Animal breeding and experiments were approved by the regional councils of the Land Rhineland-Palatinate according to the German Animal Protection Law (TSchG §4/3) and followed the guidelines for the welfare of laboratory animals. C57BL/6 J mice were bred in the animal facilities of the University of Kaiserslautern, and both sexes were analyzed at postnatal day (P) 11 ± 1, i.e., around hearing onset.

RNase-free solutions were prepared, and equipment was cleaned as described (Cadwell et al., 2017). Coronal brainstem slices were prepared (Hirtz et al., 2011) and in vitro electrophysiology was performed as described (Müller et al., 2019). Briefly, slices were perfused with ASCF (in mM: 125 NaCl, 2.5 KCl, 1 MgCl₂, 1.25 NaH₂PO₄, 2 sodium pyruvate, 3 myo-inositol, 0.44 L-ascorbic acid, 25 NaHCO₃, 10 D-glucose, 2 CaCl_2; pH 7.4 when oxygenated with carbogen). Whole-cell patch-clamp recordings from LSO neurons were performed at 36.5 ± 1°C. Patch pipettes were filled with internal solution (in mM: 140 potassium gluconate, 10 HEPES, 5 EGTA, 1 MgCl₂, 2 Na₂ATP, 0.3 Na₂GTP). The internal solution contained 0.8% RNase inhibitor (Invitrogen). The liquid junction potential was 15.4 mV and was corrected online.

Electrophysiological characterization was limited to ~10–15 min to minimize mRNA degradation. Total RNA was then extracted into the patch pipette by gently aspirating the neuronal cytoplasm. Shrinkage of the soma and entry of the nucleus into the pipette were visually monitored under DIC optics (Supplementary Figures S1A1–A3). After successful harvesting, atmospheric pressure was restored and the pipette was slowly removed from the soma (Supplementary Figure S1B). In case of contamination of the pipette by extracellular material, absence of a membrane patch at the tip of the pipette, or soma swelling, samples were discarded (Supplementary Figures S1C1,C2,D1,D2). The contents of the pipette were collected in a reaction tube containing 2 μL lysis buffer, 0.2% Triton-X, ERCC RNA Spike-In Mix (Invitrogen, 1:400,000 dilution) and RNase inhibitor. Samples were stored at −80°C and processed further en bloc (library preparation and sequencing). From a total of 263 patched neurons, 103 were retained for further analysis. Ultimately, 86 neurons passed the RNA quality control (see below). One negative control was collected per experimental day. A pipette filled with internal solution was moved so that its tip touched the slice. It was then withdrawn, and the contents of the pipette were processed as for harvested cells.

Double-stranded cDNA was generated as described, with minor modifications (Picelli et al., 2013). Briefly, 1 μL of 10 mM dNTPs and 0.5 μL of Oligo-dT primer (Supplementary Table S1) were added to each sample, and samples were incubated at 72°C for 3 min and immediately placed on ice. Reverse transcription was performed using 0.5 μL SuperScript II RT (200 U/μL, Invitrogen) supplemented with 0.25 μL RNasin (40 U/μL, Promega), 2 μL Superscript II first-strand buffer (5x), 0.48 μL 100 mM DTT, 2 μL 5 M betaine, 0.12 μL 0.5 M MgCl₂, 0.1 μL 100 μM template-switching oligonucleotide and 0.25 μL nuclease-free water. Incubation was performed at 42°C for 90 min, followed by 10 cycles at 50°C for 2 min and another incubation step at 42°C for 2 min. Enzyme inactivation was achieved by incubation at 70°C for 15 min. cDNA was pre-amplified by adding 12.5 μL KAPA HiFi HotStart ReadyMix PCR Kit (Roche), 0.25 μL IS PCR primer and 2.25 μL nuclease-free water with these thermocycling conditions: 98°C for 3 min, 18 cycles at 98°C for 20 s, 67°C for 15 s, 72°C for 6 min and final elongation at 72°C for 5 min. cDNA was purified using 0.8 X Ampure XP Beads (Beckman Coulter) and elution in 7 μL RNase free water. Quality control of randomly selected cells was assessed on an Agilent Bioanalyzer using the HS DNA Kit. On average, 450 pg. of cDNA was used as input for dual-indexed Nextera XT library preparation (Illumina). A total of 9 cycles of library amplification were performed according to the manufacturer’s recommendations. All individual single-cell libraries were pooled in equal amounts prior to purification of 100 μL pooled libraries with 1X Ampure XP Beads (Beckman Coulter) and final elution in 15 μL RNase-free water.

Prior to sequencing, the library pool was quantified using the Qubit dsDNA HS Assay Kit (Invitrogen), and the fragment size distribution was checked using the High Sensitivity DNA Kit (Agilent) on an Agilent 2100 Bioanalyzer. The library pool was additionally quantified with the PerfeCTa NGS Quantification Kit (Quanta Biosciences) and normalized for clustering on a cBot (Illumina). Libraries were sequenced on a HiSeq2500 (Illumina) using dual index reads with 88 bp read length.

Initial quality control of the raw data was performed using FastQC.¹ Reads were trimmed using Trim Galore! (v0.4.2)² to remove 3′ ends with base quality below 20 and adapter sequences. Reads were aligned to the mouse genome mm10 using STAR (Dobin et al., 2013, v2.5.2a) with a 2-pass mapping strategy per sample. PCR duplicates were detected using MarkDuplicate from the Picard tools (version 1.115).³ Reads aligned to mm10 were summarized to Gencode annotation vM2 (Harrow et al., 2012) using featureCounts (Liao et al., 2014, v1.5.0-p3) and counting primary alignments only (Li and Dewey, 2011, v1.3.1).

For quality control, filtering, and normalization of single-cell RNA-seq (scRNA-seq) data, we used the SingleCellExperiment R package (version 1.8.0, Amezquita et al., 2020) and the scran package (version 1.14.6; Lun et al., 2016) on Bioconductor. The deconvolution method was used to remove cell-specific library size biases, and size factors were calculated for each library using the ComputeSumFactors function. Of 103 aspirated cells, 17 (~16%) were discarded due to poor sequencing quality, leaving 86 neurons. Poor quality cells had >3 median absolute differences below the median for number of counts or number of transcripts. Genes with an average of at least one count in at least 5% of the neurons qualified for downstream analysis, resulting in 11,659 expressed genes. Transcripts per million (TPM) values were calculated by normalizing for gene length and sequencing depth (Zheng et al., 2020). The top 500 highly expressed genes were selected for Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis using gene set enrichment analysis (GSEA) performed with DAVID 2021 (Huang et al., 2009; Sherman et al., 2022).

Clustering was performed using SC3 (Kiselev et al., 2017) and Seurat (Satija et al., 2015). Principal component analysis (PCA) was performed using the Seurat function RunPCA (Lun et al., 2016). Normalized log₂-transformed highly variable gene (HVG) counts were used as input. t-Distributed Stochastic Neighbor Embedding (tSNE) (Kobak and Berens, 2019) plots were generated with RunTSNE.

DEGs were detected using the findMarkers function from the scran package, which uses a t-test by default. A gene was defined as a DEG if the normalized count showed a ≥ 2-fold upregulation (log₂FC ≥ 1) and a false discovery rate (FDR) ≤ 0.05 (-log₁₀FDR ≤ 1.3). DEGs were ranked by their Area Under the Receiver-Operating Characteristic curve (AUROC). DEGs were also selected for Gene Ontology (GO) enrichment analysis using GSEA performed with DAVID 2021. GO terms were condensed using REVIGO (Jiang and Conrath algorithm) (Supek et al., 2011). UpSet plots were generated using the UpSetR package (v.1.4.0) (Conway et al., 2017).

The single-cell RNA sequencing data generated during this research project has been submitted to NCBI GEO and can be accessed using the accession number GSE241761.

Lethally anesthetized mice (7% chloral hydrate, 0.01 mL/g body weight or 10% ketamine, 220 μg/g body weight plus 2% xylazine, 24 μg/g body weight) were transcardially perfused with 100 mM phosphate-buffered saline (PBS, pH 7.4) at room temperature (RT), followed by ice-cold 4% paraformaldehyde (PFA) for 20 min. Brains were removed from the skull, postfixed for 2 h in 4% PFA (4°C) and stored in 30% sucrose-PBS at 4°C overnight. For Kv7.2, Kv7.3, and osteopontin, brain fixation was performed with Zamboni solution. 30-μm-thick brainstem slices were cut with a sliding microtome (HM 400 R, MICROM) and transferred to 15% sucrose-PBS, followed by three rinses in PBS (30 min, RT). Incubation and rinsing steps were performed free-floating unless otherwise noted. For Kvβ3, a heat-induced epitope retrieval protocol was used prior to antibody treatment. For this, slices were immersed in 10 mM sodium citrate buffer (pH 6.0) plus 0.05% Tween 20 (Roth) and heated at 95°C for 20–40 min. They were then returned to RT, followed by three rinses in PBS (5 min, RT).

The Kv7.2/3 agonist retigabine (N-(2-amino-4-(4-fluorobenzylamino)phenyl) carbamic acid ethyl ester; Sigma-Aldrich) was dissolved in water and stored at 4°C. On the day of the experiment, this stock solution (100 mM) was diluted, and retigabine (30 μM) was applied 10 min before and throughout the electrophysiological measurements. To analyze drug effects, we used the same current step protocol as described above in the electrophysiology section, but sampled at 20 kHz.

Statistical analyses were performed using Excel (Microsoft) or GraphPad Prism 9 (GraphPad Software). Data were tested for normal distribution using the Kolmogorov–Smirnov test. If normally distributed, they were compared in a 2-tailed Student’s t-test. Otherwise, a Mann–Whitney-U test (unpaired data) or a Wilcoxon signed-rank test (paired data) was used. The Šidák correction was used for multiple comparisons.

For selected genes, we compared the expression level to in situ hybridization results from the AMBA.⁴ We primarily analyzed P56 sections. They were chosen because >40% of the genes of interest were missing data at P14. P14 images in the AMBA were mainly sagittal views and less informative. Therefore, they were analyzed for comparison purposes only. The images show slightly modified coronal or sagittal sections (Supplementary Figures S4, S5). They represent gene expression energy values per 200-μm voxel. To allow direct navigation to an image of interest, we have provided the hyperlinks in Supplementary Table S7.

The LSO is a brainstem nucleus containing neuron types with ascending and descending projections (Figure 1A). To gain a comprehensive understanding about the molecular and physiological complexity of this prominent nucleus, we performed patch-seq experiments in acute slices (Figure 1B). We recorded from 263 patch-clamped LSO neurons and harvested the cytoplasm in 103 cases (see Methods for exclusion criteria). Of these 103, 86 passed the quality controls and displayed 1.6 × 10⁶ mapped reads (median; interquartile range (IQR) 1.3 × 10⁶; Figure 1C). Negative controls yielded an 8-fold lower number (median: 0.2 × 10⁶; IQR: 0.2 × 10⁶). The mean mapping rate exceeded 80% and was thus >2-fold higher than in negative controls (Supplementary Figure S2A). The largest proportion of reads mapped to exons (45%), an almost 3-fold higher value than in negative controls (Supplementary Figures S2C,D). The median percentage of ribosomal RNA was 3.8% (IQR: 4.5%; Supplementary Figure S2B). With one exception, no single value exceeded 15%, indicating sufficient rRNA depletion (Dobin and Gingeras, 2015). The number of detected genes per neuron or negative control correlated with the number of reads mapped. A mean of 2,863 genes were found per neuron (median: 2.9 × 10³; IQR: 1.2 × 10³; Figure 1D) and 310 genes in negative controls, a > 9-fold lower value (median; IQR: 565; Figure 1D). The number of mapped reads, number of detected genes, mapping rate, and rRNA rate were consistent with published patch-seq datasets (Cadwell et al., 2017, 2020), demonstrating a high library quality.

A major advantage of the single-cell patch-seq approach, which is absent from bulk analyses (Koehl et al., 2004; Nothwang et al., 2006; Ehmann et al., 2013; Xiao et al., 2013; Moritz et al., 2015; Kolson et al., 2016), is the ability to focus on neurons instead of globally addressing different cell types (including glia or blood endothelium). We took advantage of this and examined the most abundant transcripts in the 86 LSO neurons. Within the cohort of 9,082 genes with ≥1 TPM, 185 and 60 showed TPM values >1,024 and > 2,048, respectively (Figure 2A). The 50 most highly expressed genes are shown in Figure 2B (32,581–2,460 TPM, inset). For 20 of them, bee swarm and violin plots are depicted in Figure 2C. Three genes showed exceptionally high expression levels, namely Ckb, Snurf, and Calm1 (Figure 2B; mean TPM: 32,581, 28,886, 24,914). They encode brain-type creatine kinase B (CKB), SNRPN upstream open reading frame protein (SNURF; SNRPN = small nuclear ribonucleoprotein polypeptide N), and calmodulin 1 (CALM1 aka CaM), respectively. Our findings are corroborated by in situ hybridization results from the AMBA which show high signal intensities in the major nuclei of the superior olivary complex (SOC) (Figure 2D). Further information on Ckb, Snurf, and Calm1 is provided in the Discussion.

Glucose consumption is a signature of brain activity (Ashrafi and Ryan, 2017) and is particularly high in the CAS (Sokoloff, 1981). As 13 of the 50 most highly expressed genes in our sample encode enzymes involved in catabolism, we examined pathways of energy demand and ATP production in more detail. We analyzed gene enrichment in pathways that completely combust glucose to CO₂ and H₂O: (1) Glycolysis (including Pyruvate oxidation), (2) Citrate cycle, and (3) Oxidative phosphorylation (OxPhos; Supplementary Table S2). The analysis included 162 genes, including Ckb (4). Among the top 500 highly expressed genes, 7% (33) belonged into these pathways and were distributed across the four categories (Figure 2E). When normalized to the set of 162, the frequency was 20%. Taken together, these numbers emphasize and specify the genes involved in ATP generation and highlight the importance of energy metabolism in the LSO.

Gene enrichment analysis revealed seven genes among the top 500 that encode v-ATPase subunits (Atp6ap1, Atp6v1b2, Atp6v6v11, Atp6v0c, Atp6v0d1, Atpv1e1, Atp6v1d; 3,589–438 TPM; #24 to #460). Twenty-three v-ATPase genes have been described (Alexander et al., 2011). v-ATPases are associated with lysosomes and primary-active proton pumps which play an important role in driving transmitter loading into synaptic vesicles (Egashira et al., 2015). Our results indicate that LSO neurons were in a metabolically active state.

To identify neuron types in the LSO, we applied unsupervised clustering. We determined the consensus matrix for 2–10 clusters using SC3 and calculated the average silhouette width for each number of clusters (widths vary from 0 to 1; a width close to 1 represents the optimal number of clusters). We obtained the largest width (0.98) with two clusters (Supplementary Figures S3A,B). Fifty-six of the 86 neurons (~2/3) belonged to cluster 1, and 30 (~1/3) belonged to cluster 2 (Figure 3A). Similarity values were close to 1 for each neuron in its own cluster and close to 0 in the other cluster, except for neuron #1 in cluster 2. We checked for consistent results using the alternative tool Seurat, which assigned 59 neurons to cluster 1 and 27 neurons to cluster 2 (Supplementary Figure S3C). Fifty-six neurons in cluster 1 overlapped between SC3 and Seurat, while 27 neurons overlapped in cluster 2. The substantial overlap demonstrated consistency between the two tools. For three clusters, the results were less consistent and the overlap between SC3 and Seurat was small (Supplementary Figure S3D). Taken together, these results confirm the best results with two clusters.

To visualize the clustering results, we performed a PCA based on HVG (Methods and Supplementary Figure S3E). PC1 explained 11% of the variance and separated both clusters, except for two neurons belonging to cluster 2. Because PC1 plus PC2 together explained only 17% of the variance, we generated a tSNE plot using PC1–PC10. Again, the neurons were clearly separated into two clusters, except for two neurons (Figure 3B; see also Supplementary Figure S3E). Overall, these results demonstrate two major clusters for our sample, regardless of the clustering method used. Below, we explore the two clusters obtained by SC3.

After dividing the 86 patch-seq neurons into two clusters, we assessed the expression level of 11,659 genes (Figure 3C). Two hundred and seventy-one genes exceeded TPM ≥ 1,000 in at least one cluster. Approximately 50% of the genes overlapped (130/271; inset), indicating considerable similarity between the two clusters. Each of the 13 genes involved in ATP synthesis shown in Figure 2B ended up in the intersection, suggesting high metabolic demands in both clusters. Among the 141 non-intersecting genes with ≥1,000 TPM (46 in cluster 1; 95 in cluster 2), only four and six, respectively, showed <100 TPM in the complementary cluster (Figure 3D). This again indicates homogeneity between the two clusters for many transcripts.

The above comparisons revealed considerable similarities between the clusters. To identify differences, we focused on DEGs (see Methods for definition). As documented in a volcano plot, 353 DEGs were identified, of which 254 were upregulated in cluster 1 and 99 in cluster 2 (Figure 4A; Supplementary Table S3). To identify the top DEGs, the DEGs were ranked by the AUROC value. In Figure 4B, the expression of the top 40 DEGs, 20 in each cluster, is visualized in heatmaps (56 neurons in cluster 1, 30 in cluster 2). Below and in Supplementary results, we provide some background information for each of these top 40 DEGs.

Taken together, DEGs#1–20 in cluster 1 plus cluster2 contain eight genes (20%) whose expression was previously described in the LSO (cluster 1: Spp1, Cacng5, Cck; cluster 2: Calca, Ucn, Nxph1, Isl1, Gata3). We now attribute their expression to neurons. The results for the remaining 80% of the genes are new. Each of the three known DEGs in cluster 1 can be associated with pLSOs. Similarly, four genes from cluster 2 can be associated with LOCs (Calca, Ucn, Isl1, Gata3). Thus, the transcriptomic results provide substantial and consistent evidence to conclude that cluster 1 and cluster 2 contain pLSOs and LOCs, respectively. If our conclusion is correct (see below), we identified 17 novel marker genes for pLSOs and 15 for LOCs (highlighted in bold in Figure 4B).

Figure 4B shows 40 DEGs ranked by the z-score. Corresponding expression levels are detailed in violin plots (Figures 4C,D), highlighting significant expression differences. We also plotted the corresponding TPM values for both clusters (Figure 4E and inset). The highest expression in cluster 2 was seen for Pgrmc1 (3,462 TPM). However, this DEG did not end up higher than #18 (Figure 4B), because 50% of cluster 1 neurons also expressed it, at levels up to 4,763 TPM (framed inset in Figure 4E). In cluster 1, TPM values were relatively low, and the highest value was found for Spp1 (273; 0 in cluster 2). Overall, 20 of the 40 DEGs showed TPM values ≤10 (15 in cluster 1; 5 in cluster 2). For these, we did not provide names (Figure 4E and inset, gene names provided only if >10 TPM).

We compared our data with AMBA’s P14 and P56 data, covering 17 of the top 40 DEGs at P14 (42.5%) and 38 DEGs at P28/P56 (95%). A developmental assessment revealed virtually no change in the expression profile of six DEGs in pLSO neurons, while two transcription factors (Rorb, Tcf4) and Cacng5 appeared to be downregulated with age. Among the marker genes for LOC neurons, seven genes (Calca, Ucn, Cit, Fxyd6, Gata3, Nfia, Atp8a1) showed lower levels at P56, with only two showing no change, indicating developmental downregulation.

Next, we analyzed the 353 DEGs according to GO annotations. Several GO terms were enriched (Figure 4F; Supplementary Table S4). More specifically, the terms “Nervous system development,” “Calcium ion binding,” and “Ion transport” appeared in both clusters. Within the GO term “Hormone activity,” which was unique for cluster 2, Calca and Ucn (DEG#1 and #2) encode CGRP and urocortin, established neurotransmitters for LOCs (cf. Introduction). The third gene in this GO term was Uts2b (DEG#25), which encodes the neuromodulator urotensin 2B (Supplementary Table S3). Urotensin 2B is involved in motoneuron function and stimulates ACh release (Huitron-Resendiz et al., 2005; Nothacker and Clark, 2005). The fourth “Hormone activity” gene was Calcb (DEG#34). It is a paralog of Calca and codes for the calcitonin related polypeptide β-CGRP (formerly known as CGRP-II; Alexander et al., 2021). β-CGRP enables calcitonin receptor binding activity and participates in regulating the cytosolic Ca²⁺ concentration. Recently, Calcb expression has been associated with a subset of LOCs (Frank et al., 2023). The fifth gene in the GO term was Igf1 (DEG#58), which encodes the insulin growth factor 1 (IGF-I). IGF-I is involved in differentiation and maturation of neurons and synapses (Nieto-Estévez et al., 2016). Finally, the sixth “Hormone activity” gene was Vgf (DEG#77), which encodes the neurosecretory protein VGF (= VGF nerve growth factor inducible). VGF is a neuropeptide precursor involved in neurogenesis and metabolic regulation.

Collectively, the results from the unsupervised clustering analysis clearly affiliate the two clusters with pLSOs and LOCs.

We identified Spp1|OPN as DEG#1 in cluster 1 (presumably pLSOs) and Calca|CGRP as DEG#1 in cluster 2 (presumably LOCs). To assess correspondence between gene expression and protein abundance, we performed immunohistochemical experiments. Immunosignals for OPN were intense in the LSO and other SOC nuclei (Figures 5A1,B1). They were spread throughout the LSO and were associated with somata and proximal dendrites of cells reminiscent of pLSOs, because of their fusiform shape and size (Figures 5A2,A3,B2,B3). Labeling was particularly strong in a cup-shaped structure within the cytoplasm and virtually absent from the cells’ nucleus. Double labeling with glycine transporter type 2 (GlyT2), used to mark the perisomatic zone, demonstrated cytoplasmic OPN localization, rather than an association with the plasma membrane (Figures 5A3,B3). Conjunctively, the results demonstrate a correspondence between high Spp1 expression and a high OPN abundance, implying a close correlation between mRNA and protein level. Moreover, OPN appears to be associated with pLSOs.

Because of the elaborate demonstration of CGRP in coronal sections (references in Introduction), we characterized the protein in horizontal sections (Figure 5C). CGRP immunosignals were strong in the LSO, as for OPN. In contrast, they were virtually absent from other SOC nuclei (Figure 5C1). Labeled somata were located along the tonotopic axis (from lateral to medial), and their size was a bit smaller than for OPN-positive somata (Figures 5C2,C3). Immunopositive dendrites were oriented perpendicular to the tonotopic axis, as described for LOCs (Friauf et al., 2019). Cytoplasmic labeling was more homogeneous than for OPN. Together, the Calca|CGRP findings imply a close relationship between gene expression and protein abundance. They also provide additional evidence that cluster 2 comprises LOCs. In the following, we will exploit the electrophysiological properties of our patch-seq neurons to assess whether cluster 1 comprises pLSOs, whereas cluster 2 comprises LOCs.

A great advantage of the patch-seq method is its potential to characterize electrophysiological parameters prior to scRNA-seq. We determined 16 physiological features prior to cytoplasm harvesting. Because of the comprehensive knowledge about such features and the differences between pLSOs and LOCs (Sterenborg et al., 2010), we assessed whether the two clusters are also distinguishable by electrophysiological parameters.

One important biophysical feature is the firing pattern. Among cluster 1 neurons, we observed two types of firing pattern. At rheobase, ~80% (44/56) neurons displayed a single AP with a short latency (Figures 6A,D; Supplementary Figure S6B; Supplementary Table S5.1). We refer to them as “pLSO—Onset.” The remaining 20% of cluster 1 neurons fired ~2 APs near rheobase (Figure 6B; Supplementary Figure S6B). The latency of the first AP was longer than for Onset neurons (7.0 ms vs. 2.9 ms; Supplementary Table S5.1). We refer to these neurons as “pLSO—Sustained.” Cluster 2 neurons displayed a delay-type repetitive firing pattern that clearly distinguished them from cluster 1 neurons (Figure 6C, Supplementary Figure S6B; Supplementary Table 5.1). Near rheobase, ~4 APs were generated, and the latency of the first AP (29.4 ms) was several-fold longer than for Onset and Sustained neurons. These neurons are named “LOC—Delayed.” A delay-type repetitive firing pattern is characteristic for LOCs, identified as such via retrograde labeling from the cochlea (Fujino et al., 1997). The authors correlated the firing pattern to A-currents.

The approximated rheobase values differed statistically between the three neuron types (Supplementary Figure S6A). With increasing current pulse amplitudes, pLSO-Onset neurons kept firing only few APs (Supplementary Figure S6B). In contrast, Sustained neurons increased their firing rate linearly, whereas Delayed neurons peaked and declined thereafter. These findings are in accordance with previous studies on pLSOs and LOCs in gerbils (Sanes, 1990; Walcher et al., 2011), rats (Kandler and Friauf, 1995; Fujino et al., 1997; Adam et al., 1999, 2001), and mice (Sterenborg et al., 2010; Walcher et al., 2011; Haragopal and Winters, 2023). Therefore, they provide independent physiological evidence that cluster 1 consists of pLSOs whereas cluster 2 is composed of LOCs.

pLSOs and LOCs can be distinguished by a sag behavior of the membrane potential in response to hyperpolarizing current steps. We found prominent sag behavior in Onset and Sustained neurons, but a virtual absence in Delayed neurons (Figures 6E,F). These findings confirm published results. pLSOs specifically show a depolarizing rebound behavior after the termination of hyperpolarizing current pulses (Kandler and Friauf, 1995; Walcher et al., 2011). Indeed, we found a rebound behavior in Onset and Sustained neurons, but not in the Delayed type (Figure 6E).

We further compared V_rest, R_in, τ_m, C_m as well as the AP peak amplitude, halfwidth, and threshold across the three neuron types. Of these seven biophysical features, R_in and τ_m were significantly higher in Sustained neurons than in Onset neurons, whereas six parameters differed between Delayed neurons and each pLSO subtype (rheobase, V_rest, R_in, τ_m, AP halfwidth, AP threshold; Supplementary Figures S6C–G,I,J). The AP amplitude did not differ significantly between the three groups (Supplementary Figures S6G,H). In a last step, we assessed six synaptic parameters, namely the peak amplitude, the rise time, and the τ_decay of sEPSCs and sIPSCs. Whereas none of the six parameters was statistically distinguishable between Onset and Sustained neurons, Delayed neurons differed significantly in 9/12 cases when compared to the two subtypes of pLSOs (Supplementary Figure S7).

Overall, the electrophysiological results, together with our findings on DEGs, provide sufficient evidence to conclude that pLSOs belong to cluster 1, whereas LOCs belong to cluster 2. Moreover, the differences between the two subtypes in cluster 1 are too small to warrant a subcategorization into two subclusters.

The above electrophysiological characterization comprised 16 features. To reduce the dimensionality of the multivariate data and to search for potentially novel subclusters present in our electrophysiological data sets, we performed a PCA based on the 16 electrophysiological features. In such analyses, novel subclusters may arise due to a combination of differences that need not be significant, but together reveal distinct cell types. The results were in accordance with—and thus confirmed—the two clusters of LSO neurons identified via mRNA sequencing. The two subtypes of pLSOs (Onset and Sustained) clustered together (Figure 7A). The PCA also showed higher heterogeneity for cluster 2 than for cluster 1 where virtually all neurons were spotted in close vicinity to each other.

PC1 explained almost 50% of the cluster formation (Figure 7A). To determine the contribution of each feature to PC1, we evaluated the loading scores (Figure 7B). Four of the 16 features described time-related PSC characteristics (feature 1, 7, 9, 11). Thus, specific kinetics of both excitatory and inhibitory ligand-gated receptors appear to be a solid physiological means to distinguish pLSOs from LOCs (faster values in pLSOs). Moreover, R_in, feature 2 on the loading score, may also be used to distinguish the two major types of LSO neurons rapidly and easily during electrophysiological characterization (cf. Supplementary Figure 6D). The lowest influence on PC1 was provided by the AP amplitude, followed by V_rest (Figure 7B; Supplementary Figures S6C,H). A vector factor map analysis showed a positive correlation of rheobase, sag amplitude, sIPSC amplitude, and C_m with PC1, whereas all other features were negatively correlated (Figures 7B,C). Interestingly, PC2 was mainly influenced by the AP amplitude and V_rest, with both vectors being almost antiparallel to one another.

To assess a linkage of the three types of firing pattern with the gene expression pattern, we reused the tSNE plot from Figure 3B and mapped the data from the electrophysiological clusters onto the corresponding neurons (Figure 7D). Remarkably, there was absolute congruency between the transcriptional cluster assignment and the electrophysiological parameters. Hence, we feel safe to conclude that cluster 1 represents pLSOs and cluster 2 LOCs. Within the pLSOs, we identified two subtypes by their biophysical differences.

Auditory brainstem neurons possess a plethora of morphological and functional features as well as molecular specializations that enable them to reliably process information in an ultrafast and temporally precise way (Oertel, 1997; Trussell, 1997, 1999; Golding and Oertel, 2012). Among morphological features are the well-known presynaptic terminals of the calyx-or endbulb-type (Held, 1893). Functional features include AP firing behavior and synaptic fidelity. Molecular specializations comprise an elaborate repertoire of voltage-gated ion channels (Hardman and Forsythe, 2009; Oertel, 2009; Reijntjes and Pyott, 2016). Of the 269 mouse genes for ion channels listed in the International Union of Basic and Clinical Pharmacology/British Pharmacological Society Guide to Pharmacology (IUPHAR/BPS GTP) (Alexander et al., 2021), our sample expressed 42% (114/269; Figure 8A; Supplementary Table S6 for cohort-related results). More specifically, 40% of voltage-gated ion channel genes were expressed (58/144; Figure 8A). For ligand-gated and other ion channels, values were 50 and 36%, respectively (38/76 and 18/49). Dividing the category “Voltage-gated ion channels” into eight subdivisions revealed expression of 43% of the potassium channels, by far the highest number (33/77; Figure 8B). Together, the results illustrate a broad repertoire of ion channels in LSO neurons, particularly potassium channels.

As potassium channels are crucial determinants of multiple physiological properties (V_rest, AP threshold, firing rate, AP duration, postsynaptic potential duration, timing), and because of the immense interest auditory brainstem neurons have received concerning those channels, we focused on this subcategory (Figure 8C). Among 33 expressed K⁺ channel genes, 18 encode pore-forming Kv αsubunits (18/40 = 45%), six Ca²⁺- or Na⁺-activated (KCa or KNa) channels (6/8 = 75%), five two-pore domain K2P channels (5/14 = 36%), and four inwardly rectifying Kir channels (4/15 = 27%).

Transcript levels for the 18 Kv α subunits and the spread across single neurons are shown in Figure 9A and Supplementary Table S6. The panel also depicts expression for nine genes encoding “Potassium voltage-gated channel regulatory subunits” (of 14 listed,⁵ Gray et al., 2015). Overall, the LSO neurons expressed a high number of Kv genes (50%; 27/54). For five genes, the mean TPM values exceeded 10 (i.e., log₂[TPM + 1] > 3.5) in at least one cluster (white-magenta heat map in Figure 9A; Supplementary Table S6). These genes were: pLSOs: Kcnab3|Kvβ3, Kcnq2|Kv7.2, Kcna1|Kv1.1, Kcnh7|Kv11.3; LOCs: Kcnq2|Kv7.2, Kcnab1|Kvβ1. Across all LSO neurons, the by far highest TPM values were seen for Kcnq2 and Kcnab3 (mean: 166 and 141). Four Kv genes qualified as DEGs in pLSOs (marked by asterisks in Figure 9A; pLSOs: Kcna1|Kv1.1, DEG#92; Kcnh7|Kv11.3, DEG#6; Kcnab3|Kvβ3, DEG#146; Kcnip1|KChIP1, DEG#67). We also determined the percentage of neurons expressing a given gene (white-green heat map in Figure 9A). Kcna1|Kv1.1 and Kcnh7|Kv11.3 were expressed by ~70% of cluster 1 neurons. For cluster 2, the percentage was relatively low: ~40% of the neurons expressed Dpp6|DPL1, the highest value observed (Figure 9A). Cluster 1 neurons displayed similarly high values for this gene.

We wondered how many Kv genes a single LSO neuron may express and assessed the heat maps vertically (threshold ≥1 TPM). On average, a single neuron expressed 17% of the Kv genes expressed in the whole sample (4.6/27; range: 0–11; Figure 9B). More specifically, values for the four cohorts were: 5.4 Onset; 5.6 Sustained; 3.0 LOCs (= cluster 2); 5.4 pLSOs (= cluster 1). Thus, pLSOs appear to express a 1.8-fold higher number of Kv genes than LOCs, implying higher complexity in the ascending LSO branch for this ion channel family. Taken together, pLSOs and LOCs show a broad and complex repertoire of Kv transcripts for α and regulatory subunits. Even within a cohort, molecular individuality seems to be a prominent feature. It appears that the complexity of the Kv channel profile in pLSOs is higher than in LOCs. Fifteen percent of the 27 Kv channel genes are DEGs, with no exception in pLSOs, emphasizing the importance of Kv channels particularly in this neuron type. To validate the gene expression results, we performed immunohistochemistry for four Kv subunits (next chapter). In one case, we also performed a series of pharmacological experiments.

Kv7.2 (aka KCNQ2) immunoreactivity was detectable in all major SOC nuclei, and labeling was predominantly somatic (Figures 10A1,A2). GlyT2 colabeling helped to identify pLSOs by their soma shape and the punctate perisomatic and peridendritic pattern described before (Figure 10A3; Friauf et al., 2019). Kv7.2 immunosignals were strong in the cytoplasm, whereas a neurons’ nucleus was virtually immunonegative. We also analyzed Kv7.3 immunoreactivity (Figure 10B). The overall pattern resembled that seen for Kv7.2, except that labeling was more prominent in the nucleus than in the cytoplasm (Figure 10B3). For both Kv7 subunits, we observed no co-localization with GlyT2 in the LSO (Figures 10A2,A3,B2,B3), indicating that Kv7.2 and Kv7.3 are not present in glycinergic axon terminals. In contrast, immunosignals for Kv7.2 and Kv7.3 co-localized with vGLUT1-immunopositive axon terminals contacting LSO neurons at the soma and in the neuropil (Figures 10C,D). Together, these results suggest that the two Kv7 subunits are specific for glutamatergic synapses and absent from glycinergic terminals. This difference between the ipsilateral excitatory and the contralateral inhibitory input may be of physiological relevance for transmitter release in the context of sound localization and should be addressed in future studies.

To address functional Kv7.2/7.3 channels in pLSOs, we performed pharmacological experiments in which retigabine was applied while assessing biophysical properties from pLSOs. The drug is a Kv7.2/3 agonist that shifts channel activation toward more negative potentials (Main et al., 2000; Wickenden et al., 2000). In both pLSO subtypes, Onset and Sustained, retigabine hyperpolarized V_rest by ~4 mV (Figures 11A1,A2,B1,B2; Supplementary Table S5.2). Moreover, it increased the rheobase by ~30%, thus reducing multiple-firing behavior (Figures 11A1,A3,A4,B1,B3,B4). In summary, the immunohistochemical and pharmacological analyses validated our patch-seq data and obtained some interesting results concerning transmitter-type association. To our knowledge, this study is the first detailed report on Kv7.2 and Kv7.3 proteins in the LSO (cf. Cooper et al., 2001; for Kv7.5, see Caminos et al., 2007).

We also validated expression for Kcnh7|Kv11.3 (DEG#6). It was expressed by ~70% of the pLSOs, whereas LOCs completely lacked expression (cf. Figure 9A). Double-immunofluorescent labeling of SOC sections for Kv11.3 and CGRP demonstrated Kv11.3 immunosignals throughout the LSO, whereas CGRP-positive cells were restricted to the center of the LSO, stretching along the tonotopic axis (Figures 12A1,B1,B2). For both proteins, immunosignals were localized in the somata and proximal dendrites (Figures 12A2–A4; see also Figure 5C3). In central aspects of the LSO, Kv11.3-positive somata outnumbered CGRP-positive ones ~2-fold (65% vs. 35%; Figure 12C2). Kv11.3 signals did not overlap with CGRP signals (Figures 12C1,C2), confirming the mRNA results and demonstrating exquisite cell-type specificity.

We finally performed immunohistochemical validation experiments for Kcnab3|Kvβ3 (DEG#146 in pLSOs). Kvβ3 immunolabeling was heavy in MNTB, SPON and LSO (Figure 12D1). The somata of MNTB neurons displayed intense immunosignals in the cytoplasm, but not in the nuclei. An opposite pattern occurred in LSO and SPON neurons (Figures 12D2–D7). LSO and SPON neurons displayed perisomatic labeling, suggesting the presence of Kvβ3-positive MNTB axon terminals. Indeed, double labeling demonstrated co-distribution of perisomatic Kvβ3 signals with GlyT2 signals, confirming presynaptic Kvβ3 in inhibitory synapses.

Besides the Kv channels, three other classes of potassium channels are known (Figure 8C). They are encoded by 37 genes, and we found expression for 15 of them (Supplementary Table S6). In LOCs, Kcnn3|KCa2.3 and Kcnj3|Kir3.1 expression was high and moderate, respectively (427 vs. 30 TPM). pLSOs showed a high expression for Kcnn1|KCa2.1 and a moderate one for Kcnj3|Kir3.1 (134 vs. 18 TPM). None of the genes was a DEG. KCa2.1 and KCa2.3 form small-conductance, Ca²⁺-activated K⁺ channels (SK channels) that have high Ca²⁺ affinity and mediate the SK currents (Bond et al., 2005; Faber, 2009; Adelman et al., 2012). They are a major contributor to the afterhyperpolarization phase of APs and regulate AP firing, including AP precision (Engel et al., 1999; Hallworth et al., 2003). Their Ca²⁺ sensitivity is provided by CaM (Saimi and Kung, 2002), for which our study shows the third highest expression level among 11,659 genes (Figure 2B; Supplementary Table S3.2). SK2 channels in outer hair cells are activated via Ca²⁺ influx through ionotropic ACh receptors activated by cholinergic MOC neurons, ultimately resulting in hyperpolarization, and turning down the cochlear amplifier (Housley and Ashmore, 1991; Guinan, 2011; Maison et al., 2013). To our knowledge, nothing is known about the role of KCa2 channels in auditory brainstem neurons.

Nav channels play a fundamental role in generating, shaping, and conducting APs (Catterall et al., 2005; de Lera Ruiz and Kraus, 2015; Hull and Isom, 2018). LSO neurons in total expressed nine Nav genes (Supplementary Figure S8A; Supplementary Table S6), of which five encoded α subunits and four regulatory β subunits (13 genes are listed altogether: Nav1.1—Nav1.9 plus Navβ1—Navβ4). Thus, 69% (9/13) of Nav genes were expressed in our sample. Mean expression levels of ≥10 TPM in at least one cluster were seen for four genes (pLSOs: Scn1a|Nav1.1, Scn8a|Nav1.6, Scn1b|Navβ1; LOCs: Scn1a|Nav1.1, Scn3a|Nav1.3, Scn8a|Nav1.6, Scn1b|Navβ1). Across all neurons, highest expression occurred for Scn1b|Navβ1, followed by Scn8a|Nav1.6, and Scn1a|Nav1.1 (mean TPM: 92, 47, 16). Three genes qualified as DEGs (pLSOs: Scn4b|Navβ4, DEG#173; LOCs: Scn2a|Nav1.2, DEG#49; Scn3a|Nav1.3, DEG#11). Scn4b transcripts were found in 25% of pLSOs and in not at all in LOCs. Scn2a was expressed by ~60% of LOCs and ~ 20% of pLSOs. Likewise, Scn3a was expressed by ~70% of LOCs and 20% of pLSOs.

At the single cell level, an average neuron expressed 33% (3.0/9) Nav genes displayed in the whole sample (range: 0–7; Supplementary Figure S8B). Corresponding values for the four cohorts were: 2.9 Onset; 2.9 Sustained; 3.2 LOCs [= cluster 2]; 2.9 pLSOs [= cluster 1]. Thus, ~3 Nav genes appear to be expressed per LSO neuron, regardless of the subtype, indicating similar numerical complexity. Nevertheless, the fact that 33% of the genes are DEGs, together with the broad percent range, implies a considerable amount of molecular individuality at the cellular level.

Of 10 listed genes encoding Cav channel α subunits, six were expressed by the LSO neurons (Supplementary Figure S9A; Supplementary Table S6). In the group of regulatory subunits (α2δ, β, γ), we found expression for 12/16 genes. Collectively, LSO neurons expressed ~70% of the listed Cav genes (18/26). Therefore, gene expression for Cav channels was more elaborate than for Kv channels, but very similar to Nav channels (Kv: 50%; Nav: 69%). In general, expression levels were also considerably higher than for Kv or Nav subunits. Across all LSO neurons, mean TPM values exceeded 200 for Cav2.3, Cav2.2, and Cavβ4 (639, 372, 218), values not seen for Kv and Nav channels. Whereas mean TPM values ≥10 in at least one cluster occurred for five Kv and four Nav genes, 11 Cav genes fulfilled these criteria. On the α subunit side, these were Cacna1b|Cav2.2, Cacna1d|Cav1.3, Cacna1e|Cav2.3, and Cacna1g|Cav3.1. Across all LSO neurons, by far the highest mean TPM value was observed for Cav2.3, followed by Cav2.2 (1,491 and 372). For α2δ subunits, genes with ≥10 TPM were Cacna2d2|Cavα2δ2 and Cacna2d3|Cavα2δ3. For the β subunits, we identified Cacnb1|Cavβ1, Cacnb3|Cavβ3 and Cacnb4|Cavβ4, and for the γ subunits, we found Cacng2|Cavγ2 and Cacng4|Cavγ4. Only two Cav genes were DEGs, one gene per cluster (LOCs: Cacna1e|Cav2.3, DEG#26; pLSOs: Cacng5|Cavγ5, DEG#12). Concerning the percentage of cells, Cav2.3 was expressed by ~70% of the LOC cohort and Cavγ5 by ~65% of the pLSO cohort.

An average LSO neuron expressed 25% Cav genes (4.5/18; α plus regulatory subunits; range: 1–9; Supplementary Figure S9B). Corresponding values for the four cohorts were: 4.5 Onset; 5.4 Sustained; 4.3 LOCs [= cluster 2]; 4.7 pLSOs [= cluster 1]. The numbers indicate a similar complexity across clusters. In summary, a large fraction of Cav genes is expressed in our LSO sample (~70%). The expression profile differs substantially between pLSOs and LOCs as well as between the two pLSO subtypes. However, the similar number of expressed genes per neuron indicates similar complexity across cohorts.

Mammalian TRP channels form a superfamily comprising six families (Venkatachalam and Montell, 2007; Talavera et al., 2008). The patch-seq LSO neurons expressed ~13 of the listed TRP genes (9/28; Supplementary Figure S10A; Supplementary Table S6). Five of them displayed a mean of ≥10 TPM in at least one cluster, namely Mcoln1|TRPML1, Pkd2l2|TRPP5, Trpm2|TRPM2, Trpm7|TRPM7, and Trpv2|TRPV2. For three genes, TPM values were ≥ 100 in at least one cohort (TRPML1, TRPP5, TRPM2). The highest level (374 TPM) was observed for TRPML1 in Sustained neurons. None of the TRP genes was a DEG, implying only minor differences between both clusters. We found only very low transcript levels, if any, for TRPC channels. This slightly contrasts with earlier findings of mGluR-activated TRP-like channels that mediate Ca²⁺ influx in the developing LSO upon high-frequency stimulation (Ene et al., 2003, 2007). The authors concluded that these Ca²⁺ channels are likely TRPC channels.

An average LSO neuron expressed 1.4 TRP genes (5% of 28 listed; range: 0–5; Supplementary Figure S10B). Corresponding values for the four cohorts were: 1.2 Onset; 1.8 Sustained; 1.6 LOCs [= cluster 2]; 1.3 pLSOs [= cluster 1]. In summary, ~1/3 of all TRP genes (9/28) were expressed in our sample. There is a remarkable similarity between the two pLSO subgroups which was not evident for Kv, Nav, or Cav channels.

CNR channels translate concentration changes of cyclic nucleotides into electrical signals (Craven and Zagotta, 2006; Napolitano et al., 2021). They are closely homologous to Kv channels and form two families, HCN channels and cyclic nucleotide-gated (CNG) channels (Craven and Zagotta, 2006). In contrast to Kv, Nav or Cav channels, CNG channels are poorly ion selective and only weakly voltage dependent. CNG channels are tetramers formed by CNGA and CNGB subunits, whereas HCN channels are tetramers formed by four subunits (HCN1–4). The latter open with hyperpolarization, are weakly K⁺ selective, and are regulated by cyclic nucleotides, e.g., cAMP. Of 10 Cng/Hcn genes listed in the IUPHAR/BPS GTP (Alexander et al., 2021), LSOs exclusively expressed HCN channels (20%, 2/10; Supplementary Figure S11A; Supplementary Table S6). None qualified for a DEG. Mean expression levels for Hcn2|HCN2 were low (< 5 TPM); only Hcn1|HCN1 displayed a mean of ≥10 TPM. The highest percentage of neurons expressing a HCN gene was 50% in Onset neurons (white-green heatmap). The other cohorts displayed substantially lower values, particularly for HCN2.

An average LSO neuron expressed 20% of the HCN genes (0.4/2; range: 0–2; Supplementary Figure S11B). Corresponding values for the four cohorts were: 0.6 Onset; 0.1 Sustained; 0.3 LOCs; 0.5 pLSOs. Collectively, gene expression for HCN channels was not high in the LSO. However, HCN channel-mediated voltage sags have been described in pLSO—Onset and Sustained neurons (Leao et al., 2006). Our recordings confirm voltage sags in both pLSO types, but we find HCN channel expression only in Onset neurons. It is plausible that mRNA levels for HCN channels (and protein turnover) are too low in Sustained neurons to be detected. This is supported by a study indicating that HCN channels are sequestered within endosomes which can serve as an alternative reservoir for HCN surface expression, independent of their biosynthesis (Hardel et al., 2008).

LSO neurons receive fast synaptic input through multiple neurotransmitter systems targeting ligand-gated ion channels. Our electrophysiological data showed that the kinetics of excitatory and inhibitory sPSCs are solid means to distinguish pLSOs from LOCs (Figures 7B,C). To further assess this difference on the transcript level, we analyzed gene expression for subunits of six major channel types: glutamatergic AMPA, NMDA, kainate, and “orphan” receptors (GluA, GluN, GluK, GluD), glycine receptors (GlyRs), γ-amino-butyric acid receptors (GABAARs), and nicotinic acetylcholine receptors (nAChRs). We found expression for 55% of the receptor subunit genes (32/58 listed in IUPHAR/BPS GTP; Figure 13A). Thirteen (67%) of them code for ionotropic glutamate receptors, three for GlyRs, 11 for GABAARs, and five for nAChR subunits (Figure 13A; Supplementary Table S6; listed are 18, 5, 19, and 16). For 19 genes, the mean expression was ≥10 TPM in at least one cluster (15 in pLSOs, 17 in LOCs).

Across all 86 neurons, the highest expression occurred for Glra1|GlyRα₁ (756 TPM), followed by Gabra1|GABAARα₁, Gabrg2|GABAARγ2, Gria3|GluA3, and Gabrb2|GABAARβ2. In comparison, GluK5, GluN, or nAChR gene expression was substantially lower (90 TPM for GluK5, the highest value). Expression for four glutamate receptor subunits of the GluD and GluK subtype was generally low, and it was absent for 5 of 18 genes listed. Slightly higher expression levels were seen for nAChRs. Five genes turned out to be DEGs (pLSOs: Gria4|GluA4 DEG#40, Grin2b|GluN2B DEG#36, Gabrb3|GABAARβ3 DEG#72; LOCs: Gria1|GluA1 DEG#68, Gabra1|GABAARα1 DEG#89). The diversity emphasizes the importance of diverse ligand-gated ionotropic receptors in LSO neurons and points to specific functions, even within a cluster. This is highlighted by the percentage of pLSOs expressing a DEG:66% for Gria4, 55% for Grin2b, and 54% for Gabrb3 (Figure 13A, white-green heatmap; 23, 17, and 20% for LOCs). Values in LOCs were 73% for Gria1 and 57% for Gabra1 (41 and 34% for pLSOs).

Concerning the question of how many genes encoding ligand-gated receptor subunits are expressed in a single LSO neuron, the mean number was 8.0 (range: 2–14; Figure 13B). Corresponding values for the four cohorts were: 8.2 Onset; 8.2 Sustained; 7.6 LOCs; 8.2 pLSOs, implying similar transcript complexity. Taken together, more than 50% of the subunits in ligand-gated ion channels are detectable at the mRNA level. The high number of DEGs (5 in 58) demonstrates clear differences between pLSOs and LOCs.

G protein-coupled receptors (GPCRs) constitute the largest family of cell-surface proteins involved in signal transmission (Rosenbaum et al., 2009). Upon agonist binding, they activate specific heterotrimeric G proteins which results in the modulation of distinct downstream effector proteins, including ion channels. The receptors are characterized by the presence of seven membrane-spanning segments in a monomer. We focused on three major types of GPCRs being activated by glutamate, GABA, or acetylcholine. For metabotropic glutamate receptors (mGluRs), there are three groups comprising eight members (group I: mGluR1, mGluR5; group II: mGluR2, mGluR3; group III: mGluR4, 6–8). Whereas group I receptors are primarily present at postsynaptic sites, group II and III mGluRs localize predominantly presynaptically. Metabotropic GABA receptors have two members (GABABR1 and GABABR2), whereas metabotropic ACh receptors (mAChRs) have five (M1—M5, encoded by Chrm1—Chrm5).

LSO neurons in general expressed 67% of the named GPCR genes (10/15; Supplementary Figure S12A; Supplementary Table S6). Four genes displayed a mean expression of ≥10 TPM in at least one cluster (GABABR1, mGluR1, mGluR3, mGluR8). By far the highest expression was seen for mGluR3 (all LSO: 521 TPM). In contrast, mGluR4 transcripts were almost absent, and those for mGluR2 and mGluR6 were undetected. Within GABABRs, transcripts prevailed for GABABR1 (25 TPM). Among the mAChRs, expression was absent for three genes and low for M2 and M4. Nevertheless, Chrm2|M2 was DEG#93 for LOCs (the only DEG in the CPCG group; TPM: 5 vs. 0). It was expressed by 27% of LOCs and 2% of pLSOs.

An average LSO neuron expressed 2.0 GPCR genes (range: 0–6; Supplementary Figure S12B). Corresponding values in the cohorts were: 1.5 Onset; 2.3 Sustained; 2.6 LOCs; 1.7 pLSOs. Together, the results demonstrate a higher GPCR complexity in LOCs.

LSO neurons display a high degree of transmitter diversity (cf. Figure 1A). To explore the diversity at the single-cell level, we analyzed gene expression of a subset of neurotransmitter-associated molecules, including neuropeptides to explore neurotransmitter phenotype. Figure 14A and Supplementary Table S6 illustrate seven neurotransmitter-associated genes expressed in the LSO (Slc17a6|vGLUT2; Slc6a5|GlyT2; Gad2|GAD65; Ddc|DOPA decarboxylase = DDC; Chat|choline acetyltransferase = ChAT; Calca|CGRP; Ucn|urocortin). The highest expression was seen for CGRP (LOCs: 1,256 TPM), followed by DDC (LOCs: 231 TPM), GAD65 (LOCs: 87 TPM), and vGLUT2 (pLSOs: 59 TPM). Four genes qualified for DEGs (LOCs: Calca|CGRP DEG#1; Ucn|urocortin DEG#2; Gad2|GAD65 DEG#32; pLSOs: Slc17a6|vGLUT2 DEG#26). Transcripts for vGLUT2, indicative of a glutamatergic phenotype, were found in 66% of the pLSOs, and transcripts indicative of a glycinergic and GABAergic phenotype in 13 and 21%, respectively (Figure 14A, white-green heatmap). Thus, the predominant transmitter in pLSOs seems to be glutamate. The opposite appears to hold for LOCs, as 67% of them showed transcripts for GAD65 and 17% for GlyT2, whereas only 20% were associated with a glutamatergic phenotype (sum >100% because of multi-transmitter neurons; see below). Our results also indicate an excitatory, rather than an inhibitory, phenotype for most Onset neurons (70% vGLUT2, only 5% GlyT2 and 18% GAD65), whereas Sustained neurons appear to substantially utilize inhibitory transmitters (42% GlyT2 plus 33% GABA vs. 50% vGLUT2). Reportedly, membrane properties of excitatory pLSOs favor integrative level coding, whereas those of inhibitory pLSOs favor time coding (Haragopal and Winters, 2023).

As the above results implied multi-transmitter properties, we analyzed our samples in more detail. In total, the 86 LSO neurons displayed 25 combinations of neurotransmitter-associated gene expression (Figure 14B, columns 1–25; see also Supplementary Table S6). Of those, 13 combinations were seen in pLSOs and 15 combinations in LOCs. A single transmitter phenotype was obvious in 56% of the pLSOs (columns 1–4). Forty-three percent of the pLSOs appeared to be purely glutamatergic (column 1), whereas 23% co-expressed one to two neurotransmitter-associated genes in addition to vGLUT2 (columns 5–8, 12, 13). Thus, glutamate was by far the predominant transmitter type in pLSOs (66%; see Set size diagram). The remaining 34% were vGLUT2 negative (columns 2–4, 9, 10, 15). Seven pLSOs (12.5%) expressed the CGRP gene, which was a bit surprising because the neuropeptide is usually affiliated with LOCs. Transcripts for ChAT and urocortin were completely absent from the pLSO cohort. The maximum number of neurotransmitter-associated genes expressed in an individual pLSO was three (columns 12, 13, 15; 9% of the neurons).

LOCs displayed a substantially higher degree of multi-transmitter properties than pLSOs. Seventeen percent of them co-expressed five genes (Figure 14B, columns 23–25). Forty percent of the LOCs co-expressed >3 genes (columns 19–25), whereas no pLSO did so. On the other side, only 7% of the LOCs displayed a single-transmitter phenotype (Figure 14B, columns 1, 2; vs. 56% pLSOs). The dominant subtype of LOCs (23%) co-expressed genes coding for GABA, CGRP, and urocortin (column 18). Across all combinations, 93 and 83% of the LOCs showed transcripts for CGRP and urocortin, respectively (see Set size diagram). This pair of neuropeptides is clearly a hallmark of LOCs. Transcripts for GAD65 occurred in 67% of the LOCs. They were always associated with co-expression, in 10 neurons for three to four additional genes (columns 20–22, 24, 25). Thirty percent of LOCs expressed Ddc|DDC, indicative of a dopaminergic phenotype (vs. 6% of pLSOs). Ddc expression was always associated with three to four other transcripts, in each case for CGRP and urocortin. At a frequency of 17% (3 neurons), LOCs expressed Chat, a surprisingly low number (see Discussion for explanation). Chat expression was invariably flanked by CGRP transcripts and in some cases also with GAD65 (columns 17, 24, 25) and/or urocortin transcripts (columns 23–25).

An average LSO neuron expressed 1.9 neurotransmitter-associated genes (range: 0–5; Supplementary Table S6). Corresponding values for the four cohorts were: 1.1 pLSO—Onset; 1.6 pLSO—Sustained; 3.3 LOCs; 1.2 pLSOs. The almost 3-fold higher value in LOCs compared to pLSOs further highlights the complex transmitter repertoire in the LOC system. Collectively, the results on neurotransmitters add further facets to the overall picture that pLSOs and LOCs differ considerably in their transmitter repertoire. Altogether, the UpSet plot in Figure 14B shows that only 3 of the 25 combinations are shared between the two clusters (columns 1, 2, 10). This implies a very clear distinction between pLSOs and LOCs concerning their transmitter repertoire. pLSOs are characteristic in that they express mainly glutamatergic or GABAergic genes. In contrast, LOCs display multi-transmitter properties, most frequently GABA plus two neuropeptides.

In a final analysis, we assessed “Super DEGs” and “Cluster similarity” (details in Methods). We restricted the analysis to the field of neurotransmission, namely voltage-gated ion channels, ligand-gated ion channels, GPCRs, and neurotransmitter-associated proteins (Figure 15). The input comprised 114 genes, of which 18 were DEGs. About 10% of the genes (12/114) were Super DEGs (6 in each cluster; Figure 15A1). Five of them encode subunits of voltage-gated ion channels (pLSOs: Kv11.3, Kvβ3, Kv1.1; LOCs: Nav1.3, Cav2.3), four encode subunits of ligand-gated ion channels (pLSOs: GluA4, GABAARβ3; LOCs: GABAARα1, GluA1), and three genes encode neurotransmitter-related proteins (pLSOs: vGLUT2; LOCs: CGRP, urocortin). No Super DEG was revealed for GPCRs.

Cluster similarity was found for ~25% of the genes (27/114; Figure 15A2). We further quantified this finding by determining an overlapping index (OI) (Pastore and Calcagni, 2019). OI values ranged from 0.95 to 0.2 and allowed to rank the genes. When the stringency was increased (by setting the OI threshold to 0.8), 12 genes remained. Six of them encode subunits of voltage-gated ion channels (Cav1.3, Cavγ2, TRPML1, TRPM2, TRPM7, TRPP5), four encode subunits of ligand-gated ion channels (GlyRβ, GluA2, GluK5, GluN1), two encode GPCRs (GABABR1, mGluR8), and none encodes neurotransmitter-related proteins.

Taken together, ~80% (90/114) of the genes involved in neurotransmission are neither Super DEGs nor do they fulfill the stringent criteria for Cluster similarity. The remaining 24 genes categorize equally to Super DEGs and Cluster similarity (12 + 12). Moreover, the 12 Super DEGs are equally split between pLSOs and LOCs. Thus, we crystallized three manageable sets of neurotransmitter-associated genes whose expression pattern differs between pLSOs and LOCs or, on the other hand, is closely similar.

The three most highly expressed genes showed expression levels of ~25,000 TPM and more, >2-fold higher than the fourth gene (Figure 2). The related proteins are brain-type creatine kinase B (CKB), SNRPN upstream open reading frame protein (SNURF; SNRPN = small nuclear ribonucleoprotein polypeptide N), and calmodulin 1 (CALM1 also known as CaM).

CKB is involved in energy homeostasis. The cytosolic enzyme transfers phosphate between ATP and various phosphogens, for example creatine, thus generating phosphocreatine and ADP and establishing a readily available ATP-buffering system. It is not surprising that Ckb expression correlates very well with gene expression for Na⁺/K⁺-ATPase α1 and α3 subunits (Jiao et al., 2021; NB: α3 is exclusive to neurons). The CKB-mediated buffer system appears to play a pivotal role in the hair bundle of vestibular hair cells (Shin et al., 2007). Reduced CKB activity is associated with hearing impairment in mice with Huntington disease (Lin et al., 2011). Interactions of CKB with the potassium-chloride transporter KCC2 have been demonstrated (Inoue et al., 2004, 2006; Li et al., 2007; Salin-Cantegrel et al., 2008). The interactions facilitate KCC2-mediated Cl-extrusion from neurons and maintain the low intracellular Cl-concentration required for effective inhibitory neurotransmission, i.e., hyperpolarization (for LSO, Balakrishnan et al., 2003). As robust inhibitory input to LSO neurons is crucial for sound localization, high CKB abundance is not surprising. Indeed, high Ckb expression in the LSO is also shown in the AMBA (Figure 2D), thus validating our data.

SNURF is a 71 amino acid-long polypeptide encoded by the two-gene locus “snurf/snrpn.” This locus also encodes the small nuclear ribonucleoprotein (associated) polypeptide N (SNRPN) (Ozçelik et al., 1992). Across seven tissues, SNRPN is most abundant in the brain (rat; McAllister et al., 1988). Extensive alternative splicing and promoter usage occurs for SNURF and SNRPN, and the full-length nature of several transcripts remains to be determined. Like other imprinted genes, the SNURF/SNRPN domain is involved in neurodevelopment (Ribeiro Ferreira et al., 2019). Developmentally related changes of DNA methylation have been described for the mouse gene (Miyazaki et al., 2009). Loss of imprinted gene expression leads to disorders, e.g., severe intellectual disability (Nicholls et al., 1998). Presumably, SNURF plays a role in pre-mRNA splicing, and in line with this, SNURF is localized to the cell’s nucleus. Despite decades of research, the biological role of SNURF is not fully understood (Basak and Basak, 2022). A Pubmed search (“snurf brain” and “snurf neuron”; 2023–10-22) revealed 14 and nine results, respectively. None was associated with papers on the auditory system, the brainstem, or the cortex. Taken together, our study discovered Snurf as a novel gene that probably plays an important role in the CAS. The intermediate expression level for Snurf in the LSO and other auditory brainstem regions, as demonstrated in the AMBA, strengthens our conclusion.

CaM is one of three calmodulin proteins in the family of EF-hand Ca²⁺-binding proteins. These ubiquitous proteins exhibit a high degree of evolutionary conservation (Wren et al., 2019). CaM-Ca²⁺ complexes mediate the control of many structurally different proteins, such as voltage-gated ion channels (e.g., Cav1.3: Johny et al., 2013; Cav2.1: Lee et al., 2002; Jiang et al., 2008; KCa2: Bildl et al., 2004; Kv7.1: Sachyani et al., 2014; Kv7.2: Alaimo and Villarroel, 2018; Nav1.5: Gabelli et al., 2014). CaM appears to be the most relevant transducer of Ca²⁺ signals (Levitan, 1999; Faas et al., 2011). Moreover, irrespective of the intracellular Ca²⁺ concentration, CaM is constitutively tethered to Kv7.2/3 channels and presumably plays a key role in Kv7 channel folding and trafficking (Wen and Levitan, 2002; review: Soldovieri et al., 2011). Notably, the present study demonstrates substantial gene expression for Kv7.2 and Kv7.3 (Figures 9, 10), pointing to such an interaction.

In the CAS, mostly analyzed at calyx of Held synapses, CaM controls several forms of synaptic plasticity (Nakamura et al., 2008; Lee et al., 2010; Lipstein et al., 2013). Calm1 expression increases significantly upon tone rearing in auditory cortex interneurons, implying a role in experience-and activity-dependent circuit refinement during the tonotopic critical period (Kalish et al., 2020). We are not aware of a CaM study in the LSO. Based on our results, we suggest investigating interactions between CaM and Kv7.2/3 channels in pLSOs. Consistent with our findings, the AMBA shows intermediate to high expression levels in the LSO (Figure 2D).

Besides Ckb, 32 other genes among the top 500 were involved in ATP synthesis (Figure 2). The prevalence of transcripts for catabolic pathways is consistent with classic reports of high energy consumption in auditory brainstem neurons that fire APs at particularly high rates (e.g., Sokoloff, 1981). Previous reports on the LSO have focused on a few metabolic proteins (Trattner et al., 2013; Brosel et al., 2018). Our study is the first to comprehensively address LSO neurons and link ATP demand to a variety of specific enzymes. In a recent scRNA-seq study in the mouse cochlea, 33 genes among the top 500 in class I spiral ganglion neurons were associated with energy metabolism (Petitpré et al., 2018). Nine of them overlapped with the 33 genes detected in LSO neurons, all of them in the OxPhos group (Ndufa11, Ndufs2, Ndufs3, Sdhb, Sdhd, Cox5a, Atp5b, Atp5f1, Atp5g3). The overlap of ~30% indicates similarity between the two groups of auditory neurons, but also considerable dissimilarity, as previously shown for mitochondrial molecules (Fecher et al., 2019). In this lineage, only 3 of the 33 genes were DEGs (cluster 1: Pfkp, DEG#21, Ckb, DEG#254; cluster 2: Acly, DEG#10), suggesting a similar efficiency of ATP synthesis in pLSOs and LOCs. An attractive hypothesis is that ATP may also be used as an extracellular messenger by maturing LSO neurons (Dietz et al., 2012; Babola et al., 2020).

There is considerable controversy over whether neurons use lactate—rather than glucose—to fuel oxidative phosphorylation (Pellerin and Magistretti, 2012; Nortley and Attwell, 2017; Li and Sheng, 2021). Lactate would be released from glial cells via the astrocyte-neuron lactate shuttle. Neuronal computation is energetically expensive, with most of the energy used to reverse ion entry that has generated synaptic potentials and APs (Harris et al., 2012; Meyer et al., 2022). LSO neurons generate APs at unusually high rates of several hundred Hertz (Brosel et al., 2018). We examined transcripts for key enzymes and obtained evidence that LSO neurons do not primarily use lactate as a carbon source for ATP generation. The expression of Slc16A2, which encodes the neuronal monocarboxylate transporter MCT2, was absent from our sample. In contrast, the expression of Slc2a3, which encodes the neuronal glucose transporter GLUT3, was high (gene #201, 955 TPM). Expression levels of other genes (neuronal Ldha|lactate dehydrogenase A, #5,658, 4 TPM; astrocytic Glo1|lactoylglutahione lyase, #1,521, 82 TPM; Glo2 = Hagh|hydroxyacylglutathione hydrolase, #119, 1,225 TPM; astrocytic Slc16A1|MCT1, #6,252, 3 TPM) confirm our conclusion that lactate-fueled OxPhos is not the major ATP producer in LSO neurons. This does not mean that LSO neurons exclusively utilize glucose (Brosel et al., 2018), and developmental changes may also be considered (Lujan et al., 2021). Collectively, our results support the conclusion of an extraordinary ATP demand in LSO neurons and suggest glucose—rather than lactate—as the primary energy source.

Our patch-seq LSO neurons clearly separated into two clusters, with 65% belonging to pLSOs and 35% to LOCs. Literature data suggest that pLSOs and LOCs together comprise ~90% of all LSO neurons (~70% + ~20%). In addition to pLSOs and LOCs, several other neuron types have been described morphologically. No categorization data are available for mice, but five types have been described in gerbils and cats and seven in rats (Helfert and Schwartz, 1987; Rietzel and Friauf, 1998). A prominent subtype are the LSO marginal cells, so named because of the location of the somata along the edges of the LSO (review: Friauf et al., 2019). Marginal cells appear to comprise only ~5% of the population, regardless of the species analyzed (Ollo and Schwartz, 1979; Helfert and Schwartz, 1986, 1987; Rietzel and Friauf, 1998). Small neurons tend to be underrepresented in patch-clamp recordings, and we did not specifically target small somata or for somata located in the marginal zones of the LSO. Nevertheless, we are confident that we did not miss any major cell type.

Based on our results, we also find it unlikely that pLSOs and LOCs form further subclusters, although we did find differences in membrane properties between pLSOs —Onset and Sustained neurons. Interestingly, Onset and Sustained neurons have been described in both vGLUT2-positive and vGLUT2-negative pLSOs (they are presumably excitatory and inhibitory, respectively; Haragopal and Winters, 2023). This complicates the identification of distinct expression profiles for excitatory and inhibitory pLSO subclusters. Nevertheless, activity-dependent maturation could manifest differences in the expression profile and result in different membrane properties between Onset and Sustained pLSOs.

Kcnh7|Kv11.3 was DEG#6 in pLSOs (Figure 9A). The gene encodes ERG3, a pore-forming subunit of ether-a-gogo-related channels. ERG channels have a predominantly somatic location, are most active after intense electrical activity, and effectively modulate neuronal excitability, thereby complementing Kv1 channels (Hardman and Forsythe, 2009). The channels are inward rectifiers with unusual gating kinetics. They open at relatively negative membrane potentials, thus affecting the depolarization phase of an AP, and they inactivate very rapidly. The highest conductance is generated during repolarization (Shepard et al., 2007; Bauer and Schwarz, 2018), when the fast recovery from inactivation takes place (Bauer and Schwarz, 2018). Several studies showed a role of ERG channels in the nervous system (Sacco et al., 2003; Niculescu et al., 2013). In ventral cochlear nucleus neurons, they play a role in setting AP threshold and AP frequency (Yildirim and Bal, 2018). In the mouse MNTB, where ERG1 and ERG3 subunits are predominant, ERG currents modulate Kv1 currents and limit AP generation around AP threshold voltages (Hardman and Forsythe, 2009).

Kcna1|Kv1.1 was DEG#92 in pLSOs (Figure 9A). Kv α subunits have received extensive attention in the auditory brainstem, and immunoreactivity has been demonstrated in several species in somata, dendrites, and axon terminals (mouse: Wang et al., 1994; Brew et al., 2003; rat: Barnes-Davies et al., 2004; gerbil: Nabel et al., 2019); three bat species: (Rosenberger et al., 2003; Pätz et al., 2022). Kv1.1 channels were most intensively analyzed. They carry a rapidly activated, low-threshold, and sustained K⁺ current (I_KL) that determines the 1st AP latency as well as the AP waveform (Kole et al., 2007; Fischl et al., 2016). MNTB neurons in Kv1.1 knockout mice display a reduced I_KL and are hyperexcitable (Brew et al., 2003). In line with this, pharmacological blockade of Kv1.1 channels in pLSOs with α-dendrotoxin (NB: blocks also Kv1.2 and Kv1.6; Oak and Yi, 2014) convert a single-firing to a multiple-firing pattern (Barnes-Davies et al., 2004). Kv1.1 knockout mice display behavioral deficits in sound localization (Robbins and Tempel, 2012; Karcz et al., 2015), and their LSO neurons show a weaker capability of encoding interaural level differences (Karcz et al., 2011). pLSOs clearly differ from LOCs regarding Kv1.1 channels. The results provide further evidence for the importance of temporal integration of excitation and inhibition during sound localization, especially in the ascending branch of the LSO.

The Kvβ3-encoding gene Kcnab3 was DEG#146 in pLSOs. Three Kvβ isoforms have been described in the brain (Heinemann et al., 1995; Rhodes et al., 1996). They have oxidoreductase activity and interact with the pore-forming α subunits of Kv1 and Kv4 channels (Dwenger et al., 2022). The assembly generally increases whole cell Kv current and modifies voltage sensitivity (Pongs and Schwarz, 2010). We are not aware of a study on Kvβ3 in the auditory system. Our results show high protein abundance in the major SOC nuclei and co-distribution of Kvβ3 and GlyT2 in axon terminals terminating on pLSO somata, most likely originating from MNTB neurons (Figure 12D).

Slc17a6|vGLUT2 was DEG#26 in pLSOs. Vesicular glutamate transporters are involved in the storage and release of glutamate. Our patch-seq results for genes expressing vGLUT confirm in situ hybridization results showing the absence of vGLUT1 and the presence of vGLUT2 in mouse LSO somata (Ito et al., 2011). They are also consistent with the detection of vGLUT2 in the LSO of rat, mouse, and owl monkey at the protein level (Blaesse et al., 2005; Haragopal and Winters, 2023) and at the mRNA level (Hackett et al., 2011). The association of vGLUT2 with pLSOs is consistent with findings in rats that vGLUT2-positive LSO neurons project into the contralateral inferior colliculus (IC; virtually no ipsilateral projection), thus representing ascending neurons (Ito and Oliver, 2010). Moreover, concerning the three classical fast neurotransmitters (glutamate, glycine, GABA; expressed by 44 pLSOs in different combinations), we show a surprisingly high 4:1 ratio between a purely excitatory and a purely inhibitory phenotype (64% vs. 16%; Figure 16; Supplementary Table S6). The remaining 20% show a mixed phenotype (predominantly glutamate plus GABA; notice rounding effect). Thus, the present results indicate a dominant glutamatergic phenotype in mouse pLSOs, whereas glycinergic and/or GABAergic neurons are much less abundant. A mixed phenotype (glutamate plus glycine) has also been described for the projections of LSO neurons into the IC of rats (Fredrich et al., 2009). The authors found that all ipsilaterally projecting neurons were glycinergic, with 60% of them also staining for glutamate. Similar to mice, ¾ of the LSO neurons in gerbils are excitatory (Mellott et al., 2022). In contrast, excitatory and inhibitory LSO neurons appear to be balanced in cats (Glendenning et al., 1992) and guinea pigs (Helfert et al., 1989). We propose that species differences underlie the different ratios.

Gria4|GluA4 was DEG#40 in pLSOs. In contrast, GluA1 subunits appear to be a marker for LOCs (DEG#68). Kinetics of glutamatergic ionotropic receptors can be in the submillisecond range (Geiger et al., 1997), and the fastest kinetics are obtained when GluA4 subunits are present (auditory brainstem: Geiger et al., 1995; Yang et al., 2011; Rubio et al., 2017). By contrast, GluA1-containing receptors are slow-gating. At calyx of Held synapses, GluA4 is the main determinant for the fast synaptic responses and indispensable for driving high-fidelity neurotransmission and conveying precise temporal information (Yang et al., 2011). GluA4 is also expressed in auditory nerve synapses (Rubio and Wenthold, 1997), and its loss impairs the acoustic startle response (García-Hernández and Rubio, 2021). These results emphasize the importance of this subunit in the CAS.

AMPAR-mediated EPSCs in the LSO are characterized by sub-millisecond decay times (0.77 ms) that are slowed upon acoustic trauma, which is mirrored by mRNA decreases for GluA4, yet increases for GluA1 (Pilati et al., 2016). In this line, our transcriptomic results imply different kinetics at glutamate receptors between pLSOs and LOCs. Indeed, our electrophysiological findings demonstrate considerably faster EPSC kinetics in pLSOs (Supplementary Figures S7D,E; Supplementary Table S5.1; Figure 16). The involvement of pLSOs in sound localization requires sub-millisecond precision in signal integration. According to our data, hyper-precise kinetics are not present in LOCs, which seem to protect the auditory nerve fibers from excitotoxic effects of acoustic overexposure, a process occurring in minutes (Accardi, 2014). Notably, pLSOs have myelinated axons, whereas LOC axons are unmyelinated (Brown, 2011).

The sixth Super DEG in pLSOs, Gabrb3, codes for GABAARβ3, the β3 subunit of GABA_A receptors (DEG#72). GABA_A receptors are heteropentamers, most often consisting of subunits from three different families in a 2:2:1 fashion (2α + 2β + 1 of either γ, δ, ε, π). The most common arrangement is 2α₁2β₂γ₂. Intermediate levels of Gabrb3 mRNA were identified in the LSO of rats and mice (Campos et al., 2001; Fischer et al., 2019), and immunosignals for β3 subunits have been described in the rat LSO (Pirker et al., 2000). Our patch-seq results confirm the expression in pLSOs. Gabrb3 loss results in severe cochlear histopathology, providing evidence for a functional GABAergic innervation of the cochlea (Maison et al., 2006). The physiological role of β3 has not yet been analyzed in the CAS. The same holds for GABAARα1, whose encoding gene was a Super DEG in LOCs. In a mammalian expression system, receptors containing the α1 subunit responded to GABA application with long-duration openings that were organized in multi-opening bursts, in contrast to brief-duration, isolated openings of receptors containing α6 subunits (Fisher, 2004). The author suggested that GABA acts as a high efficacy agonist at the former receptor type. Consequently, LOCs may be more efficiently inhibited by GABA than pLSOs (see also Fischer et al., 2019). Interestingly, a given pLSO neuron or LOC neuron displays an almost inverse expression pattern for β3 or α1 GABAAR subunits (Figure 13A), indicating substantial architectural and functional differences.

For LOCs, the Super DEGs Calca and Ucn, encoding the neuropeptides CGRP and urocortin, respectively, were DEG#1 and DEG#2 (Figure 4B). There is clear evidence that CGRP is a molecular marker for LOCs (mouse: Maison et al., 2003; Wu et al., 2018). Reportedly, CGRP is not associated with pLSOs, and our results confirm this, although we found few CGRP transcripts in 13% of these neurons (< 2 TPM; Figure 14A). Our results document the impressively complex multi-transmitter properties of LOCs at the single-cell level. The cohort displays 12 multi-transmitter variants, more than 90% (28/30) of the neurons express at least two of the seven transmitter genes analyzed, and up to five neurotransmitter genes can be co-expressed by a single neuron (Figure 14B). The dominant combination is GABA + CGRP + urocortin, which was observed in 61% of the LOCs (Figure 16; Supplementary Table S6). In 10 neurons, it was flanked by the expression of a fourth or even a fifth gene. Urocortin demonstrates an even higher cell-type specificity than CGRP, as gene expression was completely absent from pLSOs (Figure 14A). Taken together, our patch-seq results demonstrate high transmitter complexity for LOCs at single-cell resolution. They also imply functional diversity within this relatively small population of efferent neurons.

Scn3a|Nav1.3 (DEG#11 in LOCs) encodes an α subunit of Nav1.3 channels generally less common than Nav1.1, Nav1.2, and Nav1.6 channels in the central nervous system (Scott et al., 2010). Nav1.3 channels are more abundant in the fetal than the adult brain (Liao et al., 2023). Nav1.3 currents recover rapidly from inactivation, which results in low AP thresholds and high AP frequencies (Cummins et al., 2001; Liao et al., 2023). However, the distinct expression of Scn3a|Nav1.3 in LOCs does not explain the higher AP threshold compared to pLSOs (Supplementary Figure S6J). AMBA shows no Scn3a expression in the LSO at P56, and P14 data was not available (Supplementary Figure S5).

As gene expression for Nav1.3 subunits was low in pLSOs, we assessed other Nav genes and found the highest expression level for Scn8a|Nav1.6, followed by Scn1a|Nav1.1 (66 and 18 TPM; Supplementary Figure S8A; Supplementary Table S6). The LOC cohort displayed considerably lower Nav1.6 expression (12 TPM; < 5-fold). Nav1.6 subunits are abundant in the nervous system and occur in various excitatory and inhibitory neurons. They form dense clusters at axon initial segments and nodes of Ranvier where they are the dominant Nav subtype (Caldwell et al., 2000; Chen et al., 2008; Zybura et al., 2021). For example, they are the major isoform at Calyx of Held-MNTB axon terminals where the abundance increases during development (Leao et al., 2005). In the auditory nerve, Nav1.6 and Nav1.1 are placed at the AP generator site, together with the submembraneous scaffolding protein ankyrin (Kim and Rutherford, 2016). As demonstrated in several brain regions, Nav1.6 channels contribute to peak Na⁺ currents and repetitive AP firing (references summarized in Chen et al., 2008). They display persistent and resurgent Na⁺ currents, lower the AP threshold and mediate repetitive spiking (Zybura et al., 2021). Together, the above findings suggest an important role for Nav1.6 channels in initiation and propagation of high-frequency AP bursts in LSO neurons, particularly in myelinated, mature pLSOs.

Cacna1e, coding for Cav2.3, was DEG#26 in LOCs. Previously, we identified mRNA for Cav2.3 in LSO neurons (Jurkovicova-Tarabova et al., 2012). Cav2.3 is the α1 subunit of R-type Cav channels (Catterall, 2011). Together with PQ-and N-type channels (containing Cav2.1 and Cav2.2 subunits, respectively), R-type channels mediate presynaptic Ca²⁺ influx and neurotransmitter release (Ricoy and Frerking, 2014; Held et al., 2020). Among the 18 Cav genes expressed in the LSO cohort, Cacna1e|Cav2.3 demonstrated the highest value by far (mean TPMs: LOCs 1,491; pLSOs 182; Supplementary Figure S9A; Supplementary Table S6). High TPM values also occurred for Cav2.2 in both neuron types (pLSOs 414; LOCs 363). Together, we reason that LOCs likely prioritize Cav2.3 subunits over Cav2.2 subunits, whereas pLSOs favor the reverse order. On the other hand, Cacna1a|Cav2.1 expression seems be negligible in both groups (6.1 vs. 2.2 TPM). The functional impact of this needs to be determined.

Our Cluster similarity scoring finally revealed 12 genes, four of which code for TRP channels, another four for glutamate receptor molecules, two for Cav channels and one each for glycine and GABA receptor subunits (Figure 15). Among the four TRP channels were two TRPM channels (melastatin) as well as one TRPML (mucolipin), and one TRPP (polycystin) channel (TRPM2, TRPM7, TRPML1, TRPP5). TRP channels are cation channels that generally cause cell depolarization. However, they can also be associated with intracellular organelles and function as intracellular Ca²⁺ release channels (Gees et al., 2010). They play a crucial role in sensory physiology, possibly contributing to hearing (Venkatachalam and Montell, 2007). TRP channels have been poorly studied in the CAS. In the mouse ventral cochlear nucleus, activation of TRPM2 channels shifts V_rest to more positive values (Bal et al., 2020). The authors reasoned that TRPM2 channels, which are induced by oxidative stress, may play a modulatory role in setting the excitability level of stellate cells. To our knowledge, there is no study on TRPM or TRPPP channels in the SOC. Our results show that pLSOs and LOCs are similar in their TRP channel composition, as also evidenced by the absence of any DEG in this group. These results pave the way for further investigations. Of note is the relatively low percentage of neurons expressing a given TRP gene (highest value: 50% for TRPM7 in LOCs), indicating heterogeneity within a given cluster. This issue of inter-cellular heterogeneity is of interest for future studies.

Regarding Cluster similarity for glutamate receptor molecules, ~15% (4/26; including 18 ionotropic and 8 metabotropic) met the criteria, namely GluA2, GluK5, GluN1, and mGluR8. This corresponds to one gene in each major class (AMPA; kainate, NMDA, metabotropic; Figures 15A2,B). Regardless of subtype, ~75% of the LSO neurons showed GluA2 transcripts, the highest proportion across all 32 ligand-gated ion channels (Figure 13A). The Cluster similarity observed for GluA2 contrasts with the Super DEGs GluA1 and GluA4. Edited GluA2 subunits render AMPA receptors Ca²⁺ impermeable and are abundant in the brain. AMPA receptors lacking GluA2 are Ca²⁺-permeable, allowing them to serve as signaling molecules via Ca²⁺ influx, in addition to mediating excitatory neurotransmission (Man, 2011).

Kainate receptor subunits, including GluK5, have been identified in the SOC by immunohistochemistry and pharmacological analysis (Petralia et al., 1994; Löhrke and Friauf, 2002; Vitten et al., 2004). Notably, GluK5 subunits have high agonist affinity and fast kinetics (Dhingra et al., 2022). In summary, there is consistent evidence that both types of LSO neurons are equipped with fast gating glutamate receptors, and pLSOs even more so than LOCs.

Our results on GluN1 confirm immunohistochemical findings in the LSO (Nakagawa et al., 2000) and attribute the subunit to both pLSOs and LOCs. Glycine, a co-agonist at NMDA receptors, binds to GluN1 subunits, whereas glutamate binds to GluN2. Consequently, NMDA receptors require both subunits to be functional. Our study shows similar expression for both isoforms (GluN2B and GluN2D), and the majority of pLSOs express GluN2B, especially the Sustained neurons (Figure 13A). Our GluN2B results are consistent with a pharmacological report suggesting a correspondence of the NMDA receptor subunit with the period of major circuit refinement at MNTB-LSO synapses (Case and Gillespie, 2011). In particular, GluN2B subunits confer relatively long decay kinetics to NMDA receptors, and the developmental shortening of EPSCs likely results from a subunit substitution of GluN2A for GluN2B.

For genes involved in inhibitory neurotransmission, Cluster similarity was demonstrated for the ionotropic GlyR subunit GlyRβ and the metabotropic GABABR1 (Figures 13A2,B2, 14A; Supplementary Figure S12A). For GlyRβ, we found expression in 79% of the pLSOs and 67% of the LOCs. The 79% value was the highest observed in the pLSO cohort. The presence of GlyRβ transcripts is expected, as this subunit is a mandatory component of heteropentameric GlyRs and essential for their postsynaptic localization (Weltzien et al., 2012). The higher TPM values for the adult GlyRα1 compared to the fetal GlyRα2 (Figure 13A) indicate that most of the replacement of the fetal by the adult α subunits (Dutertre et al., 2012) has occurred by the time mice begin to hear, at least in pLSOs (α3 and α4 transcripts were not detected in our sample).

GABABR1 is one of two GPCR molecules activated by GABA (review: Bassetti, 2022). The level for GABABR1 transcripts was moderate in our LSO sample (~25 TPM), >10-fold higher than the very low Gabbr2|GABABR2 expression (~1 TPM; Supplementary Figure S12A; Supplementary Table S6). Since GABABRs are obligatory heterodimers composed of GABABR1 and GABABR2 subunits (review: Frangaj and Fan, 2018), the difference in expression is surprising. Based on our transcript results, we cannot localize the GABABR subunits to presynaptic or postsynaptic sites. We have recently shown that postsynaptic GABABRs mediate hyperpolarization in a modulatory manner in mouse pLSOs (Fischer et al., 2019). GABABR signaling has previously been shown to mediate long-lasting depression at inhibitory gerbil MNTB-LSO synapses (Kotak et al., 2001). This form of synaptic depression declines with age and may contribute to the use-dependent reorganization of MNTB axon arbors.

In the Cav channel group (Zamponi et al., 2015), we observed Cluster similarity for two subunits, Cav1.3 and Cavγ2. Cav1.3 plays an important role in the auditory system. It is involved in deafness (Platzer et al., 2000) and in the development of auditory brainstem structures, including the LSO (Hirtz et al., 2011, 2012). Functional Cav1.3 channels have been described in pLSOs (Jurkovicova-Tarabova et al., 2012). Therefore, the present results at the mRNA level are consistent with previous findings at the protein level.

Cavγ2, also known as stargazin or TARPγ2 (most literature refers to stargazin), is one of eight Cavγ isoforms in a functionally diverse family (Chen et al., 2007; Bissen et al., 2019). Six of the eight Cavγ isoforms belong to the TARP subfamily, including Cavγ2 (and Cavγ5). TARP stands for “Transmembrane AMPAR regulatory protein,” implying an interaction with AMPA receptors and not a mandatory role as a true Cav subunit. TARP/AMPAR complexes promote the correct trafficking of AMPARs and their targeting to the synaptic surface (review: Jacobi and von Engelhardt, 2017). TARPγ2-containing AMPARs have been localized in presynaptic axon terminals where they modulate GABA release (Rigby et al., 2015). The expression profile of Cavγ2 (and the other AMPAR regulatory subunits) and the functional characterization in the auditory brainstem await future studies. In this context, it is worth noting that Cacng5|Cavγ5 was DEG#12 in pLSOs.

Kcnq2|Kv7.2 was not on the top 12 list of genes showing Cluster similarity (its position was 18; Figure 15B2). Nevertheless, we analyzed Kv7.2 in combination with Kv7.3 by means of immunohistochemistry and pharmacology. Kv7.2 and Kv7.3 are primarily found in the brain (Baculis et al., 2020). They are the major molecular correlate of low-threshold, non-inactivating Kv channels that mediate the M current, which is hyperpolarizing and critical for preventing neuronal hyperactivity (Wang et al., 1998; Robbins, 2001; Brown and Passmore, 2009). In the calyx of Held terminals, Kv7 channels, most likely Kv7.5 homomers, control release probability through changes in V_rest (Huang and Trussell, 2011). In adition, they function to facilitate reliable high-frequency synaptic signaling (Zhang et al., 2022). Our results suggest a similar function for Kv7.2 and Kv7.3 channels in pLSOs. Whether the channels are localized in pLSO axon terminals and involved in modulating transmitter release requires further analysis.

There is increasing evidence that a variety of characteristics vary considerably among species in the SOC, including the LSO. Species-specific effects (e.g., low-frequency vs. high-frequency hearing) on structure and function are well known (Heffner and Heffner, 1992; Grothe, 2000; Fritzsch et al., 2013) and are evident in the SOC. First, there is considerable variation in the arrangement, shape, and relative size of individual SOC nuclei (Friauf et al., 2019). Second, there are species differences in the location of olivocochlear neurons in the SOC (Aschoff and Ostwald, 1987; Brown, 2011). Third, there are species differences in the quantitative contribution of LOCs and MOCs to olivocochlear fibers (Perrot and Collet, 2014). Fourth, interspecies differences may explain the inconsistencies between reports on the transmitter repertoire of LOCs (Guinan, 2011). Fifth, discrepancies in the excitatory/inhibitory transmitter ratio have been reported for pLSO neurons (Williams et al., 2022). In the heterogeneous LSO, these differences have created multi-scale, long-standing, and persistent puzzles (Guinan, 2011; Williams et al., 2022). Consequently, general statements about the LSO when pooled from different species must be interpreted with caution. We have taken this into account for the summary cartoon (Figure 16), which shows results from mice only.

For a few genes, we exemplarily validated our transcriptomic results on the protein level (Figures 5, 10, 11, 12). Furthermore, we compared them with those of a very recent study using single-nucleus sequencing of LOCs (Frank et al., 2023). The comparison revealed considerable similarity between the genes detected, thus strengthening each study. For example, both studies found expression of Col4a3 and B3galt1 in LOC neurons.

We do not want to hide the fact that the patch-seq method has its limitations. The general limitations of scRNA-seq also apply to the patch-seq variant. Contamination from the extracellular solution can lead to false positives and affect the quality of the results. To minimize this problem, we performed stringent quality controls (see Methods). The low throughput (we patched 263 neurons but collected only 103) reflects our caution in cell harvesting to minimize contamination. In addition, we took advantage of the fact that extracting the cytosol and the nucleus significantly increases the RNA yield and improves the quality of the transcriptomic data. Finally, by including negative controls and carefully monitoring the harvesting process, we were able to gain insight into potential sources of contamination. Our observations indicate that the effects of contaminants on the transcriptomic profiles were generally small. For example, ChAT expression and urocortin expression were absent in the pLSO cohort. Nevertheless, we attribute some of the results on CGRP (in 12.5% of pLSOs) and HCN (in 17% of LOCs) to contamination (see Figure 14; Supplementary Figure S11). To allow a post hoc reaction to the effects of the contamination problem, we have designed the gene tables in such a way that the readers can interactively change the TPM thresholds and thus adjust the stringency of the filtering criteria to their personal needs (cf. Supplementary Table S6).

The age of the animals is another factor influencing contamination. In patch-seq experiments in adult tissue, where a dense perineuronal network of glial cells is present, the risk of contamination is higher. This was one of the reasons why we used young mice (P10-P12) in the present study. By limiting the sampling to this age, we also obtained a unique database for the cell identity of the developing LSO neurons. Given that numerous electrophysiological, biochemical and anatomical studies have used and continue to use this age in the LSO, we believe that our database will be of great value to the research community.

Other challenges arise from the fact that, due to biological and technical limitations, only a fraction of mRNA is captured, leading to zero inflation and false-negative results. We believe that some of our results, such as not detecting HCN in all pLSOs, are affected by zero inflation.

Comparisons of results at the transcript level with those at the protein level are confronted with an imperfect correlation between the transcriptome and the corresponding proteome (Nie et al., 2007; Schwanhäusser et al., 2011; Liu et al., 2016). The mapping between high-throughput measurements at the level of the transcriptome and the corresponding proteome is complex (Gunawardana et al., 2015). Conclusions based on the transcriptome alone should therefore be treated with caution, and gene expression levels should only be used as a proxy for protein levels (Liu et al., 2016).

We assume that some of our unexpected results have biological causes, in addition to the artifacts mentioned above. For example, the lack of ChAT expression that we observed in most LOCs is consistent with the low in situ hybridization signals detected in AMBA LSO P14 sections. However, neurons in the facial nucleus and other motor nuclei are strongly labeled in the AMBA. It is possible that gene expression for ChAT is low in the LSO while protein abundance is high. Taken together, it is advisable for future studies to first extend the transcriptomic findings to the protein level in order to determine whether the described mRNA level is manifested in protein abundance and to localize the proteins in the neurons before starting functional analyses.