Whole-organism cellular gene-expression atlas reveals conserved cell types in the ventral nerve cord of Platynereis dumerilii

Abstract

The comparative study of cell types is a powerful approach toward deciphering animal evolution. To avoid selection biases, however, comparisons ideally involve all cell types present in a multicellular organism. Here, we use image registration and a newly developed “Profiling by Signal Probability Mapping” algorithm to generate a cellular resolution 3D expression atlas for an entire animal. We investigate three-segmented young worms of the marine annelid Platynereis dumerilii, with a rich diversity of differentiated cells present in relatively low number. Starting from whole-mount expression images for close to 100 neural specification and differentiation genes, our atlas identifies and molecularly characterizes 605 bilateral pairs of neurons at specific locations in the ventral nerve cord. Among these pairs, we identify sets of neurons expressing similar combinations of transcription factors, located at spatially coherent anterior-posterior, dorsal-ventral, and medial-lateral coordinates that we interpret as cell types. Comparison with motor and interneuron types in the vertebrate neural tube indicates conserved combinations, for example, of cell types cospecified by Gata1/2/3 and Tal transcription factors. These include V2b interneurons and the central spinal fluid-contacting Kolmer-Agduhr cells in the vertebrates, and several neuron types in the intermediate ventral ganglionic mass in the annelid. We propose that Kolmer-Agduhr cell-like mechanosensory neurons formed part of the mucociliary sole in protostome-deuterostome ancestors and diversified independently into several neuron types in annelid and vertebrate descendants.

In recent years, the molecular comparison of cell types has grown into an exciting new discipline. One major aim is to understand the diversification of cell types in animal evolution and to reconstruct their ancient repertoire in remote ancestors (1, 2). Cell types represent evolutionary units, defined as “sets of cells that change in evolution together, partially independent of other cells, and evolutionarily more closely related to each other than to other cells” (2). One important focus of interest is the evolution of neurons and nervous systems, with several studies comparing sensory neurons (3–6), interneurons (7), or motor neurons (8, 9). The starting point for these comparisons has been the observation that molecular regions in the developing nervous system are conserved from medial to lateral, between vertebrate and fly (10), and vertebrate and annelids (9) (Fig. 1). If these mediolateral regions already existed in urbilaterian ancestors, it is likely that they gave rise to region-specific neural cell types in these ancestors; however, given that conserved developmental regions can result in entirely different morphologies (11, 12), it is clear that any further homology statement requires an exhaustive analysis of the cell types that differentiate from the compared regions.

Fig. 1.

Comparison of mediolateral patterning, morphogenesis, and molecular neuroanatomy in vertebrate and annelid. Schemes illustrate similarities and differences between vertebrate and annelid developing trunk nervous systems. Note that, because of different infolding strategies, the basal neuropil is outside of the tube in vertebrates, but inside the cord in annelids. Hatched gray, proliferative layer; Solid gray, neuropil. (A) Generalized development of the vertebrate neural tube. hb9 expression after Dichmann and Harland (46) and phox2 expression after Talikka et al. (47). White arrow in the Middle panel indicates invagination of the neural tube centered on the neural midline. Molecular anatomy of neural tube simplified after refs. 14 and 15. (B) Generalized development of the annelid nerve cord. hb9 and phox2 expression after Denes et al. (9) and this study. White arrows in the Middle panel indicate the bilateral, lateral infolding of the DGM. Molecular neuroanatomy according to this study. DGM, dorsal ganglionic mass; dI, dorsal interneurons; KA, Kolmer-Agduhr; MN, motorneurons; sMN, somatic motorneurons; V, ventral interneurons; VGM, ventral ganglionic mass.

In vertebrates, the alar plate of the neural tube contains several sets of dorsal interneurons, and the basal plate ventral interneurons and motor neurons that generate the locomotor patterns controlling movement (Fig. 1A) (13). The distinct identities of these neuron types are established by combinations of transcription factors (14, 15), which form so-called “core regulatory complexes” (CoRCs) that activate the neural differentiation genes specific for a given cell type (2, 16). For example, a complex of the LIM homeodomain proteins Lhx3 and Islet1, the EGH homeodomain factor Hb9, and the cofactor NLI specifies somatic motor neurons (17); another complex of the Gata2 factor, together with the Tal1/Scl bHLH factor, the LIM-only protein Lmo4, and NLI specifies the inhibitory descending V2b interneurons (18) (Fig. 1A); and V2a interneurons are controlled by a CoRC comprising Lhx3 and NLI. Any large-scale comparison of cell types across phyla should thus focus on the neuron-type–specific, combinatorial expression of transcription factors. Indeed, combinations of Lhx3, Islet, and Hb9 have been found to specify somatic motor neurons also in fly and nematode (8). On these grounds, it was proposed that somatic motor neurons specified by Hb9, Islet, and Lhx3 existed in bilaterian ancestors (8, 9).

However, it was also noted that in all groups, additional motor neurons exist, specified by different, partially overlapping combinations of transcription factors (8, 14), blurring the picture of cell-type interrelationships between phyla. Obviously, clade-specific cell-type diversification events have taken place in the divergent bilaterian lineages, and cell types may have been lost in specific lineages (e.g., the “chordate bottleneck” hypothesis; see ref. 8). It follows that, to infer cell-type interrelationships between remote animal groups, it will be misleading to restrict comparisons to few candidate neurons: cell types that are even more closely related may exist elsewhere in the body, which—once included—might produce a different picture of cell-type interrelationships within or across species. Instead, cell-type comparisons should provide molecular information for as many cell types as possible in as many species as possible.

Here, we follow a systemic approach to molecularly characterize and compare all differentiated cell types in an entire animal. Our focus is on the marine annelid Platynereis dumerilii, which, in a number of recent studies (19–21) has proven to retain ancestral neural characters not found in fly or nematode, and is thus especially suited for long-range evolutionary comparisons. We find that, in Platynereis, the ventral nerve cord (VNC) is almost fully differentiated after 6 days post fertilization (dpf) in three-segmented young worms, which are composed of around 12,000 cells in total, making this stage an ideal system for in toto analysis (22). We present a technique, Profiling by Signal Probability mapping (ProSPr), which allows us to generate an atlas for around 100 genes and map gene expression onto a close-to-complete cellular model of the whole animal. We then focus on the VNC neurons to determine distinct molecular identities for 605 bilateral pairs of cells. Similarity-based clustering reveals neuron types expressing specific combinations of transcription factors. Besides several types of segmentally iterated, putative motor neuron types defined by coexpression of nk6, hb9, and islet, we also identify spatially coherent clusters of neurons that might be evolutionary related to different classes of vertebrate interneurons (Fig. 1). We propose that one of these groups, coexpressing gata1/2/3 and tal transcription factors in a highly specific manner, may relate to the similarly defined sensory Kolmer-Agduhr neurons (23) and V2b ventral interneurons (14) in the vertebrates.

Results

In Toto Molecular Profiling of Platynereis Nectochaete Larvae.

To select the most suitable life-cycle stage with a close-to-complete complement of cell types on the one hand, but as few cells as possible on the other, we determined the earliest stage when most neurons are differentiated in Platynereis. We treated developing larvae with 24-h pulses of EdU (that labels mitotic cells) from 2 to 6 dpf and found a significant decrease of cellular proliferation in the developing VNC over time, with just a few cells dividing during 5–6 dpf (Fig. S1A). In line with this finding, the vast majority of VNC cells appeared to express the neuronal markers elaV and synaptotagmin (Fig. S1B) at these later stages, indicative of neuronal differentiation. Interestingly, this stage undergoes pelago-benthic transition (24), when animals start feeding and crawling on the substrate, performing periodic alternations of movements of their lateral body appendages, called parapodia. For this reason, the Platynereis VNC at 6 dpf represents a good system to study the molecular profile of neuronal cell types in a prototype protostome central nervous system. It should be stressed, however, that the 6-dpf VNC likely also contains quiescent progenitors and glial cells interspersed between neurons; in the absence of reliable markers, however, we are so far unable to identify these.

Fig. S1.

The VNC of Platynereis is mainly composed of postmitotic neurons at 6 dpf. (A and B) Panels show ventral and side projections of Platynereis VNC (dashed red box in B) at 6 dpf. (A) Cell age as determined by sequential 24-h pulses of EdU, from 2 to 6 dpf (red, EdU; blue, DAPI). (B) VNC location and RNA localization for postmitotic pan-neuronal markers (elav and synaptotagmin), and for glutamate decarboxylase (gad), used to measure PrImR and ProSPr resolution. The asterisk labels the cell used for the analysis of resolution (red, reflection microscopy; blue, DAPI). (C) Assessment of image registration of 144 hpf (equal to 6 dpf) Platynereis larvae for different imaging isotropic resolutions. gad RNA localization was used to identify a single cell in the second ganglia of the VNC (asterisk in B) in 20 animals. For every imaging resolution (x axis), the centroid of the gad cells was calculated. The dashed red horizontal line is the average diameter of a cell (calculated as the median of all of the cells segmented at 334 nm and assuming cells as spheres). The voxels per cell (vpc) are indicated on top, which show, for every imaging resolution, how many voxels (3D pixels) fit inside an average cell. (C′) The individual points represent all of the distances, in microns, between the centroids of the cells. Box plots and violin plots represent the distribution of these distances. (C′′) Distance between the gad cells centroids, and the center of mass of their spatial distribution, calculated as the median position of the 20 cells. (D) Variation of the position of the gad cells in the three dimensions and 3D representation of the location of the gad cells after the registration. They are represented as spheres with average cell diameter. In red, the centroid of the distribution. The standard size of Platynereis dumerilii larvae at 6 dpf is 280 μm in the anterio-posterior axis.

In a first attempt to produce an expression atlas at 6 dpf, we used Profiling by Image Registration (PrImR), a technique successfully used to generate cellular resolution atlases at earlier developmental stages (25, 26). This technique overlays and aligns different expression patterns onto a common, highly stereotypic spatial framework, such as the axonal scaffold, or DAPI-stained nuclei. However, we found that, at this more variable later stage with complex body features, the variability in position of a single, uniquely identifiable cell is larger than the average cell radius, in conflict with cellular resolution (Fig. S1 B–D).

To gain cellular resolution, we reasoned that sampling of an increased number of individuals should provide enough spatial information to produce a 3D map of probabilities for the location of the expression signal that could subsequently be used to calculate its average spatial location (Fig. 2A). We refer to such representation as a signal probability map (SPM). To test whether an SPM would recover the positional information of a given cell, we compared the region of highest expression probability with the arithmetic mean position (centroid) of this cell in different individuals. For this, we manually segmented a uniquely identifiable cell in several registered individuals, determined the centroid for all cells, and calculated the center of mass of the population (Fig. S1). We then compared the center of mass to the peak probability region within the SPM and determined that the distance between them remained below one average cell radius when using more than 12 samples for generating the SPM (Fig. 2B), and that increasing the imaging resolution over 660 nm per pixel did not result in a significantly better performance (Fig. 2C).

Fig. 2.

SPMs can be used to achieve single-cell resolution after image registration at 6 dpf. (A) Scheme representing the signal reconstruction process, generating SPMs. Individual animals are registered to the reference, the signal is binarized, and overlapped to generate a probability map of the spatial distribution, which is later thresholded. (B) Effect of threshold percentage and sample size on signal reconstruction, scored as the distance of the resulting 3D object center of mass to the centroid of the cell distribution (calculated by manual segmentation), and the volume of the 3D object. From the initial dataset of 20 manually segmented cells (see text and Fig. S1), random picks were made for every imaging resolution. Each point represents the distribution of centroid-center distance and volume (mean and error bars) of 1,000 random iterations. Horizontal dotted line represents the average cell radius and vertical dotted line represents the average cell volume (calculated for the population of manually segmented cells). Results for the 550 nm per pixel imaging resolution dataset in C. (C) Effect of the imaging resolution and the number of samples on the performance of the signal reconstruction process, measured as in B. (D) Plot showing, for 99 different SPMs, the decrease rate of the signal volume in relation to the threshold, and its low correlation with the sample size and the coverage of the signal, making the use of a general thresholding method based on these parameters not adequate. (E) Quantification of SPMs based on the agreement between the samples for different molecular markers.

Taking these values into account, we generated a high-resolution reference (Fig. S2 and Methods) using DAPI scans of 153 individuals, which we used to register all individual larvae to the same 3D coordinate system. We then calculated SPMs for whole-body expression patterns of up to 100 different genes assessed by whole-mount in situ hybridization. As our primary interest was in neural cell types, this gene set comprises 59 transcription factors covering most homeodomain families, bHLH factors, Sox, Fox, and T-box families, 39 neural differentiation genes, as well as 3 proliferation markers. To infer an average expression for a given gene from the SPM, we set out to determine the specific threshold up to which the experimentally determined expression probability in a given SPM would be interpreted as real expression. We found that a general thresholding method was not adequate because of gene- and probe-dependent technical and biological variation (Fig. 2 D and E). Instead, we chose an individual threshold for each gene, so that the volume of the resulting average expression approximated the expected expression volume (that is the mean of expression volume in the registered individuals). For increased accuracy, we determined this threshold independently for each spatially separable volume of the overall expression pattern (Fig. 3). We refer to the resulting objects as minimal expression domains (MEDs). Next, we automated the preregistration (Fig. S3A), registration, and postregistration (Fig. S3B) steps, to generate, from raw microscopy files, MEDs that can be easily transformed into full gene-expression maps. We thus generated standardized whole-body gene-expression maps for all genes at 6 dpf (Fig. S4). We refer to our automated pipeline as ProSPr (Fig. S3C).

Fig. S2.

Reference for Platynereis 6-dpf larvae. (A–D) Coronal sections from ventral to dorsal (anterior up). (A) At the level of the VNC neuropil. (B) Dorsal to the VNC, at the level of the beginning of the stomodeum. (C) At the midpoint of the stomodeum, gut, and brain neuropil. (D) At the level of the posterior part of the brain neuropil and end of the gut. (E–H) Transverse sections from anterior to posterior (ventral down). (E) At the level of the brain neuropil. (F) At the level of the stomodeum. (G) At the level of the second VNC ganglion. (H) At the level of the pygidium. (I–K) Three-dimensional views of the reference from the front, left side, and back, respectively. The standard size of Platynereis dumerilii larvae at 6 dpf is 280 μm in the anterio-posterior axis.

Fig. 3.

Dissection of gene expression maps into MEDs. Schematic representation of the principle to threshold the SPMs (color-code indicates signal probability). Three-dimensional objects are automatically isolated and thresholded to approximate the amount of signal measured in the individual animals in that 3D space. As a consequence of the thresholding, new isolated 3D volumes can appear. This process is repeated for each isolated object until no further subdivision of the objects can be achieved, giving rise to MEDs.

Fig. S3.

Automation of ProSPr method. (A) Illustration of the steps and outputs of the image analysis pipeline before registration. (Upper, Left) Bright-field panel of maximum projections of every individual in the microscopy file, indicating the ID for each individual (cropped). (Lower) Explanation of the transformations and analysis of the channels. The animal autofluorescence (enhanced for an easier appreciation) can be subtracted from the signal channel. (Upper, Right) Panel of maximum projections of the oriented enhanced signal for every individual with their respective IDs (cropped). (B) Postregistration process to binarize in situ signal. (Left) Illustration of the process of masking the samples. Registered individuals are normalized and averaged [PrImR method (26)]. This average is then thresholded (automatically or manually) to create a mask. The mask is applied to the individual samples, and for each sample, the signal is quantified inside and outside the mask for different threshold values. Images show snapshots of 3D representations for the marker tbh. (Right) Quantification of the signals and difference between them. For each threshold, the signal inside and outside the mask is measured, and the difference between these values is computed. The algorithm identifies the first maxima of the difference function to assign a threshold value to the sample. (C) Flowchart of the method ProSPr. In blue and red, the most important steps and results, respectively, of the process. The standard size of Platynereis dumerilii larvae at 6 dpf is 280 μm in the anterio-posterior axis.

Fig. S4.

Ventral and lateral projections of the expression of markers included in the atlas. (A) Transcription factors directing the mayor grouping of cells in the t-SNE (see Fig. S7). (B) Transcription factors expressed in the VNC. (C) Other transcription factors. (D) Differentiation markers. (E) Proliferation signal (EdU pulses). (F) Table showing mayor animal regions where the expression of different genes has been detected. The standard size of Platynereis dumerilii larvae at 6 dpf is 280 μm in the anterio-posterior axis.

Generating Cellular Models from the Expression Atlas.

We next developed a method to reconstruct virtual cells from the gene-expression atlas (Fig. 4A). We reasoned that, given the cellular resolution that our gene-expression atlas provides, it should represent cells as units of homogenous gene expression. To reduce the complexity of the dataset of more than 15 million voxels, we grouped these into virtual cubes of 3³ voxels that we refer to as supervoxels (SV), with each SV containing the average gene-expression information of the covered volume (Fig. 4C). Given an average cell diameter of 6 µm and the imaging resolution (isotropic 0.55 µm per pixel), each cell should thus be represented by 10–40 SV. To reduce the complexity of the dataset further, we subdivided the animal into distinct body parts enriched for neural tissue, hand-segmented from the reference. These included the VNC, brain, lateral ectoderm, cryptic segment (27), and the back end, the pygidium (Fig. 4B).

Fig. 4.

Reconstruction of a cellular model from the 6 dpf Platynereis atlas. (A) Workflow of the recursive partition of hierarchical clustering to group SV into cells. (B) Manual segmentation of the reference to profile different tissues. (C) Illustration of the cellular grid to build SV (maximum projection of NBT/BCIP reflection signal in a single neuron). (D) Removal of SV with low correlation with their neighbors. (E) t-SNE of the 4,315 cells recovered by the algorithm in the segmented tissues, using all genes in the atlas. CS, cryptic segment; latEct, lateral ectoderm; Pyg, pygidium; VNC, ventral nerve cord.

With the aim of generating a cellular model, we developed a strategy for the assembly of SV into groups representing virtual cells. First, we removed SV that did not show a high expression correlation with their neighbors, because these most likely represented cellular boundaries or noise (Fig. 4D). Second, we recursively used unsupervised hierarchical clustering to find groups of SV with similar expression profiles. Because cells belonging to the same cell type can be spatially separate (e.g., on the left and right body sides, or in consecutive segments), this similarity search was based on expression correlation only and did not take into account spatial location. As third step, however, we then assessed the resulting clusters of SV for spatial coherence. Clusters with largely scattered (spatially incoherent SV distributions) were eliminated, as they likely represented clustering artifacts (Fig. 4A). Clusters of SV concentrated around one or few locations were maintained, as they likely represented real cells. Note that in most cases such spatially coherent clusters represented bilateral pairs of cells. This recursive partition and clustering algorithm was run separately for each body part and recovered a total number of 4,315 spatially coherent, minimal units of homogenous gene expression. These virtual units, mostly representing bilateral cell pairs, covered more than 80% of the total processed volume. Given that cell boundary volumes had been discarded, we are confident that our cellular atlas represents a molecular description of almost all cells of the Platynereis nervous system at 6 dpf.

Approximating Cell Types: Spatially Coherent Clusters of Cells with Similar Regulatory Signature.

We next investigated how the 4,315 virtual cells would relate to each other. For a first coarse grouping, based on expression similarity, we used t-distributed stochastic neighbor embedding (t-SNE), which transfers high-dimensional data into a space of low dimensions (28). This revealed groups of VNC and head cells as coherent entities in the 6-dpf body (Fig. 4E). In subsequent analyses, we focused on the VNC cells and among these we selected, by visual inspection, a high-confidence set of cells based on their overall morphology, expression correlation, bilateral symmetry, and size. Fig. S5 shows these 605 reconstructed, bilateral pairs of cells in the Platynereis VNC. With the aim of identifying cell types within this dataset, we used t-SNE to group cells based on their combination of transcription factors. We focused our analysis on those cell groups with most expression information (groups I to XIV in Fig. 5A, composed of cells expressing four or more transcription factors) and discarded those for which little expression information was available. Each of these gene-rich groups homogeneously expressed a small set of transcription factors—such as nk6, hb9, pax6, and islet for group I (Fig. 5B)—many of which are known to play conserved roles in regional patterning (Fig. S6).

Fig. S5.

Ventral and lateral 3D representations of the 605 VNC cells reconstructed from the atlas, and included in the analysis.

Fig. 5.

Cell-type analysis on the 6 dpf Platynereis ventral nerve cord. (A) t-SNE of the VNC cells using transcription factor expression, showing cells color-coded according to how many transcription factors they show expression for. Groups of cells with high expression information have been manually selected for further analysis. (B) Hierarchical clustering and heatmap showing the genes expressed in cells in group I. Cell types have been manually defined based on the clustering and are highlighted in different colors. (C) Spatial location (frontal and lateral views of the VNC) of the cell types in group I. Transcription factors and neurotransmitters (italics) expressed in these cells are indicated in the figure. In white, genes common to the entire group; in color, those specific for the different cell types. (D) Frontal views of the VNC showing the location of the different cell types identified from the groups in A. Genes expressed in these cell types are indicated as in D.

Fig. S6.

Transcription factor expression mapped on the t-SNE graph. (A) Transcription factors directing the mayor grouping of cells in the t-SNE. (B) Other transcription factors expressed in the ventral nerve cord.

To further relate the molecular signatures of the cells composing each group, we performed unsupervised hierarchical clustering with all markers included in the atlas. This process yielded clusters of virtual cells that expressed identical combinations of transcription factors and largely similar sets of neural effector genes (see, for example, type 1 to type 5 in Fig. 5B, and the sets of cells of similar color in Fig. 5 C and D and Fig. S7). Given the importance of the transcription factor regulatory signature in defining cell-type identity, we propose these core clusters may approximate cell types of the VNC (Discussion). We found all of these putative cell types to be spatially coherent, with specific ventral-dorsal, medial-lateral, and anterior-posterior coordinates (Fig. 5 C and D). In line with the segmented nature of the annelid nerve cord, we noted that many of them spread out over several segments in serially repeated arrangement (e.g., Fig. 5C). Interestingly, the first segment showed many unique types not iterated in the second and third segment (e.g., in groups III, IV, V).

Fig. S7.

Depiction of cell types within the VNC groups. Hierarchical clustering and heatmaps showing the genes expressed in cells in groups II to XIV (see Fig. 5). Cell types have been manually defined based on the clustering and are highlighted in different colors. Spatial locations (frontal and lateral views of the VNC) of the cell types within each group are shown.

Molecular Anatomy of the Ventral Nerve Cord.

We noted that the two axes of the t-SNE reflect the major anatomical axes of the VNC (ventral-dorsal and medial-lateral), illustrating its molecular anatomy. To explore this further, we subdivided the t-SNE plot into an arbitrary number of seven superclusters (Fig. 6A) using k-means classification (seven was found to produce consistent results). The resulting superclusters indeed demarcated medial, intermediate, and lateral parts in the nerve cord, as well as prominent dorsal “extensions” of the nerve cord (Fig. 6B). We refer to these as the medial, intermediate, and lateral ventral ganglionic mass (VGMm, VGMi, VGMl) and dorsal ganglionic mass (DGM), respectively (Fig. 6C). In molecular terms, the VGMm-i boundary coincides with the medial expression boundary of pax6, and the VGMi-l boundary with the lateral extent of hb9 and nk6 expression. The DGM prominently and specifically expresses the phox2 gene (Fig. 6D).

Fig. 6.

Medio-lateral arrangement of VNC superclusters. (A) k-means clustering (seven clusters) on the VNC cells t-SNE. (B) Spatial location of the k-means superclusters found in A (ventral and lateral views of the VNC). (C, Upper) Illustration of the molecular structure of Platynereis VNC at 6 dpf mapped on the entire body plan (ventral and lateral views). (Lower, Left) Transversal section (see Top) of the reference at the level of the second VNC ganglia, indicating the ventral structures. (Lower, Right) Depiction of the major genetically defined territories in the VNC. (D) Cell expression of transcription factors Pax6, Nk6, Hb9, and Phox2 overlaid on the t-SNE graph. (E) Cell expression of transcription factors Gata1/2/3, Tal, Dbx1, Brn3, and Isl overlaid on the t-SNE graph. (F) Spatial location of four major groups of cell types compared with those in vertebrates in this study. Frontal views of the VNC showing the cells that coexpress the transcription factors indicated at the bottom.

The partition of the 6-dpf VNC ganglia into phox2⁺ DGM and pax6⁺, nk6⁺, and hb9⁺ VGM relates back to mediolateral regions of the 48 hours post fertilization (hpf) ectoderm (Fig. S8A). At 48 hpf, phox2 is expressed more laterally than hb9 in the ventral neuroectoderm. During subsequent morphogenesis, these phox2⁺ lateral cells are folded inwards on each side of the neuroectoderm (Fig. S8 B and C). In consequence, the former lateral cells are now located inwards, in a more dorsal position, as part of the DGM. Fig. 1 explains and summarizes these morphogenetic movements, leading to the specific neuroarchitecture at 6 dpf (Fig. 6C).

Fig. S8.

Transition from medio-lateral to dorso-ventral arrangement in Platynereis VNC development. (A) Overlay of double whole-mount in situ hybridization showing a ventral view of the expression of hb9 and phox2 at 48 hpf. (B) Transversal section illustrating the expression of phox2 in the trunk neuroectoderm at 48 hpf. (C) Transversal section illustrating the lateral expression of phox2 in the dorsal ganglionic mass of the ventral nerve cord at 5 dpf. (Magnification: A, 250× relative to A4 printing; B and C, 370× relative to A4 printing.)

Analyzing the molecular anatomy of the VNC further, we found two transcription factors, tal and gata1/2/3, specifically expressed in the VGMi (Fig. 6E), and determined their coexpression (Fig. 6F). This led us to identify nine tal⁺, gata1/2/3⁺ neuron types with distinct combinations of transcription factors distributed over several t-SNE groups (III, IV, V, VI, VII, XII, and XIV). Two of these cell types express serotoninergic (coexpressing sert and trpH) and cholinergic markers, another three cholinergic only, and one GABA- and glutamatergic markers. These neuron types may relate to the gata2/3⁺, tal1⁺ V2b interneurons and the Kolmer-Agduhr cells of the ventral part of the spinal cord of the vertebrates (Fig. 1) (Discussion).

Discussion

Cellular-Resolution Expression Atlases for Complex Body Plans.

We have demonstrated that whole-body gene-expression atlases with cellular resolution can be achieved for an advanced developmental stage of an entire animal, the three-segmented young worm of the annelid P. dumerilii. At this stage, the nervous system and locomotor apparatus is largely differentiated and fully functional. So far, in other animal models similar cellular resolution expression atlases exist for selected tissues only (29, 30). Building on previous image registration approaches (25, 26), ProSPr is a highly automated and easy-to-customize technique, running in free license software (Scripts available in prospr.embl.de). In an unbiased manner, it allows subdividing the animal into genetically defined regions for which quantitative information can be extracted. Our technique is transferrable to other species that are transparent or amenable to whole-body imaging via tissue-clearing methods (31).

The high quality of the atlas is reflected by the successful reconstruction of 4,315 minimal spatial units: pairs (mostly) of virtual cells with expression information for up to 100 genes. This reconstruction is achieved by a recursive partition clustering algorithm that groups SV solely based on similar expression and then validates the resulting cells by spatial coherence. This way, a high percentage of SV corresponding to a large fraction of the total animal volume is successfully grouped as part of computer-defined cells. The bilateral symmetrical arrangement of the cells, reflecting the bilateral symmetry of the animal, makes it likely that such reconstructed cellular units correspond to real biological cells and thus validates our approach.

The Nature of Platynereis Ventral Nerve Cord Neural Cell Types.

Having detected thousands of bilateral pairs of virtual cells, we ask how these assemble into cell types. In this report, we have focused our analyses on a subset of VNC cells, representing an important part of the trunk central nervous system. To identify possible cell types in the VNC, we have used hierarchical clustering of cells grouped via t-SNE, which yielded sets of cells expressing similar combinations of transcription factors. According to a recent evolutionary definition of cell types (2), this is good indication that these cells might belong to the same type. We reasoned that cells within the same cell type might still vary in gene expression, because of distinct states of differentiation, or local somatic environment. (Note that variation because of technical or biological noise, such as bursts in transcription, is less likely because expression has been averaged over many individuals.) Using these criteria, we ended up defining 41 tentative cell types for those parts of the VNC with good expression coverage (Fig. 5 B–D and Fig. S7). Taking into account that some of these might represent quiescent progenitors or unidentified glial or epidermal cells, a careful estimate would extrapolate around 200 distinct types of neurons in the Platynereis nerve cord at this stage (relating the volume contained in the 41 cell types to the total volume of the VNC). However, given the restrictions of our dataset with a limited number of genes per cell assayed for only one stage, it is still difficult to draw a line between cells belonging to the same cell type and cells belonging to related, so-called sister-cell types (1).

We noted that a considerable number of our tentative cell types are unique to the first segment. This is in line with the observation that, while the second and third segments take part in crawling locomotion, the parapodia of the first segment do not play a major role—which would imply divergent sets of neuron types. Later, in the course of the second metamorphosis (24), segment one transforms into a head segment bearing long sensory cirri. Our results indicate that this specific fate is already foreshadowed in the rather distinct cell-type complement of segment one at 6 dpf.

Molecular Anatomy of the Annelid Ventral Nerve Cord.

The subdivision of the nervous system into genetically defined superclusters reveals its molecular neuroanatomy in an unbiased manner. Of 59 transcription factors, we identify only a handful that define coherent anatomical regions. Among them, pax6 plays the most prominent role, as it is absent from the VGMm, coexpressed with nk6 and hb9 in the VGMi, and coherently expressed in the VGMl. In addition, phox2 expression defines the DGM (Fig. 6D). Overall, this resembles the vertebrate situation, where Nk6 together with Hb9 define ventral interneuron populations in the basal plate, whereas Phox2 is found in several types of dorsal interneurons (Fig. 1) (15). Notably, however, vertebrates differ from the annelids in that Phox2 is also specifying visceral motor neurons more ventrally (15).

Our unbiased approach reveals another important feature of the annelid VNC. As apparent in Fig. 6 B and C, the intermediate and lateral parts of the ganglionic mass (VGMi and VGMl) take the rough shape of concentric rings that cross the midline along the intersegmental boundaries. This finding reflects the segmental nature of the VNC and its anatomical subdivision into spherical ganglia. It appears that the evolution of ganglia from a homogenous ventral neural mass involved some kind of “constriction” of the mediolateral architecture, moving lateral parts closer to the midline. This idea finds support for example in the medially constricting pax6 expression in Drosophila (32), which likewise possesses a ganglionic VNC.

Evolutionary Conservation of Motor and Interneurons.

Many of the transcription factors defining the presumed cell types and groups of cell types in the Platynereis VNC have previously been shown to control specification of neural cell types in the vertebrates [Tal, Gata1/2/3 (18); Beta3 (33); Brn3, Hb9, HNF6, Lhx1/5, Nk6, Pax6, Phox2 (14, 15); Dach (34)]. This information indicates that at least some of the Platynereis cell types may be specified by similar CoRCs and thus be evolutionarily related (1, 2). Four possible cases of conservation are apparent from our study. First, the nine neuron types in the VGMi that coexpress gata1/2/3 and tal (scl) (Fig. 6 E and F and Fig. S7) may relate to V2b interneurons and Kolmer-Agduhr sensory neurons in the vertebrate basal plate, which are specified by the same transcription factors. However, although the vertebrate V2b CoRC also contains Lmo4, Pdu-lmo4 is not prominently expressed in the 6-dpf VNC (Fig. S4). In addition, vertebrate V2b cells are GABAergic, whereas the Platynereis gata1/2/3⁺, tal⁺ cells show multiple transmitters including GABA. Second, VGMi contains segmentally reiterated cells coexpressing hb9, nk6, and islet (Figs. 5 B and C and 6 E and F) that may represent conserved motor neurons, given that the same three transcription factors specify motor neurons in the vertebrate neural tube (14). Third, VGMl contains pax6⁺, dbx1⁺ neurons (Fig. 6 E and F) that, by molecular identity, may relate to GABAergic commissural interneurons in the vertebrate neural tube (14). Fourth, DGM contains phox2⁺, brn3⁺, isl⁺ neurons (Fig. 6 E and F) that may correspond to sensory interneurons with similar identity in the vertebrate hindbrain (15). Notably, in none of these cases other cells were detected outside of the central nervous system showing similar combinations of transcription factors. Because at this state of analysis not much is known about the axonal projections of the Platynereis neuron types, further morphological and physiological characterization and comparison with vertebrate motor and interneuron types would be desirable. In addition, because phenotypic similarities—and thus, similarities in gene expression—can also arise by convergent or concerted evolution (2), validation of the proposed homologies will require more extensive molecular characterization, including the functional validation of transcription factor involvement in cell-type specification and, ideally, single-cell transcriptomics for an unbiased account of all cellular modules used by each type (35).

Kolmer Agduhr Cells: An Ancient Cell Type of the Mucociliary Sole?

The proposed ancient nature of a neuron-type clade specified by a Gata1/2/3-Tal CoRC finds support by recent studies showing that the same transcription factors are also expressed in Kolmer-Agduhr neurons in fish (23) and in the corresponding central spinal fluid-expressing neurons in mouse (36). Kolmer-Agduhr neurons are hair cell-like sensory neurons that contact the central canal and send out ipsilateral ascending axons (37). Interestingly, the sensory apparatus of Kolmer-Agduhr cells directly contacts Reissner’s fiber, a mucous strand in the neural canal in all vertebrates and in the chordate amphioxus (38), indicative of an ancient origin of Kolmer-Agduhr–like neurons. We have recently proposed that Reissner’s fiber is a remnant of an ancient mucociliary sole (39). Finally, both Kolmer-Agduhr cells and some Platynereis gata1/2/3⁺, tal⁺ neuron types (Fig. 6) share the synthesis and release of serotonin, at least transitorily (ref. 40 and this study). Taken together, these observations indicate that today’s gata1/2/3-tal–expressing neuron types in annelids and vertebrates may stem from Kolmer-Agduhr–like sensory serotoninergic precursors involved in the control of mucociliary locomotion.

Outlook.

ProSPr is a highly valuable method to study the cell-type complement across different developmental stages and across different species that has the potential to significantly change the direction and throughput of research in the field of evolution and development. Once established, cellular expression atlases can be constantly expanded with additional gene sets fed into the resource. Although we exemplify and illustrate the value of our resource with the inferred molecular neuroanatomy of the annelid VNC and the identification of a subset of putative cell types, a similar analysis and comparison is currently carried out for other parts of the nervous system and the entire body across different developmental stages. In addition, spatial mapping of single-cell transcriptomes (41) onto the cellular atlas has huge potential in complementing this resource with full expression profiles for all cell types, and guide the inclusion of new genes to the atlas for refined mapping. Because ProSPr can easily be applied to other stages, this will allow tracking cell types through development. Furthermore, the small size of Platynereis larvae and the relatively small overall cell number enable whole-body connectomics (42), and thus the integration of ProSPr and connectomics data into multidimensional, cellular whole-body atlases, which can finally be complemented by functional characterization of key genes and cell types using Crispr-Cas9. The possible transfer of ProSPr to other model species has the potential to boost comparisons between species and phyla, and to thus contribute to unraveling the evolutionary history of cell types.

Materials and Methods

Platynereis Breeding.

P. dumerilii larvae were obtained from an established culture at the European Molecular Biology Laboratory Heidelberg. Larvae were raised in natural sea water, at 18 °C under a 16-h light/8-h dark cycle.

Whole-Mount in Situ Hybridization.

Available transcriptome and EST data were used to design PCR primers. HotStart Taq Polymerase (Qiagen) or the Phusion polymerase (New England BioLabs) was used for amplification. PCR fragments were cloned into the pCRII-Topo vector (Invitrogen) and validated by sequencing. Animal fixations and whole-mount in situ hybridizations were performed as described previously (43).

Imaging.

Six days post fertilization, Platynereis larvae were mounted in 2,2′-thiodiethanol (TDE) according to Asadulina et al. (25), and imaged in a Leica SP8 confocal microscope using the following parameters: 40× objective, 0.75 zoom, 696 × 696 pixels, and z-step size of 0.55 μm, achieving an isotropic resolution of 550 nm per voxel. DAPI channel was imaged using the 405-nm laser and the NBT/BCIP signal using the 633 laser in reflection mode. Bright field and autofluorescence channels were acquired exciting with the 488 or the 633 laser.

Image Registration.

Image registration was performed using the DAPI stain as the reference signal, following the protocol described in Asadulina et al. (25).

Reference.

The reference was built using the DAPI channels of 153 individual larvae in an iterative registration procedure. First, a rigid transformation was applied to the individuals to position them in the same orientation, and a “rigid average” was calculated. Rigidly oriented animals were registered to this average using affine transformation, and an “affine average” was calculated. Rigidly oriented animals were registered to the affine average using an affine, followed by a deformable registration procedure, and a “deformable average” was calculated. This last step was repeated one more time to generate the final reference as the average result. The reference file can be downloaded from prospr.embl.de. Different tissues have been manually segmented from the reference using TrackEM (44).

ProSPr Pipeline.

The scripts composing the full preregistration and postregistration steps of ProSPr pipeline, coded in Fiji (45), are explained in Fig. S3 and distributed in prospr.embl.de. All gene phylogenetic analyses and raw microscopy files are available on demand. Gene expression maps and the complete atlas can be downloaded from prospr.embl.de.

Bioinformatics Analysis.

The full analysis of the gene-expression atlas has been performed using R Bioconductor. The SV generation script and the recursive partition pipeline for cellular reconstruction, are distributed in prospr.embl.de. For hierarchical clustering and t-SNE, “jaccard” distance from the “vegan” R-package was used. R-package “Rtsne” was used to run the t-SNE algorithm on the data using the following parameters: initial dimensions = 10; perplexity = 31. The method “complete” was used in the hierarchical clustering function.