Distributed control circuits across a brain-and-cord connectome

Abstract

Just as genomes revolutionized molecular genetics, connectomes (maps of neurons and synapses) are transforming neuroscience. To date, the only species with complete connectomes are worms^1-3 and sea squirts⁴ (10³-10⁴ synapses). By contrast, the fruit fly is more complex (10⁸ synaptic connections), with a brain that supports learning and spatial memory^5,6 and an intricate ventral nerve cord analogous to the vertebrate spinal cord^7-11. Here we report the first adult fly connectome that unites the brain and ventral nerve cord, and we leverage this resource to investigate principles of neural control. We show that effector cells (motor neurons, endocrine cells and efferent neurons targeting the viscera) are primarily influenced by local sensory cells in the same body part, forming local feedback loops. These local loops are linked by long-range circuits involving ascending and descending neurons organized into behavior-centric modules. Single ascending and descending neurons are often positioned to influence the voluntary movements of multiple body parts, together with endocrine cells or visceral organs that support those movements. Brain regions involved in learning and navigation supervise these circuits. These results reveal an architecture that is distributed, parallelized and embodied (tightly connected to effectors), reminiscent of distributed control architectures in engineered systems^12,13.

Extended Data Fig. 1:. central nervous system connectome generation, quality, and neuron identification

a.Workflow for serial EM dataset generation. The specimen is dissected and prepared for sectioning and EM imaging. Acquired EM micrographs are then aligned into a dataset, which is subsequently segmented into cellular fragments. b.Density of the normalized NBLAST scores of ‘proofread’ neurons in the BANC against all neurons in other connectomic datasets (different colors). We consider normalized NBLAST scores > 0.3 as high and suggest score bins to help guide data users (dashed lines). Normalized NBLAST scores are “raw” NBLAST scores divided by self-match score. All density curves are normalized to their own peak. c. We sampled 4648 postsynaptic links evenly across 67 standard neuropils^, for a false positive review (_L, left neuropil, _R, right neuropil). d.Attachment rates for presynaptic (red) and postsynaptic (navy) links to an identified cell (neuron, glia) across neuropils. We used the BANC synapse version: synapses_250226. e.The cumulative share of pre- and postsynaptic links in identified cells versus orphan fragments (not part of an identified cell). Plot is by fragment size as inferred by number of links on fragment (version 626). f. Scatter plots show the correlation between matched pairs of connected cell types in the BANC versus FAFB and MANC (and the most complete extant connectomes). Each point is a cell-type-to-cell-type normalized connection (synaptic connections from source-to-target / total number of postsynaptic links on the target cell type). FAFB-BANC: 34174 matched cell type connections, MANC-BANC: 29350 matched cell type connections. g.Confusion matrix of neurotransmitter prediction evaluated at the level of whole neurons on the held-out test set. Whole neuron prediction is based on the summed classification probabilities across all presynaptic links, selecting the most confident class. The ground-truth included 20572 neurons (from 2900 cell types, see Methods), of which 16448 were used for training and 4124 for testing. h.Users can browse BANC data via Codex (codex.flywire.ai/banc), and they can download data for programmatic analysis (via Codex, CAVE, and Dataverse at https://doi.org/10.7910/DVN/8TFGGB). i. Color-depth MIPs (maximum intensity projection images where color encodes depth) in JRC2018U space for BANC dataset neurons (version 626) available from our Dataverse archive. These can be used to search for genetic driver lines enabling functional investigation into BANC neurons, for example using NeuronBridge. Examples are shown for a specific cell type (DNa02).

Extended Data Fig. 2:. Individual DNs and ANs often influence effectors in multiple body parts.

a. Fig. 2b shows that the adjusted influence is proportional to ‘layers’ of a published graph traversal model applied to the FAFB dataset. Here we show that the adjusted influence is also proportional to the output of a different published layering algorithm. As in Fig. 2b, we used olfactory seeds annotated in the FAFB dataset. b. Distribution of presynaptic links in the VNC versus the brain, for all DNs (1313 cells) and ANs (1841 cells) in the BANC dataset. c. Distribution of segregation index values for these same DNs and ANs. Segregation index is a measure of polarization which quantifies the entropy of pre- and postsynaptic connections between the axonal and dendritic compartments of a neuron. A segregation index closer to 1 indicates a more polarized neuron. d. Here we chose three DNs and one AN that have clear behavioral effects, and we examined their adjusted influence on effector cells in different body parts. Within each subplot, each point is an effector cell, with direct connections in red. The horizontal line marks a value of 17.18, which we take as a conservative cutoff for “high influence” (see note below). All four cells have some effector influence above this cutoff. For each cell, the above-cutoff effector influences are compatible with the cell’s function. e. After discarding connections below this cutoff, we counted the number of AN and DN cell types that influence effectors in single body parts (top) or multiple body parts (bottom). The bottom plot shows only the most common 20 combinations of body parts. f. The number of AN and DN cell types that combine different numbers of body parts. Gross CNS division for combined effectors shown in color (‘both’ can appear when only one body part is targeted, because neck motor neurons can exist in both the brain and VNC). g. Same as (f), but color indicates combinations across motor classes and visceral/circulatory classes. h. The effector cell map from Fig. 2i, color-coded by adjusted influence from example ANs and DNs. Bottom right, cells are color-coded by the side of the CNS on which their efferent axon exits. Note, we chose this adjusted influence cutoff because it is the “elbow” in the cumulative distribution of AN/DN-to-effector adjusted influences involving DNs and ANs with known behavioral functions; DNs and ANs used to identify this elbow were DNa02, DNa01, DNp01, DNp02, MDN (DNp50), DNp42, DNg97, DNg100, DNg12, DNg62, DNp07, DNp10, DNg14, DNa15, DNb01, DNp37, oviDNb, DNp20, DNp22, DNp25, DNp44, DNp27, AN17A026 and AN19A018.

Extended Data Fig. 3:. Influence streams to and from AN/DN clusters

a. Tanglegram showing the relationship between two methods of sorting AN/DN clusters (Fig. 3a). The left dendrogram sorts clusters based on the similarity of their adjusted influences from sensor cell subclasses. The right dendrogram sorts clusters based on the similarity of their adjusted influence to effector cell subclasses (right). Colors denote superclusters. b. Names of studied cell types in the field, and their positions in our UMAP space, built by AN/DN direct connectivity to other neurons of the CNS. c. Our AN/DN map from Fig. 3a, with functions assigned by Braun et al. (2024). This earlier work only used direct FAFB DN-DN connectivity, and as a result, functional information was more limited than it is now. d. Adjusted influence from sensory neuron subclasses onto AN/DN neuron clusters. e. Adjusted influence from AN/DN clusters onto effector cell subclasses. f. Similarity of adjusted influence between specific sensory neurons and superclusters. Superclusters are rows; sensory neurons are columns.

Extended Data Fig. 4:. AN/DN morphologies by supercluster

a. Each subpanel shows all right-side neurons from one AN/DN supercluster in the UMAP embedding. Neuroglancer links for flight power, flight steering 1, flight steering 2, head and eye orienting, landing, threat response, proprioceptive, tactile, mating and reproduction, feeding, visceral control, probing, grooming, walking steering and walking. b. Distribution laterality index values, for each AN/DN supercluster. Each synaptic connection is signed by the anatomical side of BANC in which it is found (−1 for left, +1 for right). Laterality index is: 1 - abs(mean of the postsynaptic score - mean of the presynaptic score). Each distribution is scaled so that the area under the curve is 1.

Extended Data Fig. 5:. CNS networks’ cluster influence from sensors and to effectors

a. UMAP embedding of BANC neurons, where each point is a neuron. This analysis uses all BANC neurons that meet four criteria: they are marked as proofread, they are intrinsic neurons of the CNS (not afferents or efferents), they have >100 incoming and outgoing connections, and no part of the cell is in the optic lobe (as the optic lobes are still undergoing proofreading). In total, 29519 neurons were used for this analysis, corresponding to 88% of cell-typed central brain and/or VNC intrinsic neurons. b. Proportion of each CNS network belonging to select super classes / cell classes. c. Mean adjusted influence of AN/DN superclusters onto input neurons of the mushroom body and central complex. d. Mean adjusted influence of mushroom body output neurons and central complex output neurons onto AN/DN superclusters. e. Mean adjusted influence of sensors onto CNS networks. Visual projection neuron cell types are included, although they are not peripheral sensory neurons. f. Mean adjusted influence of CNS networks onto effector cell subclasses. g. Mean adjusted influence of each CNS network into other CNS networks.

Figure 1:. An open-source brain-and-nerve-cord connectome.

a. (left) X-ray micro-computed tomography (microCT) projection of the BANC sample following dissection, staining, and embedding for EM. (right) Surface mesh rendering of the CNS EM dataset with regions colored. A: anterior, P: posterior, D: dorsal, V: ventral, L: left, R: right. b. (top left) Aligned EM micrographs through a cross-section of the neck connective (y=92500) (magenta box in (a)). D: dorsal, R: right. (yellow box) Zoom-in of the EM data. (columns to right) Example EM image data from the BANC dataset. Neurons were automatically segmented using convolutional neural networks (CNNs)^,, with each segmented cell shaded with a different color. Mitochondria (x: 137533, y: 35220, z: 2493) and nuclei (x: 192977, y: 51679, z: 2493) (both overlaid with different colors) were segmented. Postsynaptic locations (shaded with different colors, example: x: 140988, y :36705, z: 2498) were automatically predicted and presynaptic locations (end of yellow lines) were automatically assigned using CNNs. (bottom left). The predicted neurotransmitter for the selected synapse (center of the green box) is acetylcholine. c. (top) Fraction of proofread neurons in gross divisions of the CNS. Neurons are labeled as proofread when their primary neurites or ‘backbones’ have been reviewed. (middle) Fraction of proofread neurons in the BANC matched with neurons in other connectomes, by gross divisions of the CNS. Morphological cell type level matches were confirmed by experts (teal), or matched to a likely class based on high NBLAST scores. (bottom) Fraction of true and false positive synapse predictions in different divisions of the CNS. Full CNS inventory inferred from summing counts from FAFB and MANC, and subtracting photoreceptors not captured by BANC (11468). d. Neurons were matched to metadata from previous projects by transforming their morphologies from other connectomes^,,,,, into BANC space. We used NBLAST to identify potential morphological matches. An example with DNa02 is shown, illustrating the process. Neuroglancer link for morphology, Codex link for metadata/connectivity. e. Hierarchy of cell annotations, based on previous work^,, but adopting clearer terms. Exemplified for LB1a (Neuroglancer link, Codex search) and DVm1a-c (Neuroglancer link, Codex search). See (Supplementary Data 1). f. The proportion of proofread neurons (of 114518) in the BANC by metadata label. Fast-acting neurotransmitter identities are assigned by our native BANC neurotransmitter predictions, based on. The ‘peptide’ class was added in cases where evidence from the literature supports neuropeptide expression, but our prediction is for a monoamine. In these cases we suspect the predictions are more likely to be incorrect. It is not meant to represent the number of peptidergic neurons, which would be far larger..

Figure 2:. Linking sensors and effectors through local and long-range circuits.

a. The influence of source cells on target cells is estimated via linear dynamical modeling. b. Adjusted influence (see Methods) is proportional to the number of network ‘layers’ in a graph traversal model. Direct and indirect connections are shown in red and gray, respectively. Here the source neurons are olfactory receptor neurons in the FAFB dataset, following previous work, and adjusted influence is averaged over the number of neurons in the source and target groups. Regression line in black (R²=0.94, n = 94278). c. Distribution of adjusted influence scores between all ANs (1841) and DNs (1313) and all other neurons (155936) in the dataset. Direct and indirect connections are shown in separate histograms, with the peak of each histogram normalized to its own maximum. d. Schematic of body parts associated with annotated effector cells in the BANC. Not all neurohemal organs shown. Neuroglancer link, explore on Codex here. e. Mean adjusted influence of sensory cells (columns) on effector cells. Sensory and effector cells are pooled by body part. Each row is minmax normalized to the same range (0–1). This plot summarizes data from 14410 sensory cells and 1026 effector cells. We omitted 3188 putative sensory cells whose corresponding organs could not be identified. f. Schematic: an example local loop (top) that is also linked to specific sensors via long-range connections (bottom). g. Scatterplot showing the mean adjusted influence on each effector cell from DNs versus ANs. Black, unity line. Insets: a DN soma is located in the head, whereas an AN soma is located in the body. h. An example AN and DN with strong adjusted influence on effector cells in multiple body parts. Neuroglancer link, Codex network. i. UMAP embedding of effector cells, based on the cosine similarity between the adjusted influences these cells receive from individual ANs and DNs. The major cell types in each effector cell group are listed (MNs, 833 motor neurons; ENs, 193 endocrine neurons some of which are putative). Neuroglancer link, Codex search. See (Supplementary Data 5).

Figure 3:. Clustering ANs and DNs into behavior-centric modules.

a. UMAP embedding of all ANs and DNs based on cosine similarity between their direct connectivity vectors (connections to any other proofread neuron in BANC). Neuroglancer link to ANs here and DNs here. b. Previously characterized ANs and DNs highlighted in this map (Supplementary Data 7). c. In each copy of this same map, each point is an AN or DN, color-coded by the adjusted influence that cell receives from example sensory neurons (top) or color-coded by the adjusted influence that cell sends to example effector cells (bottom). Based on these adjusted influence scores, we lumped the 27 clusters into 15 superclusters. d. Mean adjusted influence onto each AN/DN supercluster from select groups of sensory neurons. Superclusters are rows; sensory neurons are columns. A subset of visual project neurons were used to determine processed visual streams from the optic lobes^,,–, see methods. e. Mean adjusted influence from each supercluster onto select groups of effectors. Superclusters are rows; effectors are columns. f. The same map, here colored by supercluster membership. Neuroglancer link. See (Supplementary Data 4). g. Example circuit involving visceral control ANs and DNs. Neuroglancer link, Codex network. h. Example circuit involving the flight power supercluster and visceral control supercluster. Neuroglancer link, Codex network. i. Example circuit for coordinated visceral sensing and reproductive control. ANXXX986 is female-specific^,. Neuroglancer link. j. Example circuit involving a DN in the tactile supercluster. Neuroglancer link, Codex network. k. Example circuit illustrating proprioceptive input to visual neurons. Neuroglancer link, Codex network.

Figure 4:. Specializations and coordination within a functional supercluster.

a. Enlarged view of the head-and-eye orienting supercluster, taken from the UMAP embedding of all DNs and ANs (Fig. 3d). Top: cells are color-coded by their incoming adjusted influence from two different sensory sources. Same as (a), but now cells are color-coded by their outgoing adjusted influence onto three different effector cell groups. Neuroglancer link, Codex search. b. Mean adjusted influence from sensor sources, for all cell types in the head-and-eye orienting supercluster. c. Mean adjusted influence onto effector cells, for these same ANs and DNs. d. An example circuit with five cell types in the head and eye orienting supercluster. Thick arrows indicate connections with >100 synapses; intermediate arrows indicate connections with 20–100 synapses; thin arrows indicate connections with 5–20 synapses. This example was chosen to illustrate the concept of diverse but overlapping patterns of connectivity within a supercluster, as well as hierarchical interactions between cells in the same supercluster. Neuroglancer link, Codex network. e. An example circuit with two cell types in the landing supercluster (DNp10, AN06B002). This example was chosen to illustrate the concept that ANs and DNs in the same supercluster can be organized into loops. Neuroglancer link, Codex network.

Figure 5:. Interactions between behavior-centric modules

a. Mean adjusted influence of each AN/DN supercluster on every other supercluster. Values are normalized by the number of cells in each supercluster. b. Summary of the strongest adjusted influences between superclusters. c. A circuit illustrating an example of cross-cluster interactions between DNs and ANs. This circuit links cells in the proprioceptive, threat-response, and walking superclusters. Neuroglancer link, Codex network. d. Schematic example of subsumption architecture. This example has two local loops (behavior 1 and behavior 2), corresponding e.g. the control of individual legs. Behavior 3 is positioned to take control of both local loops (subsumption), contingent on some input from both sensors. Behavior 4 is positioned to subsume all other behaviors, based on some other input from both sensors.

Figure 6:. Linking CNS networks with superclusters of ANs and DNs.

a. CNS networks, obtained via spectral clustering of 51,502 backbone proofread neurons in the BANC dataset (excluding peripheral neurons and optic lobe neurons but including visual projection neurons and visual centrifugal neurons). Each panel includes a 2D kernel density estimation, a bar plot indicating the network composition, and cell count. Two pairs of networks are mirror images of each other (olfaction right/left and visual right/left), while all other networks are bilaterally symmetric, indicating high bilateral integration in those networks. Anatomical density images are normalized separately for the brain and VNC, based on a random sample of 100k synapses from each CNS network, the hotter the color the denser the synapses. b. Mean adjusted influence of each CNS network on each effector cell group (Fig. 2i). c. Strongest links between CNS networks. The size of each arrow represents the number of postsynaptic cells in that link. One weaker link is shown (155 cells), because this is the strongest output link of the central complex. d. Link strength between CNS networks, measured as the number of postsynaptic cells in that link. The color scale is capped at 2000 cells. e. Out-of-network connections, measured as the proportion of partners each cell has in another CNS network. DNs and ANs have an unusually high proportion of out-of-network connections. The area under each curve is normalized to 1. All three distributions are significantly different from each other (DN vs. other p = 1.92×10⁻⁹⁷, AN vs. other p = 6.03×10⁻⁵, AN vs. DN p = 6.74×10⁻⁴³; 2-sample Kolmogorov-Smirnov tests). f. Number of ANs and DNs in each CNS network. ANs and DNs are grouped by supercluster (Fig. 3f). g. Example circuit connecting mushroom body neurons (purple) to ANs and DNs. Neuroglancer link, Codex network. h. Example circuit connecting central complex neurons (blue) to ANs and DNs. Neuroglancer links here and here. Codex network.