Vocal and social impairments in songbirds embryonically exposed to an autism risk factor

Summary

Autism spectrum disorders (ASD) are complex neurodevelopmental conditions marked by social communication deficits. Although rodent and chicken models have yielded insights into social and sensory abnormalities, their innate vocalizations preclude the study of vocal learning impairments common to autism. Leveraging the vocal learning abilities of zebra finches (Taeniopygia guttata), we investigated how embryonic exposure to valproic acid (VPA) affects vocal development and social behavior. VPA exposure disrupted song acquisition, resulting in impaired tutor imitation, delayed song consolidation, and increased motif variability. These abnormalities co-occurred with social deficits, including reduced modulation of female-directed song and diminished social engagement, and frequency-specific hyperacusis, a sensory phenotype frequently reported in autism. Together, these findings demonstrate that embryonic VPA exposure disrupts the development of learned vocalizations and social behavior in a vocal learner. By linking early environmental insult to persistent alterations in communication-related behaviors, this work establishes a foundation for future mechanistic and therapeutic studies.

Graphical abstract

Highlights

• •Embryonic exposure to valproic acid disrupts vocal learning in songbirds
• •VPA-treated birds show reduced tutor imitation and altered song syntax
• •Early VPA exposure induces social deficits and altered sensory responsiveness
• •Songbirds offer a tractable model for environmentally induced communication deficits

Ethology; Neuroscience; Behavioral neuroscience

Introduction

Autism spectrum disorders are complex neurodevelopmental conditions affecting approximately 1% of the global population,¹ primarily characterized by deficits in social interaction and communication,² and frequently associated with language and speech impairments.^3,4 Although numerous genetic factors have been implicated in autism, its multigenic nature and the high proportion of idiopathic cases, estimated at nearly 85%,⁵ highlight the importance of additional, non-genetic contributors. Among these, prenatal exposure to certain drugs (e.g., thalidomide, misoprostol, and valproic acid), environmental toxins (e.g., chlorpyrifos), and infections (e.g., embryonic rubella) has been linked to increased risk of autism (reviewed in the study by Landrigan et al.⁶). Notably, prenatal exposure to valproic acid (VPA, 2-propylpentanoic acid), an antiepileptic medication,⁷ is one of the most well-established environmental risk factors for autism, significantly elevating the risk^8,9,10,11 and frequently resulting in persistent cognitive and language impairments, including speech delays in approximately 30% of affected children.^12,13,14 Despite these clear clinical associations, the emergence of these communication and language deficits remains poorly understood.

Addressing this gap requires animal models capable of capturing the core behavioral phenotypes of autism, particularly those involving the development of learned vocalizations. Yet, despite considerable advances, current animal models often fall short in this regard. While traditional rodent models, chicken and non-human primates have proven useful for investigating social deficits, repetitive behaviors, and sensory abnormalities,^{15,16,17,18,19,20,21,22,23,24} they are fundamentally limited by vocalizations that are largely innate and do not rely on cortical circuitry for production,^25,26,27 rendering these models unsuitable for recapitulating the complexities of speech delays and language impairments. As a result, the neurobiological basis of vocal development delays, one of the most pervasive and debilitating features of autism, remains poorly explored.

Songbirds, particularly the zebra finch (Taeniopygia guttata), provide an ethologically relevant system for examining the development of learned vocal behavior²⁸ and has begun to be leveraged to study the effects of autism risk factors on vocal development.^29,30 As members of one of the few vocal-learning taxa, zebra finches acquire their songs through social learning,³¹ and their song development shows striking parallels to human speech. These include a critical period for learning,^32,33 babbling-like subsong,³⁴ reliance on auditory feedback,³⁵ and strong social influences on vocal acquisition.^31,36,37 Successful song learning critically depends on motivation to engage with social partners, as juveniles require contingent tutor interactions,^31,38,39 learn poorly from passive acoustic exposure,³⁶ and experimental manipulations of neuromodulators such as dopamine, noradrenaline, and oxytocin alter both social engagement and vocal outcomes.^{40,41,42,43,44} Analogously, human infants’ babbling becomes more complex when caregivers provide contingent social feedback, compared with non-contingent interaction,^45,46 a process thought to be impaired in autism.^47,48 Complementing this behavioral and developmental framework, disruption of genes associated with human speech and autism (e.g., FOXP1 and FOXP2) also impairs song learning in this species,^49,50,51 supporting its relevance for studying autism-associated communication phenotypes.

Here, we examined the consequences of embryonic exposure to the environmental autism risk factor valproic acid (VPA) in zebra finches. Our results reveal that VPA exposure induces both delayed vocal development and altered social behaviors, mirroring key aspects of autism. These phenotypes include impaired tutor song imitation, delayed song consolidation, and disorganized motif syntax, along with inability to modulate vocalizations according to social context and diminished social engagement. By linking embryonic environmental insult to persistent alterations in vocal and social behavior in a vocal learning species, this work establishes a foundation for future investigations into the developmental processes and neural circuits underlying communication impairments associated with autism.

Results

Embryonic exposure to an autism-risk factor alters song development

To evaluate the effect of environmental insults on zebra finch vocal development, we performed injections of VPA directly into the bloodstream of approximately 3.5-day-old embryos (Figure 1A). At this stage the neural tube is closing in zebra finches,⁵² and regionalization of the forebrain, midbrain, and hindbrain is underway (as described in chicken embryos⁵³), and the onset of telencephalic neurogenesis is about to start.⁵⁴ Thus, this developmental window shares key similarities with embryonic day 10 of mice and day 12 of rats (see Table S1 for a detailed comparison) which has been established as the ideal time to induce autistic-like behavior in rodent models.^55,56 Remarkably, VPA exposure did not significantly change the hatching rates, or cause any significant increase in the incidence of malformations compared to control embryos at the time of hatching (Table S2). Tutor song similarity, a measure of vocal learning proficiency, was determined at 90 days post-hatching (dph), a developmental stage when songs typically stabilize. Importantly, embryonic exposure to VPA led to significantly impaired tutor song imitation compared to controls, which already produced accurate copies at this stage (Figures 1B–1E; see also Figures S1A and S1B for additional examples; one-way ANOVA, F(2,20) = 5.721, p = 0.011). This impairment persisted in within-tutor comparisons, where VPA-treated birds consistently produced poorer copies than control birds exposed to the same tutor (Figure S1C). Moreover, this effect was present at both VPA doses tested (p = 0.007 and p = 0.0388, Dunnett’s test, for 0.8 and 8 μg of VPA; hereafter referred to as VPA_low and VPA_high, respectively).

Figure 1.

Embryonic exposure to VPA leads to poor imitation and more variable songs in young adults

(A) Timeline illustrating the experimental outline; (B–D) Representative spectrogram from a single bird belonging to vehicle (B), VPA_low (C), and VPA_high (D) groups; (E and F) Similarity to tutor song (E) and self-similarity (F) in 90-day-old birds; (G) cumulative frequency of self-similarity indexes of 90-day-old birds; (H) song-stability from one day to the next is altered in VPA treated birds at this age. Inset in G shows the Kolmogorov-Smirnov distance comparing the self-similarity scores distributions of control versus VPA_low (orange column) and control versus VPA_high (purple column); E: embryonic day; D: day post-hatch; Freq.: frequency. In E, F, and H solid horizontal lines indicate the median, and dashed lines indicate the 25th and 75th percentiles. Individual data points are shown. In G, data are represented as means. ∗p < 0.05; ∗∗p < 0.01; ∗∗∗p < 0.001. n = 7 per group.

To evaluate the progression toward mature vocal patterns, we analyzed the self-similarity rate of song motifs at 90 dph relative to fully matured adult song motifs (130 dph). Importantly, VPA-exposure also affected self-similarity rates (Figure 1F; one-way ANOVA: F (2, 22) = 9.73, p = 0.0009). Specifically, VPA_low (55.6 ± 6.3%) and VPA_high (64.0 ± 6.4%) groups showed significantly lower average self-similarity compared to vehicle controls (89.5 ± 3.0%; Mean ± SEM; Dunnett’s multiple comparisons: p = 0.0007 for VPA_low vs. vehicle, and p = 0.007 for VPA_high vs. vehicle). The distribution of self-similarity scores further supported this, showing most songs from VPA-exposed animals falling below 70%, whereas most iterations from control birds scored above 80% (Figure 1G; Kolmogorov-Smirnov distance: 1.00 for vehicle vs. VPA_low, and 0.92 for vehicle vs. VPA_high; both p < 0.0001). This indicates that songs from VPA are still changing and have yet to reach their final state, unlike vehicle control birds whose songs were already very similar to their adult form by 90 dph. To further characterize the dynamic nature of song development in VPA-exposed birds, we also analyzed their day-to-day song variability (d90 vs. d91). VPA-treated birds produced highly variable songs at this stage (Figure 1H; one way ANOVA: F(2,22) = 8.86, p = 0.0015), particularly the VPA_high group (Dunnett’s multiple comparisons: p = 0.0008 for VPA_high vs. vehicle; VPA_low vs. vehicle, p = 0.15). In contrast, control birds produced largely stable songs from one day to the other. Of note, these measures were not affected by the egg opening and injection procedure itself, as vehicle-injected controls behaved identically to unmanipulated controls (Figure S2). Thus, VPA-exposed birds exhibit poorer copies of the tutor song which are also more immature and still vary from iteration to iteration.

Such variability in song motifs can arise from differences in syllable structure or from the order in which syllables are sung (hereafter referred to as syntax, used here in a non-linguistic sense), or from a combination of both. To explore these sources, individual songs were mapped as trajectories in a feature plane (e.g.,⁵⁷; see STAR methods for details). While individual songs from birds in the control group were highly reproducible, consistently following the same trajectories (Figures 2A–2C and S3A, bottom), VPA-exposed birds, regardless of the dose, sang motifs that varied from iteration to iteration (Figures 2D and G). This variability is immediately evident in the feature plane, as different iterations in VPA-treated birds follow different trajectories which can be readily observed in the form of partially overlapping loops (Figures 2E and 2H red arrowheads), resulting in a visible dispersion of trajectories when multiple iterations are overlapped in the maps (Figures 2F–2I, S3B, and S3C black arrowheads).

Figure 2.

VPA treatment increases variability in song motif temporal sequence

(A) Spectrograms of representative song motifs from a vehicle-treated bird, illustrating consistent song production.

(B) Top left: UMAP projection of latent features from all 120-ms windows from the first 100 motifs sung on day 90 by the same bird. Darkness represents density, such that lighter dots correspond to less-explored regions. S1-3: trajectories in the latent space followed by the songs shown in A. Color marks the passage of time, matching the colorbar below G.

(C) Superimposition of the trajectories followed by all 100 motifs from this bird.

(D–F) Similar representations for a VPA_low-treated bird. Red arrowheads in E highlight a trajectory passing twice over the same region, reflecting repetition within the motif. Black arrowheads in F highlight divergence across trajectories, reflecting motif variability.

(G–I) Similar representations for a VPA_high-treated bird.

(J–L) Markov chains illustrate the transition probabilities of motif sequences for a representative bird from the vehicle (J), VPA_low (K), and VPA_high (L) groups; arrow thickness indicates the relative frequency of each transition, with values labeled above each arrow.

(M) Shannon entropy of motif transition probability matrices, demonstrating the overall variability in motif syntax across groups. ∗∗p < 0.01. Solid horizontal lines indicate the median, and dashed lines indicate the 25th and 75th percentiles; individual data points are shown.

To investigate the source of variability contributing to the previously described latent features, we first analyzed the motif syntax by calculating the transition probability of one syllable to another within each rendition. Motif syntax in VPA-treated birds readily appeared less regular than those of control birds (Figures 2J–2L; see top panels of Figure S3 for data from all birds). Quantifying syntax regularity via the entropy rate of the estimated transition matrices revealed that VPA animals exhibited a significantly higher entropy rate (Figure 2M; Kruskal-Wallis; p = 0.0007; Dunnett’s multiple comparisons test p = 0.005 for both VPA_low and VPA_high), demonstrating that VPA exposure disrupts motif syntax. In contrast, control birds, as expected for crystallized songs, had an entropy rate close to zero (Figure 2M), confirming their high stereotypy across renditions. To investigate whether VPA-treated birds also exhibit variations in the song spectral composition, independent of syntax, we computed a spectral variability index⁵⁸ across motifs with identical syllable content (example in Figure 3A). The variability index did not differ between VPA-exposed and control subjects (Figure 3B; Kruskal-Wallis; p = 0.56). We next examined intersyllable gap durations to determine whether temporal aspects of song production were affected. Consistent with the spectral analyses, intersyllable gaps did not show consistent or systematic changes across vehicle- and VPA-treated and birds (Figure S4), indicating preserved timing at the level of syllable-to-syllable execution. To further assess the capacity for stable vocal production, we next examined long calls, a partially innate contact call produced in isolation that also depends on intact syrinx function and vocal motor control.⁵⁹ Spectral variability was similarly quantified across multiple renditions of these calls. Consistent with the song analysis, the spectral variability of a long call (see Figure 3C for an example) did not differ between VPA-exposed and control groups (Figure 3D; one-way ANOVA, F(2,20) = 2.42, p = 0.11). Unlike song motifs, call stability was comparable between groups at this age (Figure 3E; Kruskal-Wallis test, p = 0.99). Together, these results show that while motif structure becomes more variable, the motor control of individual syllable production appears to be largely unaffected by embryonic VPA exposure. Thus, in addition to being poorer copies of tutor songs, songs of VPA-treated birds also exhibit variability in the order of syllables at a much higher rate than those of control birds, even though the individual syllables themselves are sung consistently.

Figure 3.

VPA treatment does not affect acoustic variability of song motifs and calls

(A) Representative example illustrating comparable acoustic variability across birds, independent of treatment, for motif renditions containing identical syllable content. The upper image shows three song renditions, with red bars marking periods of high variability. The lower image displays the mean absolute deviation (MAD) derived from 16 latent acoustic features using a shotgun variational autoencoder (VAE); the bold trace represents the summed MAD across features for each analyzed sonogram.

(B) Quantification of overall motif variability reveals no significant effect of VPA treatment.

(C) Representative example showing comparable acoustic variability in a long call (from the same VPA-treated bird shown in (A). The lower image displays the MAD derived from 16 latent acoustic features using a shotgun VAE; the bold trace represents the summed MAD across features for each analyzed sonogram.

(D) Quantification of long-call variability likewise shows no significant difference between groups.

(E) Day-to-day call stability at 90 days post-hatch is not altered by VPA treatment. In B, D, and E, Solid horizontal lines indicate the median, and dashed lines indicate the 25th and 75th percentiles; individual data points are shown. n = 7–8 birds per group.

Embryonic VPA exposure delays song maturation and crystallization in zebra finches

The observed variability in syntax and low self-similarity rates indicates delayed vocal development and maturation. To determine the precise extent of this delay, we evaluated the dynamics of vocal progression from day 50–130. While tutor similarity scores progressively increased in all groups as animals developed, indicative of the imitation process associated with song learning, VPA exposure significantly affected song development overall (Figure 4A, mixed-effects model, REML; treatment: F(2.30, 44.02) = 22.69, p = 0.015; age: F(2,21) = 5.14, p < 0.0001). This effect was robust across both VPA doses (Dunnett’s multiple comparison test: p = 0.0003 for VPA_low, and p = 0.0044 for VPA_high). Fitting the dynamics of tutor-song acquisition with asymptotic exponential models revealed a significant effect of treatment on the curve time constants (nonlinear mixed-effects model; treatment: F (2,173) = 3.58, p = 0.03). Estimates showed prolonged acquisition times for VPA_low (τ = 28.4 days) and VPA_high (τ = 18.7 days) compared to vehicle controls (τ = 13.4 days; Figure 4A, dashed lines). Specifically, VPA_low showed a significant delay compared to the vehicle (Dunnett’s multiple comparison test: p = 0.014 for VPA_low) while VPA_high was not significantly different (p = 0.31).

Figure 4.

VPA exposure leads to delayed song development

(A) Tutor similarity and (B) self-similarity during song development (each day compared to day 130). Data represent means ± SEM. Vehicle controls are shown as gray circles, VPA_low as orange triangles, and VPA_high as purple triangles. Dotted lines in (A) indicate time constants (τ), while dotted lines in (B) indicate t50 values.

Given that song immaturity is characterized by variability, self-similarity rates are particularly telling: VPA exposure significantly affected song maturation, leading to consistently lower self-similarity scores in treated birds throughout development (Figure 4B; mixed-effects model, REML; treatment: F(3.84, 73.60) = 33.92, p = 0.001; age: F(2,20) = 9.732, p < 0.0001). Furthermore, a statistically significant age-by-treatment interaction (p = 0.0016) definitively shows that VPA alters the entire developmental trajectory of song progression. Quantifying the impact of VPA on song maturation dynamics by fitting sigmoidal learning curves revealed a strong significant effect of treatment on t₅₀, the dph marking the curve midpoints (nonlinear mixed model; treatment: F(2,170) = 102.9, p < 0.0001). Both VPA_low (t₅₀ = 106.4 days) and VPA_high (t₅₀ = 91.1 days) groups showed significantly delayed learning compared to vehicle controls (t₅₀ = 64.6 days; Figure 4B, dashed lines; p < 0.0001 for both; Dunnett’s multiple comparison test). No significant effect of treatment was found for the slope of the sigmoidal learning curves (F(2,17) = 1.1, p = 0.33). Collectively, these findings demonstrate that embryonic exposure to VPA delays song maturation and crystallization in zebra finches by several weeks, as evidenced by impaired tutor imitation, increased syntactic variability, and a significantly prolonged consolidation phase.

VPA impacts social behavior in zebra finches

Autism is behaviorally defined by altered social interactions. To assess social responsiveness in VPA-treated zebra finches, we examined context-dependent song features, a hallmark of male courtship behavior. More specifically, in the presence of females males produce a “directed song” with distinct features,⁶⁰ accompanied by courtship displays such as synchronized dancing.⁶¹ Consistent with these observations, the majority of vehicle-treated birds engaged in courtship and attempted copulation when presented with a female (Figures 5A–5C), whereas a smaller proportion of VPA-treated birds did so. Control birds also showed robust social modulation, significantly increasing the number of introductory notes in directed compared with undirected songs (Figure 5D; paired t test: p = 0.019) and exhibited more stereotyped fundamental frequency patterns during directed singing (Figure 5E; paired t test p = 0.022). By contrast, VPA-treated birds failed to consistently show this modulation. Despite a subset of VPA-treated animals engaging in overt courtship behaviors, such as dances, beak wipes, and copulation attempts at levels comparable to controls (Figures S6A and S6B; one-way ANOVA: courtships F(2,18) = 0.85, p = 0.46; beak wipes F(2,19) = 0.54, p = 0.59; copulation attempts F(2,19) = 0.46, p = 0.63), these birds did not consistently alter their song features based on social context. Specifically, not all VPA-treated birds showed an increase in introductory syllables per bout (Figure 5D; VPA_low p = 0.25, VPA_high p = 0.89) and we did not observe a consistent reduction in fundamental frequency variability during directed song (Figure 5E; VPA_low p = 0.35, VPA_high p = 0.55). Because group-level averages were influenced by substantial inter-individual variability, we additionally performed categorical analyses of reproductive engagement. Only 13% of VPA-treated birds attempted copulation with novel females, compared to 58% of control birds, a difference that approached statistical significance (Fisher’s exact test, p = 0.053). Consistent with this, the likelihood of a VPA-treated bird completing a full mating sequence was reduced relative to controls (odds ratio = 0.24). These findings suggest that while a subset of VPA-treated birds recognize and engage in socially relevant contexts, they fail to appropriately modulate their vocal output, mirroring key aspects of ASD-related communication deficits.

Figure 5.

VPA exposure alters social behavior in zebra finches

(A) Schematic illustrating the behavioral setup for female exposure.

(B and C) Proportion of birds engaging in courtship (B) and attempting copulation (C) across treatment groups.

(D and E) VPA-treated birds fail to modulate their songs in a social context. Connected dots represent paired measurements from the same individual, and bars indicate the group mean. (D) Shows the number of introductory notes during directed (D) versus undirected (U) singing. (E) Illustrates that song variability is reduced during directed singing in control birds but not in VPA-treated birds.

(F) Outline of the behavioral apparatus, highlighting the neutral and social perches.

(G–I) Heat maps displaying individual activity patterns for vehicle (G), VPA_low (H), and VPA_high (I) groups when a conspecific is present in the adjacent, visually accessible cage. The blue and red bars indicate the social and neutral perches, respectively. The color scale in f indicates relative occupancy during the social trial presented in G–I.

(J and K) VPA-treated birds spend less time on the social perch but not on the neutral perch, suggesting reduced social motivation.

(L) Reduced difference between time spent on the social and neutral perches in VPA-treated birds relative to controls.

(M and O) Orientation of birds during social (M) and neutral (O) perch visits. Data are plotted as time spent in each orientation relative to the window (0°).

(N and P) Time spent oriented toward the window on social (N) and neutral (P) perches, providing insight into direct engagement.

(Q) Two-dimensional UMAP projection of combined social engagement measures indicates two modes of behavior, social and anti-social. Included measures were time spent on social and neutral perches, latency, and time oriented toward the social perch). In J, K, L, N, and P, solid horizontal lines indicate the median, and dashed lines indicate the 25th and 75th percentiles; individual data points are shown. ∗p < 0.05; ∗∗p < 0.01. Scale bar in g: 5 cm. n = 7 per group, except for VPA_high group on directed singing behavior, in which only 3 birds were tested.

To obtain an independent measure of voluntary social engagement, we placed each bird in a chamber where it could choose between approaching an standardized unfamiliar male positioned across a “social perch” or an empty cage across a “neutral perch” (Figure 5F). Heat maps show that control birds concentrated their activity on the social perch, whereas VPA-treated birds spent less time in this area (Figures 5G–5I, and S5). This pattern was confirmed by quantitative analysis: VPA-exposed birds spent significantly less time in the social perch compared to controls (Figure 5J; one-way ANOVA: F(2,18) = 6.12, p = 0.006; Dunnett’s multiple comparison test: VPA_low, p = 0.0057; VPA_high, p = 0.037), whereas no significant differences were observed in time spent in the neutral perch (Figure 5K; one-way ANOVA: F(2,18) = 0.51, p = 0.62). To formally test whether perch preference differed across treatment groups, we performed a two-way ANOVA with treatment and perch type as factors. This analysis revealed a significant treatment × perch interaction for both time spent on each perch (F(2,18) = 6.693, p = 0.0067) indicating that the distribution of perch-directed behavior differed across treatment groups. Consistent with this pattern, control birds showed a significantly higher social preference index (social—neutral perch time) than VPA-treated birds (Figure 5L; one-way ANOVA, F(2,18) = 6.69, p = 0.0067). VPA-treated birds also tended to take longer to first visit the social perch (Figure S6C; one-way ANOVA, F(2,18) = 4.63, p = 0.023; Dunnett’s tests: VPAlow vs. control, p = 0.014; VPAhigh vs. control, p = 0.12) and to remain there for shorter durations (Figure S6G; Brown-Forsythe ANOVA, F(2,18) = 3.53, p = 0.058). In contrast, neither the latency to visit nor the duration of visits to the neutral perch differed between groups (Figure S6D and S6H). To formally evaluate the effects of treatment and perch type, a mixed-effects model (REML) was fitted with perch type as a within-subject factor. This analysis revealed significant main effects of treatment (p = 0.04) and perch type (p = 0.0388), as well as a trend toward a treatment × perch interaction (p = 0.0554), indicating a selective alteration of social, but not neutral, perch engagement in VPA-treated birds. Notably, despite these differences in social approach behavior, neither the total number of perch visits (collapsed across perch type; Figures S6E and S6F) nor the total distance traveled (Figure S6I) differed significantly between groups.

Although visits to the social perch provide an index of social approach, they do not directly indicate whether birds are visually attending to the conspecific. To assess active visual engagement, we quantified head orientation toward the window (Figures 5M and 5O). VPA-exposed birds spent significantly less time oriented toward the window when positioned on the social perch compared to controls (Figure 5N; one-way ANOVA: F(2,18) = 8.487, p = 0.002; Dunnett’s post hoc test: VPA_low p = 0.001, VPA_high p = 0.027), whereas no group differences were observed for head orientation on the neutral perch (Figure 5P; Kruskal-Wallis test, p = 0.45). Consistent with the perch occupancy results, a two-way ANOVA with treatment and perch type as factors revealed a significant treatment × perch interaction for time oriented toward each window (F(2,18) = 3.845, p = 0.040).

To further explore individual variability in social behaviors, a two dimensional UMAP projection was generated using time spent on social and neutral perches, time oriented toward the social window, and social perch latency (Figure 5Q). This projection revealed two distinct behavioral clusters: an antisocial mode, predominantly populated by VPA-treated birds, displaying consistently lower social engagement scores, and a social mode, largely composed of vehicle-treated control birds and characterized by high visual engagement (PERMANOVA: pseudoF (21,3) = 4.171411, p = 0.009; see Table S3 for values). Moreover, these differences were not present when birds were alone in the pretrial sessions (Figures S6J and K), as evidenced on their UMAP projection, where all individual birds are intermixed (Figure S6l; PERMANOVA: pseudoF (21,3) = 1.888645, p = 0.10). Together, these findings indicate that VPA-treated birds exhibit reduced voluntary social engagement, though with some individual variability in the extent of impairment. This pattern further strengthens the face validity of our model for autism-related social deficits.

VPA impacts auditory responses

To assess whether VPA exposure impacted basic hearing abilities and contributed to lower song similarity rates in birds, a hearing threshold test was conducted. Both VPA_low and VPA_high groups exhibited a flatter, more sensitive hearing curve (Figure 6A, orange and purple triangles). They responded to much lower sound amplitudes than the controls, especially at the lowest (0.5 and 1 kHz) and highest (8 kHz) frequencies, indicating hypersensitivity. In contrast the vehicle group exhibited a typical u-shaped auditory threshold curve (Figure 6A, gray circles), showing highest sensitivity (20–30 dB amplitude) at intermediate frequencies (2 and 4 kHz) and lower sensitivity at 0.5, 1, and 8 kHz (requiring 50–70 dB). However, despite this hypersensitivity in VPA-exposed birds, correlation tests indicated that their enhanced auditory sensitivity did not appear to influence song development performance (Figures 6B–6D). These findings suggest that altered hearing sensitivity alone is unlikely to explain the vocal deficits observed.

Figure 6.

VPA treatment induces auditory hypersensitivity without correlating with vocal learning deficits

(A) Hearing threshold curves for vehicle (gray circles), VPA_low (orange triangles), and VPA_high (purple triangles) treated birds. A two-way ANOVA revealed significant effects of frequency, treatment, and their interaction (p < 0.001 for all), indicating that VPA treatment broadly alters auditory sensitivity. Data represent means ± SEM.

(B–D) The images illustrating that auditory hypersensitivity at specific frequencies does not correlate with vocal learning performance. Scatterplots show the relationship between similarity rates and hearing thresholds at low frequencies (0.5 kHz in B, 1 kHz in C) and a high frequency (8 kHz in D). The extremely low R² values (0.001 for 8 kHz, 0.14 for 1 kHz, and 0.0006 for 0.5 kHz) indicate a negligible linear correlation between auditory sensitivity at these tested frequencies and the observed vocal learning impairments; n = 7 per group.

Discussion

This study provides comprehensive evidence that embryonic exposure to an environmental risk factor for autism, valproic acid (VPA), impairs vocal learning and social communication in zebra finches. Our findings reveal a constellation of autism-like behavioral phenotypes that mirror key diagnostic criteria in humans, thus bridging a critical gap in understanding communication deficits in autism that traditional animal models have struggled to fully recapitulate.

VPA-treated zebra finches produced significantly less accurate tutor song imitations than controls, with deficiencies persisting from juvenile development into early adulthood. Song maturation was also delayed, as treated birds failed to develop the stereotyped motif structure typical of controls by the expected age. Acoustic analyses revealed increased variability in syllable sequencing and greater entropy in motif structure. While our results align with altered vocal behavior reported by Tewelde et al.³⁰ and Pogany et al.²⁹ following late-embryonic VPA exposure, we observed distinct social phenotypes and pronounced learning deficits. These differences likely reflect methodological differences between the studies. Whereas the previous work evaluated vocal structure using the initial ten syllables of a session, a metric that may not fully represent a stabilized song motif, our analysis captured cross-rendition variability, providing a more comprehensive view of motif entropy. In addition, by controlling for peer-vocal exposure during rearing, we isolated the specific effects of VPA on tutor-based learning. Consequently, our findings reveal a more protracted and diminished learning trajectory, characterized by significant impairments in vocal consolidation. This pattern mirrors key speech and language impairments observed in autism, including delayed vocal development, poor imitation, and difficulty stabilizing speech output.^62,63,64 Notably, the convergence between VPA-induced deficits and those seen following targeted manipulation of the autism-associated genes FOXP1 within HVC and FOXP2 within Area X^49,50,51 suggests a potential commonality in the underlying neural pathways. This enticing similarity can serve to orient future research aimed at uncovering the neural mechanisms behind the behavioral alterations induced by VPA exposure in our birds.

Beyond vocal impairments, VPA-exposed birds showed reduced voluntary social engagement with a male conspecific, and failed to modulate their song appropriately in the presence of females. The observed deficits in context-appropriate singing parallels the human lack of register variation often observed in autism^4,65 and the convergence of vocal and social impairments mirrors the interdependent nature of communication challenges in autistic individuals. Given that vocal learning in zebra finches and humans is inherently social,^33,45 our findings align with the social motivation hypothesis of autism,^{47,48,66,67,68} which posits that reduced social engagement limits access to effective learning opportunities. These results therefore raise the possibility that social and vocal deficits interact in both species. In zebra finches, successful song acquisition depends on contingent tutor interactions rather than passive acoustic exposure alone,^31,36,38,39 suggesting that reduced social engagement in VPA-treated birds could plausibly constrain opportunities for effective learning. Our study extends prior work in songbirds by combining detailed analyses of social behavior with measures of vocal production and learning. Earlier studies have provided important insights by focusing on gene expression correlates⁶⁹ or on targeted manipulations of single candidate genes.^49,50,51 However, because these manipulations were typically initiated during juvenile stages (e.g., 35 days post-hatch), they were not designed to directly assess how early developmental perturbations influence the emergence of autism-relevant phenotypes or the maturation of learned vocal behavior. By integrating social, vocal, and learning measures within the same animals, our work aims to help bridge this gap and to provide a framework for linking cellular and molecular mechanisms to complex, autism-relevant behavioral outcomes. Further studies in species capable of robust vocal learning via passive playback (e.g., canaries) could help disentangle direct effects on vocal learning from secondary effects mediated by social interaction/motivation deficits.

VPA-treated birds exhibited heightened responsiveness to specific frequencies (hyperacusis), a sensory phenotype frequently reported in individuals with autism,^70,71,72 and a core diagnostic criterion.^2,73 In humans, auditory hypersensitivity has been correlated with language impairments and atypical temporal processing of complex sounds, including speech,⁷⁴ with other evidence suggesting enhanced perception of simple stimuli but difficulties with socially meaningful auditory input.⁷⁵ Nonetheless, we found no correlation between the degree of hyperacusis and the severity of vocal imitation deficits or song maturation delays, suggesting that sensory abnormalities alone do not explain the observed communication impairments.

VPA-induced impairments did not follow a linear dose-response curve, but instead showed a non-monotonic pattern, with some of the most severe behavioral effects observed at the low dose. This could reflect VPA’s known pleiotropic actions, including histone deacetylase inhibition and mitochondrial modulation, which exhibit distinct dose sensitivities.^76,77 In humans, VPA is typically administered at 10–60 mg/kg/day, yielding serum concentrations of approximately 50–100 μg/mL⁷⁸ whereas embryonic exposure is considerably lower because of limited placental transfer, plasma protein binding, and tissue distribution.⁷⁹ Within this framework, the embryonic doses used here (0.8–8 μg/g) fall within the range of physiologically relevant fetal exposures spanning low to high therapeutic conditions, particularly given that, at the time of injection, the embryo plus extra-embryonic tissues constitute only a small fraction of total egg mass. We also observed substantial inter-individual variability within each dose group, paralleling the phenotypic heterogeneity reported in autism, where large cohort studies show that VPA exposure increases risk without producing a uniform outcome.^8,9 Comparable variability has been described in rodent VPA models,⁸⁰ in which differences in genetic background and metabolic status have been proposed as contributing factors.^81,82 Although our entire protocol was designed to minimize variation, we cannot fully exclude the possibility that, at higher doses, only the least affected embryos survived to hatching, effectively truncating the upper range of impairment. At the same time, the absence of detectable group differences in survival may have been obscured by other sources of variability, such as experimenter proficiency in the injection procedure. Such factors may similarly underlie some of the variability we observed in VPA-treated zebra finches. Interestingly, low doses of VPA have been reported to produce stronger and more consistent effects than higher doses in brain organoids,⁸³ zebrafish social behavior,⁸⁴ and selected behavioral assays in mice,⁸⁵ a pattern that parallels our observations in zebra finches. Together, these findings suggest that behavioral and morphological outcomes may be mediated by partially distinct, dose-dependent mechanisms.

The timing of our embryonic manipulation (zebra finch E3-4, stages 17–21⁵²) was selected to target late neurulation and early telencephalic vesicle formation, when broad pallial territories are being established but shortly before the onset of robust telencephalic neurogenesis.⁵⁴ In chick, telencephalic ventricular injections at E3 efficiently target proliferative neuroepithelial progenitors, and can alter pallial patterning and neurogenesis as early as E4.^86,87 In mice, early telencephalic patterning at neurulation and early vesicle stages is controlled by conserved morphogen and transcription factor networks in the cranial neural plate and nascent forebrain, and disruption at this stage alters neural tube closure, progenitor pool size, and regional identity.^88,89,90,91 Accordingly, perturbations in zebra finch at E3-4 are predicted to affect early telencephalic progenitors and the specification of avian pallial circuits that are functionally analogous to mammalian cortex, including song control nuclei. Moreover, comparative mouse-chick transcriptomic analyses show that regulators implicated in human neurodevelopmental disorders are highly expressed during equivalent early telencephalic stages (HH17–21 in chick and E9.5–E10.5 in mouse),⁹² underscoring the sensitivity of this developmental window. While future studies will be required to define the precise critical window for vocal learning vulnerability, our results indicate that disruption during early embryogenesis is sufficient to produce lasting alterations in vocal development.

This study provides a rigorous phenotypic characterization and replicates key autism-linked behaviors in an animal model, particularly the impaired development of learned vocalizations, a domain that has been largely inaccessible to autism animal studies. Embryonic VPA exposure could, in principle, affect vocal and social behavior through alterations in the neural circuits governing motivation and execution, or through disruptions of motor output caused by peripheral malformations or procedural stress. Multiple lines of evidence argue against the latter interpretation. First, VPA treatment did not produce global motor impairments, but instead selectively affected behaviors expressed in social contexts. Second, treated birds showed no deficits in the stable production of individual song syllables or isolated calls, as would be expected if peripheral vocal structures or motor execution were compromised. Finally, vehicle-injected and unmanipulated controls behaved indistinguishably across the domains examined, making it unlikely that the egg opening, injection, or associated stress account for the robust and selective phenotypes observed in VPA-treated birds.

While our investigation is primarily behavioral, the robust replication of these autism-like phenotypes supports the idea that environmental insults can disrupt neurodevelopmental trajectories. Although we did not measure neural parameters directly, the persistence of these sensorimotor deficits across development, combined with the absence of impairments in long calls, suggests a central contribution rather than a primary deficit in the peripheral vocal apparatus, for example. These functional findings are therefore more consistent with alterations in the assembly or maturation of song-related circuits. One potential possibility would involve alterations in the GABAergic polarity shift or impaired long-term potentiation, as recently demonstrated in other avian VPA models.⁹³ Curiously, in zebra finches inhibition within premotor song circuits is known to be important for stabilizing learned vocal sequences while allowing ongoing plasticity in immature segments,⁹⁴ and inhibitory neurons in NCM, an auditory region critical for song memory formation, exhibit topographically organized responses to song.⁹⁵ Alterations in these inhibitory populations could therefore disrupt sensory representations of the song template. In addition, neuromodulatory systems known to influence social engagement and vocal learning outcomes in songbirds (e.g., dopaminergic or oxytocinergic pathways^40,42,44) represent plausible targets for future mechanistic studies. Taken together, our findings provide behavioral support for the zebra finch as a model to investigate circuit-level and molecular processes underlying vocal development. The observed impairments in song learning suggest particular relevance for examining mechanisms such as sensory-motor integration, memory consolidation, and circuit assembly.

A strength of the approach used here is the ability to examine developmentally emerging behaviors in a whole-animal context, enabling longitudinal assessment of VPA effects on complex social phenotypes such as vocal learning. By delivering VPA directly to the embryo, this strategy avoids indirect maternal influences, an advantage shared with other oviparous models such as chick, in which social behaviors have also been successfully examined.^21,22,23,24 While several animal systems allow behavioral tracking across development, the zebra finch offers the unique benefit of a well-characterized, learned vocal communication system that emerges through social experience. This feature enables direct analysis of how embryonic perturbations impact the acquisition and modulation of learned vocal behavior over time. Because autism is fundamentally a developmental condition, embryonic manipulations provide a direct means of probing its origins and progression. Notably, higher doses did not consistently produce more severe deficits or overt teratogenesis in hatched birds, and we did not observe the gross external malformations commonly reported in other VPA models (e.g., tail fold defects^96,97), which may reflect species-specific developmental trajectories and sensitivities.^98,99 However, because embryos that failed to hatch were not systematically examined, earlier lethal malformations cannot be excluded. The absence of overt teratogenic effects parallels the human condition, in which autism typically arises without obvious morphological abnormalities, and supports the utility of this model for isolating behavioral and neurobiological alterations in the absence of major structural confounds.

In conclusion, our study provides compelling evidence that embryonic VPA exposure induces a comprehensive array of autism-like vocal and social communication deficits in zebra finches. These findings provide a framework for investigating the biological bases of communication impairments in autism and further highlight the zebra finch as a promising system for future mechanistic and therapeutic studies. Because zebra finches acquire vocalizations through socially guided learning, this model may be particularly useful for future studies aimed at disentangling how early developmental perturbations influence the interaction between social behavior and communication development.

Limitations of the study

Some limitations of the present study should be acknowledged. First, although the VPA model is widely used to investigate neurodevelopmental mechanisms relevant to autism-related phenotypes, it represents a pharmacological perturbation and does not necessarily capture the full genetic and etiological heterogeneity of the human condition. Thus, caution is warranted when extrapolating our findings to clinical populations. Future studies combining VPA exposure with genetic or other environmental risk models may help determine whether similar behavioral and neural phenotypes emerge across distinct etiological contexts. Second, VPA was administered at a single developmental time window using two doses. While this approach was chosen to ensure experimental control and comparability with prior work, increasing the range of doses and exposure windows may yield distinct neural and behavioral outcomes. Systematic variation of these parameters will be important to determine the generalizability of the observed effects. Third, we note that using the same adult male bird as a standardized social stimulus for all experimental subjects may limit the generalization of these findings across the broader population of potential interaction partners. While this strategy reduced variability and ensured consistent stimulus conditions across experiments, future studies incorporating multiple stimulus individuals will be important to assess the robustness of the observed effects across social contexts.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Tarciso A.F. Velho (velhot@neuro.ufrn.br).

Materials availability

This study did not generate new unique reagents.

Data and code availability

The behavioral data used in this study have been deposited at Mendeley (https://doi.org/10.17632/xwyywz33zz.1) and are publicly available as of the date of publication. All code and custom functions used for behavioral analyses were written in MATLAB and Python, and are available online (https://github.com/VelhoLab/songbird-vpa). Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

We wish to thank Dr. Claudio Mello (OHSU) for his insightful comments on the first draft of this manuscript and Jack Goffinet for assistance in running the AVA code. This work was supported with grants from the Instituto Serrapilheira (Grant #1709 to T.A.F.V.), Brazilian National Council for Scientific and Technological Development (CNPq; #444971/2024-6 and #421955/2023-6 to T.A.F.V.), and the Coordination for the Improvement of Higher Education Personnel – Brazil (CAPES; financing code 001 to A.Y.M.S., R.B.G.M.C., J.A.B.A., A.S.C. and G.B.).

Author contributions

T.A.F.V. and A.Y.M.S. designed the research; A.Y.M.S., R.B.G.M.C., A.L.P.A., J.A.B.A., A.S.C., and G.B. performed the research; T.Z.L. contributed to statistical analyses; T.A.F.V., A.Y.M.S., R.B.G.M.C., and D.A.L. analyzed the data; T.A.F.V. wrote the first draft of the manuscript; T.A.F.V., D.A.L., and R.B.G.M.C. prepared the final figures; T.A.F.V. and D.A.L. edited the manuscript.

Declaration of interests

The authors declare no competing financial interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Chemicals, peptides, and recombinant proteins
Valproic acid	Sigma-Aldrich	P4543
Fast Green Dye	Sigma-Aldrich	2353-45-9
Deposited data
VPA and Vehicle Treated Songs	This paper	https://doi.org/10.17632/xwyywz33zz.1
VPA and Vehicle Treated Calls	This paper	https://doi.org/10.17632/xwyywz33zz.1
Behavioral Essays Videos	This paper	https://doi.org/10.17632/xwyywz33zz.1
Experimental models: Organisms/strains
Zebra finch (Taeniopygia guttata castanotis)	Velho Lab (UFRN)	N/A
Software and algorithms
Code for Vídeo and Behavioral Analysis	This paper	https://github.com/VelhoLab/songbird-vpa/tree/main/spatial-analysis
Code for Vocal Analysis	This paper	https://github.com/VelhoLab/songbird-vpa/tree/main/vocal-analysis
Code for vocalization dimensionality reduction	Goffinet et al.58 2021

Experimental model and study participant details

All experimental procedures were approved by the UFRN - Animal Use Ethics Committee (074.082/2017). All animals used were raised in our own colony. Animals were housed individually or as a family in wire cages (58 × 28 × 33 cm) under a 14:10 h light/dark cycle. Food and water were provided ad libitum using a standard seed mixture. Diet was supplemented every other day with boiled eggs, cuttlebone, cucumber slices and collard greens. The room temperature was maintained at approximately 25°C. For egg production, pre-established breeding pairs were provided with nests and nesting material (sterilized sisal fiber). Eggs were collected daily, replaced with dummy eggs, and transferred immediately to an artificial incubator (Grumbach Compact S84 MP GTFS, Germany). Due to low hatching rates following the egg-opening procedure, in addition to individual breeding pairs we used a larger pool of donor birds housed in flight cages to provide fertilized eggs and ensure sufficient hatchling numbers. Under these conditions, the precise biological origin of each egg could not be tracked. However, because all injection sets used eggs from the same donor population, the resulting hatchlings reflected the overall genetic variability of the colony, rather than variability due to distinct or systematically varied genetic backgrounds. Hatchlings were transferred to foster parents immediately after hatching (along with the eggshell) and were subsequently raised by both foster parents in a single family breeding cage. Because sex could not be determined prior to embryonic manipulations, embryos of both sexes were initially included; however, only males (n = 32; 8 per group) were used for behavioral analyses. An additional cohort of normally reared birds (≥90 days old) from the colony was used as social stimuli: 16 randomly selected females for directed-singing trials and one male for the social-preference test.

Method details

Embryonic manipulation

Egg windowing and artificial incubation procedures followed previously described protocols.¹⁰⁰ Briefly, for embryo injections, a small opening (approximately 1 mm in diameter) was created in the eggshell using a 26-gauge needle to access day 3 embryo’s blood vessels (stages 12–14, approximately). We then used a pulled and beveled glass micropipette (Drummond #5-000-1001-X; pulled with a Narishige PC-10 and beveled at approximately 45° with a Narishige EG-05, yielding a tip opening of 5–10 μm) filled with either valproic acid (VPA; P4543, Sigma-Aldrich) or vehicle solution colored with 0.025% Fast Green dye (2353-45-9, Sigma-Aldrich) to make it possible to visualize the dispersion of the solution into the bloodstream. Embryos receive either vehicle (PBS), 20 mg/mL, or 200 mg/mL of VPA diluted in PBS. We selected two distinct doses based on Whitsel et al.,¹⁰¹ who administered 20 μL of valproic acid (VPA) solutions ranging from 5 to 20 mg/mL (corresponding to a total of 100–400 μg per egg). For simplicity, these doses will be referred to as VPA_low and VPA_high throughout the paper. With a one-axis water hydraulic micromanipulator (MO-10 Narishige), a 40-nanoL solution (vehicle or VPA) was driven into the bloodstream of the embryo which represents 0.8 and 8 μg total for VPA_low and VPA_high, respectively. To maintain consistency, injections were made in the lower dorsal vessels following the flow that traveled directly to the embryo whenever possible, and only embryos with visually confirmed injections were considered in further experiments. After the injection, the egg was closed with silicon (Kwik-cast - Silicone Elastomer) and transferred back to the artificial incubator to continue developing.

Rearing conditions

Once hatched, hatchlings were weighed, visually inspected for possible gross morphological defects, and transferred immediately into a foster's nest. The foster parents’ cage was visually isolated from other animals in the colony to restrict the access of hatchlings to a single tutor model. Whenever possible, each foster pair received two hatchlings. Because eggs were collected daily and transferred immediately to the incubator, true siblings typically hatched at least one day apart and were therefore not assigned to the same foster parents. We did not employ a yoked design in which control and VPA-treated chicks are fostered together. This decision was made because the effects of VPA on early behavior and competitive ability were unknown; mixing treatments within the same brood could have created unequal competition among nestmates¹⁰² and potentially amplified group differences. Therefore, to standardize the early environment and eliminate competitive biases related to co-housing treatments, fostering was generally limited to a single male hatchling per cage. Following this protocol, only a few exceptions were made: one sibling pair was housed per VPA-treated group and two sibling pairs were housed in the vehicle-control group. Juveniles remained with their foster parents until dph 65, after which they were moved, together with their foster-siblings, to individual cages to continue developing until adulthood.

Vocal recordings

All animals were routinely recorded (every 10 days) from dph 50 to 130. Experimental birds were placed in recording cages placed into a soundproof box (67 × 53 × 53 cm) with water and food ad libitum under the same 14/10 h light/dark cycle experienced in the aviary. The box was connected to a sound recording system consisting of a microphone (AT803, Audio Technica) and an audio interface (M-Track Eight from M-Audio, or US-4x4 from TASCAM). Sound recordings were performed using SAP2011 software (http://soundanalysispro.com/). Each recording session lasted for approximately 3 days. After this, the animal returned to the family (parents plus sibling) or its own cage (plus sibling at later ages).

Song processing and analysis

Similarity to the tutor song was scored using SAP2011 using the asymmetric-time course similarity tool. Song stability was measured by comparing the final 50 motifs produced during the evening (typically around dph 90) with the first 50 motifs produced the following morning (onset of singing) using the SAP2011 symmetric-mean value similarity tool. The self-similarity was the result of symmetric comparisons between the first 50 motifs produced at dph 130 to the first motifs produced during specific developmental ages. When the songs were variable (early developmental stages), motifs were manually segmented, using Raven Lite 2.0 (https://ravensoundsoftware.com/raven-lite) or semi-automatically with a custom MATLAB script. Using all-to-all comparisons (n = 50 motifs x n = 50 motifs), our analysis included a total of 2500 comparisons per bird.

Song trajectories and spectral variability index

We used variational autoencoders (VAEs) to obtain low-dimensional representations of song sonograms as described in,⁵⁸ adapting code from their Autoencoded Vocal Analysis (AVA) package (https://github.com/pearsonlab/autoencoded-vocal-analysis). In essence, the algorithm aims to automatically discover the optimal encoding of fixed-duration sonogram windows into a low number of features from which the original sonograms can be optimally reconstructed.

For the analysis of song trajectories (Figures 2A–2I), we randomly selected 100 songs from each bird and trained a common ‘shotgun’ VAE model of 16 latent features on 120 ms sonogram windows (0.25–10 kHz) from all songs of all birds. For each bird, we then computed the latent means from sliding sonogram windows of 120 ms (12-ms step) and projected them on a two-dimensional space with Uniform Manifold Approximation and Projection (UMAP;⁵⁷). The darkness of each dot in the UMAP projections represents local density by gaussian-kernel density estimation. Each song was represented on this space as a trajectory connecting the projections from its consecutive sliding windows. Lines connecting jumps larger than 3 times the median distance between successive time points were removed for clarity. Trajectories are color-coded by elapsed time for individual examples (Figures 2B–2E, and 2H) or with transparency for overlaid plots (Figures 2C–2F, and 2I).

We obtained a spectral variability index as an adaptation of the variability index in.⁵⁸ For this, we first selected 100 songs with a common syntactic structure from each bird. To account for subtle differences in timing across song renditions, we adjusted all the sonograms from each bird to a common time axis using 3-knot piecewise linear time warping as described in Williams 2020¹⁰³ and implemented in the affine warp package (https://github.com/ahwillia/affinewarp). We then used randomly sampled 80-ms windows of warped sonograms from all birds to train a common VAE model of 16 latent dimensions. For each bird, we first removed the borders of the time-warped sonograms, leaving only times for which over 95% of the songs had data. We then collected 200 windows of 80 ms at time points equally spaced within this time and encoded each to their 16 latent means with the trained model. For each time point, we computed the median absolute deviation of each latent dimension and then summed them to get the total variability, resulting in 200 variability measures per bird. From these, we computed the median to obtain the spectral variability index for each bird. We followed a similar pipeline for call variability index, analyzing a variable number of calls per bird (range 62–265 calls) with 0-knot time warping and 50 windows per call.

Intersyllable gap duration analysis

Intersyllable gap duration was quantified as the time between the offset of one syllable and the onset of the subsequent syllable. Measurements were performed on the same set of 50 days–90 motifs per bird used for the motif syntax analyses.

Rhythm analysis

Rhythmic structure in vocalizations was quantified using a custom Python pipeline following the chunk-based generate-and-test (GAT) framework described by.¹⁰⁴ Audio processing and feature extraction were performed using librosa, and statistical analyses were conducted with SciPy. The analysis comprised three stages: event extraction, pulse-frequency detection, and statistical validation using Monte Carlo simulations.

For event extraction, audio files were resampled to 44.1 kHz, and up to 10,000 s of vocalizations per individual were analyzed. Syllable onsets (tᵢ) were detected using spectral flux (librosa; delta = 0.1, wait = 5, hop length = 512). Syllable offsets were defined as the point at which RMS amplitude decayed to 10% of the post-onset peak. Events shorter than 10 ms were excluded. Note durations (D_n) and gap durations (D₉) were pooled across the dataset to construct null models.

Pulse periodicity was assessed using an exhaustive GAT procedure. Candidate pulse frequencies ranging from 5.5 to 100 Hz were tested in 0.05 Hz increments. For each frequency (f; period T = 1/f), onset phases (Φᵢ) were computed as tᵢ mod T and normalized to [0,1). Using circular statistics, the mean phase (Φ^-) was calculated, and rhythmic fit was quantified using the Fourier Randomized Mean Square Deviation (FRMSD). Phase deviations were computed as the minimum circular distance between Φᵢ and Φ^-. The frequency minimizing FRMSD was defined as the Best Pulse Frequency. Statistical significance was assessed using the R-model¹⁰⁴ with 50 permutations per file. For each permutation, note and gap durations were resampled from the global pools while preserving the number of events, and onset times were reconstructed cumulatively. A p-value was computed as the proportion of randomized sequences with FRMSD values lower than the observed FRMSD. Periodicity was considered significant at p < 0.05. For songs exhibiting significant periodicity, Best Pulse Frequencies were log₁₀-transformed and grouped using hierarchical clustering (average linkage, Euclidean distance), applying a clustering threshold of 0.025 as in,¹⁰⁴ to identify distinct rhythmic populations.

Motif syntax and motif variation analysis

To assess whether the manipulated animals exhibited changes in the song syntax, used here in the non-linguistic, birdsong-specific sense to denote the temporal ordering and transition structure of sung syllables, specifically in the arrangement of their sung syllables, we calculated the transition probabilities of the 50 songs used in the similarity analysis. Syllables were labeled manually by an observer blind to experimental conditions, using the first 50 motifs produced by each bird during solo singing on the morning of day 90. A spectrogram of each bird’s first motif served as the initial reference template, and subsequent motifs were labeled accordingly using the “segmented comparison module” in Sound Analysis Pro. When a previously unobserved syllable type appeared, that motif was adopted as the new reference template for continued labeling. Manual labeling was chosen because clustering-based approaches require sufficient instances of each syllable type to generate stable acoustic categories; given the low occurrence of some syllables, automated classification could be unreliable. To minimize labeling bias, minor temporal and spectral variations within syllable types were disregarded. Subsequently, we quantified the entropy rate of each animal’s song using the previously described equation¹⁰⁵:

$$Entropyrate=-\sum _{i,j}{\pi }_{ij}{P}_{ij}{\mathit{log}}_2\left(P_{ij}\right)$$

In this equation, i and j represent different syllables while Pij denotes the transition probability between the two states, which, in our case, corresponds to the sung syllables. Additionally, πi represents the stationary distribution of the model for state i. The entropy rate values range from 0 to 1, where a value of 0 indicates a higher level of organization, and a value of 1 signifies disorganization. Therefore, as the entropy rate values approach 1, the analyzed motif exhibits greater variability.

Hearing threshold

Hearing thresholds were estimated using a previously described behavioral paradigm.¹⁰⁶ In short, birds (post-hatch day 130–140) were acclimated overnight in a sound-isolated chamber (86 × 81 × 81 cm). The following morning, we ensured the bird was perched, and turned the lights off to proceed with the stimulus presentation. Pure tones from 0.5 to 8 kHz were presented at progressively higher intensities (10 dB steps ranging from 20 to 80 dB) using an amplifier and a calibrated speaker. Each tone lasted 3 s with a 10-s interval between each presentation. Webcams without infrared filters (Logitech C920) positioned above and in front of the perch capture the animal behavior. Infrared lighting enabled the visualization within the chamber without triggering unwanted movement. Stimulation was controlled using a custom-made MATLAB routine. Stimulus presentation was synchronized with an additional infrared LED using a microcontroller (Arduino UNO). For any given sound frequency, the threshold was defined as the lowest intensity that elicited a behavioral response (head or body reorientation toward the stimulus) within 3 s after stimulus onset, averaged across three trials. To ensure scoring reliability, the same behavioral events were independently quantified by two observers. Any ambiguous cases were subsequently reviewed with a third observer, and consensus was reached in all instances regarding the presence or absence of movement.

Social preference paradigm

To assess the potential social impairments among the manipulated individuals, we conducted a social experiment to examine their interactions with an unfamiliar conspecific of the same sex. The experimental setup consisted of a cage divided into a central and two lateral chambers (Figure 5F). The experimental bird was placed in the central chamber, while another bird, termed the stimulus bird, was randomly positioned in either the left or right chamber. The experimental bird could gain visual access to the stimulus bird by perching in front of a small round window (7.5 cm diameter). This perch was referred to as the “social perch.” On the opposite side, there was another perch positioned in front of a window that led to an empty chamber. We called this perch the “neutral perch”. The windows were deliberately misaligned so that birds could only see into the lateral chambers when perched on the closest perch and while facing the window. Importantly, the stimulus bird could not initiate visual contact because its perch was positioned below the window. A single stimulus bird was used in this entire experiment.

All animals were placed in the setup for two days prior to the test for habituation. The stimulus bird were always introduced in the morning (between 8 and 9 AM) and its test lasted 1 h. Sessions were recorded using a webcam (Logitech C920) positioned above the cage and videos were processed using DeepLabCu.^107,108 Briefly, we trained a network using DLC and extracted the x/y coordinates for each assigned body part of interest, and its likelihood (prediction certainty) values related to a tracking point for each video frame. With this, we were able to extract variables related to animal engagement, bird orientation (head direction), distance, and trajectory during behavioral trials. For engagement measurements, we trained a DLC network to automatically detect the two perches by placing tracking points along them. After analyzing the videos, the moments when at least one of these tracking points vanished (i.e., a likelihood less than 0.6) indicate that an animal is there. As a result, we extracted the timestamps of the perches’ ups and downs to calculate the animals’ latencies, social and neutral stimuli engagement, and permanence. For head orientation, distance, and trajectory variables, we trained another DLC network to track animal body parts during behavioral sessions. In this case, we indicated the head, left eye, right eye, body center, and tail base as body parts to be tracked in the DLC software. To train the model, we utilized a ResNet50 architecture within the DeepLabCut (DLC) framework, focusing on skeleton-based tracking rather than elliptical body representations. Skeleton-based tracking captures posture and movement with higher precision and anatomical fidelity than ellipse-based methods by modeling explicit keypoints and their kinematic relationships, allowing robust pose estimation even during dynamic motion (e.g., blurred motions due to bird’s flight) or partial occlusions.¹⁰⁹ Ultimately, we extracted the bird orientation on each frame by having the arctangent between the head and the bird’s beak center. We normalized these measurements by their probability to evaluate the animals’ preferred head direction during the trials. Finally, we calculated the distance and plotted the trajectory of the animals having the body center tracking point measures. We calculated the Euclidean distance between the x and y coordinates of the body between each consecutive frame to have the total distance traveled in each trial. To convert pixel distances to meters, we calculated a scaling factor by relating known physical distances of reference objects in the scene to their corresponding pixel lengths in the image. This calibration was performed by measuring the length of the bar frames with a caliper. The scaling factor (mm/pixel) was then obtained by dividing the measure in meters by the pixel distance, enabling, then, the conversion of tracked pixel coordinates into meaningful physical units for subsequent analysis. To define orientation toward the stimulus, we quantified the proportion of time birds faced the transparent window based on head orientation. Given the strong lateral positioning for visual focus lateralization of avian vision, birds with head angles between 0–90° and 270–360° were considered to be oriented toward the stimuli.

Social female interactions

To investigate the impact of embryonic VPA exposure on social courtship interactions, we conducted a behavioral task where male birds were placed in a social setting with females. For this, we utilized both vocal and video recordings. Experimental males were exposed to two types of female companions: a “familiar” female, with whom they had been housed for the preceding week, and an unfamiliar female, introduced for the first time. To assess male social interactions, individual experimental males were housed overnight in the same cage used for the individual task. The following morning, each male was introduced to either a familiar or unfamiliar female for a 1-h interaction session. This procedure was repeated the next morning with the other female type, and the order of female presentation was randomized. We did not observe differences when familiar or unfamiliar females were presented, thus only averages of the two sessions are shown.

Courtship and contextual singing behavior

To assess the appropriateness of female-directed courtship behaviors in treated animals, we analyzed videotaped sessions of male-female interactions.^61,110 Courtship behavior was defined by the presence of orientation toward the female, dance movements, and beak wipes accompanied by singing, whereas copulation attempts were analyzed separately as a subsequent mating outcome. This involved quantifying specific courtship actions performed by the males, including orientation toward the female, head and side to side movements (dance), beak wipes, and mating attempts. Specific behaviors were quantified by two independent, blinded observers, and their scores were averaged. This evaluation was conducted in adult birds after they had reached sexual maturity. To examine whether VPA-exposed males could adjust their singing behavior according to social context, we also recorded song production during the female exposure. We then compared the acoustic features (e.g., song duration, number of introductory notes, fundamental frequency) of female-directed songs to undirected songs. Songs were classified as female-directed when a male oriented itself toward the female and exhibited behaviors consistent with the defined courtship repertoire. Undirected songs were those produced in a solo context.

Quantification and statistical analysis

Statistical analyses were performed using GraphPad Prism 10 (GraphPad Software, San Diego, CA, USA) and R. Analysis of variance (ANOVA), followed by Dunnett’s multiple-comparisons test, was used for mean comparisons when the assumptions of independence of observations, normality of residuals, and homogeneity of variances were met. When one or more of these assumptions were violated, nonparametric analyses were performed using the Kruskal–Wallis test. Courtship data were analyzed using Fisher’s exact test or a paired t test for within-subject comparisons, as appropriate. To assess whether VPA induced changes in the timing of vocal learning, we evaluated nonlinear mixed-effects models fitted to the song similarity development curves using the ‘nlme’ and ‘emmeans’ packages in R.^111,112 Overall, we tested for fixed effects of treatment in the temporal parameters of each model while leaving those parameters related to curve height as random effects for each bird.

Similarity to tutor vs. dph curves were fitted with an asymptotic exponential model (‘SSasymp’), with fixed effect of treatment on log-rate constant and random effects by bird for the starting and asymptotic similarities. Reported time constants are the exponential of the negative of the log-rate constants. Self-similarity vs. dph curves were fitted with a symmetric sigmoidal model (‘SSfpl’), with fixed effect of treatment on the time of mid learning (t₅₀) and the timescale (the inverse of the learning slope), and random effects by bird for the low and high similarity asymptotes.

Published: April 28, 2026