Task and behavioral performance.

A) Left, drawing of the animal performing the task. Right, schematic overview of directionality in the task: only one C-stimulus is presented per trial in one of 4 cardinal directions from the central fixation dot. This could be to the top/bottom of the central fixation, with possible gap directions to the left and right in the [L/R] task, or upward and downward in the [U/D] task. B) Example trial with a C-stimulus presented to the right, with an “Up” gap direction, where the animal has to make an upward saccade. Delay period between the mask and the onset of the target varied between 500-700 ms (randomly drawn from: 500, 550, 600, 650 and 700 ms). The saccade had to be performed in the epoch between the appearance of the choice target and the reward. Beige arrow indicates the correct choice, not actually visible to the monkey. C) Overview of performance per animal per gap direction. Black bar at the middle of the violin plot represents the 25th and 75th percentile, the white dot at the center is the median. D) Average fraction of correct responses per retraining session, after the animal had had several weeks of ‘holiday’. Each session represents the median performance of one day. E) Weight, i.e., relative impact, of the gap direction of the C-stimulus in predicting the correct response through logistic regression. Each point is the average of one session, error bars indicate 95% confidence interval from 100 bootstrap permutations. Beige data points and fit represent the up/down task, the green data the left/right task.

Neural activity in the DN during the Landolt C task.

A) MRI images of Mi and Mo, showing the electrode position in the DN. B) Example trial of extracellularly recorded DN cell in Mi aligned to C-stimulus onset (blue shaded area shows when C is visible), while the animal maintained fixation until the potential targets were presented (in this case an upward saccade was correct). Top panel depicts eye position, bottom panel shows the single unit activity of the DN cell. Pink shaded area represents juice reward pulse duration. C) Raster plot of correct trials sorted per gap direction, represented per color, of the cell presented in B. Within each direction, trials are sorted by the time of saccade onset (black dot in each trial). D) Average activity per direction in 100 ms bins in spikes per second. Left panel displays activity aligned to C-stimulus onset, right panel displays activity aligned to saccade onset. E) Preferred directions of the population of neurons. Normalized polar histograms show the fraction cells that preferred the stimulus direction (left) or the action direction (right). F) same as E for direction preference during saccade epoch. G) Normalized activity of cells displaying facilitating activity deviating from baseline in either the window after C-stimulus onset (gray bar, 0 to 800 ms, left panel) or around the saccade (gray bar, −200 to 200 ms, right panel). H) Same as in G, but for suppressed cells. I) Timing of peak / trough responses from G and H, stars indicate significance at * p=.05, ** p=.01, *** p=.001 level, according to Kolmogorov–Smirnov test; ns = not significant. For separate data sets of Mi and Mo, see Supplementary Figure 2.

Description of the sliding windows used to determine significant modulation for each category.

Encoding of task events by DN neurons.

A) Schematic of a generalized linear model (GLM) for the spike count of a given neuron as being Poisson-distributed around behavior-dependent mean rates. The neuron is predicted to have time-varying changes in activity triggered by one or more behavioral events, and these changes are flexibly modeled as weighted sums of event-aligned basis functions. The set of events that the neuron responds to, as well as any trial-type specificity of these responses, was selected so as to obtain the most parsimonious model that produced a good fit. B) Distribution across cells (two colors for the two monkeys) of a “modulation index” (modIdx) score for how much trial types influence neural responses (see also examples for individual cells in panels D and E). First, we constructed an “unspecialized” GLM model that predicts the cell’s activity rate in trial to be, independently of trial type. This is compared to a trial-type dependent GLM prediction for that cell, where the significant trial-type categories for the model were parsimoniously selected as described in the Methods. The modulation index for the cell is then defined using the ratio of trial-type specific to unspecialized model predictions:. To produce a single summary number per cell, we computed as the time- and trial-type average of . The four subplots of this panel show the distribution of modIdx scores for cells with GLM predictions where the model-selected trial category c included gap direction (top left), saccade direction (top right), past-trial gap direction (bottom left), and past-trial saccade direction (bottom right). C) Proportions of cells with GLM models that had significant dependencies on various trial types as indicated in the legend of panel A. The miscellaneous 1- to 4-category proportions were pooled over cells with infrequently occurring trial-type dependencies (<2% cells per category); for example, the “misc. 3-cat.” cells included those with simultaneous trial-type dependencies on gap, past-trial gap, and past-trial saccade conditions. D) GLM prediction for the firing rate vs. time in a trial of 3 example neurons recorded in Monkey Mi. For each cell (one column), the predicted firing rate is shown for trials of 4 different gap directions (differently colored lines), and separately for correct trials (top row) and error trials (bottom row). The onset of various events in the trial are indicated with vertical dashed lines. E) Same as D, but for 3 example cells from monkey Mo.

Decoding of trial conditions from neural firing rates.

A) Fraction of cells from which the trial outcome can be significantly decoded (see Methods) using data in 200 ms time bins aligned to various events in the trial (columns). * indicate bins in which this fraction is significantly different between the two monkeys (p < 0.05; Fisher’s exact test, 2-tailed). B-D) same as A but for the decoding of gap direction, outcome of the past trial and gap direction of the past trial, respectively. Note the higher percentage of cells with decoding of the gap occurrence shortly after the delay period in Mo (B), and the higher percentage of cells with past-outcome decoding in Mi (C).

Anatomical characterization of recording area and its input and output connectivity.

A) CTb injections in the dentate nucleus of Mi and Mo displays the site of electrophysiological recordings of its neurons at the lateral edge of the DN. B) Pictures of retrogradely labeled cells in the Principal Nucleus of the inferior olive (IO); insets show overview of IO. C) Quantification of retrogradely labeled cells in different subnuclei of the IO in Mi (blue) and Mo (pink). D) Examples of retrogradely labeled cells in the medial pons of Mi (left) and Mo (right); insets show higher magnification of labeled cells. E) Schematic representations of cerebellar cortical areas containing retrogradely labeled Purkinje cells in Mi (left) and Mo (right). F) Examples of retrogradely labeled Purkinje cells for both animals (Mi and Mo in panels on the left and right, respectively); insets show areas of the higher magnifications with the labeled cells. G) Example of anterograde CTb labeling of cerebellar fibers with dense varicosities (arrowheads) in the thalamus of Mo. Panels i and ii on the right correspond to insets in the left panel; panels iii and iv are examples of anterograde CTb labeling outside of the left panel in the anterior pulvinar and red nucleus, respectively. VL, LD, CL, APul, VPL, CM and MD indicate ventrolateral, laterodorsal, centrolateral, anterior pulvinar, ventroposterior lateral, centromedial and mediodorsal nucleus of the thalamus, respectively. dlPO, vlPO: dorsolateral-, ventrolateral Principle Olive, vlo: ventrolateral outgrowth; MAO and DAO: medial and dorsal accessory olive. Scale bars in A, B, D, F and G indicate 1250 μm, 100 μm, 1000 μm, 50 μm and 1000 μm, respectively.

Weight, i.e., relative impact, of the gap direction of the C-stimulus of the previous trial to predict the response on the next trial over time; the trials could be from either a L/R or a U/D task, hence the two conditions per panel. Thus, the same logistic regression model as in Figure 1 was applied to determine the influence of previous trial characteristics on current trial behavior (See Supplementary Table 1 for the statistical tests).

Regression analyses applied to the data from Supplementary Figure1. Only two of the regression models do significantly better than a constant model (F-test, highlighted in bold). Both of those do not survive a Bonferroni correction (α = 0.0083 for 0.05/6 test per monkey).

Neural activity in the DN during the Landolt C task separated for both animals. A) Preferred directions of the populations of neurons split between Mi (left) and Mo (right). Normalized polar histograms show the fraction of cells that preferred the stimulus direction or the action-to-gap direction, when assessing the encoding during the stimulus window (blue bars) or saccade window (black bars). B) Timing of maximum facilitating or suppressing responses from C-onset and around saccade onset for Mi (left) and Mo (right). C) Normalized activity of cells displaying facilitating (red) or suppressed (blue) activity deviating from baseline, split between Mi (left) and Mo (bright). Significant deviations after C-stimulus onset (0 to 800) and around saccade onset (−200 to 200 ms) are indicated by gray bars.

Examples of correct and incorrect trial and principal component analysis (PCA) decomposition of this population activity. A) Average firing frequency during correctly executed (black) and incorrectly executed (red) trials of an example cell recorded in monkey Mi. Lightly blue shaded areas highlight differences in firing frequency between correct (black) and incorrectly (red) executed trials. Activity is aligned to stimulus onset (left) and saccade onset (right). Thin lines indicate limits ± 1 standard deviation. B) PCA decomposition of average firing frequency aligned to stimulus onset (vertical dashed line) for the population of the DN neurons. Black lines indicate PC coefficients of average firing frequency during correct trials, red lines during incorrect trials. First three components are shown with the respective amount of explained variance at the top of the panel. C) Same as in A, but for Mo. D) Same as in B, but for Mo. Note the ramping activity after stimulus onset occurs in both Mi and Mo, but that the onset of this ramping activity in Mo precedes that of Mi in the PCA decomposition (compare left panel of D with that in B).

Ramping activity of DN neurons that modulated upon presentation of the C-stimulus. A) Average firing rate of the correct (black) and incorrect (red) trials of the population of neurons in Mi (left) and Mo (right) for which we had at least 100 trials for both types of outcomes. B) The onset of ramping was detected with the so-called Kneedle algorithm (see Methods section). Two example cells of the DN are shown in which ramping activity is detected in the first second after stimulus onset (vertical dotted lines indicate the 1 second window). The dots on top of the traces indicate the ramping onset for these individual cells. For descriptive statistics see Supplementary Table 3.

Descriptive statistics of latency to ramping onset.