The striatum, pallidum and substantia nigra are key components of the basal ganglia, and they process inputs from the entire neocortex5,6. They constitute a critical node in the cortico–basal ganglia–thalamo–cortical loop1,7,8,9,10,11,12. This recurrent network is associated with diverse functions and behaviours13,14,15,16,17, and aberrant basal ganglia function is implicated in movement disorders17,18,19, neuropsychiatric disorders20,21,22 and drug addiction23. Identifying the specific subnetworks within the loop is key to understanding how this multitude of functions and pathologies is governed. The consensus view is there are three parallel channels of information flow through the basal ganglia: associative, limbic and sensorimotor1,2,3,4 (Extended Data Fig. 1a). Previous efforts to refine the three-channel model have postulated a number of specific parallel subnetworks1,11. Yet it has not been possible to provide a definitive wiring diagram to support a more refined model owing to incomplete basal ganglia connectional data at sufficiently high spatial resolution. However, a recent multi-scale network organization of the mouse corticostriatal pathway subdivided the caudoputamen (CP) into 29 fine-scale network divisions termed domains5 (Extended Data Fig. 1a–c), a finding that suggests that there is a more granular level of organization of the basal ganglia. Here we systematically map the multi-synaptic output pathways of all CP domains through each sequential node of the loop.


Striatonigral pathway defines SNr domains

Data for the striatal output pathway analysis were produced by injection of anterograde tracer into each CP domain5 as well as the core (ACBc), medial shell and lateral shell of nucleus accumbens (Extended Data Fig. 1d), yielding a representative set of 36 injections (Extended Data Fig. 2). Coronal images containing GPe and SNr were registered, segmented, reconstructed and quantified. Network analysis of the data partitioned the striatal domains with convergent axonal fields into communities, and their terminal zones were demarcated as a new domain of GPe or SNr. The axonal reconstructions were used to make projection maps that demonstrate the underlying axonal data defining each new domain (Extended Data Fig. 1e–k).

The topographical organization of the striatonigral (direct) pathway is depicted in the projection maps (Fig. 1a, Extended Data Figs. 3a, 4a) and aligns well with data from rat and monkey8,10,12,24. Following network analysis of these data, we identified 14 domains in SNr (Fig. 1b, c, Extended Data Figs. 57, Supplementary Table 1). Most SNr domains receive convergent inputs from multiple striatal domains (Supplementary Table 1), each of which in turn receives a unique set of distinct cortical inputs5. For instance, SNr dorsal (SNr.d) and dorsomedial ( are limbic domains that collectively receive inputs from a number of limbic striatal domains (Extended Data Fig. 8a, b). In turn, these striatal domains receive inputs from limbic cortical areas that are themselves interconnected and constitute the lateral cortico-cortical subnetworks5,25, which are involved in perception of internal states and memory associated with emotion25,26. The majority of SNr domains span two or more rostrocaudal levels and several run the entire length of SNr, reflecting their inputs: striatonigral projections terminate in discrete zones within SNr that are restricted mediolaterally and dorsoventrally, but most (75%) span the entire rostrocaudal extent of SNr, forming longitudinal columns of terminal axons (Extended Data Fig. 2, 3b). Individual striatal axons conform to this topography, which was confirmed at the single-cell level using fluorescence micro-optical sectioning tomography (fMOST) (Fig. 1d, Extended Data Fig. 9) and is consistent with primate data27.

Fig. 1: SNr domain structure.
figure 1

a, Reconstructed axonal projections from each striatal domain (top) are coloured to match their source domain and projected onto SNr maps (bottom) from six representative levels of the Allen Reference Atlas51 (ARA). b, 3D striatonigral matrix (see Extended Data Fig. 7). c, SNr domain map derived from analysis of axonal projections in a. d, fMOST imaging reveals that individual striatal neurons project through the entire frontal axis of SNr (see also Extended Data Fig. 9). eg, Anterograde tracer (PHAL, pink) in striatal domain and retrograde tracer (dG-rabies-GFP, green) in thalamic domain PF.a (e) reveal that input striatonigral axonal topography matches output nigrothalamic neuronal topography from rostral through caudal SNr.medial domain (f, g). Region outlined in g is expanded (f, bottom). Images are representative of n = 2 biological replicates. Scale bars, 30 µm (f), 200 µm (g). hj, Reconstructions of these neurons (h) are shown with 4 striatal inputs (i), with the bottom left region enlarged in j. k, Rabies-labelled, SHIELD-cleared brain reveals whole SNr neurons. Right, magnified view of outlined regions. Scale bars, 1 mm (left), 200 µm (right). Collectively, images in ek suggest SNr neurons are capable of integrating most inputs to their domain at a given coronal level. Max, maximum.

Graphic reconstruction of SNr neurons reveals a dendritic morphology that allows for receipt of convergent striatal inputs from multiple CP domains by individual SNr neurons. Thalamic injection of GFP-labelled glycoprotein-deleted rabies (dG-rabies-GFP) labels neurons with extensive dendritic arbors in the SNr medial domain (SNr.m) (Fig. 1f). Reconstructions of these neurons are shown overlaid onto a composite projection map of four convergent striatal terminal fields (Fig. 1g–j). Most of these neurons’ dendritic arbors appear capable of contacting all four terminal fields. However, these neurons were reconstructed from a tissue slice 150 µm thick, resulting in truncation of the dendrites in the rostrocaudal axis. Full SNr neuronal reconstructions in SHIELD-cleared intact tissue reveal that their dendrites spread partway into adjacent rostral and caudal levels, but do not span the entire length of SNr (Fig. 1k). Collectively, these data suggest that SNr projection neurons are capable of integrating most inputs to their domain at a given coronal level, but not across the rostrocaudal length of the large domains. Thus the small changes in striatal input structure across a domain at different rostrocaudal levels (Supplementary Table 1) could result in slightly different integration profiles of neurons along the length of a domain.

Finally, we traced a striato–nigro–thalamic pathway (Fig. 1e) by injecting the anterograde tracer Phaseolus vulgaris leucoagglutinin (PHAL) in striatal domain and retrograde tracer dG-rabies-GFP in the associative subregion of the parafascicular thalamic nucleus. The results reveal close and extensive juxtaposition of anterogradely labelled striatal axons and retrogradely labelled SNr perikarya through the entire rostrocaudal extent of SNr.m (Fig. 1f). The matching input–output structure of striatal terminals and thalamic-projecting nigral neurons supports the notion of longitudinal SNr domains as a functional unit.

Parallel striatal projections to GPe

Network analysis of the striatopallidal (indirect) pathway data reveals 36 domains in GPe (Fig. 2a, Extended Data Fig. 10, Supplementary Table 2). The majority of domains (64%) have only one or two inputs, and most domains (69%) span just one coronal level of GPe. The majority of striatal domains project to GPe with restricted terminal fields that have little overlap and convergence (Extended Data Figs. 4, 11), indicating that the indirect pathway is characterized by a higher degree of specificity and parallelization compared with the direct pathway. This is reflected in the projection maps, in which more individual colours can be seen in GPe and there is more overlap of coloured terminal fields in SNr (compare Fig. 1a, Extended Data Fig. 11a). To quantify this difference, convergence of striatal inputs in the grid box datasets of GPe and SNr was analysed by tallying the number of striatal domains terminating in each grid box (Extended Data Fig. 12a). Frequency distributions of these data show that SNr receives more inputs per grid box than GPe (compare Extended Data Fig. 12b, g), approximately two times more, on average (P < 0.0001; Fig. 2b).

Fig. 2: Indirect pathway defines GPe domain structure and differs from direct pathway.
figure 2

a, Striatal axonal inputs (see Extended Data Fig. 11) define GPe domains. Striatal inputs are listed on the right. b, Analysis of grid box data reveals that the mean number of striatal domains providing inputs to each grid box differs significantly between GPe and SNr, indicating greater convergence in the direct pathway (GPe: 3.75 ± 1.79, SNr: 6.49 ± 2.97; P < 0.0001, two-tailed t-test, t = 20.26, d.f. = 1,057; data are mean ± s.d.). Total numbers of grid boxes in GPe (n = 676) and SNr (n = 649) are similar, meaning their gross volumes are similar. c, Schematic of experimental strategy for electrophysiological test of parallelism in indirect pathway and convergence in direct pathway. d, Example recording traces from RFP-tagged neurons in GPe and SNr. e, Number of neurons responding to stimulus was significantly different between GPe (n = 6) and SNr (n = 6; P = 0.0152, Fisher’s exact test, 3 mice). f, Peak amplitude of current after stimulation was significantly different between GPe and SNr neurons (GPe: 33.67 ± 24.74 pA, SNr: 267.5 ± 197.6 pA; P = 0.0348, two-tailed t-test, t = 2.876, d.f. = 5; data are mean ± s.d.). g, Simplified schematic of the direct and indirect pathways. In box plots, boxes demarcate first and third quartiles, whiskers show minimum and maximum values, the solid line is the median and the dashed line is the mean.

Source data

To functionally validate this, we performed a recording experiment combining channelrhodopsin (ChR2)-assisted circuit mapping (CRACM) with anterograde transsynaptic tracing. One CP domain was injected with adeno-associated virus (AAV) expressing ChR2 (AAV-ChR2) and a separate CP domain was injected with AAV1-Cre. The AAV1-Cre is transported and infects postsynaptic neurons in GPe and SNr, which are visualized with injections of AAV-FLEX-RFP (Fig. 2c). The ChR2-labelled axons should colocalize with RFP-labelled neurons in SNr but not in GPe; hence RFP neurons in GPe should not respond to stimulation of ChR2 axons, whereas RFP neurons in SNr should respond. Patch clamp recordings of RFP-labelled neurons were consistent with this prediction, with nearly all GPe neurons (5 out of 6) showing no response and all SNr neurons (6 out of 6) showing a response to stimulation (P = 0.0152; Fig. 2d, e), and peak current after stimulation was also significantly different (P = 0.0348; Fig. 2f). Notably, both nigral and pallidal neurons were recorded from all subjects, enabling direct within-animal comparisons. These anatomical and functional findings support a model of indirect pathway parallelization and direct pathway convergence (Fig. 2g).

Direct and indirect pathways converge in SNr

Direct and indirect pathway neurons are intermingled in striatum28, such that one domain of CP projects to specific domains of both SNr and GPe. The GPe relays indirect pathway information to SNr via a strong, direct projection29. Yet the topography and specificity of how these parallel pathways diverge and re-converge are unknown. The direct pathway has a ‘bridging collateral’ to GPe en route to the nigra30. We used Cre-dependent tracing in D1- and A2A-Cre mice to demonstrate that terminals of both direct and indirect pathways in GPe have the same topography when arising from the same striatal source, and fMOST imaging shows the same pattern at the single-cell level (Extended Data Fig. 13).

We next show that pallidonigral projections from a GPe domain converge with striatonigral axons arising from the same CP domain that serves as input source to both nuclei. Injection of anterograde tracer in limbic GPe domains 18 and 20 labels projections to nigral limbic domain SNr.d and limbic domains of CP and thalamus (Extended Data Fig. 8c–g). GPe.18 and 20 and SNr.d receive inputs from the same striatal domains (CPi.vm.v, CPi.vm.vm and CPi.vm.cvm)5. We validated this finding in a second pathway with a retrograde tracer injection in SNr.m (Extended Data Fig. 14). Labelling is seen in GPe.3, and in CP domains that innervate both SNr.m and GPe.3. Thus direct and indirect pathways arising from a common striatal source converge in SNr.

To ascertain whether homotypic direct and indirect pathways actually synapse onto the same postsynaptic neurons in SNr, we performed a CRACM–anterograde transsynaptic tracing experiment (Extended Data Fig. 8h). AAV1-Cre was injected into CP, which transsynaptically infected neurons in GPe and SNr, causing those cells to express Cre, and AAV-DIO-ChR2 was injected into GPe and AAV-FLEX-RFP was injected into SNr. RFP-expressing neurons in SNr were recorded during optical stimulation of pallidonigral terminals, evoking an inhibitory response that was capable of suppressing nigral neuronal activity (Extended Data Fig. 8i, j). The majority of recorded SNr neurons (10 out of 14) showed this inhibitory response, demonstrating re-convergence of separately processed information from the direct and indirect pathways onto individual SNr neurons (Extended Data Fig. 8k, l).

Parallel output channels of parafascicular thalamus

The thalamus is the final node in the cortico–basal ganglia–thalamic loop. Two of the densest nigral outputs are to the parafascicular (PF) and ventromedial (VM) thalamic nuclei. The PF is intricately interconnected with the other nodes of the loop, such that topographically connected subregions of cortex, striatum and nigra connect topographically with a discrete PF subregion4,31,32 (Extended Data Fig. 15a–c). Injections of anterograde and retrograde tracers in cortex, CP and SNr demonstrate this motif, showing highly specific connectivity patterns with six subregions of PF (Extended Data Fig. 15d, e). These subregions appear to be parallel output channels for integrating basal ganglia efferent signals with cortical inputs and conveying that computation to striatum and cortex. They correspond to associative (PF.a), trunk and lower limb (, upper limb (PF.ul), mouth (PF.m), limbic (PF.lim) and ventral striatal (PF.vs) domains.

For example, the ventral striatal subnetwork (Extended Data Fig. 15d, e, ventral striatal) contains infralimbic (ILA) and medial orbitofrontal (ORBm) cortex, which are interconnected and both project to ACBc and PF.vs; the ACBc also receives input from PF.vs and projects to, which in turn projects to PF.vs; the PF.vs projects back up to ILA, closing the loop. Anterograde transsynaptic tracing shows the actual synaptic specificity of the ACBc––PF.vs pathway: AAV1-Cre was injected into ACBc and AAV-FLEX-RFP into medial SNr (Extended Data Fig. 16). Labelled neurons are seen specifically in, and their axonal labelling terminates precisely where the other nodes of the ventral striatal subnetwork connect, the PF.vs (and VM.vs; Extended Data Fig. 15d–g,

VM has distinct cortical innervation topography

The SNr also projects to VM, and its terminals there are even denser than to PF. The VM also contains six output channels (Extended Data Fig. 15f–h), but with a different organizational scheme: unlike PF, with separate domains for the body sub-regions, VM has one domain (VM.s) projecting to secondary motor cortex (MOs) and another (VM.p) projecting to primary motor (MOp) and primary somatosensory (SSp) cortex. The VM.s domain projects to all somatic subregions (that is, ul, ll and tr) within MOs, whereas VM.p projects to all somatic subregions within MOp and SSp (Extended Data Fig. 15g, h). A separate domain for the mouth pathways, VM.m, projects specifically to mouth and head regions of MOp, SSp and MOs, similar to PF.m. The associative (VM.a), limbic (VM.lim) and ventral striatal (VM.vs) channels are similar in topography to homologous regions of PF. The boundaries of each of these domains were established by the position of the thalamocortical neuron groups (Extended Data Fig. 15g). Each VM domain receives a specific nigral input from at least one of the SNr domains, and this nigrothalamic input conforms to the VM domain boundaries established by thalamocortical tracing (Extended Data Fig. 15g, h).

Whole cortico–basal ganglia–thalamic loops

The cortico–basal ganglia–thalamic loop model1,2,18 is supported by electrophysiological and anatomical experiments that demonstrate segments of this network9,10,11,31,33,34,35. Yet whole circuitous loops have not been demonstrated within a single animal using any methodology. Moreover, calling the network a ‘loop’ begs the question whether it is in fact a closed-circuit loop.

We first demonstrate whole loops using the double co-injection technique for mapping interconnected network structures (Fig. 3a, Extended Data Fig. 17b). Two anterograde–retrograde tracer pairs were injected into two non-adjacent nodes of the oro-brachial subnetwork. In a serially connected four-node loop, labelling from the co-injection pairs will converge and appose in the non-injected nodes. Injections into CP and PF orofacial regions reveal overlapping labelling in the SNr oro-brachial domain (SNr.orb) and a cortical column in mouth primary motor region MOp-m/i (Fig. 3a). The same strategy demonstrates the associative subnetwork loop (Extended Data Fig. 17a). Moreover, when co-injections are placed in nodes of two separate but neighbouring loops, labelling is seen in separate, neighbouring regions of SNr and cortex (Extended Data Fig. 18), illustrating the largely parallel nature of these loops.

Fig. 3: Demonstration of the oro-brachial subnetwork loop.
figure 3

a, Double co-injection labels the loop structure of the mouth component of the oro-brachial subnetwork loop. Co-injections of anterograde–retrograde tracer pairs into striatal and thalamic oral domains (CPi.vl.v and PF.m) reveal overlapping labelling in SNr oro-brachial domain and mouth primary motor cortex MOp-m/i. Images are representative of n = 6 replicates. Scale bars, 100 µm. CTB, cholera toxin subunit B–Alexa Fluor 647 conjugate. b, Labelling of corticostriatal (green) and corticothalamic (red) neurons and thalamocortical axons (light blue) from the striatal and thalamic mouth domains reveals two zones of overlap of labelling (insets), most densely in layer 4 of MOp-m/i cortex where thalamocortical axons overlap the apical dendrites of corticostriatal and corticothalamic neurons. Representative of n = 2 replicates. Scale bar, 30 µm. c, Mice (n = 2) were injected with AAV-ChR2 into PF.m to label thalamocortical axons with channelrhodopsin and fluorophore-tagged retrobeads into CPi.vl.v to retrogradely label corticostriatal neurons. d, e, In acute slice preparation, the majority of recorded labelled neurons (9 out of 13) showed excitatory postsynaptic current response to stimulation when clamped at −70 mV (44.89 ± 42.8 pA (mean ± sd)), and none showed inhibitory postsynaptic current at 0 mV. Box plot as described in Fig. 2.

Source data

The colocalization of thalamocortical axons and corticostriatal neurons is strongly suggestive of recurrent feedback within a subnetwork loop. In the oro-brachial subnetwork, this probable closed-circuit zone lies in MOp-m/i (Fig. 3b). There, corticostriatal neurons providing input to CPi.vl.v have apical dendrites that ascend through a dense field of axons from PF.m terminating in layer 4 (Fig. 3b, insets; Supplementary Video 1). To unambiguously determine whether the loop is truly recurrent, we injected AAV-ChR2 into PF.m to opsin-label thalamocortical axons and fluorescent retrobeads into CPi.vl.v to retrogradely label corticostriatal neurons (Fig. 3c). Labelled corticostriatal neurons were patch-clamped in acute slice preparation during blue light stimulation. The majority (9 out of 13) showed an excitatory postsynaptic current to stimulation (Fig. 3d, e). The associative subnetwork loop is also demonstrated this way (Extended Data Fig. 17c). With the majority of recorded neurons exhibiting a specific monosynaptic excitatory response to thalamocortical stimulation, the cortico–basal ganglia–thalamic loop contains a substantial closed-circuit component.

Cortex sends a direct projection to SNr

The hyperdirect pathway from cortex to subthalamic nucleus (STN) to SNr was thought to be the fastest route for cortical information to reach SNr36. We demonstrate that certain cortical areas of the oro-brachial subnetwork project directly to SNr.orb (Extended Data Fig. 19). The corticonigral pathway is seen in sagittal section by injection of AAV1-Cre into MOp-m/i of Ai14 mice, labelling fine axonal fibres in caudal SNr (Extended Data Fig. 20a–d). This pathway was validated by injecting SNr.orb with AAVretro-Cre and MOp-m/i with AAV-FLEX-RFP (Extended Data Fig. 20e–g). Again, cortical axons are seen impinging deeply into SNr with topographic specificity (Extended Data Fig. 20g, h). Collaterals from this pathway are seen in STN and oromotor regions of CP, GPe and PF (Extended Data Fig. 20i–l). To verify that corticonigral axons bear boutons, we injected an AAV inducing expression of GFP-tagged synaptophysin and cytoplasmic RFP into MOp-m/i, labelling red corticonigral axons bearing green boutons in SNr.orb (Extended Data Fig. 20m). Finally, this pathway was functionally validated with anterograde transsynaptic tracing, since only functional synapses transmit AAV. Injection of AAV1-Cre in MOp-m/i of Ai14 mice labels postsynaptic SNr neurons (Extended Data Fig. 20c). This characterization of the corticonigral pathway suggests that cortex can directly activate all components of the oro-brachial subnetwork (Extended Data Fig. 20n).


This work reveals that the canonical cortico–basal ganglia–thalamic network is composed of six parallel subnetworks, each of which is organized by a number of nodes that are precisely and richly interconnected (Fig. 4a). A model of the oro-brachial subnetwork exemplifies this interconnectivity (Fig. 4c). This comports with recent findings demonstrating specific parallel basal ganglia–thalamo–cortical pathways31. Together, the data presented here are consistent with a cortico–basal ganglia–thalamic closed-loop model that has long been hypothesized1,2,18.

Fig. 4: Summary models of main findings.
figure 4

a, Cortico–basal ganglia–thalamic loop model. In cortex, two medial associative subnetworks of exteroceptive sensory areas, two lateral associative subnetworks of interoceptive limbic areas and five somatic sensorimotor subnetworks of body regions project into largely distinct striatal subnetworks, whose outputs form the more parallel indirect (striatopallidal) pathway and the more convergent direct (striatonigral) pathway. The pallidal domains send convergent projections to the same nigral domains targeted by their input striatal domains. The nigral domains then project to six regions of the parafascicular thalamus, which in turn are interconnected with the originating cortical regions. b, Model of striatal output topography, illustrating convergence in the direct pathway and parallelism in the indirect pathway. Note that the small and large points represent the true topographic patterns of source and target zones, respectively. c, The oro-brachial subnetwork model highlights the precise separation between parallel pathways, the rich interconnectivity within pathways, and the complexity of the cortico–basal ganglia circuit. Not all pathways are labelled or depicted. Cortical areas: AI, agranular insular; ECT, ectorhinal; ENT, entorhinal; GUS, gustatory; PERI, perirhinal; PIR, piriform; PL, prelimbic; PTLp, posterior parietal association; RSP, retrosplenial; TEa, temporal association; VIS, visual; VISC, visceral.

Although the pallidum and nigra have not previously been conceptualized in a ‘domain’ structure of convergent striatal afferents per se, previous experimental findings have shown the same kinds of highly dense, topographically restricted striatal terminal fields in GPe and SNr8,10,12,37,38 and topographical congruence between inputs and output zones in the striatonigral37 and striatopallidal39 pathways. Moreover, recordings of SNr neuronal activity have found that all neurons sampled within opsin-labelled direct- or indirect-pathway terminal fields responded to stimulation40 (C. J. Wilson, personal communication), indicating that our domains and the convergent striatal projections that they represent are a good proxy for synaptic convergence of striatal information.

We demonstrate that the striatonigral pathway has greater convergence than the striatopallidal pathway (Fig. 4b). Anatomical studies in rats and monkeys lend support to this finding8,10,38, and a similar pattern was found in ventral striatum41. Electrophysiological studies also demonstrated a lower degree of informational convergence in GPe42,43,44 and a higher degree in SNr34,45,46. The greater specificity of the indirect pathway is likely to have a functional significance—for example, in Parkinson’s disease. In monkeys rendered Parkinsonian by dopamine depletion, GPe neurons become responsive to a wider range of striatal inputs;43,47 in humans with Parkinson’s disease, activity in GPe neurons is reduced, as greater activity is driven through the indirect pathway48. A number of neuropsychiatric and movement disorders17,18,19,20,21,22,23 probably involve alterations in specific subnetworks of the cortico–basal ganglia–thalamic network that perform specific cognitive and behavioural functions, and the combination of malfunctions in these circuits underpin complex disorders.


Subjects and surgeries

Subjects (in toto 268 male 2-month-old wild-type C57Bl6 and Ai14 mice, Jackson Laboratories) were anaesthetized with 2% isoflurane in oxygen. Buprenorphine SR (1 mg kg−1) was administered at the beginning of the surgery as an analgesic. Glass micropipettes (10–30 μm diameter tip) filled with tracer were lowered into the target region and delivered an injection either by iontophoresis (1–10 min. infusion time, 5 μA, 7 s current pulses) or by pressure injection (20–80 nl volume). The tracers used were: PHAL (2.5%; Vector Laboratories, L-1110); AAV-GFP (AAV1-hSyn-EGFP-WPRE-bGH, Addgene); AAV-RFP (AAV1-CAG-tdTomato-WPRE-SV40, Addgene); red and green glycoprotein-deleted rabies (Gdel-RV-4tdTomato and Gdel-RV-4eGFP, I.W. laboratory) which are incapable of transsynaptic spread; AAV1-hSyn-mRuby2-Syp-eGFP (B.K.L. laboratory); AAV1-Cre (AAV1-hSyn-Cre-WPRE, Addgene 105553); AAVretro-Cre (AAVretro-EF1a-Cre, Salk Institute); Cre-dependent AAV-FLEX-GFP (AAV1-CAG-Flex-eGFP-WPRE-bGH, Addgene); Cre-dependent AAV-FLEX-RFP (AAV1-CAG-Flex-tdTomato-WPRE-bGH, Addgene 100048); AAV-ChR2 (AAV1-hSyn-ChR2-EYFP-WPRE, Addgene 26973); Cre-dependent channelrhodopsin, AAV-DIO-ChR2 (AAV1-EF1a-DIO-ChR2-EYFP-WPRE, Addgene 20298); Fluorogold (FG, 1%, Fluorochrome); rhodamine-conjugated retrobeads (Lumifluor); and cholera toxin subunit B–Alexa Fluor 647 conjugate (CTB, 0.1–0.2%; Invitrogen). Most animals received multiple tracer injection combinations with non-overlapping fluorescence profiles, creating a pool of ~700 injections. A total of 138 mice were injected specifically for the striatal-output-pathway analysis, and the remainder were used for the other experiments in this manuscript. Animals were monitored daily after surgery until their body weight was on an increasing trajectory. All methods were approved by the Institutional Animal Care and Use Committees of the University of California, Los Angeles, the University of Southern California, the University of California, San Diego, and the Huazhong University of Science and Technology.

Roster of injections for striatal output analyses

Altogether, 36 anterograde tracer injections were selected and analysed as representative injections from a data pool of 138 mice with iontophoretically delivered triple anterograde injections (PHAL, AAV-GFP and AAV-RFP) or double co-injections25,49 (AAV-GFP and CTB; AAV-RFP and FG), collectively constituting 448 injections. All domains were injected more than once across this data pool, and injections within the same domain produced nearly identical labelling patterns. A representative injection for each domain of the caudoputamen was selected from this data pool based on precision of the injection within the targeted domain, quality of the axonal labelling and histological quality. Three injections were chosen for the nucleus accumbens (ACB), one in the core (ACBc) and two in the shell, medial (ACBsh.m) and lateral (ACBsh.l). The core, medial and lateral shell have divergent connectivity patterns, and although they mainly connect with a ventral basal ganglia network (ventral pallidum, substantia innominata and ventral tegmental area), they also send limited projections to restricted regions of the dorsal basal ganglia network50. We selected injections for 30 domains of the CP, the 29 domains from the rostral (CPr), intermediate (CPi), caudal (CPc) and caudal extreme (CPext) described in ref. 5, plus a new subdivision in the CPext. The CPext previously had a dorsal (CPext.d) and a ventral domain, but based on differing input and output patterns, the ventral domain here was split into the rostral ventral (CPext.rv) and caudal ventral ( domains. Additionally, three CP injections were chosen from a level intermediate to CPr and CPi, with projections to regions of the GPe not targeted by any of the other injections included in this experiment. Based on their relative position in the lateral CP, these injection sites appeared to be rostral associations of the somatomotor domains CPi.dl.d (tr), CPi.vl.imv (ul), and CPi.vl.v (m/i); furthermore, their striatonigral projections were highly similar to those three domains. Because their striatopallidal projections terminated in regions in the GPe that were targeted by no other CP domains, they were included in the analysis of the GPe data. However, since their projections to the SNr and GPi were homologous to the CPi-level injections’ projections patterns, they were excluded from the analyses of SNr data. Collectively, 29 mice yielded the 36 representative injections (some mice yielded more than one selected injection). The other major components of the cortico–basal ganglia–thalamic network, that is, the substantia nigra compact part, internal globus pallidus, subthalamic nucleus and various other thalamic nuclei, were not analysed in the present work.

Histology and imaging

Animal subjects were deeply anaesthetized with an overdose bolus of sodium pentobarbital (Euthasol, 2 mg kg−1, intraperitoneal injection), and cardiac perfused with normal saline followed by 4% boric acid-buffered paraformaldehyde. Brains were post-fixed overnight, embedded in 4% agarose, and sectioned on a vibratome at 50 μm thickness (50–150 μm for rabies-labelled tissues), collected in a 1-in-4 manner into 4 equivalent series, and stored in cryoprotectant at −20 °C until staining. Tissue series were stained with rabbit polyclonal anti-PHAL antibody (Vector Labs AS-2300) at 1:5,000 and donkey anti-rabbit AlexaFluor 647 (Jackson ImmunoResearch, 711-605-152). Nissl substance was stained with NeuroTrace 435/455 (ThermoFisher, N21479) at 1:500 to reveal cytoarchitecture. Sections were scanned on an Olympus VS120 epifluorescence microscope running VS-Desktop software with a 10× lens (Plan Apochromat) to capture the Nissl, FG, GFP, RFP and far red tracers in multichannel photomicrographs; these images were processed for the striatofugal network analysis. High-resolution images of some tissue samples (including the rabies-labelled tissue from Figs. 1e, 3b) were captured with an Andor Dragonfly spinning disk confocal microscope running Fusion software with a 60× lens with a z step of 1 μm.

Image processing

Captured epifluorescence images were exported as large (14k × 11k pixel) multichannel tiff files (Extended Data Fig. 1e), and subsequently imported into our Connection Lens image processing software. After an initial atlas matching step, where each section was manually matched to its corresponding level of the ARA51, images containing the pallidum and nigra were registered to the mouse brain atlas (Extended Data Fig. 1f). This work exclusively references levels of the ARA (for example, ARA 81 refers to level 81 of the ARA, Bregma = −2.78 mm). Our 1-in-4 series of 50-μm tissue sections gives us a view of the brain that is every other level in the ARA, itself based on a 1-in-2 series of 100-μm sections. Therefore, we registered our tissue sections onto every other atlas level of the pallidum and nigra, and for this purpose we chose the even levels of the pallidum (that is, ARA 58–68, even levels) and odd levels of the nigra (ARA 81–91, odd levels). All tissue sections were registered to their closest ARA level (that is, a section containing GPe at ARA 61 was mapped onto either ARA 60 or ARA 62). The determination of which level a given section was assigned to was made by an experienced neuroanatomist. The process of registering to a standardized set of atlas levels provided a uniform dataset that was amenable to computational analysis. After registration, Connection Lens guided users through an interactive segmentation step, creating a binary output image of axons (black) and background (white). Since the images were previously registered, the resulting segmentations could be accurately projected onto the atlas frame (Extended Data Fig. 1g). Finally, Connection Lens applied an overlap algorithm to quantify the segmented axonal labelling by region (GPe and SNr). Each level of the GPe and SNr in the atlas depicts a single unitary region, yet we knew from the labelling patterns that the striatofugal axonal termination patterns project to a sub-region of each nucleus. We subsequently applied the grid quantification method used previously in our corticostriatal analysis5, subdividing each nucleus at each atlas level into a square grid space (105 × 105 pixels per box, equivalent to 63 μm2), and quantified the axons per grid box (Extended Data Fig. 1h). Any injection that contributed less than 1,500 pixels to a given level was excluded from the community analysis for that level. A small number of cases had labelling that slightly exceeded the 1,500-px threshold for a given level but the labelling was judged too diffuse to be a meaningful terminal field, and were similarly excluded. The surviving grid box data were subjected to network analysis to determine striatonigral and striatopallidal community structure (Extended Data Fig. 1i; see next section). The derived communities were visualized by recolouring the grid boxes according to community identity (Extended Data Fig. 1j). And finally, projection maps of striatofugal axon terminals were created by aggregating the registered segmented axonal images onto maps of SNr and GPe (Extended Data Fig. 1k).

Network analysis

The network structure of the dataset was assessed with the Louvain community detection algorithm52, obtained from the Brain Connectivity Toolbox (, and executed in Python. Louvain is a greedy, non-deterministic algorithm, with multiple runs producing differing returns of maximized modularity. Importantly, a gamma variable modulates the number of communities detected in a dataset, with smaller gamma values leading to low-dimension network structures (fewer nodes, larger communities) and larger gamma values leading to high-dimension network structures (more nodes, smaller communities). While the default gamma value of 1 is used commonly (and useful for communicating a frame of reference, as it is a de facto standard), choosing the optimal gamma value is a non-trivial problem.

In an attempt to obtain the most descriptive network partition among this parameterization and variability, we performed a survey of community structures across different gamma values. The Louvain algorithm was run 100 times per each gamma value, over a gamma range of 0 to 2 at 0.05 increments for every nucleus level. A consensus community structure (conceptually, the ‘average’ community structure53) was calculated from each batch of 100 runs at every gamma. The resultant 41 consensus community structures were compiled into a frequency histogram, to determine the most common community structures to arise over the range of gammas for a nucleus level. The true network structure of the underlying data should act as a ‘magnet’ or attractor for a stable community structure over multiple gammas; therefore an accurately characterized community structure should be obtained across multiple gamma values54, represented by peaks in the graph (Extended Data Fig. 12l–m).

We applied different analytical parameters to the direct (striatonigral) and indirect (striatopallidal) pathway data. The domains in the SNr and GPe exist in three dimensions, and for the SNr in particular likely extend across multiple levels of the nucleus. Our goal was to balance across-level similarity with high dimension domain structure, as the input data (that is, the striatofugal terminal fields) is verified higher dimensional. Moreover, we sought to parse the pallidum and nigra into more than the three classically recognized output channels.

For the direct pathway, most of the striatal domains send projections in longitudinal columns along the entire rostrocaudal extent of the SNr. This suggests that there is a relatively high degree of consistency in the community structures across adjacent levels of the SNr. However, caudally the SNr becomes physically smaller in cross-sectional area and the axonal terminal fields exhibit a higher degree of convergence than in rostral levels (Extended Data Figs. 4a, 12b, c; also see ref. 10). We quantitatively characterized this convergence to describe the degree of integration, and used these data to inform our selection of gamma values. Using the quantified grid box data, we created frequency distributions of the boxes categorized by how many different striatal inputs they received (only boxes receiving input were included in this analysis) (Extended Data Fig. 12a). We applied a minimum threshold such that inputs contributing less than 5% overlap to a box (551 pixels in a 105 pixel2 box) were excluded from that boxes’ tally of inputs. When graphed together, the histograms show that the rostral SNr levels 81–85 have negatively skewed distributions, while the caudal levels 87–91 have more platykurtic distributions with relatively fatter tails in the 10–15 inputs per box range (Extended Data Fig. 12b), a trait that becomes more apparent when the rostral levels and caudal levels are averaged (Extended Data Fig. 12c). There is no significant correlation of mode (peak value) of each histogram with rostrocaudal level as assessed by linear regression (r = 0.4252, P = 0.4006; Extended Data Fig. 12e), and there is no significant difference in mode between the rostral (mean ± s.e.m.: 7.0 ± 0.577) and caudal (7.3 ± 1.856) groups (P = 0.8796, t = 0.1715, d.f. = 4; two-tailed Welch’s-corrected t-test), verifying that the graphs have similar central tendencies (Extended Data Fig. 12c). Two-way ANOVA of the rostral and caudal groups (SNr level × amount of integration) finds a significant main effect of integration (P < 0.0001, d.f. = 14, F = 12.46) and a significant interaction of integration and SNr level (P = 0.0219, d.f. = 14, F = 2.136). Post hoc Bonferroni tests reveal the caudal group has a significantly greater proportion of boxes receiving 11 inputs (P < 0.0033, t = 3.385) and a nearly significant difference for 10-input boxes (0.05 > P > 0.025, familywise-adjusted α = 0.0033, t = 2.278). Thus, the caudal three levels of the SNr have a significantly greater proportion of their area devoted to high integration (ten or more inputs per box) (Extended Data Fig. 12d–f), indicating that the rostral and caudal SNr should be analysed with different parameters since there is probably a different, more integrative domain structure caudally.

Since the integration analysis indicates that the rostral and caudal halves of the SNr form two groups, we selected community structures that were most common through the rostral and caudal groups (Extended Data Fig. 12l). By stacking the histograms of the component levels, the peaks in the graphs reveal the most stable community structures common to all constituent levels. For the rostral SNr group, there were two clear peaks, for a six-domain and a ten-domain structure (Extended Data Fig. 12l, SNr rostral group). We have selected to present the ten-domain structure, because we sought to subdivide the SNr into as fine a coherent partition as possible, although the 6-domain structure may also be a valid way to interpret the striatonigral data as well, since it is possible there is a nested multi-scale network architecture to the striatonigral pathway as with the corticostriatal pathway. For the caudal group, there was a clear peak for the seven-domain structure (Extended Data Fig. 12l, SNr caudal group). All gamma values returning the chosen domain structure for each nucleus-level were pooled and the consensus community structure of that pool was the final network output.

Most community detection algorithms, including Louvain, impose a unitary structure on an information network such that each node belongs to one and only one community. This simplified network structure is easier to interpret, but may not be reflective of the complex nature of real-world networks, wherein a single node may participate in multiple subnetwork structures, as is probably true in brain networks. Thus, after determining the community structure for each representative level of the SNr, we manually joined together communities on adjacent levels based on continuity and similarity of inputs (Supplementary Table 1). For example, the striatal inputs to the medial region of the SNr are similar enough when comparing adjacent levels (mean Jaccard index comparing all adjacent levels = 0.61) that we determine them to form one continuous domain, the medial domain (Supplementary Table 1). The average of mean Jaccard indices for all nigral domains that span more than one level (n = 12) is 0.61 ± 0.28 (mean ± s.d.). Furthermore, all cross-level domains also have one or more inputs that consistently span the entire domain. It should be pointed out that the nigral terminations of a particular CP domain frequently shift in their mediolateral or dorsoventral position in the nigra at different rostrocaudal levels; consequently, even after our efforts to join together similar communities into unified nigral domains, many CP domains contribute projections to multiple SNr domains, such as the CPr.imd providing input to the SNr.m at levels 83, 85, 87 and 91 and to the SNr.v at level 89.

Each indirect pathway input tends to densely innervate just a few levels of the GPe. GPe boxes integrate at most 9 inputs, much more restricted than the SNr (up to 15), and mean mode of the GPe (3.33 ± 0.558 (mean ± s.e.m.); n = 6) is significantly smaller than mean mode of SNr (7.17 ± 0.872; n = 6), indicating less integration and more segregated relaying of striatal activity through the indirect pathway (P = 0.0060, t = 3.702, d.f. = 8; two-tailed Welch’s t-test). The frequency distributions of inputs per box show similarly shaped histograms across GPe levels, with decreasing mode value from rostral to caudal (Extended Data Fig. 12g, h). Linear regression of the histogram modes shows a significant correlation with GPe level (r = 0.8607, P = 0.0278), with mode decreasing towards caudal levels (Extended Data Fig. 12j), indicating that caudal GPe integrates progressively fewer inputs per box (Extended Data Fig. 12i–k). Given that the graphs vary continuously along the rostrocaudal axis, and the lower degree of continuity of striatopallidal projections along that axis, we evaluated each level of the GPe independently (Extended Data Fig. 12m). A far smaller proportion of pallidal domains exhibit cross-level structure than in the nigra (36% versus 86%). The average of mean Jaccard indices for the pallidal domains that span more than one level (n = 10) is 0.64 ± 0.21 (mean ± s.d.). No striatal domain innervates the entire rostrocaudal axis of the GPe (Supplementary Table 2).

Whole-brain 3D imaging and reconstruction of neuronal morphology

Intact brain (n = 1) was SHIELD-cleared as described55, placed in refractive index-matching solution (EasyIndex, LifeCanvas), and imaged on a LifeCanvas light-sheet microscope running SmartSPIM Acquisition software at 4× and 10× magnification. These images as well as z-stack images captured with the DragonFly confocal microscope were viewed within Aivia reconstruction software (v8.8.2. DRVision) and neurons were manually reconstructed. Geometric processing of the reconstructions was performed using the Quantitative Imaging Toolkit (, and morphometric data were obtained from the reconstructions with NeuTube (v1.0z). Descriptive statistics of the morphological features of these neurons were generated by NeuTube.

Anterograde transsynaptic tracing

This technique leverages the fact that when AAV1 is injected at sufficiently high concentration into a neuronal population, viral particles will travel down the axons and be released from the synaptic terminals where they can infect postsynaptic neurons. Detailed methodology is as described56,57. In brief, anaesthetized mice were iontophoretically injected with Cre-dependent AAV-FLEX-RFP in the target nucleus (except the demonstration of the pathway, which used pressure injection of 50 nl into medial SNr at atlas levels 81–85), and then pressure injected (20–80 nl) with AAV-Cre in an upstream nucleus. The AAV-Cre is transported anterogradely down the axons and is released from the terminals, where it transfects postsynaptic cells that have been infected with high concentrations of Cre-dependent AAV-FLEX-RFP. The scant Cre expression is sufficient to unlock strong fluorophore expression in the downstream neurons. After a three-week post-operative recovery, animals were pentobarbital-anaesthetized and perfused as above. The Cre injection site was verified by staining with mouse anti-Cre recombinase monoclonal antibody (Millipore Sigma, MAB3120) and donkey anti-mouse AlexaFluor 647 (Jackson ImmunoResearch, 715-605-150). Additionally, two Ai14 mice (Jackson Laboratories 007908) with endogenous tdTomato expression following Cre recombination were injected with AAV-Cre in mouth primary motor cortex MOp-m/i to demonstrate the corticonigral pathway (Extended Data Fig. 20a).

Experimental designs for CRACM with anterograde transsynaptic tracing

To examine convergence and parallelism in the striatonigral and striatopallidal pathways, we used CRACM58. We injected AAV1-Cre (AAV1-hSyn-Cre-WPRE; 20–80 nl, pressure) into the striatal domain CPc.d.dl, which is taken up at the injection site, transported down the axon, crosses the synapse with low efficiency and infects postsynaptic neurons in GPe.24, 26 and 32 and SNr.v. To label the Cre-expressing postsynaptic neurons with RFP, we injected those pallidal and nigral target domains with AAV-FLEX-RFP (AAV1-CAG-Flex-tdTomato-WPRE; 2–3 min, iontophoresis). We next injected AAV-ChR2 (AAV1-hSyn-ChR2-EYFP-WPRE; 2–3 min, iontophoresis) into striatal somatomotor trunk domain CPi.dl.d (tr), labelling its axons in GPe.6 and 25 and SNr.v with YFP and rendering them optically excitable. Thus in the pallidum the RFP neurons should not respond to stimulation of the ChR2 axons since they do not overlap, while in the nigra the RFP neurons should respond to stimulation of the ChR2 axons since they colocalize (Fig. 2c; n = 3).

To examine whether the direct (striatonigral) and indirect (striato–pallido–nigral) pathways re-converge onto individual SNr neurons, we injected AAV-Cre into the striatal domain CPi.dl.d (tr), which transsynaptically infected postsynaptic neurons of the indirect pathway in GPe.6 and 25 and of the direct pathway in SNr.v, causing those cells to express Cre (n = 7). In the same surgery, AAV-DIO-ChR2 (2–3 min, iontophoresis) was injected into the region of GPe.6 and 25 and AAV-FLEX-RFP was injected into SNr.v, meaning only the Cre-expressing neurons postsynaptic to terminals originating from CPi.dl.d (tr) expressed ChR2 (YFP) and RFP, respectively. Three weeks later, acute slices of nigra were prepared, and RFP-expressing neurons of the SNr (postsynaptic to the direct pathway) were patched and recorded during channelrhodopsin stimulation (the indirect pathway). If there is re-convergence of homotypic direct and indirect pathways, then the RFP-expressing neurons in SNr should respond to optical stimulation of the pallidonigral pathway (Extended Data Fig. 8h).

Two experiments were conducted to examine whether the thalamocortical axons of a cortico–basal ganglia–thalamic subnetwork synapse upon the corticostriatal neurons that feed into that subnetwork. These experiments targeted the oro-brachial subnetwork (n = 2) and the associative subnetwork (n = 4). For the oro-brachial subnetwork experiment, AAV-ChR2 was iontophoretically injected into the PF.m thalamic domain and retrobeads (30 nl, pressure) were injected into the CPi.vm.v (m/i) domain. At least three weeks later, slices containing the mouth primary cortical region MOp-m/i were collected for electrophysiological recording as described below (Fig. 3c). For the associative subnetwork experiment, AAV-ChR2 was iontophoretically injected into the PF.a thalamic domain and retrobeads were injected into the domain. Slices containing the anterior cingulate area (ACA) were collected for electrophysiological recording (Extended Data Fig. 17c).

Electrophysiological recording

At least two weeks following channelrhodopsin and tracer injections, acute brain slices were prepared for recording. Following anaesthesia, the animal was decapitated and the brain was quickly removed and immersed in ice-cold dissection buffer (cortical recording experiments (in mM): 60 NaCl, 3 KCl, 1.25 NaH2PO4, 25 NaHCO3, 115 sucrose, 10 glucose, 7 MgCl2, 0.5 CaCl2; saturated with 95% O2 and 5% CO2; pH = 7.4; for basal ganglia recording experiments: 208 sucrose, 2.5 KCl, 1.25 NaH2PO4, 26 NaHCO3, 1.3 MgCl2, 8 MgSO4 and 10 glucose). Brain slices of 200–300 μm thickness were cut in coronal plane using a vibrating microtome (Leica VT1000S). Slices were allowed to recover for 30 min in a submersion chamber filled with warmed (35 °C) ACSF containing (in mM) 130 NaCl, 3 KCl, 1.25 NaH2PO4, 26 NaHCO3, 2 CaCl2, 2 MgCl2 and 10 glucose, oxygenated with 95% O2 and 5% CO2, pH 7.2–7.4, 290–310 mOsm, and then cooled gradually to room temperature until recording. The presence of retrobead or RFP and ChR2 (YFP) labelling was examined under green and blue fluorescence excitation in the slices before recording. Patch pipettes (Kimax) with ~6–7 MΩ impedance were used for whole-cell recordings. Recording pipettes contained: 130 mM potassium-gluconate, 4 mM KCl, 2 mM NaCl, 10 mM HEPES, 0.2 mM EGTA, 4 mM ATP, 0.3 mM GTP, 14 mM phosphocreatine (pH 7.25; 290 mOsm) and biocytin (0.2%). For the thalamo–cortico–striatal experiments (Fig. 3c, Extended Data Fig. 17c) 0.1–1 μM tetrodotoxin and 0.1–1 mM 4-aminopyridine was added to the external solution for isolation and recording of monosynaptic responses to blue light stimulation. For the oro-brachial subnetwork experiment, only layer 5 neurons were recorded. Signals were recorded from red-labelled neurons with a MultiClamp 700B amplifier (Molecular Devices) running pClamp software under voltage clamp mode at a holding voltage of –70 mV for excitatory currents and 0 or +10 mV for inhibitory currents, filtered at 2 kHz and sampled at 10 kHz. Blue light (470 nm) stimulus was delivered in a 0.5–5 ms pulse, at 1–3 mW power, for 3–5 trials, delivered via a mercury arc lamp gated with an electronic shutter. Signals were analysed using Clampfit. For responding neurons, peak responses were averaged across trials, and for non-responding neurons, peak recorded amplitudes within the 1-s window following stimulation were averaged across trials.

Fluorescence micro-optical sectioning tomography

All fMOST experiments were conducted in accordance with the Institutional Animal Ethics Committee of Huazhong University of Science and Technology. For sparse labelling of striatal neurons, we created a single pAAV co-package of DNA cassettes of CMV-Cre and Cre-dependent EF1a-DIO-GFP at a ratio of 1:1,000,000, respectively, so that the final viral admixture contained one virus with both cassettes for every 1,000,000 viruses that contained only the EF1a.DIO.GFP cassette (total viral concentration 8 × 1012 genome copies per ml, from BrainVTA). This admixture was pressure injected (100 nl) into (M–L, A–P, D–V: 0.14, −1.3, −2.6), allowing high viral load for the fluorophore gene with low frequency of co-expression of Cre, resulting in sparse yet bright GFP expression. A detailed protocol has been previously described59. After 5 weeks, mice (n = 7) were anaesthetized, perfused with 0.01 M PBS and 4% paraformaldehyde, and brains were post-fixed overnight. For whole-brain imaging, brains were rinsed in 0.01M PBS solution and dehydrated in a graded ethanol series (50, 70 and 95% ethanol), submerged in gradient series of Lowicryl HM20 resin, and polymerized at a gradient temperature in a vacuum oven.

The resin-embedded whole-brain samples were imaged using fMOST, a three-dimensional dual-wavelength microscope-microtome combination instrument (see ref. 60 for a detailed description). Block imaging mode was used to slice and scan layer by layer through the whole sample in the coronal plane. GFP-labelled neurons and propidium iodide-stained cytoarchitecture were acquired at a voxel resolution of 0.32 × 0.32 × 1 μm3. The raw images were first preprocessed for intensity correction, and then the image sequence was converted to TDat, an efficient 3D file format for large volume images, to facilitate the computing of terabyte- and petabyte-scale brain-wide datasets61. We employed GTree for semi-automated, manually assisted reconstruction of neuronal morphology in 3D62. Subsequently, neuronal morphological data were mapped to the Allen CCFv3 using BrainsMapi, a robust image registration interface for large volume brain images63. Because the contours of brain regions on the propidium iodide-stained images can be more clearly identified, this greatly reduces the difficulty of accurate registration.

D1 and A2A cell-type specific tracing

For labelling D1 and D2 dopamine receptor-expressing medium spiny neurons (MSNs), adult Adora2a-Cre (GENSAT 036158-UCD, for labelling D2-MSNs) and Drd1a-Cre (GENSAT 017264-UCD, for D1-MSNs) congenic mice on a C57BL6/J background were obtained from GENSAT and backcrossed with wild type C57BL mice (Jackson Laboratories) for several generations. Surgeries were performed between 10–12 weeks of age. Subjects were anaesthetized with 2% isoflurane in oxygen. Buprenorphine SR (1 mg kg−1) was administered at the beginning of the surgery as an analgesic. Glass micropipettes (10–30 µm diameter tip) filled with AAV-DIO-EGFP (AAVDJ-hSyn-DIO-EGFP-WPRE-bGH, B.K.L. laboratory) were lowered into the target region and delivered a localized injection by pressure (50–150 nl volume) at a rate of 100 nl min−1. After 3 weeks, animals were deeply anaesthetized and perfused. Tissue sections were stained with DAPI and imaged on an Olympus VS120 epifluorescence microscope with a 10× objective lens. All procedures to maintain and use mice were approved by the Institutional Animal Care and Use Committee at the University of California, San Diego.

Statistical analysis

All standard statistical analyses were performed with GraphPad Prism v4.0c for Macintosh. Sample sizes were not predetermined statistically, but are consistent with experiments of their type. Randomization was not necessary because no independent group comparisons were made.

Reporting summary

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