Cluster replicability in single-cell and single-nucleus atlases of the mouse brain Authors: Leon French, Hamsini Suresh, Jesse Gillis Published: 2026 ABSTRACT Single-cell RNA sequencing has advanced our understanding of cellular heterogeneity. Ensuring the replicability of identified cell clusters across studies is essential for determining their biological robustness. We assess the replicability of cell clusters identified in two large mouse brain atlases, one generated using single-cell RNA sequencing and the other with single nuclei. Both profile over 4 million cells and group them into over 5,000 clusters. Using transcriptome-wide neighbor voting, we identify 2,009 reciprocally matched cluster pairs with consistent spatial localization and coordinated gene expression, which were also observed in datasets from multiple species. Reciprocal clusters are enriched in the cerebellum, where lower diversity aids replicability, while the hypothalamus’s heterogeneity limits agreement. Distinguishing close clusters is much more challenging than differentiating a cluster from most others, especially when using marker genes. By incorporating replicability data, we provide a stronger foundation for investigating the atlas-defined clusters and their biological meaning. Introduction The advent of single-cell RNA sequencing has significantly revolutionized our understanding of cellular diversity in the brain1,2. Large-scale transcriptomic atlases provide invaluable resources for dissecting cellular heterogeneity3. Validating atlas consistency is vital for their practical application in neuroscience research4. Two recent mouse brain atlases, each independently identifying over 5000 cell clusters using single-cell and nucleus sequencing, underscore the magnitude of this challenge5,6. Assessing cluster replicability, or the degree to which cell clusters identified in one study are consistent with those found in another, is critical to any claim that they represent reproducible biology, such as cell types7. Prior work by the BRAIN Initiative Cell Census Network (BICCN) has identified 70 highly replicable cell types in the mouse primary motor cortex8. This analysis leveraged seven transcriptomic datasets that differed in preparation methods (single nuclei vs. whole cells) and sequencing platforms (10x Genomics v2, v3, and SMART-Seq). Notably, the authors observed that replicability decreased as cells were subdivided into finer partitions. While single-nucleus sequencing detects fewer transcripts per cell, previous investigations have found it to possess comparable sensitivity to single-cell sequencing for identifying cell types in brain tissue9–11. However, these comparisons were limited to fewer than 12 coarse cell types, and single-nucleus sequencing may not be sufficient to detect disease-associated cell states12. From an evolutionary perspective, six retinal cell classes were shown to be highly conserved across 17 species, with transcriptomic similarity decreasing as evolutionary distance increased13. These observations raise the following questions: Do these findings extend to the thousands of cell clusters identified in the two recently published whole-brain studies? If so, what are the properties of these replicable clusters? This study deeply characterizes the replicability of cell clusters identified in the two central BICAN reference large-scale mouse brain atlases. We systematically evaluate cluster similarity through marker gene analysis and MetaNeighbor, an expression learning formalism using transcriptome-wide neighbor voting between all cells14,15. We further characterize the replicable clusters, focusing on their regional enrichment, spatial alignment, coordinated gene expression, and representation across diverse mouse and cross-species brain datasets. By rigorously assessing cluster agreement, we aim to highlight both the limitations and strengths of large-scale transcriptomic atlases in capturing cellular heterogeneity as well as providing a firm foundation for subsequent work. Results Basic properties of the single-cell mouse brain transcriptomic atlases We focused on the recently published single-cell and single-nucleus atlases that transcriptomically profiled the entire mouse brain5,6. The single-cell atlas contains transcriptomic profiles of over 4 million cells grouped into 5322 clusters. Similarly, the single-nucleus atlas assayed expression in 4.4 million nuclei that have been grouped into 5030 clusters (Fig. 1a). Both atlases used recursive graph-based clustering with marker-based stopping rules: the single-cell atlas enforced global separability between all cluster pairs, whereas the single-nucleus atlas retained only clusters with at least three discrete markers relative to other clusters within the same branch of the clustering hierarchy. The broad regional source of the cells varies, with more cells sampled from the thalamus, midbrain, brainstem, and cerebellum in the single-nucleus dataset while the single-cell atlas captured more cells in the hypothalamus and cerebral cortex (Fig. 1b). Similar differences were observed when comparing the number of clusters detected per dissection region (Fig. 1b). The number of clusters appearing within the dissection regions showed a higher correlation (Spearman rho = 0.91) between the two studies compared to cell counts (Spearman rho = 0.71). This suggests agreement in the degree of cellular heterogeneity across broad regions despite sampling differences. The cluster sizes in both atlases range from 9 to 264,669 cells in the single-cell dataset and 15 to 552,018 in the single-nucleus dataset. Consequently, a small fraction of the clusters contain most of the cells with a higher concentration of nuclei compared to the cells (Fig. 1c). For example, the ten largest clusters contain 28.9% and 52.4% of the cells in the single-cell and single-nucleus datasets, respectively. Overall, while both atlases provide extensive coverage of the mouse brain, they exhibit differences in cell sampling and cluster distribution, with a significant portion of cells concentrated in a few large clusters.Fig. 1Markers are insufficient for aligning single-cell and single-nucleus whole-brain atlases.a Summary of the two whole-brain-mouse atlases. b Bar plots showing the distribution of cells and detected clusters across dissection regions in the atlases. c Cumulative proportions of cells and clusters ranked by size, shown for single-cell (blue) and single-nucleus (red) clusters. d Schematic illustrating how author-provided markers are tested for detection of their target cluster cells relative to all clusters and similar clusters within an atlas. e Distribution of AUROC values for detecting target clusters relative to all cells (left, red) and relative to cells from the most similar non-target clusters (right, orange), shown for both atlases combined. Mean AUROC values are marked with dotted lines. f Scatter plot showing average gene expression (log2(CPM + 1)) in cells versus nuclei when averaged across cluster centroids. Each dot represents a gene, showing a strong correlation that reflects the similar overall expression profiles between the two studies (Spearman’s r = 0.98). g Heatmap of genome-wide expression profile correlations between cluster centroids from both atlases. h Euler diagram depicting the overlap between clusters identified by Liu et al. (yellow), Winter et al. (purple), and Zu et al. (pink) when mapped to the single-cell atlas. Anatomical template image is from the Allen Mouse Brain Reference Atlas42. Source data are provided as a Source data file. Markers are insufficient for aligning single-cell and single-nucleus whole-brain atlases. a Summary of the two whole-brain-mouse atlases. b Bar plots showing the distribution of cells and detected clusters across dissection regions in the atlases. c Cumulative proportions of cells and clusters ranked by size, shown for single-cell (blue) and single-nucleus (red) clusters. d Schematic illustrating how author-provided markers are tested for detection of their target cluster cells relative to all clusters and similar clusters within an atlas. e Distribution of AUROC values for detecting target clusters relative to all cells (left, red) and relative to cells from the most similar non-target clusters (right, orange), shown for both atlases combined. Mean AUROC values are marked with dotted lines. f Scatter plot showing average gene expression (log2(CPM + 1)) in cells versus nuclei when averaged across cluster centroids. Each dot represents a gene, showing a strong correlation that reflects the similar overall expression profiles between the two studies (Spearman’s r = 0.98). g Heatmap of genome-wide expression profile correlations between cluster centroids from both atlases. h Euler diagram depicting the overlap between clusters identified by Liu et al. (yellow), Winter et al. (purple), and Zu et al. (pink) when mapped to the single-cell atlas. Anatomical template image is from the Allen Mouse Brain Reference Atlas42. Source data are provided as a Source data file. While effective in the whole brain context, marker genes do not separate similar clusters Since marker genes are commonly used to annotate cell types, we evaluated the specificity and recurrence of the author-provided marker lists before examining genome-wide expression profiles of the atlases. The median number of markers per cluster was four for both atlases, with 3853 unique genes listed as markers. To test the ability of these markers to highlight their target cluster, we ranked every cell by the average z-scored marker expression and assessed whether target-cluster cells ranked higher than other cells using the area under the receiver operating characteristic curve (AUROC) (Fig. 1d). This statistic can be interpreted as the probability that the average gene expression of the markers will rank a cell of the target cluster higher than any other cell. As expected, the markers rank the target cells highly with mean AUROC values of 0.945 and 0.926 for the single-cell and single-nucleus datasets, respectively (combined distribution shown in Fig. 1e). To assess whether markers were most specific to their intended cluster, we next compared AUROC values across all clusters. Only 304 single-cell and 287 nuclei target clusters had the highest AUROC value when all clusters are tested. In other words, marker sets were often more specific to an off-target cluster. Furthermore, the top-ranked clusters collapsed to fewer than 1,000 unique clusters when all marker sets were tested within an atlas. To test whether marker sets could locally discriminate between similar clusters, we restricted the analysis to cells from the target cluster and the best off-target cluster (Fig. 1d). In this local setting, cells from target clusters were ranked below the best off-target cluster on average (mean AUROC = 0.27 to 0.28; combined distribution shown in Fig. 1e). Langlieb et al. acknowledged the difficulty in finding markers that can effectively discriminate between very similar clusters and, therefore, provided markers that also identify the cluster neighbours6. However, because our analyses focused on assessing how specifically markers identify individual clusters and distinguish them from all other clusters (including close neighbours), we used only their markers that were obtained without relaxed thresholds. Overall, these findings underscore the challenge of using a few markers for cluster identification. Having characterized the specificity of markers within the individual atlases, we next compared the similarity of clusters across datasets by testing the overlap of the marker gene lists. In this cross-atlas marker analysis, if the same clusters are in both atlases, then their markers should overlap. Using a Bonferroni correction for all 23.8 million pairwise cluster comparisons, 30 cluster pairs showed a statistically significant overlap of markers (hypergeometric test). Relaxing the multiple-testing correction to only the 4.8 million cluster pairs that share the same predominant brain region yielded 8 additional significant pairs. A further, more permissive threshold that corrects only for the cluster count in each dataset (~5000) identified 1826 pairs of clusters with overlapping markers. This limited agreement between marker genes suggests that genome-wide expression profiles are better suited to examine cluster similarity. To broadly gauge gene expression similarity between the atlases, we next compared average expression at different levels. Across the genome, when averaged across cluster centroids, the correlation between single-cell and single-nucleus atlases is high (Spearman’s rho = 0.98, Fig. 1f). In this comparison, most genes fall slightly above the diagonal, indicating higher counts per million (CPM) values in nuclei than in whole cells. This reflects the relative nature of CPM as a measure of proportion: the depletion of cytoplasm-enriched transcripts in the nuclear fraction results in a proportional increase in the remaining genes. Conversely, strongly cytoplasmic transcripts appear as outliers below the diagonal, with higher expression in the single-cell atlas. Correlating average genome-wide expression profiles across clusters shows a clear grouping within each atlas (Fig. 1g). Though there is strong per-gene correlation in average expression at the atlas level, the differences between clusters are not strong enough to override atlas-specific signals. Three companion studies from the BRAIN Initiative Cell Census Network 2.0 have aligned their single-cell clusters to the single-cell atlas of the mouse brain using advanced integration approaches. Liu et al. mapped 87.7% of the single-cell atlas clusters to at least one of their 4673 cluster-by-spatial cell groups, which were derived from single-cell methylomes. However, most of these mappings were ambiguous between the two datasets, with only 145 one-to-one cluster pairs identified. In a second study, which examined chromatin accessibility across the mouse brain, 19.5% of the single-cell atlas clusters were mapped to 1482 distinct brain cell populations from the snATAC-seq data (Zu et al.). After integration, the median proportion of cells annotated to the best-matched single-cell atlas clusters was 20.7%. A third transcriptomic study focused on spinal projecting neurons from eight regions (Winter et al.). It found that the 76 types of projecting neurons mapped to 291 of the 5322 single-cell atlas clusters. This mapping was also ambiguous, with no one-to-one mappings. By intersecting the mapped sets of clusters from these three studies, we tested the replicability of specific single-cell atlas clusters. Only 32 clusters were consistently mapped across all three studies, fewer than expected by chance (expected: 48.7, empirical p < 0.01, Fig. 1h). A larger overlap appeared when excluding the projection study, which focused on neurons. Specifically, 934 scRNA-seq clusters replicated in the brain-wide epigenetic studies (expected = 889, p < 0.001). The limited and ambiguous mappings from advanced integration methods motivated our comparison of the single-cell and single-nucleus transcriptomic mouse brain atlases. MetaNeighbor reveals cross-atlas cluster relationships Guided by the limits of marker genes and previous integration results, we used MetaNeighbor to assess replicability of the clusters in the two atlases15. As diagrammed in Fig. 2a, MetaNeighbor builds a cell-cell similarity network weighted by the Spearman correlation across a set of highly variable genes (HVGs). Cross-dataset neighbor voting based on this network is then used to provide interpretable cluster similarities (AUROC scores). Effectively, each cell is asked how well it matches to all cells in the other dataset, and this is compared with other cells within the same dataset, correcting for global data shifts. Importantly, MetaNeighbor has been optimized for large single-cell datasets14. The resulting AUROC heatmap from MetaNeighbor, when run on all 8.3 million cells, 3534 HVGs, and 10,352 clusters, shows a large block of very similar neuronal clusters. After expanding contrast for visualization of high AUROC values, a set of diagonal blocks of replicable clusters from both datasets is evident (Fig. 2b). However, at this scale, this heatmap does not capture the differences between very similar clusters because each pixel represents the AUROC values of several clusters due to resolution limits. This is clear in an expanded view of the 246 non-neuronal clusters (Fig. 2b). When zoomed in, clear blocks of similar glial clusters that span datasets contain smaller sub-blocks, indicating hierarchical structure. While similar clusters are apparent in this global comparison, focused analysis is needed to evaluate highly concordant clusters.Fig. 2Identification of reciprocally best-matched cell type clusters between the two brain atlases using MetaNeighbor.a Schematic showing the unsupervised MetaNeighbor framework used to select replicable clusters. Cell type-specific expression profiles learned from a reference dataset (single-cell) are used to predict similar cells in the target dataset (single-nucleus) and vice versa. Reciprocal top hits between the two datasets represent a set of highly replicable cell clusters between the atlases despite the compositional and technical variability in the underlying data. b Heatmap showing “one_vs_all” MetaNeighbor scores for 10,352 clusters (left), and for 246 select non-neuronal clusters (right) obtained by zooming into the bottom right corner of the complete heatmap. Clusters are labeled by dataset (top) and class for the single-cell clusters (left). Transcriptional similarity between pairs of clusters is depicted using a color scale that ranges from yellow to red for a narrow range of high replicability scores (0.8 to 1 AUROC) for better contrast. c, d, Scatterplots showing the relationship between global and local AUROC scores, using the single- cell (left) and nuclei (right) clusters as the references. The x-axis displays global scores (best one-vs-all AUROC) on a logarithmic scale, where 1 represents perfect identification. The y-axis shows local scores (best-vs-next AUROC) for the same reference clusters. Points are colored based on the density of neighboring points. e, f Histograms indicating the best-vs-next AUROC scores for non-reciprocal (darker shade) and reciprocal top hit (brighter shade) clusters from the perspective of the single-cell (blue) and nuclei (red) based reference datasets. g Bar plots showing the number of reciprocal top hit clusters across different size bins in the single-cell (blue) and single-nucleus (red) atlases. h Scatter plot comparing the log-scaled sizes of reciprocal top hit pairs between the two atlases (linear regression fit shown as blue line, grey dashed line represents the line of equality, and Spearman correlation coefficient at top left). i Bar plots marking the count of total clusters (full bar height) and reciprocal top hit clusters in each dataset (colored bars). Source data are provided as a Source data file. Identification of reciprocally best-matched cell type clusters between the two brain atlases using MetaNeighbor. a Schematic showing the unsupervised MetaNeighbor framework used to select replicable clusters. Cell type-specific expression profiles learned from a reference dataset (single-cell) are used to predict similar cells in the target dataset (single-nucleus) and vice versa. Reciprocal top hits between the two datasets represent a set of highly replicable cell clusters between the atlases despite the compositional and technical variability in the underlying data. b Heatmap showing “one_vs_all” MetaNeighbor scores for 10,352 clusters (left), and for 246 select non-neuronal clusters (right) obtained by zooming into the bottom right corner of the complete heatmap. Clusters are labeled by dataset (top) and class for the single-cell clusters (left). Transcriptional similarity between pairs of clusters is depicted using a color scale that ranges from yellow to red for a narrow range of high replicability scores (0.8 to 1 AUROC) for better contrast. c, d, Scatterplots showing the relationship between global and local AUROC scores, using the single- cell (left) and nuclei (right) clusters as the references. The x-axis displays global scores (best one-vs-all AUROC) on a logarithmic scale, where 1 represents perfect identification. The y-axis shows local scores (best-vs-next AUROC) for the same reference clusters. Points are colored based on the density of neighboring points. e, f Histograms indicating the best-vs-next AUROC scores for non-reciprocal (darker shade) and reciprocal top hit (brighter shade) clusters from the perspective of the single-cell (blue) and nuclei (red) based reference datasets. g Bar plots showing the number of reciprocal top hit clusters across different size bins in the single-cell (blue) and single-nucleus (red) atlases. h Scatter plot comparing the log-scaled sizes of reciprocal top hit pairs between the two atlases (linear regression fit shown as blue line, grey dashed line represents the line of equality, and Spearman correlation coefficient at top left). i Bar plots marking the count of total clusters (full bar height) and reciprocal top hit clusters in each dataset (colored bars). Source data are provided as a Source data file. We next summarize the full AUROC matrix by examining the best cross-dataset AUROC scores for each cluster. Almost all clusters identify another in the other atlas with an AUROC value above 0.95. Exceptions are 29 clusters in the single-cell atlas, which are enriched for oligodendrocyte-lineage clusters (n = 12, pFDR < 0.0001) and intratelencephalic/extratelencephalic glutamatergic neurons (n = 11, pFDR < 0.0001). Only two excitatory single-nucleus neuronal clusters marked by Foxp2, Myzap, and Spsb1 lack a hit with AUROC over 0.95. The median size of these clusters is above the atlas averages, suggesting these are not rare or undersampled populations. Mirroring the marker set results, high AUROC scores, which are relative to all cells, are not repeated when close clusters are compared. This is measured using the “best-vs-next” AUROC score, which tests, for a given reference cluster, if cells from the two most similar clusters in the other atlas can be discriminated based on expression similarity to the cells from the reference cluster. The average of this local AUROC score is 0.598 for single-cell and 0.627 for the single-nucleus reference clusters (Fig. 2c, d). Between the two atlases, we identified 2009 cluster pairs as reciprocal top hits, meaning they preferentially have the best AUROC scores with each other (Supplementary Data 1). When compared to their most similar clusters, these pairs have higher average “best-vs-next” AUROC scores than non-reciprocal top hits, exceeding 0.65 (Fig. 2e, f). Overlap between the author-provided marker lists is limited: only 46% of reciprocal hits share one or more markers (mean overlap of 0.54 genes), and only 25% of non-reciprocal hits have any overlap (0.28 average gene overlap). Beyond the author-provided markers, we used the MetaMarkers framework to select recurrent genes that distinguish both sides of each reciprocal pair16; this analysis yielded fewer than 6% of clusters that could be locally discriminated with an AUROC greater than 0.95 using four marker genes (Supplementary Note 1 and Supplementary Fig. 1). The resulting MetaMarkers signatures contain a median of 74 recurrent markers per reciprocal pair (range 15–990 genes, with the largest signature for microglia). While these results indicate that reciprocal top hits capture transcriptomic correspondence, reliable local discrimination of these clusters is difficult when using small sets of marker genes. We next tested how sensitive the reciprocal top hits are to dataset size and HVG selection. First, we randomly split each atlas into two subsets of 2,021,488 cells (corresponding to half of the cells in the smaller single-cell atlas). Within each atlas, MetaNeighbor correctly and reciprocally paired 80.6% of single-nucleus clusters (4054 of 5030) and 87.9% of single-cell clusters between the halves (4678 of 5322). These pairs establish a practical upper bound for cross-dataset alignment. Relative to these ceilings, the 2009 reciprocal pair count represents 49.6% of single-nucleus and 42.9% of single-cell clusters. By comparison, applying the same split-half analysis to a human single-cell retina atlas containing 3.2 million cells showed near-complete recovery, with only one of 123 cell types failing to reciprocally best match between halves17. The correct within-dataset pairs from the mouse brain atlases were enriched for the 2009 reciprocal pairs identified in the cross-dataset analyses, recovering 93.9% to 94.6% of them (p < 0.0001). Within the broader single-nucleus metaclusters, the correctly paired clusters were enriched in the inhibitory olfactory neuron metacluster, whereas pontine and medullary neuron metaclusters were less likely to match between the halves (Supplementary Data 2). At the higher class level in the single-cell dataset, the correct pairs were overrepresented in cortical neuron classes and depleted in the astrocyte-ependymal and classes containing glutamatergic neurons from the pons and medulla (Supplementary Data 3). Across the four cross-atlas comparisons between split halves, the number of reciprocal top hits was only slightly lower than in the full analysis (ranging from 1945 to 1987), and on average 80.6% of the 2009 full reciprocal pairs were recovered. These split-half comparisons establish a replicability ceiling for comparisons between two large, methodologically similar mouse brain atlases. To test sensitivity to HVG selection, we extracted the same number of HVGs using two alternative methods and also evaluated a random half of the MetaNeighbor HVGs. Using Seurat v318, only 38.8% of the genes overlapped with the MetaNeighbor set, yet we recovered a comparable number of reciprocal top hits (1973), 77.9% of which were identical to the full set. Applying MixHVG19 (on a random 5% of cells, due to R memory limits) yielded 44.9% gene overlap and 1899 reciprocal top hits, of which 1549 matched the full set (77.1%). Randomly removing half of the MetaNeighbor HVGs produced only 10 fewer reciprocal top hits, with 90.6% of them matching the full set of pairs. Finally, to refine the reciprocal pairs, we applied stricter thresholds to identify a high-confidence subset (Supplementary Data 4) and explored merging clusters to increase the number of matches. However, these approaches either reduced coverage or yielded only marginally more reciprocal top hits (Supplementary Note 2). We therefore focused on the 2009 reciprocal top hit pairs in subsequent analyses, as both their number and identity are robust to dataset size and HVG selection. Reciprocal top hits are observed across the range of cluster sizes. Larger single-nucleus clusters are more likely to be reciprocal, with a median size of 95 cells compared to 70 cells for non-reciprocal hits. In contrast, smaller single-cell clusters are more likely to be reciprocal, with a median size of 106 cells versus 134 for non-reciprocal hits. In aggregate, the majority of profiled cells are found in reciprocal best hit clusters, with 76.2% of nuclei and 50.2% of cells in these clusters (Fig. 2g). Although MetaNeighbor’s AUROC scores are independent of cluster size, the similar sizes of paired reciprocal clusters suggest comparable sampling of these paired clusters (Spearman r = 0.51, Fig. 2h). Analysis of the local context shows that larger reference clusters are more effective at distinguishing similar clusters in the other atlas. Specifically, reference clusters with more than 100 cells demonstrate better discriminatory power, with 36.4% having statistically significant “best-vs-next” AUROC scores (Wilcoxon rank-sum test, Bonferroni correction, Supplementary Fig. 2). In contrast, only 25.9% of the smaller clusters pass this threshold. Additionally, large reference clusters from the single-cell atlas demonstrate better cross-study discrimination ability compared to single-nucleus clusters. Overall, local discrimination favors larger reference clusters, particularly those from the single-cell atlas. We next intersected our reciprocal hits with the clusters found in separate integration efforts to test whether they were more likely to be observed. In the single-nucleus study, 49.2% of the 1937 clusters detected in the companion spatial Slide-seq experiment were reciprocal top hits, exceeding the expected 40% (p < 0.0001, Fig. 2i). For the three studies that mapped to the single-cell atlas, the Liu et al. mappings were depleted for reciprocal hits (36%) while the Zu et al. (44.4%) and Winter et al. (54.3%) mappings were enriched (all p < 0.0001, Fig. 2i). These results indicate that while there is some variation, our reciprocal top hits are generally observed at or above the expected rates, reflecting moderate enrichment in independent integration efforts. Characterization of the reciprocal best hit cluster pairs To better understand the 2009 reciprocal best hits, we examined their placement in the broader cell type taxonomies and their locations in the mouse brain. We tested if the reciprocal clusters were enriched or depleted for specific classes from the Yao et al. single-cell-based taxonomy. Of the 34 classes, those containing clusters of GABAergic neurons from the medulla, midbrain, and pons; glutamatergic neurons from the medulla and midbrain; and immature GABAergic neurons from the olfactory bulb were enriched for reciprocal hits. In contrast, clusters from the oligodendrocyte-lineage, astrocyte-ependymal, intratelencephalic/extratelencephalic glutamatergic, cerebral nuclei and anterior hypothalamic neurons, and GABAergic neurons from the lateral septal nucleus and hypothalamus were depleted (hypergeometric test, pFDR < 0.05, Supplementary Data 5). Across the meta-cluster classification from the single-nucleus dataset from Langlieb et al. (n = 161 with ten or more clusters), astrocytes; and neurons annotated as “thalamic projection”, “brainstem”, “medullary”, “inhibitory olfactory”, “pontine and medullary”, “tegmental and pontine”, and “midbrain and pontine” were enriched. Depleted meta-clusters were groups of brainstem neuron clusters from the pons and medulla, fibroblasts, and cholinergic neurons of the cranial nerve nuclei (Supplementary Data 6). To directly assess the spatial agreement of cluster locations, we mapped dissection annotations from the atlases to 19 common regions. For the 2009 reciprocal best hit pairs, the dissection region containing the highest proportion of cells matched between datasets for 80.8% of the pairs. For the 996 cluster pairs with cells from two or more dissection regions, the top two regions with the highest proportion of cells matched for 45.6% of the pairs. As a comparison, a set of 972 best-matching cluster pairs based on the single-cell atlas markers exhibited lower matching rates—31.6% for the top region and 13.6% for the top two regions. To examine reciprocal top hit clusters at finer spatial resolution than dissection regions, we called each cell in the MERFISH dataset used by Yao et al. using the same method with the single-nucleus cluster centroids (Fig. 3a). For each cell, we tested if calls generated from the two separate atlases corresponded to a reciprocal best hit pair. For example, seven cells that map to both sides of the smallest reciprocal hit demonstrate spatial co-localization (Fig. 3b). In this case, three cells are annotated to the single-cell-based reciprocal cluster but could not be confidently called when using the single-nucleus data as a reference (average correlation score <0.5). Across all 4.3 million cells, 3.9 and 2.0 million have been assigned to a cluster based on single-cell and single-nucleus references, respectively. Of these, 768,245 cells are assigned to both sides of a reciprocal best hit pair. Visualization of these cells across four coronal sections (Fig. 3c) reveals a striking enrichment in the cerebellum and depletion in the middle cortical layers of the cerebral cortex. The cerebellar pattern is primarily driven by the most frequent cluster pair (Ex_Gabra6_Gabrd/cluster_ID:5201), which marks 151,723 granule cells. The second most abundant reciprocal cluster pair consists of 65,194 cells annotated as cerebellar GABAergic neurons (Inh_Tfap2a_Sla/cluster_ID:5188). Although these neurons are small, cell density does not appear to drive the enrichment, as gridded images that normalize for cell density show a similar spatial pattern (Fig. 3c, bottom).Fig. 3Distribution of reciprocally matched clusters in a spatial transcriptomic atlas of the whole brain.a Schematic illustrating the workflow used to align individual cells in MERFISH sections to clusters in the single-cell and single-nucleus atlases. b MERFISH coronal section (coronal slice 18) highlighting 7 cells that are mapped to both clusters of an example reciprocally matched pair. Additionally, four cells mapped only to the single-cell side of the cluster pair are marked in blue. c Representative MERFISH sections showing the spatial distribution of cells mapped to reciprocally matched cluster pairs at single-cell (top) and coarse grid-level (bottom) resolutions. Each section displays cells’ spatial locations colored by broad CCFv3 regions (left hemisphere), and the distribution(top)/proportion(bottom) of cells mapped to reciprocally matched cluster pairs (right hemisphere). d Line plot depicting the proportion of cells mapped to reciprocally matched cluster pairs (orange dots) and detected clusters (purple dots) within coronal sections ordered from anterior to posterior along the x-axis. Stacked bar chart (bottom) showing the distribution of the detected reciprocally matched clusters across the same slices, colored by broad CCFv3 regions. e Scatterplot showing the relationship between reproducible cell calls (y-axis) and the mismatch in cluster diversity (x-axis) for each slice. Reproducible cell calls are defined as the proportion of cells mapped to reciprocal pairs. The diversity mismatch represents the difference in the number of clusters detected using single-cell versus single-nucleus references. The Spearman correlation between these measures across the slices is −0.71 (top right). f Heatmap indicating the relative enrichment of cells in reciprocally matched clusters within aligned regions from the single-cell, single-nucleus, and MERFISH spatial atlas data. Columns (regions) are ordered from highest average enrichment (left) to depletion (right). g Density plots of the distances in mm between centroids of MetaNeighbor reciprocal top hits (top) and marker-derived (bottom) cluster pairs in the MERFISH dataset. Mean centroid distances are indicated by dashed lines. Anatomical template image is from the Allen Mouse Brain Reference Atlas42. Region abbreviations: ACA anterior cingulate area, AUD auditory areas, BS brainstem, CB cerebellum, CTXsp cortical subplate, ENT entorhinal area, HPF hippocampal formation, HY hypothalamus, LSX lateral septal complex, MB midbrain, MOp primary motor area, OLF olfactory areas, PAL pallidum, RSP retrosplenial area, SSp primary somatosensory area, STR striatum, STRd dorsal striatum, SUB subiculum, TH thalamus, VIS visual areas, VISp primary visual area. Source data are provided as a Source data file. Distribution of reciprocally matched clusters in a spatial transcriptomic atlas of the whole brain. a Schematic illustrating the workflow used to align individual cells in MERFISH sections to clusters in the single-cell and single-nucleus atlases. b MERFISH coronal section (coronal slice 18) highlighting 7 cells that are mapped to both clusters of an example reciprocally matched pair. Additionally, four cells mapped only to the single-cell side of the cluster pair are marked in blue. c Representative MERFISH sections showing the spatial distribution of cells mapped to reciprocally matched cluster pairs at single-cell (top) and coarse grid-level (bottom) resolutions. Each section displays cells’ spatial locations colored by broad CCFv3 regions (left hemisphere), and the distribution(top)/proportion(bottom) of cells mapped to reciprocally matched cluster pairs (right hemisphere). d Line plot depicting the proportion of cells mapped to reciprocally matched cluster pairs (orange dots) and detected clusters (purple dots) within coronal sections ordered from anterior to posterior along the x-axis. Stacked bar chart (bottom) showing the distribution of the detected reciprocally matched clusters across the same slices, colored by broad CCFv3 regions. e Scatterplot showing the relationship between reproducible cell calls (y-axis) and the mismatch in cluster diversity (x-axis) for each slice. Reproducible cell calls are defined as the proportion of cells mapped to reciprocal pairs. The diversity mismatch represents the difference in the number of clusters detected using single-cell versus single-nucleus references. The Spearman correlation between these measures across the slices is −0.71 (top right). f Heatmap indicating the relative enrichment of cells in reciprocally matched clusters within aligned regions from the single-cell, single-nucleus, and MERFISH spatial atlas data. Columns (regions) are ordered from highest average enrichment (left) to depletion (right). g Density plots of the distances in mm between centroids of MetaNeighbor reciprocal top hits (top) and marker-derived (bottom) cluster pairs in the MERFISH dataset. Mean centroid distances are indicated by dashed lines. Anatomical template image is from the Allen Mouse Brain Reference Atlas42. Region abbreviations: ACA anterior cingulate area, AUD auditory areas, BS brainstem, CB cerebellum, CTXsp cortical subplate, ENT entorhinal area, HPF hippocampal formation, HY hypothalamus, LSX lateral septal complex, MB midbrain, MOp primary motor area, OLF olfactory areas, PAL pallidum, RSP retrosplenial area, SSp primary somatosensory area, STR striatum, STRd dorsal striatum, SUB subiculum, TH thalamus, VIS visual areas, VISp primary visual area. Source data are provided as a Source data file. Across the coronal slices, the most anterior and posterior slices exhibit higher proportions of cells assigned to reciprocal best hit clusters (Fig. 3d). This proportion is inversely correlated with the diversity of clusters detected in each slice (Spearman r = −0.56, Fig. 3d). To further investigate this relationship, we downsampled the number of MERFISH cells identified using the single-cell reference to match the number of nuclei-based calls. We then calculated the difference in detected cluster counts between the two references. Across the slices, the excess detection of single-cell clusters in this diversity mismatch measure is negatively correlated with the proportion of reciprocal best hits (Spearman r = −0.71, Fig. 3e). This association suggests that differences in sampling depth and clustering resolution can influence cluster replicability between the two atlases. Across 19 aligned dissection regions, we find that the proportion of cells mapped to reciprocal best clusters is consistent with the finer-resolution MERFISH results. Specifically, cells in the cerebellum and the hippocampal formation are more likely to be mapped to reciprocal best clusters, while the hypothalamus and the nearby lateral septal complex are depleted for cells from reciprocal clusters (Fig. 3f). Spatial proximity of reciprocally matched clusters To test similarity between reciprocally matched clusters in the spatial domain, we calculated the distances between their centroids. The average distance between the centroids of a reciprocal best hit cluster pair is 0.428 mm. In contrast, for the 972 best-matching cluster pairs derived from markers, the average distance was significantly larger at 3.12 mm (Fig. 3g). Within the 2009 reciprocal top hit pairs, the next closest clusters from the MetaNeighbor analyses are also farther apart, with a mean distance of 0.95 mm (Wilcoxon rank-sum test, p < 0.0001). The close spatial proximity of these reciprocally matched clusters is notable, as spatial information was not used during the clustering or matching processes, making this convergence especially reassuring. Coordinated gene expression across the matched clusters To provide a genome-wide comparison of the reciprocally matched clusters, we compared their per-gene expression profiles across clusters. We measured coordinated expression across the paired cluster expression centroids (Fig. 4a). To avoid biasing the results with the same genes used for pairing, we excluded those 3534 HVGs, acknowledging that some dependency persists through co-expression. For the remaining 17,450 genes, the per-gene average Spearman correlation across the 2009 reciprocal clusters was 0.357 (Fig. 4b). We randomly sampled the number of paired clusters to test if this association is sensitive to the cluster count. The average correlation remained stable across the range of sizes, suggesting that smaller atlases would provide similar levels of coordinated expression. This agrees with our earlier analyses showing that reciprocal top hits are robust to both reduced cell numbers and alternative HVG sets. To test specificity, we examined whether different gene pairs have stronger correlations than the same gene pairs between the two atlases. The average percentile ranking of same gene pairs was 94.9%, which slightly decreased as cluster pairs were downsampled (Fig. 4c). For 34.8% of the genes, their coordinated expression was higher than all correlations between that gene and any other (Fig. 4d). This metric was the most sensitive to the number of clusters, dropping to 27% on average when downsampled to one-tenth of the original clusters. Unsurprisingly, we find that genes involved in synapse-related functions were enriched for higher coordinated expression (Fig. 4e and Supplementary Data 7). This analysis highlights the robustness and specificity of coordinated gene expression in the reciprocally matched clusters.Fig. 4Robust and specific coordinated gene expression in reciprocally matched clusters.a Schematic illustrating the calculation of coordinated expression across the centroids for a specific gene. b Boxplots showing the mean same-gene expression correlation between both sides of the matched cluster pairs (y-axis) across increasingly smaller subsamples of the matched reciprocal clusters (x-axis; n = 17,450 non-HVG genes). Twenty random samples were evaluated for each subsample size. c, d Specificity metrics for the same set of 17,450 genes, depicted through boxplots of the mean percentile of same-gene correlation and the proportion of genes with the highest same-gene correlation (y-axis) across the same cluster pairs. For boxplots (b–d), the center line indicates the median, box limits indicate the upper and lower quartiles, whiskers extend to the most extreme data points within 1.5× the interquartile range, and individual points denote outliers. e Scatter plot showing the relationship between the average coordinated expression across datasets and the average specificity of coordinated expression for 3932 GO-curated gene groups. Eighty-two SynGO-annotated gene groups are highlighted in red, indicating higher reproducibility of expression across datasets than other GO groups. Source data are provided as a Source data file. Robust and specific coordinated gene expression in reciprocally matched clusters. a Schematic illustrating the calculation of coordinated expression across the centroids for a specific gene. b Boxplots showing the mean same-gene expression correlation between both sides of the matched cluster pairs (y-axis) across increasingly smaller subsamples of the matched reciprocal clusters (x-axis; n = 17,450 non-HVG genes). Twenty random samples were evaluated for each subsample size. c, d Specificity metrics for the same set of 17,450 genes, depicted through boxplots of the mean percentile of same-gene correlation and the proportion of genes with the highest same-gene correlation (y-axis) across the same cluster pairs. For boxplots (b–d), the center line indicates the median, box limits indicate the upper and lower quartiles, whiskers extend to the most extreme data points within 1.5× the interquartile range, and individual points denote outliers. e Scatter plot showing the relationship between the average coordinated expression across datasets and the average specificity of coordinated expression for 3932 GO-curated gene groups. Eighty-two SynGO-annotated gene groups are highlighted in red, indicating higher reproducibility of expression across datasets than other GO groups. Source data are provided as a Source data file. Replicability of reciprocal clusters in additional mouse brain datasets Next, we assessed the identification of the 2009 reciprocally matched cluster pairs in several additional datasets. To enable fast and memory-efficient computation of cross-dataset cell type similarities, we computed pretrained models for the single-nucleus and single-cell datasets (Fig. 5a). Applying these models to quantify replicability is equivalent to a full MetaNeighbor analysis, but with the HVGs fixed by the atlases. We first applied these models to a spatial dataset that profiled expression of 109 genes in the mouse brain using the BARseq in situ sequencing technique20. Of the 135 clusters identified in the BARseq study across their taxonomy, 54 showed the highest similarity to a reciprocally matched cluster using the single-cell model (odds ratio (OR): 1.10, p = 0.59), and 88 using the single-nucleus model (OR: 2.82, p < 0.0001). Combining the best hits from both models revealed 41 clusters with best hits to both sides of a reciprocally matched cluster pair. We then examined the spatial distribution of these clusters across the BARseq, MERFISH, and Slide-seq datasets. All but one pairing displayed good spatial concordance (Supplementary Data 8). Specifically, one pair exhibited a restricted pattern in the MERFISH dataset (0065 L5 IT CTX Glut_5, BARseq:L6 IT L), while three single-nucleus clusters were not detected in the Slide-seq data. The atlases reduced annotation ambiguity, as several BARseq clusters with uncertain annotations exhibited concordant spatial patterns (Fig. 5b). Importantly, the spatial data were independent and not used in the clustering process in the original studies or our cross-dataset comparison. Despite BARseq’s limited gene panel, its spatial patterns corroborate the re-identification of the reciprocally matched cluster pairs.Fig. 5Enrichment of reciprocally matched clusters in independent spatial data and other single-cell studies.a Schematic illustrating the application of a pretrained MetaNeighbor model to the BARseq gene expression data. b The held-out spatial data allows independent comparisons of expression patterns. Three BARseq (black points, left hemispheres) clusters with uninformative labels exhibit highly similar spatial patterns when compared to the top hit cluster locations in the MERFISH (blue points, top right hemispheres) and Slide-seq (red points, bottom right hemispheres) datasets. c Bar plots displaying the number of clusters (left) with top hit matches to either single-cell or single-nucleus reciprocal clusters (black) in different datasets. Dotplot of the odds ratios of clusters mapping to single-cell (blue circles) or single-nucleus (red circles) reciprocal clusters (right). Datasets are marked with icons to indicate the profiling method. d Heatmap indicating the relative enrichment of cells in reciprocally matched clusters within aligned regions from other mouse brain datasets. Regions are listed in the same order as Fig. 3f.e Similar visualization as in (c), but for brain datasets from 16 non-mouse species. Evolutionary relationships are depicted in the phylogenetic tree (left). f Scatter plot showing the association between phylogenetic distance from mice (MYA: million years ago, x-axis) and the mean cluster similarity AUROC score for the top hits identified by the models. g Boxplots showing the distribution of cluster replicability AUROC scores for clusters with top hits to none (n = 435), one (n = 542), or both clusters (n = 215) in a reciprocally matched pair. The center line indicates the median, box limits indicate the upper and lower quartiles, whiskers extend to the most extreme data points within 1.5× the interquartile range, and outlier points are not shown. h Summary graph showing the reciprocal matched pairs as single white nodes connected to the target clusters with top hits to both sides of the reciprocal pair. Clusters in the target datasets are color-coded by species. Nodes are arranged and labelled based on broad transcriptomic cell types (ExN excitatory neuron, InN inhibitory neuron, MGE medial ganglionic eminence, CGE caudal ganglionic eminence). Anatomical template image is from the Allen Mouse Brain Reference Atlas42. Region abbreviations: BS brainstem, CB cerebellum, HPF hippocampal formation, HY hypothalamus, MB midbrain, OLF olfactory areas, PAL pallidum, STRd dorsal striatum, TH thalamus. Brain illustrations were adapted, with modifications, from ref. 43, which is licensed under CC-BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Public domain species silhouettes were sourced from PhyloPic (www.phylopic.org). Source data are provided as a Source data file. Enrichment of reciprocally matched clusters in independent spatial data and other single-cell studies. a Schematic illustrating the application of a pretrained MetaNeighbor model to the BARseq gene expression data. b The held-out spatial data allows independent comparisons of expression patterns. Three BARseq (black points, left hemispheres) clusters with uninformative labels exhibit highly similar spatial patterns when compared to the top hit cluster locations in the MERFISH (blue points, top right hemispheres) and Slide-seq (red points, bottom right hemispheres) datasets. c Bar plots displaying the number of clusters (left) with top hit matches to either single-cell or single-nucleus reciprocal clusters (black) in different datasets. Dotplot of the odds ratios of clusters mapping to single-cell (blue circles) or single-nucleus (red circles) reciprocal clusters (right). Datasets are marked with icons to indicate the profiling method. d Heatmap indicating the relative enrichment of cells in reciprocally matched clusters within aligned regions from other mouse brain datasets. Regions are listed in the same order as Fig. 3f.e Similar visualization as in (c), but for brain datasets from 16 non-mouse species. Evolutionary relationships are depicted in the phylogenetic tree (left). f Scatter plot showing the association between phylogenetic distance from mice (MYA: million years ago, x-axis) and the mean cluster similarity AUROC score for the top hits identified by the models. g Boxplots showing the distribution of cluster replicability AUROC scores for clusters with top hits to none (n = 435), one (n = 542), or both clusters (n = 215) in a reciprocally matched pair. The center line indicates the median, box limits indicate the upper and lower quartiles, whiskers extend to the most extreme data points within 1.5× the interquartile range, and outlier points are not shown. h Summary graph showing the reciprocal matched pairs as single white nodes connected to the target clusters with top hits to both sides of the reciprocal pair. Clusters in the target datasets are color-coded by species. Nodes are arranged and labelled based on broad transcriptomic cell types (ExN excitatory neuron, InN inhibitory neuron, MGE medial ganglionic eminence, CGE caudal ganglionic eminence). Anatomical template image is from the Allen Mouse Brain Reference Atlas42. Region abbreviations: BS brainstem, CB cerebellum, HPF hippocampal formation, HY hypothalamus, MB midbrain, OLF olfactory areas, PAL pallidum, STRd dorsal striatum, TH thalamus. Brain illustrations were adapted, with modifications, from ref. 43, which is licensed under CC-BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Public domain species silhouettes were sourced from PhyloPic (www.phylopic.org). Source data are provided as a Source data file. To further assess replicability, we applied the models to ten additional mouse brain datasets (Fig. 5c). These were nearly evenly split between the use of single-cell and single-nucleus sequencing. These datasets ranged from 4,000 to 385,000 sequenced cells, with the original studies clustering the cells into tens to hundreds of groups, depending on the dataset (Supplementary Data 9). The majority of cells (mean: 77.9%, range: 33.7–98.5%) and clusters (mean: 74.3%, range: 64.3–81.4%) were most similar to at least one side of a reciprocally matched cluster pair (Fig. 5c). Consistent with the BARseq findings, the top hits across datasets were enriched for the 2009 reciprocal clusters, with stronger enrichment from the single-nucleus model (median OR: 1.91) compared to the single-cell model (median OR: 1.16). While the odds ratios were comparable between single-cell and single-nucleus preparations using the single-cell model, the single-nucleus model showed significantly higher enrichment for reciprocal matches in single-nucleus studies (median OR: 1.72 vs. 2.59, rank-sum test p < 0.05). Despite the lack of fine spatial resolution in these studies, they focused on specific brain regions or used broad dissection areas. When mapped to the regions analyzed above, the cerebellum had the highest proportion of cells from replicable clusters (71%), while the hypothalamus had the lowest (21%) (Fig. 5d). A specific example is the Steuernagel et al. dataset that was focused on the hypothalamus and had the lowest proportion of cells aligned to reciprocal clusters (34%). The recurrence and enrichment of the reciprocally matched pairs across diverse datasets reinforce their utility in identifying replicable cell clusters across brain regions and sequencing techniques. Evolutionary conservation of reciprocal clusters Next, we applied the MetaNeighbor models to 19 brain datasets from 16 species to evaluate replicability from an evolutionary perspective. Across these datasets, the number of cells and clusters was comparable to those in mouse datasets (Supplementary Data 10). On average, 74.6% of clusters, containing 74.1% of cells, were most similar to at least one side of a reciprocally matched cluster pair when the model results are combined (Fig. 5e, cluster range: 62.6–91.9%, cell range: 44.5–97.5%). Mirroring the mouse results, the datasets were enriched for the 2009 reciprocal clusters, with a stronger enrichment from the single-nucleus model (median OR: 1.97) than the single-cell model (median OR: 1.06). For the single-cell model, odds ratios were higher in studies using single-cell preparations (median OR: 1.44 vs. 0.94, rank-sum test, p < 0.05), while the single-nucleus model showed similar enrichment across both preparation types (median OR: 1.92 for single-cell preparations vs. 2.36 for single-nucleus studies, rank-sum test, p > 0.3). The species analyzed spanned a broad evolutionary range, from humans to spiders. Notably, genetic distance from mouse was negatively correlated with mean cluster similarity scores for top hits identified by the models (Spearman correlation = −0.89, p < 0.0001, Fig. 5f). Linking these findings to replicability, top hit scores were associated with the number of reciprocal clusters. Specifically, top hits that were not reciprocal clusters have a median AUROC of 0.939, which increased to 0.964 for one reciprocal hit and 0.984 for clusters with reciprocal top hits from both models (Fig. 5g). Of the 1192 cross-species target clusters, 215 had top hits corresponding to both sides of a reciprocal top hit cluster. After removing hits with inconsistent annotations at the coarse level, 104 target clusters remained (Fig. 5h). Reviewing clusters with two reciprocal top hits in four or more species highlights microglia, as well as specific clusters of oligodendrocytes, oligodendrocyte progenitor cells, Chandelier cells, and Lamp5 interneurons (Fig. 5h). Although reduced by evolutionary distance, these results suggest that the atlases and our reciprocal clusters will aid cell type identification in diverse species. Discussion This study presents a comprehensive assessment of cell cluster replicability across two large-scale mouse brain transcriptomic atlases generated using single-cell and single-nucleus RNA sequencing. Our findings reveal substantial agreement between these datasets, with a significant proportion of clusters exhibiting reciprocal best hit relationships indicative of robust cross-atlas agreement. These replicable clusters demonstrate consistent spatial localization, coordinated gene expression patterns and are observed in independent mouse and cross-species datasets. However, our analysis also highlights the challenges associated with resolving fine-grained cellular heterogeneity and the limitations of relying solely on marker genes for cluster identification. To aid further research, we provide meta-analytic marker genes, pretrained models, and reciprocal best hit clusters, facilitating confident application of the atlases. Regional differences in cluster replicability, particularly the higher concordance observed in the cerebellum compared to the hypothalamus, likely arise from a combination of factors. The cerebellum’s granule cell population has limited gene expression that prevents splitting into small clusters. In addition, it has a higher proportion of nuclear transcripts due to its small cell size, which may enhance its replicability across atlases that vary in subcellular sampling10. In contrast, the hypothalamus’s diverse cellular composition poses a greater challenge for resolving fine-grained cell types and achieving consistent cross-atlas cluster identification. The diversity-associated reduction in replicability was also observed across coronal slices. This difficulty may be further compounded by technical biases in cell isolation and sequencing processes21. Such biases led to the inclusion of single-nucleus multiome sequencing data in the single-cell atlas to better capture cell populations in the midbrain and hindbrain regions5. Our region-specific findings emphasize the need to account for regional factors—such as cellular heterogeneity, cell size, and technical limitations—when using transcriptomic atlases and help identify which regions require deeper sampling to improve cluster quality. The identification of most clusters in spatial transcriptomics datasets, along with their consistent spatial distributions, aids the transition from transcriptomic clusters to defined cell types. Our findings, particularly the proximity of reciprocal best hit cluster centroids compared to similar clusters and the repeated spatial patterns across three datasets, reinforce this. Importantly, our analyses were strengthened by the independence of the spatial data from both the original clustering and our cross-atlas comparisons. However, this independence will likely be lost as spatial data is increasingly used to support the creation of transcriptomic cell-type taxonomies22. Our results align with NIH BRAIN Initiative Director John Ngai’s view that each cell cluster is a hypothesis to be tested23. Unfortunately, it will be difficult to experimentally target these clusters as our analysis revealed limitations in resolving them in independent datasets. Marker genes, while effective for broadly classifying cell types, often lacked the specificity to discriminate between closely related clusters. Consistent with this, reliance on single markers has been shown to mislead cross-species annotations24,25. Meta-analytic analyses of cortical neuron markers reinforce these limitations, suggesting 50–200 markers per cell type16. Similarly, the “best-vs-next” AUROC scores from MetaNeighbor indicated that even reciprocal best hits sometimes struggled to distinguish between highly similar clusters. Furthermore, the average single-cell study often attains lower sampling depth for both cells and genes, further restricting the identification of fine clusters. Specifically, most of the datasets we tested the models on had a lower mean count of detected genes per cell than the atlases. This is also evident in the limited ability of the epigenetic datasets to align with the single-cell atlas. While experimentally targeting small clusters may remain challenging, integrating spatial transcriptomics offers a promising avenue for refining transcriptomic cell-type definitions. Our findings have important implications for the development and utilization of large-scale brain transcriptomic atlases. While these resources provide invaluable insights into cellular heterogeneity, careful consideration of cluster replicability is crucial for cross-study application. In practice, atlas taxonomies appear most robust when emphasizing coarser, replicable cell populations. This is supported by our split-half analysis of a retina atlas that included data from eight studies17, in which all but one of the 123 cell groups matched between halves. In the brain atlases, annotations can also incorporate brain region information where supported by spatial distributions. Together, these atlases provide a robust foundation for a reference brain atlas, acknowledging that some clusters may ultimately need to be merged or split. For example, both atlases identify a single, highly replicable microglia cluster that is conserved across species. In contrast, single-nucleus studies of the human brain from donors without neuropsychiatric or neurological conditions, as well as those with Alzheimer’s disease, identify multiple microglia clusters, suggesting that finer groupings exist in mice26–28. In our analyses, many clusters do not match across atlases, possibly due to differences in subcellular and regional sampling, technical variation in sequencing methodologies, or divergent decisions about how to discretize continuous axes of transcriptomic variation into clusters. Previous work has shown that such continuous gradients are present in the brain and can be partitioned differently across studies, potentially contributing to misalignment between atlases29–31. While we did align clusters with fewer than 20 cells, many small clusters may represent over-splitting due to technical artifacts or stochastic gene expression differences that do not reflect distinct cellular identities. Ultimately, improving cluster replicability will be essential for accurately characterizing cell types. Enhancing replicability will also strengthen the broader utility of transcriptomic atlases in advancing our understanding of brain organization and function. Methods Single-cell gene expression data Single-cell sequencing data, MERFISH spatial data, and corresponding annotations were obtained from the whole-mouse-brain atlas by ref. 5. The study used 317 adult C57BL/6J (RRID:IMSR_JAX:000664) mice and assayed gene expression using 10x Genomics v2 and v3 kits, along with 1687 nuclei profiled with the 10x Genomics Multiome platform. Final cell-by-gene expression count matrices were normalized using log(CPM + 1) where necessary. Cell clusters annotated as low quality (LQ) were removed. Annotations for dissected regions were aligned to the single-nucleus atlas by manually mapping to similar or enclosing regions. For example, the hippocampus (HIP) was mapped to the hippocampal formation (HPF), while the pons (P), hindbrain (HB), and medulla (MY) were mapped to the brainstem (BS). We used Yao et al.‘s marker genes and cluster taxonomy annotations for our analyses. Single-nucleus expression data Single-nucleus gene expression data and annotations were obtained from the Langlieb et al. atlas6. Briefly, this dataset was generated from 92 samples from 55 adult C57BL/6J mice using the 10x Genomics (v3) kit. Cluster markers were sourced from the “Whole Brain Set Cover” sheet in Supplementary Table 7 of Langlieb et al.6, using only the rows where the “Exclude up to N Nearest Neighbors” parameter was set to zero. Slide-seq spatial images were retrieved from the braincelldata.org website. Integration results from other studies Integration results from the three companion studies in the BRAIN Initiative Cell Census Network 2.0 article collection were obtained from their supplementary tables. Specifically, the “Matched AIBS 10X RNA Clusters” column in Supplementary Table 4 of Liu et al. provided mappings for single-cell DNA methylome and 3D multi-omic clusters32. Cluster (cl) identifiers that were no longer present in the updated Yao et al. cluster annotations were excluded from our analysis. Integration results from Zu et al.’s study of single-cell chromatin accessibility were retrieved from their Supplementary Table 533. For each cluster, the one with the highest label transfer score was used, resulting in 1040 unique single-cell clusters. To identify clusters in a transcriptomic study of spinal projecting neurons, we converted the AIBS_cluster_id values from Supplementary Table 5 of Winter et al. to cl identifiers34. Marker-based cell calling and cluster pairing To evaluate marker genes, we used z-scored CPM expression values across all cells in the atlas. For each marker set, we averaged these normalized expression values across its markers. This average marker expression was then used to rank the cells, allowing us to compute the AUROC score for all clusters in the dataset. The result was a cluster marker set by cluster AUROC matrix, with the diagonal representing the on-target AUROC scores. This matrix enabled the comparison of on-target versus best off-target AUROC scores across both atlases. Additionally, these scores were used to generate marker-based cluster pairings, where the single-nucleus cluster with the highest AUROC score for each single-cell marker set was paired. For single-nucleus clusters that were the best hits for more than one marker set, we took the pairing with the highest AUROC value. MetaNeighbor We used the Python version of MetaNeighbor (https://github.com/gillislab/pyMN, RRID:SCR_016727) to quantify cluster replicability15. Default settings were used for most analyses, except for enabling the ‘run fast approximate version’ flag due to the large dataset size14. HVGs were selected with MetaNeighbor’s default procedure from the 21,649 genes shared by both atlases: genes were binned into ten deciles by median expression, and within each decile the top 25% by variance were retained. The two atlas-specific highly variable sets were then intersected to yield the final gene set. The default symmetric AUROC matrix output was used to identify best reciprocal hits, and we also tested the asymmetric output for a more independent set of reciprocal hits. MERFISH spatial data Cell-by-gene expression matrices and annotations from the Yao et al. MERFISH experiment were obtained from the Allen Brain Cell Atlas (ABC Atlas, RRID:SCR_024440) (ID: MERFISH-C57BL6J-638850)5. To assign cells to single-nucleus atlas centroids, we applied the MapMyCells method used by Yao et al. from the scrattch.bigcat R package35. Briefly, for each cell, this method calculates the most correlated cluster centroid in the log2(CPM + 1) gene expression space, repeating this process 100 times after randomly subsampling the profiled genes. Cells with mean correlation values above 0.5 were assigned to clusters, in line with Yao et al.‘s minimum correlation threshold. The provided Common Coordinate Framework (CCFv3, RRID:SCR_020999) parcellation annotations were mapped to the broader dissection regions of the atlases36. For example, the pons (P) and medulla (MY) were mapped to the brainstem (BS). Gene Ontology gene sets Gene Ontology (GO, RRID:SCR_002811) gene annotations dated February 2025 were obtained from the org.Mm.eg.db R package37. Gene sets from the biological process aspect were filtered to have between 15 and 150 genes with coordinated expression data. The December 2023 version of SynGO (RRID:SCR_017330) was used to identify synaptic terms38. BARseq spatial data Single-cell expression and location data were obtained from an in situ sequencing dataset of the whole mouse brain20. We analyzed the same 1,259,256 cells that remained after quality control by the original authors, grouped into 135 clusters spanning all three levels of their hierarchical taxonomy. Instead of using HVGs, we used all 109 genes profiled in this dataset for the MetaNeighbor pretrained models from the atlases. Additional single-cell brain data from mouse and other species We assessed the utility of whole-mouse-brain atlases as a reference taxonomy by assessing the replicability of clusters across 11 mouse and 19 non-mouse brain datasets (see Supplementary Data 9 and 10 for dataset details). The human retina atlas17 used for the split half analyses was obtained from the CELLxGENE platform39. Due to the size of the atlases, we used models pretrained on each atlas to enable fast alignment of the large number of independent datasets. These pretrained models store the HVGs, scaling parameters, and cluster profiles. This allows new datasets to be rapidly projected into the same space and scored for replicability using the atlases as references. These MetaNeighbor pretrained models were built using the 3534 HVGs for the 5322 single-cell and 5030 single-nucleus clusters separately. Each independent dataset was filtered to include only genes present in the pretrained models. For cross-species comparisons, species-specific gene identifiers were converted to mouse orthologs using one-to-one orthology (sources provided in Supplementary Data 10). Target clusters in the independent datasets were mapped to their best-matched clusters, determined by the highest AUROC score calculated by the pretrained model. We then calculated enrichment for 2009 reciprocal-hit clusters in the assigned top hits using Fisher’s exact test. For each target cluster in an independent dataset, we combined results from both pretrained models by counting the number of single-cell or single-nucleus best matches that are reciprocal-hit clusters. A graph visualization of the top-hit matches for cross-species target clusters was generated using the R igraph package40. Phylogenetic divergence times between species were obtained from Timetree41 (RRID:SCR_021162). Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Supplementary Data 5 Supplementary Data 6 Supplementary Data 7 Supplementary Data 8 Supplementary Data 9 Supplementary Data 10 Reporting Summary Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Supplementary Data 5 Supplementary Data 6 Supplementary Data 7 Supplementary Data 8 Supplementary Data 9 Supplementary Data 10 Reporting Summary Source data Transparent Peer Review file Source Data Transparent Peer Review file Source Data