Single unit analysis

Spike detection, clustering and subsequent manual re-sorting were performed using the Klusta Suite (https://klusta.readthedocs.io/en/latest/https://klusta.readthedocs.io/en/latest/⁵²). After automatic spike detection and clustering, spike clusters were manually re-sorted using the KlustaViewa 0.3.0.beta1 graphical user interface into either single unit clusters or excluded from further analysis as putative artifacts. Manual sorting was based on visual inspection of the spike waveforms, on the autocorrelograms and on the distribution of the first three principal components.

Single units were classified as INs or PCs⁸ on the basis of (1) the width of the mean action potential waveform at 25% of the spike amplitude measured from baseline; (2) the first moment of the autocorrelogram in the time lag window of 0–50 ms, obtained using Matlab’s xcorr function with a temporal resolution of 1 ms; (3) the baseline firing rate during the control period of ≥ 30 s before light stimulation. For every single unit its percentile in the distribution of these three parameters was identified and used to determine a classifier coefficient for this respective single unit as follows:

Perc_{baseline rate}, percentile for the respective single unit in the baseline firing rate distribution; Perc_{1st AC moment}, percentile in the 1^st moment of the auto-correlogram distribution; Perc_{spike width}, percentile in the spike width histogram.

The classifier coefficient ranges between 1 and 100. The lower the classifier coefficient, the more a single unit resembles an ideal principal neuron while higher values indicate a more IN-phenotype. All single units with a classifier coefficient ≥ 45 were regarded as putative INs. This border was chosen somewhat arbitrarily on the basis of the distribution of classifier coefficients suggesting a bimodal character (Supplementary Fig. ^S2).

To study the response of single unit firing to light induced silencing of PVIs, we obtained peri-stimulus firing rate histograms aligned to the time point of light ignition (Fig. 1c; Supplementary Fig. ^S3). The resulting histograms were averages of ≥ 45 individual firing rate plots for every single unit. To check whether single unit firing was significantly altered by light stimulation, a Wilcoxon signed-rank test was used comparing the average firing rate in the 2 s before to the 2 s during light stimulation.

Temporal coupling of single unit activity to network oscillations was studied for the theta, gamma and ripple frequency range. First, for every spike time, the corresponding instantaneous phase of the theta, gamma and-if the spike occurred during a ripple phase—ripple oscillations was extracted from the Hilbert transform of the LFP after bandpass filtering the LFP in the respective frequency range⁵³ (theta 5–14 Hz, gamma 30–80 Hz, ripple 150–200 Hz). Spike times of all identified single units were allocated to the four different experimental conditions “light off/resting”, “light off/running”, “light on/resting”, “light on/running”. For a single unit to be included in the comparison of two conditions, the number of action potentials allocated to each of those conditions had to be ≥ 10. A Rayleigh test for circular uniformity was used to check for significant coupling (p < 0.05). To quantify the strength of spike time coupling, the spike train to field pairwise phase consistency (PPC) measure was employed^40,53. Differences in the preferred phase of firing between experimental conditions were statistically tested using a non-parametric multi-sample test for equal medians (circ_cmtest for Matlab)⁵⁴.