Long-range monosynaptic inputs targeting apical and basal dendrites of primary motor cortex deep output neurons

Abstract

The primary motor cortex (M1) integrates various long-range signals from other brain regions for the learning and execution of goal-directed movements. How the different inputs target the distinct apical and basal dendrites of M1 pyramidal neurons is crucial in understanding the functions of M1, but the detailed connectivity pattern is still largely unknown. Here, by combining cre-dependent rabies virus tracing, layer-specific chemical retrograde tracing, optogenetic stimulation, and electrophysiological recording, we mapped all long-range monosynaptic inputs to M1 deep output neurons in layer 5 (L5) in mice. We revealed that most upstream areas innervate both dendritic compartments concurrently. These include the sensory cortices, higher motor cortices, sensory and motor thalamus, association cortices, as well as many subcortical nuclei. Furthermore, the dichotomous inputs arise mostly from spatially segregated neuronal subpopulations within an upstream nucleus, and even in the case of an individual cortical layer. Therefore, these input areas could serve as both feedforward and feedback sources albeit via different subpopulations. Taken together, our findings revealed a previously unknown and highly intricate synaptic input pattern of M1L5 neurons, which implicates that the dendritic computations carried out by these neurons during motor execution or learning are far more complicated than we currently understand.

Introduction

The primary motor cortex (M1) is critical for motor execution and motor learning, especially for dexterous motor skills (Li et al. 2017; Peters et al. 2017; Papale and Hooks 2018). During the execution of movements, M1 needs to integrate various types of feedforward and feedback information, including the signal of intent from the prefrontal cortex (Goldman-Rakic 1987), preparatory signals from the premotor cortex (Thura and Cisek 2014; Svoboda and Li 2018), various levels of sensory signals (Asanuma 1981), as well as regulatory signals from the basal ganglia and the cerebellum via the thalamus (Hoover and Strick 1999; Sommer 2003). With respect to motor learning, different feedforward and feedback information to M1 are also required to drive synaptic plasticity believed to be crucial for the acquisition and consolidation of motor skills (Huber et al. 2012; Masamizu et al. 2014; Peters et al. 2014).

Pyramidal neurons in the cortex, including those in M1, have a distinct morphology in bearing two dendritic arbors. The apical dendrite extends dorsally and branches out to form tuft dendrites in layer 1 (L1), while the basal dendrites emanating directly from the soma are mainly distributed in the same layer (Spruston 2008). These two tiers of dendrites of pyramidal neurons in the cortex, as well as those of hippocampal pyramidal neurons, are now considered distinct functional compartments as they are electrotonically isolated and each is believed to be receiving different inputs (Larkman 1991; Spratling 2002; Branco et al. 2010). Furthermore, they exhibit different synaptic plasticity rules and molecular machineries (Sjostrom and Hausser 2006; Brzdak et al. 2019). There was finding suggesting that the secondary somatosensory cortex (S2) innervates M1 layer 5 (L5) pyramidal neurons spanning multiple cortical layers (Suter and Shepherd 2015). On the other hand, although it is generally believed that thalamic inputs to M1L5 neurons are mainly perisomatic (Hooks et al. 2013), it had been shown that thalamic inputs to distal apical dendrites in most cortical areas are in fact massive and highly convergent (Rubio-Garrido et al. 2009). Although there were previous studies that investigated the sources of inputs to M1 (Hooks et al. 2013; Luo et al. 2019), there has not been any attempt to systematically trace and distinguish the different monosynaptic inputs to the apical versus basal dendritic arbors of M1 neurons. Given the likely distinct functions of synaptic inputs to these two dendritic compartments in pyramidal neurons, quantifying the prominence of upstream areas that target M1 neurons and deciphering which dendritic arbor(s) they aim for is crucial in understanding how M1 controls movement and participates in the acquisition of motor skills.

It has been proposed that the integration of coincident inputs to the two dendritic compartments of pyramidal neurons could serve as a cellular mechanism of association and therefore recognition and learning (Larkum et al. 2009; Larkum 2013; Doron et al. 2020). At the same time, compared with single synaptic input models, pyramidal neuron-inspired neural networks with two input sites exhibited an increased computational power and learning ability (Kording and Konig 2000, 2001; Guerguiev et al. 2017). Recently, theoretical analysis and modeling studies demonstrated the potential of pyramidal neurons in realizing the backpropagation of errors during learning, an algorithm that is central to most contemporary artificial neural networks including the deep learning genre (Lillicrap et al. 2020). We are particularly interested in the synaptic inputs to L5 in M1, the most distinctive layer in this region, which contains heterogeneous groups of pyramidal cells that project to the thalamus, striatum, and brainstem regulating motor outputs (Weiler et al. 2008; Anderson et al. 2010) or via the spinal cord to control muscles directly (Ebbesen and Brecht 2017). Interestingly, we discovered in our previous study that during the acquisition of novel motor skills, the potentiation of synaptic inputs onto presumed tuft dendrites and basal dendrites of M1L5 neurons exhibit strikingly different temporal profiles (Li et al. 2017). This result implicates the differential and perhaps synergistic contributions of the inputs to the two different dendritic compartments of these output neurons in motor learning. In particular, conventional models of synaptic connections in sensory cortical areas (Larkum et al. 2009; Larkum 2013) would implicate that feedback or top-down signals from sensory cortices, sensory thalamus, and association areas innervate the apical dendrites of pyramidal neurons in L1, while those feedforward signals from higher motor areas and motor thalamus would target predominantly the basal dendrites.

In deciphering the source of synaptic inputs, transsynaptic retrograde tracing utilizing L5-specific Rbp4-cre mice via modified rabies virus (RV) (Wickersham et al. 2007; Guo et al. 2015) can unambiguously identify the upstream regions that send monosynaptic innervation to M1L5 neurons. On the other hand, deposition of distinct retrograde tracers onto L1 and L5, respectively, can help reveal dendritic arbor-specific inputs onto these neurons. Therefore, combining these two tracing techniques can generate an accurate and comprehensive map of the layer-specific long-range synaptic inputs to M1L5 neurons. We also assessed the synaptic strengths of layer-specific input pathways via optogenetic manipulation and electrophysiological recording. Our results revealed that, with the notable exception of secondary motor cortex (M2), most upstream nuclei innervate both dendritic compartments concomitantly with a general bias to the apical dendrites. Also, the dichotomous inputs originate from largely different neuronal subpopulations in nearly all upstream areas. Our findings thus revealed a highly intricate synaptic input pattern of M1L5 neurons and strongly implicate that the dendritic computation carried out by these neurons is far more complicated than we currently envision.

Materials and Methods

Animals

For RV tracing, Rbp4-Cre mice (Gerfen et al. 2013) were obtained from Jackson Laboratories. For the chemical tracing of different innervation to the apical dendrites and basal dendrites of M1L5 pyramidal neurons, 2-month-old C57/6 J wild-type male mice were used. The animals were bred and maintained by the Laboratory Animal Service Centre of the Chinese University of Hong Kong (CUHK). The animal room was controlled at a temperature of 23 °C on a 12-h light/dark cycle. All animals were handled in strict accordance with the CUHK guidelines, and the procedures were approved by the Animal Experimentations and Ethics Committee. All experiments were performed during the light phase (09:00–19:00).

RV Retrograde Tracing

Two-month-old Rbp4-Cre mice were used. They were anesthetized with ketamine (75 mg/kg, i.p.) and xylazine (6 mg/kg, i.p.) and placed gently in a stereotaxic frame (Narishige). Mice corneas were covered with petrolatum to prevent drying. After opening the skin and removing the temporal muscle, the wound was cleaned alternately with iodine and 75% alcohol to prevent inflammation. Then, a hole was drilled in the skull above the target area. Around 100 nL of a 2:3 volume mixture of AAV2/9-EF1α-DIO-RVG and AAV2/9-EF1α-DIO-EGFP-TVA was injected into the M1 L5 (AP: 1.2 mm, ML: 1.5 mm, and D: 1.55 mm) of Rbp4-Cre mice. Two weeks later, 50-nL RV-ENVA-△G-mCherry was injected into the same brain location. After 5 days, rabies and mCherry expression spread; then, mice were sacrificed for rabies-tracing signal detection. All the viruses were delivered by a sharp micropipette mounted on a Stoelting Quintessential Stereotaxic Injector (Stoelting) attached to a micromanipulator and then injected at a speed of 40 nL/min. The glass micropipette was held for an extra 10 min after the completion of the injection and then slowly withdrawn. After the surgery, the incisions were stitched and Baytril (5 mg/kg) and Buprenorphine (0.1 mg/kg) were applied to prevent inflammation and alleviate pain for the animals.

Retrograde Tracing by Fast Blue and CTB555

Two-month-old C57/6 J mice were anesthetized with ketamine (75 mg/kg, i.p.) and xylazine (6 mg/kg, i.p.) and placed gently in a stereotaxic frame (Narishige). Mice corneas were covered with petrolatum to prevent drying. After opening the skin and removing the temporal muscle, the wound was cleaned alternately with iodine and 75% alcohol to prevent inflammation and a craniotomy was made to expose 3 × 3 mm the cerebral surface. After retracting the dura matter, 50-nL recombinant cholera toxin-b conjugated to AlexaFluor-555 (CTB-555, Thermo Fisher Scientific, RRID: C_22843) in PBS was injected to the target site at M1L5 (AP: 1.2 mm, ML: 1.5 mm, and DV: 1.55 mm) with the injection speed of 60 nL/min. The glass micropipette was held for an extra 10 min after the completion of the injection and then slowly retreated. For the epipial application of Fast Blue (Polysciences Inc., Cat#:17740), the procedure followed previously published report (Rubio-Garrido et al. 2007). Small (~1 × 1 mm) pieces of filter paper were saturated in a 1% solution of FB in distilled water and then laid on the pial surface. The removed skull was put back in place and a layer of tissue glue was applied to the crack to facilitate tissue recovery. After the tissue glue had solidified, the muscle and skin were sutured, and Baytril (5 mg/kg) and Buprenorphine (0.1 mg/kg) were applied.

Stereotaxic Surgery for Virus Injection

Two-month-old C57/6 J mice were anesthetized and placed gently in a stereotaxic frame (Narishige) as described above until the skull was exposed and sterilized. A mounted drill was used to create holes in the skull above the injection site including VL (AP: 1.34 mm, ML:1.2 mm, and DV: 3.6 mm), VM (AP: 1.55 mm, ML:1.0 mm, and DV: 4.15 mm), S1HL L2/3(AP: 0.47 mm, ML:1.5 mm, and DV: 1.0 mm), S1DZ L5(AP: 0.47 mm, ML:2.5 mm, and DV: 1.5 mm), and M2L5 (AP: 2.45 mm, ML:1 mm, and DV: 1.45 mm). Injections of either AAV5-hSyn-hChR2(H134)-mCherry (tier ≥7 × 10¹² vg/ml, RRID: Addgene_26976) or AAV5-hSyn-mCherry (tier ≥7 × 10¹² vg/ml, RRID: Addgene_114472) were injected into the target area with a sharp micropipette mounted on a Stoelting Quintessential Stereotaxic Injector (Stoelting) attached to a micromanipulator. A volume of 50-nL virus was injected unilaterally at a rate of 40 nL/min. After the injection, the glass needle tip was left in place for an additional 10 min before slow retraction to prevent backflow and infection of dorsal tissue to the target areas. The incision of the scalp was sutured and analgesia was performed for 3 days after the injection to help recovery. Four weeks after virus injection, relevant histological and electrophysiological experiments were performed.

Histology

After 10 days RV injection, 7 days after chemical tracer injection, or 4 weeks after virus injection, mice were anesthetized with an overdose of ketamine/xylazine and transcardially perfused with 0.01 M phosphate-buffered saline (PBS, 30–40 mL) followed by 4% paraformaldehyde (PFA)/0.1 M PB pH 7.4 (30–40 mL). The brain was carefully extracted and postfixed in 4% PFA overnight (24 h). The brain was then transferred to 30% sucrose for 2 days until the brain dehydrated and sunk, and then embedded in the frozen slice embedding agent O.C.T. (SAKURA Finetek from VWR, cat#25608–930) and sectioned in coronal plane. Coronal slices were collected at 50-μm resolution in 96-well plates containing cyroprotectant (0.1-M phosphate buffer, ethylene glycol, and glycerol). For RV-tracing brain slices and AAV virus-injected brain slices, they were counterstained with 4',6-diamidino-2-phenylindole (DAPI) and cover slipped with mounting media. Slides were scanned under a microscope (C1, Nikon) in the blue and red channels and the images were converted to synthetic Tagged Image File Format (TIFF) for further analysis. When it was needed to differentiate cells in dense distributed region, separate channels could be thresholded. Considering that there was also projection from contralateral region, we counted the traced cells bilaterally. We overlapped the spliced pictures with the coronal mouse brain reference Paxinos and Franklin mouse brain atlas (4th edition, 2013), and transparent overlap processing was carried out by Adobe Illustrator CS6 (RRID: SCR_010279) to confirm the localization of the brain slices and nuclei. The spliced image to be analyzed and the corresponding coronal reference map were imported into the AI, and the map was overlaid on the scanned slice image. By stretching the reference map to the slice boundary and making multiple precise adjustments at multiple points according to the anatomical landmarks of the fibrous bundle, the atlas map could be well matched to the slice. Cell counts were carried out in the traced regions of each brain slide, and laminar counts were carried out in cortical regions. For each brain slice, the corresponding AP plane was recorded. Counters were blind to the condition and were examined for interparticipant reliability with a >95% score when counting the same brain region prior to analysis. All data were expressed as a percentage of the total number of input neurons in order to normalize the number of neurons labeled in different mice.

Electrophysiological Recording with Optogenetic Stimulation

Adult mice slice preparation was based on previously reported procedures (Ting et al. 2014). After 4 weeks of virus expression, infected male mice were deeply anesthetized and decapitated. Mice were anesthetized by isoflurane and perfused transcardially with 30-mL room temperature NMDG-artificial cerebrospinal fluid (aCSF) containing the following (in mM): 92 NaCl, 2.5 KCl, 1.25 NaH₂PO₄, 20 NaHCO₃, 10 HEPES, 25 Glucose, 5 Na-ascorbate, 2 thiourea, 3 Na-pyruvate, 10 MgSO₄, 0.5 CaCl₂, and 12 N-acetyl-L-cysteine, and the solution were saturated with carbogen (95%O₂/5%CO₂) prior to use to ensure stable pH buffering 7.4 and adequate oxygenation. After the perfusion, the mice were decapitated and the brains were gently extracted from the skull and placed into the cutting solution NMDG-aCSF for an additional 30 s. The trimmed brain was glued to a mounting cylinder and the slicing chamber was filled with NMDG-aCSF solution and carbogenated throughout the procedure. Coronal section of 300 μm was cut with a vibrotome (Campden 5100MZ-PLUS) and the sections were transferred into a holding chamber containing 34 °C NMDG-aCSF for recovery for 10 min. After the initial recovery, the slices were transferred into a new holding chamber containing HEPES-aCSF (containing, in mM: 92 NaCl, 20 NaHCO₃, 25 Glucose, 2.5 KCl, 1.25 NaH₂PO₄, 10 HEPES, 5 Na-ascorbate, 2 thiourea, 3 Na-pyruvate, 2 CaCl₂, 2 MgSO₄, and 12 N-acetyl-L-cysteine) under carbogenation. The slices were then transferred to the recording chamber with normal aCSF (containing, in mM: 125 NaCl, 2.5 KCl, 11 Glucose, 26 NaHCO₃, 1.25 NaH₂PO₄, 2 CaCl₂, and 2 MgCl₂) after 1-h recovery. To record optogenetically evoked EPSC (oEPSC) in M1 L5, following our previous report (Mu et al. 2020), we used a pipette solution containing the following (in mM): 130-mM K-gluconate, 10-mM KCl, 10-mM HEPES, 1 EGTA, 2-mM MgCl₂, 2-mM Na₂-ATP, and 0.4-mM Na₃-GTP; pH was adjusted to 7.3 with KOH and stimulating lights at 470 nm to activate ChR2 through a patterned illumination system (Polygon400 DSI-E-0470-0590-NK1 Dynamic Spatial Illuminator). The area of illumination was 60 × 100 μm in dimension and was the same in all experiments. Recorded neurons were voltage clamped at −70 mV and recordings were started after 2–3 min after a stable whole-cell configuration was obtained. In order to compare the dominance of terminals with L1 and L5, light with the same intensity and area was given separately. Access resistance (<20 MΩ) was not compensated and monitored continually throughout each experiment. Recordings were terminated whenever the input resistance increased >30% or access resistance exceeded 20 MΩ. Signals were acquired using a Molecular Devices MultiClamp 700B amplifier controlled by Clampex 10.4 software via a Digidata 1550 interface (Molecular Devices). Responses were filtered at 3 kHz, digitized at 10 kHz, and analyzed using Clampfit 10.4 (Molecular Devices).

Statistical Analysis

We assessed the fluorescence signals in 40-μm sections from five RV-injected Rbp4-Cre mice brains and seven retrograde tracer-injected C57/6 J mice brains to count cell numbers using ImageJ software. For the cells’ counting of each region or layer, we used the cell counter module in the ImageJ software and performed the manual cell count. Statistical analysis of tracing data was conducted in GraphPad Prism 8.0.2 (RRID: SCR_002798). Two-way ANOVA with Sidak’s multiple comparisons correction and one-way ANOVA with Tukey’s multiple comparison test were used for whole brain analysis and different layers’ distribution analysis. We calculated Pearson’s correlation coefficients to quantify the preference of the input patterns. Cubic spline was used to smooth the curve of distribution pattern across anterior to posterior. Wilcoxon matched-pairs signed rank test and one-way repeated measures ANOVA with Tukey’s multiple comparison test were used for the functional connectivity verification. All figures represent mean ± SEM. Significance differences were denoted by ^*P ≤ 0.05, ^**P ≤ 0.01, and ^***P ≤ 0.001.

Results

Mapping Monosynaptic Inputs Targeting Distinct Dendritic Compartments of Pyramidal Neurons

Our strategy in combining retrograde tracings and functional assessment to map the source of inputs that target the apical and basal dendrites of M1L5 neurons is summarized in Figure 1A. As shown in Figure 1B, in L5-specific Rbp4-cre mice, we first traced all the brain regions that send monosynaptic projections to M1L5 neurons via a modified RV AAV-ENVA-δG-mCherry with two cre-dependent AAV helpers, AAV-EF1a-DIO-EGFP-2δ-TVA-WPRE and AAV-EF1α-DIO-RVG-WPRE-pA (Wickersham et al. 2007; Guo et al. 2015). Only Rbp4-cre neurons express the EnvA-cognate receptor TVA and glycoprotein RVG. Subsequent infection of these neurons specifically by EnvA pseudo typed and RVG-deleted RV enables the RV to retrogradely proliferate into presynaptic cells. As neurons infected with the AAV helper encoding TVA protein express EGFP while RV-infected cells express mCherry, the starter cells can be identified by co-expression of EGFP and mCherry, while the upstream input neurons would express mCherry only. In another group of mice, we applied the retrograde tracers fast blue and chlorea toxin B 555 (CTB555) to M1 L1 and L5, respectively, to differentiate the inputs targeting the apical tuft dendrites and basal dendrites (Fig. 1C). Falsely labeled neurons due to uptake of these tracers by off-target synaptic terminals can be identified by comparing the results of the RV tracing. Finally, as illustrated in Figure 1A, following the introduction of anterograde AAV expressing ChR2 into specific upstream regions, optogenetic stimulations confined to either the distal or proximal dendritic arbors of M1L5 neurons were delivered to compare the strength of synaptic connections based on the evoked postsynaptic currents.

RV Tracing Reveals Monosynaptic Inputs to M1L5 Neurons from Multiple Brain Nuclei

In Rbp4-cre mice (n = 5), following recovery from microinjections of RV and helper viruses into M1 L5, all starter cells (105 ± 20, n = 5) were found to be restricted to M1 L5 (Supplementary Fig. 1A,B). Positive neurons expressing mCherry (Fig. 2A) could be found within M1 in L1, L2/3, and L5, in agreement with extensive local processing via inputs from interneurons in L1 and pyramidal neurons in other layers. Those outside M1 were found throughout the frontal association cortex, motor cortices, somatosensory cortices, thalamus, basal ganglia, and other subcortical regions mainly in the ipsilateral side (Fig. 2B). These neurons representing long-range inputs were quantified in both ipsilateral and contralateral sides of the brain according to the boundaries defined by the Allen Institute Mouse Brain Reference Atlas. The list of abbreviation of brain areas is shown in Supplementary Table 1. The vast majority of inputs to M1L5 pyramidal neurons originate from the frontal cortex (FrA and Lo), motor cortices (M1 and M2), somatosensory cortices (S1 and S2), other cortical plates (CC, AI, Ect, and Au), cortical subplates (CL and BLA), and various thalamic nuclei (including AM, VL, VM, Po, and PF), with the remaining inputs coming from discrete basal ganglia nuclei (GP and SNr) and a variety of other subcortical regions (including ZI, VTA, and DRN) in the ipsilateral side. Notably, the proportion of inputs from the ipsilateral motor cortices (38.0 ± 3.4%), somatosensory cortices (38.8 ± 3.5%), and thalamus (8.3 ± 0.6%) together account for more than 85% of all monosynaptic inputs (Fig. 2B). Comparatively, the contralateral brain has a much smaller contribution, which originates mainly from the motor cortices and a small extent from somatosensory cortices (Fig. 2A,B). It should be noted that, apart from pyramidal neurons, inputs from within the ipsilateral M1 also consist of local interneurons (Kubota et al. 2016).

We further analyzed the neuronal subpopulations within the motor and somatosensory cortices that send monosynaptic inputs to M1 L5. The analysis focused on the neurons in L2/3, L5, and L6 of the cortex (Supplementary Fig. 2A) covering M1, M2, S2, and S1 jaw (S1J), hind limb (S1HL), fore limb (S1FL), dysgranular zone (S1DZ), barrel field (S1BF), and upper lip region (S1ULp). We found that most input neurons were located in L2/3 and L5 rather than L6 in all ipsilateral and contralateral cortical regions examined (Supplementary Fig. 2B,C). However, while the contributions from L2/3 and L5 were similar in most of these cortical areas, there were some notable exceptions. In M1 and S1BF, the proportion of neurons labeled in L5 was significantly higher than those in L2/3, and this pattern was reversed in S1J (Supplementary Fig. 2B). In the contralateral side, only a higher proportion of neurons in L5 than L2/3 was found in S1HL (Supplementary Fig. 2C). Together, these findings indicate that L2/3 and L5 neurons from motor and somatosensory cortices are the major sources of monosynaptic inputs to M1L5 with differences in their relative contributions in some specific subregions.

Most Upstream Nuclei Target Both Distal and Proximal Dendrites of M1L5 Neurons

In the second series of experiments, after the deposition of fast blue in M1L1 and CTB555 in M1L5 (n = 7; Supplementary Fig. 3), neurons that send input to M1L1 would fluoresce in blue, while those input to M1L5 in red. Consistent with the RV retrograde tracing results, neurons labeled by these tracers were found in a large number of brain areas (Fig. 3A). However, some brain regions labeled by this method were negative in the RV tracing and might be considered false positive results probably due to absorption of the tracers by off-target fibers and terminals in L1 and L5. Therefore, only the upstream areas that were identified by both tracing methods were included in subsequent analyses.

To assess the relative contributions of different upstream areas targeting dendrites in L1 and L5, we counted the number of input neurons in each of the monosynaptically identified regions and expressed as percentages of the total number of labeled neurons in all these areas (Fig. 3A). Some notable patterns are observed. With respect to M1L1, there are five upstream regions that have the biggest contribution (>5%) and together account for more than 50% of all input neurons. These include the M2, S1HL, S1FL, S1BF, and S2. In the case of M1L5, except S1HL, the above areas and S1DZ constitute the highest proportion (>5% individually and >50% together) among all input neurons. In addition, in agreement with the RV tracing results, there are also clear inputs from other cortical areas and also a variety of thalamic areas such as VL, VM, and Po. These thalamic areas also have inputs to both dendritic arbors of M1L5 neurons. These observations mean that almost all major input regions innervate both dendritic arbors and a clear segregation of upstream brain areas targeting either compartment was not found.

To compare and analyze the distributions of inputs to M1L1 and M1L5 quantitatively, we computed the correlation coefficient between the inputs across all upstream regions mapped (Ogawa et al. 2014). In Figure 3B, each of the scattered circles represents a brain region and their relative inputs to L1 as well as L5. Overall, these input regions are highly correlated (0.926; P < 0.001), thus confirming a gross similarity in their distributions. Despite this, there are some upstream regions that exhibit a clear preference with respect to their dendritic targets. Notably, the EA, ZI, and EP innervate the dendrites in L1 only but not in L5, although the number of labeled neurons in these areas is relatively few. At the same time, there is a higher proportion of inputs from somatosensory and some other brain areas, namely S1HL, S1FL, Ect, VM, and BLA, that preferentially target L1 instead of L5 (Fig. 3A,B). On the other hand, there is a unique and significantly higher proportion of inputs from M2 that target L5 rather than L1 of M1. Consistent with the RV results, M2 is the main input from the contralateral cortex, which targets dendrites in both L1 and L5 (Supplementary Fig. 4A).

Dichotomous Inputs to M1L5 Neurons Mostly Originate from Different Subpopulations in Upstream Nuclei

We showed that there is no evidence that feedback areas, such as the cortical and thalamic sensory nuclei, target the apical dendrites specifically. Also, feedforward areas, such as higher motor areas and thalamic motor nuclei, also target both dendritic arbors. We next asked whether the dichotomous inputs targeting the two dendritic compartments originate from the same or distinct neural populations in the upstream nuclei. To answer this question, we analyzed the detailed distributions of fast blue and CTB555 single and double-labeled neurons in the main upstream regions. The co-expression of fast blue and CTB555 indicates that a neuron synapses on both dendritic arbors. Our analysis showed that, in the ipsilateral side, except the EA, ZI, and EP that innervate dendrites in L1 only (Fig. 4A), there were neurons in all other upstream areas, which were simultaneously labeled by both tracers. However, the percentage of these double-labeled neurons was in general small. As summarized in Figure 4B, the proportions of neurons in upstream nuclei that were labeled by the two tracers were all less than 15%. Among these regions, some cortical areas (FrA, M2, Cl, and DEn), thalamic areas (VM, Rh, and CM), and also DRN in the midbrain have a higher proportion of double-labeled neurons (>10%) compared with other areas. These results indicate that even within a single upstream nucleus, most neurons innervate either the distal or proximal dendrite of M1L5 neurons but not both.

To further decipher the detailed distributions of neuronal subpopulations that mediate these dichotomous inputs, we focused on the major upstream areas, namely the M2, S2, various S1 subregions, and also the thalamus. First, in agreement with the RV tracing results, most inputs from the cortical areas originate from neurons in L2/3, L5, and L6. In general, each of these layers in the cortical subregions examined (M2, S2, S1HL, S1FL, S1DZ, S1BF, and S1ULp) innervate both dendritic arbors of M1L5 neurons (Fig. 5A). However, within each layer, there are neurons that clearly target one of the two dendritic arbors only. The quantitative analysis shows that M2 L5 neurons have a significantly higher preference to dendrites in L5, while L2/3 neurons of S1HL and S1FL both have preference to dendrites in L1 (Fig. 5B). On the other hand, many L2/3 neurons in the contralateral side, which is dominated by M2, prefer dendrites in L1 (Supplementary Fig. 5A). The percentages of common neurons in the each layer of the ipsilateral and contralateral cortices are also low and are summarized in Supplementary Figure 5B,C. With respect to the thalamus, likewise, we found that the dichotomous inputs originate mostly from distinct neuronal subpopulations within the same nucleus, as shown by the representative pictures in Figure 5C. These include the most prominent thalamic sources, that is, VL, VM, and Po. Figure 5D illustrates that these neuronal subpopulations are often located in different dorsoventral or rostrocaudal territories of the thalamic nuclei.

Functional Assessment of Synaptic Connectivity Patterns

Our retrograde tracing studies have revealed a complex pattern of innervations to distal and proximal dendrites of M1L5 neurons. But there are factors that could not be revealed by the retrograde tracing that nevertheless could determine the vigor of innervation to the two dendritic arbors. These include the density of axons, number of synapses they make, and their strength. Therefore, we also assessed the synaptic strengths of pathways that show different preference to the two dendritic compartments by electrophysiological recordings, as illustrated in Figure 1. For upstream neurons that innervate preferentially the distal dendritic tuft in M1, we chose those confined in VM. As shown in Figure 6A, injection of the anterograde AAV5-hSyn-ChR2-mCherry in VM resulted in mCherry expression particularly prominent in L1 of M1, which is consistent with the retrograde tracing results as demonstrated in Figure 4B. A brief blue light stimulation elicited EPSCs (oEPSC) when delivered to targeted dendritic territories with respect to the recorded neurons in either L1 or L5. In response to the same duration and intensity of light pulses, on average, a much larger oEPSC was obtained when stimulating L1 (Fig. 6B,C). The evoked responses could be blocked by tetrodotoxin (TTX) and restored by 4-aminopyradine (4AP), suggesting that these were monosynaptic connections (Fig. 6D, Sugimura et al. 2016). Also, these oEPSCs were sensitive to the α-Amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) receptor blocker 6-cyano-7-nitroquinoxaline-2,3-dione (CNQX), indicative of their glutamatergic nature (Fig. 6E). The fact that both the rise time and decay time of the oEPSCS induced by L1 terminal stimulation were longer than those induced by L5 terminal stimulation (Fig. 6F,G) is consistent with the more distant location of the synapses.

With respect to upstream neurons that innervate preferentially the proximal basal dendrites of M1L5 neurons, we target L5 neurons in M2. Opposite patterns of terminal distributions and synaptic strength were observed. That is, mCherry expression was more prominent in L5 rather than in L1 in M1 (Fig. 6H). Consistently, the oEPSC was much bigger in amplitude when evoked in L5 than in L1 (Fig. 6I,J). Similarly, the monosynaptic and glutamatergic nature of the connections was demonstrated by the administration of TTX/4-AP and CNQX (Fig. 6K,L) although a significantly slower kinetics was found for the rise time but not the decay time (Fig. 6M,N). Finally, we also chose an upstream area that appears to send comparable innervations to M1L1 and M1L5, which is L5 of S1DZ. This was confirmed by the similar distributions of positive terminals in L1 and L5 after injection of anterograde viruses (Fig. 6O) and the comparable amplitudes of the oEPSCs (Fig. 6P,Q). Their monosynaptic and glutamatergic nature of the innervation and slower EPSC kinetics are shown in Figure 6R-U, respectively. Together, although we had not tested the inputs of all identified upstream nuclei, these electrophysiological experiments show that synaptic connection strengths match quite well with the innervation patterns we found. This finding strongly supports the reliability of the results revealed by the fast blue/CTB555 tracings.

Discussion

The main objective of the present study is to decipher the origins of neuronal populations that synapse on the two distinct dendritic arbors of L5 neurons in M1, to gain insight into the integrative functions of these output neurons. While each of the three techniques employed has its own limitation in addressing the question, combining the results from these approaches generates a consistent and therefore reliable picture. Specifically, by cross-checking against the results of RV retrograde tracing, false positive results in layer-specific chemical tracing by fast blue and CTB were eliminated. The different patterns of connectivity revealed by the tracing results were independently verified by assessing the functional connectivity via dendritic arbor-specific optogenetic stimulation. Together, our findings revealed a previously unknown but complex innervation pattern to M1 from a large number of cortical and subcortical areas. Deciphering the mode of operation or realistic modeling of the functions of the motor cortex should take these findings into consideration.

By and large, the upstream brain regions that we identified are in agreement with those reported in previous studies of mapping long-range inputs to M1 by RV (Luo et al. 2019) and fluorescent microbeads (Hooks et al. 2013). It should be pointed out that the use of Rbp4-Cre mice in our study helped in identifying those inputs that project to M1L5 pyramidal neurons only, while the traced nuclei in Luo et al. (2019) may synapse on excitatory or inhibitory neurons in M1 and across different layers. Our results confirmed that the majority of inputs to these neurons originate from various subfields of S1, S2, and M2, signifying the prominent and direct influence of a variety of somatosensory and premotor signals on the activities of M1. At the same time, motor thalamic nuclei like VL and VM as well as the higher sensory thalamic nuclei Po constitute major inputs from the thalamus. Furthermore, there are many brain regions, including the higher areas FrA, Lo, CLA, and subcortical nuclei in the basal ganglia and brainstem, also have direct, albeit more minor, monosynaptic innervation to the M1L5 neurons. Notably, we also found that the ZI, and the auditory but not visual cortex, have clear projections to these neurons, which had not been reported before and await further investigations of their significance. These various nuclei identified suggest that many descending and ascending pathways have direct access to M1L5 neurons and modulate their activities. Nevertheless, it should be pointed out that the small number of neurons we found in some subcortical areas such as GP and EA may represent scattered cholinergic neurons that constitute the nucleus basalis of Meynert, which would require further study and clarification.

An analysis of the input biases of all upstream areas reveals that a few nuclei, namely the EA, ZI, and EP, innervate the apical dendrites exclusively although the number of labeled neurons in these areas are much fewer than those in the cortex and thalamus. The most noticeable observation, however, is that most nuclei innervate both dendritic compartments. For example, FrA and Lo, most S1 subfields (e.g., S1FL, S1HL, and S1DZ) and S2 presumably conveying top-down or feedback signals innervate both apical and basal dendrites of these neurons although biased to the former. The only notable exception to this pattern is the input from M2, which has a strong bias to the basal dendrites in L5. This observation, in agreement with Hooks et al. (2013), suggests that the function mediated by M2 input is different in nature from most of other inputs. Specifically, this finding is in line with the notion that feedforward signal from premotor cortex may directly drive the firing of M1L5 neurons via targeting the more proximal, electrotonically closer basal dendrites. It is known that L5 pyramidal neurons participate in bidirectional connectivity between M1 and M2, and neurons projecting to M1 are mainly concentrated in L2/3 and L5 in M2 (Ueta et al. 2014). This is consistent with our tracing results in which we further showed that, in comparison to L2/3, M2L5 neurons exhibit a clear preference to innervate basal dendrites of M1L5 neurons.

The various motor thalamic nuclei (VL, VM, and VA) target both dendritic arbors of M1L5 neurons, and exhibit different biases to the apical and basal dendrites. In contrast to the notion that these inputs are mainly perisomatic (Hooks et al. 2013), we found a general bias to the apical dendrites. Also, our findings support previous reports that neurons in the VM send thalamocortical afferents more widely and more preferentially to layer 1 than neurons in VA and VL (Arbuthnott et al. 1990; Rubio-Garrido et al. 2009; Kuramoto et al. 2015). On the other hand, the projections from the sensory thalamus, especially Po, do not show a preference for either dendritic compartment of M1L5 neurons. We also assessed and verified the functional connectivity physiologically by optogenetic stimulation and recording of evoked synaptic currents, and confirmed that VM has a stronger synaptic influence at L1 than at L5. Since deep cerebellar nuclei and the basal ganglia modulate M1 through VL and VM, respectively (Nakamura et al. 2014; Tanaka et al. 2018), they may exert their effects via their differential innervation to the two dendritic arbors of M1L5 neurons.

It should be noted that not all target terminals took up the retrograde tracers, which would lead to errors in the quantitative estimation of the relative inputs to the apical and basal dendrites of M1L5 neurons. This problem was largely overcome by that the results were derived from the averaged data from seven mice in which the injection sites were verified to be within M1L5. Also, an important component in our strategy is to interrogate the connection strengths functionally via optogenetics and electrophysiological recordings. In the three upstream regions that we presented, namely VM, M2L5, and S1DZ L5, the oEPSC amplitudes generated by optogenetically stimulating the terminals targeting the apical or basal dendrites are in agreement with and therefore support the tracing results, namely L1 > L5 for VM, L5 > L1 for M2L5, and L1 ~ L5 for S1DZ L5, respectively. Given the long electrotonic distance between the site of stimulation at L1 and the recording at the soma, filtering of EPSC was inevitable. Unexpectedly, the prolongation of the rise time and decay time of these EPSCs did not appear to be large (no more than 50% in all cases). There might be some axons activated that innervated both dendritic arbors, reducing the differences in these parameters. Nevertheless, since the recording was made at the soma of M1L5 neurons, the oEPSC amplitudes represent a somatocentric view that reflects the relative impact of the distal and proximal synaptic inputs received at the soma (Petreanu et al. 2009).

Another level of complexity we detected in the connection patterns is that the groups of neurons in an upstream area that target these two dendritic compartments are mostly different, with only a small proportion of neurons innervating both. This observation can be found clearly in almost all upstream cortical and thalamic areas identified. Noteworthily, this pattern also applies to the individual output layers in the upstream cortical regions. For example, both L2/3 and L5 (the major output layers) of M2 and different S1 subfields contain distinct neurons that target either the apical or basal dendrites of M1L5 neurons. Our results thus provide complementary and more detailed information concerning the long-range input layer-specific projections to the motor cortex (Hooks et al. 2013). The clear division of labor from individual upstream region subdomains implies an accurate and fine control by these areas on M1L5 neurons, and strengthens the notion that the distal and proximal dendrites of these pyramidal neurons may serve different functions (Spratling 2002; Sjostrom and Hausser 2006; Branco et al. 2010; Li et al. 2017).

In recent years, neurocomputational models inspired by the spatially segregated input compartments architecture of real neurons, such as the pyramidal neurons, demonstrated an increased computational power and learning ability compared with single-input model (Urbanczik and Senn 2014; Guerguiev et al. 2017; Naud and Sprekeler 2018; Lillicrap et al. 2020). Also, this neuronal architecture could implement burst-dependent synaptic plasticity that in principle can coordinate learning via algorithms akin to back-propagation of errors (Payeur et al. 2021; Sun et al. 2021). An important point to note is that most of the identified upstream nuclei identified in the present study have themselves been shown to be reciprocally innervated by M1 projections, and also in a spatial location-specific manner (Jeong et al. 2016). Taken this into consideration, it is plausible that most brain areas reciprocally connected with M1L5 neurons convey feedforward and feedback signals simultaneously. These two streams of signal from an upstream region presumably target the two dendritic compartments differentially and from different subpopulations of neurons. This model of input connectivity contrasts the conventional model emphasized in sensory cortex, as summarized in Figure 7. Further investigations on the detailed reciprocal connection patterns between M1L5 and the different subpopulations of the upstream nuclei are needed to confirm and substantiate this model. Also, detailed knowledge of how the signals to the apical and basal dendrites from different inputs are integrated by M1L5 neurons, and how these neurons can store associations between different layers (Kording and Konig 2001; Naud and Sprekeler 2018; Lillicrap et al. 2020), are essential for understanding the significance of this cytoarchitecture in motor execution and motor learning. Nonetheless, our findings clearly indicate that the computations carried out by these neurons, attributed by the fine and intricate connections we revealed, are more complex than we currently understand.

In conclusion, our results do not support that apical and basal dendrites of M1L5 neurons receive signals from distinct nuclei. This contends against a clear segregation of inputs to M1 from clear upstream and downstream nuclei via these two dendritic compartments and implicates that M1 output neurons integrate incoming signals in a highly complex manner. In fact, it has been reported that even within the basal dendritic arbor of L5 pyramidal neurons, the properties of the more distal dendrites are different from those of the proximal thicker dendrites and may function independently (Nevian et al. 2007). Furthermore, our study did not distinguish the intratelencephalic (IT) type and pyramidal tract (PT) type L5 neurons of M1, which may receive inputs differently. Regardless, the functional connectivity information collected with respect to the dichotomous input to the two dendritic compartments of M1L5 neurons serve as a valuable reference for future studies. Further explorations should focus on how such a sophisticated synaptic connection pattern contributes to motor execution, motor learning, and other functions mediated by M1. Also, whether the picture drawn from M1L5 neurons is applicable to neurons in other layers in M1 and those in other cortical areas would be important questions to answer.

Funding

Hong Kong Research Grant Council General Research Fund (14110619, 14102221, 14167817).

Notes

We thank Ms Yue Gu for assisting histological assessment and Prof. Michael Hausser for a helpful discussion on some of the data. Conflict of Interest: The authors declare that they have no conflict of interest in this study.

References