A subset of VPA-exposed rodents (n = 8) underwent 20 days of VNS-sound pairing prior to electrophysiological recording. Previous work has varied the duration, pulse width, frequency, and intensity of vagus nerve stimulation to identify optimal parameters for inducing synaptic plasticity (Borland et al., 2016; Buell et al., 2018, 2019; Loerwald et al., 2018). Based on their findings, we delivered 500 ms, 100 μs biphasic 16 pulse, 30 Hz, 0.8 mA VNS paired with the speech sound “dad” in a double-walled sound attenuated booth. The onset of VNS preceded the onset of the speech sound by 50 ms. Pairings were randomly interleaved with silence trials for an average intertrial interval of 30 s, and stimulation was delivered via A-M Systems Isolated High Power Stimulators (Model 4100; Sequim, WA). Rats received 300 VNS-sound pairings per 2.5 h session. These methods are identical to our previous publications (Engineer et al., 2011; Borland et al., 2016, 2019, 2023; Buell et al., 2018).

Multi-unit extracellular activity was recorded from 357 sites across 10 saline (SAL)-exposed rats, 330 sites across 10 VPA-exposed rats, and 285 sites across 8 VNS-paired VPA-exposed rats. Rats were anesthetized for recordings with sodium pentobarbital (50 mg/kg body weight) and supplemented with dilute pentobarbital (8 mg/kg body weight) as needed. Vitals were monitored using MouseOx Plus Oximeter for Rodents (STARR Life Sciences; Oakmont, PA) and body temperature was maintained at 37° C. Prior to recording, rats received a tracheotomy and cisterna drain to ease breathing and reduce brain swelling. A cranial window was opened and the dura was resected to expose right AAF. Two 2×1 Parylene-coated tungsten microelectrodes (FHC, 1–2 MΩ) were lowered into layer 4/5 (~600 μm) of the cortex, and a speaker was placed 10 cm from the left ear. A stimulus set consisting of 25 ms tone pips ranging in frequency from 1-32 kHz and in intensity from 0-75 dB SPL, speech sounds, and noise burst trains (six 25 ms bursts, at a rate of 7.5, 10, 12.5, or 15 Hz) were presented (Engineer et al., 2014a, 2014b; Borland et al., 2016, 2019, 2023; Buell et al., 2018, 2019; Adcock et al., 2020b; Tamaoki et al., 2024). Speech sounds (“dad,” “bad,” “gad,” “tad,” “sad,” “rad,” and “lad”) and noise burst trains were pseudo-randomly repeated 20 times each. Neural responses passed through an RA16 preamplifier and were recorded with BrainWare (TDT). AAF was identified based on its reversed tonotopy compared to primary auditory cortex (A1) (with low frequency anterior and high frequency posterior) and fast responses (10-15 ms onset latency) (Polley et al., 2007). Similarly, a reversal in tonotopy, the loss of tuning or response strength, delay of response latency, and monotonicity was used to border AAF (Polley et al., 2007; Centanni et al., 2013; Engineer et al., 2014a, 2014b; Shi et al., 2019).

Utilizing responses to tones, we characterized receptive field properties at each recording site, including: the characteristic frequency; the lowest threshold (dB SPL) tone to evoke a response at the characteristic frequency; the bandwidth of neuron tuning at 10-40 dB SPL above the threshold; response onset and peak latency at 60 dB SPL; the percentage of recording sites responding to tone frequencies within each of 5 one-octave frequency bins (1-2, 2-4, 4-8, 8-16, and 16-32 kHz); the number of spikes evoked per tone within +/− a half octave from the characteristic frequency of the recording site; and the rate level function. For recordings to speech sounds, we compared the average driven response to the onset of the consonant (1-40 ms), response onset and peak latency, and neural classifier accuracy between groups. For recordings to noise bursts, we compared average firing rate, steady state responses to alternating bursts, response onset latency, Rayleigh statistic, vector strength, and paired-pulse ratio between groups. Steady state responses to noise burst trains were calculated by averaging the driven firing rate from bursts three onwards (366-433 ms depending on repetition rate) (Regan, 1966). Rayleigh statistic was used to describe response uniformity across time in a circular space. Recording sites with a Rayleigh statistic above 13.8 were considered phase locked since responses deviated from uniformity in synchrony with the stimulus. Vector strength was used to describe the degree of separation from uniform (i.e., the quality of the deviation) (Pandya et al., 2008; Shetake et al., 2012). Paired-pulse ratio was calculated by dividing the driven response to the second noise burst in the train by the first (P2/P1) (Debanne et al., 1996).

Because the data contained multiple levels of hierarchical nested data with an unequal number of repeated measures (e.g., recording sites), we analyzed the data with linear mixed models in R (v. 4.3.3 – 4.4.0) and RStudio (v. 2024.4.1) (Bolker et al., 2009; Pinheiro and Bates, 2009). Data was tested for normality with the Shapiro–Wilk or Shapiro-Francia test depending on its kurtosis (Mbah and Paothong, 2015). In cases where the response variable was normally distributed, the Lme4 package (v. 1.1-35.3) was used for linear mixed effect models (Bates et al., 2015). In cases where the data was not normally distributed, the glmmTMB package (v. 1.1.9) was used for generalized linear mixed effects models (glmm) (Brooks et al., 2017). The response variable (e.g., number of action potentials, rate of action potential firing, response bandwidth, latency to respond, etc.) was dependent and compared across the fixed effect of experimental group. Where applicable, the model was updated to include additional fixed effects and interactions (e.g., tone intensity, tone frequency, noise burst repetition rate). Animal and recording sites were treated as random effects, were nested, and had a fixed intercept and a random slope. When data was non-normally distributed, family was selected based on the distribution of the data. Since our response variable was neuronal activity, the Tweedie family with a log link was used for most of the analysis (Moshitch and Nelken, 2014). In cases where the data was non-normally distributed, not continuous, and did not include negative numbers or zero inflation, Beta Family and Binomial distributions were used. Akaike information criterion (AIC), Bayesian information criterion (BIC), R-squared, covariance structure, and residual diagnostics were used to compare model fit while avoiding overfitting the data. Once an appropriate model was selected, Type II Wald Chi-square tests were performed to determine if fixed effects were significant predictors of the response variable (car package v. 3.1.2) (Fox and Weisberg, 2019). Following a significant main effect, Tukeys corrected pairwise comparisons were performed with the emmeans package (v. 1.10.1) (Lenth et al., 2018). Data was visualized with GraphPad Prism (v. 10.2.3) and Matlab (v. 2022a).

During the first stage of training, rats learned to nose-poke for a 45 mg nutritionally complete sugar pellet (Bio-Serv; Flemington, NJ). After independently poking for 100 pellets per session for two sessions, rats learned to nose-poke when they detected the target word and refrain from poking during silence catch trials. As they progressed through the training stages, the hit-window for correct nose-pokes decreased from 4 s to 3 s. If the animal poked during a silence catch trial or >3 s after the sound presentation, it received a 6 s “timeout” during which the booth lights shut off and pokes elicited no feedback. Rats completed an average of 17 (± 3 SD) one-hour sessions of training on detection until they could reliably identify the target from silence catch trials, reaching a performance criterion of 75% correct for four sessions. After reaching proficiency, they were removed from training for cuff implantation surgery (Engineer et al., 2013, 2014b; Carroll et al., 2024).

After recovery from cuff and headcap implantation, rats began a series of sound discrimination tasks where they learned to discriminate the target sound “dad” from similar non-target sounds differing by initial consonant: “bad,” “gad,” “tad,” and “sad” (Engineer et al., 2013, 2014a, 2014b; Adcock et al., 2020a; Carroll et al., 2024). A subset of VPA-exposed rats (n = 8) received success-paired VNS during this task (Bowles et al., 2022). After 20 days of training with two sessions per day, all rats advanced to increasingly complex discrimination tasks involving multiple speakers, sounds that were truncated, compressed, or presented in noise (Engineer et al., 2013; Adcock et al., 2020a). These tasks are described in detail in the Supplementary Figures legend.

During the first sound discrimination task, VNS was triggered with pellet delivery after nose-poking to the target sound. The total number of stimulations was dependent on the animal’s performance and was on average 92 (± 28 SD) per session.

Behavioral performance was quantified as percent correct. Data was tested for normality with Shapiro–Wilk or Shapiro-Francia depending on its kurtosis (Mbah and Paothong, 2015). All behavioral data was normally distributed and analyzed with Bonferroni corrected two-way repeated measures ANOVA. Data was analyzed and visualized with GraphPad Prism (v. 10.2.3).

Receptive field properties describe the sensitivity and selectivity of neurons to simple sound characteristics. A failure to encode simple sound characteristics (e.g., frequency or loudness) could lead to widespread processing impairments of complex sounds. To quantify receptive field properties, we recorded AAF responses to tone pips ranging from 1-32 kHz frequency and 0-75 dB SPL intensity. When we compared the latency of response onset to each tone, there was a main effect of group (χ2 = 6.41, df = 2, p = 0.04) and post hoc comparisons revealed a significant delay in response onset among VPA-exposed rats compared to SAL-exposed rats. This delay was partially rescued in VNS-paired VPA-exposed rats who were no longer significantly different from SAL-exposed controls or untreated VPA-exposed rats (Figure 1A; Table 1). Likewise, when peak response latency was compared, a significant main effect of group was observed (χ2 = 6.79, df = 2, p = 0.03) and post hoc comparisons revealed a significant delay in peak of response among VPA-exposed rats which was partially restored among VNS-paired VPA-exposed rats (Figure 1A; Table 1). We next investigated whether response strength was altered as a function of sound intensity since children with ASD often exhibit hypo or hypersensitivity to sounds, which is often intensity dependent. When comparing response strength across sound intensity, we observed robust differences in response strength across intensities, with main effects of both group (χ2 = 105, df = 2, p < 0.0001) and intensity (χ2 = 6,017, df = 15, p < 0.0001), and group x intensity interaction (χ2 = 194, df = 30, p < 0.0001). Post hoc comparisons revealed significant between group differences at each intensity level, with VPA-exposed rats responding significantly weaker than SAL-exposed controls and VNS-paired VPA-exposed controls. VNS-sound pairing partially restored response strength across intensity levels, responding significantly stronger than untreated VPA-exposed rats, but in most cases still significantly weaker than SAL-exposed rats (Figure 1B; Supplementary Table 2).

Improvements in response latency and strength among VNS-paired VPA-exposed rats arose without changes to the sensitivity or selectivity of cortical neurons. When comparing response threshold, there was a main effect of group (χ2 = 6.48, df = 2, p = 0.03), but post hoc comparisons revealed no significant between group differences (Figure 1C; Table 1). Across all receptive field bandwidths we observed no main effects, suggesting no differences in neuron frequency tuning between groups (Supplementary Figure 1A; Supplementary Table 1). Next, utilizing the characteristic frequency of recording sites, we calculated the percentage of recording sites for each animal tuned to frequencies within each of 5 one-octave frequency bins (1-2, 2-4, 4-8, 8-16, and 16-32 kHz). There was a main effect of frequency octave (χ2 = 269, df = 4, p < 0.0001) but no effect of group, indicating there was no group differences in the percentage of AAF responding to different octaves (Supplementary Figure 1B; Supplementary Table 1). Similarly, when we compared the number of evoked spikes within each octave there was a main effect of octave (χ2 = 174, df = 4, p < 0.0001) but no significant effect of group (χ2 = 4.73, df = 2, p = 0.09), and no octave x group interaction (χ2 = 12.23, df = 8, p = 0.14) (Supplementary Figure 1C). Furthermore, we compared the spontaneous activity occurring in the 100 ms prior to tone onset, and observed no main effects, suggesting no group differences in spontaneous activity which could account for the observed differences in response strength (χ2 = 0.89, df = 2, p = 0.63). In summary, VPA-exposed rats exhibited responses to pure tones which were weaker and delayed when compared to their SAL-exposed peers. VNS-paired VPA-exposed rats exhibited a partial restoration of both response latency and response strength. These improvements were not driven by changes in the sensitivity or selectivity of AAF neurons and cannot be explained by changes in spontaneous activity or frequency representation.

The integration of both spectral and temporal information is necessary for complex sound processing. To assess the ability of AAF neurons to track rapid temporal changes, we recorded responses to four noise burst trains varying in repetition rate (7.5 – 15 Hz) (Figure 2A). When we compared the driven response evoked by each noise burst in each train, we observed robust effects of both group and noise burst repetition rate on driven response strength (Supplementary Figure 2; Supplementary Table 3). Post hoc comparisons revealed a compounding effect of repetition rate on firing rate, where VPA-exposed rats had significantly decreased firing to the third and fifth noise burst at the fastest repetition rate. These findings are further illustrated in Figure 2B which highlights the difference in steady state driven responses across alternating noise bursts in the train. Here VPA-exposed rats exhibit diminished steady state driven responses to odd and not even bursts in the noise train. VNS-paired VPA-exposed rats had a full restoration of this temporal processing degradation and responded significantly stronger than untreated VPA-exposed rats across all bursts in each train (Supplementary Figure 2; Supplementary Table 3). Since we observed weak driven responses among VPA-exposed rodents to alternating noise bursts in the trains, we further characterized temporal processing by assessing phase locking, vector strength, and paired-pulse ratio.

When we compared the ability of recording sites to phase lock, there was a main effect of repetition rate (χ2 = 8.147, df = 3, p < 0.0001), and a group x repetition rate interaction (χ2 = 12.6, df = 6, p = 0.04). On average across repetition rates, post hoc comparisons revealed a significant decrease in the average proportion of recording sites phase locked to noise burst trains among VPA-exposed rats compared to SAL-exposed controls. VNS-paired VPA-exposed rats exhibited a partial restoration of phase locking and were no longer significantly different from either group (Figure 2C; Table 1; Supplementary Table 3). To determine the phase synchrony of responses, we calculated the vector strength of recording sites. Here there was a main effect of repetition rate (χ2 = 950, df = 3, p < 0.0001), but no main effect of group (χ2 = 3.18, df = 2, p = 0.20, Figure 2D). Since phase synchrony and phase locking directly measure the successive probability of firing an action potential to repeating stimuli, we calculated paired-pulse ratio to determine whether the proportion of neurons firing action potentials differed between bursts in the train. There were significant main effects of group (χ2 = 8.02, df = 2, p = 0.01), repetition rate (χ2 = 834, df = 3, p < 0.0001), and group x repetition rate interaction (χ2 = 21.3, df = 6, p = 0.02) on paired-pulse ratio. On average across repetition rates, post hoc comparisons revealed that compared to SAL-exposed rats, VPA-exposed rats had a significantly higher probability of firing an action potential to the second burst of the noise train than the first. VNS-paired VPA-exposed rats had a full restoration of paired-pulse ratio and were no longer different from SAL controls (Figure 2E; Table 1; Supplementary Table 3). In summary, VPA-exposed rats exhibit a diminished ability to process fast repetition rate noise burst trains. This extended to a significant decrease in the proportion of phase locked neurons and a significantly higher paired-pulse ratio. These changes to temporal processing were largely restored following VNS-speech pairing. Next, we assessed whether these improvements in temporal processing led to improvements in the processing of spectrotemporally complex speech sounds.

Clinical literature on developmental disorders and specific language impairments often attribute speech perception difficulties to stop consonants specifically (Tallal, 2004; Hornickel et al., 2009). We grouped our speech sounds by manner of articulation and isolated our analysis to the four stop consonants presented (“b,” “t,” “g,” and “d”). We observed no main effect of group on onset (χ2 = 0.10, df = 2, p = 0. 49) or peak latency (χ2 = 0.96, df = 2, p = 0.61), suggesting no differences in response latency between groups (Figures 3A,B; Table 1). We next compared response strength driven by the onset of the consonant (1-40 ms), where there was a strong main effect of group (χ2 = 21.1, df = 2, p < 0.0001), and post hoc comparisons revealed significantly weaker driven activity to stop consonants among VPA-exposed rats compared to SAL controls. This deficit in response strength was completely rescued among VNS-paired VPA-exposed rats which responded significantly stronger than their untreated counterparts and were no longer different from SAL-exposed controls (Figure 3C; Table 1). Comparison of the 300 ms driven response to the vowel and the entire 400 ms stop consonant initial speech sound revealed significant main effects of group (χ2 = 19.5, df = 2, p < 0.0001; χ2 = 24.6, df = 2, p < 0.0001). Post hoc comparisons showed this was due to an increase in response strength among VNS-paired VPA-exposed rats and there was no decrease in response strength to the vowel portion of the sound for untreated VPA-exposed rats compared to SAL controls (Figures 3D,E; Table 1). These findings confirm that the rapid acoustic changes in spectral energy present in the first 40 ms of stop consonants are responsible for the degraded driven activity we observed among VPA-exposed rats. Across all speech sounds, VNS-paired VPA-exposed rats displayed a generalized increase in response strength, resulting in a full restoration of consonant processing.

Next, we examined neural discrimination of the same sounds, to determine whether VNS driven increases in response strength would impact neural discriminability of stop consonants. A neural classifier was trained to categorize spatiotemporal patterns of activity across 19 of 20 speech sound repeats, then asked to correctly identify the consonant presented from the final repeat of neural activity (Engineer et al., 2008). There was a significant main effect of group (χ2 = 9.19, df = 2, p = 0.01) on classifier performance, and post hoc comparisons revealed that recording sites from VPA-exposed rats were significantly less accurate at consonant classification compared to SAL-exposed control rats. Neural classifier discrimination was fully rescued among VNS-paired VPA-exposed rats, who were significantly different from VPA-exposed rats and no different from SAL-exposed control rats (Figure 4A; Table 1). Improved performance on the neural classifier among VNS-paired VPA-exposed rats suggests that in addition to increasing driven response strength, VNS paired with the speech sound “dad” also led to more distinct neural activity evoked by stop consonants in VPA-exposed rats.

We trained a separate set of rats to discriminate speech sounds and one group of VPA-exposed rats received success-paired VNS. Throughout the four weeks of training, we observed a significant effect of training week (ANOVA, weeks, F (3, 75) = 296.6, p < 0.0001), but no effect of treatment (ANOVA, treatment, F (2, 25) = 0.73, p = 0.49) and no week x treatment interaction (ANOVA, Weeks x Treatment, F (6, 75) = 1.205, p = 0.3133) (Figure 4B). We hypothesized that success-paired VNS may improve the generalizability of training to a variety of sound contexts, and tested the rats on additional discrimination tasks, including speech-in-noise, truncated consonants (cut to 40 ms), compressed speech, or speech sounds spoken by multiple male and female talkers. Across all tasks performance remained comparable between groups, and we failed to reveal any treatment-related differences (Supplementary Figure 4; Supplementary Table 5; additional methods in figure legend). These findings suggest that in the absence of any behavioral impairment success-paired VNS does not alter speech discrimination learning or the generalizability of training.