Developmental molecular signatures define de novo cortico-brainstem circuit for skilled forelimb movement Authors: Julia Kaiser, Payal Patel, Sam Fedde, Alexander Lammers, Matthew Kenwood, Asim Iqbal, Mark P Goldberg, Vibhu Sahni Published: 2026 ABSTRACT AbstractSkilled movement relies on descending cortical projections to the brainstem and spinal cord. The brainstem integrates cortical inputs from subcerebral projection neurons (SCPN). Here, we identify a previously unrecognized direct cortico-brainstem circuit that emerges early in development and persists into adulthood. A subset of SCPN project exclusively to the brainstem from the earliest stages of axon extension, never extending to the spinal cord. We find that during development, these cortico-brainstem neurons (CBN) can be prospectively identified by Neuropeptide Y (Npy) expression. In adulthood, Npy+ CBN project preferentially to rostral brainstem motor regions. Silencing Npy+ CBN impairs the precision and targeting of forelimb movement in a skilled reaching task. Together, these results outline how molecular programs lay the foundation for adult cortico-brainstem circuits underlying skilled movement. Our findings reveal the developmental logic diversifying descending cortical circuits and enable investigation of specific descending projection neuron subsets in motor control.Subject terms: Development of the nervous system, Motor control, Neural circuits Introduction Skilled movement depends on precise communication between the neocortex and subcerebral motor centers in brainstem and spinal cord1–3. The brainstem, once viewed mainly as a relay for posture and reflexes, is now recognized as a key integrative hub for fine motor control4. It receives dense input from diverse cortical areas5–9 and contributes critically to skilled forelimb movement10,11. Cortical output reaches these subcerebral targets through two principal populations of projection neurons: cortico-brainstem neurons (CBN), often referred to as “cortico-bulbar” neurons, which project exclusively to the brainstem, and corticospinal neurons (CSN), whose axons extend to the spinal cord but also collateralize within the brainstem. Traditionally, CBN were thought to control face and eye movements by innervating cranial motor nuclei in the brainstem, whereas CSN were viewed as the main drivers of limb and trunk movements via spinal targets1,12. However, CSN axons collateralize extensively across multiple levels of the neuraxis, including the striatum, thalamus, midbrain, and brainstem5,6, which suggests that CSN output engages both brainstem and spinal circuits in parallel. CBN, in turn, provide a direct cortical pathway to brainstem circuits involved in dexterous forelimb control13,14. Together, these findings suggest that CBN and CSN form a distributed cortico-subcerebral network that coordinates cortical, brainstem, and spinal activity to enable skilled movement. CBN and CSN reside interdigitated in layer V of cortex and extend axons through the same white matter tract. As a result, distinguishing these subpopulations has proved to be experimentally challenging as anatomical methods cannot selectively label one population without the other. There are no molecular tools to date that specifically identify CBN as distinct from CSN, leaving their precise projection patterns and functional roles in movement poorly defined. Therefore, how CBN are specified during development, and how they might specifically contribute to functional output in adulthood, largely remains unclear. Early investigations into subcerebral projection neurons (SCPN) axon development suggested that the distinction between CBN and CSN is established via large-scale developmental pruning. During early development, SCPN axons extend more broadly to a larger set of targets than will be retained in the adult. These exuberant projections are then refined through selective axon pruning so that in adulthood, specialized neocortical areas establish projections to specific subsets of targets in the brainstem and spinal cord15–17. This pruning to generate area-specific connectivity is controlled by molecular differences between SCPN residing in distinct cortical areas18. Using such molecular differences, SCPN can be reprogrammed from spinal-projecting to proximal-projecting types, effectively re-routing their axons to target more proximal levels of the neuraxis19. These recent investigations also identified a subset of proximal-projecting SCPN that emerge de novo, independent of pruning, suggesting that a subset of SCPN establishes a direct connection to the brainstem, without ever projecting to the spinal cord. An intriguing possibility is that this subset is molecularly distinct from other SCPN that extend axons to the spinal cord during early development. Such cell-intrinsic molecular delineation has been described for SCPN subsets that exhibit similar axon extension specificity at the transition between cervical and thoracic spinal segments20–22, suggesting that it might be possible to molecularly distinguish CBN from CSN during initial SCPN axon extension, prior to the large-scale axonal pruning events described above. In the present study, we establish that there is SCPN axon extension specificity at the brainstem-spinal transition zone from the earliest stages of axon extension. We find that a subset of SCPN limits its axon extension to the brainstem, establishing that there is early differentiation of CBN from CSN. We further establish that CBN can be molecularly differentiated from CSN even within the same cortical areas during early development, indicating that axon extension specificity represents a key driver of SCPN molecular diversification. This now provides unprecedented molecular access to CBN, as a distinct subset from CSN, throughout development and into maturity. Using this molecular distinction, we selectively silenced CBN (Npy+) in adulthood and reveal that this disrupts precision during a trained single pellet reaching task, causing mistargeted approaches and the emergence of novel error types not seen under baseline conditions. These findings indicate that this newly identified, developmentally defined cortico-brainstem circuit governs specific elements of skilled movement control. Finally, we use this early molecular delineation to identify anatomical differences in brainstem innervation by CBN versus CSN axons, with CBN axons preferentially innervating more rostral brainstem nuclei than CSN axons. These anatomical findings suggest circuit level substrates underlying cortical control over skilled forelimb movement. Collectively, our results identify early developmental molecular programs over SCPN diversification, which lays the foundation for participation by distinct SCPN subpopulations in anatomically and functionally distinct circuits at maturity. These findings not only deepen our understanding of SCPN development but also highlight how molecular control of early subpopulation differentiation has lasting implications for adult motor function, providing avenues to identify, define, and manipulate previously unknown cortico-brainstem circuits essential for skilled movement control. Results Cortico-brainstem neurons restrict axon extension to supraspinal levels from the earliest stages of SCPN development Subcerebral projection neurons (SCPN) extend their axons from layer V in neocortex to subcerebral targets along the rostro-caudal axis, ultimately innervating either brainstem (cortico-brainstem projection neurons, CBN) or spinal cord (corticospinal projection neurons, CSN) (Fig. 1a). Previous work has shown that from the earliest stages of axon extension into the cord, and prior to the formation of any functional connectivity, SCPN exhibit axon extension specificity at the cervical-thoracic transition zone21. This transition zone represents a critical region of differential SCPN axon extension, where specific SCPN subpopulations either continue or halt their axon extension. We hypothesized that such axon extension specificity might similarly occur at other transition zones in the white matter across the neuraxis. We analyzed the transition between the brainstem and the cervical spinal cord, i.e., the “brainstem-spinal” transition zone, to investigate whether developing SCPN axons navigating through the brainstem could similarly exhibit distinct axon extension specificity. The brainstem-spinal transition zone is established early in development. a Retrograde labeling strategy: Subcerebral projection neurons (SCPN) extend axons to subcerebral targets. At P0 (upper panel), SCPN axons can be found at caudal brainstem levels. P0 CTB injection into cerebral peduncles labels all SCPN, including both putative cortico-brainstem neurons (CBN) that extend only to the brainstem, and corticospinal neurons (CSN) that extend to the spinal cord. At P2 (lower panel), axons of CSN have extended into the spinal cord; thus, injection of CTB at C2 exclusively labels CSN. b SCPN retrogradely labeled from cerebral peduncles reside throughout the medio-lateral extent of sensorimotor cortex. c The majority of CSN labeled from spinal cord reside in medial sensorimotor cortex, with few CSN within the lateral sensorimotor cortex (zoom-in views). d Quantification of retrogradely labeled CTB+ neurons across three medio-lateral regions (cingulate, medial, lateral; vertical lines on schematic) and at four rostro-caudal levels (1.95, 2.67, 3.26, 4.23 mm; Atlas of the Developing Mouse Brain; P6). CSN were enriched in medial cortex, nearly absent in cingulate cortex, and reduced in lateral cortex, where most SCPN projected only to brainstem (linear mixed-effects model, REML; two-sided tests: medial F(1,9) = 4.10, p = 0.0736, 105.3 [−12.36, 222.9]; cingulate F(1,9) = 25.32, p = 0.0007, 105.8 [58.23, 153.4]; lateral F(1,9) = 33.27, p = 0.0003, 283.1 [172.0, 394.1]). No region × level interaction was observed (all p > 0.05) n = 6 animals for SCPN, n = 5 animals for CSN). Scale bars, (b, c) 500 µm. Data are mean ± SEM (d). Source data are provided as a Source Data file. During early postnatal development, all SCPN axons extend through the cerebral peduncle, a white matter tract within in the brainstem that they must traverse en route to the spinal cord. Within the spinal cord, corticospinal axons extend within the dorsal funiculus, a major spinal white matter tract (Fig. 1a). We asked whether all SCPN extend axons to the spinal cord during this early developmental period of axon extension, or whether a subset terminates its axon extension within the brainstem. To distinguish between these possibilities, we injected cholera toxin B (CTB), a retrograde tracer, into white matter tracts either at the level of the cerebral peduncle or into the dorsal funiculus of the cervical spinal cord. CTB labels only those neurons whose axons have already reached the injection site. Therefore, the retrogradely labeled neurons from these two distinct levels indicate how far different SCPN axons have extended in the neuraxis. The spinal injection selectively labels only those SCPN whose axons had extended beyond the brainstem and into spinal white matter. These injections were performed at developmental stages when SCPN axons are actively extending from the brainstem into the spinal cord23. By postnatal day 0 (P0), all SCPN axons have extended through the cerebral peduncle. CTB injection at this level labels all SCPN, including both CSN and putative CBN (Fig. 1a, upper panel). By P2, SCPN axons have extended into the cervical spinal cord; injections at this level therefore selectively label the CSN subset. Neurons labeled from the cerebral peduncle but not from the spinal cord would be putative CBN (Fig. 1a, lower panel). Because these injections were performed before collateral formation or pruning, differences in retrograde labeling directly reflect axon extension rather than later remodeling. CTB injection into the cerebral peduncles at P0 produced broad retrograde labeling across layer V spanning the mediolateral extent of the sensorimotor cortex (Fig. 1b), whereas injection into the spinal cord at P2 labeled fewer neurons (Fig. 1c). Dual retrograde labeling within the same mice (Extended Data Fig. 1a, b) confirmed these results. This indicates that a subset of SCPN axons terminate within the brainstem (i.e., they are CBN), revealing an early restriction of axon extension well before collateralization or target innervation. To determine how this early divergence is organized across the cortex, we quantified retrogradely labeled SCPN and CSN at both rostro-caudal and medio-lateral levels of sensorimotor cortex (Fig. 1d). CSN are predominantly located in medial cortex, where their numbers closely matched the total SCPN population (CBN + CSN). In contrast, the cingulate cortex contained only few CSN, indicating that most SCPN in this region project to the brainstem only (CBN). Similarly, the lateral cortex also contained a smaller proportion of CSN relative to all SCPN, indicating a predominance of CBN in this region. Within-animal normalization in the dual retrograde labeling experiments (Extended Data Fig. 1d) confirmed these spatial differences. Cingulate cortex contained very few CSN (2.5 ± 0.99%), indicating that nearly all SCPN in this region are CBN. In contrast, medial and lateral cortices contained higher proportions of CSN (32.9 ± 2.1% and 27.4 ± 3.98%, respectively), consistent with a predominance of CBN across both regions. Importantly, the relative enrichment of SCPN over CSN was consistent across all rostro-caudal levels examined, indicating that cortical area-specific projection neuron identity is maintained throughout the anterior-posterior axis (Fig. 1d and Extended Data Fig. 1d). It is possible that some CSN axons extend more slowly than others and that, at P2, a subset of CSN axons may not have yet reached the spinal cord, thereby appearing as CBN in these analyses despite ultimately projecting to the spinal cord. To directly test this possibility, we performed additional retrograde labeling from the cervical spinal cord at P4 (Extended Data Fig. 1e, f). If the distinction between CBN and CSN observed at P2 were due solely to a developmental delay in axon extension, a greater number of CSN would be expected to reach the spinal cord by P4 and be thus retrogradely labeled at P4. However, we observed no significant differences in the number of labeled CSN across cortical regions between P2 and P4. These results indicate that the distinction between CBN and CSN reflects true axon extension specificity rather than a simple developmental delay (Extended Data Fig. 1e, f). CSN counts at P4 were comparable to those at P2 across all cortical regions, indicating that the balance between CSN and CBN is fixed early, with CBN axons maintaining their restriction to supraspinal levels. These results further establish that the projection distinction between CBN and CSN is already stable by P2. Together, these results demonstrate that CBN axons remain confined to supraspinal levels from the earliest stages of development, establishing the brainstem-spinal transition zone as a critical site of differential SCPN axon extension specificity. Molecular specification of SCPN subpopulations occurs across two independent dimensions: axon extension specificity and cortical location The axon extension specificity of CBN and CSN suggests that their distinct projections are encoded by different molecular programs. To investigate the underlying molecular signatures that distinguish developing CBN from CSN, we retrogradely labeled SCPN with CTB injected into the cerebral peduncles at P0 (CBN + CSN), or the cervical cord at P2 (CSN only), and isolated labeled neurons by FACS for single-cell RNA-sequencing (scRNA-Seq) (Fig. 2a and Extended Data Fig. 2a). SCPN were collected at P1 and P3, and CSN at P3, representing developmental stages immediately before and after CSN axons extend into the spinal cord. While CSN axons can also collateralize in the brainstem, these early developmental stages precede the onset of axon collateralization23, and these early molecular distinctions likely reflect differential axon extension rather than collateral formation. Accordingly, the CSN annotation in these analyses was based solely on axon extension to the spinal cord, independent of any subsequent collateralization within the brainstem. Molecular specification of SCPN subpopulations based on axon extension specificity and cortical location. a SCPN or CSN were retrogradely labeled by CTB injection into cerebral peduncles at P0 (SCPN; CBN + CSN) or cervical cord at P2 (CSN), followed by FACS isolation and whole-cell scRNA-seq of cortical neurons at P1 or P3. b Integrated molecular landscape of developing SCPN and CSN, colored by sample (left panel), axon extension specificity (middle panel), or cortical location (right panel). Axon extension specificity distinguishes SCPN-only (CBN) from shared SCPN/CSN (CSN) populations; cortical location separates medial and lateral neurons. c Manual annotation of 6 unbiased clusters based on cortical location (medial, med; lateral, lat) and axon extension specificity (CBN, B; CSN, S). d Cluster stability highlights lateral CBN (B.lat) as the most stable cluster (n = 20 subsampling iterations). e Top 5 differentially expressed genes across clusters: Npy and Cartpt are the top differentially expressed genes for lateral CBN (B.lat) and CSN (S.lat), respectively (genes highlighted). f Pseudotime analysis indicated that lateral CBN (B.lat) are most developmentally advanced among SCPN clusters. Box plots show median (centre), 25–75% (box), and min-max (whiskers) (d). Source data are provided as a Source Data file. FACS-purified SCPN populations yielded high-quality scRNA-Seq data, and neurons corresponding to cortical layers V–VI, representing SCPN, were isolated for downstream analysis (Extended Data Fig. 2a–c). Integration with the adult Allen Brain Atlas24 (P56 Whole Cortex & hippocampus) showed that developing SCPN aligned most closely with adult layer 5 PT glutamatergic neurons, confirming shared projection identity (Extended Data Fig. 2d). However, when integrated exclusively with adult PT neurons, developing SCPN remained transcriptionally distinct, indicating that the molecular programs defining mature PT identity are not yet fully established at these stages. We next examined the transcriptional organization of developing SCPN to determine whether early molecular diversity relates to cortical location and axon extension specificity. In previous work, we established that SCPN in medial and lateral cortex differ in their subcerebral projections to segmental targets: lateral SCPN extend axons only to the brainstem and cervical spinal cord, whereas medial SCPN also include a distinct subpopulation that extends axons to more caudal spinal levels in the thoracic and lumbar spinal cord21. To investigate how these differences are reflected transcriptionally, we integrated our datasets (SCPN at P1, P3; CSN at P3) to generate a unified molecular landscape of developing SCPN. CSN formed a subset of the broader SCPN population, consistent with our anatomical observations that only a fraction of SCPN project to the spinal cord. These findings indicate that CSN are transcriptionally distinct from SCPN that do not extend axons to the spinal cord (Fig. 2b). Unbiased clustering of developing SCPN revealed six transcriptionally distinct and stable clusters (Fig. 2d, e). We annotated clusters along two independent axes: axon extension specificity (CBN vs. CSN) and cortical location (medial vs. lateral) (Fig. 2b and Extended Data Fig. 2c). Clusters with cells predominantly from SCPN samples but not CSN samples were annotated as “CBN”, while clusters with cells in both samples were annotated as “CSN” (Fig. 2b and Extended Data Fig. 2c). Cortical location was inferred from transcriptional similarity to medial and lateral SCPN as previously defined21. Based on these annotations, clusters were designated as cortico-brainstem (B) or corticospinal (S) and further subdivided by cortical region (med vs. lat) (Fig. 2c). Notably, both CBN and CSN could be molecularly distinguished within medial and lateral cortex, indicating that cortical location and axon targeting specificity represent distinct molecular programs that independently contribute to SCPN transcriptional diversity. Consistent with previous findings21, we observed greater transcriptional heterogeneity among medial SCPN (4 out of 6) compared to lateral SCPN (2 out of 6). Specifically, medial SCPN subdivided into one cortico-brainstem (B.med) and three corticospinal clusters (S.med, S.med1, and S.med2). In contrast, lateral SCPN segregated into two clusters: one cortico-brainstem (B.lat) and one corticospinal (S.lat) cluster (Fig. 2c). This likely reflects the broader diversity of subcerebral targets in medial cortex, where distinct SCPN subpopulations extend axons to all levels of the spinal cord, including projections past the cervical cord to thoracic and lumbar levels21. In fact, all CSN that project to thoraco-lumbar segments reside exclusively in medial cortex. In contrast, SCPN in lateral cortex extend axons exclusively to brainstem and cervical spinal levels21. Accordingly, cluster stability analysis showed overall higher stability for lateral clusters (Fig. 2d), consistent with the relative homogeneity of SCPN in lateral cortex. Given that lateral SCPN axons do not extend beyond the cervical cord21, the spinal lateral cluster likely represents cervical-projecting CSN (CSNc). It is worth noting that in the integrated UMAP, CSN appear more numerous than CBN, which contradicts the anatomical quantification in Fig. 1. This is a consequence of integrating two samples to generate the UMAP, SCPN (CBN + CSN) and CSN-only, resulting in CSN being overrepresented in the embedding. Accordingly, the anatomical data in Fig. 1 more accurately reflect true CBN and CSN abundance. We next investigated whether all SCPN differ in their developmental maturation at this stage. Since CBN limit axon extension at a time when CSN axons are still extending, we hypothesized that CBN might be more differentiated. To assess this, we performed pseudotime analysis, which arranges cells along a continuous trajectory based on how similar or different their gene-expression profiles are as they mature. This approach allowed us to place each SCPN cluster along the same developmental timeline and compare their relative developmental progression and maturation relative to each other (Fig. 2f). Pseudotime analysis confirmed that the lateral CBN cluster (B.lat) was the most developmentally advanced, whereas all other clusters showed comparable developmental states. These results indicate that analysis of SCPN developmental progression requires integrating both cortical location and axon extension specificity, and that even within the same cortical region, distinct SCPN subpopulations differ in their developmental advancement, with CBN in lateral cortex being more developmentally advanced than CSN. Finally, to identify molecular correlates of SCPN subpopulation identity, we examined the top differentially expressed genes across all clusters (Fig. 2e). Consistent with prior reports, we found Crim1 and St6galnac5, two genes previously associated with CSN in medial cortex20,21, enriched in a spinal medial cluster (S.med1). In lateral cortex, Neuropeptide Y (Npy) and CART prepropeptide (Cartpt) were the top genes differentially expressed in the cortico-brainstem (B.lat) and corticospinal (S.lat) clusters, respectively (Fig. 2e, genes highlighted). These results reveal distinct molecular signatures that define CBN and CSN subpopulations and link molecular identity to both cortical location and axon extension specificity. Npy and Cartpt expression define molecularly distinct SCPN subpopulations in postnatal cortex To validate the annotation of the six transcriptionally distinct SCPN clusters by cortical location and axon extension specificity, we focused on clusters in lateral cortex, which contains only CBN and CSNc, two spatially interdigitated SCPN subpopulations with distinct subcortical targets. We identified Npy and Cartpt as the top differentially expressed genes distinguishing CBN and CSN in lateral cortex: Npy expression was restricted to the B.lat cluster and decreased over time (mean expression level 3.93 arb. units ± 1.25 at P1 vs. 2.92 ± 1.14 at P3), whereas Cartpt was specific to S.lat and increased as CSN axons extended into the spinal cord (Fig. 3a, mean expression level 1.37 arb. units ± 0.52 at P1 vs. 1.74 ± 0.69 at P3). Temporal and spatial expression profiles of Npy and Cartpt delineate distinct SCPN subpopulations during postnatal development. a Npy is specifically enriched in cluster “B.lat”, and Cartpt in cluster “S.lat”. Expression levels differ between P1 and P3, with Npy higher at P1 and Cartpt higher at P3 (two-sided Wilcoxon rank-sum test on cells with non-zero expression: Npy W = 289,459, p < 2.2 × 10−16, n = 812 and 489 cells; Cartpt W = 153,586, p < 2.2 × 10−16, 342 and 1336 cells). b SCPN were retrogradely labeled at P0, and Npy and Cartpt expression in cortex was analyzed using RNAScope at P1, P3, P7, and P14. c Representative RNAScope images showing Npy and Cartpt expression within SCPN. d Quantification of Npy⁺ and Cartpt⁺ SCPN (normalized to CTB⁺ SCPN) across cortical regions. Npy⁺ SCPN are enriched in lateral cortex, peaking at P3 and declining by P14, while Cartpt⁺ SCPN are also enriched in lateral cortex with peak expression at P7 (two-way ANOVA; two-sided: Npy time × region F(6,18) = 4.13, p = 0.0088; Cartpt time × region F(6,18) = 3.31, p = 0.0225; n = 4 animals/timepoint; 7–8 sections/animal ( ~ 300 µm apart)). *p < 0.05, **p < 0.01, ***p < 0.001. Scale bars, (c) 50 µm. Box plots show median (centre), 25–75% (box), and min-max (whiskers) (a, d). Source data are provided as a Source Data file. We next examined the temporal and spatial expression profiles of Npy and Cartpt using RNAScope from P1 to P14 (Fig. 3b). While Npy is expressed by multiple distinct neuronal populations across the brainstem, the limbic system, and spinal cord25,26, within cortex, it is known to be expressed by interneurons27; however, recent studies have identified Npy expression in a subset of excitatory projection neurons during early cortical development28. To specifically investigate Npy and Cartpt expression in SCPN, we combined RNAScope with retrograde labeling from the cerebral peduncles and quantified expression of each gene in retrogradely labeled SCPN across four developmental time points (Fig. 3b). Overall, cortical Npy expression decreases across postnatal development, with higher expression levels at early time points that decline by P14 (Extended Data Fig. 3d). At early postnatal stages (P1–P3), although Npy is expressed broadly across both medial and lateral cortex, its expression within SCPN is largely confined to those located in lateral cortex. The proportion of Npy⁺ SCPN in lateral cortex peaks at P3 (36.9% ± 14.2), and subsequently declines sharply, with very few Npy⁺ SCPN by P14 (Fig. 3c, d, 2.2% ± 0.44). Similarly, Cartpt+ SCPN were also restricted to lateral cortex. The proportion of SCPN expressing Cartpt increased at P3 (51.95% ± 31.02) and remained comparably high through P7 before declining by P14 (Fig. 3c, d, 16.84% ± 7.80). Notably, Npy and Cartpt expression by SCPN was mutually exclusive, identifying distinct SCPN populations in lateral cortex. Importantly, both genes were only transiently expressed during early postnatal development, marking a brief developmental window in which molecular signatures align with initial stages of axon extension (Fig. 3c). Additional genes further supported the distinctions between the two SCPN subpopulations. Klhl14, previously identified as enriched in SCPN in lateral cortex20,21, labeled both CBN and a subset of CSN (Extended Data Fig. 4). This is in line with results obtained from our recently developed Klhl14-Cre knock-in mouse line that showed labeling of SCPN in lateral cortex, including a subset of CSN29. In contrast, genes with unique expression in B.lat or S.lat clusters aligned more cleanly with axon extension specificity: In lateral cortex, Sema3e (B.lat) is expressed only by retrogradely labeled SCPN relative to CSN (31% ± 8.8 in SCPN vs. 3.2% ± 2.4 in CSN), indicating specific expression by CBN (Extended Data Fig. 4). In contrast, assessing S.lat cluster genes is more challenging because retrograde labeling of SCPN includes both CBN and CSN. CSN-specific genes will thus be counted in both SCPN and CSN; however, they are expected to show greater enrichment in CSN than SCPN. In line with this, Alcam and Lrrtm3 showed higher expression in CSN than SCPN in lateral cortex (Alcam: 40% ± 21 in CSN vs. 28% ± 4.2 in SCPN; Lrrtm3: 56% ± 48 in CSN vs. 28% ± 5.6 in SCPN). This enrichment is comparable to Cartpt, indicating at least CSN-enriched expression (Extended Data Fig. 5). Collectively, these results identify genes that delineate developing CBN from CSNc in lateral cortex. To confirm the relationship between molecular identity and axon extension specificity, we assessed whether Npy+ and Cartpt+ SCPN correspond to CBN and CSN, respectively. Retrograde labeling from cervical spinal cord combined with RNAScope, revealed minimal Npy expression in CSN (Extended Data Fig. 3a, P7 1.7% ± 0.54). In contrast, Cartpt was robustly expressed in CSN predominantly in lateral cortex with a peak at P7 (Extended Data Fig. 3a–c, 42% ± 12). These findings demonstrate that Npy and Cartpt expression define distinct SCPN subpopulations within lateral cortex that differ in their axon extension specificity. Taken together, these results confirm the annotation of cortical location and axon extension specificity, delineating B.lat and S.lat clusters in our dataset. These genes define molecular identifiers that distinguish CSNc from CBN at the earliest stages of SCPN axon extension into the spinal cord. The identification of these genes does not imply that they directly control the differential axon extension specificity. In fact, this is unlikely, as their developmental expression does not align with the time at which extension decisions are made. For example, Cartpt expression in SCPN increases between P3 and P7, whereas corticospinal axons have already committed to extending beyond the brainstem by P1–P2. Even so, our findings show that these two SCPN subpopulations are distinguished by a molecular program at these early developmental stages. This in turn suggests that the differential axon extension by CBN and CSN is likely governed by differential molecular mechanisms, even if Npy and Cartpt themselves are not the causal factors. Instead, these genes function as readouts of a broader transcriptional program that ultimately establishes the divergent connectivity patterns of these two distinct subpopulations. Npy+ and Cartpt+ SCPN are distinct subpopulations in lateral cortex Transcriptomic profiling revealed two molecularly distinct SCPN subpopulations in developing lateral cortex, characterized by Npy or Cartpt expression. To validate and anatomically define these populations, we used knock-in Cre lines to selectively label Npy+ and Cartpt+ SCPN. Conditional retrograde labeling from cerebral peduncles at P0 revealed that both Npy+ and Cartpt+ SCPN are predominantly located in lateral cortex at P14 (Fig. 4a–d). In Npy-Cre mice, labeled SCPN were overwhelmingly located in lateral cortex (947.7 cells ± 100.4), with only sparse labeling in medial cortex (250.3 cells ± 122.2) and only few in cingulate cortex (30 cells ± 22). Similarly, Cartpt-Cre mice showed predominance of Cartpt⁺ SCPN in lateral cortex (444.3 cells ± 182.5), with comparatively few cells in medial (85.3 cells ± 38.1) or cingulate cortex (38.3 cells ± 7.5). These distributions closely match those observed by RNAScope and confirm that Cre expression in each knock-in line faithfully reflects endogenous gene expression. Npy+ and Cartpt+ SCPN are distinct subpopulations in lateral cortex. a Cre-dependent retro-AAV (rAAV) GFP was co-injected with control rAAV (TdTomato, TdT) at P0 into cerebral peduncles to label Cartpt+ or Npy+ SCPN in Npy-Cre or Cartpt-Cre mice, respectively. b, c Coronal section at P14 showing broad TdT+ labeling of SCPN, with Cre-dependent GFP expression restricted to lateral cortex in Npy-Cre (b) and Cartpt-Cre mice (c). d, e Quantification of GFP⁺ SCPN across cortical regions (n = 3 animals/Cre line, 7–8 sections/animal, 300 μm apart): (d) Npy⁺ SCPN are enriched in lateral cortex (one-way ANOVA; two-sided: F(2,6) = 80.49, p < 1 × 10−4; Tukey’s multiple comparisons: lateral vs cingulate p < 1 × 10−4, lateral vs medial p = 0.0002). e Cartpt⁺ SCPN show a similar pattern (one-way ANOVA; two-sided: F(2,6) = 12.75, p = 0.0069; Tukey’s multiple comparisons: lateral vs cingulate p = 0.0087, lateral vs medial p = 0.0153). f Dual retrograde labeling of Cartpt+ and Npy+ SCPN using Npy-FlpO;Cartpt-Cre double transgenic mice. Cre- and FlpO-dependent rAAVs were co-injected into cerebral peduncles at P0. g Whole-brain optical clearing of dual-labeled Npy-FlpO;Cartpt-Cre mice at P35 showing Cartpt+ and Npy+ SCPN enriched in lateral cortex. h Coronal section at P14 showing Cartpt+ and Npy+ SCPN are distinct subpopulations. i Npy⁺ SCPN reside in more superficial cortical layers than Cartpt⁺ SCPN (linear mixed-effects model, REML; two-sided test for the fixed effect: t(4.09) = −0.21, p = 0.843; n = 528 Cartpt⁺ and 2948 Npy⁺ SCPN across 3 animals). C cingulate, M medial, L lateral cortex, scWM subcortical white matter, pia pial surface. *p < 0.05, **p < 0.01, ***p < 0.001. Scale bars, (c, h) 500 µm, (h’–h”’) 50 µm. Scale bars are omitted in (g) due to tissue distortion. Cre genotypes: +/+ (homozygous), +/− (heterozygous). Data are mean ± SEM (d, e). Box plots show median (centre), 25–75% (box), and min-max (whiskers) (i). Source data are provided as a Source Data file. Since these experiments were performed using separate Cre mouse lines, they could not determine whether Npy+ and Cartpt+ SCPN represent distinct subpopulations. We thus generated Npy-FlpO;Cartpt-Cre double-transgenic mice and performed dual conditional retrograde labeling using FlpO- and Cre-dependent reporter retro-AAVs injected into the cerebral peduncles at P0 (Fig. 4f). Optical clearing at P35 confirmed that both Npy+ and Cartpt+ SCPN are located in lateral cortex (Fig. 4g–g”), with areal distributions closely matching those seen in retrograde labeling experiments performed separately in Npy-Cre and Cartpt-Cre mice (Extended Data Fig. 6a). We further mapped the location of Npy+ and Cartpt+ SCPN across isocortical areas within the Allen Mouse Brain Common Coordinate Framework (CCFv3)30, a standardized anatomical atlas that enables comparison of neuronal distributions across brain regions. We found an enrichment of both Npy+ and Cartpt+ SCPN in primary and supplemental somatosensory cortex, with additional Npy+ SCPN located within primary motor cortex (Extended Data Fig. 6b). Morphological analyses showed no difference in soma size between groups (Extended Data Fig. 6c), but Npy⁺ SCPN reside in slightly more superficial cortical layers compared to Cartpt⁺ SCPN (Fig. 4i). We also sought to determine whether Npy+ and Cartpt+ SCPN are truly distinct subpopulations or if there is some overlap between them. We analyzed the retrogradely labeled SCPN NpyFlpO;CartptCre mice to determine whether any neurons exhibited both mCherry (Npy-FlpO+) and eGFP (Cartpt-Cre+) labeling (Fig. 4h–h”). However, we found almost no double-positive neurons (1.38% ± 1.21 of mCherry+ SCPN display a very faint eGFP signal; conversely 0% of eGFP+ SCPN display any mCherry signal; n = 3 mice), demonstrating a lack of overlap between the Npy⁺ and Cartpt⁺ SCPN subpopulations. Together, these findings demonstrate that Npy⁺ and Cartpt⁺ SCPN are discrete, developmentally defined SCPN subpopulations that share a common cortical location but differ in molecular signature and laminar position. Axon extension specificity delineates Npy+ SCPN as CBN, and Cartpt+ SCPN as CSN Our findings indicate that Npy and Cartpt expression delineates axon extension differences between CBN and CSN at the brainstem-spinal transition zone early in development. However, Npy and Cartpt expression is transient in development, suggesting that this molecular distinction may not persist into adulthood, consistent with our analysis showing that developing SCPN only partially align with, but remain distinct from, adult PT neurons (Extended Data Fig. 2d). Thus, early molecular programs delineate axon extension during development, but their expression declines as circuits mature, raising two possibilities with distinct implications for how early axon extension specificity at the brainstem-spinal transition zone shapes adult connectivity. One possibility is that these initial differences reflect a temporary developmental state, with Npy⁺ SCPN later extending axons into the spinal cord and acquiring CSN-like circuitry using a subsequent set of molecular cues. Alternatively, early molecular programs may durably define cortico-brainstem circuitry, establishing stable patterns of connectivity that persist even after Npy expression declines. In that case, Npy⁺ SCPN would continue to limit their projections to the brainstem into maturity, establishing a dedicated de novo cortico-brainstem pathway. To address this and distinguish between these two possibilities, we used Npy-Cre and Cartpt-Cre mice to test whether their respective SCPN populations exhibit distinct axon extension specificity. Retrograde labeling from the cervical spinal cord (C2) at P3, followed by analysis at P14 (Fig. 5a), revealed robust labeling of GFP+ CSN in Cartpt-Cre mice, particularly in lateral cortex (103 cells ± 12.7), but only sparse labeling in Npy-Cre mice (18.3 cells ± 12.7; Fig. 5c and Extended Data Fig. 6d). These findings indicate that at this early developmental time point, Cartpt+ SCPN extend axons to the spinal cord, while Npy+ SCPN limit their axon extension to supraspinal levels. These results mirror the in situ findings that Cartpt⁺ SCPN project to the spinal cord, whereas Npy⁺ SCPN restrict axon extension to supraspinal targets, validating that Cre expression in each line captures the anatomical differences between these subpopulations. Npy+ and Cartpt+ SCPN in lateral cortex exhibit persistent axon extension specificity. a Cre-dependent retro-AAV (rAAV) GFP was co-injected with control rAAV (TdTomato, TdT) at P2 into dorsal funiculus of cervical spinal cord to label Cartpt+ or Npy+ CSN in Npy-Cre or Cartpt-Cre mice, respectively. b, c Coronal sections at P14 show no flex-GFP labeling in lateral cortex of Npy-Cre mice (b), indicating that Npy⁺ SCPN do not project to the spinal cord, whereas Cartpt-Cre mice show robust retrograde labeling of Cartpt⁺ CSN from the cervical spinal cord (c). Representative images; see Extended Data Fig. 6a for quantification (n = 3 animals). d Anterograde labeling at P1 using FlpO-dependent TdT and Cre-dependent GFP injected into lateral cortex of Npy-FlpO;Cartpt-Cre double transgenic mice. e, f Dorsal (e) and ventral (f) whole mount views of a Npy-FlpO;Cartpt-Cre mouse brain at P14, showing the distinct trajectories of Cartpt+ SCPN axons extending to the spinal cord and Npy+ SCPN axons terminate within brainstem (white arrow), representative of independent experiments with similar results. g–g” Optically cleared cervical spinal cord from one P35 Npy-FlpO;Cartpt-Cre brain shows Cartpt+ CSN axons extending into spinal cord, while Npy+ SCPN axons are absent. h–h”’ Lateral view of an optically cleared brainstem of the same sample shows termination of Npy+ SCPN axons within brainstem, while Cartpt+ SCPN axons continue beyond into the spinal cord. See also Supplementary Movie 1. i Proposed model of SCPN axon extension specificity in lateral cortex: Npy+ CBN limit axon extension to brainstem, while Cartpt+ CSNc extend axons past brainstem into cervical spinal cord. Scale bars, (b, c, e, f) 500 µm, (b’–b”’, c’–c”’) 50 µm. Scale bars are omitted in (g, h) due to tissue distortion. Cre genotypes: +/+ (homozygous), +/− (heterozygous). We next asked whether this early distinction persists as circuits mature. We performed conditional anterograde labeling at P0, a developmental stage when Npy and Cartpt are robustly expressed and Cre activity can be reliably used to access these subpopulations, by injecting both a Cre-dependent AAV and a control AAV into the lateral cortex of either Npy-Cre or Cartpt-Cre mice. The control AAV produced broad labeling of all axonal projections arising from lateral cortex. We find no discernible differences across genotypes, confirming that our injection sites encompass similar cortical areas across all injected mice (Extended Data Fig. 6g). In contrast, Cre-dependent labeling selectively revealed subpopulation-specific trajectories: Npy+ SCPN axons remain largely confined to the medulla, with minimal presence in the spinal cord, whereas Cartpt+ SCPN extend axons beyond the brainstem into the cervical, but not thoracic cord (Extended Data Fig. 6e–g). To directly compare axon extension specificity, we performed dual anterograde labeling in Npy-FlpO;Cartpt-Cre double-transgenic mice. FlpO- and Cre-dependent reporter AAVs injected into lateral cortex at P1 enabled simultaneous visualization of axons of both SCPN subpopulations (Fig. 5d,e). At P35, Cartpt⁺ SCPN axons extend within the corticospinal tract through the brainstem, while Npy⁺ axons were confined to a more rostral level (Fig. 5f, white arrow). Optical clearing of the brainstem further highlights the axon extension specificity of each SCPN subpopulation: Npy⁺ SCPN axons terminated within the pons and medulla and do not extend into the spinal cord, while Cartpt⁺ SCPN axons continued caudally through the decussation toward the cervical cord (Fig. 5h and Supplementary Movie 1). Optical clearing of the cervical cord confirmed the absence of Npy⁺ SCPN axons and showed that Cartpt⁺ SCPN projections terminate within the cervical spinal cord (Fig. 5g). Our findings propose a new model for the developmental regulation of SCPN axon extension, in which not all developing SCPN extend to the spinal cord (Fig. 5i). Instead, distinct SCPN subpopulations consistently and durably restrict their projections to supraspinal levels throughout development and into maturity. These subpopulations are not determined stochastically but are molecularly specified from early developmental stages: Npy⁺ CBN are programmed to restrict their axon targeting to supraspinal levels, whereas Cartpt⁺ CSNc extend axons beyond the brainstem to the cervical spinal cord. This indicates that molecular mechanisms, rather than cortical location alone, delineate SCPN axon extension specificity. Even though both Npy⁺ CBN and Cartpt⁺ CSNc reside in lateral cortex, they are molecularly defined to establish two distinct cortical output pathways (cortico-brainstem and corticospinal). The Npy-Cre line enables selective molecular access to CBN throughout development and into maturity. Npy+ CBN are necessary for skilled forelimb motor control Corticospinal circuits have long been considered essential contributors to skilled forelimb movement in mice via connections to spinal interneurons and motor neurons31. However, cortico-brainstem pathways also shape descending motor commands13,32, and brainstem-spinal circuits can also contribute to dexterous control in mice10,11, suggesting that the brainstem might serve as a key hub for executing complex tasks4,33. We therefore hypothesized that Npy⁺ CBN, defined during development by their restricted connectivity to the brainstem, might play a critical role in skilled forelimb movements in adulthood, particularly in precise tasks such as reaching and grasping. To test this, we selectively silenced Npy+ CBN using an intersectional viral approach in Npy-Cre mice. We injected a Cre-dependent rAAV-DIO-FlpO bilaterally into the cerebral peduncles at P0, followed by Flp-dependent hm4Di-mCherry injection into the lateral cortex at P1 (Fig. 6a). This intersectional strategy restricted expression of the inhibitory DREADD (hm4Di) and mCherry to Npy+ CBN (Npy.CBNoff group). To rule out potential contributions from putative Npy⁺ CSN, a separate cohort received matched retro-AAV-FlpO injections into the cervical spinal cord instead of cerebral peduncles, thereby controlling for any spinally projecting Npy⁺ SCPN (Npy.CSNCtrl group). As a positive control, we again used Npy-Cre mice, which do not express Cre in primary motor cortex, and co-injected AAV-Cre with the Flp-dependent hM4Di-mCherry to silence SCPN medial M1 (M1.SCPNoff group). This group served to validate DREADD efficacy in suppressing known corticospinal contributors to skilled reaching. To ensure correct targeting of our viral labeling, we analyzed the distribution of hm4Di-mCherry+ SCPN in all three groups. 3D reconstruction at P90 confirmed accurate and distinct targeting in each group (Fig. 6c–e), with mCherry+ M1 SCPN restricted to cingulate and medial cortex and mCherry+ Npy+ CBN to lateral cortex. As expected, the Npy.CSNCtrl group showed no labeling of mCherry+ CSN (Fig. 6c–e). Selective silencing of Npy⁺ CBN disrupts movement execution and targeting accuracy in skilled reaching. a Intersectional viral strategy targets hM4Di to SCPN subpopulations (M1.SCPNoff, Npy.CBNoff, Npy.CSNCtrl); b Mice are tested in a skilled reaching task at baseline (BL), silencing (JHU (DREADD ligand)), and washout (WO). c–e Targeted labeling shows Npy⁺ CBN enriched in lateral cortex (one-way ANOVA; two-sided: Npy.CBNoff F(2,45) = 64.07, p < 1 × 10−4; Tukey: cingulate vs medial p = 0.0278, cingulate vs lateral p < 1 × 10−4, medial vs lateral p < 1 ×;10−4, n = 16 animals; M1.SCPNoff F(2,6) = 4.37, p = 0.0674, n = 3 animals; Npy.CSNCtrl F(2,12) = 3.29, p = 0.0727, n = 5 animals). f Npy⁺ CBN silencing reduces reaching success (one-way ANOVA; two-sided: Npy.CBNoff F(2,45) = 8.32, p = 0.0008; Tukey: BL vs JHU p = 0.0006, JHU vs WO p = 0.0340; M1.SCPNoff F(2,6) = 8.65, p = 0.0171; Tukey: BL vs JHU p = 0.0191, JHU vs WO p = 0.0408; Npy.CSNCtrl F(2,12) = 0.64, p = 0.5467). g Silencing increases approach errors (linear mixed-effects model; two-sided; FDR-adjusted: approach t(4) = −5.72, p = 0.0185; success t(4) = 3.53, p = 0.0322; retrieval t(3.04) = 4.72, p = 0.0322; grasp t(4) = −0.34, p = 0.7500; n = 5 animals). h Markerless paw tracking to capture reach structure and kinematics. i Reach structure is altered under silencing (canonical discriminant analysis; two-sided: Wilks’ λ = 0.712, F(10,295) = 11.9, P < 2.2 × 10−16). j, k Approach-error reaches form two reach structure types (clusters C1, C2), with C2 enriched under silencing. l, m C2 reaches are slower, less direct, and less accurate, reflecting impaired execution and targeting (linear mixed-effects model; two-sided; animal random effect; FDR-adjusted; longer duration, q = 1.3 × 10−19; lower peak acceleration, q = 0.0024; greater endpoint error, q = 0.011; longer trajectories, q = 7.9 × 10−17). Scale bars, c 500 μm. Cre genotypes: +/+ (homozygous), +/− (heterozygous). Data are mean ± SEM (e, g, m). Box plots show median (centre), 25–75% (box), and min-max (whiskers) (f). Source Data are provided as Source Data File. Mice were trained on the single pellet reaching task from P35 until they reached proficiency, then tested on 3 consecutive days: at baseline, after hM4Di ligand administration (silencing), and after washout (Fig. 6b). Silencing M1 SCPN (M1.SCPNoff) impaired reaching success (Fig. 6f, 44% ± 6.9 at baseline, 28% ± 4 after JHU administration). Similarly, Npy.CBNoff mice showed a marked reduction in successful reaches during JHU administration (36.5% ± 8.9 at baseline, 22.3% ± 12.6 after JHU administration), whereas Npy.CSNCtrl mice were unaffected (29.6% ± 4.6 at baseline, 25.6% ± 7.2 after JHU administration). These results identify Npy⁺ CBN as critical to support skilled forelimb movement. We next manually classified each reach as successful or as exhibiting an error in either the approach, grasp, or retrieval phase of the movement. We then quantified the incidence of each error type across baseline and silencing sessions to see which errors are increased when Npy+ CBN are silenced. Silencing Npy⁺ CBN led to more approach-phase errors (31.7% ± 10.98 at baseline, 56.4% ± 8.16 after JHU administration) but spared grasp and retrieval accuracy (Fig. 6g and Supplementary Movies 2–5), indicating a deficit in precision of forelimb targeting. To understand how Npy⁺ CBN silencing disrupts skilled reaching, we used markerless pose estimation34,35 to follow paw trajectories using front and side views during the single pellet reaching task (Fig. 6h). This analysis uncovered distinct movement signatures depending on outcome of the reach during Npy+ CBN silencing: successful reaches followed smooth trajectories to the pellet with stable positioning near the target at endpoint, while approach error reaches showed mistargeted, variable paths and less stable paw positioning around the pellet (Extended Data Fig. 7a, b). From these aligned trajectories, we extracted 47 kinematic features capturing speed, acceleration, path geometry, timing, and digit coordination (Supplementary Data 1). Principal component analysis (PCA) over all reaches using these kinematic features revealed a clear separation between baseline and silencing conditions (Fig. 6i), indicating that Npy⁺ CBN silencing alters the overall structure of reach movements. Since Npy⁺ CBN silencing produced a selective increase in approach errors, we next examined whether this change simply represented more frequent occurrences of the same errors that occur at baseline, or whether silencing resulted in the emergence of qualitatively distinct movement errors. To address this, we specifically focused our analysis on the approach-phase of all reaches (Fig. 6j) and embedded them into a low-dimensional space defined by 33 approach-specific kinematic features. Unsupervised clustering identified two distinct reach classes: one corresponding primarily to baseline performance (C1) and a second that emerged predominantly during silencing (C2) (Fig. 6j), with both clusters represented across all mice (Extended Data Fig. 7g). Reaches in cluster C2 overall were more spatially variable (Fig. 6k) and were characterized by kinematic features reflecting disrupted movement execution and targeting accuracy (Fig. 6l, m). Specifically, reaches of this error type exhibited longer approach durations despite higher peak acceleration, increased path length, and larger endpoint displacements, often overshooting the pellet beyond the range observed even for baseline reach errors. Overall, these results suggest that silencing Npy⁺ CBN does not merely increase the frequency of existing error types but instead gives rise to a distinct, disorganized reach mode defined by degraded movement execution and reduced spatial accuracy. We next asked whether this abnormal movement pattern identified in approach errors also appeared across other reach outcomes, which would indicate a broader reorganization of reach structure. Because the initial analysis relied on kinematic features specific to the approach phase, we next tested whether the same signature was present in reaches not resulting in an approach error as well. Indeed, we found that this distinctive signature seen in approach errors was also present in non-approach-error reaches (i.e., successful reaches, grasp errors, and retrieval errors), with effect sizes and directions closely matching those observed in approach-error reaches (Extended Data Fig. 7c), indicating a similar disruption in the approach phase in all reaches. To determine whether this disruption also extended into later phases of the reach, we then analyzed all reaches using the full set of kinematic parameters spanning approach, grasp, and retrieval phases. We again observed consistent alterations in both movement execution and targeting accuracy: reaches during Npy+ CBN silencing were slower and more segmented (Extended Data Fig. 7d–f), with longer overall duration and more frequent pauses during the outward reach (Extended Data Fig. 7e). They were also less efficient (Extended Data Fig. 7f), characterized by excessive paw lift height, paw overshooting the pellet, and grasps occurring farther from the pellet. These findings demonstrate that the altered reach mode induced by Npy⁺ CBN silencing is not confined to approach errors but represents a global disruption of skilled movement control. Together, these results demonstrate that Npy⁺ CBN neurons are essential for coordinating the smoothness, efficiency, and precision of reach-to-grasp movements and more broadly suggest that cortico-brainstem circuits established early in development remain indispensable for skilled forelimb control in adulthood. Axon collateralization by Npy+ CBN vs. CSN in the brainstem differs along rostro-caudal axis Given the role of Npy+ CBN in skilled forelimb movement, we next asked how their axonal projections are organized within the brainstem. CSN generate extensive brainstem collaterals en route to the spinal cord, providing input to premotor and modulatory nuclei involved in movement control6,13,36,37. We sought to determine whether Npy⁺ CBN, whose axons terminate within the brainstem, differ to CSN in how they distribute their collaterals along the rostro-caudal axis of the brainstem. Because Npy⁺ CBN reside in lateral cortical regions that only partially overlap with primary motor areas, we first compared the brainstem collateralization of SCPN originating from medial vs. lateral cortex. This analysis established the broader organizational framework for SCPN projections before focusing on molecularly defined lateral SCPN subpopulations (Npy⁺ CBN and Cartpt⁺ CSNc). We injected AAV-hSyn-GFP into medial cortex and AAV-CAG-tdTomato into lateral cortex of C57BL6 wild-type mice at P1 and analyzed brainstem collateralization at P35. 3D reconstruction of injection sites confirmed that labeling from medial and lateral cortex was spatially distinct, corresponding to primary motor and cingulate regions versus somatosensory and insular areas, respectively (Fig. 7b and Extended Data Fig. 8d). Using our machine learning-based segmentation and quantification pipeline StARQ38, we quantified axon collateralization across major brainstem nuclei spanning midbrain, pons and medulla (Extended Data Figs. 8c and 9). We hypothesized that brainstem collateral distribution by lateral SCPN would at least partially resemble the collateralization from primary motor cortex SCPN, given its role in skilled movement control. However, their collateralization patterns differed markedly along the rostro-caudal axis (Fig. 7c, d). Medial SCPN projections were concentrated in rostral brainstem nuclei such as the superior colliculus (SC, 24.38% ± 5.19 of the overall brainstem collateral) and periaqueductal gray (PAG, 10.46% ± 4.81), whereas lateral SCPN showed a more distributed collateralization pattern extending into caudal regions, with prominent projections to midbrain reticular nucleus (MRN, 12.43% ± 4.46), PAG (6.85% ± 3.10), and SC (8.07% ± 4.76), and in the medulla to intermediate reticular nuclei (IRN, 6.41% ± 2.52), the spinal trigeminal nucleus (SPV, 8.01% ± 2.86), and the parvicellular reticular nucleus (PARN, 9.99% ± 2.56) (Fig. 7c, d, g, for abbreviations of brainstem structures see Supplementary Data 2). Even in brainstem nuclei receiving input from both medial and lateral SCPN, such as the SC, axon collaterals rarely overlap. Instead, their collateralization follows a somatotopic organization, with medial and lateral SCPN targeting distinct, spatially segregated regions within the nucleus (Fig. 7c, upper panel). Thus, even though both lateral and medial cortex SCPN function in controlling skilled forelimb movement, they exhibit differences in their brainstem innervation/collateralization (Extended Data Fig. 9). Distinct cortico-brainstem collateralization patterns of SCPN subpopulations reveal specialized connectivity. a SCPN subpopulations were labeled by anterograde AAV injections into medial and lateral cortex in the same animals (wild-type; n = 3 animals) or by Cre-dependent AAVs injected into lateral cortex of Npy-Cre and Cartpt-Cre mice (n = 3 animals/Cre line). Brainstem sections were analyzed using StARQ38, a machine-learning pipeline for automated segmentation and axon quantification. b Injection sites are non-overlapping for medial and lateral SCPN, whereas Npy⁺ and Cartpt⁺ injections overlap within lateral cortex. c Coronal brainstem sections show distinct rostro-caudal collateralization patterns of medial and lateral SCPN; within lateral SCPN, Npy⁺ CBN, and Cartpt⁺ CSN display divergent projections. d Collateralization patterns differ across SCPN subpopulations. Medial vs lateral SCPN show distinct projections (one-way ANOVA; two-sided; FDR-corrected post hoc (Benjamini–Hochberg): MRN q = 4.4 × 10−4, SC q = 1.1 × 10−30, PARN q = 2.4 × 10−8, PAG q = 0.030). Within lateral SCPN, Npy⁺ CBN show higher projection density to MRN (q = 5.9 × 10−41), PAG (q = 2.3 × 10−8), SC (q = 7.6 × 10−11), SN (q = 0.0037), and PG (q = 4.5 × 10−4), whereas Cartpt⁺ CSN preferentially target IRN (q = 0.0030), MDRNd (q = 0.012), PARN (q = 0.016), and SPV (q = 4.3 × 10−38). See Extended Data Fig. 8 for all brainstem regions. e, f Representative regions show denser Npy⁺ CBN innervation in MRN and stronger Cartpt⁺ CSN innervation in SPV and MDRNd. g Summary of SCPN connectivity across brainstem subdivisions with quantitative representation of innervation, highlighting distinct projection profiles of medial vs lateral SCPN and Npy⁺ CBN vs Cartpt⁺ CSN; circle size indicates relative innervation per nucleus. For abbreviations, see Supplementary Data 2. Scale bars, c 500 μm, e, f 100 μm. Cre genotypes: +/+ (homozygous), +/− (heterozygous). Data are mean ± SEM (d). Source data are provided as a Source Data file. Having established that cortical origin determines broad organizational differences in SCPN brainstem collateralization, we next asked whether molecularly distinct SCPN subpopulations (Npy+ CBN and Cartpt+ CSNc) within lateral cortex also differ in their projection patterns (Fig. 7a, lower panel). Cre-dependent AAVs were injected into the lateral cortex of Npy-Cre or Cartpt-Cre mice at P1, and brainstem projections were analyzed at P35. As an additional control, we co-injected a Cre-independent reporter virus in both Npy-Cre and Cartpt-Cre mice. 3D reconstruction of injection sites showed that injections across all mice comparably targeted lateral somatosensory and insular cortex (Extended Data Fig. 8b). Further, this control labeling produced similar brainstem innervation patterns across Cre lines, comparable to lateral SCPN innervation (Extended Data Fig. 8c). Cre-dependent labeling volumes for medial and lateral injections were similarly well matched across animals (Fig. 7b and Extended Data Fig. 8e). Using StARQ to quantify corticofugal collateralization across the brainstem (Fig. 7c–g), we found that Npy⁺ CBN projections were selectively enriched in rostral motor-associated regions, including the MRN (20.10% ± 5.33), SC (19.35% ± 5.15), and PAG (10.54% ± 2.03). These nuclei were also innervated by Cartpt⁺ CSNc, albeit to a lesser degree (MRN: 8.19% ± 0.62; SC: 14.25% ± 1.43; PAG: 6.16% ± 0.05). In contrast, Cartpt⁺ CSNc collaterals extended more caudally into medullary structures that received minimal Npy⁺ CBN input, including the SPV (14.79% ± 1.40), IRN (5.15% ± 0.43), and PARN (4.23% ± 1.01). Thus, while both SCPN subpopulations arise from lateral cortex, their axonal trajectories diverge along the rostro-caudal axis: Npy⁺ CBN preferentially engage rostral motor-associated brainstem nuclei, whereas Cartpt⁺ CSNc maintain broader connectivity with both sensory- and motor-related nuclei (Fig. 7g and Extended Data Fig. 8c). Whether these anatomical biases correspond to functional specialization remains to be determined, but this organization suggests that Npy⁺ CBN neurons may exert a specialized role in modulating skilled movement via rostral premotor hubs. Overall, this study reveals molecular programs that delineate axon extension differences by SCPN subpopulations that go beyond areal controls. Further, these developmentally defined subpopulations of SCPN contribute to adult motor circuitry, specifically for skilled movements. Our findings highlight how early molecular specification drives SCPN diversification to support movement regulation and underscore the critical role of cortico-brainstem connectivity in skilled movement behaviors. Discussion Skilled movement relies on precise connectivity between SCPN and their targets in the brainstem and spinal cord. This connectivity is established by both direct cortico-bulbar projections formed between the neocortex and the brainstem, as well as corticospinal projections that extend to the spinal cord while also collateralizing in the brainstem. Direct cortico-bulbar connectivity is known to arise by pruning of corticospinal projections16. Our results reveal an additional, distinct developmental mechanism that establishes de novo cortico-brainstem connectivity. We find that there is early molecular delineation between CBN and CSN during axon extension in development, and that these molecularly defined CBN do not extend axons into the spinal cord. We identify Npy and Cartpt as differentially expressed genes that distinguish CBN from CSN within the same cortical area (lateral sensorimotor cortex), defining a molecular delineation between these closely related SCPN subpopulations. We find that although expression of these genes declines during development, the distinction they define is durable: Npy+ CBN axons terminate within the brainstem, whereas CSN axons extend into the spinal cord, and this organization persists into adulthood. At maturity, Npy+ CBN and Cartpt+ CSN axons exhibit distinct topographical collateralization within the brainstem, with Npy+ CBN axons preferentially innervating more rostral brainstem nuclei, while Cartpt+ CSN axons preferentially innervate caudal regions. Our results further show that Npy+ CBN, while restricting their axons to the brainstem, are key components of the circuitry governing skilled forelimb control, particularly those mediating precise movement execution and targeting. Together, these findings redefine how distinct cortical output pathways emerge during development, showing that molecularly specified developmental programs that direct initial axon extension, rather than postnatal refinement alone, establish the circuit architecture in adulthood that is essential for skilled movement. Refinement of SCPN connectivity through axonal pruning from spinal to brainstem levels is well established15–17,39,40, and several molecular regulators of this process have been identified18,19. Our data reveal an additional layer of developmental specification: a subset of CBN restrict their axons to brainstem levels from the earliest stages of axon extension, forming a de novo cortico-brainstem circuit. Our findings suggest that an early molecular program regulates whether SCPN axons terminate within the brainstem or extend into the spinal cord, although the precise mechanisms that underlie this process remain to be elucidated. Similar axon extension control has been observed at spinal levels20,21, suggesting that comparable mechanisms might operate across the neuraxis. This further raises the question of how many distinct sites exist along the neuraxis where axonal targeting specificity establishes early diversification of SCPN connectivity, and what mechanisms govern this process during corticofugal circuit formation. The diversification of subcerebral connectivity has classically been attributed to be driven by mechanisms that direct cortical areal specification41. These molecular programs then define SCPN across cortical regions, where they control area-specific connectivity, e.g., SCPN in motor versus visual cortex express unique genes that shape their subcerebral projections in an area-dependent manner19,41–44. These programs can even be re-engaged experimentally to redirect motor-cortical SCPN to more proximal targets, underscoring the molecular plasticity of areal control19. Areal identity operates in parallel with developmental timing, which influences when neurons acquire subtype-specific molecular features and thereby determines projection range and target selection45. Our findings introduce axon extension specificity as an additional developmental axis that not only complements but also intersects with these established mechanisms of diversification. Within the same region of lateral sensorimotor cortex, CBN and CSN diverge through distinct molecular programs. Given that these two neuronal subpopulations reside within the same cortical territory, their divergent projection patterns indicate that cortical areal diversification alone cannot account for this level of circuit specialization. Therefore, SCPN residing in the same cortical territory can acquire distinct long-range projections through cell-intrinsic molecular specification. Developmental timing also potentially contributes to this organization. We find that Npy⁺ CBN are more differentiated during early postnatal stages compared to CSN. However, Npy⁺ CBN are also positioned more superficially within layer V, suggesting that they may be generated slightly later than CSN, while undergoing a more accelerated differentiation program. There is precedence for such anatomical projection divergence being governed by developmental timing in other neocortical46, as well as spinal projection neurons47. To uncover the molecular basis of axon extension specificity between CBN and CSN, we combined developmental anatomical labeling with single-cell transcriptomic profiling of purified SCPN subpopulations (i.e., all SCPN (CBN + CSN) vs. CSN only). This nested comparative profiling allowed us to distinguish CBN from CSN at early stages of axon extension specificity. We identified Npy and Cartpt as differentially expressed genes that delineate these subpopulations and distinguish CBN from CSN at the molecular level. Although Npy and Cartpt have known functions in the adult nervous system, how their developmental expression contributes to the specification or maturation of these two SCPN subpopulations remains unclear. However, additional differentially expressed genes, including Sema3e48,49 and Alcam50,51, have known roles in axon guidance and may contribute toward regulating differential axon extension or might function in subsequent steps in the establishment and/or refinement of SCPN circuitry. Further, Lrrtm3, a leucine-rich repeat protein structurally related to Lumican, a proteoglycan that we previously identified as controlling corticospinal axonal collateralization52, might function at later stages to refine subcerebral connectivity53,54. Our newly established molecular tools to distinguish CBN from CSN now provide the necessary foundation to dissect the potential function of these candidate genes and the molecular control over SCPN axon extension at the brainstem-spinal transition zone can now be investigated with precision. The transient molecular programs that distinguish CBN and CSN during development represent the early phase of a dynamic differentiation process that ultimately gives rise to their distinct circuit organization. While Npy and Cartpt are differentially expressed by these distinct subpopulations, they likely do not themselves govern the divergence in axon targeting. However, this early molecular divergence clearly aligns with the later emergence of two anatomically discrete descending cortical output pathways that persist into adulthood. Both arise from the same area of sensorimotor cortex and yet exhibit differential collateralization within brainstem, likely engaging distinct subcerebral networks. This suggests that cortical output at maturity is structured by axon extension specificity, which lays the groundwork for which specific brainstem circuits each pathway recruits. We establish that medial SCPN preferentially innervate rostral brainstem targets such as the SC and PAG, whereas lateral SCPN collateralize more caudally within reticular and sensory regions. Within lateral sensorimotor cortex itself, where Npy⁺ CBN and Cartpt⁺ CSN are intermingled, this finer molecular distinction further subdivides the pathway: CBN terminate predominantly in rostral motor-associated nuclei, while CSN project toward caudal sensory and premotor domains. Together, this hierarchical organization establishes parallel cortico-brainstem subnetworks in which anatomical position and molecular identity jointly determine adult connectivity, potentially positioning these pathways to influence different, possibly overlapping, components of skilled movement. This organization of creating and maintaining two parallel descending routes from the cortex raises an important question about the functional significance of this anatomical diversification. Skilled forelimb movement relies on precisely organized and adaptable circuits that can generate complex limb movements while continually adjusting them as the action unfolds. Rodents and primates execute these modular reach-to-grasp phases using similar strategies1,55. This behavioral complexity is reflected in the organization of descending cortical pathways. Motor and somatosensory cortical regions contribute to different components of forelimb control, and SCPN in these areas form parallel, functionally specialized routes to distinct brainstem and spinal targets that then also subserve distinct elements of this motor control13,56–62. Within this broader organizational framework, our anatomical data show that medial SCPN in rostral motor cortex project predominantly to rostral motor-associated nuclei, including the SC and MRN, consistent with previous anatomical tracing results showing limited axonal collateralization by M1 SCPN5,8,9. Lateral SCPN, in contrast, exhibit more extensive collateralization in the brainstem along the rostro-caudal axis, aligning with evidence that lateral sensorimotor cortex, including S1 CSN, shapes skilled forelimb behavior through its influence on spinal circuits31,63. What has remained unclear is whether the cortico-brainstem component of this lateral pathway contributes directly to skilled movement, as its function has been difficult to isolate from that of intermingled CSN. Using our molecular access to Npy⁺ CBN, we were able to specifically dissociate their functional contributions from CSN and establish that they are also necessary for skilled forelimb behavior. Silencing Npy⁺ CBN did not disrupt the overall sequence of the reach, i.e., approach, grasp, and retrieval remained intact. However, it selectively impaired reach kinematics, producing slower, segmented trajectories and inaccurate targeting. Future experiments similarly targeting and silencing Cartpt+ CSNc will enable a direct comparison of their respective contributions, an experiment that would provide highly informative insight into the possible division of labor between these two pathways. As with the corticospinal system, which affects both motor output and sensory modulation31,64, Npy⁺ CBN connectivity spans brainstem nuclei implicated in both functions. Npy+ CBN reside in cortical areas with known projections to brainstem regions associated with tactile feedback modulation14, suggesting that they may influence sensory processing during movement. In addition, these neurons also target premotor brainstem centers involved in shaping skilled forelimb actions13, providing a complementary pathway through which they may influence forelimb movement. The precise balance of these contributions remains unresolved, but the broad innervation by CBN axons and the selective kinematic deficits following their silencing point to a role for CBN in the online control of skilled movement. Npy⁺ CBN likely relay updated motor commands to motor-associated brainstem centers as the reach unfolds, thereby supporting the continuous adjustments required for accurate movement execution. The brainstem centers through which Npy+ CBN could exert such an influence are likely to be multifaceted. Several brainstem structures targeted by Npy⁺ CBN have known roles in guiding skilled limb movements across rodents and primates. The SC, where Npy⁺ CBN axons collateralize, is required for accurate forelimb reaching65 and contributes to proximal limb control66,67. This raises the possibility that the SC serves as a key relay through which CBN influence skilled forelimb behavior. Consistent with this idea, we find topographic differences in SC innervation not only between medial and lateral SCPN but also between CBN and CSN residing in lateral cortex, suggesting that each subpopulation may engage distinct SC domains to support specialized aspects of movement control. While our machine learning pipeline quantifies axonal projections across predefined brainstem nuclei, it is not designed to detect subnuclear or intranuclear spatial differences. Thus, finer-scale distinctions between Npy⁺ and Cartpt⁺ axonal collateralization, such as within the SC, are also likely to be biologically meaningful. While these extend beyond the resolution of the present analysis, they also will warrant future investigation. Npy⁺ CBN also collateralize within the pontomedullary reticular formation, where neurons exhibit both movement-related and preparatory activity during reaching68, and stimulation of this region can evoke activity in distal wrist muscles previously attributed solely to corticospinal pathways69. CBN also innervate the midbrain reticular formation, which has been implicated in postural regulation and skilled movement coordination70. Taken together, these converging lines of evidence indicate that multiple brainstem targets of Npy+ CBN provide substrates capable of supporting the real-time adjustments that are disrupted when these neurons are silenced. However, it must be emphasized that the scope of CBN influence may extend beyond specific individual targets, since cortical output is known to broadcast to multiple subcortical centers, including the brainstem, to coordinate ongoing motor commands6,56,71. Silencing of deep cerebellar nuclei elicits deficits in skilled forelimb movement that resemble the kinematic impairments observed with CBN silencing, raising the possibility that CBN access components of these loops to support real-time movement refinement72,73. More broadly, our findings, together with prior work, support an emerging model in which distinct SCPN subpopulations both within and across cortical regions differentially access motor- and sensory-related elements of these conserved circuits, creating multiple descending pathways through which skilled behavior is shaped. The developmentally informed molecular tools that we have established now provide molecular access for dissecting such distinct functional role(s). It is also worth noting that several of the nuclei targeted by Npy+ CBN have previously been implicated in forelimb motor control across species. In both rodents and primates, brainstem-spinal pathways are essential for dexterous reaching, and disruption of brainstem neurons that give rise to these connections impairs reach precision and grip coordination10,74–76. Brainstem nuclei such as the reticular formation and SC contain neurons whose activity correlates with arm movement and reach onset are equally important in both species68,69,77. This cross-species consistency in functional roles of CBN-innervated nuclei suggests that the broader architecture by which the Npy+ CBN circuit influences skilled forelimb movement in mice may be evolutionarily conserved beyond rodents. These findings also emphasize that cortical control over skilled forelimb movement extends beyond primary motor cortex, and that lateral cortical regions and their distinct contributions are equally indispensable. Within lateral cortex, a cortico-brainstem pathway that is molecularly and anatomically distinct from the corticospinal system influences skilled movement. Defining how CBN contribute to both sensory modulation and motor refinement will be important for understanding how skilled movement is supported. The tools developed in this study will facilitate the dissection of specific contributions from both cortico-brainstem and corticospinal pathways to skilled forelimb movements. Together, our results show that early developmental programs establish separation between cortico-brainstem and corticospinal projections, giving rise to two complementary descending routes from cortex. Importantly, transient molecular and projection differences between SCPN subtypes lay the groundwork for the mature circuit specializations that ultimately support these distinct pathways. This organization supports a view of skilled behavior as emerging from coordinated activity across multiple specialized pathways rather than a single descending hierarchy, with developmental molecular programs preconfiguring this diversification. Method Data reporting No statistical methods were used to predetermine sample size. The experiments were not randomized. Investigators were blinded to group allocation during behavioral experiments and outcome assessments wherever possible. Sample sizes refer to biologically independent animals, except where noted. Animals All mouse studies were performed in accordance with institutional and federal guidelines and were approved by the Weill Cornell Medical College Institutional Animal Care and Use Committee. Wild-type CD1 and C57BL/6J mice were obtained from Charles River Laboratories (Wilmington, MA), and transgenic lines were obtained from Jackson Laboratories: Npy-Cre (B6.Cg-Npytm1(cre)Zman/J, Stock No.: 02785178), Cartpt-Cre (B6;129S-Cartpttm1.1(cre)Hze/J, Stock No.: 028533), and Npy-FlpO (B6.Cg-Npytm1.1(flpo)Hze/J, Stock No.: 03021179). All transgenic lines were maintained on a C57BL/6J background and bred to heterozygosity with wild-type C57BL/6J mice, except for Cartpt-Cre, which was maintained as a homozygous line. The Npy-FlpO;Cartpt-Cre line was generated by crossing Npy-FlpO and Cartpt-Cre mice. Genotyping was performed using standard primer sets provided by Jackson Laboratories. The day of birth was designated as postnatal day 0 (P0), and experiments were conducted across developmental stages (P0–P14) and in adult mice ( ≥ P35), as specified for each experiment. For samples collected prior to P14, sex could not be reliably determined and data from both sexes were combined. For behavioral and anatomical studies, both male and female mice were used; sex was not included as a factor in the analysis. Animals were housed in groups of four to five under a 12 h light/dark cycle with food and water ad libitum, at ~22 °C and 40–60% relative humidity. The specific age and number of animals used for each experiment are provided in the corresponding figure legends. Viruses used in this study All viral vectors were obtained as pre-packaged AAV preparations from Addgene unless otherwise noted. We gratefully acknowledge the generous plasmid contributions of the original depositors to Addgene, who are listed alongside each construct in Supplementary Data 3. Addgene supplied the corresponding viral preparations in the indicated serotypes and titers. The AAV-fDIO-hM4Di-mCherry was obtained as a plasmid from Addgene and packaged into a viral preparation by VectorBuilder. Retrograde labeling of SCPN subpopulations Retrograde labeling from the cerebral peduncles or cervical spinal cord was performed using a pulled glass micropipette attached to a nanojector (Nanoject II/III, Drummond Scientific, Broomall, PA) at a rate of 23 nl/s as previously described20,21,80. Cholera Toxin B subunit (CTB; Thermo Scientific) was injected bilaterally at 3 injection sites at the cerebral peduncles (161 nl injected at each site in 23 nl increments; total of 483 nl) or into the spinal segment C1/C2 (both sides of the midline, total of 207 nl), respectively, under ultrasound-guided backscatter microscopy (Vevo 2100; VisualSonics, Toronto, Canada). Injections were performed at the time point of initial axon extension (P0 at cerebral peduncles, P2 at cervical spinal cord23) unless stated otherwise. For AAV-mediated retrograde labeling of Npy+ or Cartpt+ SCPN in their respective Cre reporter mouse lines, rAAV-DIO-GFP and rAAV-tdTomato were co-injected into the cerebral peduncles at P0 in Npy-Cre or Cartpt-Cre mice. For AAV-mediated retrograde labeling of Npy+ or Cartpt+ CSN, rAAV-DIO-GFP and rAAV-tdTomato were co-injected into the cervical spinal cord level C2 at P3 in Npy-Cre or Cartpt-Cre mice, respectively. For dual retrograde labeling of Npy+ and Cartpt+ SCPN, rAAV-DIO-GFP and rAAV-fDIO-mCherry were co-injected into the cerebral peduncles at P0 in Npy-Flpo;Cartpt-Cre double transgenic mice. Anterograde labeling of SCPN subpopulations For AAV-mediated anterograde labeling of Npy+ or Cartpt+ SCPN in Cre mice, Cre-dependent reporter AAV (AAV-DIO-GFP or AAV-DIO-tdTomato) was co-injected with a control reporter AAV (AAV-tdTomato or AAV-GFP) at P0 into rostro-lateral cortex (345 nl). AAV was injected unilaterally under ultrasound-guided backscatter microscopy via a pulled glass micropipette attached to a nanojector at a rate of 23 nl/s. For dual AAV-mediated anterograde labeling of Npy+ and Cartpt+ SCPN in Npy-Flpo;Cartpt-Cre mice, AAV-DIO-GFP and AAV-fDIO-tdTomato were co-injected at P1 into rostro-lateral cortex (345 nl). Fluorescent activated cell sorting (FACS) of whole cells Brains of retrogradely labeled mice were collected at P1 (SCPN group) or P3 (SCPN or CSN group). Tissue was kept in cold buffers and on ice at all times. Brains were rapidly dissected and sliced into 800 µm coronal sections using a TC-1 Tissue Chopper (Electron Microscopy Sciences). Sections containing labeled cells were selected under a fluorescent dissection scope (Nikon SMZ18), and the entire mediolateral extent of sensorimotor cortex was collected and pooled across the litter (8–10 brains/sample). A single cell suspension of the samples was obtained using enzymatic (15 min, 37 °C) and mechanical dissociation (gentle trituration using rounded glass Pasteur pipettes of decreasing diameter (600 μm, 300 μm)) followed by filtration through a 40 µm mesh (Biologix, 15-1040-1). SCPN were FACS-purified from the single cell suspension using a WOLF flow sorter (NanoCellect Biomedical) with standard settings and an adapted threshold for cell size (30000). To enhance sorting efficiency and reduce contamination, cells were first sorted at a high concentration ( ~ 6 × 106 cells/ml), followed by a secondary sort of the initially sorted cells at a lower concentration ( ~ 2 × 105 cells/ml). This two-step process enabled efficient enrichment of CTB 555-labeled cells. We collected ~15000 cells in 5 ml per sort. Cells were enriched for downstream processing by centrifugation at 80 × g for 10 min and immediately prepared for single-cell RNA sequencing. Single-cell RNA sequencing Single-cell RNA-seq samples include both biological and technical replicates, with technical replicates derived from pooled animals (8–10 animals per sample). Each sample (P1.SCPN, P3.SCPN, and P3.CSN) was sequenced using the Chromium 10× Single Cell 3’ pipeline, following the standard protocol. SCPN samples were generated from two independent experiments (biological replicates), each with two technical replicates. CSN samples consisted of two technical replicates. Briefly, single-cell suspensions of 5000–10,000 cells were loaded onto a Chromium Chip B/G and processed using the standard protocol for the Chromium Next GEM Single Cell 3ʹ GEM and Gel Bead Kits v3/v3.1. Libraries were prepared with the Chromium Next GEM Single Cell 3ʹ Library Kit v3.1 and Chromium i7 Multiplex barcodes. The sequencing libraries (n = 11; P3.CSN (2), P3.SCPN (5), P1.SCPN (4)) were assessed for quality on the Agilent TapeStation (Agilent Technologies, Palo Alto, CA, USA), quantified with a Qubit 2.0 Fluorometer (Invitrogen, Carlsbad, CA), and pooled libraries were quantified by qPCR (Applied Biosystems, Carlsbad, CA, USA). The pooled libraries were clustered on 5 lanes of a flow cell and loaded onto an Illumina HiSeq instrument (4000 or equivalent) according to the manufacturer’s instructions, sequenced in a 2 × 150 bp configuration with 8 bp single indexing as recommended by 10× Genomics. Perfusion, fixation, and tissue processing For labeling studies, mice were terminally anesthetized using ice (mice younger than P7) or a mix of ketamine (150 mg/kg; Covetrus, NDC 11695-0703-1) and xylazine (15 mg/kg%, Akorn, Inc., NDC 59399-111-50) (mice older than P7). Animals were transcardially perfused with 10 ml PBS followed by 10 ml cold 4% paraformaldehyde (PFA). Brains were dissected, post-fixed in the same solution overnight, and stored at 4 °C in PBS until further use. After cryopreservation in 30% sucrose overnight, tissues were directly embedded in Tissue-Tek OCT Compound and cut into coronal sections (50 µm, 6 series collected) on a cryostat (Leica CM1860). Free-floating sections were collected and stored at 4 °C in PBS until further processing. For in situ hybridization, brains were sectioned at 30 µm and collected on-slide across 12–14 slides. Slides were stored at −20 °C and processed further the following day. Immunolabeling of SCPN subpopulations and axons To visualize cell bodies of retrogradely labeled Cartpt+ and Npy+ SCPN in the cortex, free-floating 50 µm cortex sections were blocked in a BSA blocking solution (0.3% BSA in PBS with 0.3% Triton-X) for 30 min at room temperature. Sections were incubated in rabbit anti-GFP, 1:500 (Invitrogen, A-11122) overnight at 4 °C. Secondary antibody goat anti-rabbit Alexa Fluor 647 (Invitrogen, A-21245) was used at a dilution of 1:250 and sections were incubated for 3 h at room temperature. Following secondary antibody incubation, sections were stained with DAPI, mounted on slide, and cover slipped with Fluoromount-G (Invitrogen, 00-4958-02). Spinal sections (50 µm, axial) were processed using the same staining method to visualize axon tracts. Following conditional silencing of SCPN subpopulations using inhibitory DREADDs, primary antibody Living Colors® DsRed Polyclonal Antibody (Takara Bio USA, 632496) was used followed by secondary antibody goat ant-rabbit Alexa Fluor 546 (Invitrogen, A-11035, 1:250) to label Npy+ CBN and putative Npy+ CSN, using staining protocol described above. To visualize brainstem collaterals following anterograde labeling of Npy+ CBN or Cartpt+ CSN in Cre mice, free-floating 50 µm brainstem sections were blocked with Image-iT FX Signal Enhancer (Invitrogen, I36933) as per manufacturer’s instructions. Standard staining protocol was subsequently followed. Primary antibody rabbit anti-GFP, 1:500 (Invitrogen, A-11122), was used to label Npy+ axons and rabbit anti-RFP, 1:500 (Rockland Immunochemicals, 600-401-379) was used to label Cartpt+ axons. NeuroTrace™ 500/525 Green (Invitrogen, N21480) or NeuroTrace™ 640/660 Deep-Red (Invitrogen, N21483) Fluorescent Nissl Stain was used at a dilution of 1:200 to label neurons. Sections were then mounted on slide and cover slipped with Fluoromount-G (Invitrogen, 00-4958-02). In situ hybridization Coronal sections were processed with the RNAScope Multiplex Fluorescent v2 kit (Advanced Cell Diagnostics, 323100). The probes used were: Mm-Npy (313321-C3), Mm-Cartpt (432001-C2), Mm-Alcam (462061-C3), Mm-Lrrtm3 (461501-C3), Mm-Klhl14 (510591), Mm-Sema3e (449631-C2) (Advanced Cell Diagnostics). The CTB signal in retrogradely labeled cells was amplified by immunofluorescence using a rabbit anti-CTB primary antibody (Abcam, ab34992, 1:200). The RNA-protein co-detection workflow was performed with the RNA-Protein Co-detection Ancillary Kit (Advanced Cell Diagnostics, 323180) according to the manufacturer’s instructions. Microscopy and image preprocessing Imaging was performed using multiple microscopy setups depending on the sample type. The P4 coronal brain sections shown in Fig. 1 (50 µm) were imaged in 4 µm z-stacks at 20× magnification on a Zeiss Axio Imager M2 microscope using Stereo Investigator software (MBF Biosciences), and single-plane images were extracted from the z-stacks using the Deep Focus tool in Neurolucida. Otherwise, coronal brain sections, spanning the cortex and the brainstem, were imaged at 10× magnification on the same microscope, with images processed as single TIF files in Neurolucida software (MBF Biosciences). RNAScope samples were imaged at 20× magnification on a Nikon A1R confocal microscope, with 5 z-stacks taken at 1 µm intervals around the focal point. Whole-mount overviews of the brain (dorsal and ventral view) were captured using a Nikon SMZ18 fluorescent dissection microscope. Tissue preservation, clearing, immunolabeling, and imaging A whole brain sample with dual retrogradely labeled Npy+ and Cartpt+ SCPN (Npy-FlpO;Cartpt-Cre) was cleared, stained, and imaged on a light sheet microscope by LifeCanvas Technologies (Cambridge, MA). The PFA-fixed brain was collected using the same protocols as described above. The brain was preserved using SHIELD reagents following the manufacturer’s instructions81 and delipidated with Clear+ delipidation reagents (LifeCanvas Technologies). After delipidation, the sample was immunolabeled with goat anti-GFP (EncorBio, GPCA, 6 µg), rabbit anti-RFP (Rockland Immunochemicals, 600-401-379, 6 µg), and mouse anti-NeuN (EncorBio, MCA-1B7, 10 µg), followed by donkey anti-goat Alexa Fluor 488 (Jackson ImmunoResearch, 705-545-147, 12 µg), donkey anti-rabbit Alexa Fluor 555+ (Thermo Fisher Scientific, A32794, 12 µg), and donkey anti-mouse Alexa Fluor 647 (Jackson ImmunoResearch, 715-605-151, 20 µg). Immunolabeling was performed using eFLASH technology82, which integrates stochastic electrotransport83 and SWITCH84 on a SmartBatch+ device (LifeCanvas Technologies). For refractive index matching, the sample was incubated in 50% EasyIndex (RI = 1.52) overnight at 37 °C, followed by 100% EasyIndex for one day. Imaging was performed using a SmartSPIM axially swept light sheet microscope with a 3.6× objective (0.2 NA). The resulting TIFF image stacks were converted to Imaris files (.ims) using the Imaris File Converter, and 3D reconstructions were created using Imaris software (Oxford Instruments). For visualizing the brainstem and spinal cord in anterogradely labeled Npy+ and Cartpt+ SCPN (Npy-FlpO;Cartpt-Cre mice), a similar clearing procedure was performed in house using the SmartBatch+ active electrophoretic system (LifeCanvas Technologies). The whole brain (including spinal cord) was initially treated in SHIELD OFF for 3 days at 4 °C, followed by SHIELD ON at 37 °C for 24 h. After an initial incubation in Delipidation Buffer for 24 h, the sample underwent active delipidation for 30 h following standard protocols. Index matching was achieved with 24 h of 50% EasyIndex at 37 °C, followed by 100% EasyIndex. The spinal cord was separately embedded in 1.5% 88 °C melting point agarose (R2801, Thermoscientific) and subjected to repeated index matching within the agarose block. No additional antibody staining was applied for endogenous fluorescent proteins. Imaging was conducted on a Zeiss Lightsheet 7 (Carl Zeiss AG, Germany) using Zen Black (version 3.1), with samples immersed in low-fluorescence refractive index matching immersion oil (RI = 1.52, LDF, Cargille). Brainstem and spinal cord samples were imaged in dual channels (488, 561) at 5× magnification (0.94 × 0.94 × 3.5 µm), while whole brains were imaged at 2.5× magnification (1.83 × 1.83 × 5.23 µm). The images, captured as 16-bit CZI files, were stitched using Zen Blue (version 3.4.91). 3D reconstructions and video renderings were created in syGlass (version 1.8.2), utilizing the cut tool to remove excess imaging volume. Quantification of retrogradely labeled SCPN For P4 brains, to ensure consistent comparison across animals, we quantified retrogradely labeled SCPN/CSN using a fixed set of anatomical levels: coronal sections were selected at four selected rostro-caudal levels (1.95, 2.67, 3.27, 4.23 mm, approximated from the P6 Atlas of the Developing Mouse Brain85) for manual counting of retrogradely labeled SCPN/CSN using Neurolucida software (MBF Biosciences). To estimate cell distribution along the mediolateral axis, each cortical hemisphere was divided into 5 bins of equal width. We then quantified number of labeled neurons in the cingulate cortex (most medial bin), medial cortex (next 2 bins), and lateral cortex (outer/lateral most 2 bins). For in situ hybridization analysis at early developmental timepoints (P1–P14), CTB+ SCPN/CSN co-labeled with RNAScope signal were manually selected using FIJI/ImageJ (version 1.54f86). For Npy and Cartpt experiments, we analyzed the same four anatomically matched sections used for retrograde quantification (1.95, 2.67, 3.27, 4.23 mm), allowing direct comparison across assays. For all other RNAscope targets, we quantified the section at 2.67 mm, which consistently contained the highest density of retrogradely labeled neurons and provided the most robust signal for comparison across animals. In every case, each double-labeled neuron was assigned to one of the predefined mediolateral cortical bins using its soma coordinates, ensuring that spatial distributions were directly comparable to the tracing-based analyses. For P14 and P35 brains, to investigate spatial distribution across the entire brain, AAV-labeled SCPN/CBN/CSN were manually selected in FIJI/ImageJ across the rostro-caudal extent of the cortex (8–9 matched sections spaced 300 μm apart, ranging from +2.00 to −2.00 mm relative to Bregma87). Cell coordinates were transformed into the Common Coordinate Framework version 3 (CCFv3, 10 μm resolution)30 framework through affine transformation using CELL3D, which allows quantification of cell numbers per specific cortical region (e.g., somatosensory cortex, primary motor cortex etc.). For mediolateral analyses, cells were classified into cingulate, medial, or lateral bins using the same five-bin division described for P4. For visualization, all neurons were plotted as individual points within a 3D CCFv3 brain volume and color-coded by experimental group or animal. The mean rostro-caudal (X) and dorsolateral (Z) coordinates were calculated for each group, along with standard error, to represent cell distribution across these two axes. 3D reconstruction of SCPN injection volumes To quantify injection volumes, cortical areas with AAV-labeled signal were manually outlined in FIJI/ImageJ, and the coordinates were transferred into CCFv3 reference atlas using a similar custom MATLAB script described above (VOL3D). CCFv3-adapted cell coordinates were used to compute surface boundaries via Delaunay triangulation, followed by Laplacian smoothing to reconstruct and display the 3D injection volumes within the CCFv3 framework. The volumes were converted into a 100 µm voxel grid, retaining only enclosed voxels within the surface boundary, and injection volume was estimated by summing the volume of these enclosed voxels. To assess overlap with other volumes or selected ABA structures, enclosed voxels were compared to those in the target structure, and non-overlapping voxels were excluded. Percent overlap was calculated by dividing the number of overlapping voxels (overlap size) by the total number of voxels in the target structure. Cortical structure abbreviations are provided in Supplementary Data 4. Quantification of axons within brainstem nuclei using StARQ Utilizing the StARQ deep learning-based brainstem segmentation framework38, we fine-tuned the model on coronal brainstem sections with labeled SCPN axons (AAV-mediated anterograde labeling of SCPN in Cartpt-Cre and Npy-Cre mice) to automatically register and segment the brainstem regions of interest. Images were preprocessed using affine and non-rigid alignment. The fNissl signal was used for segmentation (registration channel) of the brainstem images to generate binary masks of brain regions with high confidence scores. Each predicted brainstem region was then assigned a unique RGB color code for signal quantification, i.e., labeled SCPN axons. The binary mask for each segmented brainstem region was used to filter out pixels in the corresponding signal channel of AAV + SCPN axons. The signal distribution was computed by multiplying the binary mask by the pixel intensities in the signal channel within the same section, effectively summing the pixel intensities within the masked region. The measured signals were normalized within each animal by dividing the signal for each region by the sum of signal across all regions in that animal. This in-animal normalization allowed for the comparison of relative signal distributions across different regions while accounting for variations in total signal strength between animals. Additionally, we measured the size of each region by calculating the total number of active pixels in the binary mask, ensuring that the area size was accounted for during signal analysis with no overt differences across experimental groups. Brainstem structure abbreviations are provided in Supplementary Data 2. Intersectional DREADD expression in SCPN subpopulations For the conditional silencing of SCPN subpopulations using inhibitory DREADDs, Npy-Cre mice received bilateral injections of rAAV-DIO-FlpO either into the cerebral peduncles at P0 for M1.SCPNoff and Npy.CBNoff groups or into the cervical spinal cord at P2 for the Npy.CSNCtrl group. The Npy.CBNoff and Npy.CSNCtrl groups received a bilateral injection of AAV-fDIO-hM4Di-mCherry at P1 into lateral cortex while the M1.SCPNoff group received a bilateral co-injection of the same construct with AAV-Cre at P1 into the M1 region of cortex to drive expression of the inhibitory DREADDs. Single pellet reaching task Following intersectional viral labeling at postnatal days 0–3 (P0–P3), adult Npy-Cre mice underwent behavioral training on the single pellet reaching task. Because retroAAV injections targeting either the cerebral peduncle (Npy⁺ CBNoff and M1.SCPNoff group) or the spinal cord (Npy⁺ CSN group) had to be performed during early postnatal development, mice could not be stratified into experimental groups based on subsequent behavioral performance. All injected mice were therefore treated identically at each behavioral timepoint. Mice were handled for 5 min daily 5 days per week using a clear acrylic tube (⌀ 4.5 cm), which was left in the home cage over night to allow mice to familiarize themselves with the tube88. After a week of handling, baseline weights were recorded, and mice were placed on food restriction. Rationing started at 0.1 g of food per gram of baseline body weight, which was adjusted over time to allow the mice to maintain weight above 85% of their baseline. Pellets were also added to each cage in addition to their daily rations to get mice familiar with them. Training for the task was divided into three phases: During habituation phase (lasting 1–3 days), mice were placed in an acrylic grasping chamber with a front-facing slit (based on Chen et al.89, gift of the Hollis lab) and allowed to explore for 10 min. Sugar pellets (20 mg, chocolate flavor, Bio-Serv) were placed in front of the slit, and forceps were used to draw the mouse’s attention to the slit. Upon eating pellets regularly from the slit, they were fed pellets directly from the forceps. Once mice consistently ate from the forceps, they were determined to be ready for shaping. In the shaping phase (lasting 1–5 days), pellets were held with forceps in front of the mouse and were offered only after the mouse reached out and touched it with a paw. Once the mice were consistently touching the pellets and hand preference was established, pellets were placed on a pole (TR1.5 - Ø1/2” Optical Post, L = 1.5”, Thorlabs) positioned 1 cm above the box floor and approximately 1.7 cm from the slit, with a slight offset (0.5 cm) to the side opposite from that of hand preference. This variant of the single pellet reaching task was specifically chosen to engage shoulder muscles during the reaching and grasping movements. Mice were rewarded for reaching towards the pole until they could consistently reach and touch the pellets, after which assistance was provided only when mice exhibited high levels of stress. Once mice were successfully reaching for pellets in an engaged manner and exhibited at least one successful grasp, we continued with the training phase (3–14 days). Reaching attempts were recorded as follows: Any touch of the pellet was recorded as an attempt. Additionally, successful reaches were counted when the pellet was successfully grasped and eaten. Sessions concluded after 25 attempts. Mice were considered fully trained after successfully grasping at least 30% of attempts on three consecutive days. If a mouse failed to meet this criterion in 10 sessions, it was classified as a non-learner and excluded from the experiment. After successful training, grasping performance was recorded for three consecutive days: To record baseline performance on day 1, mice received an intraperitoneal (i.p.) injection of 10 µL saline per 1 g body weight 30 min before recording. On day 2, mice received an i.p. injection of 10 µL DREADD ligand JHU37160 (J60; Hello Bio) for a final concentration of 1 mg/kg body weight, and grasping performance was recorded 30 min after. Mice were again injected with saline 24 h later, allowing sufficient time for the drug to be metabolized to record grasping performance after wash-out. Kinematic analysis of skilled reaching Kinematic analysis of skilled reaching was performed from front- and side-view recordings of 25 trials per session for each animal on baseline and JHU administration days. Sessions were recorded using two Basler acA1440-220 μm cameras (Basler AG, Ahrensburg, Germany) equipped with Basler C23-1216-2M F1.6 f12mm lenses (Basler AG). Both cameras were mounted on custom 3D-printed holders (gift of the Hollis lab) that positioned each camera level with the top of the reaching pole and at fixed distances from the pole ( ~ 16.5 cm for the front view, ~29 cm for the side view). Lens apertures were fully opened, and focus was adjusted to the pellet position. Recordings were acquired using Basler Pylon Viewer software with exposure set to 4800, a 720 × 540-pixel capture area, and a frame rate of 200 fps. For each session, videos were recorded continuously from before the first trial until after the final trial. Forelimb trajectories were tracked using DeepLabCut35 and processed with custom MATLAB scripts. For side views, the paw center was tracked directly; for front views, it was computed as the average position of digits 1 and 4. Tracking quality was quantified by the proportion of frames with DeepLabCut likelihood > 0.95; animals with mean tracking confidence <0.8 across sessions were excluded. Front and side videos were aligned by applying a manually measured temporal offset between cameras to synchronize reach frames across views. Frames containing a reach were identified from the side-view paw center trajectory. Reach onset was defined as the first frame where the paw crossed the slit, and the turning point as the frame of maximal paw extension, corresponding to either pellet contact or the maximum forward position. The end of the reach was defined as the first frame where the paw crossed back over the slit, or, if no return occurred, the first stable pause following peak extension. Each reach was then annotated by a blinded observer into one of four categories: success, approach error, grasp error, or retrieval error. Approach errors were defined as a reach in which digits never encircled the pellet, regardless of whether the paw contacted or brushed the pellet. A grasp error occurred when the digits closed around the pellet but failed to secure it. A retrieval error occurred when the pellet was successfully lifted but dropped before completion of the retrieval. A success was defined as lifting and transporting the pellet without dropping it. Trajectories were converted to millimeters, centered on the pellet, and smoothed with a low pass moving average filter to reduce jitter. Reaches with persistently low tracking confidence, failed segmentation, or insufficient valid samples after smoothing were excluded. From the resulting pellet-centered trajectories, 47 per-reach kinematic metrics quantifying speed and acceleration, path geometry, timing and accuracy, and digit coordination were computed (Supplementary Data 1). Bioinformatics Single-cell RNA-seq data were processed as recommended90. In brief, raw read files were processed and aligned to the mouse reference genome, mm10 (GENCODE vM23/Ensembl 98) using 10× Genomics Cell Ranger 6.0.1 count and aggr91. Using established R packages and custom-written code, cells with low read counts (empty droplets) or cells with a high number of mitochondrial gene products ( ± 3 × than median absolute deviation) were removed. Size-factor normalized logcounts were obtained92,93 and batch-corrected94. After integration of the three sample types (P1 SCPN, P3 SCPN, P3 CSN) using the top 2000 most variable genes with min. mean normalized expression of 0.00195, dimensionality reduction (UMAP) was done on the batch-corrected log counts. For cluster stability analysis, gene expression counts from P1Ped and P3Ped samples were normalized using the SCTransform function in Seurat96–98, regressing out mitochondrial gene percentages. FastMNN94 was applied to remove batch effects, and the top 48 dimensions of the corrected embedding were used to construct a kNN graph with varying values of k (4, 6, 8, 10), which were refined into SNN graphs. Clusterings were computed on each SNN graph using the Louvain algorithm at resolutions of 0.2, 0.3, 0.4, and 0.5. To determine the most stable clustering, 80% of the dataset was subsampled 20 times, repeating the clustering process. Jaccard indices (JI) were calculated for each subsample, and stable clusters were defined using the AssignStableCluster from scclusteval R package99 as having a JI > 0.8 in at least 16 out of 20 subsamples. A k value of 8 and resolution of 0.4 produced the most stable clusters (n = 6). Anchor cells between the P3SCc samples and the P1Ped and P3Ped samples were identified using the FindTransferAnchors function from Seurat. Using the identified anchors, the TransferData function was employed to classify P3SCc cells based on the stable cluster labels from the P1Ped and P3Ped dataset. The dataset was annotated with singleR100,101 using previously published single-cell data of the P0 mouse cortex to subset our dataset to Layer V neurons102. Axon extension specificity was defined as follows: clusters that were only present in the SCPN but not the CSN samples were annotated as “brainstem projecting”, while overlapping clusters were annotated as “spinal projecting”. Medial vs. lateral cortical location in our P1 sample was annotated using SingleR, leveraging microarray data from a previous study21 that compared medial vs. lateral SCPN at P1. This allowed us to label our P1 dataset based on transcriptional profiles matching the microarray time point. Marker genes between clusters were obtained using findMarker function in Seurat v4.097. Pseudotime trajectories were calculated from P1 to P3 (SCPN samples) using the destiny R package103. Diffusion components were calculated from the FastMNN-corrected embedding using DiffusionMap. This method captures cellular transitions through a non-linear dimensionality reduction technique104. Pseudotime was calculated for all cells using diffusion components with the DPT function, with the starting point set as the cell ranked first in the diffusion map’s eigenvector. To identify genes with continuous expression changes over pseudotime, a generalized additive model was fitted to each gene’s expression using the gam function from the mgcv R package105. The data were normalized with SCTransform, and PCA was performed to reduce dimensionality, followed by mutual nearest neighbors (MNN) embedding and UMAP visualization. Diffusion maps were computed on the MNN coordinates to calculate pseudotime, which was then visualized using ggplot2. Mapping to mouse adult single-cell data of whole cortex24 was done as follows: The gene expression matrix from the P1–P3 developmental dataset was combined with a mouse adult single-cell dataset (P56). FastMNN was used to align cells across the datasets, correcting for batch effects while preserving biological variation. The top 30 MNN-corrected dimensions were used to construct a kNN graph with the FindNeighbors function, which was refined into an SNN graph. Clusters were identified on this SNN graph using the Louvain algorithm in Seurat with a resolution of 0.8. For visualization, Uniform Manifold Approximation and Projection (UMAP) was computed on the MNN-corrected embedding using the RunUMAP function. Statistical analysis Statistical analysis other than those related to the single-cell RNA-seq data was performed in Prism 10.3 (GraphPad Software) and R version 4.4.1. Ordinary one-way ANOVA followed by Tukey’s HSD multiple-comparisons test was used for group comparisons (within-group differences in cortical locations, behavior, and across-group differences in brainstem regions of interest). Two-way ANOVA was used to assess the effects of time and cortical region and their interaction in RNAScope quantifications. For comparisons of retrogradely labeled SCPN vs. CSN at P4, mixed-effects models with restricted maximum likelihood estimation (REML) were used to account for variability in bregma alignment. For nested measurements (e.g., soma size and cortical depth), linear mixed-effects models (REML) were used with gene group as a fixed effect and animal as a random intercept. A two-sided Wilcoxon rank-sum test (Mann–Whitney U test) was used to compare Npy and Cartpt expression at P1 vs. P3 in the scRNA-seq dataset (cells with expression > 0 only). Kinematic analyses of single pellet reaching were performed on data from five Npy.CBNoff mice with consistently high-quality tracking; mice with unreliable DeepLabCut trajectories were excluded. Per-mouse error-type proportions (Success; Error during Approach/Grasp/Retrieval) were compared between baseline and JHU in Npy.CBNoff mice using linear mixed-effects models (day fixed; animal random), with FDR-adjusted pairwise contrasts. For multivariate kinematics (PCA), features were variance-screened, mean-imputed, and z-scored. PCA was performed on all reaches, and day effects were tested by MANOVA on the first two principal components (Pillai’s trace). Canonical discriminant analysis was derived from the same MANOVA. Univariate mixed-effects models (day fixed; animal random) were fit per feature with multiple comparisons controlled using the Benjamini–Hochberg false discovery rate (FDR). Approach-phase features were further analyzed using PCA denoising ( ≤ 15 PCs), UMAP (n_neighbors = 15, min_dist = 0.08), and k-means clustering (k = 2–6, selected by silhouette score). Cluster composition was summarized across day and animal, and cluster-defining features were identified using mixed-effects models (cluster fixed; animal random) with FDR correction. For comparisons between approach-error and other reaches, per-feature mixed-effects models (day fixed; animal random) were fit separately for each set, and paired effect sizes were compared. Multiple comparisons were controlled using FDR (q < 0.05). For comparisons of axonal projection densities across brainstem regions, post hoc comparisons were corrected using the Benjamini–Hochberg FDR procedure. Statistical tests were two-sided unless otherwise specified. P values are unadjusted unless otherwise indicated. In all bar graphs and box plots, dots represent individual animals unless otherwise indicated. In bar graphs, data are presented as mean ± SEM. In box plots, data are represented as median with 25th–75th percentiles (box) and min-max (whiskers). Statistical tests were two-sided unless otherwise specified. P values are unadjusted unless otherwise indicated. Statistical significance was defined as *p < 0.05; **p < 0.01; ***p < 0.001. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Supplementary Movie 1 Supplementary Movie 2 Supplementary Movie 3 Supplementary Movie 4 Supplementary Movie 5 Reporting Summary Peer Review file Source data Source Data Acknowledgements We thank the members of the Hollis Lab, particularly Amanda Bernstein and Joseph Minkler, and the Yoshida Lab, especially Nishtha Ranawat, Fumiyasu Imai, and Hirohide Takatani, for their technical guidance. We also extend our thanks to our student volunteers Jimena Tetuan-Perez, Nirupama Angira, Eliana Zitrin, Eytan Sieger, and Ben Selip for their valuable assistance. Our gratitude goes to Michał Bonar from NanoCellect for his support with FACS purification and to Dimitris Konstantopoulos from Genevia Technologies for assistance with transcriptomic analyses. We thank Linda M. Gerber for assistance with statistical analysis and Abigail L. Person for valuable feedback on the kinematic analyses that significantly improved the manuscript. We acknowledge the infrastructure support provided by Weill Cornell Medicine, particularly the ABC Core, and the BNI Structural and Functional Imaging Core. Author contributions J.K. and V.S. conceived and designed the study; J.K., P.P., S.F., A.L., M.K. and A.I. performed experiments and analyzed data; M.P.G. provided resources and analyzed data; J.K. and V.S. interpreted the data and wrote the manuscript; all authors reviewed and approved the final manuscript. Schematics were created by J.K. and do not contain third-party content. Peer review Peer review information Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available. Funding This work was supported by the National Institutes of Health (NINDS R01NS131662, NTRAIN/NICHD K12HD093427, S10OD028547, S10OD036432), the New York State Department of Health (C39069GG), the Wings for Life–Spinal Cord Injury Foundation, and the Craig H. Neilsen Foundation. Infrastructure support was provided by the Burke Foundation to V.S. J.K. was partially supported by a postdoctoral fellowship from the Swiss National Science Foundation (SNSF, P2EZP3_191858) and is a co-investigator on C39069GG. The Zeiss Lightsheet 7 microscope at the Optical Imaging Facility, UT Health San Antonio, was funded by NIH grant 1S10OD030383. Data availability The raw sequencing data generated in this study have been deposited in the NCBI Gene Expression Omnibus (GEO) under accession code GSE270265 and are publicly available [https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE270265]. Source data are provided with this paper. Code availability FIJI/MATLAB pipelines for 3D reconstruction of injection volumes and cell coordinates within CCFv3 space are available at [https://github.com/jkaiser87/] (VOL3D106 for injection volume, CELL3D107 for cell coordinates, based on transformation of 2D sections into CCFv3 space using AP_histology v1108). Kinematic analysis of single pellet reaching was performed using custom MATLAB code (Sahni_SPG_analysis109). The machine-learning pipeline StARQ [https://github.com/itsasimiqbal/StARQ] was used for segmentation and quantification of axons across brainstem regions. Competing interests A.I. is the founder of Tibbling Technologies and has no competing interests related to this manuscript. All other authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary information The online version contains supplementary material available at 10.1038/s41467-026-73476-4. References Associated Data Supplementary Materials Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Supplementary Movie 1 Supplementary Movie 2 Supplementary Movie 3 Supplementary Movie 4 Supplementary Movie 5 Reporting Summary Peer Review file Source Data Data Availability Statement The raw sequencing data generated in this study have been deposited in the NCBI Gene Expression Omnibus (GEO) under accession code GSE270265 and are publicly available [https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE270265]. Source data are provided with this paper. FIJI/MATLAB pipelines for 3D reconstruction of injection volumes and cell coordinates within CCFv3 space are available at [https://github.com/jkaiser87/] (VOL3D106 for injection volume, CELL3D107 for cell coordinates, based on transformation of 2D sections into CCFv3 space using AP_histology v1108). Kinematic analysis of single pellet reaching was performed using custom MATLAB code (Sahni_SPG_analysis109). The machine-learning pipeline StARQ [https://github.com/itsasimiqbal/StARQ] was used for segmentation and quantification of axons across brainstem regions.