Abstract

Corticocortical (CC) projections in the visual system facilitate hierarchical processing of sensory information. In addition to direct CC connections, indirect cortico-thalamo-cortical (CTC) pathways through the pulvinar nucleus of the thalamus can relay sensory signals and mediate cortical interactions according to behavioral demands. While the pulvinar connects extensively to the entire visual cortex, it is unknown whether transthalamic pathways link all cortical areas or whether they follow systematic organizational rules. Because mouse pulvinar neurons projecting to different areas are spatially intermingled, their input/output relationships have been difficult to characterize using traditional anatomical methods. To determine the organization of CTC circuits, we mapped the higher visual areas (HVAs) of male and female mice with intrinsic signal imaging and targeted five pulvinar→HVA pathways for projection-specific rabies tracing. We aligned postmortem cortical tissue to in vivo maps for precise quantification of the areas and cell types projecting to each pulvinar→HVA population. Layer 5 corticothalamic (L5CT) “driver” inputs to the pulvinar originate predominantly from primary visual cortex (V1), consistent with the CC hierarchy. L5CT inputs from lateral HVAs specifically avoid driving reciprocal connections, consistent with the “no-strong-loops” hypothesis. Conversely, layer 6 corticothalamic (L6CT) “modulator” inputs are distributed across areas and are biased toward reciprocal connections. Unlike previous studies in primates, we find that every HVA receives disynaptic input from the superior colliculus. CTC circuits in the pulvinar thus depend on both target HVA and input cell type, such that driving and modulating higher-order pathways follow complementary connection rules similar to those governing first-order CT circuits.

Significance Statement

Understanding the function of the pulvinar will require knowledge of its anatomical connections. Using state-of-the-art rabies tracing, we establish a map of brain-wide and CTC pulvinar connections. While the primate tectopulvinar pathway selectively targets dorsal visual areas, we find ubiquitous SC input in the mouse. This extrageniculate projection supports unconscious visually guided behavior, suggesting all visual cortical areas contribute to such functions in mice. Our results also unify long-standing anatomical hypotheses. Namely, “driver” CTC inputs are feedforward relays and adhere to the “no-strong-loops” hypothesis. “Modulator” L6CT inputs are overrepresented in the target HVA, reflecting the reciprocal connectivity described in previous bulk tracing studies. Together, these findings constitute a comprehensive map to guide future experimental and theoretical studies of pulvinar function.

Introduction

An animal’s perception and interactions with the external world rely on incoming sensory information. In mammals, the dorsal lateral geniculate nucleus (dLGN) relays information to the visual cortex, where a hierarchically organized network processes increasingly complex features. This hierarchy is defined by the laminar organization of direct CC projections (Felleman and Van Essen, 1991; Wang et al., 2012). However, the dLGN is not the only thalamocortical (TC) input. Instead, all visual areas receive projections from the pulvinar, a higher-order thalamic nucleus whose primary input is the cortex itself. This bidirectional connectivity could allow the pulvinar to route information between cortical areas indirectly (Sherman and Guillery, 2011). The pulvinar also relays ascending signals from the superior colliculus (SC) and retina (Lyon et al., 2010; Allen et al., 2016; N. Zhou et al., 2017; Beltramo and Scanziani, 2019). Without a comprehensive map of synaptic connections, however, it is unclear how these transthalamic circuits relate to the visual cortical hierarchy (Halassa and Sherman, 2019).

Viral and genetic tools make the mouse an ideal model for dissecting fine-scale input/output connections. Murine visual cortex includes V1 and higher visual areas (HVAs) with hierarchically organized feedforward and feedback connections (Wang and Burkhalter, 2007; Wang et al., 2012; Garrett et al., 2014; Harris et al., 2019). In rodents, individual pulvinar (or lateral posterior nucleus) neurons project to input layers of only 2–3 cortical areas (Nakamura et al., 2015). While pulvinar outputs are not completely independent, then, they do not broadcast indiscriminately (Juavinett et al., 2020; Blot et al., 2021). Therefore, TC projections could send distinct signals to HVAs depending on the organization of their corticothalamic (CT) inputs. CT axons are distributed topographically in the pulvinar, aligning roughly with the cell bodies of corresponding TC projections (Tohmi et al., 2014; Bennett et al., 2019; Juavinett et al., 2020). While coarse input/output organization is reciprocal, overlapping regions could still support interareal relay (Shipp, 2003; Sherman and Guillery, 2011). For a transthalamic pathway to transmit sensory information, however, CT neurons must be capable of driving activity. Layer 5 (L5CT) “drivers” are necessary for pulvinar responses and receptive field structure, whereas layer 6 (L6CT) “modulators” are not required for visual activity but can affect frequency and gain of responses (Sherman, 2016; Kirchgessner et al., 2021). If cortico-thalamo-cortical (CTC) circuits form reciprocal loops, however, the postsynaptic influences of L5CT and L6CT projections would have vastly different effects on circuit dynamics. The “no-strong-loops” hypothesis predicts that to avoid uncontrolled oscillations, L5CT drivers should not form reciprocal connections with TC projections (Crick and Koch, 1998). While this constraint holds true for first-order thalamic nuclei, it remains unproven in higher-order CTC circuits. Therefore, describing both the cortical regions and cell types connecting to each pulvinar→HVA population is essential.

Determining fine-scale CTC circuit organization requires projection-specific tracing of the relationships between pulvinar inputs and outputs (“TRIO”; Schwarz et al., 2015). Previous studies applied this technique to individual pulvino-cortical projections (Blot et al., 2021; Leow et al., 2022; Miller-Hansen and Sherman, 2022), but it is unclear whether those findings can be generalized. To reveal the organizational principles governing CTC connections across the visual hierarchy, we targeted pulvinar axons in each of five functionally mapped HVAs (LM, AL, RL, AM, and PM) for Cre-dependent, monosynaptic rabies tracing and quantified their brain-wide inputs. Consistent with previous reports of AL and PM projections, we find that V1 provides the largest source of driving L5CT input to every pulvinar→HVA pathway, confirming the long-standing theory that CTC connections are indirect feedforward relays (Sherman and Guillery, 2011; Blot et al., 2021). L5CT pathways from HVAs specifically avoid reciprocal connections, consistent with the “no-strong-loops” hypothesis (Crick and Koch, 1998). In contrast, L6CT inputs are distributed across V1 and HVAs, and they preferentially form reciprocal CTC loops. Driving and modulating pulvinar inputs thus form complementary antireciprocal and reciprocally biased pathways, respectively. Regardless of the target HVA, all pulvinar projections receive SC input. These results show that the pulvinar is a feedforward pathway between cortical areas that reflects the directionality of the existing cortical hierarchy and that each of these parallel pathways could integrate sensorimotor information from the SC.

Materials and Methods Animals

Thirty-three adult female and male C57BL/6J mice (The Jackson Laboratories) aged 6–10 weeks were used in this study. One mouse derived from a cross between Ntsr1-Cre GN220 GENSAT BAC transgenic and INTACT (R26R-CAG-loxp-stop-loxp-Sun1-sfGFP-Myc) mice was used to validate the L5 specificity of the Ctip2 antibody (Gong et al., 2007). All experimental procedures followed procedures approved by the Salk Institute Animal Care and Use Committee.

Viruses

AAV5-Ef1α-Cre-WPRE for retrograde Cre expression in pulvinar→HVA projections(2.22 × 1012 GC/ml; UNC Vector Core)

AAV8-CAG-FLEx-TCB (AAV-FLEx-TVA) for projection-specific expression of the avian TVA receptor with mCherry (2.04 × 1013 GC/ml; Salk GT3 Core)

AAV8-CAG-FLEx-oG-WPRE-SV40pA (AAV-FLEx-oG) for projection-specific expression of optimized rabies glycoprotein (1.55 × 1013 GC/ml; Vigene Biosciences)

EnvA + RVΔG-eGFP for transsynaptic labeling of pulvinar inputs with GFP (3.29 × 108 TU/ml; Salk GT3 Core)

To trace the inputs to specific projection populations, a mixture of AAV-FLEx-TVA and AAV-FLEx-oG was injected into the pulvinar. To prevent glycoprotein-independent retrograde labeling of CT neurons, the AAV-FLEx-TVA was diluted to between 8.16 × 1011 and 4.04 × 1012 GC/ml with Hanks buffered saline solution and then mixed in a 1:1 ratio with undiluted AAV-FLEx-oG. In control experiments with no AAV-FLEx-oG, higher concentrations of AAV-FLEx-TVA resulted in nonspecific RVΔG-eGFP expression in L5CT and L6CT cells at the AAV5-Cre injection site (Fig. 2). Since CT neurons in the target HVA expressed Cre, direct infection of their axons with AAV-FLEx-TVA permitted nonsynaptic uptake of RVΔG. After reducing the TVA concentration, however, RVΔG-eGFP+ neurons were no longer observed in the cortex or any other distal structures. We frequently saw a small number of GFP+ astrocytes at the AAV-Cre injection site, usually in upper cortical layers and with distinct morphology. These were present regardless of AAV-FLEx-TVA concentration and were not counted in the experimental brains. Control injections without Cre but including glycoprotein did not result in any long-range labeling, even at the higher TVA concentrations which failed the no-glycoprotein control. RVΔG-eGFP+ neurons were occasionally found in the hippocampus immediately dorsal to the pulvinar, presumably as a result of virus reflux along the pipette track; nevertheless, these were unlikely to be starter cells since they did not express Cre and did not result in labeled inputs in control injections.

Surgical procedures

Mice were anesthetized with isoflurane (2% induction; 1.5% maintenance) and mounted in a stereotax (David Kopf Instruments Model 940 series). Using a carbide burr, a small craniotomy was made over the left pulvinar (−1.8 to −2.3 from bregma; −1.2 to −1.6 lateral from the midline). Injection coordinates varied depending on the cortical projection region. The virus was pressure-injected via syringe at a rate of 10 nl/min through a tapered glass pipette (25–30 μm tip inner diameter) at a depth of 2.3–2.5 mm below the pial surface. To prevent backflow of the virus into the hippocampus and cortex, the pipette was left in place for at least 10 min before retraction. During the same surgery, a custom metal headframe was attached to the skull with dental cement (C&B Metabond, Parkell) at an angle approximately level with the surface of V1 (Juavinett et al., 2017). Either buprenorphine-SR (0.5–1 mg/kg) or meloxicam (4 mg/kg) were administered subcutaneously for post-op analgesia. Mice recovered for 1–4 d in their home cage with a suspension of ibuprofen (30 mg/kg) in water prior to intrinsic signal imaging (ISI).

After ISI mapping, blood vessel landmarks were used to target injections of AAV5-Ef1α-Cre-WPRE to one of six HVAs (PM, AM, RL, AL, LM, or POR). We chose AAV5 for its high propensity to infect long-range thalamocortical axons. In comparison, AAVretro-Cre caused marked toxicity and reduced cell labeling at the cortical injection site, and when diluted to nontoxic concentrations, its efficiency was comparable with AAV5-Cre. The virus was injected at two depths between 200 and 500 μm below the pial surface. The total volume of AAV-Cre ranged between 35 and 75 nl and was chosen based on the size of each HVA to maximize coverage without spreading into adjacent visual areas. The skull thinning required for ISI is appropriate for short survival times but leads to infection and tissue degradation after approximately a week. Due to the long survival time required for expression of oG, we removed the skull after thinning and placed a molded cement cover over the craniotomy to preserve the tissue health. A subset of animals was imaged without any skull thinning to guide injections and then subsequently reimaged with skull thinning for complete HVA mapping prior to euthanasia.

Three weeks after injection of AAV-Cre, 350 nl of EnvA + RVΔG-eGFP was injected into the pulvinar at the previous site at 2–3 depths between 2.3 and 2.6 mm below the pial surface. We rarely observed TVA expression along the previous pipette track, but to avoid infecting any cells in the hippocampus or cortex dorsal to the pulvinar, we injected rabies at a slight mediolateral angle. Buprenorphine-SR (0.5–1 mg/kg) was administered subcutaneously, and mice received ibuprofen in their water (30 mg/kg) as they recovered in their cages for 10 d until tissue harvesting.

Intrinsic signal imaging

After viral injections in the thalamus, HVAs in the cortex were identified using ISI through a thinned skull as previously described (Garrett et al., 2014; Juavinett et al., 2017). Mice were anesthetized with 0.1–1% isoflurane and sedated with (1 mg/kg) chlorprothixene hydrochloride injected intramuscularly. Anesthesia was continuously monitored and adjusted to maintain a lightly anesthetized state (1–1.5 breaths/second). Most experiments resulted in maps with seven visual areas (V1, PM, AM, RL, AL, LM, and P/POR) successfully segmented. These maps were used to guide cortical injections and to align postmortem tissue to the HVA borders. Occasionally, the initial map of visual cortex identified a subset of these areas. In these cases, the partial map was sufficient to target a cortical injection, and ISI was repeated just prior to the experimental endpoint to obtain a complete map that was suitable for postmortem tissue alignment. The border between P and POR was usually ambiguous, so these injections were targeted to the positive field sign area directly behind LM, which likely included both areas.

Exclusion criteria

Since the pulvinar is topographically organized with respect to its cortical inputs and outputs, the accuracy of targeting for both helper virus injection and the RVΔG injection was crucial to avoid biasing the measured input samples. Animals were excluded if the two injections did not overlap, in which cases we observed reduced starter cell populations and little to no transsynaptic labeling. We excluded animals in which starter cell locations were restricted or heavily biased to the tectal-recipient caudal pulvinar, as the corresponding cortical inputs were either absent or only present in caudally projecting area POR. We also excluded animals in which starter cells were biased compared with the overall distribution of starter cells for a given projection population, which aligned with the topographical organization that has been previously reported (Bennett et al., 2019; Juavinett et al., 2020). We excluded rare experiments in which surface blood vessels could not be accurately aligned. While we attempted to target P and POR as well, difficulty in mapping these areas with ISI prevented us from restricting AAV5-Cre to a single area. Due to this ambiguity in cortical target, these experiments were also excluded from analysis.

Histology

Mice were killed with an intraperitoneal injection of Euthasol (15.6 mg/ml) and then transcardially perfused with phosphate-buffered saline (PBS) followed by 4% paraformaldehyde (PFA). After dissection, brains were postfixed in 2% PFA and 15% sucrose in PBS for 16–24 h at 4°C and then submerged in 30% sucrose for an additional 16–24 h at 4°C.

The left visual cortex and hippocampus were dissected from the rest of the brain and sectioned tangentially on a freezing microtome. After the first 250 μm section which contained the surface blood vessels, the remaining cortical tissue was sectioned in 50 μm increments. Immunohistochemical staining of the free-floating cortical sections amplified the eGFP and mCherry signals and provided a laminar marker (Ctip2) to distinguish L5 from L6. After blocking at room temperature in PBS with 10% normal donkey serum (NDS) and 0.5% Triton X-100, tissue was incubated for 16 h at 4°C with goat anti-GFP (1:500; 600-101-215; Rockland Immunochemicals), rabbit anti-dsRed (1:500; 632496; Takara Bio USA), and rat anti-Ctip2 (1:1,000; ab18465; Abcam) in PBS with 1% NDS and 0.5% Triton X-100. The tissue was then incubated for 6 h at room temperature with donkey anti-goat conjugated to Alexa Fluor 488 (1:500; A-11055; Thermo Fisher Scientific), donkey anti-rabbit conjugated to Alexa Fluor 568 (1:500; A-10042; Thermo Fisher Scientific), and donkey anti-rat conjugated to Cy5 (1:500; 712-175-153, Jackson ImmunoResearch) in PBS with 1% NDS and 0.5% Triton X-100, followed by a counterstain with 10 μM DAPI in PBS. Tangential sections were mounted on gelatin-subbed slides, air-dried, and then coverslipped with polyvinyl alcohol mounting medium containing 1,4-diazabicyclo-octane (PVA-DABCO).

The remaining brain tissue was sectioned coronally on a freezing microtome in 50 μm increments from the olfactory bulb to the end of the cerebellum. Free-floating sections were blocked in PBS with 5% NDS and 1% Triton X-100 for 1 h at room temperature, then incubated with goat anti-GFP (1:500; 600-101-215; Rockland Immunochemicals) and rabbit anti-dsRed (1:500; 632496; Takara Bio USA) in PBS with 1% NDS and 0.5% Triton X-100 for 16–24 h at 4°C. The tissue was then incubated in the appropriate secondary antibodies for 2 h at room temperature and then counterstained with DAPI. Coronal sections were mounted on glass slides with PVA-DABCO.

Image processing and registration

All slides were imaged on an Olympus BX63 epifluorescence microscope with a 10× objective. Cells in coronal sections were counted manually and assigned to brain areas based on cytoarchitectonic landmarks by an expert or automatically registered to the Allen Common Coordinate Framework (CCF; Wang et al., 2020) after transforming sections using the SHARP-TRACK program (Shamash et al., 2018) and custom MATLAB scripts. Medial TRN cells were sometimes mistakenly assigned to VPL due to slight misalignments at the nuclei border. Therefore, those cells were manually checked and compared with cytoarchitectonic landmarks. The percentage of cells assigned to major brain areas was comparable between these two registration methods.

Z-stacks of tangential cortical sections were processed in the Olympus cellSens software using the Extended Focal Imaging method and then converted for additional processing using ImageJ software (NIH). Cortical sections were virtually “flattened” using radial blood vessel alignment (Kim et al., 2020), which corrected for the curvature of the cortex and aligned cells in deep layers with the HVA assignments defined at the pial surface. This alignment was achieved with sequential affine matrix transformations of deeper sections onto superficial sections using the Landmarks Registration plugin. When apical dendrites could be matched between images, they were included as landmarks in addition to radial blood vessels. The anatomical reference and HVA border map obtained during ISI were similarly warped onto the aligned image stack using common surface blood vessel landmarks present in the top tangential section. Cells in each transformed section were registered to the aligned HVA map and assigned to a cortical layer based on the borders of the Ctip2 signal in individual sections. The retinotopically organized, positive field sign patch designated as RL usually extended into the S1 barrel field as has been previously reported (Zhuang et al., 2017), so the anterior border was modified from the automatically segmented borders so that it aligned with the posterior edge of the imaged S1 barrel fields. All RL injection sites were still contained within the truncated borders. Area A lies in between RL and AM (Wang and Burkhalter, 2007; Garrett et al., 2014), but it was never segmented in our ISI. The medial border of RL was therefore estimated to lie along a line extending from the gamma barrel in between rows C and D to V1. The delineation of area A and AM relative to the posterior parietal cortex (PPC) is debated (Lyamzin and Benucci, 2019; Gilissen et al., 2021); in this study, we defined A as the region between RL and AM which is visually responsive and retinotopically organized. The area outside of these functional borders and posterior to somatosensory cortex was defined as PPC.

Input cells were frequently labeled in the nonvisual areas contained in tangential sections. To assign these areas, we used a 3D volume created from the registered coronal sections to define the edges of the dissected cortex in the CCF. This dissection boundary was then mapped onto the CCF flatmap, which was aligned to the cut edges of the tangential stack.

Visual areas P and POR were only partially mapped by intrinsic signal imaging due to these areas extending behind the lambda suture and, in the case of POR, laterally out of the imaging ROI. The anterior borders with LM and V1 were confirmed with ISI mapping, but the remaining boundaries of POR were defined by the transformed flatmap and the border of the entorhinal cortex which was evident from the cytoarchitecture. We observed a positive field sign band extending behind V1 as has been reported previously (Zhuang et al., 2017). In postmortem tissue, this region frequently had TVAmCh+ pulvinar axons in layer 4, confirming that this area is separate from V1, which only receives matrix-type projections from the pulvinar terminating in layers 1 and 5a (N. Zhou et al., 2018). Rather than defining the whole posterior edge of the cortex as area P, we used the most lateral edge of the dense DAPI band in RSPv layer 2/3 to delineate the border between P and RSP.

Analysis and statistical methods

Because the glycoprotein was not labeled with a fluorescent marker, we considered any GFP+ cell near the injection site to be a putative starter cell. All cells outside of this region were classified as inputs and expressed as a fraction of the total input population for each brain region (total inputs summarized in Table 1). Brain-wide inputs to each projection population were compared with a Kruskal–Wallis test, with post hoc Dunn–Šidák correction for significant comparisons.

Table 1.

Summary of input and starter cell counts

Table1-1

Brain-wide input counts for all animals from areas defined by the Allen Institute CCF. Download Table1-1, XLSX file.

To visualize the relative positions of SC inputs to different projection populations (Fig. 4C), cells were registered to the CCF using the forward transforms obtained from tissue registration described previously. For each target HVA, we pooled cell positions across animals without weighting so that brains with less efficient SC input labeling did not skew the distribution. These cells were divided into four 500 µm bins covering the anterior/posterior extent of SC and plotted with the atlas boundaries obtained at the center of each bin (−3.05, −3.55, −4.05, −4.55 mm from the bregma).

Benavidez et al. (2021) described a systematic organization of SC inputs and outputs that varies along the radial extent of SC. To determine whether the SC inputs to pulvinar share this trend for cortical topography, we defined an angle, theta (θ), representing cell positions along a polar axis where the ventral boundary of the SC bin is 0° and the midline is 90°. Probability density functions were fit to the theta values for the pooled, binned samples of SC cells using a kernel smoothing function with boundary correction at 0° and 90°.

L5CT and L6CT density maps were generated by binning the continuous distributions of cell positions for each animal in hexagonal bins with a width of 180.6 μm. Example brains for each target area are shown. Cartesian distances of input cells to the injected area were calculated from the transformed tissue. Distances of all cells from each target area were fit with a Gaussian kernel smoothing function to estimate the probability density.

A null distribution of cell distances was generated by shuffling the target area 10,000 times and recalculating the cell distances from the shuffled target locations. To test whether observations of L5CT or L6CT cells near the injection site were different than would be observed by chance, we compared the true proportion of cells within a 400 μm radius to the proportions calculated from the shuffled distributions. We calculated one-tailed p values for L5CT and L6CT to test for antireciprocity and reciprocity, respectively. p values were adjusted using the Benjamini–Hochberg procedure for control of false discovery rate. Any p values calculated as 0.0000 are reported as p < 1.0 × 10−4.

Outliers with large proportions of inputs in SC conversely had lower overall proportions of visual cortical input, so we normalized HVA inputs to the total visual cortical inputs for discretized area comparisons. Fractions of visual cortical input to each projection target were compared for L5CT and L6CT. To visualize the relative strengths of these HVA connections, we plotted the median input values in a heatmap. Color scales normalized for each target area (column), resulting in a visualization of the relative strengths of each input compared with the strongest input, in arbitrary units.

To test for significant input/output associations for L5CT connections, we performed a Fisher’s exact test on the raw cell counts with Monte Carlo simulation (9.6 × 107 shuffles) to estimate a p value ± 0.0001 with 95% confidence. L6CT sample sizes were much higher than L5CT, so a Pearson’s χ2 test for independence was applied to the L6CT input/output matrix. The shuffling procedure for cell distances was repeated for the cells’ HVA assignments to generate a null distribution of area counts to compare against projection-specific distributions.

ResultsBrain-wide mapping of pulvinar input/output connections

We characterized the brain-wide inputs to specific pulvinar→HVA pathways using projection-specific, Cre-dependent, monosynaptic rabies tracing (Fig. 1). After Cre-dependent AAV helper viruses (AAV8-Flex-TVA and AAV8-DIO-oG) were injected into the pulvinar of wild-type mice (n = 24, 4–5 per target HVA; Table 1), we used ISI to generate retinotopically defined borders of five HVAs (PM, AM, RL, AL, and LM; Fig. 1A,B). These functional maps guided precise, restricted injections of AAV5-Cre into each HVA, which retrogradely infected pulvinar neurons and drove Cre-dependent expression of the avian TVA receptor and rabies glycoprotein (oG; Fig. 1B). EnvA + RVΔG-eGFP was injected into the same location in pulvinar after 21 d. Animals were killed 10 d later to allow for sufficient transsynaptic rabies labeling. Because HVA borders vary between individual animals, we aligned each functionally defined ISI map to postmortem, tangentially sectioned, cortical tissue for quantification of GFP-expressing inputs (Fig. 1B,D). Noncortical brain areas were also sectioned and processed for quantification of brain-wide inputs (Fig. 1E).

Figure 1.

Targeting visual areas for pulvino-cortical input tracing. A, Intrinsic signal imaging and automatic segmentation of HVA borders aligned to in vivo blood vessels (animal RL2). B, Schematic of viral injection strategy for projection-specific input tracing. Top, AAV5-Cre injected into individual HVAs in wild-type mice retrogradely infects pulvinar projection neurons expressing Cre-dependent helper viruses. Pseudotyped, G-deleted rabies expressing eGFP is injected into the pulvinar to label presynaptic inputs to specific pulvino-cortical projection populations. Bottom, Visual cortex is dissected and sectioned tangentially to align tissue with the in vivo blood vessels and HVA map. The remainder of the tissue is sectioned coronally and processed to quantify brain-wide inputs. C, Models of cortico-thalamo-cortical connectivity. D, Tangential cortical sections aligned to the HVA map. Left, Surface blood vessel landmarks (yellow arrows) in postmortem tissue aligned with in vivo image. Right, GFP+ cortico-pulvinar inputs in deeper sections aligned with radial blood vessel landmarks. E, Putative starter cells in the pulvinar expressing TVAmCh and RVΔG-eGFP. All scale bars, 200 µm.

Cre-dependent rabies tracing in reciprocally connected brain areas requires careful selection of viral serotypes and concentrations to avoid both false-positive and false-negative labeling of input cells (Lavin et al., 2020). To optimize combinations of AAVs, we conducted numerous troubleshooting and control experiments (see Materials and Methods and Fig. 2). In our hands, cortical injections of AAVretro-Cre, commonly used for retrograde targeting (Tervo et al., 2016), resulted in pronounced tissue damage and complete absence of rabies-labeled CT neurons at the injection site. Lowering the concentration of AAVretro-Cre reduced cortical damage, but Cre expression in the pulvinar with sufficiently diluted AAVretro was no greater than with AAV5. To rule out Cre-independent rabies spread, we injected helper AAVs and EnvA + RVΔG-eGFP while omitting AAV5-Cre and confirmed that no cells outside of the immediate injection site expressed GFP. In reciprocally connected thalamocortical circuits, however, AAV-FLEx-TVA injections into the pulvinar can mediate retrograde infection of Cre-expressing CT neurons in the target area, which are then capable of expressing TVA at levels permitting direct uptake of EnvA + RVΔG from axon terminals in the pulvinar. Because this glycoprotein-independent labeling is facilitated by Cre expression, it is essential to conduct control experiments with AAV-Cre while omitting Cre-dependent AAV-oG. In initial tests, we frequently observed GFP+ CT neurons concentrated at the AAV-Cre injection site. These cells almost never had a detectable mCherry signal, underscoring that the lack of visible TVA expression is an inadequate quality control criterion. In a comparison of different helper virus serotypes, we found that AAV8 injections in the pulvinar mediated the least amount of retrograde uptake from CT neurons. Using two separate AAVs for TVA and oG rather than a bicistronic vector, we titrated the concentration of AAV8-FLEx-TVA until control injections without AAV8-DIO-oG resulted in negligible RV labeling in the target cortical area (Fig. 2).

Figure 2.

High titer AAV-FLEx-TVA permits retrograde, glycoprotein-independent rabies spread to Cre-expressing CT neurons. A, Schematic of viral injection strategy for control injections without glycoprotein. B, HVA-projecting pulvinar neurons express TVAmCh (red) and RVΔG-eGFP (green) for high (5.1 × 1012 GC/ml; left) and low (4.1 × 1011 GC/ml; right) titer AAV-FLEx-TVA injections. All scale bars, 200 µm. C, In the absence of glycoprotein expression, astrocytes (yellow arrows) express GFP with both high (left) and low (right) AAV-FLEx-TVA titer, but CT neurons (white arrows) express GFP only when AAV-FLEx-TVA titer is high. D, Number of GFP-labeled CT neurons (top) and astrocytes (bottom) at the AAV-Cre injection site with varying AAV-FLEx-TVA titer. Bars indicate mean cell counts. Open circles represent counts for individual animals.

Using the reagents and viral titers established in control experiments, we observed putative starter cells in the pulvinar consistent with the previously reported topography of HVA projections (Bennett et al., 2019; Juavinett et al., 2020). The starter cell coverage was sensitive to the volume and stereotaxic coordinates of helper virus injections, such that small or off-target injections led to biased labeling of starter cells and cortical inputs. To ensure that each HVA projection was adequately sampled, we injected a larger volume of helper virus that spread beyond the borders of the pulvinar (see Materials and Methods). We excluded off-target injections with incomplete coverage of the pulvinar compared with unbiased distributions (Bennett et al., 2019). Because the proportion of starter cells in the pulvinar was 82.0 ± 2.3% (mean ± SEM; Table 1), we consider the resulting inputs to largely reflect the connections of the pulvinar nucleus. However, a small proportion of starter cells were located in adjacent higher-order thalamic nuclei such as the laterodorsal (LD) and the posteromedial (POm) nuclei, which also project to HVAs. Individual PM and AM cases with high starter cell proportions in LD and POm did not have noticeably different input distributions compared with cases with starter cells more restricted to the pulvinar.

Figure 3 illustrates the input from all major brain regions to pulvinar neurons projecting to each of five HVAs expressed as a proportion of all rabies-labeled input neurons (see Extended Data Table 1-1 for complete area counts). Input from each of these areas did not depend on the cortical target of the traced pulvinar neurons (p > 0.05), except for SC (p = 6.9 × 10−3; Kruskal–Wallis test). Consistent with previous bulk retrograde tracing, we found that the largest source of input to the pulvinar pooled across projection groups was the cortex (41.4 ± 1.3% mean ± SEM), followed by the thalamic reticular nucleus (15.1 ± 0.9%, mean ± SEM), SC (11.0 ± 1.3%, mean ± SEM), and pretectal nuclei (10.8 ± 0.6%, mean ± SEM).

Figure 3.

Brain-wide inputs to the pulvinar are similar for different projection targets. A, Example coronal sections of Cre-dependent rabies labeling in two example experiments with transformed CCF atlas borders. Scale bars, 500 μm. Top, RVΔG-eGFP+ inputs (green) in CTX and TRN and TVAmCh+ pulvinar axons (red) in CTX, STR, and TRN in animal RL2. Bottom, Starter cells projecting to AL are labeled in the pulvinar with TVAmCh (red) and RVΔG-eGFP (green) in animal AL1. B, Brain-wide inputs to the pulvinar from major divisions of the brain, expressed as a fraction of the total inputs. Markers represent individual animals, colored according to the pulvinar projection target. Bar lengths represent the mean of the sample obtained by pooling projection targets together. For all areas except for SC, the distribution of inputs for different projection targets was not significantly different (p > 0.05; Kruskal–Wallis test). Black error bars represent the pooled mean ± SEM. For SC, the median input for each projection target is displayed by a vertical line. CTX, cortex; dLGN, dorsal subdivision of the lateral geniculate nucleus; HB, hindbrain; HY, hypothalamus (other); MB, midbrain (other); PTN, pretectal nuclei; Pulv, pulvinar; RSP, retrosplenial cortex; S1BF, primary somatosensory barrel field; SC, superior colliculus; STR, striatum; TH, thalamus (other); TRN, thalamic reticular nucleus; vLGN, ventral subdivision of the lateral geniculate nucleus; ZI, zona incerta.

The superior colliculus provides input to all projection populations

After the cortex, the SC is the second largest source of excitatory input to the pulvinar (Fig. 3B). With direct retinal input to its superficial layers, the SC forms an ascending visual pathway through the pulvinar that bypasses the dLGN and is conserved across mammalian species (Harting et al., 1973; Abramson and Chalupa, 1988; N. Zhou et al., 2017; Petry and Bickford, 2019). In macaques, this pathway projects to dorsal stream visual cortical areas (e.g., V3 and MT), but never ventral areas V2 and V4 (Lyon et al., 2010). Surprisingly, we find SC inputs to every pulvinar→HVA population in the mouse (proportion of total input from SC to pulvinar→LM: 23.9 ± 1.1%, AL: 11.7 ± 1.2%, RL: 6.9 ± 0.7%, AM: 6.4 ± 1.8%, PM: 10.3 ± 3.4% median ± median absolute deviations; Fig. 4A,B). LM-projecting pulvinar neurons receive a significantly larger proportion of their input from SC than do RL- and AM-projecting pulvinar neurons [p(LM vs RL) = 0.04, p(LM vs AM) = 0.006, p > 0.05 for all other pairwise comparisons, Kruskal–Wallis test with Dunn–Šidák post hoc correction]. These results are consistent with previous work showing that SC targets the caudal region of the pulvinar, which in turn projects to lateral extrastriate areas (P, POR, LM, and LI; N. Zhou et al., 2017, 2018; Beltramo and Scanziani, 2019; Bennett et al., 2019). We only included injections that covered the cortico-recipient pulvinar but did not require the presence of starter cells in the tecto-recipient caudal pulvinar. Because of these exclusion criteria (see Materials and Methods), SC→pulvinar→LM input in this study is underestimated.

Figure 4.

Superior colliculus targets all pulvino-cortical projection populations. A, SC input to PM- (magenta), AM- (purple), RL- (cyan), AL- (green), and LM-projecting (orange) pulvinar neurons. Bar lengths correspond to the median SC input expressed as a fraction of total inputs. Input to LM-projecting neurons was significantly greater than to AM-projecting (padj = 0.0067) and RL-projecting (padj = 0.0434) neurons (Kruskal–Wallis nonparametric rank test with Dunn–Šidák post hoc correction for multiple comparisons). All other pairwise comparisons were nonsignificant. B, Example coronal sections of GFP+ inputs in the SC from experiments targeting pulvinar projections to LM (left, LM1) and AM (right, AM5). Transformed CCF borders are superimposed. Scale bars, 500 μm. zo, zonal; sg, superficial gray; op, optic; ig, intermediate gray; iw, intermediate white; dg, deep gray; dw, deep white; PAG, periaqueductal gray. C, Top left, Coronal atlas of SC layers at −3.56 mm from the bregma, adapted from the Allen Mouse Brain Atlas (Wang et al., 2020). Polar coordinates define mediolateral position in the SC (Benavidez et al., 2021), where θ = 0° represents the ventral border of the SC and θ = 90° represents the midline. Top right, Smoothed point distribution of inputs labeled in the SC from all PM-projecting experiments combined. Middle, AM (left) and RL (right) input distributions. Bottom, AL (left) and LM (right) input distributions. D, Input from each SC layer, expressed as a fraction of total inputs. Markers correspond to an individual animal and are colored according to pulvinar projection target. Vertical bars indicate group medians. E, Estimated probability distribution function of θ for superficial and deep SC inputs to each pulvinar→HVA population.

Next, we examined the laminar distribution of SC inputs to each HVA-projecting population (Fig. 4C–E). Here, we find SC inputs in both superficial and deep lamina (Fig. 4B–D). Inputs, especially to AL-, RL-, and AM-projecting pulvinar neurons, were often found near the border of the optic (op) and intermediate gray (ig) layers, similar to a group of projections in squirrels described by Fredes et al. (2012) that target the ipsilateral rostral pulvinar. Overall, we find a low proportion of inputs in the sg layer compared with previous studies using nonspecific retrograde tracers in the pulvinar (Roth et al., 2016; N. Zhou et al., 2017; Scholl et al., 2021). Since the sg layer projects to the caudal pulvinar (or the inferior pulvinar in primates), this discrepancy is also best explained by incomplete spread of the helper AAVs in thalamus. (Harting et al., 1972; Benevento and Standage, 1983; Abramson and Chalupa, 1988; Stepniewska et al., 2000;