Neural oscillations in the fronto-striatal network predict vocal output in bats

Abstract

The ability to vocalize is ubiquitous in vertebrates, but neural networks underlying vocal control remain poorly understood. Here, we performed simultaneous neuronal recordings in the frontal cortex and dorsal striatum (caudate nucleus, CN) during the production of echolocation pulses and communication calls in bats. This approach allowed us to assess the general aspects underlying vocal production in mammals and the unique evolutionary adaptations of bat echolocation. Our data indicate that before vocalization, a distinctive change in high-gamma and beta oscillations (50-80 Hz and 12-30 Hz, respectively) takes place in the bat frontal cortex and dorsal striatum. Such precise fine-tuning of neural oscillations could allow animals to selectively activate motor programs required for the production of either echolocation or communication vocalizations. Moreover, the functional coupling between frontal and striatal areas, occurring in the theta oscillatory band (4-8 Hz), differs markedly at the millisecond level, depending on whether the animals are in a navigational mode (that is, emitting echolocation pulses) or in a social communication mode (emitting communication calls). Overall, this study indicates that fronto-striatal oscillations could provide a neural correlate for vocal control in bats.

Fig 1. Properties of echolocation pulses and communication calls produced by bats.

(a) Exemplary acoustic recording including an isolated call and a vocalization train. Zoomed-in views have been included in (b) and (c) to show spectrograms of the syllable train and the isolated echolocation pulse. Panels (d) and (e) display two further examples of isolated vocalizations (communication calls in this case). Panel (f) shows a combined histogram of sound duration and peak frequency in echolocation and communication sounds. Note that these two call types are well segregated in the frequency domain. The latter is also noticeable in the average call spectra shown in (g). Data underlying this figure can be found at https://doi.org/10.12751/g-node.6a0d94.

Fig 2. LFPs during vocalization in the CN and FAF.

(a) Mean LFP (± SEM) of all isolated communication calls (n = 628) studied. Signals from all three channels of the striatum were pooled together thus rendering a higher number of responses for the striatum than for the FAF. (b) Mean LFP (± SEM) obtained during the production of isolated echolocation pulses (n = 493) in the striatum. (c) and (d) Colormaps showing the mean of z-scored LFPs in the FAF across cortical depths, 500 ms before and after communication calls (c) and echolocation pulses (d). Data underlying this figure can be found at https://doi.org/10.12751/g-node.6a0d94. CN, caudate nucleus; FAF, frontal auditory field; LFP, local field potential.

Fig 3. Spectral differences in neural activity obtained in the CN during echolocation and communication production.

(a)–(b) Power spectrogram in the CN during communication (a) and echolocation (b). Mean values of 10,000 randomization trials are displayed in each case. (c) Colormap representing the Cliff’s Delta values of echolocation versus communication comparisons at each time point and frequency. Gray outlined regions mark areas with a medium effect size (Cliff’s Delta > 0.33 [60]). Red colors indicate more power in the LFPs during echolocation than communication. Blue regions indicate the opposite trend. The LFPs underlying this figure can be found at https://doi.org/10.12751/g-node.6a0d94. CN, caudate nucleus; LFP, local field potential.

Fig 4. Time-frequency differences in power distributions across FAF channels, depending on the vocalization type.

(a)–(d) LFP spectrograms of four illustrative channels of the FAF for the communication condition (n = 10,000 randomization trials, see Methods). (e)–(h) Spectrograms obtained in the same four example channels during echolocation. (i)–(l) Colormaps of Cliff’s Delta values obtained when comparing the time-frequency dynamics in the echolocation and communication conditions in the four example channels. Black highlighted regions indicate large effect size (d > 0.47). Gray indicates medium effect size (d > 0.33) [60]. (m)–(o) Mean Cliff’s Delta values across FAF depths. Mean values were obtained for all the frequencies that composed the theta (4–8 Hz), beta (12–30 Hz), and gamma (30–80 Hz) bands, represented in panels m, n, and o, respectively. This figure was created based on data that can be found at https://doi.org/10.12751/g-node.6a0d94. FAF, frontal auditory field; LFP, local field potential.

Fig 5. LFP power differences during the production of LHF communication calls and echolocation pulses.

LHF communication calls carry pronounced energy at both low (<50 kHz) and high frequency (see median vocalization spectra in (a)). LF calls carry power only at low frequencies, while echolocation pulses carry power at high frequencies. (b) Cliff’s Delta effect size measures obtained in the CN when comparing LHF versus echolocation sounds. (c)–(d) Similar to panel (b), but for the FAF at 300-μm and 800-μm depths, respectively. (e)–(f) Average Cliff’s Delta across FAF depths in the beta and gamma ranges, respectively. Overall, the results obtained when comparing neural activity related to LHF and echolocation call production resembled those obtained when pooling data from all communication calls (see Fig 4; for results of comparing LHF and LF calls, see S6 Fig). Data underlying this figure can be found at https://doi.org/10.12751/g-node.6a0d94. CN, caudate nucleus; FAF, frontal auditory field; LF, low-frequency; LFP, local field potential; LHF, low- and high-frequency.

Fig 6. LFP signals leading to vocalization can be used to predict vocal output.

(a) Prediction accuracy calculated using a binary SVM classifier (see Methods) trained with LFP information (filtered by frequency band) occurring before vocalization in the echolocation and communication conditions (all communication calls were pooled together). Models were trained with half of the data (n = 5,000 randomization trials in each vocalization condition). The other data half was used for calculating the models’ prediction accuracy. (b) Same as panel (a), but in this case the models had to classify post-vocalization activity. Note that in the post-vocalization condition, prediction accuracy dropped in the striatum. In the FAF, accuracy was still highest in deep layers in the gamma range. (c) Same as (a) and (b), but here the model had to predict a third dataset corresponding to pre-vocalization activity in trials with contaminated post-vocalization time (training set was the same as in (a)). Even in this case, FAF signals rendered good predictions about ensuing vocal output. The SVM was computed based on data that can be found at https://doi.org/10.12751/g-node.6a0d94. CN, caudate nucleus; FAF, frontal auditory field; LFP, local field potential; SVM, support vector machine.

Fig 7. Functional coupling between the FAF and the CN during vocalization.

Time-frequency resolved coherence between the striatum and four exemplary channels of the FAF at (a) 100-μm; (b) 300-μm; (c) 500-μm; and (d) 800-μm depths during communication (n = 628 trials). Black lined regions refer to the 95th percentile of all computed coherence values during vocalization. (e) Time resolved coherence strength between both structures across cortical depths in theta and alpha (panel (f)) during the production of communication calls. Mean coherence values across frequencies in each range were calculated. (g)–(j) Coherograms in four example channels during echolocation pulse production (n = 493 trials). (k)–(l) Same as (e) and (f) but for the echolocation case. During echolocation production, pronounced coherence in theta in the top-to-middle layers was found 200 ms after call onset. This temporal pattern differs from that observed during the production of communication calls. The coherence was computed based on data that can be found at https://doi.org/10.12751/g-node.6a0d94. CN, caudate nucleus; FAF, frontal auditory field.

Fig 8. Spiking activity in the CN and FAF during vocalization.

Spiking probability (computed as numbers of spikes per trial per bin, binsize = 3 ms) in the CN 500 ms before and after communication ((a), n = 628 trials) and echolocation ((b), n = 493). Spiking across all channels recorded in the FAF during communication (c) and echolocation (d). In the FAF, during both types of vocalizations, distinct spiking activity could be identified in deep layers. One reason for the small increase in spiking activity in response to the vocalization could be due to the sparse distribution of vocalization relevant neurons in the FAF. Data underlying this figure can be found at https://doi.org/10.12751/g-node.6a0d94. CN, caudate nucleus; FAF, frontal auditory field.

Fig 9. Spike-phase locking across vocalization conditions.

(a) dVS obtained before vocalization onset in the communication-surrogate condition in the CN; (b) echolocation-surrogate condition; and (c) echolocation-communication condition. (d)–(f) dVS values computed for all recorded channels in the FAF in the three conditions mentioned above. Statistical differences were tested by comparing VS distributions (Wilcoxon rank-sum tests with Bonferroni correction, *p < 0.001, see Methods). Data underlying this figure can be found at https://doi.org/10.12751/g-node.6a0d94. CN, caudate nucleus; dVS, difference in vector strength; FAF, frontal auditory field; VS, vector strength.

Fig 10. Visual abstract depicting the main results presented in this study.

“SG,” “G,” and “IG” indicate a putative subdivision of the FAF into supragranular, granular, and infragranular layers, respectively. Note that we do not report data on the directionality of the connection between both regions, and thus functional coupling is displayed with a double arrow. Different electrophysiological parameters such as LFP power measurements, spike-phase locking and LFP–LFP coherence demonstrate that fronto-striatal circuits can predict ensuing vocal output in bats. CN, caudate nucleus; FAF, frontal auditory field; LFP, local field potential.