To better understand the evolutionary history of the ZAL2^m chromosomal rearrangement, we generated additional sequence data from a rare, ‘super-white’ (ZAL2^m homozygote) bird (Horton et al., 2013; Sun et al., 2018). We generated variable fragment size libraries consisting of 150 bp paired-end reads (insert size of 300 bp and 500 bp) and 125 bp mate pair reads (insert size of 1 kb, 4–7 kb, 7–10 kb, and 10–15 kb). We performed whole-genome sequencing of an additional 62 birds (49 white-striped birds and 13 tan-striped birds sampled from a variety of locations around the U.S.) (Materials and methods, Supplementary file 1). White-striped birds, which are heterozygous for the rearrangement (ZAL2/2^m), were sequenced at higher coverage than tan-striped birds (ZAL2 homozygotes) so that we could obtain sufficient reads to separate ZAL2 and ZAL2^m alleles in white-striped individuals (average mean depth coverages were 41.5 × vs 28.4 × for white- and tan-striped birds, respectively, Supplementary file 2). Genomic variants were called according to the guidelines of Genome Analysis Toolkit (GATK) (ver. 4.1) (Materials and methods), leading to the discovery of a total of 11,382,994 single nucleotide polymorphisms (SNPs). None of the samples showed evidence of family relationships when we computed relatedness estimates between individuals. Consequently, we used all samples in the subsequent analyses. We found a significantly higher number of polymorphic sites within white-striped birds than tan-striped birds exclusively for ZAL2/2^m chromosomal regions (Figure 1—figure supplement 1). Nucleotide diversity of the rearranged region of the ZAL2/2^m chromosomes was elevated in white-striped birds compared with tan-striped birds, suggesting distinctive patterns between the two plumage morphs (Figure 1a).

Among the total SNPs identified, 12.6% (N=1,439,991) resided on scaffolds we have previously assigned to the ZAL2/2^m chromosome (Sun et al., 2018). Principal component analysis (PCA) of these ZAL2/2^m SNPs revealed distinct clusters corresponding to the morphs (Figure 1b). The first principal component (PC1), which explained 6.7% of the variation in the data, clearly separated tan- and white-striped birds, with the lone super-white individual (ZAL2^m/2^m homozygote) as a clear outlier. In contrast, other available phenotypic information, including sex and geographic origin of samples, did not show meaningful variation with the principal components, and other PCs had little explanatory power (Figure 1b). Tests for admixture also failed to identify significant population substructures by geographical origin of samples (Figure 1—figure supplement 2). This lack of population structure is unsurprising, as 35 of the 63 samples (56%) were from birds that were migrating, and, thus, the breeding location of these birds is unknown.

We examined several genomic features of the ZAL2^m chromosome using the additional genomic resources we generated. We first performed a de novo genome assembly of the super-white bird, employing newly generated sequence data, to study the ZAL2^m chromosome with an assembly derived entirely from a bird homozygous for the ZAL2^m chromosome. The total assembly size was 1058 Mbp (N50 length of 3.1 Mbp, longest scaffold 27 Mbps), comparable to that of the ZAL2/2 reference assembly (1052 Mbp, N50 scaffold length of 4.86 Mbp, longest scaffold 45 Mbp) (see Supplementary file 3 for more details). There were 160 putatively ZAL2^m-linked scaffolds (Materials and methods), with a total length (110.99 Mbp) comparable with that of ZAL2-linked scaffolds from the reference assembly (108.5 Mbp Tuttle et al., 2016). Despite this similarity in total length, however, the average length of the individual ZAL2^m-linked scaffolds was significantly shorter than scaffolds on other chromosomes in the super-white assembly (p<0.001, Mann-Whitney U-test). It was also shorter than the average scaffold length on the ZAL2 chromosome in the ZAL2/2 reference assembly (Figure 2a). We did not observe such a pattern in the other chromosomes of similar size when comparing between the two assemblies (Figure 2a). This result was consistent with the presence of repetitive DNA sequences on ZAL2^m causing more assembly breaks compared with the ZAL2/2 reference genome. We found evidence that the ZAL2^m chromosome contained more repeat elements and was especially enriched for long terminal repeat elements (2.4 Mbp vs 2.1 Mbp) and interspersed repeats (5.8 Mbp vs 5.5 Mbp), compared with the ZAL2 chromosome. The number of these repeat elements is likely to be underestimated, given that the ZAL2^m assembly is highly fragmented. Additionally, we found that ZAL2 and ZAL2^m had accumulated a higher proportion of structural variants (insertions and deletions) compared with other chromosomes (Figure 2b).

Next, using the large amount of newly generated population genomic data, we examined patterns of SNPs on ZAL2^m alleles separately from those on ZAL2 via haplotype phasing using fixed differences between the two chromosome types (Materials and methods). We found that the total number of genetic variants was approximately 12-fold reduced on the ZAL2^m alleles compared with ZAL2 alleles (29,420 vs 367,466 SNPs and 3,921 vs 32,479 indels on ZAL2^m and ZAL2, respectively, after excluding singleton variants). The mean nucleotide diversity was similarly reduced on the ZAL2^m chromosome compared with the ZAL2 chromosome (0.0104% vs 0.1141% for ZAL2^m vs ZAL2, respectively).

Analyses of the genetic variants on the ZAL2 and ZAL2^m alleles showed evidence of only weak genetic degeneration. The ratio of non-synonymous to synonymous fixed differences inside the rearranged region was slightly, but significantly, elevated for ZAL2^m-derived compared with ZAL2-derived fixed differences (Figure 2c), which is consistent with either positive selection or inefficient purifying selection on ZAL2^m. We found that the ratio of non-synonymous to synonymous nucleotide diversity (π_N/π_S) was significantly increased on ZAL2^m compared with ZAL2 (p=1.3 × 10^–13, Mann-Whitney U-test, Figure 2d). The minor allele site frequency spectrum (SFS) for the ZAL2^m synonymous and non-synonymous sites showed a large proportion of singleton variants and an irregular decay of allele frequency as the minor allele count increases (Figure 2e), also suggesting reduced efficacy of purifying selection on ZAL2^m.

Assuming the mutation rates of the two chromosomes are similar, the ratio of effective population size (N_e) between the ZAL2^m and ZAL2 can be approximated by the ratio of nucleotide diversity of synonymous sites between the ZAL2 and ZAL2^m. The proportion of N_e between ZAL2^m and ZAL2 is 0.12 (±0.01), which is threefold lower than the expected ratio of 0.33 (because the ZAL2^m chromosome is 1/3 as frequent as the ZAL2, given ‘balanced’ morph frequencies observed in the wild). This lower proportion suggests that N_e of ZAL2^m has undergone further reduction than expected from the census size in the population, consistent with the effects of reduced recombination. We noted that the linkage disequilibrium between variants on ZAL2^m exhibited the classic decay with distance (Figure 2f), indicating at least some level of recombination. Taken together, these results are consistent with reduced, but not entirely eliminated, recombination on the ZAL2^m chromosome.

We next examined whether degeneration inside the rearranged region on ZAL2^m has resulted in globally reduced expression of the alleles carried by the ZAL2^m supergene variant. To do so, we used multiple large RNAseq datasets from a variety of tissues in birds sampled from different geographic locations and times of year (see Materials and methods, Table 1). As predicted and consistent with what was previously reported (Sun et al., 2018), we found evidence of consistently reduced expression of the ZAL2^m alleles in 8/10 types of tissue (Figure 2g). We next tested for an association between the number of accumulated mutations (non-synonymous, synonymous, and in the promoter region) on ZAL2^m and allelic bias (AB) in expression of the ZAL2^m alleles within each tissue, which would link genetic degeneration within or near genes to reduced expression of ZAL2^m. We found evidence that allelic bias in gene expression was associated with the rate of non-synonymous fixed differences (X²(1)=9.97, p=0.00159), although the effect size was exceedingly small (marginal r²=0.0020) (Figure 2—figure supplement 1a). Neither the rate of synonymous fixed differences (X²(1)=1.0098, p=0.315, Figure 2—figure supplement 1b) nor the number of fixed differences within 1 kb upstream of the transcription start site (X²(1)=0.8992, p=0.343, Figure 2—figure supplement 1c) were associated with allelic bias. Thus, the overall reduction in expression of the alleles carried by the ZAL2^m supergene is weakly associated with an increased number of non-synonymous fixed nucleotide changes within genes. The limited and weak nature of the effect suggests, however, that the pattern of gene expression may have been affected also by other factors, for example ongoing selection (thus manifested as nucleotide polymorphism), selection at more distal sites, and/or epigenomic mechanisms, such as differences between ZAL2 and ZAL2^m in DNA methylation or histone modification (see Sun et al., 2021).

Although the level of genetic diversity was overall reduced on ZAL2^m, it was exceptionally elevated in one region corresponding to ~3 Mbps spanning 5 scaffolds. This region, henceforth referred to as the ZAL2^m outlier region (Figures 2 and 3), includes at least 15 protein-coding genes that are well conserved as single copy genes across 13 avian species (Table 2). On average, nucleotide diversity in ZAL2^m across this region was 0.001, which is tenfold higher than the mean nucleotide diversity of ZAL2^m and even exceeds the nucleotide diversity in the corresponding region within ZAL2 by 3.15-fold.

To first examine whether this region has recently experienced an introgression event that could have caused the observed patterns, we constructed a phylogenetic tree of this region. The resulting tree exhibited the same topology as those from other regions of the ZAL2^m, thus providing no support for introgression (Figure 3b). We have not yet resolved the distribution of repetitive elements in the two chromosomes. Thus, to avoid errors in the phylogenetic analysis resulting from the rapid turnover of repetitive elements, we also performed the analysis using only the exons of single-copy orthologous genes, similar to analyses done by Stolle et al., 2022. The results (Figure 3—figure supplement 1) did not provide support for an introgression event and therefore did not change our conclusions. In addition, we calculated the D-statistic (ABBA-BABA test, which tests for genomic regions that are discordant with the species tree Martin et al., 2015) to test for evidence of gene flow from another species from the same genus, Harris’ sparrow (Zonotrichia querula). Sliding window estimates of the D-statistic did not show any differences in patterns between haplotypes (Figure 3—figure supplement 2), suggesting that the multiple haplotype structures in ZAL2^m were not introduced by gene flow from this closely related species or from the ZAL2 chromosome. Nonetheless, the ZAL2^m outlier region showed longer branch lengths than did both the corresponding region on the ZAL2 chromosome and a randomly selected region of ZAL2^m, reflecting the high genetic diversity within this region (Figure 3b). Tajima’s D was significantly higher throughout the ZAL2^m outlier region compared with the genomic background (p<0.001, permutation test) (Figure 3a).

The increase of nucleotide diversity, high Tajima’s D, and long branch length all suggest that balancing selection has impacted the ZAL2^m outlier region. Concordantly, we identified a signature of balancing selection in the ZAL2^m outlier region on the basis of allele frequency spectrum (β statistics) computed using BetaScan, which detects signatures of balancing selection by looking for an excess of genomically proximate SNPs with a similar allele frequency (Siewert and Voight, 2020). Visual examinations of the ZAL2^m outlier region also revealed the presence of potential long-range haplotypes that likely reflect long-term balancing selection and several recombination events within the haplotypes (Figure 3c). Notably, we identified two major putative haplotypes, which we refer to here as 1 and 2, consisting of 17 and 30 ZAL2^m chromosomes, respectively. A possible third haplogroup was also identified, although the sample size (n=2) was too small to warrant further analyses. The genetic differentiation between the two ZAL2^m outlier haplogroups was lower than the differentiation between either haplogroup and ZAL2 (Figure 3d), indicating that the haplogroups have evolved since the split between ZAL2 and ZAL2^m. This interpretation is also consistent with the inferred phylogeny (Figure 3b). Note that the sequencing depth for the two haplogroups did not differ significantly (p=0.78, Figure 3—figure supplement 3). The ZAL2^m outlier regions exhibited an excess of non-synonymous SNPs compared with synonymous SNPs (Figure 2d), as well as an excess of intermediate frequency minor alleles (Figure 3—figure supplement 4). Neither geographic location of collection nor sex was associated with any distinct patterns by haplogroups in a PCA analysis using genetic variants in the region (Figure 3—figure supplement 5; although note the breeding location of the birds from two of the locations was unknown). Together, these results solidly implicate balancing selection in the evolution of the ZAL2^m outlier region. In other words, multiple haplotypes of the ZAL2^m chromosome are maintained within the population of white-striped birds via balancing selection.

We looked for phenotypic consequences of the putative ZAL2^m haplogroups using the subset of white-striped birds from the sequencing samples for which we had phenotypic data (n=6 for haplogroup 1, n=12 for haplogroup 2, see Zinzow-Kramer et al., 2015; Horton et al., 2014 for details). White-striped birds with the H1 vs the H2 versions did not differ significantly in aggressive behavior (measured during simulated territorial intrusions), gonad size, tarsus length, cloacal protuberance volume, or any other measure when using the Benjamini-Hochberg correction for multiple testing to avoid false positives (more than 50 tests were performed).

We next tested for effects of haplogroup on the expression of genes inside the ZAL2^m outlier region. In the subset of white-striped birds for which we had haplogroup information, we examined RNAseq data from two brain regions, the hypothalamus (Hyp) and the ventromedial arcopallium (AMV, the functional homolog of the medial amygdala, formally named the nucleus taeniae of the amygdala) (see Materials and methods and Supplementary file 4 for details). We found that the gene GREB1 (growth regulating estrogen receptor binding 1) was more highly expressed in haplogroup 2 (unadjusted p=0.0195) in AMV. Additionally, there was a trend for the gene KCNS3 to be more highly expressed in both Hyp (unadjusted p=0.0511) and AMV (unadjusted p = 0.1567). This gene encodes a voltage-gated channel subunit that in humans and mice is specific to fast-spiking parvalbumin (inhibitory) neurons (Georgiev et al., 2014; Miyamae et al., 2021). Although no genes were significantly differentially expressed at the genome-wide level, our current findings provide potentially interesting candidates to explore further in the context of behavioral differences between the morphs.

Our work has provided a refined map of genetic differentiation between the ZAL2 and ZAL2^m chromosomes (Figure 4—figure supplement 1). Interestingly, three scaffolds near the chromosomal end (i.e. at the beginning of the p-arm, as well as at the end of the q-arm in the genomic coordinate of the ZAL2 chromosome), exhibited both decreased F_ST values and a reduced rate of fixed genetic differences (D_f). One possible explanation for these findings is that these regions represent a younger evolutionary stratum, which is possible because ZAL2^m contains a series of nested inversions that occurred at different times (Lahn and Page, 1999; Xu et al., 2019; Bergero et al., 2007). Another possibility is that the inflated within-ZAL2 nucleotide diversity (see Figure 3a) within the sampled population reduced the F_ST estimate.

By leveraging our resource of a haplotype map of both the ZAL2 and ZAL2^m chromosomes, we investigated signatures of positive selection in each chromosome. We identified 216 20 kB regions on ZAL2 exhibiting a significantly elevated H-statistic (empirical p-values <0.05), which is a measure of the average length of the stretches of linkage disequilibrium between the haplotypes (Schlamp et al., 2016; Messer Lab — Resources, 2014, Materials and methods). One notable stretch, which included four top candidate 20 kB regions (Scaffold NW_005081582.1, 480–520 kb and 920–960 kb) showed evidence of positive selection on both the ZAL2 and ZAL2^m chromosomes (blue shaded area in Figure 4—figure supplement 2). On the basis of previous literature (Thomas et al., 2008), this region on ZAL2^m may be placed on the end of that chromosome. The elevated H-statistic in this region cannot be explained by the recombination rate, as recombination is low across the entire length of ZAL2^m (Figure 4—figure supplement 3). On the ZAL2 chromosome, in contrast, recombination on the chromosome ends is increased (Figure 4—figure supplement 3), as has been reported in other species (Nachman and Churchill, 1996; Barton et al., 2008). We did not, however, observe an elevated H-statistic in these regions (Figure 4—figure supplement 2a). Note that a structurally resolved chromosome-level assembly does not yet exist for the white-throated sparrow. Thus, the precise location of these regions on the ZAL2^m chromosome still need confirmation.

We also identified a long stretch (~6 Mbp) showing an overall elevated H-statistic in another region of ZAL2 (red shaded area in Figure 4—figure supplement 3a, Supplementary file 5). This region exhibited the lowest estimates of nucleotide diversity within ZAL2 and the highest estimates of F_ST between the two chromosome types. Overall, there were 68 genes located in regions with a significant (p<0.05) H-statistic for ZAL2^m and 109 genes in regions with a significant H-statistic for ZAL2 (Supplementary file 5), meaning that these genes are inside regions showing evidence of a selective sweep. These observations suggest that selective sweeps took place on both ZAL2 and ZAL2^m chromosomes after the rearrangement events.

Some of the behaviors that differ between morphs, namely aggression, are predicted by the expression of several genes located inside the ZAL2/2^m rearrangement, such as ESR1 and VIP (Merritt et al., 2020; Maney et al., 2020; Horton et al., 2020; Zinzow-Kramer et al., 2015; Horton et al., 2014). We hypothesize that the ZAL2^m supergene region is enriched for alleles that have been shaped by antagonistic selection. Specifically, we predict alleles within the ZAL2^m supergene region that are beneficial for birds of the white-striped morph, but disadvantageous for birds of the tan-striped morph. Similarly, we predict that the ZAL2 chromosome is enriched for genes with alleles that benefit birds of the tan-striped morph (Maney et al., 2020).

To test this hypothesis, we examined whether allelic bias in the expression of genes inside the rearrangement was associated with differential expression of the same genes between the tan- and white-striped morphs (lists of genes showing differential expression or allelic bias can be found in Supplementary file 6). Here, allelic bias is defined as the differential expression of the ZAL2 versus ZAL2^m alleles in white-striped birds. Allelic bias in expression would indicate a role of cis-regulatory variants—i.e., variants in non-coding promotor, enhancer, or silencer regions regulating a gene—in functionally altering the expression of that gene. For these tests, we used data on gene expression in the brain only, as heart and liver tissue samples were limited to white-striped birds.

First, we found that among the genes that were differentially expressed between morphs, more of these genes are located inside the rearranged region on ZAL2/2^m than expected by chance, pointing toward cis- (as opposed to trans-) regulatory differences underlying morph differences in expression (X² tests showed highly significant effects for both brain regions in adults and nestlings of both sexes, with false discovery rate, FDR, correction to account for the multiple statistical comparisons performed using the Benjamini-Hochberg correction) (Figure 4a). To explore this further, we next tested whether the differentially expressed genes showed greater allelic bias than genes without differential expression. Consistent with what we previously reported using only males (Sun et al., 2018), we found that differential expression significantly predicted the degree of allelic bias in expression of that gene (X²(2)=664.16, p<2.2 × 10^–16, controlling for sequencing batch and brain region, see Materials and methods). In addition, tan-biased genes showed greater ZAL2 allelic bias and white-biased genes showed greater ZAL2^m allelic bias than did genes that were not differentially expressed between morphs (post-hoc T>W vs T=W: z=–16.87; post-hoc W>T vs T=W: z=19.63, p<2.2 × 10^–16; post-hoc W>T vs T>W: z=26.35, p<2.2 × 10^–16) (Figure 4b). These findings suggest that gene expression differences between the morphs are driven, in part, by evolutionary changes in the regulatory regions of genes captured by the rearrangement.

We also looked for a relationship between gene expression and signatures of positive selection on both the ZAL2 and ZAL2^m chromosomes. We predicted that positive selection on cis-regulatory alleles on ZAL2^m would, on average, increase the expression of the ZAL2^m alleles and that positive selection on ZAL2 would increase the expression of ZAL2. We therefore asked whether genes located in regions with evidence for a selective sweep on either ZAL2 or ZAL2^m (as indicated by the H-statistics for 50 kb windows) were also likely to exhibit bias in gene expression in the brain. For the ZAL2^m chromosome, we found that the ZAL2^m H-statistic was a significant, positive predictor of bias in brain gene expression toward the ZAL2^m allele, controlling for sequencing batch and brain region (X²(2)=40.17, p=1.893 × 10^–9; t=5.114, p=3.24 × 10^–7) (see Materials and methods) (Figure 4c). Likewise, we found that genes that exhibited evidence of positive selection on ZAL2 showed allelic bias toward the ZAL2 allele (X²(2)=24.231, p=5.475 × 10^–6; t=–3.151, p=0.00164) (Figure 4d). Taken together, these results suggest that gene expression on ZAL2 and ZAL2^m is driven by selection for cis-regulatory alleles that benefit each morph.

Our analyses uncovered many candidate genes that may contribute to the phenotypic differences between the tan- and white-striped morphs. To identify candidates, we performed gene ontology analysis on the set of genes that were either (1) within a region that exhibited evidence of positive selection on either ZAL2 (n=109 genes) or ZAL2^m (n=68 genes) (See Supplementary file 7), (2) exhibited allelic bias in six or more tissues (of the eight total brain tissue analyses) (n=249 genes), or (3) exhibited consistent allelic bias across all samples from either one brain region, sex, or age group (i.e. tissue-, sex-, or age-specific expression) (See Supplementary file 7, n=117 genes). The rationale for the inclusion of this final group is that genes that show a tissue-specific pattern of allelic bias are strong candidates for genes with adaptive significance, as changes in gene regulation have long been shown to play a role in adaptive evolution (Wray, 2007).

Using this set of 409 unique genes, we found the strongest enrichment for genes in the categories ‘behavior’ (p_FDR = 0.0252) and ‘metabolic process’ (p_FDR = 0.0252). The behavior category included genes such as GRIK2 (glutamate receptor ionotropic, kainate 2), GRM1 (metabotropic glutamate receptor 1), NRXN1 (Neurexin-1), OPRM1 (µ-opioid receptor), HTR1B (5-hydroxytryptamine receptor 1B, i.e. serotonin receptor), MEIS1 (Meis homeobox 1), and PAK5 (serine/threonine-protein kinase PAK 5). Additionally, there was enrichment for several gene ontology categories related to metabolism, with 205/409 (50.1%) of these candidate genes involved in the broad category ‘metabolic processes’ (p_FDR = 0.0210). We also found significant enrichment in the category ‘regulation of multicellular organismal process’ (p_FDR = 0.0342). This category included several additional interesting candidate genes, including ESR1 (estrogen receptor alpha), NR2E1 (nuclear receptor subfamily 2 group E member 1), HCRTR2 (orexin receptor type 2), SYNDIG1 (synapse differentiation-inducing gene 1), DISC1 (disrupted in schizophrenia 1), and ESRRG (estrogen-related receptor gamma).

Allelic bias toward the ZAL2 allele could result from decreased expression of the ZAL2^m allele due to degeneration. Allelic bias in the other direction, toward ZAL2^m, is more consistent with adaptation (Martinez-Ruiz et al., 2020). We therefore examined relationships between allelic bias and positive selection. We found that 97/359 (27.8%) genes exhibiting a signal of allelic bias showed a bias toward the ZAL2^m allele. Four of those 97 genes also overlapped regions exhibiting a significant H-statistic on ZAL2^m: PDSS2 (all trans-polyprenyl-diphosphate synthase), XRN2 (5'–3' exoribonuclease 2), RPS27A (Ubiquitin-40S ribosomal protein S27a), and TCTE1 (T-complex-associated-testis-expressed 1).This result implicates these four genes in adaptive ZAL2^m phenotypes, providing novel candidate genes for further exploration.

Looking at our results from a different perspective, of the 409 genes that exhibited evidence either of allelic bias or positive selection (by the H-statistic), 55 exhibited evidences of both. These genes are especially interesting candidates for mediating phenotypic differences between the morphs. Three of these genes were located in areas of elevated H-statistic on both the ZAL2 and ZAL2^m chromosomes: PDSS2, ATP6V1C2 (ATPase H+Transporting V1 Subunit C2), and GHRL1 (Grainyhead Like Transcription Factor 1). Both ATP6V1C2 and GHRL1 were in the shared region highlighted in blue in Figure 4—figure supplement 2, suggesting that this region may be an especially important evolutionary hotspot in the divergence between the morphs. Overall, our analysis revealed a number of candidate genes for which positive selection appears to have affected the regulation of genes on both chromosomes.

We performed whole genome sequencing of 63 samples (WS: N=49, TS: N=13, and SW: N=1), which included 28 birds captured during the breeding season near Argyle, Maine (ME), USA, 25 birds captured in Atlanta, Georgia (GA), USA during fall migration (November and December), and 10 postmortem samples opportunistically collected from birds who died from building strikes in Chicago, Illinois (IL), USA (see Balakrishnan et al., 2014) during the spring or fall migrations. Sequencing reads from these samples, as well as reads from three previously described samples, were mapped to the reference genome assembly (GCF_000385455.1) using Bowtie2 (ver. 2.3.5) with the very-sensitive-local option. Possible PCR duplicates were removed using Picard tools (ver. 2.19). SNP and INDEL calling were conducted using GATK Haplotypecaller (ver. 4.1.2) with the ERC GVCF option and joint genotyping of all samples were performed with Genotype GVCF option. We filtered out SNPs with any missing information, MAF <0.05, meanDP <5, or meanDP >80. Raw sequencing reads are available on NCBI SRA (BioProject PRJNA818012).

To complement the available short-read sequencing data from the female super-white bird (Sun et al., 2018), we generated additional paired-end reads from the same individual (150 ×; insert size of 300 bp and 500 bp) to improve the assembly quality. Genomic DNA was extracted from a 200 mg liver sample from the super-white bird using the Qiagen DNEasy Blood and Tissue DNA kit. Additionally, mate-pair libraries of different insert sizes (insert size of 1 kb, 4–7 kb, 7–10 kb, and 10–15 kb) were prepared and sequenced by the Brigham Young University Genome Sequencing Center. Raw sequencing reads are available on NCBI SRA (BioProject PRJNA818012).

Using these data, we constructed a whole genome de novo assembly of Z. albicollis, including the ZAL2^m chromosome. Paired-end sequencing reads were trimmed by Trimmomatic v.0.32 (Bolger et al., 2014), and error correction of the trimmed sequencing reads was conducted by Lighter v.1.1.2 (Song et al., 2014). Initial contig assembly and scaffolding was conducted by ABySS v.2.1.5 (Jackman et al., 2017). We used Gapcloser (Xie et al., 2014) to fill the gaps emerging from scaffolding process. The total assembly size was 1058 Mbp (N50 length of 3.1 Mbp), consisting of 160 scaffolds (111.76 Mbp) belonging to the ZAL2^m chromosome with the longest scaffold length of 5.1 Mbp (Supplementary file 3).

To discriminate scaffolds that originate from the ZAL2 vs the ZAL2^m chromosome, we mapped the super-white assembly against the genome of the House sparrow (Passer domesticus), which is the most closely related species with chromosome level assembly using LASTZ v1.03. Scaffolds uniquely mapped to the homologous House sparrow chromosomes with >30% coverage and >85% identity were retained. In the case of multi-mapping scaffolds, we used more stringent criteria with >70% coverage and >85% identity. We conducted the same procedure using the tan-striped reference assembly. The list of matched scaffolds was highly consistent with our previous study, with the addition of several new scaffolds (n=15, sum of scaffold lengths = 129.1 kb). To distinguish sequences inside vs outside of the rearranged (i.e. inverted) regions, we computed the average frequency of heterozygous SNPs in sliding windows of 25 kb size with 1 kb step.

Utilizing the large amount of newly generated whole genome sequence data, here we identified putatively fixed genetic differences between the ZAL2 and ZAL2^m chromosomes. Briefly, we identified positions at which the genomes of the tan-striped (ZAL2/2) and super-white (ZAL2^m/2^m) birds are homozygous for different alleles while the white-striped (ZAL2/2^m) birds are heterozygous. Note that the probability of this allelic pattern occurring by random chance is 2.6x10^–23 given the sample size, according to a binomial test. Following this procedure, we obtained a total of 931,424 SNPs and 97,375 insertions and deletions (InDels) between ZAL2 and ZAL2^m chromosomes, increasing the number of putatively fixed SNPs between the two chromosomes beyond previous publications (Tuttle et al., 2016; Sun et al., 2018). Variant call format files of fixed differences (both SNPs and InDels) are available at DOI: 10.6084 /m9.figshare.19395146. As we expected, a vast majority (N=930,588; 99.91%) of the fixed differences reside in the scaffolds that we previously predicted to be inside the rearrangement (Sun et al., 2018). The remaining fixed differences were found in scaffolds that were either too short or that mapped ambiguously to multiple chromosomes. FastEprr was used to estimate the recombination rate for non-overlapping 50 kbp sliding window with default setting (Gao et al., 2016).

We performed read-based haplotype phasing of the sequencing data using reads from white-striped individuals (ZAL2/2^m heterozygotes). Briefly, using the 931,424 single nucleotide fixed differences between ZAL2/2^m, we assigned sequence reads from the heterozygous birds to the corresponding chromosome of origin (i.e. either ZAL2 or ZAL2^m) if the paired-end reads were mapped to a region that overlapped at least one fixed difference. Reads from white-striped birds were mapped to N-masked (masking the putative fixed differences between ZAL2 and ZAL2^m) reference assembly to avoid mapping bias and assigned to their chromosome of origin using SNPsplit v.0.3.2 (Krueger and Andrews, 2016). Because the two supergene variants differ in the amount of genetic diversity they carry, and because read mapping algorithms are sensitive to the number of differences between the reference sequence and the read, our approach risks a slight mapping bias. We believe that this slight bias is unlikely to have fundamentally changed our results.

On an average, 8.5% of all reads from white-striped birds were assigned to either the ZAL2 or ZAL2^m chromosomes. In comparison, the size of the chromosomal inversion is estimated to be approximately 9.5% of the total genome (Thomas et al., 2008). The ZAL2 and ZAL2^m assigned reads were extracted respectively using Bedtools bamtofastq. To call variants for both ZAL2 and ZAL2^m chromosomes, the reads were remapped to the reference genome assembly using Bowtie2 v. 2.3.5. Variant calling was conducted using GATK Haplotypecaller v.4.1.2 with the ERC GVCF option. Vcftools was used to filter out SNPs with any missing information, MAF <0.05, meanDP <5, or meanDP >80. Accession information for all raw RNA sequencing data used is available in Table 1.

Estimates for nucleotide diversity (π), between population divergence (d_XY), density of fixed differences (d_f), between population difference in allele frequency (F_ST), and Tajima’s D were computed in non-overlapping sliding windows sized 10 kb, 25 kb, and 50 kb, respectively. If a data set has substantial missing points, the estimation of nucleotide diversity is often biased (Korunes and Samuk, 2021). Thus, calculation of nucleotide diversity and d_XY from the variant call format (VCF) file can be over-estimated since missing sites are not distinguished from invariant (monomorphic) positions in variants-only VCFs. To account for potential inflation of population summary statistics, we considered an average breadth of coverage across samples for each window when we computed nucleotide diversity, d_XY, and d_f. Nucleotide diversity of protein coding sequence was computed using SNPGenie (Nelson et al., 2015).

For phylogenetic reconstruction of ZAL2^m outlier windows, we used RaxML v.8.0.2 for 200 kb windows. The targeted window was selected based on the Tajima’s D values and haplogroup assignments were determined by manually inspecting genotype plots. For the control region, we concatenated 10 randomly selected regions of 20 kb windows (200 kb). The evolutionary model was set to GTRGAMMA with acquisition bias correction using conditional likelihood method (-m ASC_GTRGAMMA –asc-corr=lewis). The medium ground finch (Geospiza fortis) was used as an outgroup species. We conducted bootstrap with 100 replicates.

To examine whether introgression from other species could account for the ZAL2^m outlier window, we computed the D-statistic (Martin et al., 2015) using as ingroup genomes the ZAL2 chromosome, the Harris' sparrow chromosome 2, and the ZAL2^m chromosome (P1, P2, and P3, respectively). We used the medium ground finch as the outgroup species (P4). The D-statistic was computed for four individual white-striped birds with high sequencing coverage.

The analysis of linkage disequilibrium (LD) decay was performed across the entire chromosomal rearrangement. We calculated pairwise r² values between variant sites using PopLDdecay (Zhang et al., 2019).

We performed H-scan (Schlamp et al., 2016; Messer Lab — Resources, 2014) to identify soft and hard sweeps, inferred by extended tracts of homozygosity, using phased haplotypes of ZAL2 and ZAL2^m. H-scan outputs were separated into 25 kb non-overlapping windows. As a representative summary statistic, we chose the maximum H-statistic value for each window. To calculate the empirical p-value, we first binned genomic windows into 50 SNP increments based on the number of SNPs. When a window included more than 100 SNPs, all the windows were merged into one bin. On the basis of the ranking of summary statistics in each bin, we calculated an empirical p-value for each window. Candidate regions of positive selection were defined as those with an empirical p-value less than 0.05.

We examined both gene expression divergence and allele-specific expression using three RNAseq data sets from white-throated sparrows (four types of tissue from four separate samples of birds, see Table 1). The first data set consisted of gene expression in male white-throated sparrows collected during the breeding season. In those birds, the samples comprised two brain regions, the Hyp and the AMV (Sun et al., 2018; Zinzow-Kramer et al., 2015), a functional homolog of the medial amygdala (Mello et al., 2019). The second data set consisted of gene expression in the same two brain regions (Hyp and AMV) in adult females collected during the breeding season as well as male and female nestlings (see Sun et al., 2021 for details). The third data set consisted of gene expression in heart and liver tissue in white-striped males collected during migration, then housed in captivity under two lighting conditions (long days and short days) and sampled at two time points during the day (ZT6 and ZT18) (Horton et al., 2019). In statistical analyses, the females and nestlings of each sex were treated as separate ‘batches’ of RNAseq such that there were five total batches: brain samples from adult males, adult females, nestling males, nestling females, and liver/heart samples from adult males. Each tissue was nested within batch to account for repeated sampling from the same individuals.

RNAseq reads were aligned to a reference genome N-masked for putative fixed differences for both chromosomal polymorphisms present in this species (i.e. ZAL2/2, ZAL3^a/3^a,) using HiSat2 (Kim et al., 2015). To examine allele-specific expression, we used SNPsplit v.0.3.2 to assign mapped reads to ZAL2 or ZAL2^m for the white-striped samples. Transcripts were quantified using StringTie (Pertea et al., 2016) and differential expression and allelic-specific expression analyses were performed using DESeq2 (Love et al., 2014). To test for differential expression between the morphs, we used the following model in DESeq2: ‘design = ~morph’. To test for allelic bias (AB), we used the size factors generated for each sample in the previous step, and then used the following model: ‘design = ~individual + allele’. To perform hypothesis testing, we used linear mixed models with tissue nested in sequencing batch as random effects using the R package lme4. To test for the effect of interest, we then performed a likelihood ratio test to compare a full model to a reduced model with the factor of interest removed using the anova function. Where applicable, we used the summary function to perform post-hoc comparisons between multiple groups.

To test for significant allelic bias in expression within each sample type, we used a Wilcoxon rank sum test to test whether the ratio of ZAL2^m to ZAL2 expression differed significantly from 0.5, applying an FDR correction for multiple testing. To test for an association between allelic bias and the number of mutations within a gene, we computed the per-base number of mutations by dividing the number of mutations by the gene length and then sorted the genes into deciles. We also computed the number of fixed mutations within 1 kB upstream of the transcription start site (TSS). We then tested whether there was an association between allelic bias and the decile rank or number of mutations within 1 kB upstream of the TSS using the following linear model: Log2 AB ~Variable + (1|Batch:Tissue). To test for overrepresentation of morph-biased genes located inside the rearrangement, we performed a two-sample test for equality of proportions for each brain sample type only. For this test, we compared the proportion of differentially expressed genes, out of the 1007 genes inside the rearranged region on ZAL2, to the proportion out of the 13,369 genes located elsewhere in the genome. For this comparison, we used the prop.test function in R to perform a two-proportions Z-test, applying an FDR correction for multiple testing. To test whether the morph bias in expression of a gene significantly predicted allelic bias for that gene, we grouped genes into three categories: those that were more highly expressed in the white-striped morph (W>T), those more highly expressed in the tan-striped morph (T>W) and those that that did not differ significantly between morphs (T=W). Using the allelic bias for each gene, we then tested whether allelic bias differed among these categories using the following linear mixed model: Log₂ AB ~MorphBias_Category + (1|Batch:Tissue), followed by Bonferroni-corrected pairwise post-hoc tests using the ‘ghlt’ function in multcomp package. Note that heart and liver tissue samples were obtained from white-striped birds; because we did not have data on differential expression by morph for these tissues, they were excluded from this analysis, as well as the tests for overrepresentation described above.

To test whether haplogroup affects gene expression, we merged the gene count matrices for the male (H1: n=7, H2: n=3) and female (H1: n=3, H2: n=3) white-striped birds that were also included in the whole genome sequencing dataset and tested for an effect of haplogroup using the following model: design ~sex + haplogroup. Only one gene in Hyp (Geranylgeranyl Diphosphate Synthase 1, GGPS1) and one gene in AMV (RUN And FYVE Domain Containing 4, RUFY4) were differentially expressed at the genome-wide level. Neither of these genes are located in the ZAL2^m outlier region and only GGPS1 is located on ZAL2. Thus, we report only unadjusted p-values for genes inside the ZAL2^m outlier region that were differentially expressed (both unadjusted and adjusted p-values for all genes in the ZAL2m outlier region are reported in Supplementary file 4).

To examine the relationship between the H-statistic and allelic bias in expression, we computed the average H-statistic for each gene. Each gene was assigned the H-statistic value of the 20 kb bin overlapping the gene (or the average of multiple 20 kb bins, if a gene overlapped two bins). We then used a multiple linear regression model to examine the relationship between the Log₂ASE and the log₂H-statistic of a gene for both ZAL2 and ZAL2^m separately, using the following model: Log2 AB ~ Log₂H-statistic + (1|Batch:Tissue). Additionally, we performed functional enrichment analyses for the genes in regions with a significant H-statistic using ToppFun with human homolog HGNC gene names (Chen et al., 2009).