We first sought to characterize differences in song nuclei gene expression specializations relative to surrounds in juvenile males and females with and without E2 treatment. We used RNA-Seq data from a previous study by our lab, on the effects of E2 manipulation on the song system in both males and females²⁵. This E2-treatment program masculinized females sufficiently for them to produce song as adults in another cohort of birds²⁵. Birds were sacrificed as juveniles at post-hatch day 30 (PHD30), after a one hour period of silence to limit activity-dependent gene expression in song nuclei and surrounding motor regions; this developmental age was chosen because it is around the time when both males and E2-treated females start to sing (e.g. subsong) and all four song nuclei are sufficiently developed to be visually apparent in histological sections²⁵. The four major song nuclei (HVC, LMAN, RA, and Area X) and their adjacent surrounding motor regions (DN, AN, LAI, and Str respectively; Fig. 1a) were dissected using laser capture. The E2-treated females in this study had similarly sized song system nuclei as males, indicating that E2 treatment prevented atrophy. This past study mapped the RNA-Seq reads to an older genome assembly lacking the W sex chromosome²⁵.

With this data, we first remapped RNA-Seq reads (from n = 3 birds per group) to a new zebra finch assembly (GCF_008822105.2) with both the Z and W chromosomes. In our quality control analyses, hierarchical clustering of sample expression vectors revealed two technical outliers, one HVC from a vehicle-treated male HVC and one RA from a E2-treated female, which we excluded from further analysis (Fig. S1). Similar to previous PCA and hierarchical clustering²⁵, the remapped RNA-Seq data expression levels with outliers removed still resulted in separation of song nucleus and surround and some E2-treated female samples, without sufficient resolution at finer group levels (Fig. S1). To address our questions at finer resolution and in an unbiased way, we performed weighted gene correlation network analysis (WGCNA), where we decomposed the transcriptome into actively expressed gene modules using hierarchical clustering of the gene adjacency matrix. This matrix describes the inferred structure of gene networks within our data and was calculated by soft thresholding the matrix of gene-to-gene expression correlations across samples (Fig. S2)³⁵. After initial hierarchical cluster-based module assignment, genes were iteratively reassigned to the module whose aggregate expression they best correlated with until no additional genes met the WGCNA reassignment threshold (Fig. S3 and Methods for parameterization). This optional reassignment was performed to ensure that genes are matched to the aggregate measure that best represents their expression.

Of the ∼21,000 annotated zebra finch genes, 13,220 were well expressed in the finch telencephalic brain regions sampled, a comparable number of genes to what we have seen expressed in adult zebra finch telencephalon^9,36. These 13,220 were assigned to 14 co-expression modules (Fig. S4a). The median observation of unassigned transcripts was 22-fold lower than the median observation of module assigned transcripts (4.86 vs. 0.22 FPKM). Two modules clearly marked single samples from two different birds (Fig. S4a,b - arrows), indicative of technical overfitting; these two modules (not birds) were excluded from further analysis, resulting in a more visibly diverse pattern within and across modules (Fig. S4c,d). The remaining 12 modules contained 12,444 genes in total were lettered in descending order of size A through L, containing from 4,890 to 127 constituent genes each, and were dynamically expressed across brain regions and treatment groups (Fig. 1c)

To understand song system transcriptional specialization in the context of the gene modules, we calculated the module eigengene (MEG) expression values for each of the 12 modules (Fig. 1d). MEGs are the first principle component of variance of all genes in a module and are the aggregate measure for each module’s expression across samples. We then tested for statistically significant correlations between MEG expression for each module and the song nuclei specializations relative to their respective surrounds. Each song nucleus had unique and overlapping specialization of genes in 2 to 4 modules of the 12 (Fig. 2a). The LMAN specialization relative to the surrounding AN correlated with the expression of gene modules B, F and I; HVC specialization relative to DN correlated to the expression of gene modules B, F, and G; RA specialization relative to LAI consisted of gene modules C and L; and Area X specialization relative to the surrounding Str consisted of gene modules C, F, G, and I. Area X was the only song nucleus that did not have a specialized module unique to it relative to other song nuclei (Fig. 2a).

In non-vocal learning juvenile females, interestingly LMAN was specialized relative to the AN by the same gene modules as in males (B, F, and I) as well as an additional module G (Fig. 2b); RA was specialized by module A as in males, but not module L and by additional modules A and G. In contrast, neither juvenile female HVC nor Area X exhibited significant gene module expression specializations relative to their surrounds. We interpret these findings to indicate that: LMAN and RA atrophy later in juvenile female development; that module A and G genes could be associated with the start of this atrophy; HVC and Area X are likely the first to atrophy or not develop; and lack of any gene module specialization in them at this age could mean that they would be more sensitive to estrogen prevention of vocal learning loss.

To assess the effects of chronic exogenous estrogen on the developing song system, we first performed a control analysis of modules in the E2-treated juvenile males. In the E2-treated juvenile males, the song nuclei specialized modules overlap to those seen in vehicle-treated males (Fig. 2a vs 2c); differences were that module I in LMAN and Area X and module L in RA were no longer present; and module G newly appeared in HVC, module A appeared in RA as seen in females, and module E also appeared in RA. These differences in gene module specializations in the male song system are consistent with the E2-treatment regimen used causing a slight decrease in vocal learning accuracy in males²⁵.

In contrast, E2-treated females more closely mirrored the gene module specializations seen in the vehicle treated male song system (Fig. 2a vs 2d). Specifically, module G was not associated to RA while module C became more strongly correlated; module G expression was strongly specialized in HVC; and module C appeared specialized in Area X. Despite these changes to be more like males, in E2-treated females, module A in RA maintained its specialization as seen in vehicle-treated females (Fig. 2b vs 2d), and module E appeared in RA as in E2-treated males (Fig. 2c vs 2d); a weak correlation of a module I in HVC and module K in RA appeared (Fig. 2d). That is, the transcriptional response of female song nuclei to E2-treatment appeared to be far more dramatic than in males, consistent with estrogen inhibition in males having only a limited impact on song nuclei and vocal learning²⁵.

We performed an additional test for E2-induced changes to gene module expression in females by comparing E2-treated song nuclei in females to the combination of E2-treated surrounds from the same animals, vehicle-treated female surrounds, and vehicle-treated female song nuclei (Fig 2e). This analysis represents a comparison between samples from vocal learning capable female samples and non-vocal learning female samples at each node. It revealed 1 to 3 modules in each song nucleus of E2-treated females that were a subset of some of the strongest correlated modules found in control males (Fig. 2a vs 2e). These were: module F in LMAN; module G in HVC; module C in RA; and modules C, F and G in Area X. That is, these are four core modules for song nuclei that are the most sensitive to E2-induction or enhancement in females and most associated with the presence of vocal learning in females.

We next tested whether some modules could be explained by sex differences in the brain. We compared vehicle treated male and female song nuclei and surrounds separately, and found that module E strongly correlated with higher expression in all male brain regions relative to females (Fig. 2f,g). Among the song nuclei, module G eigengene expression was significantly higher in male HVC and Area X; conversely, module C expression was higher in female LMAN and RA (Fig. 2f). Among the surrounds, module H eigengene expression (not significant in any other comparison) was significantly higher in female DN and module C was significantly higher in female AN and Str (Fig. 2g). These findings indicate the existence of a sex specific, telencephalon-wide shift in the transcriptome represented in these data by module E. This broad sexual dimorphism is further refined by regionally patterned sex differences like those observed for module G in HVC.

We next quantified the magnitude of module expression differences and observed several gene modules with clear region-specific expression patterns (Fig. 2h-k). Module A, which appeared specialized in RA relative to LAI in several comparisons, was highly expressed in all Area X and Str samples, regardless of treatment or sex (Fig. 2h). Interestingly, FOXP2, a gene critical for spoken-language and vocal learning in humans³⁷ and songbirds³⁸, was a potent member of module A, correlating with the eigengene at r² = 0.92 across all samples. Module C, which was specialized to several song nuclei including RA, HVC, and Area X relative to their surrounds and sexually dimorphic in LMAN and Str, was most highly expressed in arcopallial samples, especially RA, with some increase in HVC (Fig. 2i). Module F, which was specialized to LMAN relative to its surround in all such comparisons, was highly expressed exclusively in LMAN regardless of sex and E2-treatment (Fig. 2j). Module G, which was specialized to HVC in a sex and E2 dependent manner, was only highly expressed in HVC samples, with males having higher expression than females, and females further split by treatment (Fig. 2k).

We sought to understand the cumulative biological function of genes among the modules. To do this, we mapped the zebra finch genes to their 1:1 human orthologs where possible and then used human gene annotation to examine the Gene Ontology (GO) functions enriched within each module’s constituent genes. Of the 12,444 module assigned genes, 7,909 (63%) had 1:1 human orthologs annotated in Ensembl. We found GO terms significantly enriched in module G (Table 1), which included “DNA binding transcription factor activity”, “cell differentiation”, “anatomical morphogenesis”, “cell-to-cell signaling”, and “positive regulation of multicellular organism growth”, indicating that module G genes specialized in male and E2-treated female HVC and Area X potentially enact a proliferative developmental program. Other significantly enriched terms were “extracellular matrix structural component”, “external side of the plasma membrane”, and “extracellular space”, indicating that this module may also act to restructure the extracellular matrix, perhaps to accommodate new cells. Module E, which was differential between the sexes for both song nuclei and surrounds, had 6 significantly enriched terms, of which 3 pertained to DNA damage repair: “nucleolus”, “transcription, DNA templated”, and “U2-type precatalytic spliceosome” (Table 2). These results in module E indicate that there is a telencephalon-wide sexually dimorphic gene expression program in finches that likely acts within the nuclear environment. A full table of GO enrichments by module can be found in supplemental data (supplemental_sig_go_enrich.csv).

We next asked if any of the modules were enriched for genes previously determined to be convergently specialized in songbird song nuclei and human speech brain regions^7,9. We performed this analysis by treating the different sets of convergently specialized genes as gene ontology terms. These gene lists included the transcriptional convergence between RA and human dorsal laryngeal motor cortex (dLMC), HVC and dLMC, LMAN and dLMC, Area X and the anterior caudate, and Area X and the anterior putamen (Gedman et al. Fig3B,D⁹). We found that each of these convergent gene lists was enriched for the modules specialized to or highly expressed in the finch brain region used to generate the list (Fig3a). We found that module B, specialized to both LMAN and HVC (Fig 2a), was enriched for the convergently specialized genes in human dorsal laryngeal motor cortex (dLMC; Fig 3a)⁹. Module C, specialized to RA (Fig 2a,i), was enriched for the genes convergently specialized in dLMC and RA (Fig 3a) known to match the lower layers of the cortex⁹. Module A, highly expressed in Area X and Str (Fig 2h), was enriched for genes convergently specialized in Area X and human anterior striatum (both caudate and putamen; Fig. 3a).; Module I, specialized to Area X (Fig 2a), was also even more strongly enriched for the same convergences as module A (Fig 3a). These findings indicate that the genes previously identified as convergently regulated between songbird song brain regions and human speech brain regions are components in the larger specialized gene networks identified here using WGCNA.

We next determined whether any modules were enriched in genes from specific chromosomes. We performed a bootstrapped analysis, randomizing the mapping between genes and modules 50,000 times to approximate null distributions. P values for each chromosome-module pairing were then FDR corrected. To do this as accurately as possible, we performed this analysis using revised chromosomal assignments and structure from the most recent zebra finch genome assembly, bTaeGut_1.4.pri, which better resolves the microchromosomes³⁹ (see Fig. S6 for the less resolved bTaeGut_W.pat.v2 assembly). Surprisingly, we found that each of the 12 modules were enriched for genes from at least one chromosome (Fig. 3c). The most striking enrichment was of module E genes on the Z and W sex chromosomes, with nearly all expressed W genes and ∼⅔ of expressed Z genes being members of module E. This is consistent with module E exhibiting higher expression in the male telencephalon regardless of treatment or region (Figs. 2d-e). For the autosomes, we observed significant enrichments of two HVC specialized gene modules on chromosome 1A; module B (specialized in male HVC and human dLMC); and module G (specialized in male HVC and E2-treated female HVC and sexually dimorphic in vehicle-treated HVC). These results are particularly interesting given that zebra finch chromosomes 1 and 1A are believed to be the result of a songbird-specific fragmentation of the ancestral chromosome 1 found in chickens^40,41.

Module A, the largest module which was highly expressed in Area X and adjacent striatum, was enriched across 5 macro-chromosomes (chr1, 2, 3, 4, and 10). Module F, which was strongly expressed in LMAN, was enriched on chromosomes 2 and 7. Modules C (most strongly expressed in the arcopallium and part of RA and Area X specializations) and I (a component of LMAN specialization) was enriched on chromosome 6 on microchromosome 35. Module D, a component of the LMAN specialization in E2-treated females, was enriched across 8 small microchromosomes (chr22, 23, 25, 27, 28, 29, 30, 31). 4 of the 5 smallest modules in gene counts were enriched in the microchromosomes: Module H, which was sexually dimorphic only in DN, was enriched on microchromosome 28; module K was significantly enriched on microchromosome 37 and in the “other” category, which includes all remaining unnamed DNA scaffolds in the assembly such as further microchromosomes; module J was also enriched in “other”.

The telencephalon-wide, sexually dimorphic expression of the sex chromosome dominated module E is one of the strongest features in the present dataset; Module E eigengene is highly expressed in all male samples and depleted in all female samples regardless of brain region or treatment (Fig. 4a vs 4c-d). To better understand the relationship between the sex chromosomes and module E, we examined the distribution of membership in module E with the sex chromosomes separated. WGCNA allows us to consider gene membership in a module as a continuous variable, rather than a binary variable, by correlating each gene’s expression profile to the module eigengene. Doing this for module E, we found that Z transcripts were positively correlated to the module eigengene while W transcripts were anticorrelated (Fig. 4e). This is consistent with Z chromosome transcripts being generally depleted in females relative to males, while W chromosome transcripts were only expressed in female brains. This reduction in Z chromosome transcript abundance is consistent with the finding that diploid Z chromosomes in male birds do not have one copy inactivated to compensate for gene dosage, unlike the X inactivation in female mammals^30,42.

Given the sexually dimorphic expression of module E across brain regions and the enrichment of the sex chromosomes within that module, we separately compared the expression of sex chromosome genes in the non-vocal motor surrounds of vehicle-treated control males and females regardless of WGCNA assignment to better understand sex chromosome expression without the influence of vocal learning specialization or E2 response. To do this, we correlated expression of each sex chromosome gene to the sexual dimorphism, such that female-enriched transcripts were positively correlated and male-enriched transcripts were anticorrelated (Fig. 4f,g). In these gene sets, we identified between 73 and 82 significantly expressed W chromosome genes and between 433 and 527 significantly depleted Z chromosome genes for each of the surround regions in males relative females. Examining the union of these gene sets revealed that a total of 95 W and 694 Z chromosome genes were differentially expressed between sexes in at least one brain region, 62% and 65% of annotated sex chromosome genes respectively. Conversely, the intersection of these regional gene sets contained 58 significantly expressed W chromosome genes and 306 significantly depleted Z chromosome genes in all non-vocal regions in females, representing 38% and 29% of annotated sex chromosome genes, respectively (Fig. 4h-i, Table S2). This indicates that there is a telencephalon wide set of sex enriched and depleted sex chromosome genes in addition to regionally patterned sex chromosome gene expression.

We also observed strong enrichments of several of the smaller gene modules, J and K, on microchromosomes. Examining the MEG expression for these modules, we found that both were expressed brainwide in non-overlapping sets of animals. Module J was highly expressed in all samples taken from animal-b, an E2-treated female (Fig. 4b); similarly, module K was most highly expressed across all samples from animal-g, an E2-treated male, and exhibited intermediate expression in all samples from animal-f, a vehicle-treated female (Fig. 4c). Though not enriched on microchromosomes, Module L exhibited similar animal-specific high expression in animal-a, an E2 treated female (Fig. 4d). These animal and chromosome specific shifts in the transcriptomes could represent the systemic effects of allelic chromosomal structural variation which WGCNA was able to capture due to our broad sampling of regions across the telencephalon.

As it is unlikely that the gene modules act independently of each other, we sought putative interacting genes between two modules of interest, module G specialized in HVC and Area X and module E dominated by sex chromosome genes. To do this, we again replaced binary in-or-out module membership with continuous module membership by correlating each genes’ expression to the relevant module eigengene³⁵. This allowed us to quantify the extent to which any gene was associated with any module, regardless of initial assignment³³. Looking across all assigned genes for our modules of interest, we identified two outlier genes, PDE8B and HABP4, which were the most E associated gene assigned to module G and the most G associated gene assigned to module E, respectively (Fig. 4j). Both genes were significantly upregulated in male HVC relative to its surround, but not in female HVC of either treatment. PDE8B catalyzes the hydrolysis of the second messenger cAMP and mutations to the gene cause an autosomal dominant form of striatal degeneration in humans⁴³. HABP4 is an RNA binding protein, known to repress the expression and subsequent DNA binding of MEF2C⁴⁴, a Z chromosome transcription factor which has undergone accelerated evolution in songbirds¹⁰ and whose repression by FOXP2 is critical for cortico-striatal circuit formation in mice related to vocal behaviors⁴⁵. Both PDE8B and HABP4 are found on the Z chromosome and HABP4 was one of the 694 Z transcripts significantly depleted across the telencephalon in vehicle-treated females relative to males (Table S2). These findings indicate that Z chromosome genes may be subject to multiple gene regulatory programs; both the telencephalon wide depletion driven by reduced copy number and specialized upregulation in HVC for vocal learning behavior.

To reduce the 344 genes in module G to the putative drivers of masculine HVC development, we next examined the relationship of continuous membership in module G (correlation to MEG-G) to gene expression specialization in HVC at the level of single genes. We did this by correlating gene expression to the specialization of HVC in males (vehicle- and E2-treated) or E2-treated females for each of the following four vocal learning comparisons: 1) male HVC specialization relative to the surround; 2) E2-treated female HVC specialization relative to the surround; 3) HVC sexual dimorphism in vehicle-treated controls; and 4) E2-treated female HVC specialization relative to vehicle-treated female HVC. For all four comparisons, we found the higher the correlation of module G genes to the MEG, the higher the correlation with the vocal learning specialization (Fig. 5a-d). This means that their expression was higher in male or E2-treated female HVC relative to the appropriate non-vocal learning controls across all comparisons. We identified genes of interest as being strongly correlated (r² ≥ 0.5) to both the module G eigengene and vocal learning specialization (Fig. 5e-h, higher magnification view of colored boxes in Fig. 5a-d). All genes of interest exhibited a positive correlation to vocal learning specialization across all four comparisons (Fig. 5a-d). We next generated a core gene list for module G involvement in E2- and sex-dependent HVC development by intersecting these four vocal learning specialized gene sets (Fig. 5i). We found a core set of 14 genes that strongly marked vocal learning capable HVC in all comparisons and strongly correlated with the aggregate of module G expression (Table 3, Fig. S6).

We examined the chromosomes of origin for these 14 core genes and found that three (GHR, RGS7BP, and THBS4) were on the Z sex chromosome (Fig. 5j). This was a >3-fold significant enrichment over chance of Z chromosome transcripts (p = 0.009, upper-tailed hypergeometric test) against a background of module assigned genes. This result was statistically significant regardless of the background gene set, when using all genes (∼21,000) or only module G members (344). These three Z chromosome module G genes in HVC not only exhibited >50% reduced expression in control female HVC relative to males, but also exhibited upregulated expression in male HVC relative to the surrounding DN regardless of treatment and upregulation in E2-treated female HVC relative to either the surround or vehicle-treated female HVC (Fig. S7). In the context of our cross-region sex chromosome analysis (Fig. 3c-f), only RGS7BP was sexually dimorphic outside of the song system, being depleted in all female surround regions. Neither GHR (growth hormone receptor) nor THBS4 were significantly depleted in any surround region comparison. Taken together, these results indicate that the 3 core genes in vocal learning capable HVC on the Z chromosome are subject to additional E2 sensitive transcriptional regulation in HVC, separate from the Z chromosome transcript depletion seen throughout the female brain.

Of the 14 core genes total, several have been previously studied in the brains of other species which may inform their role in vocal learning systems. THBS4 encodes a secreted extracellular-matrix glycoprotein necessary for appropriate neuronal migration in the mouse⁴⁶ and is elevated 6-fold in the human cortex compared with non-vocal learning primates⁴⁷. Human EDA2R was recently identified as a top correlate of cognitive performance and brain size⁴⁸; it was also found in a human GWAS study to correlate with circulating estrogen and testosterone levels⁴⁹. Rare mutations in human PHETA1 lead to Lowe oculocerebrorenal syndrome, that includes pathophysiology in seizures, mental retardation, and structural brain abnormalities^50,51. SIX2 is a homeobox domain containing transcription factor that governs early brain and craniofacial development and provides neuroprotection from dopamine injury^52,53. GHR encodes a transmembrane receptor whose activation controls cell division⁵⁴. The gene which encodes GHR’s ligand, growth hormone, is interestingly duplicated and undergoing accelerated evolution in the genomes of songbirds, including the zebra finch⁵⁵. Based on these findings, we consider GHR as the most likely candidate gene related to the E2 sensitive atrophy of HVC in females. Based on the findings in these gene sets, we hypothesize that without excess estrogen in females, HVC expansion is prevented by not specializing the growth and neuronal migration promoting genes in module G to the HVC lineage by late development. This is potentially enacted by depleting necessary gene products from the Z sex chromosome, such as GHR, which are already present in only one copy.