scRNAseq was performed on 4 dpf larval zebrafish, an age corresponding to well documented studies of spinal circuity and patch clamp analysis of neuronal control over swimming behavior (Bhatt et al 2007, Liao & Fetcho 2008, McLean et al 2008, Satou et al 2009, Menelaou & McLean 2012, Wang & Brehm 2017, Menelaou & McLean 2019, Kishore et al 2020, Satou et al 2020, Wen et al 2020). For each of two duplicate experiments, approximately 150 spinal cords were isolated, followed by dissociation into single cells. In each case, ∼10,000 to 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)(see Methods). After applying standard quality control filters and data integration, a total of 11762 spinal cord cells were obtained 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 (Fig. 1A). The neuronal markers were elavl4 and snap25a (Fig. 1B) and glia markers were gfap, slc1a2b, myrf and sox10 (Fig. 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+) (Fig. 1C). Both glial types were composed of multiple clusters, indicating further diversity. The remaining 3 clusters (clusters 6, 11, and 17, corresponding to 11% of total cells) showed mixed expression of neuronal and glial markers. They could not be assigned to either group with confidence and were therefore 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 re-grouped into 33 clusters using Seurat (Fig. 2A). To assign individual neuronal identities to these transcriptome clusters, we used a combination code that relied on the co-expression of neurotransmitter biosynthesis/transporter genes along with differentially expressed marker genes (DEGs; Fig. 2; Table 1; Supplementary Table 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 (Fig 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 & Kullander 2004, Goulding 2009, Grillner & 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 (Fig. 2C, Table 1). Two types of sensory neurons, the glutamatergic Rohon-Beard (RB) cells and GABAergic Kolmer-Agduhr (KA) cells (Fig. 2A & C), are each represented by two distinct clusters. 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 (Fig. 2A & C). The relative abundance of the cell types associated with individual clusters (Table 1) were 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 data 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, Bohm et al 2022) that we were unable to identify, likely due to differences in expression level of canonical markers at different developmental stages. It is probable that these interneuron populations are present in the clusters of unidentified cells.

Our analysis also resolved subtypes within neuronal classes. For example, the two clusters assigned as KA neurons shared a common set of markers for the cerebrospinal fluid contacting interneuron (pkd1l2a/pkd2l1; Fig. 3A). However, assignment to subtype can be made on the basis of differential expression of urp1 versus sst1.1, previously shown to label KA+ and KA functional groups, respectively (Fig. 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 have been classified into two sub-populations, Types I and II, based on morphology, molecular and functional heterogeneity (Bhatt et al 2007, McLean & Fetcho 2009, Ampatzis et al 2014, Menelaou et al 2014). The molecular feature differentiating the two is the higher expression level of vsx2/shox2 maker genes in Type I compared to Type II v2a (Kimura et al 2006, Menelaou et al 2014, Menelaou & McLean 2019). This distinction was reflected in the vsx2/shox2 expression pattern among the three v2a clusters in our dataset (Fig. 3B). The two strong vsx2/shox2 -expressing clusters likely represented Type I v2a neurons, while the third cluster likely represents Type II (Fig. 3B). We further identified gjd2b, the gene encoding the gap junction protein connexin 35.1 8 subunit, as enriched specifically in the Type II v2a cluster (Fig. 3B). This finding is consistent with previous immunohistochemistry in adult fish demonstrating selective expression of gjd2 in Type II v2a axons (Carlisle & Ribera 2014, Pallucchi et al 2022). Subtype assignment based on differential expression patterns of these markers were validated with in situ hybridization in the larval Tg(vsx2: Kaede) fish (Fig. 3C). Expression of the Kaede fluorescent protein driven by vsx2 promoter labeled Type I v2a with strong fluorescence and Type II with weak signal (Fig. 3C). Probes against gjd2b colocalized only with those with weak fluorescence, confirming its specific expression in the Type II v2a subtype (Fig. 3C). The two subtypes of v2a interneurons have been shown to be recruited in swimming behavior of different speeds, with Type II tied to the escape response (Bhatt et al 2007, McLean & Fetcho 2009, Menelaou & McLean 2019). The availability of transcriptomes for these neuronal sub-groups 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 (Fig. 5A). The first set was based on components of the cholinergic pathway that included slc18a3a, (Fig. 5A) and chata (Fig. 2B). The second set of Mn markers included the transcription factors mnx1, mnx2b and isl1 (Fig. 2C, Fig. 5A)(Appel et al 1995, Hutchinson & Eisen 2006, Zelenchuk & Bruses 2011, Asakawa et al 2012, Seredick et al 2012). Overall, ∼27% of the profiled neuronal single cell transcriptomes corresponded to Mns (1354 cells) (Fig. 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 cells represent Mns based on the same canonical markers used for whole spinal cord, giving rise to 7790 cells for sub-clustering (Fig. 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 (Fig. 5C). A small cluster (3.6% of total) expressing genes gfra1a and tbx3b (Supplemental Fig. 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 & 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 & Brehm 2017, Wen et al 2020). We examined the top DEGs among the clusters (supplemental Table 2), and found two markers, chga and nr2f1a, with non-overlapping expression pattern that could reflect this broad classification (Fig. 5C). chga was enriched in one distinct cluster, while nr2f1a was present in the majority of the remaining cells (Fig. 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 (Fig. 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 (Fig. 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 & Kawakami 2011, Wen et al 2020). Strong chga labeling colocalized with the GFP labeled-CaP in each hemi-segment (Fig. 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 (Fig. 5F). These results firmly established chga as the marker for PMns, leaving the chga negative clusters representing SMns (Fig. 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 & 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 (Fig. 6A). In situ hybridization using the three probes labeled SMns with different dorsal-ventral positions in the motor column (Fig. 6A). The foxb1b+ SMns occupied positions dorsal to alcamb+ SMns. bmp16+ SMns shared a location similar to alcamb+ SMns, but were far fewer in number (Fig. 6A). We next applied alcamb and foxb1b probes to two fluorescent transgenic lines that each preferentially labeled a subset of SMns with different peripheral targets. The Tg(isl1: GFP) predominantly labels SMns that innervate dorsal muscles, while Tg(gata2:GFP) labels SMns that innervate ventral muscles (Appel et al 1995, Meng et al 1997, Higashijima et al 2000). The expression of alcamb overlapped with GFP+ SMns in the Tg(gata2: GFP) fish (Fig. 6B), while foxb1b co-localized with those in Tg(isl1: GFP) (Fig. 6C), suggesting that alcamb/foxb1b expression distinguishes SMns with ventral versus dorsal target fields. These results are consistent with past functional studies that have shown that SMns display a gradient of features reflected in their anatomical position, morphology, muscle target field size, membrane biophysical properties and synaptic transmission strength (McLean et al 2007, Wang & Brehm 2017, Bello-Rojas et al 2019, Wen et al 2020).

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, based on both average levels of gene expression and proportion of cells expressing individual genes, we obtained a list of 508 candidates that showed over 30% difference between the two groups of Mns (Supplemental Table 2). About two thirds of the 191 genes enriched in the SMns (118/191) encoded either transcriptional factors or RNA binding proteins, neither of which are likely candidates for functional distinctions between Mn types. Instead, the higher levels of these classes in SMns may reflect the less well differentiated state that accompanies the later birthdate of the SMns. Consequently, attention was focused on the 317 genes enriched in PMn transcriptomes, in order to search for transcriptional correlates that potentially account for stronger synaptic release and higher firing frequencies that distinguish PMns from SMns.

A glimpse into the biological pathways individual DEGs involved in was provided by Gene Ontology (GO) enrichment analysis in the aspect of “biological process” (The Gene Ontology Consortium, 2021)(Ashburner et al 2000). The molecular function of DEGs was manually annotated by referencing evidence-based database ZFIN (Bradford et al 2022). Significantly enriched GO terms for PMn DEGs were predominantly associated with 3 broad categories of biological processes, synaptic function, ion channels/transporters and ion homeostasis, and ATP generation (Supplemental Fig. 2). We focused on the categories of synaptic function and ion channels/transporters, due to their published roles in excitability and synaptic transmission. DEGs encoding 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 protein such as Synuclein, Bassoon and Piccolo (Fig. 7A). Top DEGs among the second category were a trio of 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 a subunit (cacna1ab), a Kv3.3 potassium channel a subunit (kcnc3a), and a sodium channel ý4 subunit (scn4ba) (Fig. 7B & C) (Eggermann et al 2011, Lewis & Raman 2014, Zhang & 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 ý subunit scn4ba in PMns was validated by in situ hybridization (Fig. 7D). scn4ba probes specifically labeled dorsally-located Mns with large sizes in the Tg(mnx1:GFP) fish (Fig. 7D). These represented the PMns, as shown by co-labeling of individually labeled CaP, MiP and RoP Mns (Fig. 7D & 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 sub-type specific antibody efficiently labeled NMJ synaptic terminal marked by α-bungarotoxin (α-Btx, Fig. 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 (Fig. 7F1). Kcnc3 antibody staining showed a specific signal reduction in synapses located in the ventral-most musculature (Fig. 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 co-express 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 (Fig. 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 neuronal components of the escape circuit. Specifically, the Type II v2a excitatory interneuron and the CoLo inhibitory interneuron have both been shown in electrophysiological recordings to form direct connections with the PMns and to selectively engage in escape response versus swimming behavior (Hale et al 2001, Higashijima et al 2004c, Kimura et al 2006, Liao & Fetcho 2008, McLean et al 2008, Satou et al 2009, Ampatzis et al 2014, Menelaou & McLean 2019, Satou et al 2020). Examination of the cassette expression revealed high levels in these two interneuron types compared to other interneuron types not involved in escape (Fig. 8A). Some of the interneuron types not involved in escape also expressed kcnc3a /cacna1ab channel types, but not the full cassette. Among the cassette members it appeared that the scn4ba sodium channel β subunit was the limiting element for neurons of escape circuit (Fig. 8A & 7C).

Enrichment of the ion channel cassette in all members of the ultrafast escape circuit lends support to the idea that it serves as a general molecular code for fast and strong synapses. 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 (Fig. 8B). This further suggests that a synaptic cassette works collectively with the ion channel cassette to create a strong neuronal circuitry controlling the escape behavior. These results are incorporated into a simplified model that potentially accounts for the differential circuitry that regulate separate behaviors (Fig. 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 the standards approved by Institutional Animal Care and Use Committee at Oregon Health & Science University (OHSU). 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 mins 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). 4 dpf larvae were fixed in 4% paraformaldehyde at 4 °C for 4 hrs. Zebrafish Kcnc3a channel was labeled using a polyclonal antibody originally generated against the human Kcnc3 protein (ThermoFisher 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 40x/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, 100 μm respectively). Intact spinal cords were transferred to L15 media and washed 3 times with fresh media. The spinal cords were incubated with 0.25% trypsin solution (in 1xPBS 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 4,000-5,000 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 Supplemental Table 3, with a representative example shown (Supplemental Fig. 3). 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 (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. 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 & Satija 2019, Choudhary & 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 data sets we used t distributed stochastic neighbor embedding (t-SNE) (van der Maaten & Hinton 2008) or uniform manifold approximation projection (UMAP) (McInnes & Healy 2018) implemented through Seurat.

Datasets were combined using Seurat’s integration pipeline (Stuart et al 2019), considering 9,000 variable genes. To examine the correspondence between duplicates, exploratory clustering was done in the combined datasets (Supplemental Fig. 4A & B). Intermixing of data points from different samples was inspected both visually, and by plotting the distribution against one another (Supplemental Fig. 4A & 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 (Supplemental Fig. 4A). 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 (Supplemental Fig. 4B).

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 (Fig. 5A & 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 (Supplemental Fig. 5A). Differential expression analysis performed independently with either Mn source yielded highly reproducible sets of the DEGs (Supplemental Fig. 5B), 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 (Supplemental Fig. 6), 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 (Supplemental Fig. 6). Overall, these contaminants accounted for ∼2.1% of the total cell population isolated from the spinal cord. No cells in the FACS sorted data set 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 gene ontology (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.

Raw fastq files and unprocessed aligned data can be accessed for free through the Gene Expression Omnibus (accession number GSE232801). All code used in data processing and figure creation has been deposited, and is freely available on github (https://github.com/JimmyKelly-bio/Single-cell-RNA-seq-analysis-of-spinal-locomotor-circuitry-in-larval-zebrafish).