Rules of amino acid convergence: Not how many, but who in avian vocal learning clades

Single amino acid variants (SAVs) may provide clues to understanding evolution of traits. A complex trait that has evolved convergently among species is vocal learning, the rare ability to imitate sounds heard and an important component of spoken-language. Here we assessed whether convergent vocal learning bird species have convergent SAVs (CSAVs) that could be associated with their specialized trait. We analyzed avian genomes and identied CSAVs in vocal learners, but also in most species combinations tested. The number of CSAVs among species was proportional to the product of the most recent common ancestor (MRCA; origin) branch lengths of the species in question, and vocal learning birds did not exceed the overall proportion in most test. However, genes with identical CSAVs (iCSAVs) in vocal learning species were uniquely enriched in ‘learning’ functions, and a subset of iCSAV genes were under positive selection and had enriched specialized regulation in vocal learning and their adjacent brain subdivisions. Several top candidate genes converge on the cAMP signaling pathway, including DRD1B and PRKAR2B. Our ndings suggest a complex mechanism of amino acid convergences and specialized gene regulation upon which selection acts for specialized convergent traits.


Introduction
Single amino acid variants (SAV) are one of the potential drivers of evolution for various traits. For example, the Forkhead box P2 (FOXP2) transcription factor has two well-known human-speci c SAVs which might have been positively selected for learning behavior related to language 1,2 . Mutant mice humanized for the two SAVs of FOXP2 showed more advanced learning abilities 3 and alterations of cortico-basal ganglia circuits 4,5,6 , which play critical roles in spoken-language 7 ; and mice containing a heterozygous missense mutation that causes speech syllable apraxia in humans also showed syllable sequencing de cits 8,9 . A crucial component of spoken-language is vocal learning, the ability to produce vocalizations through imitation, and is a convergent trait observed in only a few animals, including songbirds, parrots, and hummingbirds among birds, and bats, dolphins/whales, seals, elephants, and humans among mammals 7,10−13 . Both vocal learners and vocal non-learners share an auditory pathway that controls auditory learning, while only the vocal learning birds and humans have been found to share a specialized convergent forebrain pathway that controls vocal learning [11][12][13][14] . Supporting the hypothesis of independent origins of vocal learning, the recent genome-scale phylogenetic tree reported by the Avian Phylogenomics Project showed that the three avian vocal-learner lineages are indeed not monophyletic 15 . Even though songbirds and parrots are relatively closely related, the closest lineage to songbirds 15 , sub-oscines, is a vocal non-learning lineage In the rst genome-scale analyses for vocal learning in the avian lineage, genes with positively selected changes in zebra nch (a songbird) compared to chicken were identi ed 16 . Some of the positively selected genes were in ion channels, which are known to control neurological function, behavior and disease 16 . However, the comparison was made narrowly between only one vocal learner (zebra nch) with one vocal non-learner (chicken), which is a very distant 15 relative, like a marsupial is to a placental mammal.
The big bang of draft genome sequences of the Avian Phylogenomics Project, consisting of 48 avian species that represent 34 orders of birds 17 , provided an unprecedented opportunity to investigate genetic features speci c to polyphyletic vocal learning clades. These studies found convergent brain gene expression specializations in vocal-learning birds and human 11,12,14 . We also found mutually exclusive convergent amino acid substitutions unique to vocal learners, using a novel method (Target-speci c Amino Acid Substitutions [TAAS] analysis) 17 . However, the study overlooked several viewpoints reported around that time for molecular convergences 18,19 ; it did not separately test for identical (i.e. true convergence) versus different (divergent) convergent amino acid substitutions; it did not test for preponderance of convergences and divergences over random control sets of species; and it did not test for possible in uences of close phylogenetic relationships.
Here, we investigated basic rules of molecular convergent evolution and their functions in various combinations of avian species, including vocal learners. We improved methods for comparative genomics of convergent and divergent amino acid variants among species, and tested whether vocal learning birds have more molecular convergences or divergences than control sets. We discovered phylogenetic features associated with the number of convergent and divergent substitutions among species beyond those of previous studies 19,20 , and the underlying nucleotide variant changes associated with these amino acid substitutions. We found a preponderance of higher changes in vocal learners when considering the most recent common ancestor branch lengths, and we did found an enrichment in learning functions, positive selection, and specialized gene expression in vocal learning brain regions and the subdivisions they reside for a subset of genes with amino acid convergences of vocal learning birds.

Results
Convergent amino acid variants speci c to avian vocal learners. Among the genomes of 48 avian species spanning most orders 17 , we compared 6 species from the three vocal learning orders or suborders (songbirds: zebra nch, medium ground nch, and American crow; parrots: budgerigar, and kea; and hummingbirds: Anna's hummingbird) with 41 vocal non-learning birds (Fig. 1a, Fig. S1) 15 . Ri eman, a close relative of songbirds and suboscines, was initially excluded because of the uncertainty of its vocal learning ability, although assumed to be a vocal non-learner 15 . These were short-read genome assemblies and do not have all repetitive and GC-rich regions assembled, but do have most protein coding genes assembled 21  were used, instead of 'convergent' and 'divergent' or 'parallel' 20,24 , as the term 'convergent' can be ambiguously interpreted as pertaining to only identical CSAVs or both identical and divergent CSAVs. Out of 154 AVL-CSAV sites, 25 sites (16%) were identical CSAVs (AVL-iCSAV) and 129 sites (84%) were different CSAVs (AVL-dCSAV) among vocal learners relative to vocal non-learners (Supplementary Data 1). For example, the 253rd site of B3GNT2 was a Type 1 (AVL-iCSAV) site with asparagine (N) observed in all avian vocal learning species and histidine (H) in all vocal non-learning species; while the 217th site of SMRC8 was a Type 3 (AVL-dCSAV) site with glutamine, valine, and leucine (Q, V, and L) observed in avian vocal learners and isoleucine (I) in all vocal non-learners (Fig. 1c).
Avian vocal learners did not show a preponderance of identical convergent single amino acid variants. We next tested whether avian vocal learners have a higher frequency of convergent substitutions relative to control sets of species. Considering the polyphyletic relationship of the 6 vocal learning species examined, we designed 2 types of clade-speci c control sets: 1) Random controls consisting of 1,000 different species combinations with 6 randomly selected target species from 3 independent lineages without considering any traits; 2) Core controls consisting of all possible 59 sets given the phylogeny with 6 target species having at least 2 vocal learning clades and 1 non-learning clade, and 2 sets with different numbers of target species with other convergent traits (6 birds of prey from 4 clades and 15 waterbirds from 4 clades: Supplementary Note 2; Supplementary Table 2). We conducted the CSAV analysis for this total of 1,056 random and core control species combinations, and identi ed convergent identical (Ctrl-iCSAV) and different (Ctrl-dCSAV) amino acid substitutions in all control sets.
As an extension to the previous studies on convergent evolution in reptile and mammalian lineages 18-20 that tested pair-wise combinations of two species (Supplementary Note 3), we found strong correlations between iCSAVs and dCSAVs tested in higher dimensional combinations of species (Fig. 1d,e). Although higher than the expectation (20.48 and 23.04 respectively) according to the regression with random and core control sets, the number of convergent substitutions in vocal learning birds was not an outlier (adjusted p > 0.05, Bonferroni Outlier Test 25 ) from the trend observed in the control sets (Fig. 1d,e). Several outliers did exist among the control sets, with the highest residual being 32.46 in one of the random control sets (4 Passeriformes, budgerigar, and falcon), and 17.61 in a core control set (3 songbirds, Anna's hummingbird, and 2 land fowls; Fig. 1d,e). These species combinations of 2 control sets with the highest residuals, however, do not share any known convergent traits as far as we are aware. In the groups of species with other known convergent traits (birds of prey and waterbirds), both iCSAVs and dCSAVs were less frequently observed (Fig. 1e). These ndings support that identical convergent single amino acid substitutions are widespread, and their numbers vary in different species combinations that does not appear to readily correlate with known convergent traits.
Amino acid convergences are associated with product of origin branch lengths. We sought a measure of molecular convergence that controls for phylogenetic relationships. According to previous studies on mammalian or drosophila genomes 24 , and vertebrate mitochondrial genomes 19 , fewer convergent substitutions are expected with the greater phylogenetic tree branch distance. However, the correlations found in these studies showed high variations, which makes it di cult to identify the outliers. Here, we took into consideration additional phylogenetic features, including the relationship between the convergent variants versus the most recent common ancestor [MRCA] branch of each clade (origin branches), the terminal branches, and the nodes of the tree (Fig. 2, top row; Supplementary Note 4; Supplementary Fig. 1). We observed strong and signi cant correlations between CSAVs and the product of MRCA branch length lengths (POB) for both random control ( Fig. 1f and Fig. 2) and core control sets ( Fig. 1g and Supplementary Fig. 4). The correlation was also observed for both iCSAV and dCSAV (Fig. 2a). Much weaker correlations of CSAVs were observed with the product of terminal branches (PTB), distances among terminal branches (DTB) and terminal nodes (DTN; Fig. 2b-d). In contrast to the iCSAV versus dCSAV correlation analyses, in the POB phylogenetic analyses the number of CSAVs in avian vocal learners was a signi cant outlier relative to random control species sets (adjusted p < 1.41e-08; Fig. 2a). The outlier position of CSAVs was driven mostly by the dCSAVs (adjusted p < 1.64e-07), but also combined trend for the iCSAVs in vocal learners (Fig. 2a). Two other species combinations were outliers, but different from the iCSAV versus dCSAV analyses: (4 Passeriformes, chicken, and duck) and (3 songbirds, Anna's hummingbird, carmine bee-eater, and downy woodpecker). However, the avian vocal learners were not a signi cant outlier relative to core control sets only ( Supplementary Fig. 4). These ndings suggest that POB value can largely explain convergent variants at the amino acid level, where the longer their ancestral branch lengths the greater frequencies of convergence amino acid variants, and that vocal learning birds are somewhat different in this regard relative to other species combinations.
Convergent amino acid substitutions can arise from complex nucleotide variants at multiple sites. To investigate what types of codon and nucleotide variants can cause CSAVs, we modi ed the CSAV method to detect convergent single codon variants (CSCVs) made up of 3-nucleotides and convergent single nucleotide variants (CSNVs) in those codons ( Fig. 3a,b; Supplementary Note 5). Similar to the CSAV analysis, the codon and nucleotide variants were classi ed into 2 identical types and 2 different types ( Fig. 3a,b). A CSAV can theoretically originate from identical single nucleotide variants at a homologous codon position (CSNVs) or complex multiple nucleotide variants (CMNV) at different codon positions (Fig. 3c). However, some CSNVs can also give rise to no change in the amino acid, namely synonymous substitutions (Fig. 3d). We checked for overlaps among CSAVs, CSCVs, and CSNVs to trace the source of the convergent amino acid substitutions at the codon and nucleotide levels.  (Fig. 3e). An example of a nonsynonymous CSAV explained by a CSNV is in the 253rd codon site of B3GNT2, where all vocal learners had the same convergent nucleotide (A), codon (AAT), and amino acid (N, Asparagine) sequence relative to all vocal non-learners (e.g. C; CAT or CAC; and H, Histidine; Fig. 3f). An example of a nonsynonymous CSAVs explained by a CMNV is the 482nd site of LRRN4, where all of vocal learners showed identical amino acid convergence (AVL-iCSAV) to Histidine (H), while their codons consist of CAC or CAT for vocal learners and TTT, TAT, TAC, CGC, and TCG for vocal non-learners (Fig. 3f).
In the random and core control species sets, although we found different total numbers of CSCVs and their corresponding CSAVs and CSNVs, their relative proportions (%) were similar to each other and to that of vocal learners (Fig. 3e); about 1/3 of CSAVs of random and core control sets originated from CSNVs at each homologous nucleotide site, while 2/3 originated from multiple nucleotide changes at different nucleotide sites (Fig. 3e). These ndings suggest that amino acid convergences originate not only from identical single nucleotide substitutions at each homologous site but also from complex nucleotide substitutions at multiple sites within a codon.
Convergent variants at all levels are best explained by the product of MRCA branch lengths. Next, we performed correlation tests between nine types of sequence variants (three types of convergences [all, identical, and different] at three levels [amino acid, codon, and nucleotide]) and four types of phylogenetic features (POB, PTB, DTB, and DTN; Supplementary Note 6). As expected, all nine types of convergent variants (CSAVs, iCSAVs, dCSAVs, CSCVs, iCSCVs, dCSCVs, CSNVs, iCSNVs, and dCSNVs) were highly correlated with each other, in both random control (Fig. 4) or core control species sets ( Supplementary  Fig. 5). For the phylogenetic features, the POB showed the strongest correlation with all variant types, where the others (DTB, DTN, and PTB) were either weaker or not correlated at all (Fig. 4). Correlations were overall weaker in the core control sets of species ( Fig. 4 vs Supplementary Figs. 5,6), presumably due to a smaller number of species combinations than the random control sets. Unlike the CSAVs, the numbers of vocal learner-speci c variants did not exceed that of control sets for the other variant types (Fig. 4).
Post hoc analyses of vocal learner substitutions in Ri eman. Ri eman and more broadly the New Zealand Wrens, a close relative of vocal learning songbirds, have been assumed to be a vocal non-learner 15 . Although ri eman was excluded from the initial CSAV search, we can ask whether its sequences match those of vocal non-learners as assumed. We applied principal component analysis (PCA) and phylogenetic analysis for the 154 AVL-CSAV sites and the subset of 25 AVL-iCSAV sites (Supplementary Note 7). PC1 and PC2 accounted for 53% and 66% of the total variances of the AVL-CSAV sites and AVL-iCSAV sites, respectively (Fig. 5a,b). The vocal learning birds clustered away from the vocal non-learning group as expected (Fig. 5a,b). For the AVL-CSAV sites, ri eman clustered with the vocal non-learners (Fig. 5a). For the AVL-iCSAV subset, ri eman was separate from the two groups, but was still closer to vocal non-learners (Fig. 5b). Phylogenetic analysis of these AVL-CSAVs was consistent with the PCA results, where instead of branching with its closest relatives, the songbirds, ri eman was on a branch outside and next to the vocal learners (Fig. 5c,d). These results support the assumption that ri eman is a vocal non-learner. The results also suggest that some of the convergent sites and genes could be associated with their convergent trait.
Functions of convergent genes. To investigate the biological functions of genes with convergent variants in vocal learners and in control sets, gene ontology (GO) analysis was performed on 9,513 (1,057 species sets * 9 variants types) gene lists with 1 or more convergent variants (Supplementary Note 8). Among them, at least one signi cant (adjusted p < 0.05) GO term was found for 1,038 gene lists (10.9%). We further found a positive correlation between the number of signi cant GO terms and the number of genes with convergent variants in each list (Fig. 6a); however, we found negative correlation between the average adjusted p value of signi cant GO terms and the number of genes with convergent variants (Fig. 6b).
In vocal learning birds, we did not nd any GO enrichment for the total AVL-CSAV gene list. However, the AVL-iCSAVs gene list subset was signi cantly enriched for 'learning' (GO:0007612, adjusted p = 0.042). Four genes were responsible for this enrichment (DRD1B [also known as DRD5], LRRN4, PRKAR2B, and TANC1; Fig. 6c). The identical amino acid convergences of DRD1B, PRKAR2B, and TANC1 also showed identical codon convergences (AVL-iCSCVs) in all vocal learners, while that of LRRN4 showed different synonymous codon changes (AVL-dCSCV; Fig. 6d). Of the 1,056 control species combinations, only one control gene list (a Ctrl-dCSCV gene list) showed signi cant enrichment for 'associative learning' (GO:0007612, adjusted p = 0.035); the associated set of species (Fig. 6e) did not include any vocal learners, but a convergent variant in LRRN4 contributed to this functional enrichment (Fig. 6f). The ndings indicate that convergent genes in vocal learners do function in the brain and for learning.
Convergent amino acid sites are xed and positively selected. We checked for additional evidence of whether the CSAVs in vocal learners are reliable, by checking for sequence assembly artifacts and SNPs within species. We used the dbSNP database of a representative vocal learner (zebra nch; n = 1,257 samples; build 139) and a vocal non-learner (chicken; n = 9,586 sample, build 145; Supplementary Note 9). At the 154 AVL-CSAV sites, zebra nches showed complete xation without any nonsynonymous polymorphisms. However, one missense SNP was found in the chicken OTOA gene (c.2581A > G, p.Thr861Ala) resulting in an amino acid change identical to that of vocal learners ( Supplementary Fig. 7). We also validated xation of the convergent substitution in DRD1B by PCR of genomic DNA and sequencing of 3 male and 3 female zebra nches and chickens ( Supplementary Fig. 8). These ndings indicate that the vast majority (99.4%) of the convergent amino acids we identi ed in vocal learners are the result of true species-speci c variants.
In addition, to consider positive selection on the convergent sites, we performed dN/dS analysis with the branch-site model for CSAV genes in the avian vocal learning species and their closest control set (songbirds, parrots, and swifts; Supplementary  Fig. 9). We found that under half of the iCSAV sites showed signs of positive selection (Likelihood ratio value (D) > 0, ω 2 .foreground > 1) in the vocal learning birds (10 of 25 genes, 40%) and the closest control set (12 of 26 genes, 46%), with 12% (3) and 23% (6) being statistically signi cant, respectively (adjusted p < 0.05, posterior probability > 0.5). These ndings suggest that a subset of iCSAV genes in different species combinations have been positively selected whether the species share a convergent trait or not, and it does not seem to need a greater number of positively selected sites for the avian vocal learning ability.
Genes with convergent identical changes are specialized in vocal learning and the associated brain subdivisions. We next tested if the AVL-CSAV genes are expressed in vocal learning brain circuits. We analyzed 8 brain transcriptome data sets, which include genes that show singing-regulated expression (increased or decreased) in song learning nuclei of songbirds 26 (Area X, HVC, LMAN, and RA), differential expression (increased or decreased) in song nuclei compared to their surrounding non-vocal motor brain regions (NUC vs SUR), one song nucleus compared among the other song nuclei (NUC vs other NUCs), and the surrounding regions of each song nucleus compared to the other surrounding regions (SUR vs other SURs), from independent experimental data sets (DEG_2014: microarray method in 2014 14,26 , DEG_2019: micro-dissected RNA sequencing in 2019 27 , and DEG_2020: laser capture microscope with RNA sequencing in 2020 27 ; Supplementary Note 11, Supplementary  Fig. 10, and Supplementary Data 2). Relative to the average of all genes (8,295 avian orthologous genes) measured, we found no enrichment of AVL-CSAV genes (0-13%) among the singing regulated genes or song nuclei specialized genes relative to the surrounding brain regions, whether positively selected or not (Fig. 7a). However, in two independent transcriptome experiments, we found 60 to 100% of the positively selected AVL-iCSAV (AVL-PS-iCSAV) genes were enriched among the differentially expressed genes in one song nucleus relative to the others, and some of those were also enriched to a lesser degree in the adjacent surrounding brain subdivision relative to the others (Fig. 7a). These enrichments were not found for the AVL-dCSAV genes, not for any positively selected gene set in the closest related control set (Fig. 7a), nor for all control sets even before positive selection analyses (not shown). Out of 4 song nuclei, Area X involved in song learning showed the highest number of differentially regulated genes out of singleton orthologous genes of birds or genes with amino acid convergences speci c to avian vocal learners in comparisons among a song nucleus and the other song nuclei (NUC vs other NUCs, DEG_2020; Supplementary Data 2, Fig. 7b).
Out of 25 AVL-iCSAV genes, a total of 8 genes (32%) had positively selected sites in vocal learners and differential expression speci c to a song nucleus and surrounding brain subdivision: B3GNT2, DRD1B, FNDC1, HMGXB3, MTFR1, PIK3R4, PRKAR2B, and SMPD3 (Table 1). These include two genes, DRD1B and PRKAR2B, revealed in the GO analyses for learning functions (Fig. 6c, d). Further DRD1B has specialized up-regulation speci c to adult Area X compared to its surrounding striatum (Table 1 and Fig. 7c) [28][29][30]  Convergent evolution between vocal learning birds and human in the same genes. Lastly, we investigated amino acid convergences between vocal learning birds and humans. We performed SAV analysis for 18,565 singleton orthologous genes in primates, and discovered 11,317 genes (61.0%) contained 126,087 human-speci c amino acid substitutions relative to other primates (Human-SAV; Fig. 8; Supplementary Note 12), including the well-known p.Thr303Asn and p.Asn325Ser amino acid substitutions in FOXP2 speci c to humans ( Supplementary Fig. 11). Since this is over 60% the coding genes in humans, it is unlikely that the majority of these substitutions are associated with differences between humans and other primates. Narrowing them down to our candidate gene list to those found in avian vocal learning species, out of 141 AVL-CSAV genes, 125 singleton orthologous genes were identi ed in human (Supplementary Data 3). Among the 125 genes, 85.6% (107 genes; well over 61%) contained 396 Human-SAV sites compared to other non-human primates (Fig. 8). For random and core control sets of birds, on average 64.8% and 77.9% included Human-SAVs, respectively (Fig. 8). Despite this lower average % in control sets, several random and core control species sets still showed higher percentages with Human-SAV (Fig. 8). These ndings are consistent with analyses among birds, that although vocal learning birds and humans have convergences in the same genes, the total number is not higher than some control sets that may not have trait convergences with humans.
We scanned for overlaps of the amino acid substitutions in vocal learning birds and in humans. Out of 107 bird-primate orthologous genes with 120 AVL-CSAV sites among birds and 396 Human-SAV sites among primates, only SETD4 had an overlapping site, in the 103 rd position of the primate and avian peptide alignment (Fig. 9a). However, the amino acid Arginine (R) in one avian vocal learning bird (Budgerigar) was identical to that of non-human primates, as opposed to Glycine (G) in human; vocal non-learning birds had a Glutamine (Q; Fig. 9a). In contrast to this result, we found that of the 107 genes with avian vocal learner convergences and with human-speci c substitutions among primates, 79 showed 111 conserved protein domains containing at least one of AVL-CSAV and Human-SAV site, respectively (Supplementary Note 13). This included four genes with identical convergent substitutions in vocal learning birds, enrichment for learning in the GO analyses, with positive selection and differential expression between song learning nuclei (Figs. 9; DRD1B, LRRN4, PRKARB, and TANC1). DRD1B showed one AVL-iCSAV site and one Human-SAV site located in the c-terminal intracellular region that interact with DRD2 31 , a gene also associated with a spoken language disorder 32 (Fig. 9b). These convergent variants speci c to vocal learning birds and humans, but at different sites in the signaling domain of the DRD1B receptor along with its specialized expression serve as candidates for changing dopamine receptor function in human speech and songbird vocal learning circuits 28,33 . These ndings indicate that in more distantly related species, convergence in genes associated with convergent traits may occur with lineage-speci c convergences occurring in different amino acid sites, but in the same protein domain.

Discussion
As the primary structure of proteins, amino acid substitutions can contribute to various traits including human language [34][35][36] .
Our ndings give us new insights into convergent evolution of amino acid substitutions, and possible in uence on convergent traits. We discovered correlations between the frequency of amino acid convergence with the product of ancestral branch lengths. These amino acid convergences originate from identical but also complex nucleotide substitutions in each codon across species. Our nding that around 60% of convergent amino acid variants (e.g. codons) originate from more complex patterns of convergent nucleotide variants amongst species suggest that identical protein coding convergence due to different nucleotide convergence is more common. Remarkably, although vocal learners did not have a higher preponderance of identical amino acid convergences above background levels, we nd that a subset of the sites and the associated genes have been positively selected upon and have specialized expression between different brain subdivisions. To explain our ndings, we propose a hypothesis of selection on a background of convergent substitutions for convergent traits.
Our phylogenetic analyses suggest that the background level and rate of convergent substitutions is a function of the product of substitutions rates along the MRCA branches of each clade. Only in the MRCA analyses did we nd correlations between the phylogenetic feature, POB, and convergent substitutions in species from multiple independent lineages, where other analyses and studies have attempted and failed to nd 24,37−39 . Our positive selection, functional association, and gene expression analyses suggest that selection occurs on some of these convergent substitutions to contribute to evolving novel, convergent traits, in our case vocal learning. Similarly, we predict that some of the other convergent substitutions we found in birds of prey or waterbirds will be associated functionally with those traits. According to this hypothesis, it is not about how many genes show convergence, but which speci c genes (e.g. who) show convergence, as the most important factor to consider.
Our ndings of an association between convergent identical amino acid changes in vocal learners and specialized expression speci c to a vocal learning nucleus and the associated surrounding brain subdivision was both intriguing and perplexing to us. If anything, we were testing a more logical outcome of amino acid convergence in genes that show singing-regulated gene expression or specialized expression in vocal learning brain regions relative to the surrounding brain subdivisions. But the unexpected relationship with vocal learning and brain subdivision specialization we believe is real, as we replicated multiple times, and there is 100% overlap of the most signi cantly selected genes in vocal learners and brain subdivision gene expression speci city. These ndings suggest that there is selection of protein coding sequence changes in vocal learners for a set of genes that have brain region speci c expression, particularly in the striatum. Further, one of the striatum-speci c genes, DRD1B, also had specialized up-regulation in Area X of the striatum, suggesting further regulatory genomic region changes. Often coding and regulatory genomic sources of trait evolution are pitted against in each other as alternatives 40 , but our ndings suggest that they could synergistically in uence evolution of each other. Studies in our group are underway to nd the regulatory regions of these genes, and to determine what non-coding sequence changes are the cause of their specialized regulation. Based on convergent variants, positive selection, and differential gene expression in song nuclei and the surrounding regions, we suggest 8 key candidate genes for associations with the vocal learning ability in birds (Table 1).
When searching for convergent substitutions among species, we believe our approach of multi-wise comparisons and the product of the MRCA branch lengths (POB) maybe more informative than past approaches. Previous studies found correlations between convergent identical and different substitutions (previously called convergent and divergent substitutions) between pairs of species among reptiles 18 or mammals 20 . We further nd that such a relationship exists in higher dimensional combinations of species, but this type of analyses does not control for species relationships. Several other studies found that the rate or number of convergent identical and different substitutions decreases with increasing genetic distance between two lineages 19,24 . Our ndings with the product of the MRCA branch lengths in convergent polyphyletic clades suggest that the deeper in time their common ancestor, the more likely to nd higher proportions of detectable convergences at the amino acid, codon, and nucleotide levels. These analyses provided a new null hypothesis of convergent evolution according to phylogeny.
The biological function of genes with amino acid convergences speci c to avian vocal learners gave us new insights into the potential molecular mechanisms of vocal learning. The four convergent learning-related genes with AVL-iCSAV sites includes the DRD1B dopamine receptor associated with learning 41 , and LRRN4 that affects long lasting memory 29,42 , fundamental traits of vocal learning 43 . TANC1 regulates dendritic spines and spatial memory 44 . At the mechanistic level, DRD1B, through its G-protein, regulates activity of adenylyl cyclase's synthesis of cAMP in the cell membrane 30,45 ; PRKAR2B, or Protein Kinase cAMP-dependent Type II Regulatory Subunit Beta, is an enzyme that activates cAMP-dependent protein kinase (PKA) inside the cell 46 . Additionally, one of the most well-known genes that PKA inhibits is involved in learning, including vocal learning 47 , namely the cAMP response element binding protein (CREB1), a transcription factor responsive to cAMP signaling via PKA, which regulates genes that converts short-term memories into long-term memories 48 . The combined ndings suggest that some genes with convergent identical amino acid changes may have a nexus at targeting the cAMP signaling pathway associated with the vocal learning ability ( Supplementary Fig. 12).
Although our study illuminated novel ndings, it spurs on ideas for future studies. Vocal learning species could share other convergent traits besides vocal learning 7,13,49,50 , and the identi ed genes could be associated with these other traits. The basic rules of convergent evolution we discovered in protein coding regions leave open the possibility that similar or different rules apply to non-coding regions. The greater association of ri eman CSAVs with vocal non-learning species could be further tested with brain and behavior studies, to see if indeed they do not have a vocal learning forebrain circuit or advanced vocal learning behavior. We identi ed new candidate genes and speci c nucleotide substitutions that can be genetically manipulated when the technology is more advanced 3,51 to test possible causal roles in the evolution and function of vocal learning. It will be useful to determine if the convergent rules we identi ed here are speci c to birds, or are more widespread across life forms.

Declarations
Author contributions CL, SC, KK, HK, and EDJ designed the study. CL and EDJ collected all of raw data for this study. CL, KK, HK, and EDJ developed algorithms and programs to identify convergent variants. CL, HK, and EDJ designed random and core control sets by considering phylogenetic relationships among vocal learning birds, and developed algorithms and programs to calculate phylogenetic features. CL performed most of analyses. CL and DY performed PCA and phylogenetic analysis to estimate the vocal learning ability of Ri eman, and conducted xed difference analyses for public SNP resources of zebra nch and chicken. JH and HL validated xed differences of convergent amino acid variants in zebra nch and chicken. GG, JA, EH, MR, OW, ARP, and EDJ provided DEG pro les based their independent studies with DEG analyses on song nuclei and those surrounding regions. CL and EDJ wrote the draft paper and all of authors reviewed it. HK and EDJ supervised this study.  dCSAV (Types 3+4, x-axis) of random and core control species sets, respectively. (f, g) Correlation plots between iCSAV (Types 1+2) and product of origin branch lengths (POB) of random and core control sets, respectively. Different combinations of species with known convergent traits are indicated by colored symbols. Statistics calculated as a linear regression, with adjusted R2 and p values using the 'lm' function in the R package (ver. 3.5.1)52.

Figure 2
Amino acid convergence amount is correlated to the product of origin branch lengths. Shown are regression analyses of amino acid convergence for (a) total CSAV (Types 1-4), (b) iCSAV (Types 1+2), and (c) dCSAV (Types 3+4) in the vocal learning set and 1,000 random control sets of avian species with four phylogenetic tree features (top row): product of origin branch lengths; product of terminal branch lengths; distance among terminal branches; and distance among terminal nodes. In the example type of tree branches, red lines show the branches used for the calculations and red text the species clades that have a convergent trait. Right, legend key for control species combinations and species with known convergent traits, and the four types of CSAV (based on pattern in Fig. 1b). Statistics calculated as a linear regression, with adjusted R2 and p value using the 'lm' function in the R package (ver. 3.5.1)52.      or differentially expressed genes (DEG) in song learning nuclei or adjacent brain subdivisions of the zebra nch brain (y-axis) that have convergent amino acid coding sequences and have been positively selected, in vocal learners and the closest control set of species (x-axis). DEGs collected from three independent sources based on microarray14,25, micro-dissected RNA sequencing27, and laser capture microscope RNA sequencing27 data sets from 2014, 2019, and 2020 analyses, respectively (Supplementary Note 11). NUC vs SUR: song nucleus compared to its surrounding non-vocal motor brain regions. NUC vs NUC: a song nucleus compared to another song nucleus. SUR vs SUR: a surrounding region of a song nucleus compared to another song nucleus. (b) Songbird brain diagram showing the song learning system. Yellow, forebrain song learning brain regions with SRG and DEGs measured. Grey, other song learning nuclei. Grey arrows, connections between the song nuclei. Red-up arrow and blue-down arrow indicates number of AVL-iCSAV genes supported by differential expressions in 'NUC vs other NUCs' and 'SUR vs other SURs' comparisons / in the DEG_2020 data source. (c) DRD1B mRNA expression pattern in zebra nch Area X and surrounding striatum at 3 different development time points, with specialized expression (white arrows) appearing by adulthood. Image used with permission from Kubikova et al. 28. Proportion of genes with human speci c amino acid substitutions. Human-SAV, is % of genes with human speci c amino acid substitutions compared to 24 non-human primates (n=18,852 total genes analyzed). Non-human primate SAV, is % of genes with non-human primate speci c amino acid substitutions in each of 23 non-human primates compared to all other primates including human. Human SAV in AVL-CSAV, is % of genes with avian vocal learner-speci c substitutions that overlap with genes containing human-speci c substitutions (n=107 out of 124 orthologs identi ed between vocal learning birds and humans). Human-SAV in random Ctrl-CSAV, same type of results but for 1000 random control sets of avian species. Human-SAV in core Ctrl-CSAV, same type of results but for 61 core control sets of avian species.