Coordinated cerebellar climbing fiber activity signals learned sensorimotor predictions.

Mice

All experimental procedures using animals were carried out with the approval of the Duke University Animal Care and Use Committee. All experiments were performed during light cycle using adult mice (>p60) of both sexes, randomly selected from breeding litters. All mice were housed in a vivarium with normal light/dark cycles in cages with 1–5 mice. Imaging experiments were performed using Tg(PCP2-Cre)3555Jdhu mice (Jackson Labs, 010536; n=12). Optogenetic mapping experiments were conducted in PCP2-Cre animals crossed with Ai35(RCL-Arch/GFP) (Jax 012735; n=5 mice). Single unit recordings during the cue prediction condition were performed in wild type C57BL/6J mice (n= 8 mice). Additional behavioral experiments where imaging was not performed were conducted in wild type C57BL/6J mice (n=18). We used two exclusion criteria for animals in this study: (1) poor recovery or other health concerns following surgical intervention or (2) missed virus injection, as determined by in vivo imaging and post-hoc histological analysis. Surgical Procedures 3–10 hours prior to surgery, animals received dexamethasone (3mg/kg) and ketoprofen (5mg/kg). Surgical procedures were performed under anesthesia, using an initial dose of ketamine/xylazine (50mg/kg and 5mg/kg) 5 minutes prior to surgery and sustained during surgery with 1.0–2.0% isoflurane. Toe pinches and breathing were used to monitor anesthesia levels throughout surgeries. Body temperature was maintained using a heating pad (TC-1000 CWE Inc.). Custom-made titanium headplates (HE Parmer) were secured to the skull using Metabond (Parkell). For imaging experiments, a 3mm diameter craniotomy was made over the lobule simplex approximately 1.4mm lateral and 2.8mm posterior to lambda, and glass cover slips consisting of two 3mm bonded to a 5mm coverslip (Warner Instruments No. 1) with index matched adhesive (Norland No. 1) were secured in the craniotomy using Metabond. Buprenex (0.05mg/kg) and cefazolin (50mg/kg) were administered following surgery twice a day for two days. Following a minimum of 4 recovery days, animals were water deprived for 3 days, or until body weight stabilized at 85% of initial weight, and were habituated to head restraint (3–5 days) prior to behavioral training. For imaging experiments, the glass cover slip was removed following behavioral training, and mice were injected (WPI UMP3) with AAV1.CAG.Flex.GCaMP6f.WPRE.SV40 (UPenn vector core, titer = 9.40×10 12 or 7.60×10 12 ). 150 nL virus diluted 1:1–1:5 in ACSF was injected at a rate of 30nl/min and a depth of 150 μm at 1–3 sites in dorsal lobule simplex. Imaging was performed beginning 14 days following injection. For in vivo pharmacology experiments, a 3mm craniotomy was performed over lobule simplex of headposted mice, and the dura mater was peeled back at the center of the craniotomy. 10uL of combined NBQX (300 μM) and CPP (30 μM) was applied into a well surrounding the craniotomy on the surface of the cerebellum 20 minutes prior to behavioral training. In 8 of 14 experiments, MCPG (30 μM) was also included to block metabotropic glutamate receptors. Because there no significant differences in performance across these two groups, all data were pooled for analysis. During behavior, a second drug application of 10uL was administered 20 and 40 minutes into the task, and depending on the duration of training, a third dose was applied after 1 hour. Craniotomies were subsequently covered with silicone elastomer (WPI, Inc.) prior to returning animals to their home cage. To quantify the spread of pharmacological agents in the cerebellum, 10 μL of fluorescein dye (Sigma-Aldrich #F6377, 1mM in aCSF) was applied to the craniotomy in 3 animals with the same method and timecourse as for drug applications. Following standard histological processing and imaging (below), labeling was quantified by creating a binary pixel mask thresholded at 30% of the maximum fluorescence value for each experiment. Dye labeling was then measured using the area of these pixel masks ( Supp. Fig. 4 B,C ).

Behavior

During behavioral training, animals were head-fixed and placed in front of a computer monitor, lever and reward delivery tube. Animals were trained to self-initiate trials by depressing the lever using their right forepaw, and required to successfully hold the lever in the down position for randomized intervals ranging between 500ms and 5 s on the cue reaction paradigm before performing the cue prediction paradigm. For both paradigms, a high contrast hold cue was present at all times, including the intertrial interval (ITI), and transitioned 90 degrees to the release cue on each trial at the instructed time of release until the animal either released the lever or 1 second had passed. Training sessions lasted for 90 minutes, and learning in the cue prediction task occurred with the range of 309–700 trials. Lever releases within one second of the release cue were rewarded immediately at the time of lever release (0.01 M saccharine). Immediately following any lever release, a solenoid was engaged to prevent lever press during the ITI. Following the ITI (3–6 s), the solenoid is lowered, allowing the mouse to self-initiate a new trial. During training, a 1–3 second ‘timeout’ was implemented to punish early lever releases. No timeouts were used in fully trained animals for imaging or behavior data collection sessions. For reward omission sessions, 20% of randomly determined correct lever releases were unrewarded. Animals used for imaging experiments performed the cue reaction condition with a mean peak percent correct of 79.7 ± 0.23% achieved in 26 ± 2 training days. On imaging days animals performed a range of 150 to 580 trials per session with a mean of 273.9 ±13.1 trials. Behavioral parameters including lever press, lever release and licking were monitored using Mworks ( http://mworks-project.org ) and custom software written in MATLAB (Mathworks). To assess the degree of lever control by the mice, the dynamics of lever press and release trajectories were compared to lever presses initiated by the experimenter where the lever was allowed to return to the rest position on its own ( Supp. Fig. 11 ). These results demonstrate that mice moved the lever both faster and slower than its intrinsic kinematics, and thus had full control over its trajectory. Licking was measured with electrical contact circuit. Inter-trial intervals ranged from 3 to 5 seconds.

Calcium Imaging

Wide-field imaging: Single photon imaging was performed using a customized microscope (Sutter SOM) affixed with a 5x objective (Mitutoyo, 0.14NA) and CMOS camera (Qimaging, Rolera em-c 2 ). Excitation (470 nm) was provided by an LED (ThorLabs, M470L3), and data were collected through a green filter (520–536 nm band pass, Edmund Optics) at a frame rate of 10Hz, with a field of view of 3.5×3.5mm at 1002×1004 pixels. Two-Photon Imaging: Two-photon imaging was performed with a resonant scanning microscope (Neurolabware) using a 16x water immersion objective (Nikon CFI75 LWD 16xW 0.80NA). Imaging was performed using a polymer to stabilize the immersion solution (MakingCosmetics, 0.4% Carbomer 940). A Ti:Sapphire laser tuned to 920nm (Spectra Physics, Mai Tai eHP DeepSee) was raster scanned via a resonant galvanometer (8 kHz, Cambridge Technology) onto the brain at a frame rate of either 30 Hz with a field of view of either 1030 μm × 581 μm (796 × 264 pixels) or 555 μm × 233 μm (796 × 264 pixels), or 15.5Hz with a field of view of 555 μm × 452 μm and (796 × 512 pixels). Data were collected through a green filter (510 ± 42 nm band filter (Semrock)) onto GaAsP photomultipliers (H10770B-40, Hamamatsu) using Scanbox software (Neurolabware). A total of 12 mice were used for imaging experiments (10 mice for wide-field, 11 mice for two-photon, 12 mice total).

Single Unit Recordings and Optogenetics

Acute single unit recordings were performed in awake animals by performing a craniotomy over lobule simplex and inserting a multi-electrode silicone probe (Neuronexus, A4×8–5mm-100–400-177-A32, 4 shanks, 8 site/shank at 100 μm spacing) using a Cerebus multichannel acquisition system (Blackrock Microsystems, Salt Lake City). For single unit recordings obtained in the cue prediction condition ( Supp. Fig. 6 ), chronically implanted electrode arrays were used (Dual drive movable electrode bundles with 8 tungsten electrodes (23 μm) in each cannula or 16 tungsten electrodes bundle in one cannula, Innovative Neurophysioloy Inc). Electrode arrays were implanted using stereotaxic coordinates to target lobule simplex at AP 6.2; ML 2.0. Electrode bundles were inserted into lobule simplex at a depth of approximately 0.2–0.3 mm. The implant was encased in Metabond for stability. For both recording conditions, continuous recording data was bandpass filtered with a 2-pole Butterworth between 250 Hz and 5 kHz and referenced against an electrode with no spikes using Spike2. Single units were isolated by amplitude thresholding. Template waveforms were defined and characterized by their width and peak, and PCA of waveforms was done in off-line in Spike2. Complex spikes and simple spikes were discriminated as in de Solages et al. 50 , first on the basis of their stereotypical waveform. Complex spikes typically had a multi-wavelet form, including a large positive peak within 6ms following spike initiation. Manual identification of 10–20 complex spikes was used to generate a mean template waveform, which was then compared to all other spikes for a given unit using the Spearman rank order correlation coefficient. The combination of Spearman coefficients and the magnitude of the positive waveform deflection was used to segregate complex and simple spikes. The presence of a post-complex spike pause (20 ms or more) in simple spike firing was verified by cross-correlogram for all isolated single units. Complex spike rates for the cue prediction condition were normalized to the baseline firing rate determined in a one-second window during the ITI according to: (FR-baseline)/baseline) ( Supp. Fig. 6 ). For optogenetic mapping of the dorsal cerebellum, an optical fiber (0.39 NA, 400 μm core multimode, ThorLabs FT400EMT) coupled to a 532 nm laser (Optoengine, MGL-III-532) was positioned above the cranial window using a micromanipulator (Scientifica PatchStar).

Histology

Mice were deeply anesthetized with ketamine/xylazine (200mg/kg & 30mg/kg respectively, IP) then perfused with PBS then 4% paraformaldehyde. 100um sagittal sections were cut using a vibratome (Pelco 102). Cerebellar slices were then mounted using a mounting medium (Southern Biotech Fluoromount-G or DAPI Fluoromount-G) and imaged with a fluorescence microscope (Nikon Eclipse 80i).

Data

Analysis and Statistics Behavior: Behavior sessions were only analyzed within the time range of active task performance. Accordingly, the last trial of each session was determined by the occurrence of either two consecutive failed trials (in which the mouse did not release the lever in response to the cue) or two consecutive trials with post-ITI duration to press longer than 1.5x the session average. Reaction times were measured using lever releases within the reward window as well as those occurring 200ms prior to the reward cue to account for predictive responses. Reaction times are plotted as a binned average of five trials. Initial reaction times were calculated according to the y-intercept of the linear regression of the first fifty trials. Final reaction time at the end of each session was calculated by averaging the reaction times of trials 200–250. The symmetry of reaction time data about the sample mean was measured to test the extinguishment of learning after following cue prediction sessions according to skewness. Sample skewness was defined by s = E ( μ - X ) 3 σ 3 where μ is the mean of X , E (∗)is the expected value of the quantity ∗, and σ is the standard deviation of X . Thus, for a given sample x , this calculates: 1 n ∑ i = 1 n ( x i - x - ) 3 1 n ∑ i = 1 n ( x i - x - ) 2 3 for x i ∈ x where x - is the sample mean. Movement trajectories were measured using a lever affixed to a rotary encoder (US Digital) with a 60mm radius and 1250 pulses per revolution. Encoder values were collected at 1000 Hz, and trajectories were calculated by up-sampling at 200 μs (5000 Hz) using nearest-neighbor interpolation. Pulse number was converted to degrees, and the vertical displacement in millimeters was calculated using the chord length of the leaver displacement angle. 20–80% rise times were calculated using normalized average trajectories. Presses and releases were sorted according to duration in a window from 200ms before threshold crossing to the time of a threshold crossing half way between the top and bottom of the total lever displacement. Wide-field imaging: Imaging analysis were performed using custom MATLAB code. Regions of interest (ROIs) were selected within empirically defined forelimb movement region of lobule simplex according to the spread of GCaMP expression. Window location and the lobule identity were identified according to folia patterns visible through the cranial window, landmarks recorded during surgery, and post-hoc histology. Baseline fluorescence (F) was measured on a trial-by-trial basis during the inter-trial interval (ITI) as the mean fluorescence 900 to 200ms before trial initiation. Normalized fluorescence (ΔF/F) was calculated according to the cumulative activity within an ROI, or on a pixel by pixel basis. Lever hold times 1 s after the visual cue were exceedingly rare, and thus excluded from analysis. Lever releases between 200ms and 1000ms following the visual cue were classified as correct, and releases prior to visual cue were classified as early. Lever releases ≤200 ms following the visual cue were considered too fast to be reactions to the cue based on average reaction time distributions. Peak ΔF/F was measured on a trial by trial basis in the time window from 100ms before to 400 ms after lever release. Note that differences in the timing of single trial calcium transient peaks produce smaller amplitudes for ΔF/F timecourses as compared to reported peak ΔF/F measurements. Spearman’s correlation between lick rate and peak ΔF/F within sessions was calculated according to the lick rate within 500 ms following each lever release and the peak calcium transient for each trial. For this correlation analysis, trials without licking were removed, and only sessions with 7 or more trials in each condition were included. Meta K-means analysis and Cluster Correlations: Images were first registered to reduce motion artifacts, and then thresholded at 70% of maximum intensity for each frame to remove background noise. Images were downsampled 5-fold in both X and Y. Baseline F was defined as the averaged fluorescence across the entire movie for each pixel, and used to normalize change in fluorescence (ΔF/F). ΔF/F was then re-normalized to the maximum ΔF/F during the entire movie for each pixel. The repeated k-means clustering algorithm (meta-k-means) separated pixels to cluster centroids based on a pairwise correlation distance function using Pearson’s linear correlation coefficient (r). The final clusters were determined by first thresholding all the k-means results at 800 out of 1000 runs, and then by merging highly correlated clusters based on Dunn’s index. Clusters occupying less than 3% of the total imaging field were excluded from further analysis. Intra-cluster and inter-cluster correlation coefficients were calculated between 100 ms before and 300 ms after lever release for both correct and early trials on a frame by frame basis. To test for differences in the relationship between spike rates and correlation coefficients between trial types, paired clusters from each distribution were resampled with replacement and fit with a line to measure the y-intercept 1000 times. Statistical significance was computed according to the 95% confidence interval of the distribution of y-intercept difference values (difference between y-intercepts 0.0375 ± 0.0013, 95% confidence interval [0.0349, 0.0402]). Two-photon imaging: Motion in the XY plane was corrected by sub-pixel image registration. To isolate signals from individual PC dendrites, we utilized principal component analysis (PCA) followed by independent component analysis (ICA). Final dendrite segmentation was achieved by thresholding the smoothed spatial filters from ICA. A binary mask was created by combining highly correlated pixels (correlation coefficient > 0.8) and removing any overlapped regions between segmented dendrites. Notably, image segmentation using these criteria did not extract PC soma, which were clearly visible in many single- and two-photon imaging experiments. Fluorescence changes (ΔF) were normalized to a window of baseline fluorescence (F) between 500ms and 100ms preceding trial initiation. Responses were categorized as significant (p