Nature Neuroscience

A cortical circuit for audio-visual predictions


All animal procedures were approved by and carried out in accordance with guidelines of the Veterinary Department of the Canton Basel-Stadt, Switzerland. C57BL/6 and PV-Cre mice, female and male, between the ages of 3 and 4 months and group-housed by gender were used in our studies. Mice were housed on a 12-h light/dark cycle in cages with horizontal running wheels at an ambient temperature of between 20 °C and 25 °C and humidity between 40% and 60%.


Surgeries were performed as described previously52. In brief, mice were anesthetized using a mix of fentanyl (0.05 mg kg−1), medetomidine (0.5 mg kg−1) and midazolam (5 mg kg−1). A craniotomy of either 5 mm or 3 mm in diameter was made over V1; a glass coverslip was super-glued in place; and a custom-machined stainless steel head bar was implanted.

AAV injections

Injections consisted of 100–250 nl of AAV vector with a titer in the range of 1012–1014 genome copies per ml. The coordinates of the injections in V1 were 2.7–2.8 mm lateral from the midline and 2.8–3.0 mm posterior from bregma. For AuC injections, the coordinates were 4.4 mm lateral from the midline and 2.6–2.8 mm posterior from bregma, and the injection pipette was rotated to be perpendicular to the brain surface. For somatic imaging in V1, we used AAV2/1-EF1α-GCaMP6f for V1 PV-Cre excitation; for FMI, we used AAV2/1-EF1α-ChrimsonR-tdTomato53; and, for AuC axon imaging, we used AAV2/1-EF1α-GCaMP6s54.


For postmortem histological analyses, mice were transcardially perfused with 4% paraformaldehyde in PBS. Brains were isolated and maintained in 4% paraformaldehyde at 4 °C on a shaker overnight. The fixed tissue was then rinsed with PBS and sectioned into 70-µm- or 100-µm-thick slices using a vibratome. Sections were mounted and sealed with DAPI ProLong mounting medium. Sections for all mice were imaged using a Zeiss AxioScan.Z1 slide scanner at ×10 magnification (Zeiss Zen blue edition software). All images used for quantification of the number of neurons expressing GCaMP were acquired at ×20 magnification, 5-µm step, z-stack images using a confocal microscope (VisiView version 3.3 software). Atlas overlays for histological images were adapted from ref. 50. Atlas images were first aligned to both rhinal fissures and the external capsule of coronal sections, and, subsequently, the thickness of the cortex was adjusted to fit each individual mouse. Confocal ex vivo histology images were acquired for all mice.

Quantification of AAV spread

Injections of AAV2/1-EF1α-GCaMP6s-WPRE in AuC for axonal imaging in V1 also result in axonal uptake and expression in V1 neurons that project to AuC. To quantify what fraction of the axons in V1 could come from retrogradely labeled V1 neurons, we used a separate set of five mice for histological quantification. Mice were injected with AAV2/1-EF1α-GCaMP6s-WPRE in AuC and sacrificed for histological analysis time-matched to the start of the imaging experiments. We performed a histological quantification using confocal images of fixed tissue in a region corresponding to the location of our two-photon imaging window. We then quantified the number of neurons per slice volume (656 µm × 656 µm × 32 µm). We found infected neurons in V1 in two of five mice with a mean ± s.e.m. across mice of 2.6 ± 1.9 neurons and five infected neurons in one of five mice in secondary visual areas (1 ± 1, mean ± s.e.m. across mice) (Extended Data Fig. 2a,b). Given that the number of axons we were able to image in V1 in a volume of 200 µm × 200 µm × 40 µm was more than two orders of magnitude larger (day 1: 1,054.8 ± 117.8, day 2: 893.2 ± 91.6, day 3: 1,008.2 ± 121.9 and day 4: 1,025.6 ± 130.0; mean ± s.e.m.), retrogradely labeled V1 neurons are unlikely to account for a substantial fraction of the axons recorded in V1. Note that the comparison by volume is not entirely straightforward as one would need to estimate the average fraction of total V1 volume that the axon of a given V1 neuron would be visible in. However, additionally arguing against a contamination by axons of V1 neurons is the fact that expression levels in retrogradely labeled neurons tend to be far lower than at the primary injection site55. Thus, although we cannot exclude that some of the axons in our dataset originated from retrogradely labeled V1 neurons, the vast majority of them were likely AuC projection axons.

Two-photon imaging

Functional imaging of GCaMP6-expressing neurons was performed using a modified Thorlabs B-Scope. The illumination source for two-photon imaging was a femtosecond infrared laser (Spectra-Physics) tuned to a wavelength of 910 nm. A 12-kHz resonance scanner (Cambridge Technology) was used for line scanning to acquire data at a frame rate of 60 Hz at a resolution of 400 × 750 pixels. In addition, we used a piezo actuator (Physik Instrumente) to acquire images at four different depths by moving the objective (Nikon ×16, 0.8 NA) in 15-µm steps between frames, thereby reducing the effective frame rate per imaging plane to 15 Hz.

Optogenetic stimulation during two-photon imaging

The methods for simultaneous two-photon imaging and optogenetic stimulation were described previously24,56. In brief, the illumination source for the ChrimsonR stimulation was a switchable 637-nm laser (OBIS, Coherent) used at an average power of 11 mW and triggered using a TTL pulse. A dichroic mirror (ZT775sp-2p, Chroma) was used to combine two-photon and optogenetic stimulation light, and a long-pass dichroic mirror (F38-555SG, Semrock) was used to filter GCaMP6 emission from illumination light. To prevent stimulation light artifacts, the 637-nm laser was synchronized to the turnaround times of the resonant scanner when data were not acquired. To reduce the influence of ringing artifacts in the amplifier, signals were digitally band-pass filtered at 80 MHz using a 1.6-GHz digitizer (NI-5772, National Instruments) and an FPGA (PXIe-7965, National Instruments) to implement a digital Fourier filter.

Conditioning paradigm


Mice were handled by the experimenter every day for at least 1 week before being introduced to the virtual reality (VR). Water restriction began 1 week before the start of experiments in which a water reward was delivered, and mice received 1 ml of water per day. Three to five days before the experiment, mice were exposed and habituated to head fixation in the VR and rewarded with sunflower seeds after each exposure period. Mice were considered habituated when they voluntarily walked onto the experimenter’s hand and did not resist head fixation. During experiments, mice received supplemental water after conditioning if they had not consumed at least 1 ml in water rewards. Mice were monitored to ensure they maintained at least 80% of their original body weight. For V1 soma imaging, one cohort of five mice underwent optogenetic experimentation in the VR context on day 1, followed by 5 d of conditioning, followed by a final day of optogenetics. A second cohort of five mice had optogenetic experimentation after only 5 d of conditioning. For AuC axon imaging, 20 mice were conditioned for 4 d. One mouse was removed from the analysis on day 4 owing to insufficient image registration. Of these mice, eight were PV-Cre and were also used for optogenetic and visual-context-only experiments.


Auditory stimuli consisted of either 16.1-kHz or 10.5-kHz pure tones presented at approximately 65 dB SPL29. The three visual stimuli used were a sinusoidal grating, a geometric pattern of triangles and a geometric pattern of ovals. One of the associated stimuli (a and b) was always the grating, but the pairing of the stimuli was otherwise randomized and counterbalanced across animals. For paired conditions, the auditory stimulus was 1 s in duration, followed immediately by a visual stimulus 1 s in duration, followed immediately by a reinforcement: a—water reward, b—air puff. For visual-stimulus-only conditions, the visual stimulus was presented for 1 s and never reinforced. Approximately 25% of trials were the Vx condition during the first four conditioning days (day 1, Va: 24.5% ± 0.2%) and ~14% of trials on day 5 (Va: 13.8% ± 0.5%). The occurrence of Vc as a fraction of all un-cued visual stimulus trials was day 1: 50.1% ± 0.3 % and day 5: 33.9% ± 0.5%. On day 5, AbVa occurred for ~14% of all cued visual stimulus trials (AbVa: 13.8 ± 0.6). Values reported are mean ± s.e.m. For axonal imaging, the visual-only paradigm was performed 1 d before and after conditioning as well as after the audio-visual paradigm on conditioning days (Fig. 2i). Stimuli consisted of full field grating presentations of eight orientations with a stimulus duration of 2 s and a gray (mean-luminance) inter-stimulus interval of 3 s. Optogenetic stimulation of AuC axons for FMI experiments was performed 1 d before and 1 d after the conditioning paradigm as described above. Stimuli were also presented occasionally on the same day as optogenetic stimulation for a couple of reasons. First, we wanted to obtain a relative measure of V1 neuron responsivity to natural visual stimulation and to artificial optogenetic stimulation of AuC axon input on the same day. This allowed us to control for whether neurons were different in their excitability in general before versus after conditioning or showed more specific changes in their responsiveness to visual stimuli.


Mice were head-fixed and free to locomote on an air-supported polysterene ball. A virtual tunnel designed with low-contrast gray checkered walls was projected onto a toroidal screen surrounding the mouse and yoked to linear displacement of the ball. From the mouse’s perspective, the screen encompassed a visual field of approximately 240° horizontally and 100° vertically. One speaker was placed on the left side and one on the right side of the VR for presentation of auditory stimuli. The VR system was otherwise constructed as described previously52. A water spout was placed in front of mice, and licking was detected using a custom-made electrical circuit in which a mouse closes the circuit whenever its tongue contacts the metal spout or water droplet57. The resulting voltage was thresholded to calculate licking events.

Image analysis

Regions of interest (ROIs) for soma were obtained using custom semi-automated image segmentation software. ROIs for axons were obtained in an automated process as previously described in Mukamel et. al.58 using a combination of principal and independent component analysis and image segmentation modified in-house. Fluorescence traces across time were then calculated as the mean pixel value in each ROI per frame. ΔF/F was calculated using median-normalized traces and filtered as described previously59. For axonal imaging, data came from the same location in the brain using blood vessel patterns for alignment, but individual axons were not matched across imaging time points.

Data analysis

Data analysis was performed using custom-written MATLAB (MathWorks) code. To quantify differences between response curves during visual stimulation (Figs. 1c,g, 2c and 3e and Extended Data Figs. 2c, 3b,c,d and 5d,e g), ΔF/F was compared in a response time window (11 frames, 267−1,000 ms after visual stimulus onset) with a baseline subtraction during auditory stimulation (mean activity in a window preceding visual stimulus onset, 10 frames, −667 ms to 0 ms) bin by bin for one-frame (66.7-ms) time bins using a paired t-test (P < 0.05). Dots above response curves indicate significant difference for at least three consecutive bins. For quantification of responses during visual, auditory, optogenetic or sham stimulation, ΔF/F was averaged over the response time window (11 frames, 267−1,000 ms after stimulus onset) and baseline subtracted (mean activity in a window preceding stimulus onset, ten frames, −667 ms to 0 ms) (Figs. 1d,h, 2g and 3d,e and Extended Data Figs. 1b–i, 2a, 3b,c,e, 4c–e and 5a,c f,h). Mean neural activity is an average across trials and neurons. Mean behavioral data are an average across trials and mice. Licking and running were quantified during the response time window (Fig. 1f and Extended Data Figs. 2c and 3b,c). For quantification of visually responsive axons (Fig. 2g–i), ΔF/F during the response time window was compared to ΔF/F during the baseline window. Normalized suppression of AuC axons was quantified as the difference between the response to the stimulus with and without optical stimulation of V1 PV neurons, normalized by the mean response to the stimulus without optical stimulation (Fig. 2f). The response difference index was computed by subtracting the response during the visual stimulus after the auditory cue (Aa,b,oVa,b,o) from that during the visual stimulus presented alone (Va,b,o) (Fig. 1d,h and Extended Data Figs. 1f,i, 3b,e and 5f), the visual stimulus after the paired cue (AaVa) from that during the unpaired cue (AbVa) (Fig. 1h and Extended Data Fig. 3c and 5h) or the visual stimulus after the unpaired cue (AbVa) from that during the visual stimulus alone (Va,b,o) (Fig. 1h) or (AaVb) from (Vb) (Extended Data Fig. 3e) and normalized to the mean visual response alone (Va,b,o) on day 1 of conditioning. Note that we used a subtractive measure normalized by day 1 responses to avoid division by 0 problems. For classification of V1 neurons as excited by or inhibited by AuC stimulation, we split the population of neurons into two groups. Those with a response greater than 0 were included in the excited-by group, and those with a response less than 0 were included in the inhibited-by group (Fig. 3e and Extended Data Fig. 5d–h). For running speed matching (Extended Data Fig. 3b,c), an iterative resampling procedure was used: the fastest and slowest trials were successively removed in the stimulus conditions with higher and lower average running speeds, respectively, until average running speed in the condition with the initially higher average running speed was lower than in the condition with the initially lower average running speed. For Fig. 3d,e, early in conditioning is day 1 of experiment (first exposure to conditioning stimuli), and late in conditioning is the average of the visual responses on days 3 and 4 of conditioning. For the no-reinforcement paradigm (Extended Data Fig. 1g–i), mice were exposed to two sets of stimuli as in the reinforced experiments, AaVa and AbVb, but, as neither condition was reinforced, visual and auditory cue responses were calculated by averaging across both conditions (AoVo is the average of AaVa and AbVb; Vo is the average of Va and Vb; and Ao is the average of Aa and Ab).

Statistics and reproducibility

All statistical analyses were performed in MATLAB using custom-written software. Sample sizes were chosen to match typical numbers used in animal behavioral experiments. All data acquired were included in the analysis, with the exception of one mouse that was removed from Fig. 2g,h owing to technical difficulties displaying stimuli during conditioning. Changes in the number of mice (and neurons) across time points are the result of technical difficulties that prevented the acquisition of data in some mice (Supplementary Table 1). Data were first tested for normality using a Lilliefors test, and, when the null hypothesis could not be rejected (ho: data come from a normally distributed population), parametric tests were used. Otherwise, non-parametric tests were used. Paired t-tests or rank-sum tests were used for analyses with matched samples. For all unmatched samples, data that failed to reject the ho in the Lilliefors test, unpaired t-tests were used (for example, comparisons of axon responses on different conditioning days). Error shading and bars indicate s.e.m. unless otherwise stated in the figure legends. All statistical tests were two tailed. Scattered data were quantified using correlation coefficients, denoted as r, and coefficients of determination were computed by taking the square of r. For a summary of all statistical tests used, n values and exact P values, see Supplementary Table 1. No statistical methods were used to determine sample sizes, but sample sizes were selected based on typical sample sizes used in the field. All imaging and behavioral data were acquired from multiple experimental series. Data were additionally subdivided into multiple smaller groups to ensure that effect directions (for example, activity suppression and response differences) were maintained. All efforts to reproduce our results were successful. C57BL/6J mice were assigned randomly to experimental groups defined by injection location and experimental procedure. PV-Cre mice were assigned to optogentic experiments based on their genotype, and our stimulation protocol included randomization of activation laser and sham stimulations. The experimenter was not blinded to group allocation of mice for two-photon and behavioral data but was blinded to mouse identity and cortical region for quantification of histological analyses.

Reporting Summary

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

Source link

Related Articles

Leave a Reply

Back to top button