A Selective Projection from the Subthalamic Nucleus to Parvalbumin-Expressing Interneurons of the Striatum

Visual Abstract


Introduction
Influential models of the functional organization of basal ganglia (BG) circuits and their thalamic and cortical partners, such as the "direct/indirect pathways" scheme (DeLong, 1990;Y. Smith et al., 1998) and a modification accentuating the "hyperdirect pathway" (Nambu et al., 2002), consider the striatum and subthalamic nucleus (STN) to be the primary input nuclei of the BG. The same models assume that striatum and STN are not monosynaptically connected. Accordingly, the primary targets of striatal and STN projecting axons are typically listed as the external globus pallidus (GPe) and BG output nuclei, i.e., internal globus pallidus (or entopeduncular nucleus in rodents) and substantia nigra pars reticulata (Y. Smith et al., 1998;Emmi et al., 2020). That striatum is not commonly deemed a target of STN outputs is at variance with anatomic evidence. Indeed, studies of neuronal populations using "classical" tracers support the existence of moderate subthalamostriatal projections in rats, cats and monkeys (Beckstead, 1983;Kita and Kitai, 1987;Nakano et al., 1990;Y. Smith et al., 1990), although data can be obfuscated by the bimodal (mixed anterograde and retrograde) transport of these tracers and their potential uptake by fibers of passage (Y. Smith et al., 1998). On the other hand, single-neuron tracing indicates that most rat STN neurons innervate striatum, and sometimes to a greater degree than innervation of BG output nuclei (Koshimizu et al., 2013). Furthermore, recent studies employing sophisticated viral vector-mediated cell labeling in mice suggest that STN neuron axons can form synapses with striatal neurons (Wall et al., 2013;Guo et al., 2015;J.B. Smith et al., 2016;Klug et al., 2018;Choi et al., 2019). Importantly, there has been no physiological characterization of the incidence and strength of subthalamostriatal neurotransmission.
Understanding the organization and impact of the subthalamostriatal projection is challenged by the diversity of cell types within striatum. Approximately 90-95% of striatal neurons can be classified as spiny projection neurons (SPNs), which may then be further divided into, for example, populations giving rise to the direct pathway or indirect pathway (Gerfen and Surmeier, 2011). The remaining ;5-10% is made up of interneurons that are either cholinergic or GABAergic, the latter of which also exhibit substantial heterogeneity in form and function (Tepper et al., 2010(Tepper et al., , 2018. Stemming from this diversity, there is much scope for axons to engage one or more cell types in a biased manner. Indeed, the connections between striatal neurons are often selective for cell type, supporting a complex mix of reciprocal and nonreciprocal interactions across the wider microcircuit. Some of the extrinsic inputs to striatum, arising from both cortical and subcortical areas, also exhibit selectivity for target cells (Silberberg and Bolam, 2015;Tepper et al., 2018;. Currently, it is unclear how the STN innervates and influences the striatal microcircuit, including the extent to which the subthalamostriatal projection is selective for target neuron type(s). Resolving these issues requires definitions of the structural substrates and physiological properties of this projection, while accounting for cellular diversity in striatum. To address this, we conducted monosynaptic retrograde tracing from distinct populations of striatal neurons in transgenic mice, quantifying the connectivity from STN neurons to SPNs, GABAergic interneurons and cholinergic interneurons. To complement these anatomic studies and gain insights into neurotransmission, we used ex vivo electrophysiology and optogenetics to interrogate the responses of a corresponding range of striatal cell types to activation of subthalamostriatal axons. Our results collectively reveal that the subthalamostriatal projection is highly selective, providing relatively rich and efficacious glutamatergic inputs to GABAergic parvalbumin (PV)-expressing interneurons.

Materials and Methods
Experimental procedures were performed on mice and were conducted either at the University of Oxford in accordance with the Animals (Scientific Procedures) Act, 1986 (United Kingdom), or at Rutgers University with the approval of the Institutional Animal Care and Use Committee (IACUC) in accordance with public health service (PHS) policy on humane care and use of laboratory animals. All experimental work adhered to Society for Neuroscience Policies on the Use of Animals in Neuroscience Research. All mice were group housed, had ad libitum access to food and water, and were maintained on a 12/12 h light/dark cycle. Zeitgeber time (ZT; with ZT0 = lights on in the animal facility) of the experiments was ZT1-ZT9.

Animals and related procedures for anterograde labeling and monosynaptic retrograde labeling of neurons in vivo
Adult male and female mice, aged 2.5-8 months, were used for these experiments. Eight lines of transgenic mice were used; all were bred to a C57Bl6/J background, and only mice heterozygous/hemizygous for the transgene(s) were used in experiments. To target glutamatergic neurons of the STN for anterograde labeling, we used a VGluT2-Cre mouse line (B6J.129S6 (FVB)-Slc17a6 tm2(cre)Lowl /MwarJ; The Jackson Laboratory; RRID:IMSR_JAX:028863). To retrogradely label STN neurons innervating striatal neurons as a whole, we used a "double transgenic" line produced by crossing VGAT-Cre mice (Slc32a1 tm2(cre)Lowl /J; The Jackson Laboratory; RRID:IMSR_JAX:016962) with ChAT-Cre mice (B6;129S6-Chat tm2(cre)Lowl /J; The Jackson Laboratory; RRID:IMSR_JAX: 006410). To retrogradely label STN neurons innervating more restricted populations of neurons in striatum, we used the following lines: Drd1a-Cre mice (B6.FVB(Cg)-Tg For stereotaxic intracerebral injections of adeno-associated virus (AAV) and rhabdovirus vectors in mice, general anesthesia was induced and maintained with isoflurane (1.0-3.0% v/v in O 2 ). Animals received perioperative analgesic (buprenorphine, 0.1 mg/kg, s.c.; Ceva) and were placed in a stereotaxic frame (Kopf Instruments). Wound margins were first infiltrated with local anesthetic [0.5% w/v bupivacaine (Marcaine); Aspen]. Body temperature was maintained at ;37°C by a homeothermic heating device (Harvard Apparatus). To anterogradely label glutamatergic STN neurons with green fluorescent protein (GFP), we used a glass micropipette (internal tip diameter of 15-22 mm) to unilaterally or bilaterally inject ;33 nl (per site) of a AAV2-CAG-FLEX-GFP vector [titer: 3.7 Â 10 12 vg/ml; University of North Carolina (UNC) Vector Core] into the STN of VGluT2-Cre mice, using the following stereotaxic coordinates: 1.90 mm posterior of Bregma, 1.75 mm lateral of Bregma, and 4.75 mm ventral to the brain surface. To minimize reflux, the micropipette was left in place for ;20 min after the injection. Mice were maintained for 28-35 d after surgery to allow for neuron transduction and labeling with GFP; they were then humanely killed and perfused (see below). To carry out monosynaptic retrograde tracing from neurons in the dorsal striatum, we made sequential use of a single Cre-dependent "helper virus" [AAV5-DIO-TVA V5 -RG; bicistronically expressing TVA receptor fused to a V5 tag, and the rabies glycoprotein (RG); titer: 2.2 Â 10 13 vg/ml; Ährlund-Richter et al., 2019] and a "modified rabies virus" ([EnvA]-SADDG-EGFP; pseudotyped with EnvA, RG deleted, and expressing enhanced GFP; Fürth et al., 2018;Ährlund-Richter et al., 2019). In a first step, we used a glass micropipette to unilaterally inject 60-120 nl of helper virus into the central aspects of the dorsal striatum of Cre-expressing mice, using the following coordinates (caudal approach at an angle of 20°to vertical): 0.00 mm anterior of Bregma, 2.20 mm lateral of Bregma, and 2.70 mm ventral to the brain surface. To minimize reflux, the micropipette was left in place for ;10 min after the injection. Allowing 21 d for neuron transduction, Cremediated recombination, and the generation of "starter" neurons expressing all components necessary for retrograde labeling of their presynaptic partners, we then as a second step injected 60-120 nl of modified rabies virus into the same striatal locations, using the following coordinates (vertical, no angle): 1.00 mm anterior of Bregma, 2.20 mm lateral of Bregma, and 2.50 mm ventral to the brain surface. To minimize reflux, the micropipette was left in place for ;10 min after the injection. Mice were maintained for 7 d after surgery to allow for starter neuron infection and retrograde trans-synaptic labeling of neurons providing monosynaptic inputs to starters. Thereafter, mice were killed with pentobarbitone (1.5 g/kg, i.p.; Animalcare) and transcardially perfused with 20-50 ml of 0.05 M PBS, pH 7.4 (PBS), followed by 30-100 ml 4% w/v PFA in 0.1 M phosphate buffer, pH 7.4 (PB). Brains were removed and left overnight in fixative at 4°C before sectioning.
As a control for the Cre and TVA dependence of neuronal labeling in in monosynaptic retrograde tracing experiments, we injected helper virus and/or modified rabies virus into the striata of adult wild-type C57Bl6/J mice (n = 2). As previously reported for the vectors we use here (Ährlund-Richter et al., 2019), these injections resulted in negligible numbers of starter neurons and retrogradely-labeled neurons (data not shown).
A version of design-based stereology, the "modified optical fractionator," was used to generate unbiased cell counts and determine the proportions of a given population of neurons that expressed certain combinations of molecular markers (Abdi et al., 2015;Dodson et al., 2015;Garas et al., 2016). All stereology, imaging for stereology, and cell counting, was performed using Stereo Investigator software (v. 2019.1.4, MBF Bioscience, RRID:SCR_002526). Acquisition of tissue images for stereological sampling was conducted on an AxioImager.M2 microscope (Zeiss) equipped with an ORCA Flash-4.0 LT digital CMOS camera (Hamamatsu), an Apotome.2 (Zeiss), and a Colibri 7 LED light source (type R[G/Y]B-UV, Zeiss). Appropriate sets of filter cubes were used to image the fluorescence channels: AMCA (excitation 299-392 nm, beamsplitter 395 nm, emission 420-470 nm); Alexa Fluor 488 (excitation 450-490 nm, beamsplitter 495 nm, emission 500-550 nm); Cy3 (excitation 532-558 nm, beamsplitter 570 nm, emission 570-640 nm); and Cy5/DyLight 649 (excitation 625-655 nm, beamsplitter 660 nm, emission 665-715 nm). Images of each of the channels were taken sequentially and separately to negate possible crosstalk of signal across channels. In order to quantify retrogradely-labeled subthalamostriatal neurons, we first defined (using a 10Â objective lens; 0.45 NA; Plan-Apochromat, Zeiss) the borders of the STN according to the expression of FoxP2 (Abdi et al., 2015); the full extent of the STN was imaged for each series examined in each mouse. To quantify striatal starter neurons, we first defined (using a 10Â objective) the outer boundaries of regions within striatum that contained neurons immunoreactive for V5, an indicator of neurons transduced with the helper virus (and thus, potential starter cells). After delineating the borders of STN and the boundaries of striatal regions containing starter neurons, images for stereological sampling were acquired using the optical fractionator workflow in Stereo Investigator, employing a 2-mm-thick "guard zone," and an unbiased 2D counting frame and grid frame of 600 Â 600 mm (i.e., 100% of the region in the X, Y plane was sampled). Z-stacked images across a 10-mm-thick "optical disector" were acquired using a 20Â objective lens (0.8 NA; Plan-Apochromat), and images or "optical sections" were taken in 1-mm steps to ensure no loss of signal in the z-axis. Captured images were then analyzed offline. A neuron was only counted once through the series of optical sections when its nucleus came into sharp focus within the disector; neurons with nuclei already in focus in the top optical section of the disector were ignored. The use of stereology, and this optical disector probe in particular, ensured that we could generate robust and unbiased cell counts in a timely manner. For a given molecular marker, X, we designate positive immunoreactivity (confirmed expression) as X1, and undetectable immunoreactivity (no expression) as XÀ. A neuron was classified as not expressing the tested molecular marker only when positive immunoreactivity could be observed in other cells on the same optical section as the tested neuron. Striatal starter neurons were defined by their co-expression of immunoreactivity for V5 and rabies-encoded GFP. To ensure a high level of precision in the cell counts, data were only included from individual mice when the Coefficient of Error (CE; using the Gundersen method) was 0.1 with a smoothness factor of m = 1 (West et al., 1991;Gundersen et al., 1999). The CE provides an estimate of sampling precision, which is independent of biological variance. As the value approaches zero, the uncertainty in the estimate precision reduces. The number of sections counted per mouse was thus dependent on variability; sections/series were added to the analysis until the CE 0.1 for each animal. Images for figures were acquired with a confocal microscope (LSM880, Zeiss). All image adjustments were linear and applied to every pixel.
For stereotaxic intracerebral injections of AAV vectors in mice, general anesthesia was induced and maintained with isoflurane (1.0-3.0% v/v in O 2 ). Animals were placed in a stereotaxic frame (Kopf Instruments). Wound margins were first infiltrated with local anesthetic (0.25% w/v bupivacaine, with 1:200,000 epinephrine [Sensorcaine], Hospira). Body temperature was maintained at ;37°C by a homeothermic heating device. To express channelr-hodopsin2 (ChR2) in STN neurons, we used a 1 ml microsyringe (7000 Series, Hamilton) to unilaterally inject ;45 nl of an AAV vector (all from UNC Vector Core) into the STN of mice, using the following stereotaxic coordinates: 1.88 mm posterior of Bregma, 1.68 mm lateral of Bregma, and 4.50 mm ventral to the brain surface. To minimize reflux, the microsyringe needle was left in place for ;10 min after the injection. For experiments using wildtype or PV-tdTomato mice, we injected an AAV5-CAMKIIa-hChR2(H134R)-EYFP vector (titer: 10.6 Â 10 12 vg/ml) into the STN. For experiments using NPY-GFP mice, we injected an AAV5-CAMKIIa-hChR2(H134R)-mCherry vector (titer: 2.7 Â 10 12 vg/ml) into the STN. For experiments using ChAT-Cre mice, we injected an AAV5-CAMKIIa-hChR2 (H134R)-EYFP vector into the STN. To visualize striatal cholinergic interneurons in the same ChAT-Cre mice, we unilaterally injected (ipsilateral to transduced STN) 600 nl of an AAV5-CAG-FLEX-tdTomato vector (titer: 5.0 Â 10 12 vg/ ml; UNC Vector Core) into the dorsal striatum, using the following stereotaxic coordinates (200 nl per site at three depths along the dorsal-ventral axis): 0.70 mm anterior of Bregma, 1.80 mm lateral of Bregma, and 3.20, 2.60, and 2.20 mm ventral to the brain surface. To minimize reflux, the microsyringe needle was left in place for ;10 min after the most dorsal injection. Animals received postoperative analgesic (buprenorphine SR-LAB, 0.1 mg/kg, s.c.; ZooPharm) and were maintained for 42-56 d after surgery to allow for neuron transduction; the mice were then deeply anesthetized and perfused for ex vivo recordings and anatomic verification of ChR2 expression.
For electrophysiological testing, single slices of striatum were transferred to a recording chamber and perfused continuously with oxygenated ACSF maintained at 32-34°C. Slices were visualized using infrared gradient contrast video on a microscope (BX50WI, Olympus) equipped with a 40Â water-immersion objective lens (LUMPIanFL/ IR, Olympus), an iXon EMCCD camera (DU-885K-CS0, Andor), a halogen light source (TH3 power supply, Olympus), and imaging software (Solis v 4.4.0, Andor). A mercury lamp (BH2-RFL-T3 power supply, Olympus) coupled to appropriate filter cubes (U-M41001, Olympus; 49 008, Chroma Technology) was used to visualize native fluorescence (of GFP, mCherry and/or tdTomato) in the slices, which in turn ensured accurate targeting of desired cell types in areas of striatum that were traversed by axons expressing fluorescent fusion proteins of ChR2. Somatic patch-clamp recordings of some individual striatal neurons were made in a whole-cell configuration (in voltage-clamp mode at a holding voltage (V h ) of À70 mV, and/or in current-clamp mode) using glass pipettes that were filled with a K-gluconate-based solution composed of (in mM): 130 K-gluconate, 10 KCl, 2 MgCl 2 , 10 HEPES, 4 Na 2 ATP, 0.4 Na 3 GTP. Either 0.2% w/v biocytin (B4261, Sigma) or 1.5 ml/ml Alexa Fluor 594 dye (A10438, ThermoFisher Scientific) was added to the pipette solution to facilitate, in particular, the post hoc anatomic identification of SPNs. The osmolarity and pH of this Kgluconate pipette solution were 290-295 mOsm and 7.3, respectively. Whole-cell recordings from other striatal neurons were made in voltage-clamp mode (V h À70 mV) using glass pipettes that were filled with a CsCl-based solution composed of (in mM): 125 CsCl, 2 MgCl 2 , 10 HEPES, 4 Na 2 ATP, 0.4 Na 3 GTP, plus 1.5 ml/ml Alexa Fluor 594 dye. The osmolarity and pH of this CsCl pipette solution were 290-295 mOsm and 7.3, respectively. Irrespective of internal solution, pipettes typically exhibited a DC impedance of 3-5 MV measured in the recording chamber. Somatic patch-clamp recordings were obtained using a Multiclamp 700B amplifier (Molecular Devices, RRID:SCR_018455) and ITC-1600 digitizer (Instrutech), with AxoGraph software (v. 1.7.4, RRID:SCR_ 014284) used for data acquisition and analysis. Electrode signals were low-pass filtered (Bessel filter) at 1 or 10 kHz for voltage-clamp or current-clamp recordings, respectively, and sampled at 20 kHz. Most SPNs were recorded in wild-type mice, and were classified as SPNs by their characteristic morphologic properties and/or intrinsic electrophysiological properties (Gertler et al., 2008). When patched with pipettes containing the K-gluconatebased solution, SPNs were filled with either biocytin or Alexa Fluor 594 dye for post hoc anatomic analyses and verification of cell type. Current-clamp recordings of spontaneous activity and responses to somatic current injections were also useful in confirming SPN identity. All SPNs patched with CsCl-based pipette solution were filled with Alexa Fluor dye for post hoc identification. Different types of striatal interneuron were targeted for recording according to their expression of native tdTomato or GFP fluorescence in slices from PV-tdTomato mice, NPY-GFP mice and ChAT-Cre mice.
To optically stimulate ChR2-expressing axons in striatum, brief flashes (2 or 5 ms) of blue light (470 nm) were generated by a TTL-controlled LED (LB W5SN-GZJX-35-Z, Mouser Electronics; housed above the microscope condenser) and delivered to the slice as wide-field illumination. The intensity of the light flashes, as measured just above the slices, was ;1.5 mW/mm 2 . To avoid overt desensitization of ChR2, successive light flashes (single flash or, in some cases, 10-to 20-Hz trains of five flashes) were delivered at intervals of 15 s. We used two analytical approaches, based on averaging and latency, to assess whether optical stimuli evoked reliable monosynaptic responses in striatal neurons. Analysis of membrane dynamics, i.e., EPSPs or EPSCs, was performed on averaged responses from nine stimulus trials. Peak EPSP/EPSC amplitude was measured as the amplitude of the averaged event relative to averaged prestimulus baseline. Indicative amplitude thresholds for reliable detection of EPSPs and EPSCs were .0.3 mV and .2.5 pA, respectively. We considered evoked EPSPs/EPSCs to be putative monosynaptic responses when their onsets occurred at ,6 ms from the onset of light flashes. We also analyzed a time window of 6-12 ms from flash onset to test for disynaptic/polysynaptic responses. We sometimes observed brief (,0.5 ms) "switching artifacts" occurring within 0.1 ms of the onset and/or offset of light flashes; these artifacts were readily distinguished from evoked EPSPs/EPSCs. Series resistance was regularly monitored, but not compensated, during recordings; if there was .20% change in series resistance, neurons were excluded from further analyses. Theoretical liquid junction potential was estimated to be À10.3 mV, and was not corrected off-line in current-clamp recordings. After filling of neurons with biocytin or Alexa Fluor dye, slices were removed from the recording chamber and immersed in 4% w/v PFA in PB for 1-2 d at 4°C before further anatomic processing.

Brain slice processing for histofluorescence and immunofluorescence
To verify recorded cell types and ChR2 expression in striatum, free-floating tissue slices (250-300 mm thick) were washed in Triton PBS and processed for histofluorescence and immunofluorescence. Biocytin-filled striatal neurons were revealed by incubating slices overnight at room temperature in Triton PBS containing Alexa Fluor 405-conjugated streptavidin (1:500 dilution; S32351, ThermoFisher Scientific). Axonal ChR2 expression was revealed by incubating slices for 1-2 h in Triton PBS containing 10% v/v NDS (D9663, Sigma, RRID:AB_2810235), then overnight at room temperature in Triton PBS containing 1% v/v NDS and Alexa Fluor 488-conjugated rabbit anti-GFP (1:500; A-21311, ThermoFisher Scientific, RRID: AB_221477). For post hoc verification of ChR2 expression in and around the STN, the expression of native enhanced yellow fluorescent protein (EYFP) fluorescence was compared with the borders of STN defined by FoxP2 immunolabeling: slices were incubated for 1-2 h in Triton PBS containing 10% v/v NDS, then overnight at room temperature in Triton PBS containing 1% v/v NDS and rabbit anti-FoxP2 (1:1000; HPA000382, Atlas Antibodies, RRID: AB_1078908), then washed in PBS, and then incubated at room temperature for 4-5 h in Triton PBS containing Alexa Fluor 594-conjugated donkey anti-rabbit IgG (1:1000; A-21 207, Thermo Fisher Scientific, RRID:AB_141637). After binding of streptavidin/antibodies, and final washing in PBS, slices were mounted on glass slides, cover-slipped in Vectashield Mounting Medium, and imaged on a confocal microscope (FluoView FV1000, Olympus). For our ex vivo electrophysiological/optogenetics experiments, we only include data obtained in acute striatal slices from mice in which ChR2 expression was verified post hoc to be well restricted to STN neurons. The identities of recorded and biocytin-/dye-filled SPNs were verified according to the presence of densely spiny dendrites.

Statistical analysis
For each experiment, descriptions of critical variables (e.g., number of mice, neurons, and other samples evaluated) as well as details of statistical design and testing can be found in Results and the figure legends. For monosynaptic retrograde tracing experiments, the ratio (x/y) of the stereologically-estimated total number of striatal starter neurons (x) to the estimated total number of retrogradely-labeled STN neurons (y) was calculated for each animal (Do et al., 2016;Choi et al., 2019), and is referred to as the "normalized connectivity index." Graphing and statistical analyses were performed with GraphPad Prism (v8.4.2 or v5.03, RRID:SCR_002798). The Shapiro-Wilk test was used to judge whether datasets were normally distributed (p 0.05 to reject). All datasets used for comparative statistical testing were determined to be normally distributed. To test for differences between two datasets that were dependent or independent, we used paired or unpaired t tests, respectively. For testing three or more datasets that were dependent or independent, we used one-way repeated measures ANOVAs or one-way ANOVAs, respectively (both with Tukey's post hoc tests where appropriate). Significance for all statistical tests was set at p , 0.05 (exact p values are given in the text). Data are represented as group means 6 SEMs unless stated otherwise, with some plots additionally showing individual samples (mice or neurons) as appropriate.

Data availability
Data generated and analyzed for this study will be made available upon request to peter.magill@ndcn.ox.ac.uk or to koostib@gmail.com.

Subthalamic nucleus neurons project to, and form synapses with, neurons in dorsal striatum
Evidence of a subthalamostriatal projection in mice is somewhat conflicting; one anterograde tracing study explicitly refutes its existence (Schweizer et al., 2016), whereas other retrograde tracing studies clearly support its existence (Tervo et al., 2016;Klug et al., 2018;Choi et al., 2019). We used a combination of anterograde and retrograde tracing strategies, actioned through intracerebral injections of "conditional" viral vectors in transgenic mice, to provide further context on this issue and elucidate the structural basis for direct interactions between genetically-defined cell types in STN and striatum.
Most STN neurons in rodents and other mammalian species are glutamatergic and robustly express the Slc17a6 gene encoding the synaptic protein vesicular glutamate transporter 2 (VGluT2; Y. Smith and Parent, 1988;Albin et al., 1989;Barroso-Chinea et al., 2007;Rico et al., 2010;Schweizer et al., 2016). To reveal the axonal projections of glutamatergic STN neurons, we stereotaxically injected AAV vectors expressing GFP in a Cre recombinase-dependent manner into the STN of adult Slc17a6-Cre (VGluT2-Cre) mice (Fig. 1A). The somata, dendrites and axons of transduced neurons were then revealed with immunoreactivity for GFP. This anterograde tracing strategy allowed for highly-selective transduction of neuronal somata and dendrites throughout the rostrocaudual extent of STN ( Fig. 1B-D), delineated according to the borders of a group of densely-packed neurons immunoreactive for the transcription factor FoxP2 (Abdi et al., 2015). We occasionally observed sparse transduction of a few neuronal somata in the neighboring zona incerta and parasubthalamic nucleus ( Fig. 1B-D). However, we did not observe any GFP-expressing (AAV-GFP1) somata in brain structures known to robustly innervate striatum, including the cerebral cortex, intralaminar and motor thalamus, basolateral and central amygdala, ventral tegmental area or substantia nigra pars compacta (Y. Smith et al., 1998;Pan et al., 2010). The AAV-GFP1 axons of transduced STN neurons traversed rostrally into the dorsal striatum, where they diffusely arborized over several millimeters (Fig. 1E-H). The ventral striatum (nucleus accumbens) was also innervated by AAV-GFP1 axons, but the projection was comparatively scant (Fig. 1E, F). In contrast, we observed dense AAV-GFP1 axonal plexuses within the GPe (Fig. 1H), a major target of STN neurons (Y. Smith et al., 1998). Within dorsal striatum, the AAV-GFP1 axons exhibited both "en passant" and "aux terminaux" boutons that were immunoreactive for VGluT2 ( Fig.  1I-L). The results of these anterograde tracing studies support the existence of a glutamatergic projection from the STN to the dorsal striatum in mice.
To complement these anterograde labeling studies, we conducted monosynaptic retrograde tracing from neurons in the dorsal striatum (Callaway and Luo, 2015;Ährlund-Richter et al., 2019). This entailed the sequential injection of a Cre-dependent helper virus and a modified rabies virus into the dorsal striatum of Cre-expressing mice ( Fig. 2A). We first aimed to elucidate the extent to which STN neurons innervate striatal neurons as a whole. The vast majority of striatal neurons are GABAergic, with the exception being cholinergic interneurons (Tepper et al., 2010(Tepper et al., , 2018Silberberg and Bolam, 2015). These GABAergic neurons can be selectively accessed for study by exploiting their expression of the Slc32a1 gene encoding the vesicular GABA transporter (VGAT), whereas cholinergic neurons can be targeted via their expression of the choline acetyltransferase (ChAT) gene. Thus, to achieve our aim, we unilaterally injected the Cre-dependent helper virus [bicistronically expressing TVA receptor fused to a V5 tag, and the rabies glycoprotein (RG); Ährlund-Richter et al., 2019] into the central aspects of the dorsal striatum of double transgenic adult VGAT-Cre:ChAT-Cre mice (n = 4). Allowing time for Cre-mediated recombination, and the generation of "starter" neurons expressing all components necessary for retrograde labeling of their presynaptic partners, we then injected modified rabies virus (RG deleted, pseudotyped with EnvA, and expressing enhanced GFP) into the same striatal locations ( Fig. 2A). We visualized striatal starter neurons by their co-expression of immunoreactivity for V5 and rabies-encoded GFP (Rb-GFP; Fig.   2A-D). By virtue of stereotaxic targeting, we did not observe starter cells in the nucleus accumbens. The majority of starter neurons in dorsal striatum [86.9 6 1.0% (mean 6 SEM); n = 5055 total starter neurons counted from 4 mice] expressed nuclear immunoreactivity for Ctip2, a marker of most SPNs but not striatal interneurons (Arlotta et al., 2008). A small proportion of starter neurons (2.4 6 0.4%) instead expressed somatic immunoreactivity for ChAT. A third group of starter neurons (10.7 6 0.9%) expressed . Immunoreactivity for the transcription factor FoxP2 (magenta) was used to facilitate delineation of STN and other brain regions. cp, cerebral peduncle; Ctx, cortex; fr, fascicularis retroflexus; Hc; hippocampus; Pf, parafascicular nucleus of the thalamus; Po, posterior thalamus; Str, striatum; VPL, ventral posterolateral nucleus; VPM, ventral posteromedial nucleus; ZI, zona incerta. B-D, Immunofluorescence signals for anterogradely-labeled neurons (AAV-GFP) and FoxP2 in coronal sections from the same mouse as shown in A. Note that AAV-GFP1 somata and dendrites were located in "rostral" (B), "central" (C), and "caudal" (D) parts of STN. A standard stereotaxic reference (approximate distance relative to Bregma) is given for each rostrocaudal level. E-H, Immunofluorescence signals for AAV-GFP1 axons in the dorsal striatum (boundaries marked by dashed lines) and other forebrain regions of the same mouse as that shown in A-D. Sections are ordered from most rostral (E) to most caudal (H). Note the relatively dense plexus of AAV-GFP1 axons within the external globus pallidus (GPe), which is well known to be a major target of STN neurons (H). ac, anterior commissure; NAc, nucleus accumbens. I, J, AAV-GFP1 axons exhibit boutons, both "en passant" and "aux terminaux," in dorsal striatum. Immunoreactivity for NeuN (orange) was used to reveal the somata of striatal neurons. K, L, Highermagnification confocal micrographs of boxed zones in J (with NeuN channel omitted). AAV-GFP1 axonal boutons (arrowheads) expressed immunoreactivity for the vesicular glutamate transporter VGluT2 (cyan). Scale bars: A, 750 mm; B-D, 200 mm; E-H, 750 mm; I, 100 mm; J, 5 mm; K, L, 1 mm.
Research Article: New Research neither Ctip2 nor ChAT, suggesting they were GABAergic interneurons. Retrogradely-labeled neurons that monosynaptically innervated the starter neurons were identified by their expression of immunoreactivity for Rb-GFP, but not V5 ( Fig. 2A-D). Importantly, Rb-GFP1 neurons were observed throughout the rostrocaudual extent of STN ( Fig. 2E-G). A high proportion (93.7 6 0.8%) of these subthalamostriatal neurons (n = 314 total neurons counted from 4 mice) expressed nuclear immunoreactivity for FoxP2. Using unbiased stereological sampling methods, we estimated the total numbers of striatal starter neurons and Rb-GFP1 STN neurons in each VGAT-Cre:ChAT-Cre mouse (Fig. 2H). The average number of starter neurons per mouse was estimated to be 25 275 6 2820 whereas the average number of Rb-GFP1 STN neurons was 1570 6 137. To evaluate the relative connectivity from these stereological estimates, we calculated a "normalized connectivity index" for each mouse (Do et al., 2016;Choi et al., 2019). On average, the normalized connectivity index in the VGAT-Cre:ChAT-Cre mice was 0.065 6 0.010. Thus, in general, striatal starter cells far outnumbered the retrogradely-labeled STN neurons. The results of these retrograde tracing studies suggest that glutamatergic STN neurons innervate and form synaptic connections with a host of neurons in the dorsal striatum.

Subthalamic nucleus neurons rarely provide inputs to striatal spiny projection neurons
Striatal SPNs are heterogeneous in form and function. The direct/indirect pathways model of BG circuit organization (DeLong, 1990;Y. Smith et al., 1998) posits a dichotomy in striatal output, instantiated by direct pathway SPNs (dSPNs) and indirect pathway SPNs (iSPNs). Monosynaptic retrograde tracing studies have revealed some quantitative differences in the extrinsic inputs to dorsal striatal dSPNs and iSPNs, although there are some clear inconsistencies (Wall et al., 2013;Guo et al., 2015;Fürth et al., 2018). To elucidate the extent to which STN neurons innervate dSPNs, we unilaterally and sequentially injected Cre-dependent helper virus and modified rabies virus into the dorsal striatum of adult Drd1a-Cre mice (n = 5; Fig. 3A). The majority of striatal starter neurons expressed Ctip2 (85.0 6 5.6%; n = 2261 total neurons counted from 4 mice). Of these Ctip21 neurons, only a tiny fraction (0.6 6 0.2%) also expressed somatic immunoreactivity for preproenkepahlin (PPE), a selective marker of iSPNs (Lee et al., 1997;Gerfen and Surmeier, 2011;Garas et al., 2016;Sharott et al., 2017). These data indicate a highlyselective transduction of dSPNs by the two viruses in the Drd1a-Cre mice. Focusing on STN, we then observed Rb-GFP1 neurons throughout its rostrocaudual extent (Fig. 3B-D). A high proportion (93.8 6 1.6%) of these subthalamostriatal neurons (n = 223 total  . Subthalamic nucleus neurons rarely provide inputs to striatal spiny projection neurons. A, Retrograde labeling of STN neurons that monosynaptically innervate direct pathway spiny projection neurons (dSPNs). Inset, a helper virus and modified rabies virus were unilaterally injected into the dorsal striatum of adult Drd1a-Cre mice. Neurons in STN that innervate starter dSPNs express rabies-encoded enhanced GFP (Rb-GFP, green). Main image, Coronal section from a Drd1a-Cre mouse, showing immunofluorescence signals for V5 (cyan) and Rb-GFP. Note the enriched co-localization of V5 and Rb-GFP signals, indicative of starter neurons, in the dorsal striatum (boundaries marked by dashed lines). B-D, Retrogradely-labeled (Rb-GFP1) input neurons were located in rostral (B), central (C), and caudal (D) parts of the STN; all images from a single Drd1a-Cre mouse. Immunoreactivity for FoxP2 (magenta) was used to facilitate delineation of STN (dashed lines). C, Inset, Higher-magnification confocal image of Rb-GFP1 neurons within the white boxed area in C; the Rb-GFP1 neurons often co-expressed FoxP2 (arrowheads). E, Retrograde labeling of STN neurons that monosynaptically innervate indirect pathway SPNs (iSPNs). As for scheme in A, but with use of Adora2a-Cre mice to target iSPNs. F-H, Retrogradely-labeled input neurons were located in rostral (F), central (G), and caudal (H) parts of the STN; all images from a single Adora2a-Cre mouse. I, Stereologically-estimated numbers of striatal starter neurons in Drd1a-Cre mice (n = 5) and Adora2a-Cre (n = 5) mice, respectively, as well as retrogradely-labeled STN neurons in these mice (each neurons counted from 5 mice) expressed FoxP2. To elucidate the extent to which STN neurons innervate iSPNs, we conducted the same monosynaptic retrograde tracing procedure in Adora2a-Cre mice (n = 5; Fig. 3E). The majority of striatal starter neurons expressed Ctip2 (94.8 6 1.5%; n = 2551 total neurons counted from 5 mice), and of these Ctip21 neurons, almost all (97.7 6 0.2%) co-expressed PPE, together indicating a highly-selective transduction of iSPNs in the Adora2a-Cre mice. Again, we observed Rb-GFP1 neurons throughout the rostrocaudual extent of STN (Fig. 3F-H), and a high proportion (95.4 6 2.9%) of these subthalamostriatal neurons (n = 112 total neurons counted from 5 mice) expressed FoxP2.
The average number of striatal starter neurons per Drd1a-Cre mouse was stereologically estimated to be 10,470 6 2597 whereas the average number of Rb-GFP1 STN neurons was 1088 6 350 (Fig. 3I). The average number of striatal starter neurons per Adora2a-Cre mouse was estimated to be 8263 6 2139 whereas the average number of Rb-GFP1 STN neurons was 477 6 124 ( Fig. 3I). On average, the normalized connectivity index in Drd1a-Cre mice (0.098 6 0.011) was significantly higher (p = 0.0183, unpaired t test) than that in Adora2a-Cre mice (0.060 6 0.007). Taken together, these anatomic data suggest that STN neurons modestly innervate both dSPNs and iSPNs, and that the innervation of dSPNs is comparatively greater.
To gain further insight into the properties of STN inputs to SPNs, we used ex vivo electrophysiology and optogenetics to test for neuronal responses to selective activation of subthalamostriatal axons. Specifically, we made visualized whole-cell patch-clamp recordings from SPNs in dorsal striatal slices from adult wild-type mice that had received STN injections of an AAV expressing the lightactivated ion channel channelrhodopsin2 (ChR2) fused to enhanced yellow fluorescent protein (EYFP) under the control of the Camk2a (CaMKIIa) promoter (Fig. 3J). We then delivered single, brief flashes (2 or 5 ms) of blue light (470 nm, ;1.5 mW/mm 2 ) to the slices to evoke neurotransmitter release from ChR2-expressing STN axons (Fig. 3J). This optogenetics-based strategy allowed for well-circumscribed transduction of STN neurons with ChR2, as revealed with EYFP (Fig. 3K,L). The identities of recorded SPNs were verified according to morphologic criteria (as revealed after cell filling with fluorescent dye or biocytin; Fig.  3M) and, in cases where we used a K-gluconate-based pipette solution for recordings, additionally confirmed by their distinctive intrinsic membrane properties ( Fig. 3N; Gertler et al., 2008). We took care to ensure that each neuron was recorded in an area of the striatal slice that was innervated by ChR2-expressing axons (Fig. 3M). In this first level of analysis, we did not attempt to further classify the recorded neurons as dSPNs or iSPNs. We recorded 29 SPNs (from 6 mice) in both current-clamp mode and voltage-clamp mode (V h À70 mV) using a K-gluconate pipette solution. Upon delivery of blue light flashes (5 ms) to the surrounding striatal tissue, none of these SPNs exhibited evoked responses (EPSPs or EPSCs, respectively) at latencies of ,12 ms that should capture monosynaptic and polysynaptic responses (Fig. 3O). In using a K-gluconate pipette solution, we were mindful of the possibility that inadequate "space clamping" could lead to false negatives, such that excitatory inputs to the distal dendrites of neurons might not be readily detected with recordings at their somata (Koos et al., 2004). To help address this, we recorded another 20 SPNs (from 4 mice) in voltage-clamp mode using a CsCl-based pipette solution that reduces electronic distance effects (Koos et al., 2004). Again, none of these SPNs exhibited monosynaptic or polysynaptic responses to blue light flashes of 2 ms (Fig. 3P). In summary, 0 of 49 recorded SPNs (from 10 mice) exhibited a detectable response to selective activation of subthalamostriatal axons (Fig. 3Q). These electrophysiological data indicate that STN inputs to striatal spiny projection neurons are rare and, when assessed by somatic recordings, poorly efficacious.

Subthalamic nucleus neurons commonly provide reliable inputs to striatal parvalbumin-expressing interneurons
Monosynaptic retrograde tracing has revealed some quantitative differences in the glutamatergic cortical and thalamic inputs to dorsal striatal dSPNs and PV interneurons (Choi et continued circle or square indicates an estimate from an individual Drd1a-Cre mouse or Adora2a-Cre mouse, respectively). J, Main strategy for combined ex vivo electrophysiological and optogenetic interrogation of synaptic connections between STN neuron axons and striatal SPNs. A viral vector (AAV-CamKIIa-ChR2-EYFP) expressing the light-activated ion channel channelrhodopsin2 (ChR2) fused to a fluorescent reporter (EYFP) was first injected into the STN of wild-type (C57BL/6) mice. Visualized whole-cell patch-clamp recordings were then made from SPNs in acutely-prepared tissue slices from these mice, with brief pluses of blue light (470 nm) being used to selectively stimulate the axons of transduced ChR2-expressing STN neurons. K, Coronal tissue section from a wild-type mouse, showing a circumscribed group of transduced neurons (AAV-ChR2-EYFP, green) in the STN (dashed lines). Hc; hippocampus; Th, thalamus. L, Higher-magnification image of STN and neighboring brain regions from section shown in K. M, SPN filled with Alexa Fluor 594 dye (red) during ex vivo whole-cell patch-clamp recordings, surrounded by ChR2-expressing axons of STN neurons (AAV-ChR2-EYFP, green). N, Current-clamp recordings (K-gluconate-based pipette solution) of a neuron classified as an SPN, as per its characteristic voltage responses (top) to somatic injection of 900-ms pulses of hyperpolarizing or depolarizing current (bottom, from À100 to 1225 pA, in 25-pA steps). O, Voltage-clamp recordings (V h = À70 mV) of an SPN, patched with K-gluconate solution, that showed no responses to optical stimulation of ChR2-expressing STN axons; nine individual trial traces ( al., 2019), raising the possibility that STN inputs to these two striatal cell types might also differ. To elucidate the extent to which STN neurons innervate PV interneurons, we unilaterally and sequentially injected Cre-dependent helper virus and modified rabies virus into the dorsal striatum of adult PV-Cre mice (n = 9; Fig. 4A). The majority of striatal starter neurons expressed somatic immunoreactivity for PV (95.5 6 1.3%; n = 357 total neurons counted from 9 mice), attesting to a highly-selective transduction of PV interneurons by the two viruses. We then observed Rb-GFP1 neurons throughout the rostrocaudual extent of STN (Fig. 4B-D). A high proportion (90.7 6 3.5%) of these subthalamostriatal neurons (n = 112 total neurons counted from 9 mice) expressed FoxP2. The average number of striatal starter neurons per PV-Cre mouse was stereologically estimated to be 640 6 107 whereas the average number of Rb-GFP1 STN neurons was 262 6 60 ( Fig. 4E). On average, the normalized connectivity index in PV-Cre mice was relatively high (0.430 6 0.067). Taken together, these anatomic data suggest that STN neurons robustly innervate striatal PV interneurons.
To gain further insight into the properties of STN inputs to PV interneurons, we made whole-cell patch-clamp recordings from PV interneurons, as visualized by their native red fluorescence in striatal slices from adult mice expressing the reporter protein tdTomato under the control of the Pvalb promoter (PV-tdTomato mice), and tested their responses to optical activation of subthalamostriatal axons expressing ChR2 fused with EYFP (Fig. 4F). When recorded in current-clamp mode using a K-gluconate pipette solution, PV interneurons (n = 14, from 5 mice) exhibited intrinsic membrane properties typical of this cell type, i.e., conforming to a "fast spiking" phenotype ( Fig. 4G; Tepper et al., 2010). Upon delivery of single brief flashes (2 ms) of blue light, 8 of 14 PV interneurons recorded using K-gluconate consistently exhibited short-latency EPSPs (Fig. 4H) and/or EPSCs (Fig. 4I). The average peak EPSP amplitude was 1.55 6 0.42 mV (n = 6 responsive PV interneurons), and did not cause the interneurons to fire action potentials. The average peak EPSC amplitude was 35.22 6 7.31 pA (n = 7 responsive PV interneurons with V h = À70 mV; Fig. 4J). We recorded another 15 PV interneurons (from 3 mice) in voltage-clamp mode using a CsCl-based pipette solution. The blue light flashes (2 ms) evoked EPSCs in 10 of 15 PV interneurons; the average peak EPSC amplitude was 97.14 6 18.40 pA (Fig. 4J). The average latency to the onset of evoked EPSCs was 3.90 6 0.21 ms (n = 17 responsive PV interneurons; Fig.  4K); because of the short-latency onsets of the evoked EPSCs, we considered all these responses of PV interneurons to be putatively monosynaptic (see Materials and Methods). In summary, 18 of 29 recorded PV interneurons (from 5 mice) exhibited clear and consistent responses to the selective activation of subthalamostriatal axons (Fig.  4L). The glutamatergic nature of the inputs from STN was verified in a subset of voltage-clamp recordings (n = 5 PV interneurons tested); bath application of antagonists of AMPA/kainate-type and NMDA-type ionotropic glutamate receptors (CNQX and AP5, respectively, both at 10 mM) greatly diminished the amplitudes of evoked EPSCs (Fig.  4M,N). Finally, we explored the short-term plasticity profiles of these subthalamostriatal inputs by delivering trains of five light flashes at interflash intervals of 50 or 100 ms (12 responsive PV interneurons; Fig. 4O-R). As compared with the EPSCs evoked by the first flashes in the trains, the amplitudes of subsequent EPSCs were successively reduced with optical stimulation at 20 Hz ( Fig. 4O,P) but not at 10 Hz (Fig. 4Q,R). Taken together, these electrophysiological data indicate that STN inputs to striatal PV-expressing interneurons are relatively common and efficacious.

Subthalamic nucleus neurons rarely provide inputs to two types of striatal neuropeptide Y-expressing interneurons
In rodent striatum, GABAergic interneurons expressing neuropeptide Y (NPY) are comprised of at least two major cell types that can be distinguished by a host of properties (Tepper et al., 2010(Tepper et al., , 2018. The most common type of NPY interneuron is defined by, among other features, its co-expression of somatostatin and nitric oxide synthase. Because these NPY/SOM/NOS interneurons exhibit characteristic low-threshold spikes (LTSs) ex vivo, they are also commonly termed "LTS" interneurons (Tepper et al., 2010(Tepper et al., , 2018. Dorsal striatal SOM and PV interneurons exhibit dissimilarities in their glutamatergic cortical and thalamic inputs Choi et al., 2019), again raising the possibility that STN inputs to these two striatal cell types might also differ. To elucidate the extent to which STN neurons innervate SOM interneurons, we unilaterally and sequentially injected Cre-dependent helper virus and modified rabies virus into the dorsal striatum of adult SOM-Cre mice (n = 5; Fig. 5A). The majority of striatal starter neurons expressed somatic immunoreactivity for somatostatin (98.0 6 1.0%; n = 481 total neurons counted from 5 mice), attesting to a highly-selective transduction of SOM interneurons by the two viruses. Of these SOM-expressing interneurons, a large fraction (96.7 6 1.7%) also expressed immunoreactivity for NOS, indicating that the vast majority of starter neurons likely correspond to LTS interneurons. We then observed Rb-GFP1 neurons throughout the rostrocaudual extent of STN (Fig. 5B-D). A high proportion (97.8 6 2.2%) of these subthalamostriatal neurons (n = 49 total neurons counted from 5 mice) expressed FoxP2. The average number of striatal starter neurons per SOM-Cre mouse was stereologically estimated to be 2019 6 216 whereas the average number of Rb-GFP1 STN neurons was 196 6 47 (Fig. 5E). On average, the normalized connectivity index in SOM-Cre mice was 0.098 6 0.025. Taken together, these anatomic data suggest that STN neurons modestly innervate striatal NPY/SOM/NOS-expressing LTS interneurons.
To gain further insight into these neuronal connections, we made whole-cell patch-clamp recordings from LTS interneurons, as visualized by their native green fluorescence in striatal slices from adult mice expressing humanized Renilla GFP under the control of the Npy promoter (NPY-GFP mice; Ibáñez-Sandoval et al., 2011), and tested their responses to optical activation of subthalamostriatal axons expressing ChR2 fused with the reporter protein mCherry (Fig. 5F). When recorded in current- When recorded in current-clamp or voltage-clamp modes, none of these LTS interneurons exhibited monosynaptic or polysynaptic responses to the local delivery of brief (5 ms) flashes of blue light (Fig. 5H).
The second major type of NPY-expressing cell in mouse striatum, the so-called neurogliaform (NGF) interneuron, does not express SOM or NOS, and does not exhibit LTS activity (Ibáñez-Sandoval et al., 2011). Our use of SOM-Cre mice for monosynaptic retrograde tracing allowed for highly-selective access to LTS interneurons but, in doing so, precluded structural analyses of STN inputs to NGF interneurons. Previous work has shown that LTS interneurons and NGF interneurons differ considerably in their responses to cortical and thalamic inputs (Assous et al., 2017;. It is not known whether the responses of LTS and NGF interneurons to STN inputs also differ. To address this, we also made whole-cell patch-clamp recordings from NGF interneurons, as visualized in striatal slices from adult NPY-GFP mice, and tested their responses to optical activation of ChR2-expressing subthalamostriatal axons (Fig. 5F). When recorded in current-clamp mode using a K-gluconate pipette solution, NPY-expressing NGF interneurons (n = 4, from 2 mice) were readily classified as such by their distinctive membrane properties ( Fig. 5I; Ibáñez-Sandoval et al., 2011). When recorded in current-clamp or voltageclamp mode, none of these NGF interneurons responded to the delivery of brief (5 ms) flashes of blue light (Fig. 5J).
In a final set of experiments, we recorded another nine NPY interneurons (from 3 mice) in voltage-clamp mode using a CsCl-based pipette solution (note this solution precluded their electrophysiological classification as LTS or NGF interneurons). Again, none of these interneurons responded to 5-ms blue light flashes (Fig. 5K). In summary, 0 of 27 recorded NPY interneurons (from 4 mice) exhibited a detectable response to selective activation of subthalamostriatal axons (Fig. 5L). These electrophysiological data indicate that STN inputs to striatal neuropeptide Y-expressing interneurons are rare and, when assessed by somatic recordings, poorly efficacious.

Subthalamic nucleus neurons rarely provide inputs to striatal cholinergic interneurons
Monosynaptic retrograde tracing has revealed some quantitative differences in the glutamatergic cortical and thalamic inputs to dorsal striatal cholinergic interneurons as compared with SPNs (Guo et al., 2015) and PV interneurons (Klug et al., 2018), highlighting the prospect that STN inputs to cholinergic interneurons might also differ. To elucidate the extent to which STN neurons innervate cholinergic interneurons, we unilaterally and sequentially injected Cre-dependent helper virus and modified rabies virus into the dorsal striatum of adult ChAT-Cre mice (n = 5; Fig. 6A). The majority of striatal starter neurons expressed somatic immunoreactivity for ChAT (90.9 6 3.6%; n = 513 total neurons counted from 5 mice), confirming a highly-selective transduction of cholinergic interneurons by the two viruses. We then observed sparse Rb-GFP1 neurons scattered throughout the rostrocaudual extent of STN ( Fig. 6B-D). A high proportion (82.1 6 6.0%) of these subthalamostriatal neurons (n = 26 total neurons counted from 5 mice) expressed FoxP2. The average number of striatal starter neurons per ChAT-Cre mouse was stereologically estimated to be 2052 6 339 whereas the average number of Rb-GFP1 STN neurons was 104 6 29 (Fig. 6E). On average, the normalized connectivity index in ChAT-Cre mice was 0.052 6 0.014. Taken together, these anatomic data suggest that STN neurons modestly innervate striatal cholinergic interneurons.
To gain further insight, we made whole-cell patch-clamp recordings from cholinergic interneurons, as visualized by their native red fluorescence in striatal slices from adult ChAT-Cre mice that had received striatal injections of a Credependent AAV expressing tdTomato, and tested their responses to optical activation of subthalamostriatal axons expressing ChR2 fused with EYFP (Fig. 6F). When recorded  in current-clamp mode using a K-gluconate pipette solution, these cholinergic interneurons (n = 17 neurons from 4 mice) exhibited intrinsic membrane properties typical of this cell type, including spontaneous firing in the absence of current injection (Fig. 6G,J; Goldberg and Reynolds, 2011). When recorded in current-clamp or voltage-clamp mode, none of these cholinergic interneurons exhibited monosynaptic or polysynaptic responses to the local delivery of brief (2 ms) flashes of blue light (Fig. 6H,J). We also recorded another eight cholinergic interneurons (from 3 mice) in voltage- clamp mode using a CsCl-based pipette solution. Again, none of these interneurons responded to the 2-ms optical stimuli (Fig. 6I). In summary, 0 of 25 recorded cholinergic interneurons (from 7 mice) exhibited a detectable response to selective activation of subthalamostriatal axons (Fig. 6K). These electrophysiological data indicate that STN inputs to striatal cholinergic interneurons are rare and, when assessed by somatic recordings, poorly efficacious.

Comparative analyses of subthalamostriatal connections
A strength of our work here is the ability to directly compare the relative levels of STN innervation of multiple cell types in dorsal striatum, using both structural and electrophysiological readouts of connectivity. Taken together, the results of our monosynaptic retrograde tracing studies showed that the connectivity from STN neurons to striatal PV interneurons is significantly higher (;4-to 8-fold) than that from STN to any of the four other striatal cell types examined, as well as from STN to striatal neurons as a whole (Fig. 7A). Of equal note, the connectivities of dSPNs, iSPNs, SOM interneurons and ChAT interneurons were quantitatively similar when compared altogether (Fig. 7A). The collective results of our ex vivo electrophysiological/optogenetics experiments showed that, of the five striatal cell types tested, only PV interneurons (and 62% of them) exhibited robust monosynaptic excitatory responses to selective activation of subthalamostriatal axons (Fig.  7B). We conclude that there is a highly selective and impactful glutamatergic projection from the STN to striatal PV interneurons.

Discussion
Here, we provide structural and electrophysiological evidence of a remarkable target selectivity in the mouse subthalamostriatal projection. Our data converge to support the concept that glutamatergic STN neurons are positioned to directly and powerfully influence striatum by virtue of their enriched innervation of parvalbumin-expressing interneurons.

Anatomical readouts of subthalamostriatal connectivity
Monosynaptic retrograde tracing from genetically-defined cell types has generated important new insights into the sources and organization of extrinsic inputs to striatum (Wall et al., 2013;Guo et al., 2015;J.B. Smith et al., 2016;Fürth et al., 2018;Klug et al., 2018;Monteiro et al., 2018;Choi et al., 2019;Melendez-Zaidi et al., 2019). Using this approach, we provide comparative analyses of the STN innervation of five cell types in dorsal striatum. We observed that the STN innervates dSPNs, iSPNs, PV interneurons, SOM (LTS) interneurons, and cholinergic interneurons. Importantly though, the connectivity from STN neurons to striatal PV interneurons stands apart in being of comparatively high magnitude. The precise substrates for this striking selectivity are unclear. It could arise from many combinations of convergent, divergent, and exclusive "one-to-one" connections. A higher connectivity index could thus reflect one or more structural configurations, including (but not limited to): an individual PV interneuron is more likely to receive inputs from multiple STN neurons; an individual STN neuron is more likely to innervate multiple PV interneurons; and/or a relatively larger fraction of PV interneurons is innervated by STN. Past studies of dorsal striatum employing monosynaptic retrograde tracing have compared "whole-brain" inputs to two or three striatal cell types, and some have reported that STN provides inputs to SPNs and/or interneurons (Wall et al., 2013;Guo et al., 2015;J.B. Smith et al., 2016;Klug et al., 2018;Choi et al., 2019). All these studies indicate that inputs from STN are relatively scant, at least when compared with other glutamatergic inputs from cortex and thalamus. None of these studies reported significant differences in the innervation of multiple striatal cell types by STN neurons. This apparent discordance with our results might arise from differences in tools used, the cells transduced for tracing, and/or the metrics analyzed. Unlike other studies, we: (1) injected a single helper virus, avoiding ambiguities in defining starter cells with all components necessary for retrograde labeling of their presynaptic partners; (2) quantified starter cell "specificity" for all mouse lines; (3) employed unbiased stereology to estimate total numbers of starter cells and input neurons; and (4) used a normalized connectivity index focused on STN only, rather than expressing counts as a fraction of wholebrain inputs. We conclude from our anatomic experiments that there is a highly-selective projection from the STN to striatal PV interneurons. Accordingly, the STN joins a growing list of subcortical structures that, although not considered canonical sources of inputs to striatum, selectively target striatal interneurons; exemplars include the GPe Mallet et al., 2012) and peduculopontine nucleus . We speculate that selective innervation of striatal interneurons is a common circuit motif.

Electrophysiological readouts of subthalamostriatal connectivity
Building on the anatomic data above, we used a combination of ex vivo electrophysiology and optogenetics to gain the first direct insights into the incidence and strength of subthalamostriatal neurotransmission. We observed that, of the five striatal cell types tested, only PV interneurons commonly exhibited monosynaptic excitatory responses to activation of subthalamostriatal axons. The STN inputs to PV interneurons were glutamatergic, robust and reliable, with no or moderate attenuation when driven at frequencies similar to STN neuron firing rates in vivo i.e., 10-20 Hz (Mallet et al., 2008a, b;Deffains et al., 2016). As such, the anatomic and electrophysiological readouts of connectivity provided strong corroborative evidence of a highly-selective innervation of PV interneurons by the STN. However, the lack of responsive SPNs, NPY-LTS interneurons and cholinergic interneurons was at variance with anatomic data showing a modest innervation of these cell types. This variance could arise from the limitations of each technique. For example, retrograde tracing could give "false positives" if the rabies virus were taken up by non-synaptic structures. Then again, retrograde labeling of STN neurons could occur via single/few synapses that are weak, "silent" and/or located at the distal dendrites of these striatal neurons, thereby giving "false negatives" in electrophysiological recordings made at their somata. Our use of a CsCl-based pipette solution should have minimized electronic distance effects (Koos et al., 2004), but does not preclude undetectable inputs. There are other cases of mismatches between the anatomic connectivity, as defined by monosynaptic retrograde tracing, and electrophysiological connectivity of striatal circuits (Choi et al., 2019). Nevertheless, our results reinforce the utility of monosynaptic retrograde tracing (when highly selective for a given cell type, as was the case here) for generating predictive hypotheses about circuit connectivity that can then be further tested with complementary methods.

Implications for striatal microcircuits
The monosynaptic connection from the STN to PV interneurons has important implications for the organization of activity in striatal microcircuits. GABAergic PV interneurons can exert powerful effects on their SPN targets, for example, delaying or negating action potential firing (Koós and Tepper, 1999;Tepper et al., 2004). It follows that suprathreshold excitation of PV interneurons by STN inputs could provide a novel substrate for "feed-forward" inhibition in striatum. The efferent connections of PV interneurons in turn suggest that dSPNs, iSPNs, and NPY/SOM/NOS-expressing LTS interneurons (and some PV interneurons) would be subject to this feed-forward inhibition (Gittis et al., 2010;Planert et al., 2010;Tepper et al., 2010Tepper et al., , 2018Szydlowski et al., 2013). Notably, cholinergic interneurons are not extensively targeted by PV interneurons (Szydlowski et al., 2013) or STN neurons.

Subthalamostriatal combinatorics
The STN as a whole innervates all other BG nuclei, as well as discrete parts of midbrain, brainstem, thalamus and cerebral cortex (Y. Smith et al., 1998;Emmi et al., 2020). Because STN neurons exhibit markedly heterogeneous structural (axonal), physiological and molecular properties (Bevan et al., 2000;Sato et al., 2000;Koshimizu et al., 2013;Wallén-Mackenzie et al., 2020), it seems unlikely that all STN neurons innervate all of these diverse targets. It is currently unclear whether all STN neurons innervating striatum also innervate, via axon collaterals, one or more other BG nuclei. Singleneuron tracings in rats suggest the majority of subthalamostriatal neurons additionally innervate GPe, entopeduncular nucleus and substantia nigra, but there are further combinations of projections (Koshimizu et al., 2013). Single-axon reconstructions in monkeys present another scenario in which about one fifth of STN neurons might innervate striatum but no other BG nuclei; this interpretation is however challenged by incomplete labeling of axon collaterals (Sato et al., 2000). Our anatomic data show that at least five types of striatal neuron receive inputs from STN, albeit to differing degrees, but it remains to be determined whether an individual subthalamostriatal neuron can innervate more than one type of striatal neuron; the potential number of combinations of subthalamostriatal connections is large. Our experiments were focused on central aspects of dorsal striatum (caudate-putamen), thus precluding conclusions as to whether the same patterns of subthalamostriatal innervation also hold for other territories of dorsal striatum and/or ventral striatum (nucleus accumbens). Notably, the functional properties of PV interneurons vary across striatal regions (Garas et al., 2016;Monteiro et al., 2018), and this heterogeneity might extend to their inputs from STN.

Wider circuit context
The subthalamostriatal projection is not included in the direct/indirect pathways scheme nor (to our knowledge) any other model of the functional organization of corticalbasal ganglia-thalamocortical circuits. A feed-forward flow of information through these circuits is central to the direct/indirect pathways scheme (DeLong, 1990;Y. Smith et al., 1998). Conceptually, the striatum lies "upstream" of STN in the indirect pathway (Gerfen and Surmeier, 2011). One implication of this arrangement is that the routing back of STN output to striatum must occur through polysynaptic pathways, classically those incorporating the BG output nuclei and their thalamic effectors. There is scope for these sequential connections to degrade information carried by outgoing STN signals before it reaches striatum. In contrast, the monosynaptic projection from STN to striatum that we elucidate here offers a substrate by which striatal neurons, and particularly PV interneurons, can be quickly updated on STN activity dynamics with minimal distortion of signal. Given that STN selectively innervates GABAergic PV interneurons, which in turn innervate and powerfully control GABAergic iSPNs, this permutation of subthalamostriatal feedback could subserve a homeostatic function along the indirect pathway; increased STN firing would ultimately result in less iSPN output and thence, augmented GABAergic output from prototypic GPe neurons (Abdi et al., 2015) that would, in turn, restrain STN activity. This homeostatic function might also serve to counteract STN hyperactivity emerging in Parkinsonism (Mallet et al., 2008a, b); ultimately though, homeostatic control is likely overwhelmed, as also reflected by increased iSPN output and hypoactive prototypic GPe neurons after chronic dopamine depletion (Abdi et al., 2015;Sharott et al., 2017). All permutations of rapid monosynaptic subthalamostriatal feedback have implications for another influential model of BG functional organization that emphasizes the so-called hyperdirect pathway and is based on the cortical driving of STN neuron output before SPN output (Nambu et al., 2002). For example, when a voluntary movement is about to be initiated by cortical mechanisms and the hyperdirect pathway drives STN neurons, the swift and direct routing of these STN signals back to striatum could influence the subsequent/ongoing engagement of dSPNs and iSPNs that mediate the sequential selection and cancelation of desired and competing motor programs, respectively (Nambu et al., 2002). In this way, subthalamostriatal feedback could contribute to wider circuit dynamics that help ensure that movements are initiated, executed and terminated with appropriate timing (Nambu et al., 2002). We conclude that the cell type-selective innervation of striatum by glutamatergic STN neurons, as detailed here, is positioned to fulfil diverse and likely unique roles within BG circuits.