thought to be essential for modulation of pacemaker activity and for shaping the spatiotemporal dynamics of cholinergic input to the sinoatrial and atrioventricular nodes [31, 32]. Our findings suggest that in zebrafish, these neurons are dispensable for the earliest phases of functional control. They likely refine or diversify cardiac modulation as they integrate into the circuit. This sequence may reflect a developmental principle: early reliance on extrinsic parasympathetic fibers to ensure inhibitory capacity, followed by later recruitment of intracardiac ganglia for fine-tuned control.

The final stage in the developmental sequence is the engagement of vagal sensory neurons (VSNs). By 12 dpf, sensory axons reliably reach the sinoatrial and atrioventricular plexuses, but encoding analyses revealed that VSNs can track heart-rate fluctuations even at earlier ages, likely through indirect coupling via vasculature, gill innervation or innervation of other tissues with heart rate-coupled dynamics (Fig. 6-8). Across spontaneous fluctuations, responses to visual threat, local optogenetic perturbations, and optogenetically-induced tachycardia, a single biphasic kernel accounted for neural responses. Neurons differed only in phase, tiling the kernel at systematic offsets. This phase-shifted organization sharpened with development, evolving into a bimodal distribution centered at opposite polarities. Such a population code allows continuous coverage of heart-rate dynamics while remaining robust to the source of modulation. The emergence of this stimulusinvariant representation demonstrates that VSNs contribute a stable and generalizable code for cardiac interoception.

We term this computational property of heart rate encoding, chronoreception by analogy to chemoreception and baroreception, and consistent with the use of the word “chronotropic” in reference to changes in heart rate [35]. To our knowledge, such heart rate-responsive cells have not been previously described. Hemodynamically, it makes sense to be able to ascertain heart rate directly because, along with contractility, it is one of the two variables that determine cardiac output. Regardless of whether heart rate is detected at the heart or in other structures, the heart rate encoders are present across conditions and are inputs to the brain.

Together, these results reveal a temporally ordered program for assembling heart–brain circuits. First, descending motor vagus fibers establish a minimal efferent control structure with a hindbrain premotor nucleus providing inhibitory drive (Fig. 9, 5dpf). Second, sympathetic inputs add modulatory power to gain and dynamics (Fig. 9, 7dpf). Third, vagal sensory neurons innervate the heart, they implement a distributed, phaseshifted code for heart rate feedback even before stable heart innervation (Fig. 9, 12dpf). Each stage introduces new computational capabilities without disrupting earlier ones, allowing larvae to maintain functional control as circuit elements mature. This progressive layering provides a framework for understanding how regulation of internal organs is assembled during development.

The sequence we describe parallels and opportunities for the study of autonomic development in mammals. In rodents and humans, autonomic innervation of the heart arises prenatally, making it difficult to observe the onset of function in vivo [7]. Fetal studies suggest that vagal tone emerges early and that sympathetic influences strengthen later in gestation [36]. Sensory encoding of blood pressure has been inferred from nodose baroreceptor recordings in mammals, which respond to beat-to-beat pressure oscillations [37, 38], but direct demonstration of heart rate dynamic encoding has been lacking. Our results provide the first continuous, cell-resolved account of how HR is encoded in a vertebrate vagal sensory ganglion.

Recent studies have shown that interoceptive signals from the heart influence emotion, cognition [39], and behavior [40], and disruptions in cardiac sensory coding have been implicated in conditions ranging from anxiety to dysautonomia [41, 42]. The discovery of a canonical, phase-shifted kernel code for heart rate provides a mechanistic basis for how the brain may reconstruct cardiac dynamics from distributed afferents. In disorders such as postural orthostatic tachycardia syndrome [43, 44] or long-COVID dysautonomia [45], where feedback control of heart rate is impaired, alterations in this encoding scheme may contribute to pathophysiology. Similarly, aberrant development of sympathetic motifs or premotor inhibitory control could destabilize cardiac regulation. By establishing zebrafish as a model in which all three limbs of the autonomic circuit can be interrogated optically, our work opens the door to dissecting how genetic, developmental, or environmental perturbations disrupt interoceptive coding.

We have shown that the vertebrate nervous system assembles neural circuits for heart–brain regulation through a staged sequence that first establishes descending control, then incorporates sympathetic modulation, and finally engages sensory nerves. Each stage introduces distinct computational motifs: proportional kernels in the motor vagus, proportional-integral-derivative motifs in the sympathetic ganglia, and phase-shifted population codes in vagal sensory neurons. This architecture ensures that cardiac control is robust, adaptable, and ready to support survival as soon as larvae begin to interact with their environment. More broadly, these findings provide a framework for probing the developmental logic of interoceptive circuits and for exploring how their disruption contributes to autonomic disease.

Here we report the number of animals used for results that do not explicitly have the number of animals in the main figure.