Developing mouse inhibitory neuron single-cell transcriptomes reveal distinct modes of cell-type diversification

Abstract

The cerebral cortex depends on a diverse repertoire of inhibitory neurons, yet how this diversity emerges during development remains unclear. Rare inhibitory subtypes are often underrepresented in single-cell RNA-sequencing datasets, limiting resolution of their developmental trajectories. Here we developed a computational pipeline to enrich and integrate rare cell types across datasets and applied it to somatostatin-expressing (SST⁺) inhibitory neurons, the most diverse inhibitory class in cortex. We generated Dev-SST-v1 and Dev-SST-v2, transcriptomic reference maps comprising more than 55,000 mouse SST⁺ neurons. These maps identify three major SST⁺ inhibitory neuron groups—Martinotti cells (MCs), non-Martinotti cells (nMCs) and long-range projecting (LRP) neurons—each defined by a distinct developmental strategy. MCs commit early, whereas nMCs diversify progressively. LRPs follow a contracting trajectory, with one transient subtype eliminated by programmed cell death. Together, these findings establish three distinct modes of SST⁺ inhibitory neuron diversification, including a previously unrecognized contracting mode.

Researchers discover a transient inhibitory cell type, and reveal that the developing cortex builds inhibitory diversity through three routes—early commitment, gradual diversification and selective elimination.

Main

Cortical inhibitory neurons are highly diverse and coordinate information processing, network synchrony and circuit homeostasis. Their developmental dysregulation has been linked to disorders including epilepsy, schizophrenia and autism^1,2. This underscores the need to understand how these neurons develop and specify.

Advances in single-cell and single-nucleus RNA sequencing (sc/snRNA-seq) have transformed the study of brain tissue, which contains heterogeneous cell types. However, adequate representation of cortical inhibitory neurons in most datasets remains particularly challenging due to their rarity. Resolving their developmental trajectories requires either extensive enrichment or sequencing of large populations—approaches that are not standard in most experiments.

In the adult mouse cortex and hippocampus, inhibitory neurons comprise only ~15% of neurons, whereas ~85% are excitatory glutamatergic neurons¹. Within this minority population, SST⁺ neurons are especially heterogeneous at transcriptional and functional levels, making specific subtypes particularly hard to capture^3,4. Although large-scale atlases have improved annotation through hierarchical classification schemes, rare inhibitory cell types remain underrepresented. For example, 16% SST⁺ types are represented by <150 cells, compared with only 2.7% excitatory pyramidal neuron types^3,5.

The challenge of poor representation is further compounded in developmental datasets^6–8. Early developmental transcriptional profiles are dominated by immature signatures, obscuring cellular diversity^6–9. For SST⁺ neurons, which include Martinotti cells (MCs), non-Martinotti cells (nMCs) and long-range projecting (LRP) neurons^1,6,8,10, this has limited reconstruction of fine-scale developmental trajectories.

To address this problem, we developed a computational pipeline to improve detection and integration of rare neuronal cell types during development. The pipeline comprises three modules—rare cell extraction, dataset integrability assessment and hierarchical clustering for atlas construction. Using cortical SST⁺ neurons as a model system, we generated two transcriptomic resources, Dev-SST-v1 and Dev-SST-v2, spanning the mouse brain from embryonic day 16.5 to postnatal day 10. Together, they include more than 55,000 SST⁺ neurons, representing a sevenfold increase in sampling over a previous study⁶.

Our analysis reveals three principal SST⁺ groups—MCs, nMCs and LRPs—each with distinct modes of diversification. MCs follow an invariant trajectory in which early developmental clusters align closely with adult cell types. nMCs follow an expanding trajectory in which a single neonatal cluster gives rise to multiple adult types. LRPs follow a contracting trajectory—two early postnatal clusters persist until P5, after which one is selectively lost through programmed cell death (PCD), leaving a single adult type. Together, these results establish Dev-SST-v1 and Dev-SST-v2 as a comprehensive resource for inhibitory neuron development and reveal diverse strategies of cell-type diversification in the developing mammalian cortex.

Results

Developing somatostatin-expressing (SST⁺) inhibitory neurons are underrepresented in most scRNA-seq datasets

SST⁺ inhibitory neurons represent 3–4% of cortical neurons yet are among the most diverse classes^1,3,4,11. In the adult mouse neocortex, scRNA-seq studies identified 19 supertypes and 45–73 types or clusters^3–5. To resolve the development of these neurons, we built a high-resolution single-cell atlas using a three-module computational pipeline (Fig. 1a).

Fig. 1. Integrative scRNA-seq developmental atlas for SST⁺ inhibitory neurons.

a, Schematic of the three-step integrative pipeline for generating an atlas. Step 1—selection of rare cell classes (SST⁺) from publicly available scRNA-seq datasets. Step 2—assessment of dataset integrability using anchor points, k-NN graph evaluation, and k-BET testing. Step 3—construction of the new developmental SST⁺ atlas (Dev-SST-atlas) by integrating accepted datasets, reclustering and QC. b, Barplot showing the distribution of SST⁺ cells identified in each dataset in module 1, highlighting variability in SST⁺ cell representation across datasets. c, UMAP visualization of Dev-SST-v0, comprising 19,874 SST⁺ cells, annotated by sample types, main groups and subtype labels transferred from ref. ⁶. d, UMAP visualization of Dev-SST-v0 annotated by iterative clustering parameters set in this study. e, UMAP embedding of Dev-SST-v0 annotated using MapMyCells with ABC Atlas³, overlaying mapping probabilities for each cell. f, Stacked barplot depicting the relative abundance of subclasses in Atlas-v0, restricted to cells with mapping high probabilities (>0.9). DS, datasets.

In module 1, we surveyed available scRNA-seq datasets from the developing mouse cortex spanning embryonic day (E)12.5 to postnatal day (P)10 (Extended Data Table 1), totaling more than 1.5 million cells^6,8,12–26. After quality control (QC; Methods; Extended Data Fig. 1a), SST⁺ clusters were identified by expression of Sst and Lhx6. Among 18 datasets (DS1–DS17, DS25 (refs. ^6,8,12–26)) analyzed, 7 datasets contained more than 1,000 SST⁺ neurons (Extended Data Table 1a and Fig. 1b). The largest dataset⁶ contained ~19,000 SST⁺ cells and was used to generate a reference atlas, Dev-SST-v0. We annotated Dev-SST-v0 by label transfer from ref. ⁶ (Fig. 1c,d, Extended Data Fig. 1a–c and Supplementary Fig. 1a,b). To assess integration quality, we quantified anchor-point distributions and found normalized mean anchor values across datasets (Supplementary Fig. 1c,d).

Extended Data Table 1.

Mouse cortical developmental datasets

Mouse cortical developmental datasets.

Extended Data Fig. 1. Computation pipeline for the generation of Dev-Sst-v2.

a, Module 1 flowchart outlining the filtering criteria, QC modules and steps for isolating high-quality cortical Sst⁺ cells. b, Module 2 flowchart illustrating the integrable testing using Dev-Sst-v0 as the reference atlas. c, Module 3 flowchart detailing the additional quality control steps applied to Dev-Sst-v0, following the reclustering algorithm used to generate Dev-Sst-v2.

We then mapped developmental cells to the adult inhibitory neuron reference using MapMyCells³ (RRID: SCR_024672). At the subclass level, more than 75% of Dev-SST-v0 cells aligned with adult SST⁺ inhibitory neurons—Sst 053, Sst-Chodl 056 (Fig. 1e,f), with high confidence (>0.9 probability). Approximately 15% of cells mapped to Pvalb 052 or olfactory-striatal inhibitory neurons, 045 OB-STR (Fig. 1e). SST⁺ cells in Dev-SST-v0 were enriched using fluorescence-activated cell sorting (FACS) from Sst^cre; RCE animals, which contain minimal (<1%) Pvalb⁺ cells⁶. This suggests that some developing SST⁺ neurons may transcriptomically resemble adult Pvalb⁺ and olfactory-striatal interneurons. Overall, module 1 accurately extracts rare cell types.

Generalized computational pipeline to assess integrability of different datasets

Module 2 of our pipeline examines structural similarities across distinct datasets (Fig. 1a and Extended Data Fig. 1b). Because only seven datasets contain more than 1,000 SST⁺ (Extended Data Table 1), we generated eight additional datasets spanning E16.5 to P5 (DS18–DS24, DS26; Extended Data Table 1), each containing more than 1,000 SST⁺ cells (Fig. 2a). We applied the computational pipeline of module 1 to these new datasets (Fig. 2a).

Fig. 2. Computational pipeline to assess the integrability of different datasets.

a, Tabular overview of datasets used to test integrability with Dev-SST-v0, including both newly generated and published datasets. The table lists SST⁺ cell counts, normalized mean anchor points, k-BET (k-NN batch effect test) rejection rate and fractions of cells within Dev-SST-v0 (k-NN, k = 60) for each dataset. b, Barplot of normalized mean anchor points calculated using Seurat’s CCA integration function between each dataset (subsetting randomly 1,000 cells, n = 4 comparisons) and Dev-SST-v0. Datasets with >1,686 mean anchor points (horizontal dashed line) were considered integrable. Error bars represent mean ± s.e.m. c, Fraction of cells in each dataset within Dev-SST-v0 neighborhoods with increasing k value, n = 16 datasets. The dashed line (k = 60). d,e, Dot plot of fraction of cells within k-NN (k = 60; d) and k-BET rejection plotted against anchor points rates for each dataset (e). Datasets passing the three integrability criteria (in green) have mean anchor points of >1,686, fraction within k-NN > 50% and k-BET rejection rate of <0.5. f,g, Box plot of mean count of RNA (f) and the number of genes detected from datasets assessed passing either 0/1 (n = 4 datasets) or 2/3 criteria (n = 16 datasets; g). Box plots—centerline, median; box bounds, 25th and 75th percentiles; whiskers, 5th and 95th percentiles. ***P < 0.001 (f, P = 0.000068; 0.000002, P = 0.000002) two-tailed unpaired ttest. h, Percentage of discarded cells post-QC from each dataset. i, UMAP visualization of all DS and highlight of DS passing 0/1 criteria. j, Stacked barplot of cells assigned as MC, nMC, LRP or not assigned in DS passing 0/1 or 2/3 criteria.

At stage E13.5 and earlier, few SST⁺ cells arrive in the cortex (Supplementary Fig. 1e,f). Sampling from medial GE samples effectively captured newly born SST⁺ interneurons (Supplementary Fig. 1g–i). While these postmitotic neurons could be assigned to broad classes, cluster-level mapping probabilities were substantially lower than those from E16.5 onward (Supplementary Fig. 1i–k). We therefore focused subsequent analyses on samples from E16.5.

We then tested whether SST⁺ neurons from different datasets could be integrated using the following three criteria: transcriptomic similarity, local structural consistency and batch-mixing quality. First, we quantified anchor points after randomly subsampling 1,000 SST⁺ cells per dataset and integrating each dataset with Dev-SST-v0 using Seurat canonical correlation analysis (CCA; Methods). Dev-SST-v0 contains four reference samples from ref. ⁶, and the mean anchor points were calculated. We set the acceptance threshold at >1,686 normalized mean anchor points, corresponding to the 25th percentile of the anchor distribution (Fig. 2a,b and Supplementary Fig. 1d). Second, we assessed local structural similarity using k-nearest neighbors (k-NN) derived from the integrated principal component (PC). For each dataset, we calculated the fraction of cells with neighborhoods matching those in Dev-SST-v0. We set the acceptance threshold to >50% cells within the same k-NN structure (k = 60), matching the value observed in Dev-SST-v0 (Fig. 2c,d). Third, we calculated the k-BET rejection rate (Methods), a standard measure of batch mixing and required k-BET < 0.5 for acceptance (Fig. 2a,e). Of the 13 datasets tested, 7 passed all three criteria and were used to generate the developmental SST⁺ atlas in module 3.

To determine whether these criteria reflected dataset quality, we compared datasets passing ≤1 criterion with those passing ≥2. Datasets meeting at least two criteria showed significantly higher mean RNA counts and gene numbers per cell (Fig. 2f,g, ***P < 0.001), as well as lower cell loss during QC (<20% versus >40%; Fig. 2h). In lower-quality datasets, about 25% retained cells could not be confidently assigned to MC, nMC or LRP subclasses (Fig. 2i,j), whereas nearly all cells from datasets passing ≥2 criteria were classified successfully. Thus, the three integrability criteria serve as stringent quality filters.

The parameters defined in module 2 are generalizable and can be applied to other contexts. As a first example, we applied module 2 to snRNA-seq data from 13 adult mouse spinal cord samples (Data availability; Supplementary Fig. 2a). Using the control sample with the largest number of cells as a reference atlas, we assessed integrability of the remaining samples using the same three criteria (Supplementary Fig. 2b–e). As a second example, we constructed a developmental atlas of cortical PV⁺ interneurons, another rare inhibitory population. From six datasets, we identified PV⁺ interneurons, selected the dataset with the largest cell number as the reference atlas and applied the same integration criteria (Supplementary Fig. 3a,b). Four datasets passed all three criteria and were integrated to generate Dev-PV-v1 (Supplementary Fig. 3c,d). Using MapMyCells, we found that more than 85% of Dev-PV-v1 cells mapped to adult 052 Pvalb Gaba with probability >0.9 (Supplementary Fig. 3e,f). Together, these examples demonstrate that module 2 provides a robust and generalizable framework for assessing dataset integrability.

Integrability criteria do not depend on sequencing depth

To identify factors that influence integrability of independent datasets, we searched for dataset-specific features in experimental setup and computational analyses. As sequencing read depth differs greatly between laboratories and samples, we examined whether this could influence integrability.

We took two datasets, DS21 and DS24, and performed both shallow and deep sequencing. Shallow sequencing targeted a sequencing depth of 5,000 reads per cell; the actual mean reads achieved was 8,227 ± 794. Deep sequencing targeted 50,000 reads per cell; the actual mean reads achieved was 61,384 ± 20,484 (Extended Data Fig. 2a). Then, we performed the integrability test (module 2) and quantified anchor points with Atlas-v0, fraction of cells within k-NN and the k-BET rejection test (Extended Data Fig. 2a–d). Despite an ~7.5-fold difference in sequencing depth between shallow-sequenced and deep-sequenced samples, both passed the integrability criteria. Our analysis indicates that sequencing depth alone is not enough to explain failed integration, and suggests that other parameters may influence integrability.

Extended Data Fig. 2. Integrability of the same dataset with variations in sequencing depth.

a, Normalized mean anchor points for DS21 and DS24 under shallow (S) and deep sequencing protocols. b, Fraction of cells in shallow- and deep-sequenced datasets within Dev-SST-v0 neighborhoods with increasing k value. Both shallow and deep sequencing protocols met the acceptance threshold of >50% (k = 60) (dashed line). c,d, Dot plot of (d) fraction cells within k-NN (k = 60) and (e) k-BET rejection rates plotted against anchor points for shallow and deep sequencing. Datasets of either deep and shallow sequencing depth that passed most integrability criteria (in green) have mean anchor points >2000, fraction within k-NN > 50%, and k-BET rejection rate <0.5.

Dev-SST-v2 reveals 26 clusters of cortical SST⁺ inhibitory neurons

We integrated SST⁺ cells from datasets that passed module 2 criteria into Dev-SST-v0 using Seurat CCA to generate two resources, Dev-SST-v1 (passing two criteria) and Dev-SST-v2 (passing three criteria). Additional QC steps removed doublets, stressed cells and clusters composed predominantly (>60%) of cells from a single dataset (Methods). After filtering, Dev-SST-v1 contained 63,572 SST⁺ cells, whereas Dev-SST-v2 contained 59,658 cells (Extended Data Fig. 3a,b).

Extended Data Fig. 3. Construction, comparison, and validation of two developmental SST⁺ interneuron atlases (Dev-SST-v1 and Dev-SST-v2).

a,b, UMAP of SST⁺ neurons integrated datasets that passed either ≥2 criteria (a, Dev-SST-v1) or all 3 criteria (b, Dev-SST-v2) in module 2 using Seurat CCA. The final Dev-SST-v1 and Dev-SST-v2 atlases contained 63,572 cells (stages E16.5, E18.5, P1, P2, P5, P7, P10) and 59,658 cells (stages E16.5, E18.5, P1, P2, P5), respectively. c, Confusion matrices derived from random forest classification summarizing how true cell identities in each atlas version were assigned to predicted types. Diagonal entries represent correct classifications (accuracy), while off-diagonal entries capture misclassifications (cross-talk). d,e, Per-class comparisons between versions (Δ = v2 − v1) for diagonal accuracy (d) and cross-talk (e). Compared to Dev-SST-v1, Dev-SST-v2 consistently has higher accuracy and lower cross-talk across Martinotti (MC), non-Martinotti (nMC), and LRP populations. LRPs: n = 5; MC: n = 22; nMC: n = 10. f, Saturation analysis assessing sampling sufficiency across 2,560–48,656 subsampled cells. For each subsample, four metrics were computed: (1) fraction of conserved clusters, (2) fraction of conserved differentially expressed genes, (3) absolute difference in mean expression, and (4) the robust Hausdorff distance. Inflection points from second-derivative analyses identified saturation at 8,639–11,151 cells. All major developmental stages (E16.5, E18.5/P1, P5) exceed this threshold. Error bars represent ± s.e.m (d–f). g, Box and violin plot of Jaccard index of all clusters in Dev-SST-v2. The cutoff of 0.75, red dashed line, was used for cluster retention. n = 20 subsamplings, each time sampling 80% of the cells without replacement. Box plots—centerline, median; box bounds, 25th and 75th percentiles; whiskers, most extreme values within 1.5× interquartile range.

To compare clustering performance between the two resources, we generated confusion matrices from random forest predictions (Methods; Extended Data Fig. 3c). Dev-SST-v2 showed higher or maintained per-class accuracy and reduced cross-talk among LRP, MC and nMC populations relative to Dev-SST-v1 (Extended Data Fig. 3d,e), and was thus used for subsequent analyses.

We then asked whether sampling was sufficient at each developmental stage. Saturation analyses based on cluster conservation, conservation of differentially expressed genes, mean expression differences and robust Hausdorff distance identified saturation points between 8,639 and 11,151 cells (Extended Data Fig. 3f). At the three major developmental stage groups—E16.5, E18.5/P1 and P5—Dev-SST-v2 exceeded these thresholds.

Dev-SST-v2 has three major SST⁺ populations—MC, nMC and LRP (Fig. 3a)—consistent with previous work⁶. These populations were classified with high accuracy (90–95%; Fig. 3b). Most cells (>75%) were mapped with high confidence (>0.9 probability) to adult cortical SST⁺ subclasses 053 Sst and 056 Sst-Chodl using MapMyCells (Fig. 3c,d). Iterative clustering identified 26 stable SST⁺ clusters (Fig. 3e and Extended Data Fig. 3g, Jaccard index >0.75), roughly thrice more than previously reported for comparable developmental stages⁶. Each cluster displayed a distinct gene expression profile (Extended Data Fig. 4a,b,e). Compared with our previous work, this increase was driven primarily by greater resolution among MCs, which expanded from 3 to 18 clusters, whereas nMCs increased from 5 to 6 and LRPs remained as two cortical types (Fig. 3e). We identified two cortical LRP types, LRP1 and LRP2.1, consistent with previous results. A third minor group, LRP2.2, represented <3% LRPs and likely reflects striatal interneuron contamination because it strongly expressed Nkx2-1, a marker of early postnatal striatal but not cortical SST⁺ interneurons²⁷ (Extended Data Fig. 4c).

Fig. 3. Dev-SST-v2 reveals 26 clusters of cortical SST⁺ inhibitory neurons.

a, UMAP visualization of Dev-SST-v2. SST⁺ cells are grouped into three main populations—MCs, nMCs and LRPs, across developmental stages (E16.5, E18.5, P1, P2 and P5). b, Confusion matrix displaying high classification accuracy (>0.9 probability) for distinguishing MC, nMC and LRP populations. c, Fraction of cells in Dev-SST-v3 matched with high confidence (>0.9 probability) to adult cortical SST⁺ inhibitory neuron subclasses using the MapMyCells algorithm, with subclass annotations 053 Sst and 056 Sst-Chodl dominating. d, Stacked barplot comparing relative abundances of main cell types between Dev-SST-v0 and Dev-SST-v2, highlighting the stability of population structure across atlases. e, Iterative clustering of Dev-SST-v2 reveals 26 stable clusters of cortical SST⁺ interneurons (>0.75 Jaccard index). Clustering results demonstrate the expansion of developmental diversity, identifying 18 MC types, 6 nMC types and 2 LRP types (LRP1 and LRP2.1). Three rare clusters, HIn1, HIn2 and LRP2.2 (<1% of all cells), are noncortical.

Extended Data Fig. 4. Gene expression differences among distinct clusters within Dev-Sst-v1.

a,b, Bubble plots of marker genes in clusters of (a) MCs (b) and nMCs. c, Barplot showing the relative fraction of LRP types within the LRP population. LRP2.2, 2%. d, UMAP projections of LRP types from Dev-Sst-v1. e, Bubble plots showing marker gene expression of each LRP. Please note LRP2.2 exhibits high expression of Nkx2.1, a marker for striatal.

To validate CCA-based integration, we applied Waddington optimal transport (WOT), which models probabilistic transitions across time points (Extended Data Fig. 5). WOT independently confirmed the inferred developmental relationships—E16.5 clusters showed the strongest transition probabilities to corresponding P1 clusters, and P1 clusters to their P5 counterparts (Extended Data Fig. 5a). Strong enrichment of diagonal over off-diagonal couplings further supports that these developmental mappings reflect biological continuity rather than integration artifacts (Extended Data Fig. 5b–d).

Extended Data Fig. 5. Waddington optimal transport analysis validates developmental transitions inferred from CCA-based integration.

a, Schema of Waddington optimal transport (WOT) probabilistic transitions between successive developmental stages (E16.5 → P1 → P5). b, Strong enrichment of diagonal couplings relative to off-diagonal transitions was observed for LRP (n = 2), nMC (n = 6), and MC (n = 18) clusters, indicating that cells within each cluster predominantly transition into the expected lineage-matched clusters across time. Box plots—centerline, median; box bounds, 25th and 75th percentiles; whiskers, most extreme values within 1.5× interquartile range. c–e, Confusion matrix of WOT probability for E16 to P1 and P1 to P5 transition for (c) LPR, (d) nMC, and (e) MC.

Cortical SST⁺ interneurons display heterogeneous modes of diversification

To determine when SST⁺ interneuron types emerge during development, we used MapMyCells to assign developmental MCs and nMCs to the adult subclasses and supertypes defined in ref. ³ (Fig. 3a,c and Extended Data Fig. 6a–c). Our premise is that developing SST⁺ neurons with strong transcriptomic similarity to adult cell types are likely to mature into those identities. At the subclass level, almost all (>99%) developing MCs are assigned to the adult 053 Sst subclass (Extended Data Fig. 6a). For nMCs, 85% of cells mapped to their adult counterparts (Extended Data Fig. 6a). At the supertype level, mapping was strongest at P5 (Extended Data Fig. 6b,c). By P1, all 19 adult SST⁺ supertypes could already be assigned to developmental MC and nMC populations (Extended Data Fig. 6d).

Extended Data Fig. 6. Assignment of developmental MCs and nMCs to adult subclasses and supertype using MapMyCells.

a, Left, stacked barplot representing fraction of MCs and nMCs that are assigned to adult subclasses. Assignments with >0.9 or <0.9 probability are annotated as ‘asssigned’ or ‘unassigned’, respectively. Right, assignment of different subclasses with probability >0.9. b,c, Stacked barplot representing fraction of (b) MCs and (c) nMCs at different stages that are assigned with a high probability (>0.9) to adult supertypes. d, Stacked barplot representing developmental SST interneurons assigned to each supertype with a high probability (>0.9).

Nearly all developmental cells mapping to the adult 052 Pvalb subclass were nMCs (Fig. 3c), comprising 26.5% of all nMCs (Extended Data Fig. 6a). This may reflect that some SST⁺ supertypes are transcriptomically similar to Pvalb⁺ cells^4,11. Recent work showed that during development, the Pvalb type 735 segregates from a branch of cells belonging to the Sst subclass²⁸. These findings suggest that some developmental SST⁺/nMCs clusters share more transcriptomic similarity with the adult Pvalb than the Sst supertypes.

Next, we examined whether developmental clusters show one-to-one correspondence with adult supertypes. We defined a one-to-one match as >65% cells in a developmental cluster mapping to a single adult supertype (>0.9 probability). Among the 18 developmental MC clusters, 7 showed a clear one-to-one relationship with adult Sst supertypes (Fig. 4a,b). In contrast, only one of six developmental nMC clusters showed such a match; most mapped to multiple adult supertypes (Fig. 4c,d).

Fig. 4. MCs and nMCs diversification during development.

a, UMAP of Dev-SST-v2, focusing on developmental interneuron clusters by adjusting the axis limits (UMAP_2, −5 to 5 instead of −5 to 10 in Fig. 3). Labeling highlights only MC for visualization. Developing MC clusters mapped to adult SST⁺ interneuron supertypes using MapMyCells with a high-confidence threshold (>0.9). Clusters with a one-to-one (1:1) match to adult supertypes are highlighted. b, River plot of developmental MC clusters and their corresponding adult supertypes, illustrating 1:1 matches. c, UMAP of developmental interneuron clusters by adjusting the axis limits (y-axis UMAP_2, −5 to 5 instead of −5 to 10 in Fig. 3). Labeling highlights only nMC for visualization. nMC clusters mapped to adult SST⁺ supertypes, highlighting limited 1:1 matches. Notably, nMC clusters mapped to both SST⁺ and Pvalb⁺ adult subclasses. d, River plot of developmental nMC clusters and their corresponding adult supertypes, illustrating a more flexible developmental trajectory with multiple clusters contributing to individual adult supertypes. Only one of six nMC clusters showed a 1:1 relationship with adult supertypes. e, Bubble chart (top) and barplot showing the composition of adult SST⁺ supertypes derived from developmental MC and nMC clusters. Ten supertypes were exclusively composed of MC-derived cells, while three were exclusively nMC-derived cells. Supertypes with mixed compositions were primarily assigned to nMC adult types. f, FCs in the relative abundance of each Sst supertype from P1 to P5 and P1 to P56.

To further examine this relationship, we calculated the relative contribution of MC and nMC clusters to each adult Sst supertype (Fig. 4e). Thus, 10 of 19 adult supertypes were composed predominantly of developmental MCs (>75%), whereas only 3 were predominantly nMC derived. The MC-dominated supertypes correspond to adult cell types that are also well-defined as MCs by morphology, electrophysiology, transcriptomics and spatial distribution. By contrast, mixed developmental compositions were more common among adult nMC-related supertypes. These analyses indicate that most adult MCs emerge from developmental MC clusters, whereas adult nMCs can arise from both developmental MC and nMC populations.

To assess the biological relevance of these clusters, we performed MERFISH on P5 tissue using a customized ~500-gene probe set designed to classify developmental clusters accurately (Extended Data Fig. 7a and Supplementary Table 1). Many clusters showed distinct spatial distributions (Extended Data Fig. 7b)—six were enriched in infragranular cortical layers, whereas five were restricted to ventral regions such as insular and piriform cortex (Extended Data Fig. 7c,d).

Extended Data Fig. 7. P5 MERFISH spatial validation of developmental SST⁺ clusters.

a, Confusion matrix on classifying accuracy of customized probe set of ~500 genes used in MERFISH. This gene set is optimized to classify Dev-SST-v2 clusters with high accuracy. b, Spatial cell-type assignment using MERFISH. Most developmental clusters exhibited distinct and regionally enriched spatial distributions across the P5 cortex. c, Six clusters were confined to infragranular layers (L5/L6). d, Five clusters localized predominantly to ventral cortical regions, including insular and piriform cortices. e,f, To assess whether the spatial organization of developmental clusters resembles that of their adult counterparts, we compared P5 MERFISH data with adult SST⁺ MERFISH from ref. ⁵. Eight developmental clusters—four infragranular and four ventral—were examined alongside their one-to-one adult Sst supertypes. All pairs displayed highly similar regional localization patterns in both developmental and adult tissue.

Several developmental clusters corresponded one-to-one with adult Sst supertypes (Fig. 4). We asked whether their spatial localization at P5 resembled that observed in the adult. Among eight developmental clusters—four enriched in infragranular layers and four in ventral cortex—we found that each closely matched the spatial distribution of its corresponding adult supertype (Extended Data Fig. 7e,f). This pronounced anatomical specificity indicates that the clusters identified in Dev-SST-v2 capture biologically meaningful cell types. The early segregation of ventral interneurons by P5, before PCD, further suggests that this organization arises during migration.

Finally, we asked when the relative abundance of SST⁺ interneuron types is established. We quantified the fraction of each adult Sst supertype in published adult cortex data and compared it with relative abundance at P1 and P5 within MC and nMC populations. Fold-change (FC) analysis revealed that, except for three supertypes, most Sst supertypes changed substantially in relative abundance between P1 and P5 and between P1 and adulthood (Fig. 4f). Thus, consistent with previous findings²⁸, the proportions of most SST⁺ supertypes are not fixed early, but continue to shift during postnatal maturation. Overall, these results indicate that SST⁺ interneurons diversify through multiple developmental trajectories, with some adult supertypes emerging relatively late.

LRP1 and LRP2.1 are distinct cell types with divergent trajectories

Our previous study⁶ linked developmental LRP1 to the adult Sst-Chodl-64 and Sst-Chodl-65 subtypes, while LRP2.1 (formerly LRP2) was associated with Sst-Chodl-63 in the datasets mentioned in ref. ⁴. To refine this classification, we used MapMyCells and found that LRP1 corresponds to the Sst-Chodl-Gaba-2 (supertype 241), whereas LRP2.1 aligns with Sst-Chodl-Gaba-4 (supertype 239; Fig. 5a). Spatial data⁵ showed a striking anatomical segregation—Sst-Chodl-Gaba-2 (supertype 241) is confined to the cortex, while Sst-Chodl-Gaba-4 (supertype 239) is found exclusively in the striatum (Fig. 5b). In contrast, at P5, we found that some Sst-Chodl-Gaba-4 (supertype 239) are found in the cortex (Fig. 5c). Notably, Sst-Chodl-63 from the dataset of ref. ⁴, which accounts for <5% of all LRPs, also maps to Sst-Chodl-Gaba-4, further indicating that LRP2.1 cells possess a transcriptomic signature similar to adult striatal interneurons (Extended Data Fig. 8a).

Fig. 5. LRP1 and LRP2.1 are distinct cortical cell types.

a, UMAP of developmental LRP clusters, highlighting distinct mappings of LRP1 to adult supertype 241 Sst-Chodl-Gaba-2 and LRP2.1 to adult supertype 239 Sst-Chodl-Gaba-4. b, Adult spatial distribution of LRP supertypes in the ABC Atlas MERFISH data⁵. Please note that supertype 241 Sst-Chodl-Gaba-2 is exclusively cortical, while supertype 239 Sst-Chodl-Gaba-4 is restricted to the striatum. c, P5 spatial distribution of P5 LRP supertypes using customized MERFISH data from this study showing mapping of 0239 Sst-Chodl-Gaba-4 in the cortex. d, UMAP of P5 cortical LRP clusters and striatal SST⁺ interneurons labeled by region of tissue, label transfer from Dev-SST-v2 and cell cluster. e, Bubble plot of key marker genes distinguishing LRP2.1 and striatal SST⁺ interneurons at P5. f, UMAP showing expression of key marker genes in cortical LPRs and striatal SST⁺ interneurons. g, UMAP of E18.5 GE scRNA-seq data from ref. ²⁹, re-analyzed and labeled by cell cluster. h, UMAP of E18.5 GE cells showing expression of key marker genes at E18.5 that segregate cortical LPRs and striatal SST⁺ interneurons. i, Bubble plot of key genes distinguishing LRP2.1 and striatal SST⁺ interneurons at E18.5. j, Pseudotime analysis of LRP1 and LRP2.1 using diffusion map space and developmental stage data, showing two distinct branches. k, Expression of shared and branch-specific genes along pseudotime progression.

Extended Data Fig. 8. Characterization of LRP1 and LRP2.1.

a, UMAP plot of Sst-Chodl clusters (LRP) found in ref. ⁴ using annotation in that paper (left) and annotation assigned by MapMyCells (right). Furthermore, 63_Sst_Chodl cells are predominantly assigned as supertype 0239 Sst Chodl. b, P6 cortical tissues injected with AAV-eGHT-600m-sYFP2 at P0 and staining with GFP. c, P6 Sst^cre cortical tissues injected with AAV-eGHT-dio-h600-sYFP2 at P0 and stained with GFP and NOS1. Scale bar = 100 µm. d,h, Confusion matrix calculating accuracy of distinguishing LRP1 and LRP2.1 in Dev-Sst-v1 using 3 genes (d) Sst/Nos1/Dach1 or (h) Sst/Nos1/MafB. e,f,i,j, Co-immunostaining of (e,i) GFP/Dach1 or (f,j) GFP/MafB with quantification of relative abundance of LRP1 and LRP2.1 in Sstcre using mice injected with AAV2,9-eGHT-dio-600m-sYFP2 at P0 and immunostained at (e,f) P5 or (i,j) at P7. Yellow and white arrowheads denote LRP1 and LRP2.1, respectively. g,k, Barplots represent quantification of LRP1 and LRP2.1 assessed by staining with Dach1, Mafb, or combined-mean result of both methods at (g) P5 and (k) P7. The numbers of the bars represent the cell number quantified from each condition, n = 3 mice per time point. Both methods give similar results. l, Relative abundance of 19 SST⁺ interneuron supertypes at P14 in control and Bax cKO mice, datasets re-analyzed from ref. ³⁶. No significant differences in the relative abundance of these supertypes were observed. m, PCA plots of hLRPs from ref. ³⁹ labeled by stages post-transplantation.

As LPR2.1 cells map transcriptomically to adult striatal interneurons, we then assess whether they are transcriptomically different clusters from striatal SST⁺/Nos1⁺ interneurons. We isolated SST⁺ striatal interneurons at P5 (Fig. 5d), performed scRNA-seq and integrated these data with all P5 LPR cells annotated in the Dev-SST-v2 dataset (Fig. 5d). We found that cortical LPR2.1 cells clustered in close proximity to both cortical LPR1 cells—previously defined long-range GABAergic neurons—and striatal SST⁺ interneurons (Fig. 5d). Although SST⁺ striatal interneurons shared several transcriptomic features with cortical LPR2.1 cells, including being assigned an LPR2.1 by label transfer approaches using the labels from Dev-SST-v2 (Fig. 5d, middle), the two populations showed distinct cluster boundaries (Fig. 5d, right) and differential expression of key genes such as Rora, Kncip4, Erbb4, Cacna1c and Maf (Fig. 5e,f).

Using recent data that have sequenced cells from E18.5 mouse ganglionic eminences (that is, MGE, LGE and CGE) cells²⁹, we annotated these datasets and found that LPR1, LPR2.1 and striatal SST⁺ interneurons are segregated into three distinct clusters, similar to the P5 samples (Fig. 5g) with clear differences in gene expression patterns (Fig. 5h,i). These analyses indicate that LPR2.1 cells are not striatal cells, but they have partial transcriptomic overlap with cortical LPR1 and striatal SST⁺ interneurons. LPR2.1 and striatal interneurons are transcriptomically different in the GE before LPRs migrate into the cortex.

Having established that cortical LPR2.1 cells constitute a distinct cortical population, we then asked whether they follow a developmental trajectory separate from that of LPR1. We performed pseudotime analysis using the expanded Dev-SST-v2 dataset. First, we identified 335 genes (by ANTLER; Methods) grouped into four distinct modules involved in neuronal differentiation, axon development and synapse assembly (Supplementary Table 2). Pseudotime trajectory generated diffusion maps revealed two clearly separated branches corresponding to LRP1 and LRP2.1 (Fig. 5j). At early pseudotime points, these two clusters were distinct, each expressing a unique set of branch-specific genes (Fig. 5k). These results suggest that LRP1 and LRP2.1 represent two fundamentally distinct cortical cell types rather than different cell states.

LRP2.1 is a transient cell population that declines between P5 and adulthood in the mouse cortex

Our analyses indicate that most LRP2.1 cells do not persist in the adult neocortex, despite their unique developmental trajectory. Both adult cortical transcriptomic dataset⁴ and spatial data⁵ (Fig. 5b,c) show that <5% LRPs map to LRP2.1 in the adult brain (Extended Data Fig. 8a). In contrast, Dev-SST-v2 reveals that LRP2.1 constitutes ~25% of all LRPs at E16.5, increasing to ~55% by P5 (Fig. 6a). Given that cortical inhibitory neurons do not migrate back to the striatum after reaching the cortex³⁰, and LPR2.1 is transcriptionally distinct to striatal interneurons as early as E18.5 (Fig. 5g), we hypothesize that LRP2.1 may be a transient cell type.

Fig. 6. LRP2.1 is a transient cell type in both mouse and human neocortex.

a, Relative proportions of LRP1 and LRP2.1 in Dev-SST-v2. b, Validation of LRP1 and LRP2.1 relative abundance, re-analyzed from ref. ²⁸. c, Experimental schema of labeling LRPs by AAV2/9-eGHT-600m-dio-sYPF2 in SSTcre mice at P0. d,e, Co-immunostaining for GFP and Dach1 (LRP2.1 marker) at P5 (d) and P7 (e). White arrowheads indicate double-positive cells (LRP2.1). Scale bar = 100 µm. f, Proportions of LRP1 and LRP2.1 quantified by immunostaining at P5, P7 and P21, n = 3 animals per time point. g, Relative abundance of LRP1 and LRP2.1 at P2, P7 and P14 re-analyzed from control mice and P14 in Bax cKO mice, datasets from ref. ³⁶. h, Relative abundance of human LRP1 and LRP2.1 in developing neocortex re-analyzed from independent datasets^37,38. i, Relative abundance of LRP1 and LRP2.1 in hPSCs-derived MGE interneurons xenotransplanted in the mouse neocortex, re-analyzed from ref. ³⁹. The number on each bar represents the number of LRPs found in each stage. j, PCA plot of hLRP1 and hLRP2.1, re-analyzed from ref. ³⁹. *P < 0.05, **P < 0.01, ***P < 0.001 using Fisher-exact test, correction with multiple testing, n = cell number of each type indicated on the graph.

To track these populations beyond P5, we re-analyze snRNA-seq datasets²⁸, which span developmental stages E13.5 to P60. Although LRPs were rare in these datasets, LRP2.1 abundance peaks at P5 before rapidly declining, stabilizing to <10% in adulthood (Fig. 6b). To ensure that this sharp reduction at P7 does not reflect selective vulnerability of these cells during tissue dissociation, we examined LRP1 and LRP2.1 abundance in the cortex at P5, P7 and P21 using histology. We labeled LRPs with a modified enhancer virus targeting the Nos1 locus, AAV2/9-eHGT-600m-sYPF2 (ref. ³¹). In early postnatal stages, we noted that this enhancer also transiently labels pyramidal neurons due to the low and transient expression of Nos1 during development³² (Extended Data Fig. 8b). To restrict expression to SST⁺/NOS1⁺ neurons, we subcloned the enhancer into a Cre-dependent construct, and generated pAAV-eHGT-600m-dio-sYPF2 (Extended Data Fig. 8c). We injected AAV2/9-eHGT-600m-dio-sYPF2 into the lateral ventricles of P0 Sst^Cre mice and (Fig. 6c), and confirmed that all fluorescently labeled cells at postnatal stages are NOS1⁺ (Extended Data Fig. 8c).

Dev-SST-v2 atlas revealed that a combination of three markers—either Sst/Dach1/Nos1 or Sst/Mafb/Nos1—distinguishes LRP1 and LRP2.1 with >80% accuracy (Extended Data Fig. 8d,h). To differentiate between LRP1 and LRP2.1 histologically, we performed co-immunostaining with GFP/Dach1, a positive marker for LRP2.1 (Extended Data Fig. 8e,i), and GFP/MafB, a positive marker for LRP1 (Extended Data Fig. 8f,j). Control experiments confirmed that quantifying LRP1 and LRP2.1 fraction by either GFP/Dach1 and GFP/MafB staining produces similar results (Extended Data Fig. 8g,k).

At P5, we found that LRP1 and LRP2.1 were present in nearly equal proportions, consistent with Dev-SST-v2 quantifications (Fig. 6a,d–f). By P7, however, LRP2.1 abundance decreased substantially (Fig. 6e,f, ***P < 0.001), and, by P21, only ~18% LRPs were LRP2.1 positive (Fig. 6f, ***P < 0.001). This is consistent with our previous findings⁶, which showed that only 10% of LRPs were Dach1⁺ in the cortex at P28. Together, these results demonstrate that LRP1 and LRP2.1 exist in nearly equal proportions at P5. However, by adulthood, LRP2.1 has largely disappeared, supporting the hypothesis that this cell population is transient and selectively eliminated after birth.

PCD selectively eliminates LRP2.1 in the mouse cortex

In the mouse neocortex, postnatal PCD removes ~40% of cortical GABAergic neurons³³. This process has been considered nonselective, affecting all interneuron subtypes uniformly, occurring between P7 and P12 (refs. ^33–35). Recent work described in ref. ³⁶ shows that in Fezf2-deficient mice, in which layer 5 excitatory pyramidal neurons are absent, some L5/6 SST⁺ interneuron subtypes adjust their relative abundance through selective PCD. We therefore hypothesized that selective PCD may be a mechanism shaping the composition and diversity of LRP populations.

To investigate whether LRP2.1 is eliminated by PCD, we re-analyzed snRNA-seq datasets from ref. ³⁶, which included inhibitory neurons from control (Nkx2.1^Cre;Bax^fl/+;Rosa26^LSL-h2b-GFP) mice at P2, P7 and P14, and Bax cKO (Nkx2.1^Cre;Bax^fl/fl;Bak^−/−;Rosa26^LSL-h2b-GFP) mice at P14, of which interneurons do not undergo PCD. In controls, LRP2.1 abundance declines sharply between P2 and P7, disappearing entirely by P14 (Fig. 6g). In Bax cKO mice, however, LRP2.1 persists at P14 at levels comparable to P2 (~25%; Fig. 6g, ***P < 0.001), indicating that its loss in wild-type (WT) animals is due to PCD. No significant changes were observed in the abundance of other SST⁺ interneuron supertypes (Extended Data Fig. 8l), supporting the conclusion that LRP2.1 is selectively eliminated. Together, these findings provide evidence that inhibitory neuronal diversity can be sculpted through cell-type-specific postnatal elimination.

LRP2.1 emerges in the human developing cortex and declines by birth

To determine whether LRP2.1 is conserved in humans, we re-analyzed scRNA-seq datasets from fetal, pediatric and adult human cortex from two independent studies^37,38. Similar to mice, hLRP1 and hLRP2.1 were present in nearly equal proportions at early stages. hLRP2.1 constituted ~50% LRPs at gestational weeks (GW) 18–20 but declined sharply by birth (Fig. 6h). Data from hPSC-derived MGE interneurons xenotransplanted into mice³⁹ confirmed this trend—hLRP2.1 was most abundant at 1-month post-transplantation but declined sharply by 3-month post-transplantation (Fig. 6i,j and Extended Data Fig. 8m). Because PCD in the human central nervous system begins around GW20 and concludes by birth⁴⁰, these analyses strongly suggest that hLRP2.1 is selectively eliminated during this period. The selective removal of LRP2.1 by cell-type-specific PCD is likely a conserved mechanism across mammalian cortices.

At P4–P6, LRP2.1 neurons are connected to cortical and thalamic excitatory neurons

To determine if this transient cell type, LRP2.1, is synaptically connected within the brain circuit, we investigated the presynaptic partners between P4 and P8. Layer (L) 5 SST⁺ neurons receive transient thalamic input during early postnatal stages^41,42, and these SST⁺ cells are born as early as E10.5 (ref. ⁴¹). Thus, we asked whether LRP neurons at P5 also receive thalamic input and when LPR1 and 2.1 are born.

To examine the monosynaptic connections on LRPs, we injected SST^Cre mice at P0 with two types of AAV-helper viruses, both carrying Cre-dependent expression cassettes for green fluorescent protein (GFP), optimized rabies G protein and the envelope receptor, TVA. To label all pan-SST⁺ cells (interneurons and LPRs), the helper virus used a general synapsin promoter (AAV2/9-syn-Flex-nGToG—nuclear GFP, TVA and optimized G protein). To label LPR specifically, we used the same enhancer for Nos1 (AAV2/9-600m-Flex-nGToG). At P4, we injected rabies virus expressing TdTomato and Flp recombinase (N2c-ΔG-Tdt-Flp) into the cortex and analyzed the number of TdTomato⁺ presynaptic neurons, normalizing to the number of starter cells (TdTomato⁺/GFP⁺ cells—double-positive) per brain at P8 (Extended Data Fig. 9a).

Extended Data Fig. 9. Rabies-based monosynaptic tracing reveals that LRPs receive strong thalamic and subplate inputs during early postnatal development.

a, Experimental strategy. SST^Cre mice were injected at P0 with Cre-dependent AAV helper viruses encoding GFP, TVA, and optimized rabies G protein. The synapsin-promoter helper virus (AAV2/9-syn-Flex-nGToG) labeled all SST⁺ cells (pan-SST⁺), whereas an LPR-specific enhancer helper virus (AAV2/9-600m-Flex-nGToG) selectively labeled LPRs. At P4, N2c-ΔG-Tdt-Flp rabies virus was injected into somatosensory cortex, and histology was performed at P8. Starter cells were defined as TdTomato⁺/GFP⁺ double-positive neurons. b,c,e,f, Mice were injected with (b,c) synapsin-promoter helper virus or (e,f) LPR-specific helper virus. In both conditions, extensive TdTomato⁺ (presynaptic) neurons were detected in thalamus and deep cortical layers. d, Quantification of number of presynaptic cells normalized to 100 starter cells for pan-SST⁺ & LPR starter populations (**P < 0.01) and thalamic (***P < 0.001), assessed by linear mixed-effects model with brain identity included as a random intercept. P values were calculated using Wald z-tests for fixed effects and adjusted for multiple comparisons using the Holm method. g, Mean starter cell density from synapsin-promoter helper virus or LPR-specific helper. Each dot represents the average of 8 sections in each brain, n = 4 brains per condition (d,g). Box plots—centerline, median; box bounds, 25th and 75th percentiles; whiskers 5th and 95th percentiles.

Consistent with previous findings⁴¹, SST^Cre mice injected with the synapsin-promoter helper virus (starter cells = all SST⁺ cortical cells) showed extensive labeling of thalamic and deep-layer cortical neurons (Extended Data Fig. 9b,c). Mice injected with the LPR-specific helper virus showed extensive labeling of thalamic neurons, deep-layer cortical neurons and subplate neurons (Extended Data Fig. 9e,f). We compared the density presynaptic input cells in either SST⁺ or LPRs as starters and found that there were significantly presynaptically connection in the subplate (**P < 0.01) and the thalamus (***P < 0.001) when LRPs were starters (Extended Data Fig. 9d,g). Similar to SST⁺ interneurons⁴¹, LRPs receive substantial thalamic input and are monosynaptically connected to subplate neurons during early postnatal development.

To evaluate when LPRs are born, we injected thymidine analog 5-ethynyl-2′-deoxyuridine (EdU) in time-mated pregnant mice on embryonic days E11.5, E12.5, E13.5 or E14.5. At P5, we assessed the fraction of Nos1⁺/EdU⁺ (LRP1) or Nos1⁺/Dach1⁺ /EdU⁺ (LPR2.1) cells (Supplementary Fig. 4). Similar to previous work⁴³, we found that most LPRs are born around E12.5 (Supplementary Fig. 4a–e). The percentage of EdU⁺ LRP2.1 compared to LPR1 is significantly less at E11.5 (P < 0.05), while is higher at E13.5 (P < 0.05), indicating that LRP2.1 is born later than LPR1 (Supplementary Fig. 4e). Moreover, the peak birthdate for LRP1 and LRP2.1 are E12.5 and E13.5, respectively, unlike L5b SST⁺ interneuron, which is reported to be E10.5 (ref. ⁴¹).

Next, we compared thalamic innervation between LRP1 and LRP2.1 neurons. We labeled LRP neurons using an enhancer-driven AAV (AAV2/9-eHGT-600m-sYPF2) and performed co-immunostaining for GFP, Dach1 (a marker for LRP2.1), and VgluT2, which marks thalamic presynaptic terminals, at P4, P8 and P14 (Fig. 7a–e). At P4, both subtypes showed dense VgluT2⁺ puncta surrounding their soma and proximal dendrites (Fig. 7a,c,e). Thalamic innervation markedly declined by P8 and was largely absent by P14 (Fig. 7b,d,e), similar to the previous report for L5b SST⁺ cells⁴¹.

Fig. 7. LRP1 and LRP2.1 neurons receive comparable transient thalamic input but differ in cortical excitatory connectivity.

a–d, Confocal images of LRP neurons labeled with an enhancer-driven AAV (AAV2/9-eHGT-600m-sYPF2) and immunostained for GFP, VgluT2 and Dach1 at P4 and P14. Both LRP1 (a and b, GFP⁺/Dach1⁻, white arrowheads) and LRP2.1 (c and d, GFP⁺/Dach1⁺, white arrows) neurons showed dense VgluT2⁺ perisomatic puncta at P4, which declined markedly by P8 and were largely absent by P14. e, Quantification of VgluT2⁺ puncta density (spot number per μm² cell area) across ages. ***P < 0.001, two-factor ANOVA. P4, n = 3 animals, 34 cells LRP1 and 29 cells LPR2.1; P8, 3 animals, 65 cells LRP1 and 37 cells LPR2.1; P14, n = 4, 78 cells LRP1 and 26 cells LPR2.1. Box plots—centerline, median; box bounds, 25th and 75th percentiles; whiskers, most extreme values within 1.5× interquartile range. f, Schematic of thalamocortical slice preparation. g,h, Representative thalamus-evoked EPSCs in LRP1 (g) and LRP2.1 (h) neurons at P5–P6. i–k, Evoked EPSC success rate (i), latency jitter (coefficient of variance of latency; j) and evoked EPSC amplitude (k) did not differ between subtypes, n = 16 animals, LRP1 = 15 cells; LRP2.1 = 8 cells. l,m, Representative traces of sEPSCs recorded from LRP1 (l) and LRP2.1 (m) in thalamocortical slices, with post hoc immunostaining confirming subtype identity. n,o, sEPSC frequency (n) and amplitude (o) were similar between LRP1 and LRP2.1. p,q, Cortical slices (with severed thalamic afferents)—thalamic stimulation failed to elicit EPSCs in either subtype, n = 16 animals, LRP1 = 15 cells (p), LRP2.1 = 8 cells (q). r, sEPSC frequency from cortical slices with severed thalamic afferents was significantly lower in LRP2.1 neurons, n = 29 animals, LRP1 = 49 cells, LRP2.1 = 15 cells; **P < 0.01, two-tailed unpaired t-test. s, sEPSC amplitude remained unchanged; n = 29 animals, LRP1 = 49 cells, LRP2.1 = 15 cells. Error bars presented as mean ± s.e.m. (i–k, n, o, r, s). ANOVA, analysis of variance; VB, ventrobasal; CV, conduction. The scale bars (a,c,m,q) represent 10 μm.

To assess the synaptic strength of thalamic inputs onto LRPs, we performed whole-cell patch-clamp recordings from LRPs in thalamocortical slices preserving corticothalamic connectivity (Fig. 7f–h and Supplementary Table 3). At P5–P6, electrical stimulation of the thalamic axons reliably evoked excitatory postsynaptic currents (EPSCs) in LRP1 (GFP⁺/Dach1⁻) and LRP2.1 (GFP⁺/Dach1⁺) neurons, with comparable success rates (Fig. 7i), latency jitter (Fig. 7j), coefficient of variation of latency⁴⁴ and evoked EPSC amplitudes (Fig. 7k). Although EPSC latency slightly differed between LRP1 and LPR2.1, this timing difference is unlikely to reflect differences in synaptic input, as the evoked EPSC amplitude, rise time, jitter and success rate were indistinguishable (Supplementary Table 3). As thalamocortical axon conduction velocity varies along the fiber trajectory during early postnatal development due to differences in myelination⁴⁵, the observed latency difference likely reflects effective conduction delay rather than synaptic strength. Consistent with this interpretation, the frequency and amplitude of spontaneous EPSCs (sEPSCs), which reflect the combined contribution of thalamic and intracortical excitatory inputs, did not differ between the two subtypes (Fig. 7l–o).

Because rabies tracing indicated that LRPs received subplate connections, we asked whether LRP1 and LRP2.1 differ in their cortical inputs. For that, we performed patch-clamp experiments in cortical slices, where we severed the thalamocortical afferents. We verified the loss of thalamic connection by stimulating the thalamus and failing to elicit EPSCs in either LRP subtype (Fig. 7p,q and Supplementary Table 3). Under these conditions, the frequency of sEPSCs was significantly reduced in LRP2.1 neurons compared with LRP1 (Fig. 7r, **P < 0.01), while amplitude remained similar (Fig. 7s). These results show that, although both LRP1 and LRP2.1 receive comparable thalamic input, they differ in their cortical excitatory connections. Notably, LRP2.1 neurons are part of the cortical and thalamic circuits before their disappearance.

Discussion

Modular computational pipeline for analyzing rare cell types

Rare cell types are sparsely represented in individual scRNA-seq datasets, making it necessary to aggregate data across studies. However, integration across experiments introduces the risk of conflating biological variation with batch effects. Our modular pipeline addresses both problems by first enriching specific cell types from diverse datasets and then testing whether datasets generated under different conditions can be reliably integrated.

Successful cross-study integration depended strongly on technical compatibility. Datasets that differed substantially in tissue preparation, dissociation protocol or library chemistry often failed integrability thresholds and were excluded. This reflects current limitations of cross-study integration rather than a specific restriction of our approach. To make this explicit, we generated two atlases with different stringency levels—Dev-SST-v1, which includes more external data, and Dev-SST-v2, which retains only datasets passing all three criteria. Dev-SST-v2 showed greater cluster stability and lower cross-talk, supporting its use for downstream analyses.

The pipeline is also generalizable. We applied it to developmental PV⁺ interneurons and adult spinal cord datasets, showing that the same framework can be used across tissues, ages and experimental paradigms. To facilitate reuse, we provide the pipeline as open-source R/Nextflow-Conda code (Data availability and Code availability) together with a public web interface (https://shiny.gbiomed.kuleuven.be/AtlasLL/) for exploring gene expression, developmental stages and clusters. By combining new and published datasets, this framework enabled construction of the largest developmental atlas of SST⁺ interneurons to date.

Comparison of Dev-SST-v2 to our previous work

Dev-SST-v2 extends our previous atlas⁶ by increasing Sst⁺ sampling approximately sevenfold and improving integration through MapMyCells-based cell-by-cell annotation. Incorporation of the adult reference from ref. ³, including cortical and extracortical interneurons, enabled broader subtype classification than in our earlier study. We also clarified that the LRP2.1 population, initially annotated as striatal, is instead a transient cortical population with transient molecular similarity to striatal interneurons. Together, these advances reveal developmental states that were not resolved previously.

Distinct modes of diversification within cortical SST⁺ interneurons

Using Dev-SST-v2, we examined the timing and modes of diversification of SST⁺ cortical interneurons and found substantial differences in diversification strategies across SST⁺ cells. Two of these modes—invariant (predetermined) and expanding (stepwise) diversification—have been described previously in other neuronal types⁴⁶—the predetermined (invariant) mode has been reported for pyramidal neurons¹⁷, chandelier neurons⁴⁷ and Meis2⁺/CGE-derived interneurons⁴⁸, whereas expanding trajectories were observed in retinal ganglion cells⁴⁶. The invariant and expanding modes coexist within the Sst⁺ subclass and correspond to MC and nMC clusters, respectively, while a contracting mode characterizes the LRP population, which transitions through a transient developmental state before convergence.

An intriguing implication of our findings is that nMC neurons retain the capacity to alter their identity postmitotically. Supporting this idea of fluid model of cell-fate acquisition, a recent report shows that a fraction of mouse SST⁺ interneurons, when transplanted into human organoids, begins to express Pvalb⁴⁹. Given that some developing nMCs map to adult Pvalb⁺ cells, we speculate that these postmitotic populations are intrinsically amenable to changes in their identity.

Convergent model of diversification in LRP neurons

A major finding of our study is the contracting trajectory of LRPs. LRP1 and LRP2.1 are transcriptionally distinct by P5, but re-analysis of Bax cKO datasets indicates that LRP2.1 is selectively eliminated from P7 onward through PCD. Developmental apoptosis is a general feature of cortical interneuron maturation, but it has largely been viewed as indiscriminate pruning. Our data instead suggest that cell death can act with subtype specificity, providing a mechanism to sculpt inhibitory diversity with high precision.

PCD is a general developmental hallmark of cortical inhibitory neurons^33–35. This form of cell pruning was thought to target inhibitory neurons indiscriminately. We show here that PCD can also target specific inhibitory populations, providing a mechanism for fine sculpting of cortical circuits. Moreover, this mechanism appears to be conserved in humans as hLRP2.1 peaks at GW18–20 and is eliminated by birth, coinciding with the completion of PCD in the central nervous system⁴⁰. The parallel trends in mice and humans suggest that selective pruning of LRP2.1 is evolutionarily conserved and fine-tunes the repertoire of inhibitory neurons.

Functional implications of the transient LRP2.1 neurons

The transient nature of LRP2.1 raises questions about its function. Transient neuronal populations, such as subplate neurons, Cajal–Retzius cells and early (first wave) oligodendrocytes, are known to influence brain development, acting as scaffolds, and they shape circuits and connectivity^50–53. We show that LRP2.1 neurons, similarly, participate in the circuit before their elimination.

Our rabies tracing experiments indicate that LRP neurons are monosynaptically connected to multiple subplate neurons. The frequency of cortical sEPSCs in LRP2.1 neurons is lower than in LRP1 neurons, suggesting potential differences in intracortical excitatory input between the two LRP subtypes. Given that late-born subplate neurons undergo apoptosis while early-born ones persist to form layer 6b⁵⁴, we speculate that LRP subtypes may similarly engage with persistent versus transient circuit partners. Future studies examining the specific presynaptic partners of LRP2.1, as well as the consequences of LRP2.1 failing to undergo PCD in adulthood, could provide deeper insights into the function of this transient neuronal population.

In conclusion, our study presents a modular scRNA-seq computational pipeline for enriching and integrating rare neuronal populations, leading to the construction of the most comprehensive developmental atlas of SST⁺ cortical interneurons to date. We reveal three distinct modes of neuronal diversification—invariant, expanding and contracting. Notably, we describe a unique contracting trajectory in LRPs, where early diversity is downscaled through selective PCD. Additionally, we identified a transient cell type, termed the long-range, inhibitory, minority population neuron, which is evolutionarily conserved. These findings provide a data-rich foundational basis for understanding how neuronal diversity is generated and refined during the development of the mammalian cortex.

Methods

Mouse models

The mouse lines used in this study are as follows: WT C57BL/6J (strain 000664, RRID: IMSR_JAX:000664), Sst^Cre (B6N.Cg-Sst^{tm2.1(cre)Zjh/J}; strain 018973, IMSR_JAX:018973)⁵⁵, RCE (B6.Gt(ROSA)26Sor^{tm1.1(CAGEGFP)Fsh}; strain 032037, RRID: MMRRC_032037-JAX)⁵⁶, RC:FLTG (B6.Cg Gt(ROSA)26Sor^{tm1.3(CAG-tdTomato,-EGFP)Pjen/J}; strain 026932, RRID: IMSR_JAX:026932)⁵⁷ and Nkx2.1^flp (Nkx2-1^{tm2.1(flpo)Zjh/J}; strain 028577, RRID: IMSR_JAX:028577)⁵⁸. All adult mice were housed in groups and kept on a reverse 12-h light/12-h dark cycle regardless of genotype. Only time-mated pregnant female mice were housed individually. Male and female mice were used in all experiments. SST⁺ cells labeled by Sst^Cre mice were crossed with RCE reporter mice containing a floxed-stop cassette upstream of enhanced GFP (eGFP) in the ROSA gene locus to obtain Sst^Cre;RCE^fl/+. In some experiments, we isolated MGE-derived cells by using Nkx2.1^flp;RC:FLTG. All mice were maintained on a C57Bl/6 background and housed under a 12-h light/12-h dark cycle with ad libitum access to food and water. All animal procedures were approved by the KU Leuven animal welfare committee and the Scripps animal welfare committee under approved institutional animal protocols numbers KUL ECD 091-2020 and 090/2022, and Scripps 17-0025-3.

Tissue dissociation

Male and female SstCre/+;RCE mice were used for E16.5, P1 and P5 cortical dissociations. For E16.5, a full litter of five to eight mice was pooled together. For P1 samples, tissue from three animals was pooled; for P5 cortical samples, tissue from two animals was pooled; and, for P5 striatal samples, tissue from three animals was pooled. Tissue preparation was similar to ref. ⁶. These include samples from DS19, DS20, DS21, DS22, DS23, DS24 and DS26. Briefly, all solutions used in this procedure have an osmolarity of 290 ± 20 mOsm kg⁻¹. For E16.5 and P1, tissue from animals was pooled together. For P5, tissue from two animals was pooled together. The neocortex of E16.5, P1 and P5 SST^cre/+;RCE mice was dissected in an ice-cold HBSS buffer supplemented with 25 mM glucose.

Tissue was cut into small pieces and digested for 10 min at 34 °C with carbogen oxygenation using 30 ml of dissociation buffer solution, pH7.5, with the following composition: 0.2 mg ml⁻¹ pronase E (Sigma-Aldrich), 50 mM of trehalose, 10 mM glucose, 0.8 mM kynurenic acid, 0.05 mM APV, 0.09 M Na₂SO₄, 0.03 M K₂SO₄ and 0.014 M MgCl₂. After 10 min of incubation, E16.5 tissue was incubated at 34 °C for an additional 5 min with 375 µl of 10 mg ml⁻¹ DNase, while P1 and P5 tissue were incubated for an additional 5 min with 375 µl of 10 mg ml⁻¹ DNase and 750 µl of 10 mg ml⁻¹ Liberase (Sigma-Aldrich). Postenzymatic digestion tissues were washed once in an ice-cold dissociation buffer without enzyme. Tissue was mechanically triturated into single cells in 0.5 ml (for E16.5) or 5 ml (for P1 and P5) of OptiMEM (Invitrogen) supplemented with 2% trehalose, 10 mM glucose, 0.8 mM kynurenic acid and 0.05 mM APV using a volumetric pipette. For P1 and P5 tissue, myelin was removed by overlaying the single-cell suspension with 20% Percoll and centrifuging at 700g for 10 min at 4 °C. The cell pellet was resuspended in 400 µl of OptiMEM with 2% trehalose, 10 mM glucose, 0.8 mM kynurenic acid and 0.05 mM APV and tissue chunks were eliminated by passing cell suspension through a 40-mm cell strainer (BD Aria Fusion).

For sample DS18 only, pregnant C56BL/6 dams carrying E18.5 embryos (n = 3 animals), previously injected intraventricularly with AAV expressing eYFP, were anesthetized with isoflurane and killed by decapitation. The uterus was removed, the embryos were moved into cold Leibovitz’s L15 medium supplemented with 7 mM HEPES and each embryo was decapitated. From each brain, a section of the neocortex, corresponding roughly to S1, was dissected out with a microknife, avoiding the underlying striatal tissue. Tissue from four to five embryos of the same litter was pooled. Tissue sections were dissociated into single-cell suspensions using the Worthington tissue dissociation kit; the papain, DNase and albumin ovomucoid inhibitor solutions were reconstituted and oxygenated according to the manufacturer’s instructions. Tissue was digested in papain/DNAse for 15 min at 37 °C with carbogen oxygenation, triturated in EBSS using 1-ml lo-bind pipette tips, and then the cell suspension was mixed with albumin ovomucoid inhibitor/DNAse and centrifuged at 230g for 5 min in a swinging bucket centrifuge. Cells were resuspended in OptiMEM-trehalose medium (OptiMEM supplemented with 50 mM trehalose, 10 mM glucose, 0.5 mM kynurenic acid, 0.025 mM APV, 4 mM MgCl₂, 0.2 mM HEPES). Debris was removed by overlaying the suspension on top of a 5% OptiPrep gradient and centrifuging at 100g for 10 min. The cell pellet was resuspended in 500 µl of OptiMEM-trehalose media, and DAPI and DRAQ5 were added to a final concentration of 1 µM.

For dataset DS23, after myelin removal, the cell pellet was resuspended in 1 ml of 3% glyoxal with the following composition: 2.8 ml nuclease-free H₂O, 0.79 ml 100% EtOH, 0.31 ml 40% glyoxal (Sigma-Aldrich, 50649), 30 µl acetic acid, pH 4–5, for 15 min at 4 °C shaking. After incubation, cells were centrifuged at 800g for 10 min at 4 °C and resuspended in 500 µl of 1× PBS supplemented with 1% BSA and tissue chunks were eliminated by passing cell suspension through a 40-mm cell strainer (BD Aria Fusion).

FACS followed by 10x Genomics scRNA-seq

GFP⁺ and DAPI⁻ for live samples and GFP⁺ and Fixable Viability Dye eFluor 780 negative cells for fixed samples were isolated by FACS with a flow cytometer (BD Aria Fusion) using a 100 µm nozzle, about 3,000–5,000 events per s, at purity mode, and collected directly in 600 µl (cold at 4 °C) of OptiMEM with 2% trehalose, 10 mM glucose, 0.8 mM kynurenic acid and 0.05 mM APV solution in a 1.5-ml LoBind (Eppendorf) tubes precoated with 3% BSA overnight. FACS gating was set with a negative control of WT mice (Supplementary Fig. 5). We limited the FACS procedure to a maximum sort time of 25 min to ensure high cell survival. We collected 20,000–30,000 cells for each sample. We centrifuged the collected cells at 600g for 5 min at 4 °C and resuspended the FACS cells in 40 µl of OptiMEM with 2% trehalose, 10 mM glucose, 0.8 mM kynurenic acid and 0.05 mM APV. To determine cell number and viability, 5 µl of trypan blue was added to 5 µl of cells and cells were counted on a hemocytometer. In all experiments, we proceed to scRNA-seq only if >80% viability is observed. About 12,000 isolated single cells for each sample were loaded onto a custom droplet encapsulation platform, Hy Drop⁵⁹, for single-cell capture. cDNA library preparation was performed following the 10x Genomics commercial kit CG000315. RNA-seq was performed using BGI MGISEQ-2000 and Illumina NovaSeq 6000 (v1.5).

For sample DS18, the cell suspension was sorted for live eYFP⁺ cells using an Astrios EQ sorting flow cytometer with a 100-µm nozzle at two-way purity. Sorted cells were collected in 0.5 ml of cold FACS media (OptiMEM supplemented with 50 mM of trehalose, 10 mM glucose, 0.5 mM kynurenic acid, 0.025 mM APV, 0.2 mM HEPES). Approximately 20,000 cells were collected. Cells were centrifuged at 200g for 15 min at 4 °C in a swinging bucket centrifuge and resuspended in about 20 µl of FACS media. A total of 5 µl of the cell suspension was taken out for cell counting and viability with a Countess, and we only used suspensions with >80% live cells. Approximately 9,000 isolated single cells per sample were loaded onto a 10x Genomics single-cell chip for GEM formation and cDNA library preparation. RNA-seq was performed using an Illumina NextSeq 200 P2 flow cell.

Shallow and deep scRNA-seq

cDNA libraries were sequenced at two different depths. For shallow sequencing, we targeted a sequencing depth of 5,000 reads per cell; the actual achieved read number was 8,227.5 ± 793.5 (mean ± s.d.). For deep sequencing, we targeted 50,000 reads per cell; the actual read number achieved was 61,384 ± 20,484.

Data preprocessing

Sequencing data were prepared for analysis using the 10x Genomics Cell Ranger pipeline. First, Illumina and BGI BCL output files were demultiplexed into FASTQ format files. Feature counts were computed for individual GEM wells. STAR aligner⁶⁰ was applied to perform splicing-aware alignment to the mm10 reference genome, and then reads were bucketed into exonic, intronic and intergenic categories. Reads were classified as confidently aligned if they corresponded to a single gene annotation. Only confidently aligned reads were carried forward to UMI counting. A cell-calling algorithm, in conjunction with the barcode rank plot, was applied to filter out low-RNA-content cells and remove empty droplets. The output is a read count matrix. For publicly available datasets, we downloaded raw count matrices. We used the R package Seurat (v4.3.0.1) to perform the bulk of our subsequent data analysis, including filtering, normalization, scaling and other downstream processing.

Sample processing and filtering of Sst⁺ cells (module 1)

Each sample was first processed separately for cell filtering and identification of cortical Sst⁺ neurons. We discarded any genes detected in fewer than ten cells before filtering low-quality cells. Cells were retained if they met the following criteria: (1) unique gene count >700, (2) mitochondrial gene content <10% and (3) ln(total UMI) > mean(ln(total UMI)) − 2 s.d.(ln(total UMI)). To assess cell cycle phase differences in our dataset, we used the ‘CellCycleScoring’ function from Seurat. Cell cycle phase scores were calculated for each cell using predefined gene lists for the S phase and G2M phase. Normalization and scaling of reads were performed using the R package Seurat, with the functions ‘NormalizeData()’ and ‘ScaleData()’. During scaling, the number of total genes, percentage of mitochondrial gene expression and cell cycle phase difference score were regressed out to remove unwanted sources of variation from the data. Highly variable genes were defined using the Seurat function ‘FindVariableFeatures()’ and principal component analysis (PCA) was performed with the function ‘RunPCA()’. Doublets were identified using the scDblFinder⁶¹ (v1.12.0) algorithm with doublet rate set to 0.07. Cell clustering annotations were provided to scDblFinder, derived by applying ‘TransferData()’ function to the cluster labels of our previously published dataset⁶ using anchors identified by ‘FindTransferAnchors()’ function on top 40 PCs. The optimal number of PCs was estimated by performing a cross-validation on PCA⁶² and the resulting number of components was used as input to compute a k-NN graph further refined into a shared NN by the Seurat function ‘FindNeighbors()’. Clusters were defined using the Louvain algorithm implemented in ‘FindClusters()’ function. The Uniform Manifold Approximation and Projection (UMAP) embedding was also computed by running ‘RunUMAP()’ with the estimated optimal PCs. We selected cells belonging to clusters expressing Sst gene, discarding ‘stressed’ clusters, characterized by high expression of heat shock genes and lower average UMI, and clusters expressing Meis2, QC metric similar to ref. ⁶.

Integrability testing of datasets (module 2)

To test whether any given dataset can be used to generate an integrative atlas, we selected only datasets containing at least 1,000 high-quality Sst⁺ cells, as identified using previously described methods. The integrability of each dataset with our reference atlas (Atlas-v0) was evaluated using the following three criteria: anchor points, neighborhood composition and k-BET⁶³ rejection test.

Anchor points

The integration anchors between the testing datasets and Atlas-v0 were identified using the ‘FindIntegrationAnchors()’ function in the Seurat package, with reduction = ‘cca’. For each dataset, we subset 1,000 cells and compute the mean number of anchors for each testing dataset relative to Dev-SST-v0. This is termed a normalized mean anchor point, as we always subset 1,000 cells. A normalized mean anchor count >1,686 was set as the acceptance criterion.

Neighborhood composition

The testing dataset was integrated with Atlas-v0 using the ‘IntegrateData()’ function based on previously identified anchors to generate a corrected gene expression matrix. This matrix was scaled using ‘ScaleData()’, and PCs were computed using ‘RunPCA()’. Pairwise Euclidean distances were then calculated across cells in the testing dataset and cells in the reference atlas (Dev-SST-v0) within the integrated PCA space. For each test cell, the minimum distance to reference cells was recorded and added as metadata.

k-NN computation

NNs were identified using the ‘dbscan::kNN()’ function in PCA space. For each k value, the proportion of cells meeting the threshold of reference atlas cell representation was computed.

To evaluate neighborhood composition, we calculated the proportion of test dataset cells whose NNs contained at least half the expected fraction of reference atlas cells across a range of k values (5–200, in increments of 5). Datasets in which at least 50% of test cells had at least half the global fraction of reference atlas cells among their NNs (k = 60) were considered to meet the integration criteria. To assess the similarity of local transcriptomic structure between datasets, we constructed a k-NN graph for each dataset (k = 60). Two datasets were considered to have similar neighborhoods when at least 50% of the NNs (k = 60) of corresponding cells overlapped.

k-BET rejection rate

The k-BET⁶³ rejection rate (v0.99.6) was calculated to evaluate how well the testing dataset and reference atlas (Atlas-v0) were integrated in the corrected PCA space. A rejection rate below 50% was set as the acceptance criterion.

Only datasets meeting all three criteria (anchor points, neighborhood composition and k-BET rejection rate) were included in the generation of Atlas-v1.

Data integration and annotation (module 3)

Individual samples were integrated into a single object using CCA⁶⁴. Anchors were identified across samples with ‘FindIntegrationAnchors()’ function with parameters reduction = ‘cca’, k.filter = 100, k.score = 15 and anchor.features = 2000. A batch-corrected gene expression matrix was then computed using ‘IntegrateData()’ function. The resulting matrix was scaled using ‘ScaleData()’ function, regressing for number of total genes, percentage of mitochondrial gene expression and cell cycle score. PCA was performed with ‘RunPCA()’ function, and the optimal number of PCs was determined through cross-validation. Clustering was computed with ‘FindNeighbors()’ and ‘FindClusters()’ functions, based on the optimal number of PCs estimated. UMAP embedding was computed using ‘RunUMAP()’ function. Cells within ‘stressed’ clusters characterized by high expression of heat shock genes and low average UMI, doublets or clusters showing high expression of mitochondrial genes were further removed. Resulting cells were annotated into major groups of cortical Sst+ neurons (LRP, MC, and nMC) using cells annotated in the base atlas. This annotation was achieved by applying the ‘TransferData()’ function on anchors identified between each sample and the base atlas using the ‘FindTransferAnchors()’ function. Each cell was annotated to the group that achieved the maximum score.

Iterative clustering of Dev-SST-v0, Dev-SST-v1, Dev-SST-v2

Clustering algorithm was performed separately on each major group of cortical Sst⁺ neurons (LRP, MC and nMC). To compute clusters, we followed an iterative approach characterized by the following steps:

1. Sample filtering—remove samples with <20 cells.

2. Data integration and clustering—perform data integration using CCA and apply clustering as previously described. Clustering was computed at resolutions of 0.1, 0.2, 0.3, 0.4 and 0.5 using a k-NN graph built with k = 20 using the optimal number of PCs. Duplicated clustering (clusters from different resolutions that result in the same cell annotation) was considered only once.

3. Clustering stability evaluation—we performed 20 data subsamplings, each time sampling 80% cells without replacement, and repeated the clustering procedure for all selected resolutions on each subsample. Jaccard indices were calculated for each cluster at each resolution by comparing the original clusters with those obtained from each subsample. We defined a stable cluster as one with a Jaccard index 96 >0.75 in at least 75% of subsamples. A stable set of clusters was defined as one where >70% of cells belonged to stable clusters and stable clusters accounted for >60% of total clusters. These parameters were selected based on recommendations from ref. ⁶⁵. If no clustering was found to be stable, the iteration was stopped, and the cluster label obtained in the previous iteration was assigned as the final identity of all the cells. Otherwise, among stable clusterings, we selected the resolution that produced the higher percentage of cells within stable clusters. If two or more clusterings resulted in the same percentage, the resolution with the higher number of clusters was chosen.

4. Cluster identity evaluation and merging—clusters with <300 for MCs and nMCs and 150 cells for LRPs were merged with their closest cluster—centroids were calculated in batch-corrected PCA space and the closest cluster was defined as the one with the smallest Euclidean distance across centroids. For each resulting cluster, marker genes were identified by comparing gene expression with all other cells not belonging to the cluster. FindMarkers() function from Seurat was used with parameters logfc.threshold = 1 and only.pos = TRUE. Genes with an adjusted P value <0.01 were selected and the sum of the min (−log₁₀(p.val_adj), 20) values across all selected genes was calculated (DEScore). If the resulting value was ≤60, the cluster was merged with its closest cluster as previously described. The DEScore cutoff of >60 was retained from our previous work⁶.

5. Process iteration—after identity evaluation, if only one cluster remained, the iteration was stopped and the cluster label obtained in the previous iteration was assigned as the final identity of the cluster. If more than one cluster remained, the process was repeated from steps 1 to 5 for all clusters with 600 cells or more for MC and nMC, and 300 for LRPs. Clusters with <600/300 cells were returned as the final identity.

6. After all clustering iterations ended, cells removed at step 1 of each iteration were annotated. The k-NN constructed during iteration 1 was used to identify the NNs for each removed cell, and each cell was assigned to the cluster most represented among its NNs. Finally, we computed marker genes for all clusters using ‘FindAllMarkers()’ function from Seurat with parameters logfc.threshold = log₂(1.5), min.pct = 0.2 and only.pos = TRUE. Based on the resulting gene list, we built a hierarchical clustering dendrogram of clusters using Euclidean distance. We then identified the fifth percentile of the distribution of cophenetic distances across all cluster pairs, where the cophenetic distance is defined as the dendrogram height at which the branches of the two clusters merge. For all pairs of clusters with cophenetic distance below this threshold, we computed the DEScore, as previously defined, in a one-versus-one cluster comparison. If the two clusters had a DEScore ≤60, they were merged into a single cluster.

Random forest classification and atlas comparison

To evaluate and compare the performance of two independently generated scRNA-seq atlases (Dev-SST-v1 and Dev-SST-v2), we trained random forest classifiers to predict cell-type identities from gene expression profiles. Classification models were constructed separately for each major interneuron family (LRPs, MC and nMC). For every cell type within each family, we computed confusion matrices quantifying how frequently each true class was assigned to each predicted class.

From these matrices, we extracted per-class performance metrics—diagonal accuracy is defined as the fraction of cells correctly predicted for a given class (true-positive rate per class). Cross-talk is defined as (1 − (diagonal accuracy)), representing the degree of misclassification to other classes. For each atlas (v1 and v2), we summarized these metrics and computed per-class differences (Δ = v2 − v1) for both diagonal accuracy and cross-talk. Positive Δ (diagonal) values indicate improved classification accuracy in v2, while negative Δ (cross-talk) values indicate reduced confusion between classes.

Cell annotation validation

To validate cell identities, we performed fivefold cross-validation, taking 20% data as a test set and the remaining 80% as the training set. We computed differentially expressed genes for all clusters in the five training sets using ‘FindMarkers()’ function with options logfc.threshold = ln(1.5), only.pos = TRUE, min.pct = 0.5. We then proceed in a pairwise fashion, using ‘randomForest’⁶⁶ R package (v4.7-1.1). We trained a random forest classifier with 100 trees on each pair of clusters in the training data using only the top ten differentially expressed genes, ranked by log(FC), from each cluster. This classifier was used to predict the identity of cells in the test set. After fivefold cross-validation, we have a predicted identity for every cell for each pairwise comparison of clusters. The above procedure was repeated for a total of ten iterations. From the consistency of the outcome across these iterations, we extracted a measure of the stability of the cell classification. We used cell_type_mapper, the backend Python package of the MapMyCell tool, to perform cell-type annotation on custom datasets.

WOT analysis

We performed WOT⁶⁷ analysis to infer lineage relationships and cell-fate transitions across developmental time points in the major subgroups of cortical Sst⁺ neurons (LRP, MC and nMC). All analyses were conducted using the WOT command-line interface (v1.0.8). For both MC and nMC subsets, the cluster sizes were imbalanced. To address this, we computed the median cluster size across all clusters and randomly subsampled this number of cells from each cluster. This approach allowed us to generate a more balanced dataset for fate prediction. The subsampling procedure was repeated thrice, and WOT analysis was performed independently on each random subset. First, transport maps were generated between consecutive developmental stages using the optimal_transport function. As input, we used a log-normalized count matrix restricted to highly variable genes to reduce noise and focus on the most informative features. The growth_iters parameter was set to 3 to allow estimation of population growth rates across iterations, thereby improving transport map stability. Next, fate probabilities were computed with the fates function, applied to the resulting transport maps. This procedure estimated, for each cell, the probability of transitioning into a given cluster identity at subsequent time points. To facilitate interpretation, cells at each time point were grouped according to their cluster assignments, and for each cluster we calculated the median fate probabilities toward all possible cluster identities at the following time point. For MC and nMC subsets, mean values across the three subsampled datasets were reported.

Cell annotation to adult counterparts

To annotate our data to their adult counterparts, we used the MapMyCells³ algorithm, using the 10× whole mouse brain (CCN20230722) as reference taxonomy and using the hierarchical mapping as chosen algorithm. All annotations of cell identity in the adult reference were used; we did not restrict the dataset by brain region or cell class. Briefly, to annotate unlabeled cells to their corresponding class, subclass, supertype or cluster, the algorithm compares the gene expression of each unlabeled cell with the average expression profile of each child node at every level of the hierarchy. This comparison is performed using marker genes that effectively discriminate among child nodes, assigning a vote to the child node with the highest Pearson correlation. For robustness, the process is repeated 100 times, randomly selecting 90% of all marker genes every time. A cell is assigned to the child node with the highest number of votes, and cell labels were considered valid if 90 of 100 annotations assigned the cell to the same child node. We then defined a 1:1 matching between one of our clusters and an annotated supertype if at least 65% of cells within the cluster were assigned to that supertype, and if the majority of all cells annotated to that supertype originated from that cluster, representing at least 30% of all cells annotated to the supertype.

Cortical and striatal integration

SST⁺ interneurons were sorted from striatum and processed as previously described (module 1). Seurat label transfer analysis was performed to map cells to the major clusters (LRP, MC or nMC) using the FindTransferAnchors() function followed by the TransferData() function. Over 90% of cells were LRPs, and these cells were selected and integrated with P5 cortical LRP from Dev-SST-v2. To place all cells assigned as LRP in the same UMAP embedding, we performed FastMNN analysis⁶⁸ using the RunFastMNN() function from SeuratWrappers package (v0.3.1). Differentially expressed genes between the cortical LPR2.1 and striatal interneurons were identified using the FindMarkers() function, with parameters min.pct = 0.2 and logfc.threshold = 1.

Trajectory and pseudotime analysis

We performed trajectory analysis on the LRP clusters. First, we computed gene modules that captured distinct sources of variation within the data. We used the R package ANTLER⁶⁹ (v0.9.0) to generate these modules. Computation began with a normalized counts-per-million matrix. Genes present in fewer than ten cells were discounted from the computation. Genes were filtered before clustering to retain only those genes with a Spearman correlation of at least 0.3 with at least five other genes. These genes were clustered hierarchically using Spearman correlation as a dissimilarity metric. This hierarchical clustering was repeated over many iterations, with a heuristic algorithm determining an optimal number of modules. At each iteration, modules were discarded if they failed to meet validity thresholds on the expression level of individual genes. At the terminal iteration, all modules met the validation criteria and were returned as the final set of gene modules. To understand the module contents, we performed functional enrichment on each module and examined Gene Ontology terms. We selected 335 genes from modules associated with terms including ‘neuron differentiation’, ‘axon development’ and ‘synapse assembly’. We computed diffusion maps with destiny⁷⁰ R package (v3.12.0) on the corrected expression matrix, considering only the selected genes. The eigenvalues of the diffusion map transition matrix were used as diffusion components for low-dimensional embedding. The pseudotime for each cell was computed from the full transition matrix using the ‘DPT()’ function and represents a notion of diffusion distance from a root cell. The root cell was identified as the cell from the earliest age samples that had the highest number of early sample cells in its neighborhood (computed on diffusion components). From this analysis, we inferred linear branches diverging from an initial shared state. We then identified genes showing continuous changes over pseudotime by fitting each gene’s expression vector to pseudotime using a generalized additive model, with the ‘gam()’ function from ‘mgcv’⁷¹ R package (v1.9-0). Genes were classified as early genes (negatively correlated with pseudotime) or late genes (positively correlated with pseudotime). Late genes were further subdivided into those expressed only in one branch.

Saturation analysis

We subsampled cells without replacement at percentages ranging from 5% to 95% in 5% intervals. For each percentage, we generated ten different subsamples. To estimate the saturation point, we calculated several metrics. First, we computed robust Hausdorff distance between each subsample and the original dataset in the batch-corrected PCA space, selecting the 50th largest value as the distance measure. Next, we identified marker genes for each cluster in each subsample and recorded the total number of marker genes. Additionally, we calculated the absolute differences in mean expression for marker genes between the subsamples and the entire atlas for each cluster. The mean of differences across clusters was summed over all differentially expressed genes. Finally, we assessed the fraction of clusters that maintained their identity in each subsample. For this, cluster centroids were calculated based on gene expression vectors, and clusters from the whole dataset were matched to the centroids of the subsamples using ‘MetaNeighborUS()’ function from MetaNeighbor⁷² R package (v1.18.0). Clusters with an AUROC score of >0.9 when matched to their centroids in the subsample were considered to have retained their identity. For each subsampling percentage, we computed the mean scores for all metrics (robust Hausdorff distance, number of differentially expressed genes, sum of absolute differences in mean expression and percentage of clusters preserving their identity) across the ten subsamples. The inflection point, indicating the saturation percentage, was identified as the point where the second derivative of the metrics changed sign.

Cloning of viral vector and generation of AAV

The AAV enhancer plasmid, pAAV-eHGT-m600-sYFP2 (ref. ³¹), was provided by Boaz Levi (Allen Institute). To generate pAAV-eHGT-m600-dio-sYFP2, sYFP2 was removed with restriction enzymes BamH1 and EcoRV, and the backbone was isolated by gel purification. The insert, dio-sYFP2, with a 20-bp homology on both 5′-end and 3′-end to the backbone was synthesized by Twist Bioscience as a gene fragment. Backbone and insert were assembled by NEBuilder HiFi DNA Assembly (NEB, E2621L) per manufacturer’s recommendation and transformed into NEB Stable competent cells (NEB, C3040I). Using the transfer vectors pAAV-eHGT-600m-sYFP2 and pAAV-eHGT-600m-dio-sYFP2, AAV2/9 particles were custom generated by BrainVTA at a viral titer of >3 × 10¹³ vg ml⁻¹.

Using the transfer vectors pAAV-eHGT-m600-sYFP2 and pAAV-eHGT-m600-dio-sYFP2, AAV2/9 particles were custom generated by BrainVTA at a viral titer of >3 × 10¹³ vg ml⁻¹. The rabies helper plasmid pAAV-syn-Flex-nGToG was obtained from the Charité Vector Core. To generate the LRP-restricted helper plasmid pAAV-600m-Flex-nGToG, the Flex-nGToG cassette was subcloned by PCR from pAAV-syn-Flex-nGToG into the pAAV-eHGT-m600-dio-sYFP2 backbone. Newly generated plasmids and associated sequence information are available on Addgene (249627 and 249628) respectively. AAV2/9 particles were custom produced by VectorBuilder, at titers >1 × 10¹³ vg ml⁻¹.

Neonatal injection of AAV

AAV injections were performed on male and female P0 WT CD1 mice or SST^Cre/+ mice under hypothermia-induced anesthesia. Injections into the lateral ventricle were conducted using glass micropipettes (World Precision Instruments, 1B120F-4) connected to a pressure microinjector device (World Precision Instruments, µPUMP). The injection site was identified at half the distance from the lambda suture to each eye. The glass micropipette was inserted perpendicularly into the skull to a depth of ~2 mm, with a slight decrease in resistance indicating entry into the lateral ventricle. Approximately 1 µl of the virus (AAV2/9-eHGT-600m-SYFP2 or AAV2/9-eHGT-600m-dio-SYFP2) was rapidly injected into each ventricle. After the injections, pups were placed on a warming pad to recover until they could move normally and were then returned to their mother.

Neonatal injection of AAV and rabies virus

Male and female SST^Cre/+ mice received rabies helper AAVs containing a Cre-dependent GFP, TVA and optimized G protein (Flex-nGToG), injected into the lateral ventricles of Sst^Cre mice at postnatal day 0 (P0). Each brain received 2 µl of helper virus (final titer 5 × 10¹² vg ml⁻¹). Two helper viruses were used to generate starter cells from either all SST⁺ neurons or LRP SST⁺ neurons, n = 3 animals per condition. For pan-SST⁺ targeting, pAAV-syn-Flex-nGToG (titer > 1 × 10¹³ vg ml⁻¹) was obtained from the Charité Vector Core. For LRP-restricted targeting, pAAV-600m-Flex-nGToG was subcloned by PCR from pAAV-syn-Flex-nGToG using pAAV-eHGT-600m-dio-sYFP2 as the backbone. AAV2/9 particles were custom produced by VectorBuilder at titers of >1 × 10¹³ vg ml⁻¹. At P4, the rabies pseudovirus N2c-ΔG-Tdt-Flp (Charité Vector Core; titer = 1.06 × 10⁹ IU ml⁻¹) was injected into the somatosensory cortex using a pulled glass capillary and a nanoinjector (NanoLiter 2020; World Precision Instruments). Three injections of 50 nl each were delivered at 25 nl min⁻¹. Injection sites were defined relative to the intersection of the superior sagittal and transverse sinuses—injection 1, DL = 3.1 mm, ML = 2.2 mm, depth = 0.6 mm; injection 2, DL = 3.3 mm, ML = 2.1 mm, depth = 0.6 mm; injection 3, DL = 3.5 mm, ML = 2.0 mm, depth = 0.6 mm. Injection accuracy was verified by co-injecting fluorescent beads and by post hoc histological validation. Starter cells (GFP⁺ and TdTomato⁺ cells) and presynaptic connected cells (TdTomato⁺ cells) were quantified by an image segmentation script written in Python.

Rabies quantification and statistics

Three male and female brains per starter line were analyzed, comprising a total of 54 sections, with 9 sections per brain. For each brain, the number of Red⁺ cells per mm² in each region of interest was normalized to 100 starter cells and averaged across all sections to yield section-level mean densities per region and starter type. Differences between starter lines were assessed using linear mixed-effects models with starter type as a fixed factor and brain identity as a random intercept to account for repeated sections. Separate models were fit for each anatomical region and for each outcome (normalized Red⁺ density and Double⁺ density). Statistical significance was assessed using a linear mixed-effects model with brain identity included as a random intercept. P values were calculated using Wald z tests for fixed effects and adjusted for multiple comparisons using the Holm method. Analyses were performed in Python (statsmodels, matplotlib).

Histology and immunochemistry

Adult and postnatal pups were deeply anesthetized with an intraperitoneal injection of ketamine (87 mg kg⁻¹ body weight) and xylazine (13 mg kg⁻¹ body weight) and transcardially perfused with ice-cold PBS followed by 4% paraformaldehyde (PFA). The brains were isolated and incubated overnight in 4% PFA on a slow-speed rocking platform at 4 °C. Only for VgluT2 staining, postnatal pups were transcardially perfused with ice-cold PBS followed by 9% glyoxal, 8% acetic acid solution (pH4) as a fixative. Tissue was washed with PBS, incubated in 10% followed by 30% sucrose in PBS and cut on a freezing microtome into 100-μm coronal sections, and stored in cryoprotectant solution (30% glycerol, 30% ethylene glycol in PBS) at −20 °C or processed or free-floating immunohistochemistry. Coronal sections were stained using the following primary antibodies: rabbit anti-Dach1 (1:500; Proteintech, 10914-1-AP; RRID: AB_2230330), chicken anti-GFP (1:3,000; Aves Lab, GFP-1020), rabbit anti-MafB (1:500; Sigma-Aldrich, HPA005653), rabbit antinitric oxide synthase 1—NOS1 (1:1,000; Immunostar, 24287; RRID: AB_572256) and guinea pig anti-VgluT2 (1:1,000; Merck, AB2251). Tissue was blocked with 2% BSA, 10% goat serum in 0.25% Triton X-100 PBS for 1 h at room temperature. Primary antibodies were incubated overnight at 4 °C in 2% BSA, 10% goat serum in 0.25% Triton X-100 PBS. The following secondary antibodies conjugated to Alexa Fluor dyes were used: goat antichicken IgY (H + L) Alexa Plus 488 (Thermo Fisher Scientific, A32931), and donkey antirabbit IgG (H + L) Alexa Plus 647 (Thermo Fisher Scientific, A32795). Secondary antibodies were all used at 1:1,000, incubated for 2 h at room temperature, followed by DAPI (5 μM; Sigma-Aldrich, D9542) for 10 min at room temperature. Sections were washed thrice for 10 min with 0.1% Triton X-100 PBS after secondary and DAPI staining. Brain sections were mounted in Mowiol-DABCO (25% Mowiol; Sigma-Aldrich, 81381; 2.5% DABCO). Brain sections were imaged using an inverted Ti2-E Nikon microscope with a CrestOptics X-Light V3 spinning disk. Subsequent image processing was performed using Fiji or MATLAB.

Birthdating of LPR1 and LPR2.1

Time-mated CD1 pregnant mice were injected at E11.5, E12.5, E13.5 or E14.5, and male and female pups were analyzed at P5. Time-mated CD1 pregnant mice (plug day designated as E0.5) received three intraperitoneal injections, at 4-h intervals, of the thymidine analog EdU (30 mg kg⁻¹; Abcam, ab219801) on embryonic days E11.5, E12.5, E13.5 or E14.5. At postnatal day 5 (P5), pups were transcardially perfused with ice-cold PBS, followed by 4% PFA. Brains were dissected, postfixed in 4% PFA for 3 h at 4 °C, washed in PBS, embedded in agarose and sectioned at 60 µm on a vibratome (Leica, VT1200S) for free-floating immunohistochemistry. EdU incorporation was detected using the manufacturer’s click-chemistry protocol (EdU Click Chemistry Assay Kit; Abcam, ab219801). Briefly, sections were permeabilized for 30 min, then subjected to the click-reaction mixture containing TBS, CuSO₄, iFluor 488 azide and EdU additive solution. Sections were subsequently washed in PBS with 0.1% Triton X-100 (3 × 10 min) and immunostained with primary antibodies against Nos1 (1:1,000; Immunostar; RRID: AB_572256) and Dach1 (Proteintech, 10914-1-AP; RRID: AB_2230330). Antibodies were directly conjugated using the CoraLite Plus Flexible 2.0 labeling kits (Sanbio, KFA503-50 and KFA504-50). Images were acquired on an inverted Nikon Ti2-E microscope equipped with a CrestOptics X-Light V3 spinning disk confocal unit. Postacquisition processing and quantification were performed using custom MATLAB scripts.

Image analysis and quantification of LRP1 and LRP2.1 by histology

The following two types of staining were used for the analysis: (1) GFP and Dach1, and (2) GFP and MafB. LRP1 was defined as GFP⁺ cells expressing high levels of MafB but low levels of Dach1. In contrast, LRP2.1 was defined as GFP⁺ cells expressing low levels of MafB but high levels of Dach1. Quantification of Dach1 and MafB expression levels was performed using MATLAB with a customized script, which is publicly available on GitHub. Briefly, GFP⁺ cells were selected based on a manual thresholding function determined by the user. For each GFP⁺ cell, the centroid was identified and expanded into a circular region with a 4-pixel radius. The mean intensity signal of either MafB or Dach1 within this circular region was then calculated. For Dach1, mean intensity values in the GFP⁺ cells that were at least 2× above the background intensity were classified as high Dach1. For MafB, as all GFP⁺ cells expressed MafB, the intensity values in GFP⁺ cells were fitted to a normal distribution. Cells with MafB intensity levels above the 75th percentile (mean + 0.524 s.d.) were classified as high MafB. For quantification of EdU in LPR1 or LPR2.1, Nos1 channel was used as initial segmentation mean intensity signals were calculated for Dach1 and EdU channels. Quantification of Vglut2 puncta area was performed by thresholding the GFP channel to compute cell area; thresholding of theVglut2 staining was used to quantify Vglut2 spot area.

Slice preparation for electrophysiology

Acute brain slices were prepared from male and female SST^cre/+ mice injected with AAV2/9-eHGT-600m-dio-SYFP2. At P4–P6, mice were anesthetized by hypothermia and decapitated. Brains were rapidly dissected in ice-cold, oxygenated (95% O₂ and 5% CO₂) sucrose-based cutting solution containing 83 mM NaCl, 2.5 mM KCl, 1 mM NaH₂PO₄, 26.2 mM NaHCO₃, 3.3 mM MgSO₄, 0.5 mM CaCl₂, 22 mM D-glucose and 72 mM sucrose (pH 7.4). To preserve the thalamocortical pathway, brains were cut at a 55° angle relative to midline in combination with an additional 10° ramp. Somatosensory thalamocortical slices (350 μm) were prepared (Leica, VT1200) and allowed to recover in cutting solution at 32 °C for 30 min, followed by at least 30 min at room temperature before recordings. Initially, tetramethylrhodamine crystals (3000 MW; Thermo Fisher Scientific) were introduced into the thalamus to guide selection of slices with intact thalamocortical pathway.

Electrophysiology patch-clamp recordings and analysis

Slices were transferred to a submerged chamber (Warner Instruments) and continuously perfused (3–4 ml min⁻¹) with oxygenated artificial cerebrospinal fluid (ACSF, 32–33 °C) containing 127 mM NaCl, 2.5 mM KCl, 2.5 mM NaH₂PO₄, 2 mM CaCl₂, 2 mM MgCl₂, 25 mM NaHCO₃, 25 mM D-glucose. Bicuculline (20 μM) was added to the ACSF in all recordings. Recording electrodes (3–6 MΩ) were pulled from borosilicate glass capillaries using a horizontal P-1000 Flaming/Brown micropipette puller (Sutter Instrument) and filled with an internal solution containing 126 mM CsMeSO₃, 10 mM HEPES, 2.5 mM MgCl₂, 4 mM ATP, 0.4 mM GTP, 10 mM creatine phosphate, 0.6 mM EGTA, 5 mM QX-314-Cl and 3 mg ml⁻¹ biocytin, adjusted to pH 7.25 with CsOH. Data were acquired using a Multiclamp 700B amplifier (Molecular Devices), digitized at 20 kHz, and low-pass filtered at 3 kHz through a Digidata 1440 A interface (Molecular Devices). pClamp 10 software (Molecular Devices) was used for acquisition. Input resistance, series resistance and holding current were continuously monitored to ensure recording stability and data quality. Series resistance (<20 MΩ) was compensated by 70–75%. sEPSCs were recorded in whole-cell voltage-clamp mode at a holding potential of −70 mV for 3 min. For evoked EPSC recordings, an extracellular stimulation electrode (cluster microelectrodes, FHC) was positioned onto the internal capsule axons exiting the ventrobasal thalamic nucleus. To ascertain activation of single axons, we first determined the minimal stimulation intensity required to successfully evoke EPSC in 50–60% of trials (10–45 V, A-M Systems Model 2100). This voltage was subsequently used for evoked recordings. Evoked EPSCs were analyzed using Easy Electrophysiology, and sEPSCs were analyzed using Mini Analysis software (Synaptosoft).

Statistics and reproducibility

No statistical method was used to predetermine sample size. Sample sizes were chosen based on the prior experience with comparable scRNA-seq, histological, rabies tracing and electrophysiological experiments, and were selected to obtain sufficient numbers of high-quality cells, biological replicates, sections or recordings for downstream analyses. Male and female mice were used in all experiments.

No data were excluded from the analyses, except for predefined technical and QC criteria described in the ‘Data preprocessing’ section. These included cell viability thresholds before scRNA-seq, filtering of low-quality cells, putative doublets, stressed clusters, non-Sst⁺ cells or clusters, datasets failing predefined integration criteria and electrophysiological recordings that did not meet stability criteria.

The experiments were not randomized. Animals and samples were assigned to experimental groups based on genotype, developmental stage, anatomical region, viral condition, birthdating time point or starter line, as required by the experimental design. Data collection was not randomized or formally blocked. No randomization was used in the organization of experimental conditions or stimulus presentation. The investigators were not blinded to allocation during experiments and outcome assessment. Covariates including developmental stage, genotype, anatomical region, viral condition and brain identity were accounted for in the relevant experimental designs or statistical analyses, as appropriate for each experiment.

Statistical analyses were performed using R and Python. Differential gene expression, clustering, integration, classification, trajectory, saturation and transport analyses were performed using the indicated computational workflows. Before statistical testing, data distributions were assessed for normality, and equality of variances was assessed using a variance test. For comparisons in which the data were not normally distributed, a nonparametric test was used. Where applicable, individual data points are shown in Figs. 2f,g, 6a,b,f,g,h,i and 7e and Extended Data Fig. 9d,g to illustrate data distribution.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41593-026-02387-w.

Supplementary information

Supplementary Information ^{(6MB, pdf)}

Supplementary Figs. 1–5.

Reporting Summary ^{(78KB, pdf)}

Supplementary Tables 1–4 ^{(73.5KB, xlsx)}

Supplementary Table 1: List of genes used in MERFISH. Supplementary Table 2: ANTLER gene modules used for pseudotime analyses. Supplementary Table 3: Electrophysiological parameters of thalamocortical eEPSC at P4–P6. Supplementary Table 4: Summary of data and statistical analyses.

Author contributions

M.L., F.F.R., E.M., G.B. and L.L. conceptualized the study. M.L., E.M., F.F.R., R.M., E.N., G.B., D.K., K.W., M.R., H.B., J.X.D. and L.L. performed the methodology. M.L., E.M., F.F.R., R.M., G.B., H.B., K.W., M.R., J.X.D. and L.L. performed the investigation. M.L., E.M., F.F.R., D.K. and G.B. conducted the analysis. F.F.R., M.L., G.B., R.M., E.M. and L.L. contributed to writing the methods. M.L., F.F.R., G.B. and L.L. contributed to writing, reviewing and editing the original and final draft of the manuscript. L.L., A.T. and G.L. acquired funding. L.L., A.T., K.W. and G.L. performed resource management and were responsible for supervision.

Peer review

Peer review information

Nature Neuroscience thanks Christian Mayer and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Funding

This work was supported by the Research Foundation—Flanders (FWO) Junior Research Project (G057121N), FWO-Odysseus Award (G0E9121N), FWO/FNRS EOS (G0I7822N), iBOF (IBOF/25/010) and VIB Tech Watch grant (Techwatch—Vizgene; to L.L.), FWO Junior/Senior Research Projects (G074823N and G0A3Y24N to A.T.), International Foundation for Research in Paraplegia, Research Grant (P188 to A.T.), National Institutes of Health (grants R01NS121223 and RF1MH126719 to G.L. and F31NS118982 to J.X.D.) and FWO PhD fellowship (1192822N to E.M.). The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.

Data availability

All previously published scRNA-seq data are downloaded and available on GEO per Extended Data Table 1. All plotted values and statistical analyses are summarized in Supplementary Table 4. scRNA-seq data generated in this study are accessible on GEO (DS18—GSE280655, DS19–22—GSE271016, DS23–24—GSE271947, DS26—GSE315700 and P5 striatal interneuron sequencing data—GSE315702). Dev-SST-v2 can be accessed through a web interface at https://shiny.gbiomed.kuleuven.be/AtlasLL/. Raw MERFISH (spatial data) are deposited on BioStudies, accession S-BSST3005, at https://www.ebi.ac.uk/biostudies/studies?query=S-BSST3005. All other types of data and source data files to generate each graph are deposited on Zenodo at https://doi.org/10.5281/zenodo.20139187 (ref. ⁷³). Any other types of data can be requested through the corresponding author.

Code availability

All customized analysis codes written in R, Python, Nextflow (Conda environment) and MATLAB used to generate Dev-SST-atlas and individual figure plots/panels can be found on GitHub at https://github.com/Limlab-VIBCBD/Dev-SST-atlas.

Competing interests

All authors declare no competing interests.

Extended data

is available for this paper at https://doi.org/10.1038/s41593-026-02387-w.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41593-026-02387-w.