scRNAseq was performed on 4 dpf larval zebrafish, an age corresponding to many studies of spinal circuitry and electrophysiological analysis of neuronal control over swimming behavior (Bhatt et al., 2007; Liao and Fetcho, 2008; McLean et al., 2008; Satou et al., 2009; Menelaou and McLean, 2012; Wang and Brehm, 2017; Menelaou and McLean, 2019; Kishore et al., 2020; Satou et al., 2020; Wen et al., 2020). For each of two duplicate experiments, ~150 spinal cords were isolated, followed by dissociation into single cells. In each case, ~10,000–15,000 spinal cord cells were sequenced to a depth of 40,000 reads per cell and aligned to an improved zebrafish reference genome that is more inclusive of 5′ and 3′ untranslated regions (Lawson et al., 2020; Figure 1—figure supplement 1; see ‘Materials and methods’). After applying standard quality control filters and removing contamination from outside the spinal cord (Figure 1—figure supplements 2 and 3), data integration resulted in a total of 11,762 spinal cord cells from the combined datasets.

Graph-based, unsupervised clustering of the entire spinal cord transcriptome gave rise to 30 clusters, 27 of which were readily identifiable on the basis of established neuronal or glial markers forming the two broad categories of cell types (Figure 1A). The neuronal markers were elavl4 and snap25a (Figure 1B) and glia markers were gfap, slc1a2b, myrf, and sox10 (Figure 1C). The glial cells (46% of total cells) fell into two non-overlapping groups, corresponding to astrocytes/radial glia (gfap+/slc1a2b+), and oligodendrocytes (sox10+/myrf+) (Figure 1C). Both glial types were composed of multiple clusters, indicating further diversity. The remaining three clusters (clusters 6, 11, and 17, corresponding to 11% of total cells) showed mixed expression of neuronal and glial markers. It is unclear whether this population reflects a true cell type or instead a group of unremoved doublets. Given that these clusters could not be assigned to either neurons or glia with confidence, they were excluded from the subsequent analysis. Additionally, while the glia data are available as a resource, they were not analyzed further in this study.

The neuronal population identified by elavl4+/snap25a+ expression was regrouped into 33 clusters using Seurat (Figure 2A). To assign individual neuronal identities to these transcriptome clusters, we used a combination code that relied on the coexpression of neurotransmitter biosynthesis/transporter genes along with differentially expressed marker genes (DEGs; Figure 2, Table 1; Supplementary file 1). The first code provided clear separation of neuronal populations into four principal categories: glutamatergic (slc17a6a+/slc17a6b+), glycinergic (slc6a5+), GABAergic (gad1b+/gad2+), or cholinergic (chata+/slc18a3a+) types (Figure 2B, Table 1). The second code relied on not only established markers for both zebrafish and mouse spinal neurons, but also new markers identified in this study. Generating the list of candidate markers was aided by previous studies that profiled the neurotransmitter identity of morphologically distinct neurons in the larval spinal cord (Bernhardt et al., 1990; Hale et al., 2001; Higashijima et al., 2004a; Higashijima et al., 2004c). In addition, since many aspects of the transcriptional program that establish the spinal neuronal circuit have been shown to be evolutionarily conserved among vertebrates (Kiehn and Kullander, 2004; Goulding, 2009; Grillner and Jessell, 2009), we cross-referenced our data with recent mouse spinal cord sequencing data to search for homologous marker genes (Delile et al., 2019; Blum et al., 2021).

Using the combination code, we identified specific classes of sensory neurons, Mns and interneurons (Figure 2C, Table 1). Two types of sensory neurons, the glutamatergic Rohon–Beard (RB) cells and GABAergic Kolmer–Agduhr (KA; also referred to as cerebrospinal fluid-contacting neurons [CSF-cN] cells), are each represented by two distinct clusters (Figure 2A and C). Cholinergic clusters formed a prominent group corresponding principally to Mns. Clusters of interneurons corresponded to the inhibitory v1, v2b, and v2s, and excitatory v0v, v2a, and dl1, and dl2 types (Figure 2A and C). The relative abundance of the cell types associated with individual clusters (Table 1) was consistent with those published for in vivo labeling experiments (Higashijima et al., 2004b; Kimura et al., 2006; Ampatzis et al., 2014; Gerber et al., 2019; Satou et al., 2020), suggesting that our preparation protocol and analysis are robust and unbiased in recovering spinal cell populations. Of the 33 clusters in our neuronal dataset, we were able to assign identities to 22 clusters with confidence. There are a number of cell types described for in larval zebrafish, such as the excitatory v3 interneuron or the inhibitory v0d interneuron (Satou et al., 2020; Böhm et al., 2022), that we were unable to identify likely due to lower expression levels of canonical markers at this developmental stage. It is probable that these interneuron populations are present in unassigned clusters.

Our analysis also resolved subtypes within neuronal classes. For example, the two clusters were assigned as KA neurons, which shared a common set of markers for the cerebrospinal fluid-contacting interneuron (pkd1l2a/pkd2l1; Figure 3A). However, assignment to subtype can be made on the basis of differential expression of sst1.1 versus urp1, previously shown to label KA′ and KA′′ functional groups, respectively (Figure 3A; Djenoune et al., 2017; Yang et al., 2020).

Similarly, the glutamatergic v2a interneurons consist of multiple clusters that represented distinct subtypes. This class of interneurons has previously been divided into two subpopulations, types I and II, based on morphology, molecular and functional heterogeneity (Bhatt et al., 2007; McLean and Fetcho, 2009; Ampatzis et al., 2014; Menelaou et al., 2014). One molecular feature differentiating the two types is the higher expression level of both the vsx2 and shox2 marker genes in type I compared to type II v2a (Kimura et al., 2006; Menelaou et al., 2014; Hayashi et al., 2018; Menelaou and McLean, 2019). In our dataset, both the vsx2 and shox2 expression patterns differed among the v2a clusters (Figure 3B). The two clusters with strong vsx2 and shox2 expression likely represent type I v2a neurons, while the third cluster with weak vsx2 expression and no shox2 expression likely represents type II (Figure 3B). For this cluster, we further identified gjd2b, the gene encoding the gap junction protein connexin 35.1δ subunit, as an additional marker gene (Figure 3B). This finding is consistent with previous immunohistochemistry in adult fish demonstrating selective expression of gjd2 in type II v2a axons (Carlisle and Ribera, 2014; Pallucchi et al., 2022). Subtype assignment based on expression patterns of these markers was validated using in situ hybridization in larval Tg(vsx2: Kaede) fish (Figure 3C). Expression of the Kaede fluorescent protein driven by vsx2 promoter labels type I v2a with strong fluorescence compared to weakly labeled type II (Figure 3C). Probes against gjd2b colocalized only with those neurons exhibiting weak fluorescence, confirming its specific expression in the type II v2a subtype (Figure 3C). The two subtypes of v2a interneurons both make direct connections with the PMns and are recruited during high-speed swimming but differ in terms of connection strength and the type II v2a interneuron is recruited more effectively across the range of high swimming speeds (Kimura et al., 2006; Bhatt et al., 2007; McLean and Fetcho, 2009; Menelaou and McLean, 2019). The availability of transcriptomes for these neuronal subgroups provides an opportunity to further mine for differential molecular features responsible for the functional distinction.

In addition to validating established markers for neuronal types and subtypes, our scRNAseq analysis revealed novel marker genes for identifying interneuron transcriptomes, as shown for the commissural local (CoLo) and dorsal longitudinal ascending (DoLA) interneurons.

We next focused on transcriptome comparison within Mn populations to examine their molecular heterogeneity. As a first approach, we isolated Mn clusters from the whole spinal cord dataset based on the overlap of two sets of marker genes (Figure 5A). The first set was based on components of the cholinergic pathway that included slc18a3a (Figure 5A) and chata (Figure 2B). The second set of Mn markers included the transcription factors mnx1, mnx2b, and isl1 (Figure2,5s 2C and 5A; Appel et al., 1995; Hutchinson and Eisen, 2006; Zelenchuk and Brusés, 2011; Asakawa et al., 2012; Seredick et al., 2012). Overall, ~27% of the profiled neuronal single-cell transcriptomes corresponded to Mns (1354 cells) (Figure 5A). As a complementary approach, we also generated samples enriched for a larger number of Mns to increase the statistical power of the clustering analysis. This was achieved through fluorescence-activated cell sorting (FACS) of spinal cells prepared from the fluorescent transgenic fish line, Tg(mnx1:GFP), that broadly labels Mns (Flanagan-Steet et al., 2005; Bello-Rojas et al., 2019). Clustering analysis of the sorted data showed that >92% of the FACS-sorted cells represent Mns based on the same canonical markers used for whole spinal cord, giving rise to 7790 cells for subclustering (Figure 5B). We integrated the datasets from both the computationally isolated and experimentally purified Mn populations and performed clustering analysis using Seurat. A total of 9144 cells were grouped into 10 clusters (Figure 5C). A small cluster (3.6% of total) expressing genes gfra1a and tbx3b (Figure 5—figure supplement 1) was almost entirely sourced from the FACS-sorted datasets (>94%). In addition, it shared markers with a population of non-skeletal muscle Mns recently described in mouse spinal cord scRNAseq datasets (Blum et al., 2021). Therefore, this cluster was not included in further comparisons among Mns that control skeletal muscle contraction.

The transcriptionally distinct clusters were next linked to previously known Mn types. Two broad types of Mns have been described for larval zebrafish, the PMn and the SMn, that are commonly distinguished by birth date, progenitor lineage, and morphological features such as size, location, and periphery innervation pattern (Eisen et al., 1986; Myers et al., 1986; Menelaou and McLean, 2012; Bello-Rojas et al., 2019). There are four large-sized PMns (CaP, MiP, vRoP, and dRoP) in each hemi-segment of the spinal cord, each innervating approximately one-quarter of axial muscle target field (Bello-Rojas et al., 2019; Wen et al., 2020). By contrast, 50–70 SMns are in a more ventral location and display a gradient of sizes and functional properties (Myers et al., 1986; Westerfield et al., 1986; McLean et al., 2007; Asakawa et al., 2012; Wang and Brehm, 2017; Wen et al., 2020). We examined the top DEGs among the clusters (Supplementary file 2) and found two markers, chga and nr2f1a, with non-overlapping expression pattern that could reflect this broad classification (Figure 5C). chga was enriched in one distinct cluster, while nr2f1a was present in the majority of the remaining cells (Figure 5C). Taken together, these two mutually exclusive markers represented ~95% of Mn population.

In situ hybridization labeling with probes against chga and nr2f1a revealed spatially segregated Mn groups in the Tg(mnx1:GFP) fish. Specifically, chga probes labeled dorsal Mns that were large in size, consistent with them being the primary group (Figure 5D). nr2f1a labeling was absent in these cells, but was distributed in larger number of smaller Mns that were more ventrally located in the motor column (Figure 5D). These segregated patterns of expression strongly suggested that chga+ cluster represented the PMns, while nr2f1a marked the major SMn populations.

For further validation, we used the chga probes for in situ hybridization analysis in Tg(SAIG213A; EGFP) fish, in which a single PMn in each hemi-segment, the dorsal projecting CaP, was fluorescently labeled among all the Mns (Muto and Kawakami, 2011; Wen et al., 2020). Strong chga labeling colocalized with the GFP-labeled CaP in each hemi-segment (Figure 5E). We also labeled the MiP and RoP Mns in the spinal cord using the sparse labeling approach and observed high level of chga signal in both PMn types using in situ hybridization (Figure 5F). These results firmly established chga as the marker for PMns, leaving the chga-negative clusters representing SMns (Figure 5G). The number of cells associated with SMn clusters was ~18-fold in excess that of the PMn cluster, consistent with previous cell counts of SMns versus PMns in larval zebrafish (Eisen et al., 1986; Myers et al., 1986; McLean et al., 2007; Menelaou and McLean, 2012; Bello-Rojas et al., 2019).

In contrast to the single cluster associated with the PMn type, SMns were composed of multiple transcriptionally distinct groups. Approximately 95% of SMns fall into three groups that were distinguished by differential expression of three marker genes, foxb1b, alcamb, and bmp16 (Figure 6A). In situ hybridization using the three probes labeled SMns revealed two closely stratified layers that showed differences in dorsal-ventral positioning within the motor column (Figure 6A). The foxb1b+ SMns occupied the more dorsal position of our labeled secondaries while both alcamb+ and bmp16+ SMns shared a more ventral location (Figure 6A). We next tested alcamb+ and foxb1b+ SMns for potential correspondence to the different subsets of SMns labeled with two transgenic lines, Tg(isl1: GFP) and Tg(gata2: GFP) (Appel et al., 1995; Meng et al., 1997; Higashijima et al., 2000). Labeling with in situ probes revealed that expression of alcamb overlapped with GFP+ SMns in the Tg(gata2: GFP) fish (Figure 6B), while foxb1b colocalized with those in Tg(isl1: GFP) (Figure 6C). Thus, foxb1b/alcamb expression distinguishes previously identified SMn subsets. The third cluster, corresponding to the bmp-expressing class of SMns, were fewer in number and the targets are not known. However, bmp signaling has been linked to specification of slow muscle, raising the possibility that this muscle type receives innervation by bmp+ Mns (Kuroda et al., 2013).

We performed differential expression analysis comparing PMn and SMn transcriptomes in order to identify candidate genes that might account for their functional distinctions previously established using electrophysiology. After applying thresholds to the Mn clusters, based on average levels of gene expression, p-value, and the proportion of cells expressing individual genes, we obtained a list of 508 candidates that showed at least a 30% difference in average expression levels between Mn types, 317 of which were higher in primaries. To guide our identification of genes that play potential roles in governing excitability and synaptic transmission differences between the two groups of Mns (Supplementary file 2), we used Gene Ontology (GO) enrichment analysis for the PMn-specific DEGs (The Gene Ontology Consortium, 2021; Ashburner et al., 2000). We focused on significantly enriched GO terms for DEGs that were associated with three broad categories of biological processes – synaptic function, ion channels/transporters and ion homeostasis, and ATP generation (Figure 7—figure supplement 1) – due to their roles in excitability and synaptic transmission. Of the 37 GO terms identified for the PMn DEGs, 24 fell into one of these three broad functionally relevant categories (Figure 7—figure supplement 1). We next manually annotated the molecular function of DEGs by referencing evidence-based database ZFIN (Bradford et al., 2022) in order to identify specific functionally relevant genes differentially expressed between PMn and SMn. For the PMn type, DEGs encoding a large number synaptic proteins included those of the core exocytotic machinery such as isoforms of VAMP, SNAP25, and Syntaxin; regulators of exocytosis, including Synaptotagmins, NSF, Complexins, and RIM; synaptic vesicle proteins, and synaptic structural proteins such as Synuclein, Bassoon, and Piccolo (Figure 7A). Comparisons of expression levels indicate that no synaptic genes were specifically enriched in SMns, but 29 were enriched in PMns compared to SMns (Figure 7A). Additionally, five synaptic genes were shared at approximately equivalent levels between Mn types (Figure 7A). Similarly, eight different voltage-dependent ion channel subunits were preferentially expressed in the PMn, with the highest levels corresponding to the P/Q calcium channel, the β4 sodium channel subunit, and the Kv3.3 potassium channel. Four other ion channel types were shared by both SMns and PMns. As with the synaptic genes, no ion channel candidates were specifically enriched in the SMn type. The absence of unique ion channel and synaptic gene enrichment in SMns contrasted with the high levels of transcriptional factors and RNA binding proteins, neither of which are likely candidates for conferring functional distinctions between Mn types. Approximately two thirds of the 191 genes enriched in the SMns (118/191) were of these categories. The low levels of candidate ion channels and synaptic genes in the SMns might be expected on the basis of greatly reduced levels of synaptic transmission reflected in quantal content and release probability compared to PMn (Wang and Brehm, 2017). These results were consistent even when sorted and unsorted data sources were examined independently (Figure 7—figure supplement 2).

Of particular interest among the genes specific to the PMn type were the trio of highest expressing voltage-dependent ion channel types that have been individually linked in previous studies to augmenting transmitter release and/or AP firing rate. Those channel types include a voltage-dependent P/Q-type calcium channel α subunit (cacna1ab), a Kv3.3 potassium channel α subunit (kcnc3a), and a sodium channel β4 subunit (scn4ba) (Figure 7B and C; Eggermann et al., 2011; Lewis and Raman, 2014; Zhang and Kaczmarek, 2016). The identification of the cacna1ab channel as a top DEG in PMns corroborated previous studies firmly establishing the P/Q-type as the presynaptic active zone calcium channel mediating Ca²⁺-dependent release specifically in the PMns (Wen et al., 2013; Wen et al., 2020). Enriched expression of the sodium channel β subunit scn4ba in PMns was validated by in situ hybridization (Figure 7D). scn4ba probes specifically labeled dorsally located Mns with large sizes in the Tg(mnx1:GFP) fish (Figure 7D). These represented the PMns, as shown by colabeling of individually labeled CaP, MiP, and RoP Mns (Figure 7D and E). The zebrafish kcnc3a has been shown to be expressed preferentially in PMn at embryonic ages (Issa et al., 2011). We tested its expression at 4 dpf fish using immunohistochemical staining. A kcnc3 subtype-specific antibody efficiently labeled NMJ synaptic terminal marked by α-bungarotoxin (α-Btx, Figure 7F2). Since the PMn and SMn axons track along with each other and form synapses with the same postsynaptic receptor clusters (Wen et al., 2020), further resolution was needed to distinguish the PMn and SMn synaptic terminals. For this purpose, we ablated individual CaPs by transiently expressing the phototoxic KillerRed protein (Formella et al., 2018), followed by light inactivation at 2 dpf. By 4 dpf, the ablation of CaP was complete as indicated by the absence of its soma and processes (Figure 7F1). Kcnc3 antibody staining showed a specific signal reduction in synapses located in the ventral-most musculature (Figure 7F2), the target field shared by the CaP and numerous other SMns. This result strongly supports the expression specificity of kcnc3a channel in the PMns.

These specific calcium, potassium, and sodium channel isoforms coexpress specifically among the four types of PMns. We term this collective unit of three different voltage-dependent ion channel types as a ‘channel cassette.’ Over 55% of the cells in PMn cluster express the cassette compared to <3% in the SMns over all (Figure 7C). Together with other synaptic DEGs revealed by the analysis, our results suggested a molecular logic for the functional specialization in PMn synapses, such as strong release and high-frequency firing, that are uniquely associated with escape behavior.

Further insights into the potential contribution the ion channel cassette to fast synaptic function specialization came from examination of their differential expression pattern in other circuitry components involved in high-speed swimming and escape. Those cell types include the CoLo inhibitory and v2a excitatory interneurons, both of which provide direct synaptic connections to the PMn type. Both type I and type II v2a neurons participate in high swim speeds, but the type II is more effectively recruited over the range of higher speeds (Hale et al., 2001; Higashijima et al., 2004c; Kimura et al., 2006; Liao and Fetcho, 2008; McLean et al., 2008; Satou et al., 2009; Ampatzis et al., 2014; Menelaou and McLean, 2019; Satou et al., 2020). Additionally, electrophysiological studies have shown that type II v2a neurons fire faster and more reliably than type I v2a neurons (Menelaou and McLean, 2019). As seen in the PMn type, the coexpression of the ion channel cassette is high in both type II v2a and CoLo interneuron types compared to other interneuron types (Figure 8A). Some of those interneuron types expressed individual components of the cassette, most often kcnc3a or cacna1ab channel types, but not the full cassette. Indeed, the scn4ba sodium channel β subunit is the strongest delineator among the three for inclusion into the ion cassette classification (Figures 7C and 8A).

Top DEGs identified for the PMns that encode synaptic proteins also showed higher levels of expression in type II v2a and CoLo interneurons compared to other cell types, including vamp1a, vamp1b, syt2a, syt2b, snap25a, stxbp1a, and cplx2l, all central players in transmitter release (Figure 8B). This further strengthens the proposition that a synaptic cassette works collectively with the ion channel cassette to create a strong neuronal circuitry controlling the escape behavior and fast swimming behavior. These results are incorporated into a simplified model that potentially accounts for the differential circuitry that distinguish regulation of slow swimming from the power circuits (Figure 8C).

The transgenic line Tg(mnx1:GFP) was provided by Dr. David McLean (Northwestern University). Tg(vsx2:Kaede) was provided by Dr. Joseph Fetcho (Cornell University). Tg(SAIG213A;EGFP), Tg(islet1:GFP) and Tg(gata2:GFP) were maintained in the in-house facility. Zebrafish husbandry and procedures were carried out according to the standards approved by the Institutional Animal Care and Use Committee at Oregon Health & Science University (OHSU) (IP00000344). Experiments were performed using larva at 4 dpf. Sex of the larva cannot be determined at this age.

In most cases, we used the Gal4-UAS system to achieve mosaic expression by co-injection of two plasmids into single-cell embryos: one containing Gal4 driven by cell type-specific promoters and the other containing fluorescent reporter genes under the control of UAS element. The mnx1 or vAChT promoter drove the expression in Mns (mnx1:Gal4 plasmid provided by Dr. McLean, Northwestern University, and vAChT:Gal4 provided by Dr. Joe Fetcho, Cornell University). drmt3a promoter (drmt3a:Gal4 provided by Dr. Shinichi Higashijima, National Institute of Natural Sciences, Japan) was used for expression in glycinergic inhibitory interneurons. mCherry driven by HuC promoter was used to label spinal neurons, including the DoLA interneurons. Injected fish were screened on 3 dpf for sparse fluorescent neurons that could provide detailed morphology.

We transiently expressed the phototoxic KillerRed protein in Tg(SAIG213A:EGFP) fish by injecting a plasmid expressing KillerRed driven by UAS promoter (Addgene plasmid #115516; a gift from Marco Morsch). Fish with KillerRed expression in CaP were identified. Individual CaPs were ablated by light inactivation at 2 dpf by illuminating for 10 min with a 560 nm laser set at high power. This completely bleached KillerRed fluorescence and induced visible blebbing in CaP terminals. Fish were grown to 4 dpf, and the ablation was confirmed by the absence of GFP-labeled soma and neurites.

Whole-mount immunohistochemistry was performed as described previously (Wen et al., 2020). Also, 4 dpf larvae were fixed in 4% paraformaldehyde at 4°C for 4 hr. Zebrafish Kcnc3a channel was labeled using a polyclonal antibody originally generated against the human Kcnc3 protein (Thermo Fisher Scientific PA5-53714) at a concentration of 2.5 μg/ml. Polyclonal anti-GFP (Abcam ab13970) was used at 1 μg/ml. Alexa Fluor-conjugated secondary antibodies (Thermo Fisher Scientific) were used at 1 μg/ml. To mark the location of synapses, 1 μg/ml CF405s-conjugated α-Btx (Biotium) was included in the secondary antibody incubation to label postsynaptic acetylcholine receptors.

Fluorescence in situ RNA hybridization (FISH) was performed on whole-mount 4 dpf larva using the multiplexed hybridization chain reaction RNA–FISH bundle (HCR RNA-FISH) according to the manufacturer’s instructions (Molecular Instruments; Choi et al., 2016). Probe sets for zebrafish alcamb, nr2f1a, chga, scn4ba, foxb1b, bmp16, gjd2b, and pnoca were custom-designed based on sequences (Molecular Instruments) and used at 4 nM each. Fluorescent HCR hairpin amplifiers were used at 60 nM each to detect the probes. GFP and mCherry fluorescence in the transgenic lines and transient labeled neurons survived the FISH protocol with signal loss mostly limited to the periphery. Residual fluorescence was sufficient to mark the location of soma in the spinal cord without the need for additional amplification.

After staining, fixed larval were mounted in 1.5% low melting agarose and imaged on a Zeiss 710 laser-scanning microscope equipped with an LD C-Apochromat ×40/1.2 n.a. objective. Z-stacks of confocal images were acquired using Zen (Carl Zeiss) imaging software and presented as either maximal intensity projection or single focal planes as indicated in the figures (ImageJ, National Institutes of Health).

Single-cell suspensions were prepared from Tg(SAIG213A;EGFP) fish for the two full spinal cord datasets and from Tg(mnx1:GFP) for the two FACS-sorted Mn enrichment datasets.

About 150 4 dpf larva were euthanized in 0.02% tricaine and individually decapitated behind the hindbrain. They were incubated with 20 mg/ml collagenase (Life Sciences) in a buffer containing 134 mM NaCl, 2.9 mM KCl, 1.2 mM MgCl₂, 2.1 mM CaCl₂, and 10 mM Na-HEPES (pH 7.8) at 28°C for 2 hr, with intermittent trituration using a p200 pipette aid at 0, 0.5 hr, and 1 hr of the incubation. To release spinal cords from remaining tissue, the final triturations were done using fire-polished Pasteur pipettes with decreased opening sizes (300, 200, and 100 μm, respectively). Intact spinal cords were transferred to L15 media and washed three times with fresh media. The spinal cords were incubated with 0.25% trypsin solution (in 1× PBS containing 1 mM EDTA) at 28°C for 25 min. The digestion was terminated by adding 500 μl stop solution (L15 with 1% fetal bovine serum). The tissue was collected by spinning at 400 × g for 3 min at 4°C, washed once with L15, and resuspended in 200 μl of L15 media. Spinal cord cells were dissociated by triturating the digested tissue with fire-polished Pasteur pipettes with 80–100 μm opening. The solution was filtered through a 35 μm strainer into a siliconized collection tube. The suspension was examined on a microscope for cell count, Trypan blue staining-based viability test, and proportion of dispersed single cells. Samples with a viability above 70% were used for sequencing.

Single-cell suspension prepared from Tg(mnx1:GFP) fish were FAC-sorted for EGFP+ cells using a 100 μm nozzle on a BD inFlux cell sorter (Flow Cytometry Shared Resource, OHSU). Cells were collected in 100 μl PBS containing 0.2% bovine serum albumin in a siliconized tube.

Single-cell capture, cDNA synthesis, and library preparation were performed by the Massive Parallel Sequencing Shared Resource at OHSU using the 10X Genomics Chromium v3.0 reagent kit. Single-cell suspension for the two full spinal cord replicates targeted 10,000–15,000 cells. For the two FACS-sorted samples, 4000–5000 cells were targeted. Replicate samples were prepared from different clutches of animals. Libraries were sequenced on an Illumina NovaSeq 500 instrument to an average read depth of ~40,000 per cell.

Cellranger v6.1.1 (10X Genomics) was used for the reference genome generation and alignment. Our reference genome was generated by modifying a preexisting reference genome, Lawson v4.3.2 (Lawson et al., 2020), which we edited to add an EGFP sequence as an artificial chromosome and to correct a selection of gene names and 3′ untranslated region (UTR) annotations. A full account of all changes made to the Lawson reference genome can be found in Supplementary file 3, with a representative example shown (Figure 1—figure supplement 1). Alignment to this reference genome was performed using Cellrangers count function with expect-cells set to the targeted number of cells for each replicate.

Count matrices were processed using the Seurat v4.0 package for R (https://www.satijalab.org/seurat/; Hao et al., 2021). Genes with expression in less than three cells were excluded from further analysis. Initial quality control was performed on each sample independently. Cells were kept for further analysis if they had a number of unique genes between 400 and 4000, UMI counts between 1500 and 9000, and <5% mitochondrial gene content. No further doublet removal methods were applied. Data was normalized using the Seurat Sctransform v2 package in R, generally following the procedure outlined in the Introduction to SCTransform v2 regularization vignette (Hafemeister and Satija, 2019; Choudhary and Satija, 2022). Principal components were calculated, followed by nearest-neighbor graph calculation using ANNoy implemented through Seurat. Clustering used the Leiden community detection method implemented with the FindClusters function and the Leidenalg Python package (Traag et al., 2019). To visualize clustered datasets, we used t-distributed stochastic neighbor embedding (t-SNE) (Maaten and Hinton, 2008) or Uniform Manifold Approximation Projection (UMAP) (McInnes et al., 2018) implemented through Seurat.

Datasets were combined using Seurat’s integration pipeline (Stuart et al., 2019), considering 9000 variable genes. To examine the correspondence between duplicates, exploratory clustering was done in the combined datasets (Figure 1—figure supplement 2A and B). Intermixing of data points from different samples was inspected both visually and by plotting the distribution against one another (Figure 1—figure supplement 2A and B). For the two full spinal duplicates, one cluster of cells exhibited a strong bias towards a single replicate, with over 90% of the cells sourcing from a single replicate (Figure 1—figure supplement 2A). These cells were not included in the combined dataset for downstream analysis to control for technical and biological variability. The FACS-sorted duplicates had no obvious outliners (Figure 1—figure supplement 2B).

The combined motor neuron dataset was generated by integrating Mns extracted from the full spinal dataset and the FACS-sorted dataset, based on expression of canonical Mn markers (Figure 5A and B). While the relative populations of SMns did vary between the two sample sources, there was no Mn subtypes that could not be identified independently in both methods of Mn isolation (Figure 7—figure supplement 2). Differential expression analysis performed independently with either Mn source yielded highly reproducible sets of the DEGs (Figure 7—figure supplement 2), further validating the results from the combined dataset.

Specific cell types in the neuronal subset of the spinal data were annotated on the basis of significantly differentially expressed genes. Exploratory clustering analysis was first conducted to filter out cells not of spinal cord origin. One small cluster in the whole spine dataset (0.6% of cells) expresses tph2/ucn3l/slc18a2 at high level compared to all the rest of the clusters (Figure 1—figure supplement 3), marker genes for serotonergic raphe nucleus in the hindbrain (Oikonomou et al., 2019; Bradford et al., 2022). They reflected a trace amount of hind brain tissue during spinal cord dissection, and were removed from the analysis. Contaminating muscle cells were also removed based on expression of my1pfa/tnnt3b/actc1b myosin/troponin/actin genes (Figure 1—figure supplement 3). Overall, these contaminants accounted for ~2.1% of the total cell population isolated from the spinal cord. No cells in the FACS-sorted dataset were identified as originating from contamination from outside the spinal cord.

Differential gene expression was calculated using a Wilcoxon rank-sum test implemented with the FindMarkers or FindAllMarkers functions in Seurat. Genes were considered to be enriched in a cluster if they had a log2 fold change >0.38 (corresponding to >30% enrichment in expression level), expression in at least 30% of cells in one of the groups being compared, and a Bonferroni adjusted p-value of <10^–5. A more conservative adjusted p-value of <10^–10 was used for comparison between PMns and SMns.

DEGs comparing PMns and SMns were used to generate lists of GO terms enriched in the PMns using the EnrichR package in the aspect of ‘biological processes’ (Chen et al., 2013; Kuleshov et al., 2016; Xie et al., 2021). GO terms were considered significantly enriched with adjusted p-value<0.05.