Feature-Specific Organization of Feedback Pathways in Mouse Visual Cortex

SUMMARY

Higher and lower cortical areas in the visual hierarchy are reciprocally connected [1]. Although much is known about how feedforward pathways shape receptive field properties of visual neurons, relatively little is known about the role of feedback pathways in visual processing. Feedback pathways are thought to carry top-down signals including information about context (e.g., figure-ground segmentation, surround suppression) [2–5], and feedback has been demonstrated to sharpen orientation tuning of neurons in the primary visual cortex (V1) [6, 7]. However, the response characteristics of feedback neurons themselves and how feedback shapes V1 neurons' tuning for other features such as spatial frequency (SF) remain largely unknown. Here, using a retrograde virus, targeted electrophysiological recordings and optogenetic manipulations, we show that putatively feedback neurons in layer 5 (hereafter ‘L5 feedback’) in higher visual areas, AL (anterolateral area) and PM (posteromedial area), display distinct visual properties in awake head-fixed mice. AL L5 feedback neurons prefer significantly lower SF (mean: 0.04 cycle per degree or cpd) compared to PM L5 feedback neurons (0.15 cpd). Importantly, silencing AL L5 feedback reduced visual responses of V1 neurons preferring low SF (mean change in firing rate: −8.0%), while silencing PM L5 feedback suppressed responses of high-SF preferring V1 neurons (−20.4%). These findings suggest that feedback connections from higher visual areas convey distinctly tuned visual inputs to V1 that serve to boost V1 neurons' responses to SF. Such like-to-like functional organization may represent an important feature of feedback pathways in sensory systems and in the nervous system in general.

RESULTS

Retrograde virus CAV2-Cre leads to Cre-recombinase expression specifically in feedback neurons

At least nine cortical areas that receive retinotopically-organized input from V1 have been identified in the mouse [8]. Superficial neurons in two areas, AL and PM, have been shown to possess distinct visual properties: AL neurons prefer relatively low SF while PM neurons prefer high SF [9–11]. Both areas send significant feedback to V1 [12, 13], however functional properties of these feedback neurons have not been investigated.

We sought to examine the visual properties of neurons in AL and PM providing feedback to V1 and their functional impact on V1 neurons. To this aim, we optogenetically manipulated AL and PM feedback neurons by using canine adenovirus type 2 (CAV2-Cre) to express Cre-recombinase in a retrograde manner [14]. By injecting CAV2-Cre in V1 in Cre-reporter mice [15], we induced tdTomato and opsin expression specifically in neurons projecting to V1.

To test Cre-driven protein expression based on CAV2-Cre, we co-injected CAV2-Cre with cholera toxin subunit B conjugated with Alexa Fluor 488 (CTB-488), a widely used retrograde tracer, into V1 in tdTomato reporter mice. Tangential sections show tdTomato (CAV2-Cre-driven, red) and CTB-488 (green) co-labelling in several areas, including lateromedial (LM), AL, PM, anteromedial (AM) and retrosplenial (RS) areas (Figure 1A). Coronal sections revealed that CAV2-Cre-labelled feedback neurons were found almost exclusively in L5 while CTB-488-labelled cells were found in L5 and superficial layers. Layer 5 neurons accounted for 78%, 91% and 92% of tdTomato-labelled neurons in LM, AL and PM, respectively. The proportions of co-labelled neurons of all tdTomato-labelled neurons in L5 were 80% for both AL and PM (Figure 1C, red). The proportions of co-labelled neurons of CTB-488-labelled neurons in L5 were 55% and 40% for AL and PM, respectively (Figure 1C, green). Thus, CAV2-Cre labels feedback neurons primarily in L5 in AL and PM, with lower density and greater lateral spread (Figure S1A) compared to CTB-488.

Figure 1. CAV2-Cre-mediated labelling of feedback neurons projecting to V1.

(A) Tangential section showing retrograde labelling of neurons in areas providing projections to V1 using CAV2-Cre (left) and CTB-488 (middle; right: merged). Area abbreviations: LI (laterointermiediate), LM (lateromedial), AL (anterolateral), P (posterior), S1 (somatosensory), AM (anteromedial), PM (posteromedial), RS (retrosplenial).

(B) Coronal sections showing laminar distribution of feedback neurons.

(C) Percentage of CTB-488 co-labelling in CAV2-Cre+ cells (red) and that of CAV2-Cre co-labelling in CTB-488+ cells (green). Top: all layers. Bottom: L5 only. Error bars: mean ± SEM, in 3 mice. See also Figure S1.

AL vs. PM feedback neurons display distinct spatial frequency preferences

To characterize visual properties of feedback neurons, we used a channelrhodopsin-2 (ChR2) guided targeted recording method, adopted from [16]. In Ai9 and Ai32 double-reporter mice, injection of CAV2-Cre into V1 led to the expression of both tdTomato and ChR2 in feedback neurons in higher visual areas (Figure 2A). AL and PM were located using tdTomato signal through intact skull in vivo (Figure S1B) and craniotomies were made to allow simultaneous laser delivery and unit recordings in AL and PM, in awake head-fixed mice (Figure S2A–B) [17].

Figure 2. Channelrhodopsin-2 guided targeted recording of AL and PM feedback neurons in vivo.

(A) Experimental approach.

(B) Laser responses of feedback (blue) vs. neighbor (magenta) neurons in terms of change in firing rate during laser and latency of first spike during laser period. Some cells could not be classified (yellow). AL: circles, PM: triangles, filled: cells with spiking during laser, open: cells without spiking during laser period (no response).

(C–F) Example AL feedback neuron. (C) Response to laser. (D) Spatial frequency (SF) turning curve. PrefSF: preferred SF, based on Gaussian fit (red). (E) Spike time histogram and raster plot during 0.01-cpd grating. (F) Temporal frequency (TF) tuning curve. Gray lines: baseline firing rate (solid: mean, broken: mean ± SEM).

(G–J) Example PM feedback neuron. Same layout as in C–F.

(K–M) Visual properties of feedback vs. neighbor neurons in AL (red circles) and PM (cyan triangles). (K) Peak SF, (L) Peak TF, (M) Linearity. Error bars represent mean ± SEM, **p<0.01. See also Figure S2.

Laser stimulation robustly activated neurons in AL and PM (Figures 2C, 2G). Our criteria for feedback cells were neurons with robust and sustained laser responses with immediate increases in spiking that lasted throughout the laser period. These criteria were based on previous evidence [16] showing that only ChR2-expressing neurons respond to photo-excitation with sustained increases in firing rate. Neurons that were recorded at similar depths but did not show sustained responses were categorized as neighbor cells. As we could not definitively rule out polysynaptic activation, we refer to the first group as 'putatively feedback' (hereafter ‘feedback’), and the second group as 'neighbor'. Photo-excitation distinctly affected feedback and neighbor cells (Figures 2B, S2C–D). Results described in Figure 2 and S2 are from L5 neurons (recording depth: 420 – 570 µm).

Figure 2C–F shows an example AL feedback neuron that preferred low SF (0.01 cpd), high temporal frequency (TF, 8Hz) and displayed significant modulation by the grating frequency (F1 modulation; Figure 2E). In contrast, an example PM feedback neuron (Figure 2G–J) preferred high SF (0.16 cpd), low TF (1 Hz) and showed little to no F1 modulation.

Visual properties of AL and PM feedback and neighbor cells were analyzed using two-way ANOVAs with cell group (feedback vs. neighbor) and visual area (AL vs. PM) as factors (see also Table 1). The analyses revealed a highly statistically significant effect (p = 0.0005) of visual area on peak SF with significant differences between AL and PM feedback neurons (post-test: p = 0.0054, Figures 2K, S2E). Visual area had significant effects on linearity (F1/F0; p = 0.0140; Figure 2M), impact of running on visual responses (p = 0.0216), and there was a trend for an effect on peak TF (p = 0.0755; Figure 2L). Cell group was not a statistically significant factor in any property examined (p > 0.05). Fisher's Exact Test revealed that the null hypothesis of homogeneity between FB neurons vs. neighbors could not be rejected in terms of their peak SF (AL: p = 0.65; PM: p = 1.00).

Table 1.

Summary of Electrophysiological Data (corresponds to Figure 2)

Feedback neurons	AL	PM
Peak SF (cpd)	0.04 ± 0.01, n=13	0.15 ± 0.03, n=10
Preferred SF (cpd)	0.05 ± 0.01, n=13	0.15 ± 0.03, n=10
Peak TF (Hz)	3.4 ± 0.8, n=14	1.5 ± 0.6, n=11
Preferred TF (Hz)	3.3 ± 0.8, n=14	1.8 ± 0.7, n=11
OSI	0.77 ± 0.07, n=12	0.53 ± 0.09, n=10
DSI	0.46 ± 0.10, n=12	0.56 ± 0.11, n=10
Linearity (F1/F0)	0.90 ± 0.12, n=12	0.68 ± 0.06, n=10
Baseline firing rate (Hz)	2.4 ± 0.9, n=12	5.3 ± 1.8, n=8
Maximum visually evoked firing rate (Hz)	8.9 ± 2.2, n=12	13.0 ± 3.2, n=8
Running effect on max. firing rate (% change)	106 ± 26, n=10	29 ± 6, n=5
Neighbor neurons	AL	PM
Peak SF (cpd)	0.08 ± 0.03, n=10	0.16 ± 0.02, n=10
Preferred SF (cpd)	0.08 ± 0.03, n=10	0.15 ± 0.02, n=10
Peak TF (Hz)	3.4 ± 1.4, n=10	1.9 ± 0.5, n=13
Preferred TF (Hz)	2.5 ± 0.7, n=10	1.7 ± 0.5, n=13
OSI	0.64 ± 0.13, n=8	0.81 ± 0.08, n=12
DSI	0.55 ± 0.16, n=8	0.48 ± 0.09, n=12
Linearity (F1/F0)	1.05 ± 0.13, n=10	0.69 ± 0.09, n=10
Baseline firing rate (Hz)	4.2 ± 1.5, n=10	6.4 ± 1.9, n=10
Maximum visually evoked firing rate (Hz)	14.5 ± 4.0, n=10	13.8 ± 2.2, n=10
Running effect on max. firing rate (% change)	65 ± 27, n=7	16 ± 21, n=9

Taken together, we found that AL L5 feedback neurons preferred lower-SF gratings compared to PM L5 feedback neurons. In general, AL L5 neurons preferred lower-SF gratings and their responses were more strongly influenced by F1 modulation and running, compared to PM L5 neurons.

Like-to-like functional impact of AL vs. PM feedback silencing on V1 neurons

Previous approaches in examining the role of feedback have employed silencing methods such as cooling and optogenetic interneuron activation [3–6, 18] that inactivate the entire area, silencing not only feedback neurons projecting to V1 but also neurons that project to other areas. Therefore, the effects observed in these studies may not be specific to feedback effects in V1.

We designed two approaches to selectively silence AL or PM neurons providing feedback to V1. In Approach 1, CAV2-Cre was injected into V1 in tdTomato mice to localize AL, then a second virus AAV-FLEX-ArchT was injected into AL to express the inhibitory opsin, archaerhodopsin-T (ArchT), specifically in AL feedback neurons (Figure 3A). In Approach 2, we injected CAV2-Cre into V1 in Ai9 and Ai39 double-reporter mice, leading to tdTomato and halorhodopsin (HaloR) expression in neurons projecting to V1 (Figure 3B). For AL feedback silencing, we used both approaches. For PM feedback silencing, we used Approach 2 only. AL feedback silencing results using the two approaches did not differ significantly (2-way ANOVA, effect of silencing method, p = 0.1731, see also Figure 3I), thus data were pooled unless otherwise noted. Figure 3C–E shows the results of a control experiment where direct effects of laser were tested on putatively feedback neurons in PM using Approach 2. Spiking including visually evoked activity was completely silenced in all putatively feedback neurons tested (4 of 4 neurons). Using these approaches, we silenced feedback neurons in either AL or PM while simultaneously recording from V1 neurons.

Figure 3. Functional impact of silencing AL vs. PM feedback on V1 neurons.

(A–B) Experimental approaches 1 (A) and 2 (B). For AL feedback silencing, both approaches were used. For PM feedback silencing, approach 2 was used.

(C–E) Control experiment showing direct silencing effects on putatively feedback PM neurons (4 cells). (C) Normalized firing rates in laser-off (black) vs. laser-on (green) trials. Solid: mean, broken: mean ± SEM. Change in firing rate (%) during laser as a function of time in 250-ms bins (D) and laser power (E).

(F–I) Effects of silencing AL feedback. (F) SF tuning curve of a V1 neuron preferring 0.04 cpd with/out AL feedback silencing (black: control, green: laser). Raster plot (G) and spike time histogram (H) of the cell’s response to 0.04-cpd gratings with/out laser. (I) Change in firing rate for preferred stimulus during AL feedback silencing for V1 neurons preferring low vs. high SF. Blue: results using approach 1. Black: results using approach 2. Bold: neurons with statistically significant laser modulation. Numbers in parentheses: number of neurons. (J–M) Effects of silencing PM feedback. Same layout as F–M.

For C–F and J, error bars: mean ± SEM; ***p<0.001, **p<0.01, *p<0.05, ^ns p>0.05. See also Figures S3–4.

AL feedback silencing significantly reduced visual responses of V1 neurons preferring low SF (0.02–0.04 cpd) by 8.0% on average (paired two-sample t-test, p = 0.0068; Figure 3I, left), whereas V1 neurons preferring higher SF (0.08–0.32 cpd) failed to show a consistent effect as a group (p = 0.4614; Figure 3I, right). Figure 3F–H shows an example V1 neuron whose response to its preferred low-SF (0.04 cpd) grating was reduced by silencing AL feedback. Silencing AL feedback exerted a greater effect in the second half of the visual response (Figure 3H), a tendency that was observed in most V1 neurons showing statistically significant effects of AL feedback silencing (Figure S3). Silencing AL feedback also led to statistically significant (p < 0.05) enhancement of visual responses in two high-SF preferring V1 neurons (Figures 3I, S3–4).

In contrast to AL feedback effects, PM feedback silencing markedly suppressed responses of high-SF (0.08–0.32 cpd) preferring V1 neurons by 20.4% on average (paired two-sample test, p = 0.0147; Figure 3M, right), while low-SF preferring V1 neurons showed little to no effect of this manipulation (p = 0.2916; Figure 3M, left). Figure 3J–L illustrates an example V1 neuron whose response to its preferred high-SF (0.16 cpd) grating was reduced by PM feedback silencing. The impact of PM feedback silencing was immediate and lasted throughout the visual stimulus duration (Figure 3L), a trend that was observed in most neurons with significant effects of PM feedback silencing (Figure S3). A two-way ANOVA with cell group (low- vs. high-SF preferring cells) and silencing area (AL and PM) as factors revealed a significant interaction (F(1, 74) = 5.446, p = 0.0223) and post-tests indicated that the effects of AL vs. PM feedback silencing were statistically significantly different (p = 0.01) among high-SF preferring V1 cells. Therefore, our findings indicate that AL and PM feedback pathways exert distinct effects on high-SF preferring V1 neurons.

To characterize the specificity of feedback effects, we analyzed the effects on V1 neurons' activity during preferred vs. non-preferred gratings (Figures 3F, 3J and S4). V1 neurons showed significantly greater effects during presentation of preferred stimuli compared to non-preferred stimuli.

We found that AL L5 feedback neurons prefer low-SF and they enhance V1 neurons' responses to low-SF, whereas PM L5 feedback neurons prefer high-SF and they amplify V1 neurons' responses to high-SF stimuli. Taken together, our findings suggest that AL and PM feedback distinctly affect responses of V1 neurons, consistent with like-to-like organization.

DISCUSSION

Using a novel retrograde virus and optogenetic techniques in awake mice, we were able to selectively record and manipulate the activity of layer 5 neurons in higher visual areas providing feedback to V1. We found that AL L5 feedback neurons prefer significantly lower SF compared to PM L5 feedback neurons. Furthermore, we found that the functional impact of feedback is like-to-like: low-SF preferring V1 neurons' responses are enhanced by AL feedback while high-SF preferring V1 neurons' activity is amplified by PM feedback.

Our findings are similar to previous reports of distinct SF and TF preferences by AL and PM L2/3 neurons [9–11]. To our knowledge, our study is the first to report these differences between AL and PM using electrophysiological techniques which provide a finer temporal resolution compared to calcium imaging. Our data further demonstrate that AL and PM display these differences specifically in L5 feedback neurons. We also found other previously unreported differences: AL L5 neurons were more strongly modulated by the phase of gratings and by spontaneous running of the animal, compared to PM L5 neurons. These findings suggest that AL and PM may play distinct roles in visual processing and behavioral-state modulation of visual activity.

We found that silencing AL L5 feedback reduced visual responses in low-SF preferring V1 neurons, whereas silencing PM L5 feedback suppressed activity in high-SF preferring V1 neurons, suggesting like-to-like functional impact of feedback. This points to the possibility that different V1 populations may be under distinct feedback influences.

Since CAV2-Cre labelled primarily L5 neurons, our findings are limited to feedback neurons in L5. It is possible that the impact of L2/3 feedback may be different, as L5 neurons preferentially receive L5 feedback while L2/3 neurons preferentially receive L2/3 feedback [19–20]. Effects of feedback may also be different if a higher proportion of feedback neurons could be silenced. Viral methods targeting L2/3 and with higher efficacy should be used to address these questions in future studies. It is also possible that feedback neurons in nearby cortical areas may have been partially inactivated in silencing experiments using the tdTomato-HaloR reporter approach. However, our dual-virus approach allowed us to spatially restrict ArchT expression to AL feedback neurons, and we note that the results were similar between the two approaches. Future studies should also examine the effects of AL and PM feedback on other properties such as orientation, TF tuning and surround suppression, given that mouse V1 receptive fields display inhibitory surround mechanisms [3, 4, 21, 22].

Our findings provide several new insights into the role of feedback pathways in vision. Similar to the properties of feedforward inputs from V1 to AL and PM [23], we demonstrate that feedback neurons from these areas projecting to V1 also display distinct visual properties. Our findings indicate that feedback from these areas have like-to-like functional impact on V1, consistent with stream-specific organizational schemes that have been demonstrated for cortico-geniculate feedback pathways in monkeys [24–26]. Thus, these schemes may exist at multiple levels and they may support feature-specific top-down processes such as selective attention [27, 28]. Synaptic mechanisms underlying these effects remain unknown. Our finding that PM feedback effects are immediate while AL feedback effects accumulate more slowly suggests that some feedback effects may be mediated by polysynaptic circuits. Considering emerging evidence that mouse V1 may be organized into clusters of neurons tuned to similar visual features [29, 30], our findings lend further credence to the intriguing possibility that visual information may be processed by multiple distinct streams in both feedforward and feedback directions in the visual system.

STAR METHODS

CONTACT FOR REAGENT AND RESOURCE SHARING

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Carey Y. L. Huh (careyhuh@gmail.com).

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Animals

The following Cre-dependent reporter mice were obtained from the Jackson Laboratories: tdTomato reporter, Ai9 (B6.Cg-Gt(ROSA)26Sor^{tm9(CAG-tdTomato)Hze}/J, stock number: 7905), tdTomato reporter, Ai14 (B6;129S6-Gt(ROSA)26Sor^{tm14(CAG-tdTomato)Hze}/J, stock number: 7908), channelrhodopsin-2 reporter, Ai32 (B6;129S-Gt(ROSA)26Sor^{tm32(CAG-COP4*H134R/EYFP)Hze}/J, stock number: 12569), halorhodopsin reporter, Ai39 (B6;129S-Gt(ROSA)26Sor^{tm39(CAG-hop/EYFP)Hze}/J, stock number: 14539). Homozygous transgenic mice were bred to wildtype C57BL/6J mice to generate heterozygous offspring, or bred together to generate double-reporter mice that are heterozygous for two transgenes. Experiments were performed on heterozygous mice (both female and male) that were 34–190 days old. Typically, animals were group-housed (separated by sex, maximum of five mice per cage) under standard food, temperature, light and cage conditions. Animals for in vivo recordings were group-housed until headplate surgery. After headplate surgery, they were placed in single-housing. All surgical procedures followed animal care guidelines approved by Stanford University’s Administrative Panel on Laboratory Animal Care (APLAC). For all surgeries, body temperature was maintained at ~37.5°C by a feedback-controlled heating pad and eyes were covered with opthalmic ointment to prevent drying.

METHOD DETAILS

Viral procedures

Canine adenovirus type 2 (CAV2-Cre) virus was generated by E. J. Kremer and provided by the Montpellier vector core (PVM). The final titre was 1.25 × 10¹² viral particles per ml. AAV1-FLEX-ArchT-GFP was generated and provided by the University of North Carolina at Chapel Hill virus vector core. The final titre was 1.2 × 10¹³ viral particles per ml. All viral procedures and handling of virus followed the Biosafety Guidelines approved by Stanford University’s APLAC and Administrative Panel of Biosafety (APB) for biosafety level 2.

Injection procedure

Mice were anesthetized with ketamine (50 mg/kg), dexmedetomidin (3.75 µg/kg) and isoflurane (4% for induction, 1.5–2% for maintenance) and placed in a stereotaxic frame. Scalp was retracted and a small burr hole was made at the injection site using a pneumatic drill. The coordinates used for targeting V1 were 2600–2700 µm lateral from midline suture, 300–500 µm anterior to the anterior border of the transverse sinus, on the left hemisphere. Virus and/or cholera toxin subunit B conjugate (see below for volume used for specific experiments) were loaded into a glass pipette beveled at the tip (tip inner diameter: 20 µm; Cat. # 5-000-2005, Drummond) and injected at the speed of 50 nl/min using a microsyringe pump controller (UltraMicroPump III, WPI). Injections were performed at two depths: 600 and 300 µm below brain surface. Bone wax was applied on the burr hole and skin was sutured closed. After surgery, mice were injected with buprenorphine (0.05 mg/kg) and warmed saline (0.9% sodium chloride) and monitored for post-operative health.

Anatomy experiments

Using the procedure above, tdTomato reporter (Ai9 or Ai14) mice were injected with CAV2-Cre virus and 0.1% (w/v, in saline) cholera toxin subunit B conjugated with Alexa Fluor 488 (CTB-488, Cat. # C-34775, Life Technologies) into a single site in monocular V1 on the left hemisphere. CAV2-Cre and CTB-488 were mixed together and loaded into the same injection pipette. For flatmounted cortex, a total of 100 nl of virus and 1µl of CTB-488 were injected per animal; for coronal sections, 50 nl of virus and 450 nl of CTB-488 were injected per animal. The brain was harvested 14–20 days after injection.

For flatmounted cortex, the animal was anesthetized with isoflurane, decapitated and the brain was extracted, hemisected and fixed in 4% paraformaldehyde (PFA) solution for 10–30 min. Then, cortex on the left hemisphere was separated from the rest of the brain with microspatulas, flattened between two glass coverslips and fixed overnight at 4°C. The flattened cortex was trimmed, embedded in agar and sectioned in 150 µm in PBS using a vibratome (VT1000S, Leica). Tangential sections were mounted on slides and imaged using a 4X objective on a fluorescence microscope (Axioskop, Zeiss). Images were stitched together using a custom software based on Panorama tools (panotools.sourceforge.net).

For coronal sections, the brain was extracted, hemisected and fixed in 4% PFA solution overnight. The brain was sectioned coronally in 50 µm using a vibratome. Sections were mounted on slides using DAPI-containing fluoromount (Fluoromount-G, SouthernBiotech). Images were taken using a 10X objective on a confocal microscope (A1 confocal laser system, Nikon). Cells were counted using the cell counter plugin in Fiji [31]. Cortical layers were identified using DAPI staining (not shown) following previously used criteria [32]. CAV2-Cre-tdTomato displayed a greater lateral spread in neurons labelled, thus the analysis was restricted to locations with spatial overlap between the two populations (200–300 µm window). CAV2-Cre-tdTomato+, CTB-488+ and co-labelled neurons were counted in three sections per animal, in 3 mice.

In vivo electrophysiology and optogenetic manipulations

Time intervals between virus injection and recordings were as follows. For ChR2 experiments, recordings were performed 42–72 days after the CAV2-Cre virus injection. For feedback silencing experiments, for the first approach, CAV2-Cre was injected first, followed by injection of the AAV virus 2–3 weeks later. Recordings were performed 21–43 days after AAV injection for the first approach and 44–81 days after CAV2-Cre injection for the second approach.

All recordings were performed in awake head-fixed mice, free to run on the spherical treadmill (floating styrofoam ball). Mice were given at least three days to recover from the headplate surgery before habituation. For habituation, mice were placed on a spherical treadmill on the recording setup and head-fixed for 0.5–1 hr each day, typically for 3 days. Recordings were performed from the same craniotomy for 1–3 days. After the first day of recording, the craniotomy was covered with silicone (Kwik-sil adhesive, WPI), which was removed before the next day’s recording. Mice were given at least three hours to recover from the craniotomy surgery before recording.

Unless otherwise noted, juxtacellular recordings were obtained using standard blind loose-patch techniques, following previously reported procedures [17]. Glass electrodes (4–8 MΩ) were filled with aCSF composed of the following (in mM): 130 NaCl, 24 NaHCO₃, 10 glucose, 3.5 KCl, 1.5 MgCl, 1.25 NaH₂PO₄ and 2.5 CaCl₂ (pH 7.3). The pipette with positive pressure was lowered into the brain under visual guidance and advanced along a 55-degree axis in short 1–2 µm pulses. When an abrupt increase in the pipette resistance was observed, the positive pressure was released to achieve a tight seal with the cell. Recordings were performed with a Multiclamp 700B amplifier (Molecular Devices) and digitized by ITC-18 (InstruTECH). Data acquisition was controlled by custom software written in Igor Pro (Wavemetrics). After the pipette was lowered below the dura, agarose solution (low melting-point agarose, 2% in PBS) was applied to the well of the headplate and allowed to harden for increased stability.

For ChR2-guided targeted recordings, homozygous Ai9 and Ai32 mice were bred together to generate double-reporter mice that express tdTomato and ChR2 under the control of Cre-recombinase. Following the injection procedure described above, CAV2-Cre virus (60–70 nl) was injected into V1 on the left hemisphere. Two or more weeks after the virus injection, the mouse was re-anesthetized, placed in a stereotaxic frame, and skull was exposed to visualize tdTomato fluorescence using a tandem-lens epifluorescence macroscope [33]. The locations of AL and PM, identified using the fluorescent signal and putative locations based on previous studies [8], were marked on the skull. A metal headplate was centered over AL and PM and fixed to the skull using opaque dental cement (C&B Metabond, Parkell). On the day of recording, a craniotomy of ~400 µm in diameter was made over AL or PM based on the skull markings, using a pneumatic drill. For optogenetic stimulation, an optical fiber, emitting a circular laser spot of 100 µm in diameter, was encased in a glass pipette painted with black opaque paint and lowered to the depth of 0–100 µm, relative to brain surface. The recording pipette was lowered into the same craniotomy as the optical fiber. For optogenetic stimulation, laser was delivered (520 nm, power: 0.6–2.6 mW) for 2 s every 8 s. For recording of putatively feedback neurons, the recording pipette was moved in small (~5 µm) steps until a unit clearly responding to laser was encountered and recorded. The presence of a unit was monitored using an audio monitor. For recording of neighboring neurons, units were recorded in loose-patch mode in similar depths as feedback neurons and tested for laser responses. Putatively feedback and neighbor cells were recorded at depth range of 420 – 570 µm. Mice viewed gray screen during laser response tests. Then, a series of visual stimuli were shown to characterize the unit’s visual properties (see Visual stimulation below).

For feedback silencing experiments, two approaches were used. In the first approach, Ai9 mice were injected with CAV2-Cre (60 nl) into V1. Two-to-three weeks later, the location of AL was identified using tdTomato signal. Then, a second virus, AAV1-FLEX-ArchT-GFP (50 nl) was injected into AL at two depths: 600 and 300 µm below brain surface. AL laser stimulation site and V1 craniotomy site were marked on the skull. In this approach, the specificity of silencing was achieved by restricting the inhibitory opsin expression to AL only, in addition to having a small laser spot. The second approach involved breeding together homozygous Ai9 and Ai39 mice to generate double-reporter mice that express tdTomato and HaloR under the control of Cre-recombinase. CAV2-Cre virus (60–70 nl) was injected into V1 on the left hemisphere. Two or more weeks later, AL and PM locations, identified using tdTomato signal, and that of V1 craniotomy site (≤250 µm away from V1 injection site) were marked on the skull. In this approach, the specificity of the silencing was achieved by restricting laser delivery to that area. These approaches allow us to target optogenetic manipulations specifically to the neurons in AL and PM that send projections to V1. While these V1-projecting neurons may also send projections to other areas, these bifurcating projections are thought to be relatively rare [34].

For both approaches, a headplate was centered over V1, AL and/or PM and fixed to the skull using opaque dental cement. On the day of the recording, two craniotomies were made, one over AL or PM for optogenetic silencing and the other over V1 for unit recordings, each ~400 µm in diameter. For optogenetic silencing, an optical fiber (100 µm dia.) encased in black-painted pipette was lowered into AL or PM to the depth of 0–175 µm relative to brain surface. A recording pipette was lowered into V1 craniotomy. Loose patch recordings were obtained from V1 neurons at depths of 208–710 µm below brain surface. The units’ visual properties were characterized using sinusoidal gratings (see Visual stimulation below) with or without silencing of feedback neurons in AL or PM. For these experiments, laser was delivered (520 nm, power: 1.6–3.4 mW) for 3 s every 14 s.

For control experiments testing the effectiveness of optogenetic silencing of putatively feedback neurons, both the optical fiber and recording pipette were lowered into the PM craniotomy and laser was delivered for 2 s every 8 s (520 nm, 2.7 mW). The recording pipette was moved in small (~5 µm) steps until a unit clearly silenced by laser was encountered and recorded. In these and other preliminary experiments, we determined that laser power of 0.5–1.2 mW and 0.9–2.7 mW was sufficient for stimulating and silencing feedback neurons’ spiking, respectively, and the laser’s direct effect on neurons was limited to ~500 µm from the center of the laser spot. In our measurements based on tdTomato signal in vivo, AL-V1 distance was ~2100 µm (average value based on 23 mice) and PM-V1 distance was ~1500 µm (24 mice).

To monitor running speed of the mouse, we placed an optical mouse on the anterior pole of the spherical treadmill. The output was read with custom software written in Presentation (Neurobehavioral Systems) and converted to an analog signal by a National Instruments NI USB-6008 board.

Visual stimulation

Visual stimuli were presented on a gamma-corrected 27” LED monitor (120 Hz refresh rate, ASUS VG278HE, mean luminance: ~75 cd/m²), placed 24 cm from the mouse’s right eye and subtending ~100° × ~70°(width × height) of visual space. Stimuli were generated by custom scripts written in Presentation (Neurobehavioral Systems). A uniform gray screen was presented outside visual stimulation periods. Visual stimuli were full-screen sinusoidal gratings (contrast: 99%), varied at different directions, spatial and temporal frequencies. For characterization of visual properties, each trial lasted for 4 s and visual stimuli lasted for 2 s. First, each cell’s optimal orientation/direction was determined by presenting drifting gratings at eight directions (SF: 0.08 cpd, TF: 2 Hz). Second, the cell’s SF tuning curve was obtained with drifting gratings of various SF (0.02–0.32 cpd, TF: 2 Hz) at the neuron’s optimal direction. In some cases, the experiment was repeated with stimuli that included lower (0.005, 0.01 cpd) or higher (0.64, 1.28 cpd) SF to obtain the full curve. Third, the cell’s TF tuning curve was obtained using drifting gratings of various TF (0.5–8 Hz) at the neuron’s optimal direction and SF. In some cases, the experiment was repeated with stimuli that included higher TF (16, 32 Hz). In SF and TF tuning experiments, six different conditions (e.g., 5 SF conditions + gray screen) were presented in a random order without replacement.

For experiments with optogenetic silencing of feedback neurons during visual stimulation, each trial lasted for 7 s and visual stimuli lasted for 2.5 s. Laser stimulation (3 s) was applied 250 ms before the onset of visual stimuli and lasted until 250 ms after the visual stimulus offset. In earlier experiments, six visual stimulus conditions (e.g., 5 SF conditions + gray screen) were run with or without optogenetic manipulation, to examine the effect of feedback silencing on the V1 neuron’s SF curve. We found that in most cases, feedback silencing mostly affected neurons’ responses to their preferred SF and exerted little to no effects on responses to non-preferred stimuli. We rarely observed shifts in tuning curves. Based on these observations, in later experiments each cell’s preferred direction, SF and TF were determined first, then the effect of feedback silencing was tested on three conditions: preferred stimulus (preferred direction, SF and TF), non-preferred stimulus (non-preferred direction, preferred SF and TF), and gray screen. For both versions of the experiment, different visual conditions were presented randomly without replacement, and laser was applied every other trial.

QUANTIFICATION AND STATISTICAL ANALYSIS

For analysis of running speed, the raw speed signal was first convolved with a 100 ms boxcar filter. Trials were categorized as running trials if the mean running speed of the mouse during the visual stimulus period was greater than 1 cm/s, a threshold similar to that used in previous studies [35, 36]. Stationary trials were defined as those during which the mean running speed was less than 0.5 cm/s. Figure S2B shows an example unit recording in area PM along with the mouse’s running speed. Mice used in this study remained stationary most of the time, thus our results are from stationary periods only, unless otherwise noted.

To characterize whether a cell was visually responsive, we compared the firing rates (one number per trial) during presentation of the cell’s preferred visual stimulus with its baseline firing rates (during gray screen), using a two-sample t-test. If the visually evoked firing rate was significantly higher than the baseline firing rate (p < 0.05), the cell was classified as visually responsive. If not, the cell was deemed not visually responsive and excluded from further analysis. According to this criteria, 25 of 35 AL neurons (71%) and 23 of 34 PM neurons (68%) recorded were considered visually responsive. The proportion of visually responsive neurons was similar among feedback vs. neighbor neurons. In comparison, 69 of 73 V1 neurons (95%) recorded were considered visually responsive. We classified units with peak-to-trough time of <0.5 ms as putatively fast-spiking units, and those with greater peak-to-trough lengths as regular spiking units. We note that 1 of 16 AL FB units and 2 of 16 PM FB units were classified as fast-spiking. GABAergic feedback neurons have previously been documented [37].

Analyses of visual properties were performed as follows. Cells were included in the general pool of analyzed cells if they passed the criteria of being visually responsive in any one of the tuning curves (SF, TF or orientation), but not every cell was included in the analysis of every property. For SF and TF tuning curves, the mean firing rate for each condition was plotted with SF or TF conditions on the x-axis (log scale) and the condition that elicited the maximum firing rate was defined as the peak SF or TF. The tuning curves were fitted with a Gaussian and the SF or TF value at the peak of the fitted curve was extrapolated and defined as preferred SF or TF. To derive OSI and DSI, an orientation/direction tuning curve was constructed with mean firing rate (baseline-subtracted) as a function of eight directions. From this curve, orientation selectivity index (OSI) was calculated as (R_pref − R_ortho)/(R_pref + R_ortho), where R_pref is mean firing rate in response to the preferred orientation and R_ortho is in response to the orthogonal orientation (preferred orientation + pi/2). Direction selectivity was calculated as (R_pref − R_opposite)/(R_pref + R_opposite). Linearity was calculated according to published procedures [38], as follows. Spikes during drifting gratings at the preferred SF and orientation (2 s duration) were binned into 100 ms intervals to generate a spike time histogram (baseline-subtracted). A discrete Fourier transform was applied to compute F1/F0, the ratio of the first harmonic (response at grating TF of 2 Hz) to the 0^th harmonic (mean response). For recordings where there were sufficient numbers of stationary and running trials, we compared the mean firing rate during the neuron’s preferred stimulus in stationary vs. running periods and computed the percent change in visual response due to running (i.e., (R_running − R_stationary)/R_stationary × 100). The summary of electrophysiological data can be found in Table S1.

For feedback silencing experiments, we excluded data from narrow-spiking V1 cells (peak-to-trough time <0.5 ms), to limit our sample to putatively excitatory neurons. In addition, only cells with ≥8 repeats per condition were included. Typically 10–12 repeats were run per condition. For determining whether there was a statistically significant effect of feedback silencing on individual V1 neurons, firing rates for three visual stimulus conditions were considered: preferred stimulus, non-preferred stimuli and baseline. For earlier experiments where the effect of laser was tested on the neuron’s SF curve, the non-preferred stimulus was gratings with non-preferred SF and preferred direction. For later experiments, the non-preferred stimulus was gratings with non-preferred (orthogonal) direction, preferred SF and TF.

A two-way ANOVA was performed on data from each cell, with stimulus condition and laser as two factors, and post-tests with multiple-comparison corrections were applied. If laser led to a significant change in the neuron’s response to its preferred stimulus according to the post-test (p < 0.05), the effect of feedback silencing was considered to be statistically significant for that cell. AL and PM feedback silencing had statistically significant effects on seven V1 neurons each. For group analysis, we divided V1 neurons into two groups based on their SF preference: 0.02–0.04 cpd (low SF) vs. 0.08–0.32 cpd (high SF) preferring neurons. We computed the mean change in firing rate (% change in firing rate during preferred stimulus) during laser for each neuron. We then performed a two-way ANOVA with groups (low- vs. high-SF preferring neurons) and silencing area (AL vs. PM) as factors, with post-tests. We also analyzed laser-off vs. laser-on mean firing rates (in log₁₀, Hz) for each group and silencing area and performed paired two-sample t-tests. Individual cell data are presented as scatter plots. Spike time histograms with/without laser are binned at 100 ms.

In cases where feedback silencing failed to show statistically significant effects on V1 neuron groups in paired two-sample t-tests at significant level of 0.05 (Figure 3I, 3M), we performed power calculations. This analysis revealed that for the effect sizes observed, the number of observations needed to have a power of 0.8 was 338 and 93, for AL feedback silencing effect on high-SF preferring V1 neurons and for PM feedback silencing effect on low-SF preferring V1 neurons, respectively. The Cohen's d were 0.15 and 0.29, respectively.

Normality of data was confirmed using D’Agostino and Pearson normality test and parametric tests were used. Non-paired comparisons were performed with unpaired t-tests and regular two-way ANOVAs. For determining whether there were significant effects of feedback silencing on individual cells, regular two-way ANOVAs were performed with Sidak’s multiple comparisons post-tests with adjusted p values. Paired two-sample t-tests were also used. Data were analyzed using Matlab (MathWorks), Prism 7 (GraphPad) and R [39]. Data are represented as mean ± SEM (n=number of neurons) unless otherwise noted.

DATA AND SOFTWARE AVAILABILITY

Raw and analyzed data are deposited in Mendeley Data at the following DOI. http://dx.doi.org/10.17632/cct9pt82cw.1

KEY RESOURCES TABLE

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Bacterial and Virus Strains
Canine adenovirus type 2 (CAV2-Cre)	Montpellier vector core, E. J. Kremer [14]	N/A
AAV1-FLEX-ArchT-GFP	University of North Carolina at Chapel Hill virus vector core	N/A
Chemicals, Peptides, and Recombinant Proteins
Cholera toxin subunit B conjugated with Alexa Fluor 488	Life Technologies	Cat#C-34775
Deposited Data
Raw and analyzed data	This paper	http://dx.doi.org/10.17632/cct9pt82cw.1
Experimental Models: Organisms/Strains
Ai9 Mice (tdTomato reporter)	Jackson Laboratories	Stock#7905
Ai14 Mice (tdTomato reporter)	Jackson Laboratories	Stock#7908
Ai32 Mice (channelrhodopsin-2 reporter)	Jackson Laboratories	Stock#12569
Ai39 Mice (halorhodopsin reporter)	Jackson Laboratories	Stock#14539
C57BL/6 Mice (wildtype)	Charles River	Strain#027
Software and Algorithms
Panorama tools	N/A	panotools.sourceforge.net
Fiji	[31]	https://fiji.sc/
Presentation	Neurobehavioral Systems	https://www.neurobs.com/
Matlab	MathWorks	https://www.mathworks.com/
Prism 7	GraphPad	https://www.graphpad.com/
R	R Core Team, 2017	http://www.R-project.org/

The information is provided in a separate document.