Short Time-Scale Sensory Coding in S1 during Discrimination of Whisker Vibrotactile Sequences

Abstract

Rodent whisker input consists of dense microvibration sequences that are often temporally integrated for perceptual discrimination. Whether primary somatosensory cortex (S1) participates in temporal integration is unknown. We trained rats to discriminate whisker impulse sequences that varied in single-impulse kinematics (5-20-ms time scale) and mean speed (150-ms time scale). Rats appeared to use the integrated feature, mean speed, to guide discrimination in this task, consistent with similar prior studies. Despite this, 52% of S1 units, including 73% of units in L4 and L2/3, encoded sequences at fast time scales (≤20 ms, mostly 5-10 ms), accurately reflecting single impulse kinematics. 17% of units, mostly in L5, showed weaker impulse responses and a slow firing rate increase during sequences. However, these units did not effectively integrate whisker impulses, but instead combined weak impulse responses with a distinct, slow signal correlated to behavioral choice. A neural decoder could identify sequences from fast unit spike trains and behavioral choice from slow units. Thus, S1 encoded fast time scale whisker input without substantial temporal integration across whisker impulses.

Fig 1. Whisker stimuli and behavioral apparatus.

(A) Schematic of training apparatus, showing the rat’s right whiskers resting on the moveable stimulus panel. (B) Panel kinematics for fast, medium, and slow impulses. Circles indicate maximum velocity. (C) Panel kinematics for FFF, FMS, SMF, and SSS sequences. Data for this panel are in S1 Data. (D) Mean speed, total duration, and first pulse peak velocity for the four sequences. SMF and FMS sequences had similar mean speed and duration (dashed lines).

Fig 2. Behavioral performance on FFF-FMS-SMF-SSS discrimination task.

(A) Discrimination performance for one example rat, across 13 d of training (44–50 trials for each stimulus per day). FMS and SMF stimuli were first introduced on Day 0. The rat reliably discriminated FFF from SSS stimuli but treated FMS and SMF stimuli identically and at chance. The rat responded similarly to all stimuli when the panel was fixed, and thus was not discriminating based on piezo auditory cues. (B) Mean performance (± SEM) for all rats across all behavior sessions. (C) Relative right drink port choice as a function of mean panel speed over the entire 150-ms sequence. Each symbol is a different rat (n = 8). (D) D-prime analysis of FFF versus SSS discrimination in fixed-panel control experiments versus normal sessions. Solitary points show rats not tested on the fixed-panel control. Data for this figure are in S1 Data.

Fig 3. Behavioral performance on FSFS-SFFS discrimination task.

(A) Panel kinematics for SFFS and FSFS sequences, showing both different-intensity and same-intensity versions. (B) Behavioral performance across all behavioral sessions, for the two rats trained on this task. Open symbols: sessions using the different-intensity version of the stimuli. Filled symbols: the same-intensity version. Both rats could discriminate the different-intensity version but not the same-intensity version. (C) D-prime analysis of discrimination performance for the same two rats (circles: 60W, squares: 58B), across all behavioral sessions. Symbols are mean ± SEM across sessions. Data for this figure are in S1 Data.

Fig 4. S1 recordings during discrimination behavior.

(A) Schematic of multi-tetrode chronic microdrive. (B) Cluster separation for one recording site (top) with mean spike waveforms for three simultaneously recorded single units (bottom left). Bottom right, mean spike waveform for all fast-spike (FS) and regular-spike (RS) single units. (C) Laminar distribution of recorded units. (D) Population peri-stimulus time histogram (PSTH) for all temporally modulated units by layer and stimulus type. Different sequences have different onset times for impulses 2 and 3 (colored ticks). Data for this figure are at crcns.org repository (accession ssc-4).

Fig 5. Sequence responses for example units.

(A and B) L4 multi-unit and L5b RS single unit with phasic response to each impulse. (C and D) L6 multi-unit and L5b RS single unit with increasing firing rate during the stimulus period. Each panel shows the spike raster and PSTH across trials, for one stimulus sequence. Vertical lines: onset of each impulse. Data for this figure are at crcns.org repository (accession ssc-4).

Fig 6. Classification of S1 units by stimulus regression.

(A–C) Stimulus regression for three example units. Top, PSTH in 5 ms time bins. Bottom, stimulus panel speed integrated over 5, 15, or 180 ms, which was the best fit stimulus integration window for each unit. Right, regression of firing rate on integrated stimulus speed. (D) Coefficient of determination (R²) for all stimulus integration windows with a significant regression, for each unit with a significant regression to at least one window. Black: Fast and Medium time scale units (best integration window <55 ms). Blue: Slow units. Cells are sorted by peak R² and by sign of the regression slope for the best integration window. (E) Number of units with each best integration window and positive or negative regression slope. (F,G) Laminar distribution of units by best integration window. Data for this figure are in S1 Data.

Fig 7. Stimulus coding by fast time scale units.

(A) Population PSTH (mean ± SEM) for Fast units with 5, 10, and 15–20 ms best integration windows. (B) Population PSTH for all individual F, M, or S impulses, irrespective of sequence membership, for all Fast units. Dashed line: pre-impulse firing rate. (C) Left: net evoked rate for individual impulses, calculated as post-impulse rate–pre-impulse rate. Right: mean rate across the entire sequence above pre-stimulus baseline, as a function of mean panel speed. Symbols show mean ± SEM across units. Line: regression. (D) Population PSTH for Medium units for FFF, FMS, SMF, and SSS sequences. (E) Net evoked rate for individual impulses and mean rate across the sequence for Medium units. Conventions as in C. Firing rate was suppressed by all impulses and sequences. Data for this figure are at crcns.org repository (accession ssc-4).

Fig 8. Choice coding by slow time scale units.

(A) Top: population PSTH (mean ± SEM) for Slow Positive units across all layers. Responses were indistinguishable between FFF, FMS, SMF and SSS trains. Bottom: population PSTH for individual F, M, and S impulses, irrespective of sequence membership. (B) Population PSTH for slow negative units, showing lack of any impulse-evoked firing rate modulation. (C) Population PSTH for slow positive units in L5a and L5b, separated by stimulus type and drink port choice. Slow Positive units fired more on right-choice trials for all stimuli. (D) Difference in evoked rate between right- and left-choice trials, measured 5–50 ms after start of the final impulse, for all Slow Positive units (left) or Fast and Medium units (right). Number of units in each layer is shown at bottom. Open symbols, baseline rate before sequence onset for the same trials. * p = 0.022; ** p = 9.5 x 10^-4; *** p = 7.5 x 10^-5, paired t test comparing rate on right versus left choice trials. (E) Population PSTH averaged across all four sequences, for right- versus left-choice trials, for Slow Positive units in L5 (top), and for Fast units in L4 (bottom). Bar shows times when rate is significantly different between right- and left-choice trials by sliding t test (p < 0.05). The distribution of nose poke withdrawal times is shown for the same trials. Data for this figure are at crcns.org repository (accession ssc-4).

Fig 9. Population decoding of sequence identity and behavioral choice.

(A) Decoder architecture for stimulus decoding. Each neuron was represented by a one-versus-all (OVA) classifier, trained by logistic regression to calculate the probability of each stimulus given a single-trial spike train (blue trial). The stimulus with the highest summed probability across neurons was taken as the population prediction. (B) Average performance of stimulus decoder with 10 ms time bins, constructed from all Fast and Medium units or all Slow units. The title reports average percent correct classification across all four stimuli. Entries along the diagonal are percent correct, and rows sum to 1.0. (C) Percent correct performance for each unit in the Fast/Medium model in (B), separated by best integration window. Vertical line, chance prediction of 25%. (D) Average performance of stimulus decoder with a single 150 ms time bin. (E) Average performance of a Fast/Medium stimulus decoder with 10-ms bins, using rate-normalized or time-scrambled spike trains. (F) Performance of behavioral choice decoders, built from Fast and Medium units or Slow units, using two different bin sizes. Chance performance is 50%. (G) Choice prediction for each unit in the Slow model, using a single 100-ms bin prior to nose poke withdrawal. Units are separated by depth and response type. Best choice prediction was by Slow Positive units in L5b. Data for B–E are in S1 Data.