Task rule and choice are reflected by layer-specific processing in rodent auditory cortical microcircuits

Abstract

The primary auditory cortex (A1) is an essential, integrative node that encodes the behavioral relevance of acoustic stimuli, predictions, and auditory-guided decision-making. However, the realization of this integration with respect to the cortical microcircuitry is not well understood. Here, we characterize layer-specific, spatiotemporal synaptic population activity with chronic, laminar current source density analysis in Mongolian gerbils (Meriones unguiculatus) trained in an auditory decision-making Go/NoGo shuttle-box task. We demonstrate that not only sensory but also task- and choice-related information is represented in the mesoscopic neuronal population code of A1. Based on generalized linear-mixed effect models we found a layer-specific and multiplexed representation of the task rule, action selection, and the animal's behavioral options as accumulating evidence in preparation of correct choices. The findings expand our understanding of how individual layers contribute to the integrative circuit in the sensory cortex in order to code task-relevant information and guide sensory-based decision-making.

Fig. 1. Experimental design, learning curves and chronic CSD recording during auditory-based decision-making in a shuttle-box.

a Illustration of the two-way avoidance shuttle-box training with chronic recordings in behaving Mongolian gerbils. Subjects were trained to respond to two different pure tone frequencies (1 and 4 kHz; conditioned stimulus—CS) in a Go/NoGo task design to avoid an unconditioned stimulus (US—mild foot shock). During the discrimination phase the contingency of the CS can be either “Go” (CS+) or “NoGo” (CS−) leading to four possible behavioral outcomes (hit, miss, correct rejection—Corr. Rej., false alarm—FA). Right, Illustration of consecutive CS within a trial, length of the observation window (6 s), interstimulus interval (1.5 s) and behavioral choices. (Gerbil and loudspeaker images taken and modified from https://www.freepik.com/06/2019) (b). Averaged conditioned responses (CR) to both CS in the detection and discrimination phase as a function of training sessions (detection/discrimination: n = 9/8). During detection (gray area), hit rates reach almost 80% for both “Go”-stimuli (1 and 4 kHz). At the beginning of the discrimination phase (yellow area), conditioned responses dropped for both stimuli (<10% hit rate). The performance gradually increased reaching almost 80% for the hit rates, while false alarm rates stayed around 20%. Error bars indicate the standard error of mean (±s.e.m.). Single dots indicate CR rates of individual subjects. Supplementary Fig. 1 shows corresponding d′ learning curves. c Histogram with distributions of the averaged CR reaction times over all trials separately for the detection (top) and discrimination (bottom) phase and hits (red) and false alarms (blue). The majority of CR’s happen after the second CS presentation. d In vivo multichannel LFP recordings were obtained by single-shank silicon probes chronically implanted perpendicular to the surface of the auditory cortex targeting all cortical layers (I–VI). From laminar LFP signals single-trial current source density (CSD) distributions were calculated (here shown is a CSD averaged over 30 repetitions). During CS presentation (200 ms) tone-evoked CSD components appeared as current sink (in blue) and source (in red) activity reflecting the well-known feedforward information flow of sensory information in the A1^,. Supplementary Fig. 2 shows stability of CSD profiles recorded over the training period.

Fig. 2. Stimulus-related activity during different training phases.

a Representative example of an averaged CSD profile across all trials of the detection (left) and discrimination (right) phase of one subject. The CSD profiles show the tone-evoked activity after the first presentation of both conditioned stimuli within a trial (top: 1 kHz, bottom: 4 kHz; tone duration: 200 ms; indicated by dashed bar in upper left panel). Evoked CSD patterns between the two pure tones frequencies showed no obvious differences during the detection phase but yielded considerably different CSD patterns during discrimination for the CS+. b Corresponding raw AVREC traces (z-norm.) for the detection (left) and discrimination (right) phase for the two conditioned stimuli. The shaded error bars indicate the standard error of mean (±s.e.m) of the averaged AVREC traces. The raw traces between the two pure tones frequencies showed no obvious differences during the detection phase, but considerably different activity between CS+ and CS− trials during discrimination. c RMS values of the AVREC (time window of 500 ms beginning at each tone presentation and z-normalized) shown for each of the four consecutive CS and separated by the different behavioral outcomes during the two task phases (detection/discrimination: n = 9/8). Box plots represent median (bar) and interquartile range, and bars represent full range of data. Significance bar indicate differences revealed by pairwise testing (one-way rmANOVA; p < 0.05). Schematic illustration of the evoked cortical activity in dependence of stimulus frequency and task rule are shown in gray inserts.

Fig. 3. Behavioral choices and contingency are both reflected in population activity of the A1.

Averaged AVREC RMS values (500 ms window at CS onsets) plotted with respect to the conditioned stimuli and behavioral choice. a Left, During the detection phase evoked activity was significantly higher during hit trials compared to miss trials independent of the stimulation frequency (detection/discrimination: n = 9/8). Right, In the discrimination phase, cortical activity was strongest during correct hit trials and lowest during correct rejections. During trials of incorrect behavioral choices (miss/false alarm) tone-evoked activity was characterized by intermediate amplitudes and did not differ. Box plots represent median (bar) and interquartile range, and bars represent full range of data. Dots represent outliers. Significance bars indicate differences revealed by a two-way rmANOVA and corresponding posthoc tests with Holm-corrected levels of significance (see Supplementary Table 1) (b). In summary, cortical activity was generally higher in trials in which animals showed a conditioned response in comparison to trials where animals stayed in the compartment. Cortical activity differed strongest between correct behavioral choices, namely hits and correct rejections.

Fig. 4. Representation of contingency, not frequency revealed in synaptic population activity of granular input layers.

Parameters of interest were analyzed on a single-trial level using generalized linear-mixed effect models (a). Logistic regression curves show the probabilities of the presented CS (1 and 4 kHz as the dependent variable) for individual subjects (gray) and as an average (blue). The box plots above and below the curves represent the mean (bar), interquartile range (box), and the full range of data (whiskers). The AVREC trace RMS did not predict the frequency of the conditioned stimuli (1 and 4 kHz) during the detection phase (left, R²m = 0, R²c = 0, ns.). During discrimination an increase in the AVREC trace RMS significantly indicated that the 1 kHz “Go” stimulus was played (R²m = 0.16, R²c = 0.30; p < 0.001). Hence, auditory cortical activity in response to the same conditioned stimuli differed in dependence of the task. b GLMM’s were applied to RMS values measured within single-cortical layers (I/II, III/IV, Va, Vb, and VI). The illustration of the cortical column below indicates the GLMM predictability based on data from corresponding layers to the binary behavioral choice combinations. The color illustrates the effect size for the model-based R²m (gray = no effect to red = strong effect). The top R²m value (R²max) depicts the best fit result for all layers tested. In the detection phase, the two CS+ used as binary class in the GLMM revealed no significant prediction for any particular cortical layer. During the discrimination phase we observed a moderate prediction of the model with R²m = 0.12 for the granular input layers. For detailed results of each GLMM see Supplementary Table 2.

Fig. 5. Layer-specific contribution to behavioral choice.

a GLMM and logistic regression analysis was used to predict the behavioral choice of the subjects. Left, During the detection phase RMS values of the AVREC (z-norm.) in the 500 ms time window around the CS presentation which was initiating a hit response was significantly higher compared to the fourth CS during miss trials (R²m = 0.32, p < 0.001). Middle, This was also true for the discrimination phase, although with a more moderate effect size (R²m = 0.18, p < 0.001). When comparing data from “NoGo” trials, false alarm and correct rejections could be predicted with a high effect size (R²m = 0.27; p < 0.001). The box plots above and below the curves represent the mean (bar), interquartile range (box) and the full range of data (whiskers). b The illustration of the cortical column below indicates the GLMM predictability based on data from corresponding layers to the binary behavioral choice combinations. The color illustrates the effect size for the model-based R²m (gray/red scale). GLMM predictions for each layer showed that cortical activity from all layers were moderate to good predictors (R²m = 0.1–0.25; p < 0.001). Higher effect sizes were observed particularly at deeper layers Va, Vb, and VI, for the two possible choices (hit/miss). This finding was independent of the actual spectral content of the presented stimulus (1 kHz/4 kHz; see Fig. 3). During the discrimination phase, granular and supragranular layers appear to be important for the differential representation of the behavioral choice in “Go”-trials (R²m = 0.14–0.18; p < 0.001). For “NoGo”-trials, the GLMM revealed that false alarms are accompanied by significantly higher activity in all cortical layers except of layer VI compared to correct rejections (R²m = 0.17; p < 0.001). Supragranular layers were also the best predictor between false alarms and correct rejections classes. For detailed results of each GLMM see Supplementary Table 3.

Fig 6. Representation of choice accuracy across layer-specific population activity in A1.

a Predictability of correct (left) and incorrect (right) choices during the discrimination phase were modeled by GLMM and logistic regression. Correct “hit” responses can be very robustly predicted by higher RMS values of the AVREC trace in the time window before the actual decision compared to the time window at the trial end during correct rejection responses (R²m = 0.45; p < 0.001). In contrast, the two incorrect choices “false alarms” and “miss” were not predictable by the GLMM (R²m = 0.04; n.s.). The box plots above and below the curves represent the mean (bar), interquartile range (box) and the full range of data (whiskers). b The illustration of the cortical column below indicates the GLMM predictability based on data from corresponding layers to the binary behavioral choice combinations. The color illustrates the effect size for the model-based R²m (gray/red scale). Activity from all cortical layers contributed to the differential cortical activation between the correct choice classes, while the largest effect size was found for supragranular layers (R²m = 0.51; p < 0.001). In accordance with the insignificant GLMM result on the overall columnar activity measured by the AVREC, also no cortical layer activity could predict the two incorrect choices (false alarm/miss). For detailed results of each GLMM see Supplementary Table 4.

Fig 7. Time-resolved GLMM-based effect sizes of behavioral outcomes reflecting accumulating evidence over the trial duration.

GLMM-based R²m values for behavioral choices are plotted for time bins before an animal’s reaction (small inset top left). Dashed lines indicate small, moderate, and large effect sizes, while the color of circles indicates the corresponding p value of each GLMM (black: p < 0.001, dark gray: p < 0.01, light gray: p < 0.05 and white: n.s.). We found R²m values to generally increase over the trial duration until a behavioral choice option was made. During detection, infragranular layers Va–VI showed moderate R²m values even at 2–3 CS+ presentations before the actual response was commuted. Layers III/IV and I/II only allowed moderate predictions of the animal’s choice at the CS+ presentation preceding the reaction. During discrimination, the predictability between hits and misses were considerably less pronounced and time-resolved. Activity in cortical layers I/II and III/IV, however, allowed to correctly predict the occurrence of correct rejections of up to three CS presentations before the animal’s reaction. Such temporally dispersed evidence was particularly pronounced in upper layers in contrast to false alarm trials. Largest effects were found when comparing hit vs. correct rejection responses revealing the accumulating evidence of choice accuracy over the trial duration. For incorrect decisions, the model showed no change of low predictability over the trial length.