We performed snRNA-seq from dopaminergic nuclei (defined by Dat-Cre expression) in LRRK2^G2019S heterozygous knock-in mutants and control littermates using a similar technique to our previously published dataset (Azcorra et al., 2023) but modified to utilize a chip-based microfluidic sorting method which minimizes nuclear stress. This improved protocol enabled increased yield and sample quality (Figure 1A, Figure 1—figure supplement 1A). Doing so enabled us to generate a dataset with 28,532 profiled cells (after quality control filtering) with a median UMI count of 7750.5 and median of 3056 genes. Downstream clustering resulted in 22 unique clusters (Figure 1B), of which two clusters (16 and 21, 911 total cells) were determined to represent possible dopaminergic/glial doublets based on co-expression of glial markers in these populations and were excluded from downstream analyses (shown in gray in Figure 1B, Figure 1—figure supplement 1F). The 20 remaining clusters all showed expression of key pan-dopaminergic markers including Slc18a2 (Vmat2), Ddc, Th, and Slc6a3 (Dat) (Figure 1C). These clusters were observed in both males and females, and across genotypes, in similar proportions (Figure 1—figure supplement 1B–E). These subtypes showed distinct expression signatures of many previously described markers of DA neuron subtypes (Figure 1D; Poulin et al., 2020). However, our increased yield and sample quality also allowed explicit detection of known populations that have eluded most prior single-cell studies, such as DA neurons expressing Vip (Poulin et al., 2020; Poulin et al., 2014), thereby demonstrating the strength of this pipeline in detecting rare DA neuron subtypes. Importantly, all the described subtypes showed high expression of midbrain floorplate markers (Figure 1E), with exception of clusters 0 and 17, which possibly represent Slc6a3+ neurons in the supramammillary and/or premamillary regions (Soden et al., 2016; Nouri and Awatramani, 2017). Thus, our increased number of subtypes over prior studies represents increased granularity among classic midbrain DA neuron subtypes, rather than detection of non-floorplate derived populations that express DA markers.

Given the large number of clusters discovered in our dataset, we first sought to organize these subtypes into groups in an unbiased manner. By creating a cluster dendrogram (Figure 2A), we were able to map an approximation of the relative relatedness of different subtypes, which segregated into three main families based on major branch points. By running differential expression on each node of the cluster dendrogram (i.e. exploring differentially expressed genes (DEGs) between the two arms emerging from a given node), we were then able to create a stepwise expression code that can uniquely define any individual cluster (Figure 2A and G). The earliest division point among our clusters was defined by genes Prkcd and Lef1, which are highly expressed only in cluster 17 (Log2 fold changes = 6.279 and 7.246 respectively, BH-corrected p-values ≈ 0; note, such p-values are calculated using default Seurat differential expression functions which treat each cell as an independent sample thus potentially inflating statistical significance). Next, a family of clusters defined by Gad2/Kctd16 (deemed Gad2 family, Log2 fold changes = 5.318 and 4.014 respectively, BH-corrected p-values ≈ 0) separated from all other clusters, which expressed significantly higher levels of Ddc and Slc6a3 (Log2 fold changes = 1.419 and 1.643 respectively, BH-corrected p-values ≈ 0). The first major branch point between Ddc-high/Slc6a3-high clusters was then defined by cells expressing Sox6/Col25a1 (Log2 fold changes = 2.720 and 2.305 respectively, BH-corrected p-values ≈ 0) versus those expressing Calb1/Ndst3 (Log2 fold changes = 2.025 and 3.368 respectively, BH-corrected p-values ≈ 0), thus creating the Sox6 family and Calb1 families, respectively. This division between Sox6 and Calb1 is consistent with our previous work (Poulin et al., 2014), as well as a prior large single-cell profiling study that found this division to be the central branch point among DA neurons across several species including humans (Kamath et al., 2022; Siletti et al., 2023). Cluster divisions can be visualized by plotting the co-expression of DEGs taken from the branch points labeled in red (Figure 2C–F) on the UMAP plot. Of note, while these branch point markers describe the overall trends of gene expression, these genes are not perfectly distinct or universal among the members of these cluster families. For example, clusters 1, 11, and 13 show expression of both Sox6 and Calb1 (Figure 1D). The co-expression of these markers has previously been described in a fraction of mouse DA neurons (Poulin et al., 2014; Anderegg et al., 2015; Pereira Luppi et al., 2021). Furthermore, because the calculations of differential expression at each node take into account only the descendants of that branch point, markers defining a branch point may also be strongly expressed in clusters outside this comparison. Thus, while the final marker(s) listed for each subtype in Figure 2A are not unique to that population, following the sequential branches leading to this marker in a stepwise fashion will specifically enrich for this population (Figure 2G). By doing so, we provide a strategy for potential genetic access to any individual DA subtype using intersectional logic gates.

To enable comparisons of DA populations across existing and future literature, we next sought to better define genetic signatures for each of the individual putative subtypes and to provide a common language for discussing these populations. To do so, we utilized two complementary approaches. First, we explored the top differentially expressed genes with positive relative expression in each cluster (Figure 2B). While many of these markers are expressed in more than one population, some (e.g. Vip) are almost entirely unique to a single cluster. As a complementary approach, we performed differential expression among clusters within each of our three subtype families, thereby creating a shorthand identity based on the cluster family and top DEGs (Figure 2G). Within the Sox6 family, these clusters are Sox6 ^Tmem132d (listed as cluster 1 in Figures 1 and 2), Sox6^Arhgap28 (cluster 2), Sox6^March3 (cluster 3), Sox6^Tafa1 (cluster 4), Sox6^Kcnmb2 (cluster 8), and Sox6^Vcan (cluster 10); in the Calb1 family, these are Calb1^Pde11a (cluster 5), Calb1^Kctd8 (cluster 6), Calb1^Ptprt (cluster 7), Calb1^Sulf1 (cluster 9), Calb1^Stac (cluster 11), Calb1^Chrm2 (cluster 12), Calb1^Sox6 (cluster 13), Calb1^Lpar1 (cluster 14), Calb1^Ccdc192 (cluster 18), Calb1^Gipr (cluster 19); in the Gad2 family, these are Gad2^Syndig1 (cluster 0), Gad2^Egfr (cluster 15), and Gad2^Ebf2 (cluster 20). Cluster 17 (defined by Lef1) did not have an associated cluster family (Figure 2A). These shorthand identities provide a simple nomenclature for our clusters, which we have utilized throughout the remainder of this paper.

After establishing genetic signatures for each putative subtype, we were next able to correlate our cluster identities to those described in another recent study (Azcorra et al., 2023), as well as the consensus subtypes proposed in Figure 2G; Poulin et al., 2020. Doing so provides context for discussing our putative DA neuron subtypes with regards to prior literature. For example, cluster Sox6^Tafa1 was found to be equivalent to a population we previously defined by high expression of Anxa1, which holds notably distinct functional and anatomical properties (Azcorra et al., 2023). Several of the subtypes that emerged in our new dataset also appear to be novel divisions within clusters that previously showed evidence of internal heterogeneity based on their cluster stability metrics (Azcorra et al., 2023). For instance, a Vip + cluster (Calb1^Gipr) emerged from within a highly variable cluster (defined as #10 in Azcorra et al., 2023), and additional subdivisions emerged from previously described unstable Sox6+ clusters (most notably #3 in Azcorra et al., 2023; Figure 2G). This emergence of unique populations, such as Calb1^Gipr, is likely due to the increased yield and sample quality of our new dataset which has enabled better detection of small populations and gene markers with lower expression levels. This is further supported by the high cluster stability of the new Calb1^Gipr (Figure 2—figure supplement 1A and B), resolving the instability of the equivalent cluster in Azcorra et al., 2023. We re-clustered our snRNA-seq data into the clusters defined in Azcorra et al., 2023 and found our clusters aligned with or further subdivided many of the previously defined clusters (Figure 2—figure supplement 2A). Lastly, to further corroborate our cluster identities, we used an independent bioinformatic platform, Scanpy, to re-cluster our nuclei (Figure 2—figure supplement 2B). Comparing the Seurat and Scanpy clusters using a Sankey diagram (Figure 2—figure supplement 2C) shows a high correlation between the Seurat and Scanpy clusters.

Having identified multiple clusters of midbrain DA neurons, we next resolved to determine the spatial distribution of these populations using MERFISH (Chen et al., 2015), a form of spatial transcriptomics that offers subcellular resolution imaging-based RNA quantification. We assembled a 500 gene panel which included 120 of the top DEGs identified within the snRNA-seq dataset, as well as 380 general markers of neuronal and non-neuronal cell populations (Supplementary file 1). Seven coronal slices were collected from adult mouse midbrains and processed for MERFISH (Figure 3A). The coronal slices were selected containing ventral midbrain regions between –2.9 mm to –3.9 mm from bregma. Cells were segmented using the CellPose algorithm based on DAPI and Poly-T staining and passed through quality control filters to remove cells with too large of a volume (potential doublets) or too few transcripts detected (false cells; Figure 3—figure supplement 1A–C). This resulted in 429,713 identified cells (Figure 3B). Cells from the seven sections displayed significantly correlated gene expression and minimal batch effects (Figure 3—figure supplement 1D). Cells were integrated in Seurat using reciprocal principal component analysis (RPCA) algorithm and then clustered, resulting in 40 final clusters comprising distinct classes of neurons, oligodendrocytes, oligodendrocyte precursor cells (OPCs), astrocytes, microglia, and meningeal cells (Figure 3B). These clusters could be attributed to 25 functional groups based on the expression of cell type markers from Supplementary file 1 and their spatial distribution. Glia-associated genes were mostly contained within their expected cell types: oligodendrocytes (Mog, Cnp), OPC (Pgfra, Olig2), astrocytes (Aqp4, Gfap), and microglia (Aif1, Tmem119). However, we noted the unexpected distribution of the oligodendrocyte-specific transcript Mbp in non-glial cells which we believe might indicate the presence of this mRNA in glial-processes that overlap non-glial cells (Bradl and Lassmann, 2010). Within the neuronal class, we identified 16 excitatory clusters (presence of Slc17a6 or Slc17a7), 7 inhibitory clusters (presence of Slc32a1, Gad1 or Gad2), one serotonin neuron cluster, and one DA neuron cluster (Figure 3B-D, Figure 3—figure supplement 1G). Although our gene panel was designed to identify neuronal heterogeneity of the midbrain, it had sufficient resolution to delineate seven excitatory neuron classes present in distinct layers of the cortex and six cortical interneuron classes (Figure 3—figure supplement 1E, F). Cortical excitatory neurons were absent from subcortical structures such as the thalamus, hypothalamus, midbrain, and hindbrain. This fundamental dichotomy between cortical and subcortical neurons in the adult mouse brain is supported by recent spatial transcriptomic studies (Zhang et al., 2023; Yao et al., 2023; Langlieb et al., 2023).

A population of DA neurons was identified among our 40 clusters based on the expression Th, Slc6a3, and Slc18a2 (Figure 3C). The location of the dopaminergic cluster (blue in Figure 3D) considerably overlaps with Th expression (Figure 3E). This cluster comprised 4532 cells, was entirely located in the midbrain, and was also enriched for Ddc, Dlk1, and Drd2 (not shown). Within this population, we detected expression of genes associated with glutamatergic and GABAergic neurotransmission, in line with previous reports of multi-lingual characteristics of DA neurons (Trudeau et al., 2014; Descarries et al., 2008; Morales and Root, 2014; Morales and Margolis, 2017; Tritsch et al., 2016; Conrad et al., 2024), for instance, the glutamatergic marker Slc17a6 (Vglut2) and GABAergic markers Gad2 and Slc32a1 (Vgat) (Figure 3—figure supplement 1G). Markers for cholinergic (Chat) and serotonergic (Tph2) neurotransmission were not detected in putative DA neurons and the glutamatergic transporters Slc17a7 (Vglut1) and Slc17a8 (Vglut3) were also absent. We next wanted to characterize the neuroanatomical location of this DA neuron cluster. We manually delineated neuroanatomical boundaries on each of the slices for seven midbrain regions known to contain most dopamine neurons including SNc, VTA, RR, caudal linear nucleus (CLi), interfascicular nucleus (IF), rostral linear nucleus (RLi), and periaqueductal gray (PAG). These boundaries were delineated based on cell density (dapi), white matter landmarks, and Th transcript distribution using the Allen Brain Atlas as a reference (Figure 3—figure supplement 2A). We found that cells in our DA neuron cluster were located in the following locations: 35.2% in the SNc, 19.8% in the VTA, 7.8% in the CLi, 15.3% in the RR, 2.0% in the IF, 0.6% in the RLi, 4% in the PAG and 15.3% located elsewhere (Figure 3—figure supplement 2B). Finally, to explore the diversity of DA neurons identified by MERFISH, we subclustered the 4532 cells to resolve 12 distinct subgroups (Figure 3—figure supplement 2C–F). For clarity, we designated the clusters identified solely with MERFISH transcript levels as MER-0–11. Many of the clusters were characterized by increased expression of known subtype markers such as Calb1 (MER-6, MER-8 and MER-10), Aldh1a1 (MER-5), Ndnf (MER-0), Gad2 (MER-7 and MER-9), and Slc17a6 (MER-2).

To provide the spatial distribution of the clusters identified using snRNA-seq, we sought to project the MERFISH dataset onto the snRNA-seq UMAP space. We first compared the expression level of the 500 genes utilized in our panel to select features for integration with comparable expression levels. We selected 281 genes to be included based on similarity in the average counts per cell between the datasets. Some excluded genes potentially reflected faulty probes or improper assignment of RNA transcripts to segmented cells. For instance, Sox6 transcripts were not detected across the entire brain (data not shown). We successfully projected the MERFISH dataset onto the snRNA-seq UMAP space and selected for 2297 dopamine neurons with a cell similarity score >0.5 (mean value, 0.538), increasing confidence in the correct assignment to snRNA-seq defined clusters. Almost all clusters were represented by more than 20 cells in our MERFISH dataset (Figure 4A, Figure 4—figure supplement 1A). Clusters Calb1^Ccdc192 and Calb1^Gipr were not resolved in this label transfer either due to their relatively smaller number or location in an under-sampled region. Indeed, the Vip-expressing Calb1^Gipr has been shown to be located in the caudal midbrain (PAG/DR) (Poulin et al., 2014), a region not well covered in our MERFISH experiment. Comparing snRNA-seq clusters with the ones obtained by clustering the MERFISH dataset alone showed a strong correspondence (Figure 4—figure supplement 1B). Half of the snRNA-seq clusters were predominantly comprised of single MERFISH clusters (10/20) and 25% (5/20) were comprised of two MERFISH clusters, demonstrating that the integration with the snRNA-seq dataset allowed us to further refine our MERFISH-based classification (Figure 4—figure supplement 1B). We then imputed expression data from the snRNA-seq dataset onto the MERFISH cells to better visualize spatial expression levels. The imputed gene expression is extrapolated from anchors established from pairwise correspondences of cell expression levels between MERFISH and snRNA-seq datasets. The imputed data across the genes used for data integration correlated strongly with MERFISH transcripts, particularly for Calb1, Gad2, and Aldh1a1 genes (Figure 4—figure supplement 1C-F, Figure 4—figure supplement 2). Altogether, the dataset integration allowed us to map the neuroanatomical location of 18 of the clusters identified by snRNA-seq as well as impute expression of the whole transcriptome.

Clusters were sorted into Gad2, Sox6, and Calb1 families based on our hierarchical dendrogram (Figure 4A; Gad2 n=124 cells, Sox6 n=887 cells, Calb1 n=590 cells). Imputed Gad2, Sox6, and Calb1 gene expression within DA neurons showed a distinct medio-lateral and rostro-caudal distribution (Figure 4C–E). The spatial distribution of imputed Sox6 and Calb1 transcripts within DA neurons correlated well with previous reports (Poulin et al., 2014; Pereira Luppi et al., 2021; Panman et al., 2014) whereas the distribution of the Gad2 family is predominantly observed in the rostral linear/posterior hypothalamus region and CLi, and to a lesser extent in the VTA (Figure 4B, Figure 4—figure supplement 3). Each family was overrepresented in a neuroanatomically-defined area; for instance, the SNc is composed of 78.4% of neurons from the Sox6 family (Figure 4—figure supplement 3), consistent with previous reports (Pereira Luppi et al., 2021). However, each family was also found across multiple neuroanatomical regions (Figure 4—figure supplement 3A). We also observed spatial heterogeneity within families. For example, one branch of the Calb1 family dendrogram (Figure 4, label in shades of yellow-orange defined by higher Lepr expression) was enriched in the dorsal VTA, whereas the other Calb1 branch (Figure 4, label in shades of red defined by higher Ntm expression) populated mostly the ventromedial VTA. Similar observations could be made for the Sox6 family, where subtypes displayed biased distributions. This led us to explore the distribution of each individual cluster and representative markers toward the goal of generating a granular map of the midbrain DA system.

The hierarchical clustering of the Sox6 family reveals two major branches (Figures 2A and 4A): (1) Branch 1 defined by higher levels of Slc44a5 and composed of subtypes Sox6^Tmem132d and Sox6^March3 and (2) Branch 2 defined by Cntnap5 and composed of subtypes Sox6^Tafa1, Sox6^Arhgap28, Sox6^Kcnmb2, and Sox6^Vcan. We observed an important discrepancy between these two branches, with Branch 1 being located more dorsal and medial than Branch 2 (Figure 5A). Interestingly, this dorsoventral organization breaks down in the caudal SNc and RR where neurons from both branches are mostly intermixed. Only minor differences in location were observed between clusters of Branch 1, maybe reflecting their molecular similarities. Of note, whereas Sox6^Tmem132d is predominant in the rostral SNc and VTA for this branch, neurons of Sox6^March3 are more numerous in the RR (Figure 5B, Figure 4—figure supplement 3B). Within Branch 2, the molecular identity correlated with the medio-lateral distribution. For instance, neurons of cluster Sox6^Arhgap28 were found to be more lateral in the SNc compared to neurons of Sox6^Tafa1, Sox6^Kcnmb2, or Sox6^Vcan. This medio-lateral distribution of neurons of the SNc ventral tier might reflect the differential topographical distribution of nigrostriatal projections previously reported (Wu et al., 2019; Poulin et al., 2018).

The Calb1 family was the most molecularly diverse which was also reflected in its localization (Figure 6). Two major branches of Calb1 DA neurons were identified: (1) Branch 1 (Figure 6A–B) containing subtypes Calb1^Ptprt, Calb1^Sulf1, and Calb1^Sox6 and (2) Branch 2 (Figure 6C–D) containing subtypes Calb1^Chrm2, Calb1^Kctd8, Calb1^Stac, Calb1^Pde11a, and Calb1^Lpar1. Broadly speaking, Calb1 subtypes were predominantly localized to the VTA and CLi with some neurons showing localization to the dorsal SNc, SNpl, or RR (Figure 4—figure supplement 3). Within the VTA, Branch 1 Calb1 neurons tended to localize to the dorsal VTA (Figure 6A–B) while Branch 2 Calb1 neurons showed distinct ventromedial localization (Figure 6C–D). In more caudal sections the difference was even more pronounced, although intermingling was always observed. Within Branch 1, subtypes Calb1^Ptprt and Calb1^Sox6 were largely absent from the SNc dorsal tier and RR, whereas subtype Calb1^Sulf1 was present. Some notable distinctions were observed in both groups, specifically subtypes Calb1^Sulf1 and Calb1^Lpar1 which were each observed in significant numbers in the dorsolateral SNc/SNpl region. Of these, only Calb1^Lpar1 expressed significant Slc17a6 (data not shown), making it a likely candidate for the SNpl Vglut2 + DA neuron population found to innervate the tail of the striatum (Poulin et al., 2018). Indeed, these laterally located neurons have been genetically targeted with a Slc17a6-Cre driver, and display deceleration-correlated responses, and robust responses to physically aversive stimuli (Azcorra et al., 2023; Poulin et al., 2018).

Within the Gad2 family, we were able to resolve two subtypes. Subtypes Gad2^Egfr and Gad2^Ebf2 had very similar spatial distributions and were prominent in the posterior IF, RLi, and CLi (Figure 7, Figure 4—figure supplement 3). These regions have been associated with small DA neurons of lower Th and Slc6a3 levels. Subtype Gad2^Egfr neurons were more plentiful than cluster Gad2^Ebf2, with the latter being absent in more caudal sections. All Gad2 subtypes expressed Slc32a1, opening the possibility that these neurons co-release GABA. Further, they also expressed Slc17a6, revealing potentially multilingual DA neuron subtypes.

With cluster identities and locations defined, we next sought to explore the downstream gene expression characteristics of DA neurons in Lrrk2 mutants. The role of Lrrk2 in PD pathophysiology remains unclear (Singh et al., 2019), and Lrrk2 mutations are often proposed as being an indirect source of dysfunction for the DA system, such as through its high expression in glial or striatal cells (Chen et al., 2020; Choi et al., 2015; Parisiadou et al., 2014; Cook et al., 2017; West et al., 2014). However, previous studies have shown alterations in DA signaling or gene expression even when the mutant LRRK2 was expressed specifically in DA neurons (Pallos et al., 2021; Liu et al., 2015). Thus, as a first step in demonstrating the possibility that mutant Lrrk2 in DA neurons themselves can contribute to dysfunction, we mapped the expression of Lrrk2 across all DA neuron subtypes (Figure 8A). We found detectable expression across most subtypes, with the notable exception of Gad2^Egfr neurons.

Given the distinctive properties of different DA neuron subtypes, we next sought to test if phenotypic changes in Lrrk2 mutant mice are driven by changes in the relative proportions of said subtypes. While previous studies have shown no overt DA neuron loss, it is possible that there was a selective reduction of a subtype(s) that would be undetectable by previous methods. We first graphed the proportions of each subtype within our control and mutant datasets (Figure 8B), and found remarkable similarity across these samples, consistent with previous reports showing no DA neuron loss in these mice (Yue et al., 2015). To further assess for changes in cell state among clusters across conditions, we next compared the similarity of each cluster to all other clusters across conditions (Figure 8C). MetaNeighbor analysis revealed that the most similar population to each wildtype cluster was its corresponding subtype in the Lrrk2 mutant samples. This suggests that the general subtype organization is largely impervious to LRRK2^G2019S perturbation, and that our clustering scheme is likely more reflective of cell type rather than cell state.

To assess pan-dopaminergic gene expression changes as a function of Lrrk2 genotype, we next looked at globally differentially expressed genes across our mutant and control samples. Calculating differentially expressed genes across all clusters revealed a large number of significantly enriched or diminished genes as a function of Lrrk2 mutation (1646 genes in total with BH corrected p-value <0.05, Wilcoxon rank sum test). The results of this differential expression analysis are plotted in Figure 8—figure supplement 1A, with full results and statistics available in Supplementary file 2. Notably, most of these changes appear to be small in magnitude; among FDR-significant genes, only 51 showed increased expression and seven showed decreased expression with log2 fold changes greater than 0.5 (Figure 8—figure supplement 1A). Two genes with high enrichment in Lrrk2 mutants, Mir124a-1hg and AC149090.1 were both of particular interest due to their potential relevance to PD pathogenesis. AC149090.1 is orthologous to the human PISD gene, which is involved in autophagy and implicated in mitochondrial dysfunction (Buckley et al., 2023; Zhao et al., 2019; Thomas et al., 2018). This gene has been proposed as a sensitive marker of biological aging in the brain (Buckley et al., 2023; Jin et al., 2023), an interesting finding given that age is the single largest risk factor for PD. Also of note, Mir124a-1hg is a host gene (though not the exclusive gene) for miR-124, a microRNA with purported neuroprotective effects that has been found to be downregulated in PD patients or some PD mouse models (Zhang et al., 2022b; Yao et al., 2018; Han et al., 2019; Angelopoulou et al., 2019). While the opposite effect was seen here (i.e. an upregulation of this gene in our model), one might speculate that miR-124 is initially increased during toxic metabolic insults as a compensatory response, given its proposed neuroprotective effects. In our prodromal model without observable degeneration, this could thus represent an early sign of cell stress. Conversely, in PD patients or overtly degenerative models, a lack of compensatory miR-124 or fulminant cell death among vulnerable cells could result in an observed decrease in miR-124 expression.

We next performed pathway enrichment analyses using Gene Set Enrichment Analysis (GSEA) and gene ontology (GO) to compare our Lrrk2 mutant samples to our controls. Using GSEA, we found 18 gene sets with significant differential expression (FDR-adjustedp-values <0.05). The top 13 of these enriched pathways (those with NES scores >1.8) are shown in Figure 8D. While a broad range of biological domains are found among these results, many of the top enriched pathways were related to energy production via mitochondrial oxidative respiration, remarkably similar to prior comparisons between vulnerable and resistant DA neuron populations (Pereira Luppi et al., 2021). Mitochondrial energy production pathways, and specifically the extraordinary bioenergetic burdens of DA neurons, have been heavily implicated in PD vulnerability (González-Rodríguez et al., 2021; Surmeier et al., 2017). Thus, the enrichment of these pathways in Lrrk2 mutants is particularly intriguing as it may reflect that the structural and functional alterations reported with LRRK2 mutations in patients (Mortiboys et al., 2010) and preclinical models (Singh et al., 2021; Liu et al., 2021) further predispose these cells to dysfunction and ultimately degeneration. Utilizing GO, we again found many pathways displaying significant differential expression between Lrrk2 mutants and controls. Among these, the top pathways are all associated with synaptic organization and functions (Figure 8E). This was particularly intriguing given that pre-synaptic dysfunction in Lrrk2 mutants such as endocytosis (Nguyen and Krainc, 2018; Soukup et al., 2016; Islam et al., 2016) or axonal cargo trafficking impairments (Boecker et al., 2021) could underpin the deficits in DA signaling observed in these mice (Skelton et al., 2022; Pischedda and Piccoli, 2021). Of note, decreased endocytosis has been observed in DA neurons but not cortical or hippocampal LRRK2^G2019S primary neurons (Pan et al., 2017). These results, although in vitro, suggest cell type-specific synaptic effects on DA neurons, consistent with their vulnerability in LRRK2-related PD.

Given that global gene expression differences across our conditions appear to hold relevance for putative mechanisms of PD pathogenesis, we next sought to address the question of whether gene expression within any given subtype is specifically altered in our PD model. Although Lrrk2 mRNA seemed uniform across subtypes, downstream effects could be subtype-specific secondary to differential Lrrk2 kinase activity, distinct kinase targets, or superimposed differences in other properties of each subtype that culminate in distinct downstream transcriptional alterations. Given the proposed vulnerability of Sox6 + DA neurons and relative resilience of Calb1 + neurons in PD (Gaertner et al., 2022; Kamath et al., 2022; Pereira Luppi et al., 2021), we first compared these two cluster families across Lrrk2 mutants versus control mice to see if any changes in DA neurons are specific to vulnerable or resistant cell types. We found 729 DEGs among the Sox6 family, and 679 among the Calb1 family (BH corrected p-value <0.05). Among those genes, we found 327 DEGs were shared between the Sox6 and Calb1 families (Figure 8—figure supplement 1B). The full list of DEGs and statistics is available in Supplementary file 2. Notably, AC149090.1 was highly significant in both cohorts, implying that the pan-DA differential expression of this gene was not driven by a particular subset of DA neurons. To further compare changes within cluster families, we once again performed GSEA and GO but on groups of clusters in isolation (Figure 9A–B). In line with our previous results, GO revealed pathways that largely corresponded to synaptic and axonal function, but with higher significance of results in the Sox6 family than in Calb1 or Gad2 families (Figure 9A). Within the Sox6 family, GSEA analysis showed an upregulation of energy production and metabolism pathways similar to those observed at the global level, as well as a downregulation of intercellular communication pathways (Figure 9B, top five upregulated and downregulated pathways shown for each family). Finally, to extend the granularity of our analyses, we applied the same GO pipeline to two individual clusters, Sox6^Tafa1 and Calb1^Stac, which have the highest expression of Anxa1 in the SNc and VTA, respectively (Figure 9—figure supplement 1A). The Sox6^Tafa1 SNc subtype was of particular interest to us given recent results showing that Anxa1 + ventral tier SNc neuronal activity is selectively correlated with acceleration in mice running on a treadmill, leading to the hypothesis that degeneration of these neurons may contribute to motor deficits seen in PD (Azcorra et al., 2023). The Calb1^Stac VTA subtype was chosen as a comparator due to its similar expression profile, including key markers Anxa1 and Aldh1a1, but within the relatively resistant VTA. Similar pathways were observed in each subtype, including several pathways once again corresponding to synaptic function, however with substantially higher significance values observed in the Sox6^Tafa1 cluster. Of note, while this enrichment is intriguing, the associations to such pathways cannot be interpreted as unique to this population, as technical limitations confound direct comparisons of DEGs among individual clusters due to variable cluster sizes and internal heterogeneity (see Methods for more details).

Next, we compared the synaptic compartment and biological functions of DEGs (BH adjusted p-values <0.05) in Lrrk2^G2019S mutants versus controls using SynGO, a database for systematic annotation of synaptic genes (Koopmans et al., 2019). We focused primarily on the Sox6 and Calb1 families, but also on cluster Sox6^Tafa1 given that this population (SNc neurons with highest expression of Anxa1) is specifically linked to movement (Azcorra et al., 2023). Of the DEGs in the Sox6 family, Calb1 family, and Sox6^Tafa1 cluster, 24.82%, 20.19%, and 29.41% of these genes had synaptic localizations, respectively, with more significant synaptic associations among populations localized to ventral tier SNc (Figure 9C, Figure 8—figure supplement 1C). Further compartment categorization revealed that among these synaptic DEGs annotated by SynGO, 45.83% of Sox6 family (11.38% of total DEGs), 64.0% of Sox6^Tafa1 (18.82% of total DEGs) and 48.82% of Calb1 family (9.86% of total DEGs) DEGs have presynaptic localizations, raising the possibility of enhanced disruption of presynaptic function in acceleration-correlated DA neurons. From these presynaptic DEGs, 20/77 in the Sox6 family, 5/16 in Sox6^Tafa1, and 16/62 in the Calb1 family have specific annotations for active zones, the primary sites of evoked DA release. Annotations based on the biological functions showed that 20.83%, 28.0%, and 15.74% of Sox6, Sox6^Tafa1, and Calb1 DEGs, respectively, are associated with processes in the presynapse (Figure 8—figure supplement 1C). Overall, greater proportions of DEGs are associated with presynaptic locations in cells from vulnerable DA neurons (Sox6 family, and in particular, Sox6^Tafa1), compared to less vulnerable ones (Calb1 family).

Given the differential effects of Lrrk2 mutation across subtypes, we also sought to map the association of gene expression in our dataset with co-expression of PD-associated risk loci identified by GWAS (Nalls et al., 2019) using a polygenic enrichment score method, scDRS (Zhang et al., 2022a). We first calculated a PD GWAS risk score for each individual cell, which revealed clear bias for higher risk scores among clusters that localize to the ventral SNc (Figure 9F). We next sought to extend the cell-level calculations of scDRS to explore the relative PD risk among different subtypes. To do so, we calculated the mean PD GWAS risk score for each cluster and utilized bootstrapping to generate a corresponding 95% confidence intervals for these mean values, allowing us to determine if the observed means were significantly different from the null. We confirmed significant risk associations in several clusters, particularly those in the Sox6 family (Figure 9G). Interestingly, the Gad2 family of clusters all showed significant negative associations with PD risk loci (Figure 9G). Of note, the highest scDRS scores were in clusters Sox6^Tafa1 and Sox6^Vcan, the only two SNc clusters that express Anxa1. By calculating the average scores for each cluster family as a whole, as well as these two Anxa1-expressing SNc clusters, we found that only the Sox6 family (and Anxa1-expressing SNc clusters within it) show significant associations with PD risk loci at a population level (Figure 9H).

Finally, due to the large number of potential comparisons to be made across clusters, cluster families, and genotypes, we have also developed an online tool that allows for exploration of our datasets, which we have termed Dopabase (URL: dopabase.org). Utilizing this tool, users have the ability to access many analysis and plotting functions to continue to explore populations of interest.

Analysis done using Python (v 3.10.13 Wolf et al., 2018 - RRID:SCR_008394). Cells were clustered using the scanpy toolkit (Version #1.10.1 RRID:SCR_018139; https://scanpy.readthedocs.io/en/stable/). Outputs from CellRanger were read into Scanpy using scanpy.read_10 x_mtx(). Cells were filtered with the same parameters as described above in the Seurat workflow. Datasets were individually clustered using the Leiden clustering algorithm. The four datasets were then integrated using scanpy.tl.ingest.(). Each dataset was mapped onto Control 1. The exact workflow can be found at the Awatramani Lab Github page.

To assess the molecular signature similarity between the clusters from Control and Lrrk2^G2019S samples, we first corrected for dropouts the expression matrix of each sample using Adaptively-thresholded Low Rank Approximation (ALRA) (Linderman et al., 2022). Then, using MetaNeighbor standard pipeline (Crow et al., 2018), we picked the intersect of variable genes across all but the top decile of expression bins for Control and Lrrk2 samples. On the 5282 selected genes, MetaNeighbor was performed to assess an AUROC score of clusters similarity. The exact workflow can be found at the Awatramani Lab Github page.

To evaluate pathways and gene sets enriched in subtypes or across conditions, we utilized the fgsea (v3.19; https://bioconductor.org/packages/release/bioc/html/fgsea.html; RRID:SCR_020938) and clusterProfiler (v4.8.3; RRID:SCR_016884 – https://doi.org/10.18129/B9.bioc.clusterProfiler) R packages. Pathways for GSEA analysis (mouse canonical pathways gene sets) were obtained from MSigDB (v2023.2.Hs, 2023.2 .Mm; https://www.gsea-msigdb.org/gsea/msigdb; RRID:SCR_016863). Custom R scripts were created for visualization of GSEA results and are available in the deposited code for this paper. fgsea was run using the following parameters: nperm = 1000, stats = ranks, minSize = 25, maxSize = 500. Differential expression calculations for fgsea package input were performed using Seurat’s FindMarkers command to generate a ranked list of genes, with filtering out of low-expression genes based on a minimum of 10% detection in either population being compared.

For gene ontology, the biological processes gene sets were obtained using the org.Mm.eg.db R package (v3.19; https://bioconductor.org/packages/release/data/annotation/html/org.Mm.eg.db.html; RRID:SCR_002774), and GO (part of clusterProfiler package) was run using default settings with BH adjustment of p-values and utilizing FDR-adjusted significant (p<0.05) DEGs for the respective comparisons. A custom genetic background was set for every individual comparison calculated to ensure hypergeometric testing was not simply enriched for cell type specific pathways; for comparisons of subsets within the dataset (i.e. a family or cluster) across conditions, a minimum detection level of 10% of cells was used to define the genetic background. These same thresholds were applied to filter the DEG lists used as input for GO. Results were visualized using the enrichplot R package (v3.19; https://www.bioconductor.org/packages/release/bioc/html/enrichplot.html; RRID:SCR_006442).

The scDRS score represents the relative association for each individual cell’s expression profile (among all other cells in the dataset) with PD risk loci by utilizing the underlying SNPs and associations described in GWAS summary statistics. Gene expression data (‘corrected’ UMI counts) was extracted from the SCT assay. Expression matrices, cell metadata, and feature metadata were loaded using Pandas (v2.2.2; https://pandas.pydata.org/; RRID:SCR_018214) in Jupyter (Jupyter: v7.2.1; https://jupyter.org/install; RRID:SCR_018315) notebook. The data were converted into an AnnData object using the anndata package (AnnData v0.10.8; https://anndata.readthedocs.io/en/latest/; RRID:SCR_018209) and stored in HDF5 format. Normalization and scaling were performed using the Scanpy package (v1.8.1) (PMID: 29409532). Total counts for each cell were normalized to 10,000 reads using the sc.pp.normalize_total function and log-transformed using sc.pp.log1p. The sc.pp.highly_variable_genes function (parameters: min_mean = 0.0125, max_mean = 3, and min_disp = 0.5) was used to identify highly variable genes. Finally, the anndata object was scaled using the sc.pp.scale function, with a max_value parameter of 10.

GWAS summary statistics for Parkinson’s Disease risk were obtained from EBI GWAS Catalog Study ID: GCST009325 (PMID: 31701892) and processed using the MAGMA tool on FUMA’s online platform (v1.5.2; https://fuma.ctglab.nl/snp2gene; RRID:SCR_017521; RRID:SCR_005757) to generate a scored list (z-scores) of significant genes. Default settings were used with a 10×10 gene window. Scored gene list was processed using the munge-gs function of the scDRS package (v1.0.2; https://martinjzhang.github.io/scDRS/) to create.gs gene sets (PMID: 34431100) and loaded into the Jupyter Notebook using the scdrs.util.load_gs function. To align with the anndata object data, the gene set’s human genes were mapped to mouse orthologs and intersected with genes present in the dataset. Preprocessing was completed using the scdrs.preprocess function, binning genes by mean and variance with parameters n_mean_bin = 20 and n_var_bin = 20. Scoring for Parkinson’s disease was performed using the scdrs.score_cell function (parameters: ctrl_match_key = mean_var, n_ctrl = 1000, weight_opt = vs, return_ctrl_raw_score = False, and return_ctrl_norm_score = True).

Since scDRS does not include a native method for population-level p-values, we calculated a mean scDRS score for each cluster or family by averaging its component cells and assessed for significance by creating a bootstrapped confidence interval for these means. A score of 0 represents the null of no association between gene expression and PD risk loci, and thus if the 95% confidence interval does not overlap 0, the mean scDRS score for a given group can be regarded as significant as there is a less than 5% chance of the true group mean containing the null. Bootstrapped 95% confidence intervals for mean score within clusters or cluster families were calculated in R using 10,000 permutations for each group and plotted as mean scDRS score per group with CI represented as error bars.

Parkinson’s Disease MAGMA Analysis Ranked Genes dataset can be found at 10.5281/zenodo.13076447.

SynGO analysis protocol is described in detail at https://doi.org/10.17504/protocols.io.x54v92r94l3e/v1.

To integrate MERFISH data with snRNA-seq data we utilized the weighted-nearest neighbour algorithm defined in Stuart, Butler, et al (Stuart et al., 2019). Briefly, anchor points were established between the two datasets using the FindTransferAnchors and MapQuery functions in Seurat. This allowed for a projection of the MERFISH data into the scRNA-seq space and a transfer of cluster identity between the two. Importantly, a label transfer prediction score was assigned to each MERFISH cell and cells with scores greater than 0.5 were classified as high confidence cells. Whole genome expression was imputed onto MERFISH cells using snRNA-seq data using the TransferData function in Seurat using the snRNA-seq normalized data as a reference. Cluster identifications were compared between transfer data and MERFISH subclustering using a Sankeyplot in R. Briefly, the number of cells that transferred between two pairs of clusters were calculated and expressed as a proportion within each MERFISH cluster. Only flows >10% were included in the visualization. Spatial visualizations of DA subtypes were generated using the ImageDimPlot function in Seurat and overlaid on the Allen Brain Atlas mouse anatomical template or the Paxinos mouse brain coronal axis using affinity designer. Gene expression values were obtained using the ImageFeaturePlot function in Seurat. Code can be found on the Awatramani Lab Github page (https://github.com/AwatramaniLab/Gaertner_Oram_etal; copy archived at AwatramaniLab, 2025).