Skip to main content
PMC home page
eLife logoLink to eLife
. 2019 Sep 17;8:e47216. doi: 10.7554/eLife.47216

Learning is enhanced by tailoring instruction to individual genetic differences

David G Mets 1,2,✉, Michael S Brainard 1,2,3,4,✉
Editors: Ronald L Calabrese5, Ronald L Calabrese6
PMCID: PMC6748825  PMID: 31526480

Abstract

It is widely argued that personalized instruction based on individual differences in learning styles or genetic predispositions could improve learning outcomes. However, this proposition has resisted clear demonstration in human studies, where it is difficult to control experience and quantify outcomes. Here, we take advantage of the tractable nature of vocal learning in songbirds (Lonchura striata domestica) to test the idea that matching instruction to individual genetic predispositions can enhance learning. We use both cross-fostering and computerized instruction with synthetic songs to demonstrate that matching the tutor song to individual predispositions can improve learning across genetic backgrounds. Moreover, we find that optimizing instruction in this fashion can equalize learning differences across individuals that might otherwise be construed as genetically determined. Our results demonstrate potent, synergistic interactions between experience and genetics in shaping song, and indicate the likely importance of such interactions for other complex learned behaviors.

Research organism: Other

eLife digest

Some people do better at school than others, and some of this difference comes down to genes. But do genes place fixed limits on an individual's academic potential? Or is it possible to increase or decrease the impact of genes by changing how a person is taught? One possibility is that individuals learn best in different ways, and that tailoring instruction to suit individual learning styles could improve learning outcomes. But despite widespread interest in this idea, testing it systematically has proven challenging.

Mets and Brainard have therefore taken a new approach by testing the idea in a songbird called the Bengalese finch. Birdsong is a complex behavior learned in a similar way to human speech. Young birds listen to a tutor song – usually that of their father – and learn to mimic it through trial and error. But some songbirds learn better than others. By swapping eggs between nests, Mets and Brainard show that genetic offspring often learn the father’s song more accurately than birds fostered in from another nest. This might be because the father and offspring share genetic characteristics that contribute to the sound of the father’s song. Birds with the same genes will thus find it easier to learn the same song.

Alternatively, it could be that father birds spend more time teaching their genetic offspring than young they have fostered. To control for this possibility, Mets and Brainard played computer-generated songs to juvenile birds from different nests that had all been raised by non-singing females. Some of the songs had a fast tempo, others were slow, and a third set were in between. The results showed that juveniles learned most successfully when the training song had a similar tempo to their father’s song. This was true even though none of the birds had ever heard their father sing.

The findings thus suggest that tailoring instruction to suit an individual's natural learning tendencies – which depend on their genes – can enhance learning. Without knowing about this effect, it would be easy to assume that some of the songbirds in the current study were simply poorer learners than others. But in fact, optimizing instruction for each individual’s genetic background reduced the differences between individuals. If learning in humans is similar to vocal learning in birds, there could be broad implications for education.

Introduction

Recent studies in human populations have demonstrated strong genetic influences on academic achievement (Branigan et al., 2013; Lee et al., 2018; Okbay et al., 2016; Rietveld et al., 2013). This raises the question of whether genes place immutable bounds on achievement or whether experiential factors could amplify or dampen the impact of genetic variation. A particularly intriguing possibility is that customizing instruction based on individual differences in learning styles or genetic predispositions could improve learning outcomes (Asbury and Plomin, 2013; Moser and Zumbach, 2018; Pashler et al., 2009; Plomin, 2014). However, despite widespread interest in the potential value of ‘personalized’ instruction, it has been difficult to evaluate this idea in human populations, where it is challenging to control genetic variation, manipulate experience and quantify outcomes.

Birdsong affords an attractive system for studying how tailoring instruction based on genetic differences influences learning outcomes. Song, like human speech, is a complex vocal behavior that is learned during early life from adult vocal models (Doupe and Kuhl, 1999). Young birds listen to a 'tutor song’ (usually that of their father), and through practice develop vocalizations that closely match this target (e.g. Figure 1A). Although much prior work has focused on how experience shapes song learning, genetic predispositions also contribute to learned song structure at both the species and individual levels (Fehér et al., 2009; Gardner et al., 2005; Marler and Peters, 1977; Marler and Peters, 1982; Marler and Peters, 1988; Mets and Brainard, 2018a; Mundinger, 1995; Mundinger and Lahti, 2014; Podos et al., 1999; Soha and Marler, 2000; Thorpe, 1954). For example, while many species can learn aspects of heterospecific song, birds will preferentially acquire species typical song structure (spectral content and ordering of syllables) when tutored with a combination of heterospecifc and conspecific songs (Marler and Peters, 1977; Marler and Peters, 1982; Marler and Peters, 1988; Mundinger, 1995; Mundinger and Lahti, 2014; Podos et al., 1999; Soha and Marler, 2000; Thorpe, 1958). Consistent with this, deviations between a tutor song and species typical song can lead to poor copying of the tutor song (Lahti et al., 2011; Marler and Peters, 1988; Podos, 1997; Podos et al., 2004). Moreover, we have previously found that even within a single-species colony of Bengalese finches (Lonchura striata domestica), there is a strong heritable predisposition for individuals to produce songs at differing tempos (Mets and Brainard, 2018a). The presence of such heritable biases for learned song structure, together with the ease of controlling instructive experience and quantifying learning outcomes, renders song learning in Bengalese finches particularly suitable for testing whether tailoring instruction in accordance with individual genetic predispositions can enhance learning.

Figure 1. There is broad variation in the quality of learning across individuals.

Figure 1.

Open in a new tab

(A) Spectrograms illustrating learning of song from different tutors. Adult song is composed of a series of discrete units of sound (syllables) separated by silence, with each syllable containing characteristic spectral content. Two fathers’ songs (top: i and ii) had distinct syllable structure that was well learned by their offspring (bottom: i and ii). (B) Spectrograms of songs exemplifying the range of variation in learning present under conventional rearing and tutoring conditions. Similarity between the learned songs of individuals (Tutee songs) and the song of their tutor (Tutor song) is indicated by Song Divergence (SD) scores at left. SD indicates spectral content from a tutor song that is missing from the learned song, such that lower SD scores indicate better learning (see Materials and methods). The song of a bird that had no exposure to a tutor (isolate song) is shown for comparison. (C) The distribution of learning outcomes for 142 birds that were reared and tutored in their home nests. A gamma distribution fit to the data (blue line) and the 50th percentile (green) and 80th percentile (red) indicate that learning outcomes were skewed towards good learning (smaller SD values). Red and black arrows correspond to the learned ‘tutee’ songs and ‘isolate’ song presented in panel B.

Results

Within our genetically heterogeneous Bengalese finch colony, we found that there was a broad range in the quality of song learning. Many juveniles that were reared conventionally in their home nests and tutored by their genetic fathers learned to copy tutor song flexibly and with high fidelity (e.g. Figure 1Ai–ii, tutor vs. learned song). However, across individuals, learning could range from nearly perfect (e.g. Figure 1B, tutor song vs. tutee song, top) to extremely poor (e.g. Figure 1B, tutor song vs. tutee song, bottom; compare with isolate song). We quantified song learning using the Song Divergence (SD), a measure that estimates how much of the spectral content of syllables in the tutor song is absent from the learned song (Mets and Brainard, 2018b); hence, an SD of 0 indicates a song that perfectly matches the tutor song, while increasing values of SD (quantified in bits) indicate songs that are progressively worse copies of the tutor. The SD is computed in an automated fashion from a large set of syllables randomly selected from each bird’s song, and provides a measure that corresponds well with human assessments of song learning across a broad range of learning quality (see Materials and methods) (Mets and Brainard, 2018b). Prior work indicates that variation in the quality of song learning (Figure 1C) could be influenced by both experiential and genetic factors (Chen et al., 2016; Doupe and Kuhl, 1999; Fehér et al., 2009; Gardner et al., 2005; Marler and Peters, 1988; Marler and Peters, 1982; Marler and Peters, 1977; Mets and Brainard, 2018a; Mundinger, 1995; Mundinger and Lahti, 2014; Podos et al., 1999; Soha and Marler, 2000; Tchernichovski et al., 1999; Thorpe, 1958). Here, we were interested in the possibility that a component of this variation could be explained by an interaction between these factors - specifically, whether matching instructive experience to genetic predispositions of individuals could improve learning outcomes.

To investigate whether alignment between tutoring experience and individual genetic predispositions influences learning outcomes, we compared how well a tutor’s song was learned by his own genetic offspring relative to how well it was learned by birds that were cross-fostered from other nests (Figure 2A). We reasoned that many of the genetic factors that contribute to the structure of a father’s song would also be passed on to his offspring. Hence, we expected that a tutor’s song would be better matched to the genetic predispositions of his own offspring than to those of cross-fostered birds (Plomin et al., 1977). We hypothesized that if the correspondence between tutor song instruction and individual genetic biases influences learning, then home-reared birds would learn better than cross-fostered birds. Eight breeding pairs served as parents and provided tutoring to both their own home-reared progeny and to cross-fostered birds (see Materials and methods). To ensure that cross-fostered birds were not exposed to the songs of their genetic fathers, we transferred eggs to the nests of foster parents within 36 hr of laying, prior to the development of the peripheral auditory system (Murray et al., 2013; Yamasaki and Tonosaki, 1988). For both cohorts, juveniles were reared with their tutors from hatching until adulthood (~120 days of age), at which point their songs were recorded for analysis. In order to assess any learning differences in the context of the same tutor song and parental environment, we conducted paired comparisons of the median quality of learning within each nest for home-reared versus fostered juveniles.

Figure 2. Learning outcomes are better when juveniles are tutored by their genetic fathers.

(A) Schematic representation of the experimental design. Parental pairs (middle row, blue birds) reared both their own genetic offspring (home-reared, left), and foster birds from different genetic backgrounds (cross-fostered, right). (B) Examples of learning outcomes for two nests. (B, top) Example songs from the resident male tutors in each nest. (B, middle, bottom) Example songs from birds that were either home-reared (B, middle) or cross-fostered (B, bottom) in nests where males sang the tutor songs indicated in B, top. The examples of learned songs are from birds that had the median Song Divergence (SD) scores for their cohort (home-reared or cross-fostered). Syllable labels are provided to facilitate comparisons, but do not reflect automated SD scores used to quantify song similarity. (C) Paired plot of median SD scores for home-reared and cross-fostered cohorts. Across eight nests, birds that were home-reared learned significantly better than birds that were cross-fostered (n = 8 parental pairs, 52 cross-fostered birds, 45 home-reared birds; Wilcoxon signed-rank test, p<0.005). Median SD scores for nests corresponding to spectrograms in panel B are shown in red.

Figure 2.

Figure 2—figure supplement 1. Pedigree of all tutor pairs used in cross-fostering experiments.

Figure 2—figure supplement 1.

Squares indicate male birds while circles indicate female birds. Colored symbols indicate members of breeding pairs that were used as tutoring families for cross-foster experiments presented in Figure 2. Membership in a specific breeding pair is indicted by shared color and shared numbers. Triangle symbols indicate purchases from an outside vendor. A total of five purchases from outside vendors are indicated by numbers. Purchases were made over a period of 7 years ranging from 2001 to 2007.
Open in a new tab

Across nests, genetic offspring (home-reared birds) learned the tutor song significantly better than cross-fostered birds (Figure 2B,C; p<0.005, Wilcoxon signed-rank test). This suggests that better alignment between the acoustics of the tutors’ songs and the individual genetic predispositions of their offspring improved learning outcomes for home-reared birds. However, learning in these experiments also could have been facilitated for home-reared birds by genetic contributions to other aspects of instructive experience, such as the amount or quality of tutoring directed at offspring versus fostered birds.

To eliminate the confound of potential variability in individual interactions with the tutor, we next used a computer tutoring paradigm (Mets and Brainard, 2018a; Tchernichovski et al., 1999) to hold both the acoustic structure of the tutor song and the number of song exposures constant across all individuals (Figure 3A). We limited exposure to auditory stimuli other than the computer tutor song by transferring eggs within 36 hr of laying to nests where hatchlings were reared to independence by non-singing female foster parents. At ~45 days post-hatch, juveniles were transferred to sound chambers where they were computer-tutored with identical songs, yielding a cohort of 20 birds from 13 different breeding nests, all with the same tutor song instruction.

Figure 3. Learning outcomes are influenced by the relationship between tutor song tempo and genetic bias.

Figure 3.

Open in a new tab

(A) Schematic of computer tutoring. Individuals from different genetic backgrounds (red and blue nests) were raised by non-singing females to prevent exposure to their fathers’ songs and were then tutored using computer-controlled playback of identical tutor stimuli. (B, top) Spectrogram of a full motif from the synthetic computer tutor song with a tempo of 8.5 syl/s. (B, bottom) Example songs for three tutees from different nests displaying variation in learning outcomes for song spectral content. The SD scores for the song repertoires of these individual birds are shown next to the example spectrograms. These tutees (from top to bottom) came from nests where the fathers’ song tempos were similar to (8.5 syl/s), faster than (9.8 syl/s), and slower than (7.2 syl/s) the tutor song. (C) The distribution of learning outcomes for 20 birds drawn from 13 different genetic backgrounds. A gamma distribution fit to the data (blue line) and the 50th percentile (orange) and 80th percentile (green) indicate that learning outcomes were skewed towards better learning. Red arrows correspond to the learned songs presented in panel B. (D) Mean (gray bars) and median (red dots) SD scores for birds that were grouped by the tempos of their genetic fathers’ songs. Learning was best for birds that had a genetic bias to sing near the tempo of the tutor song (fathers’ song tempos from 7.5 to 9.5 syl/s), while learning was worse for birds that were biased to sing slower than the tutor song (<7.5 syl/s; * denotes p<0.02, one-tailed t-test) or faster than the tutor song (>9.5 syl/s; p=0.06, one-tailed t-test). Additionally, birds that were biased to sing faster than the tutor song learned better than birds that were biased to sing slower than the tutor song (** denotes p<0.02, two-tailed t-test). All p-values were corrected for multiple testing using the Holm-Bonferroni procedure.

Despite exposure to controlled computer tutoring experience, the distribution of learning outcomes was similar to that generated by live tutoring; some birds learned songs for which the spectral content of syllables closely resembled the tutor stimulus (Figure 3B, top row of 'tutee songs', and 3C) while other birds learned songs that had little resemblance to the tutor stimulus (Figure 3B, bottom row of 'tutee songs', and 3C). Thus, differences in parental behavior and in individual experience with the live tutor could not account for observed variation in the quality of learning.

In contrast, we found that a significant amount of variation in the quality of learning could be explained by how well the tutor stimulus matched individual genetic biases of juveniles. Previous work (Mets and Brainard, 2018a) demonstrated that juveniles learn songs with tempos that are strongly biased towards the tempos of their fathers’ songs, even when they have never heard their fathers sing. We therefore estimated the genetic bias for tempo of individuals from different nests as the tempo (in syl/s) measured for their fathers’ songs. Across the juvenile birds that were computer tutored, the individual biases for tempo ranged from 5.5 syl/s to 12.5 syl/s. Hence, the computer tutor stimulus, which was presented at 8.5 syl/s, was better matched to the genetic biases of some individuals than for others. We measured the quality of learning for computer tutored birds in three groups: birds that had biases for tempo that were slower than (<7.5 syl/s), similar to (7.5–9.5 syl/s), or faster than (>9.5 syl/s), the tempo of the tutor song (8.5 syl/s). Across these groups, the birds that had biases for tempo that were most similar to the tutor song learned best (Figure 3D). Thus, as with live tutoring, the quality of learning for spectral content of a controlled, synthetic song could be explained in part by the alignment between instructive experience and individual genetic bias.

To more explicitly test whether matching the tempo of the tutor song to the genetic bias of individual birds could enhance learning, we carried out an additional computer tutoring experiment in which we assessed how varying the tempo of the tutor song influenced the quality of learning for birds from three different genetic backgrounds that were biased to sing at differing song tempos - one near the lower decile of song tempos produced in our colony (7.18 syl/s father song tempo), one near the median (7.89 syl/s), and one near the upper decile (10.31 syl/s). We tutored groups of juveniles from each of these ‘slow’, ‘medium’ and ‘fast’ genetic backgrounds with songs that had identical spectral content but that were presented at ‘slow’, ‘medium’ and ‘fast’ tempos (6.5 syl/s, 8.5 syl/s, and 10.5 syl/s; see Materials and methods) (n = 28 animals; Schematized, Figure 4A). Synthetic tutor songs that differed in tempo were constructed by varying the gaps between syllables, without altering either the durations or spectral content of the syllables themselves (Figure 4—figure supplement 1). Hence, every bird was tutored with stimuli that had the same spectral content and number of syllables. Correspondingly, in order to compare learning across groups tutored with different tempo songs, we used the Song Divergence, which quantifies how well syllable spectral content is learned independently of song temporal structure. The differences in quality of learning that we observed therefore reflect influences of the rate at which syllables are presented on an orthogonal aspect of song structure - how well the spectral content of syllables is learned.

Figure 4. Matching tutor song tempo to the genetic bias of an individual results in better learning.

(A) Cohorts of individuals from ‘slow’, ‘medium’ and ‘fast’ genetic backgrounds (left) were tutored with ‘slow’, ‘medium’ and ‘fast’ synthetic tutor songs (bottom). The normalized, average Song Divergence for each of the nine cohorts is indicated in grayscale, with darker shades indicating better learning. (B) Across the three genetic backgrounds, learning outcomes were best when tutor-song tempo and genetic bias were matched (purple vs. orange and green). If unmatched, tutor song was learned better if tutor song was slower than genetic bias for song tempo (orange vs. green). For all panels, error bars are standard error. ** denotes p<0.02, two-tailed t-test. All p-values were corrected for multiple testing using Holm-Bonferroni procedure.

Figure 4.

Figure 4—figure supplement 1. Birds from a genetic background biased to sing at a medium song tempo learn best when provided with a tutor song at a medium song tempo.

Figure 4—figure supplement 1.

(A) Spectrograms of each unique syllable that made up the synthetic tutor song, and corresponding labels for identification. Synthetic tutor songs that differed in tempo were constructed by varying the gaps between syllables, without altering either the durations or spectral content of the syllables themselves. (B) Examples of learned songs produced by three birds from a genetic background biased to sing at a medium song tempo (7.8 syl/s) that were tutored on songs with slow, medium or fast tempos (Figure 4A, ‘medium’ population). Although syllables from the tutor song were recognizable in each of the learned songs, the best learning was for the individual that was tutored with a medium tempo song (8.5 syl/s), most closely matched to the genetic background. Summary data are quantified in Figure 4 and Figure 4—figure supplement 1.

Figure 4—figure supplement 2. Quantification of song learning for specific tutor-song tempo father-song tempo pairings.

Figure 4—figure supplement 2.

(A) Cohorts of individuals from each of three genetic backgrounds (‘slow’, cyan; ‘medium’, red; ‘fast’, blue) were tutored with songs of one of three tempos (‘slow’, 6.5 syl/s; ‘medium’, 8.5 syl/s; ‘fast’, 10.5 syl/s). For each of the three genetic backgrounds, birds that heard a tutor song that was matched to their genetic background learned best. Bars indicate the average Song Divergence for birds from each of the nine categorical combinations of tutor-song tempo and father-song tempo (as shown in Figure 4A). While the average learning across all tutor song tempos was worse for the slow genetic background than for the fast genetic background, the slow birds actually learned better than the fast birds when each group was provided with the optimal tutor song for that group (compare slow birds tutored at 6.5 syl/s with fast birds tutored at 10.5 syl/s). (B) Genetic background (father’s song tempo) considered alone explained only 4.1% of the variation in SD scores across the population of tutored birds. (C) Tutor song tempo (slow, medium or fast) considered alone explained only 0.28% of the variation in SD scores. For all panels, error bars indicate standard error.

Figure 4—figure supplement 3. Birds did not deviate from good syllable copying in order to optimize song tempo.

Figure 4—figure supplement 3.

Data are from 20 birds tutored with a synthetic song that had a tempo of 8.5 syl/s. These individuals were drawn from a diversity of genetic backgrounds. (A) Quality of song copying (Song Divergence in bits) as a function of learned song tempo. If birds were learning faster songs at the expense of good syllable copying, we would expect increasing Song Divergence as song tempo increases, however this trend is not apparent (expectation indicated in inset). (B) Quality of song copying as a function of the absolute difference between the learned song tempo and the tutor song tempo. If birds were learning to match the tutor song tempo at the expense of good syllable copying, we would expect the songs with the smallest difference between learned tempo and tutor tempo to have the largest SD scores (expectation indicated in inset). In contrast, a regression between these two variables (red) results in a significant positive slope. This regression was fit with the ordinary least squares algorithm and results in a slope of 0.577, with a p value of 0.0013, and an r-squared of 0.446.
Open in a new tab

Joint consideration of tutor song tempo and individual genetic bias revealed a strong interaction between the two; for each of the three genetic backgrounds, the best learning was achieved if tutor song tempo was ‘matched’ to the genetic bias (Figure 4A,B; example spectrograms illustrated in Figure 4—figure supplement 1 and individual groups quantified in Figure 4—figure supplement 2A). Within the ‘unmatched’ groups, birds learned better if they were tutored with a song slower than their genetic bias (Figure 4B, p<0.02, two-tailed t-test, Holm-Bonferroni correction for multiple comparisons). These three categories (tutor song faster than father’s song, tutor song matched to father’s song, and tutor song slower than father’s song) accounted for 39.8% of the variation in song learning outcomes (ANOVA, p<0.008, Holm-Bonferroni correction for multiple comparisons). In contrast, genetic background considered alone, and tutor song tempo considered alone, explained only 4.1% and 0.28%, respectively, of the variation in learning (Figure 4—figure supplement 2B,C). These findings demonstrate that much of the quality of song learning can be explained by the interaction between tutor song experience and genetic bias.

We considered whether the interaction between tutor experience and genetic bias could reflect a trade-off between learning song spectral structure and optimizing tempo. In particular, for some species, it has been suggested that individuals strike a balance between producing syllables with broadband spectral content and producing those syllables at a faster tempo (Lahti et al., 2011; Podos, 1997; Podos et al., 2004). This observation raises the possibility that worse SD scores in our study reflect in part a potentially advantageous sacrifice in the quality of spectral copying in order to optimize song tempo. To examine this possibility, we first tested whether SD scores were worse for individuals that sang faster songs than for individuals that sang slower songs, as might be expected if there was a trade-off between the quality of spectral copying and maximizing song tempo - a song feature that has been identified in some studies as more attractive to females and therefore favored (Ballentine, 2004; Nowicki and Searcy, 2005). We did not find this trend in our data (Figure 4—figure supplement 3A). We then tested whether SD scores were worse for individuals that sang songs more closely matched in tempo to the tutor song, as might be expected if there was a trade-off between the quality of song spectral copying and the ability to match tutor song tempo. In contrast to this possibility, we found that there was significantly better copying of syllable spectral content in birds that also more closely matched song tempo to the tutor song (Figure 4—figure supplement 3B). This is consistent with the idea that the ability of an individual to match the tutor song tempo facilitates the learning of syllable spectral content, and argues against the possibility that birds with poor learning of spectral content are optimizing some other song feature.

Discussion

Together, the cross-fostering and computer tutoring experiments show that tailoring an instructive stimulus to match the genetic bias of an individual can enhance learning. It is noteworthy that our estimates of individual bias derived exclusively from the tempo of the father’s song; hence, a more accurate estimate of bias, incorporating other genetic factors such as maternal contributions, would potentially enable even more effective tailoring of instruction. Our results additionally indicate that a failure to take into account how experience and genetics interact can lead to erroneous conclusions about the immutability of genetic constraints on individual differences in learning. For example, absent consideration of this interaction, birds that are genetically biased to sing slow songs appear to be inherently worse learners than birds that are biased to sing fast (e.g. averaged across all stimuli, slow birds learn worse than fast birds; see Figure 4—figure supplement 2). However, when instruction is individually tailored, it is apparent that ‘slow’ birds can learn as well as, or better than, ‘fast’ birds (Figure 4A and Figure 4—figure supplement 2; slow birds tutored with slow songs learned better than fast birds tutored with fast songs). Hence, customization of instruction can not only improve learning outcomes, but in so doing may also attenuate differences across individuals that otherwise might have been construed as genetically determined.

These results also inform an understanding of how genetic factors shape the cultural transmission of song within a population of birds. Previous work has indicated that songbirds are subject to genetic constraints at the species level that bias birds to learn better from songs that are more species-typical (Gardner et al., 2005; Lahti et al., 2011; Marler and Peters, 1988; Podos, 1997; Podos, 1996; Podos et al., 2004), and it has been proposed that such constraints might contribute to the long-term stability of a given species’ song. According to this idea, learning drives variation in songs across individuals, but genetic factors ‘pull’ all birds back toward a single species-specific song ‘template’, thus reducing drift in the population level song structure over generations (Fehér et al., 2009; Lachlan et al., 2018; Lachlan and Slater, 1999; Lynch, 1996).

The observation that individual Bengalese finches are biased to learn better from some songs than from others is broadly consistent with the idea that innate factors might constrain ‘drift’ away from species-typical song models. However, our finding that within this genetically heterogeneous population, each bird or nest may have a different genetic bias, indicates that there is no single song model toward which all individuals are drawn (indeed, the median song tempo in our colony was 8.5 syl/s, yet this ‘species typical’ song is a poor model for birds from faster and slower nests). Rather, our findings suggest an alternative possibility - that whatever genetic biases toward different song structure are present between nests or families could be preserved or amplified across generations, potentially contributing to gradual divergence of song structure within distinct subpopulations of birds.

These results also demonstrate that learning of complex skills like song can be shaped by both direct and indirect influences of parental genes. For home-reared birds, genetic factors bias the acoustic structure of the father’s song, including tempo, but also contribute to a similar bias in his offspring (Mets and Brainard, 2018a); this establishes an alignment of experience and genetics that we have shown can enhance learning. These findings support the idea that, for human families, a similar synergistic interaction between the home environments that are shaped by parental behavior, and the heritable predispositions of their children, contributes to the observed clustering within families of achievement in verbal, quantitative, musical and athletic domains (Meredith, 1973; Plomin et al., 1977; Tan et al., 2014; Vinkhuyzen et al., 2009). Such interactions may be especially potent for behaviors like speech and other motor skills, where ‘tutoring’ in the early home environment plays a critical role in shaping the perceptual and motor systems underlying performance (Kuhl, 2010).

More broadly, our findings highlight the critical role that gene-experience interactions can play in determining complex phenotypes. When considered alone, neither genes nor experience had a clear impact on song learning, but their interaction explained nearly 40% percent of variation in learning outcomes. Detection of this strong interaction required detailed knowledge of individual genetic predispositions and experience, information that is often hard to obtain in the context of human studies (Halldorsdottir and Binder, 2017). Nevertheless, a better understanding of such interactions between genetics and experience will likely be required to fully elucidate the mechanisms driving individual-to-individual variation in complex learned behaviors.

Materials and methods

Subjects

Subjects were male Bengalese finches (Lonchura striata domestica). 239 birds were reared and tutored by live birds and 47 additional birds were reared by foster females and tutored by computer. These animals were bred in our colony. Other than efforts to maintain some separation between lineages, mating pairs comprised randomly selected male and female birds. All protocols were reviewed and approved by the Institutional Animal Care and Use Committee at the University of California, San Francisco. Data on tempo from a subset of the cross-foster and computer tutored birds was presented in a previous study (Mets and Brainard, 2018a).

Audio recording and initial processing

For audio recording, birds were single-housed in sound isolation chambers (Acoustic Systems). Songs were digitally recorded at a sampling frequency of 32 kHz, and a bit depth of 16. Recording microphones were placed in a fixed position at the top of the cage housing the bird. Prior to further analysis, all songs were high-pass filtered at ~500 Hz using a digitally implemented elliptical infinite impulse response filter with a passband edge frequency of 0.04 radians. All recordings used for analysis were acquired during early adulthood (90–120 days post hatch).

Cross-fostering

To create populations of birds that were tutored by a live bird but never heard the song of their genetic fathers, eggs were taken from parents within 36 hr of laying and transferred to foster nests where hatchlings were reared to adulthood. For each individual, the specific foster nest was randomly selected from a set of 8 foster nests within our breeding colony. All birds that were studied in these experiments were offspring of breeding pairs from our large and genetically heterogeneous colony. The breeding pairs themselves were established from birds that were acquired over a multi-year period from outside vendors, or were bred in house (see for example Figure 2—figure supplement 1). Given the structure of our colony, the cross-fostered birds were less related to their tutors than were the home-reared birds, but in many cases, the cross-fostered birds were more or less distant ‘cousins’ of their tutors. Our expectation is that effects we reported for these experiments (consistently better learning for genetic-offspring than for less-closely related cross-fostered birds) would have been even more pronounced had the cross-fostered birds been drawn from entirely different colonies or other outside sources.

Quantification of song learning

The quality of song learning was measured using Song Kullback-Leibler Divergence or Song Divergence (Mets and Brainard, 2018b). This is a largely automated measure that estimates the degree to which the spectral content of a tutor song is copied by a tutee, and has been shown to have a good correspondence with human assessments of song learning (Mets and Brainard, 2018b). Song Divergence is computed by creating a statistical model that captures the song spectral content of both tutor and tutee and estimating the divergence between these two models. This estimate of divergence is directional; the Song Divergence calculated in the tutor-tutee direction indicates the amount of song spectral content (in bits) present in the tutor song that is not present in the tutee song ("missing content") but does not capture spectral content from the tutee song that is not present in the tutor song ("improvisation"). The model for each bird in our study was generated from a corpus of 60 song bouts as follows. Songs were segmented into discrete units of sound (syllables) separated by silence using an automatically determined amplitude threshold. For each syllable, the spectral content was extracted by calculating a power spectral density (PSD). Here, we used a single PSD per syllable, removing information about syllable temporal structure. Compared to other semi-automated methods for evaluating song learning, this allows assessment of the copying of song spectral content independently of song temporal structure (Burkett et al., 2015; Mandelblat-Cerf and Fee, 2014; Mets and Brainard, 2018b; Tchernichovski et al., 2000). These PSDs were transformed into a syllable-syllable similarity space and a Gaussian mixture model (GMM) was then fit to the distributions of these syllables. The Kullback-Leibler Divergence between two GMMs corresponding to songs from a tutor and a tutee is the Song Divergence. In the case of a synthetic tutor stimulus where there is only a single variation of the song, Song Divergence was calculated by comparing the song corpus of a tutee to the song corpus of a bird that had learned the tutor song especially well. The song corpus of the same ‘tutor’ bird was used for all such calculations.

Song tempo calculation

Song tempo was quantified as the average number of syllables produced per second of song, a measure we have previously used to identify song tempo as heritable (Mets and Brainard, 2018a). Discrete units of sound separated by silence (syllables) were identified based on amplitude. First, an ‘amplitude envelope’ was created by rectifying the song waveform and then smoothing the waveform through convolution with an 8 ms square wave. Threshold crossings of this amplitude trace were then used to identify periods of vocalization. Thresholds were set heuristically to result in segmentation that corresponded with syllable onsets and offsets apparent in human examination of spectrograms. Once the threshold was established, ‘objects’ were identified as uninterrupted regions longer than 10 ms over which the amplitude envelope exceeded threshold. Any objects separated by a gap of 5 ms or less were merged producing a final set of objects that were defined as syllables. A series of syllables that had no gaps larger than 250 ms was considered a song bout. For each bird, tempo was then quantified as the number of syllables present in a song bout divided by the duration of the song bout, averaged across at least 60 bouts of singing.

Computer tutoring

To create populations of birds that had controlled tutoring experience, eggs were taken from parents within 36 hr of laying, prior to neural development (Murray et al., 2013; Yamasaki and Tonosaki, 1988), and were then raised by pairs of non-singing foster mothers housed in sound isolation chambers. While the female foster mothers do not sing, they produce unlearned calls and influence other aspects of a hatchling’s experience, including feeding and social interactions, that could plausibly influence hatchling development and the quality of subsequent computer driven song learning. In order to ensure that any such influences of female fostering on song learning did not contribute systematically to our results, we used 14 different female foster nests (each with two foster mothers) and randomly assigned eggs from experimental nests to these foster nests. In order to prevent any ‘batch effects’, only a single hatchling was raised at a time in any given foster nest. Foster mothers raised the juveniles until they were able to feed themselves. At independence (usually 35–40 days post hatch) birds were moved to an acoustic isolation chamber with an audio recording system and a computer tutoring apparatus, based on an approach that previously has been demonstrated to drive song learning (Mets and Brainard, 2018a; Tchernichovski et al., 1999). At 45 days post-hatch, the tutoring apparatus was activated, allowing birds to access a tutor song. The apparatus consisted of a perch activated switch that caused playback of a tutor stimulus (see below). Each perch hop elicited a single playback of the tutor stimulus. Birds were allowed to playback 10 songs, three times a day (morning, noon, and evening). Playback of tutor song was limited to 30 songs per day based on previous work indicating that this was near an optimal value to maximize the quality of song learning in this paradigm (Tchernichovski et al., 1999). It sometimes took birds a few days to begin consistently actuating song playbacks. Nevertheless, all birds elicited at least 90% of the available song playbacks during the tutoring period (median = 96%), and there was no relationship between the number of playbacks that an individual heard and the quality of song learning. The computer tutoring apparatus was implemented with custom LabView software (National Instruments, "EvTutor") that is provided in supporting materials for this paper. Birds remained in the tutoring apparatus until 120 days post-hatch. For experiments involving different tutor song tempos, the tempo for an individual was randomly selected from three possible tempos (see below).

Computer tutor stimulus

The synthetic stimulus used during playback tutoring was the same as that used in our previous work (Mets and Brainard, 2018a; Mets and Brainard, 2018b). To create a naturalistic but controlled learning stimulus, a synthetic song used for computer tutoring was derived from songs sampled from our Bengalese finch colony. The synthetic song was composed of 9 categorically distinct syllables that were chosen to reflect a range of different syllable types found in Bengalese finch song (i.e. short ‘introductory’ syllables, noisy syllables, syllables with harmonic structure and constant or modulated frequency, etc.). The tutor stimulus consisted of a series of introductory syllables followed by three repetitions of a stereotyped sequence of syllables, or ‘motif’ (shown in Figure 3B). The gaps between syllables were chosen to reflect naturalistic means and standard deviations based on the distribution of gap durations found in normal Bengalese finch song. Correspondingly, the tutor song stimulus had a relatively natural prosody compared to a stimulus in which the time between syllable onsets is fixed. The 8.5 syl/s tutor song stimulus arose naturally out of this process, as 8.5 syl/s is close to the median song tempo present in our colony. The 6.5 and 10.5 syl/s tutor stimuli were created by proportionally increasing or decreasing only the inter-syllable gap durations, resulting in songs with identical spectral content presented at different tempos.

Statistics

All statistical testing in this study was carried out in consultation with the biostatistics consultancy of the UCSF Clinical and Translational Science Institute. No birds were removed from the study. As no subjective measurements were made, no blinding was performed. When used, statistical tests were appropriate to the data presented and the data, to the limit of detection, were consistent with the assumptions of the tests. For all ANOVA analyses, estimates of variance explained are Omega squared. For each experimental group, p-values for statistical tests were corrected for multiple testing within that group using the Holm-Bonferroni procedure, an extension of the Bonferroni correction which retains the same family-wise error rate while reducing the false negative rate relative to traditional Bonferroni (Holm, 1979). Unless otherwise indicated, all tests were two-tailed. For all cases, where a one-tailed test was conducted because of a directional hypothesis, the threshold p-value for the test was considered to be significant at the 0.025 level to reduce the false positive rate to be equivalent to a two-tailed test. For all tests, the statistical significance was not impacted by the use of a one-tailed versus a two-tailed test. For the cross-fostering experiments presented in Figure 2, the median within-nest SD scores for home-reared birds were compared to the median within nest SD scores for cross-fostered juveniles. Median values were used as a summary statistic for this comparison because the SD scores were approximately gamma distributed (Figure 1C). The effect of home-rearing on song learning was tested using a Wilcoxon signed-rank test, a paired test which tests the sign of the change across experimental conditions and not the magnitude of the change. This enables a test for the impact of home-rearing relative to cross-fostering while controlling for nest-specific or tutor-song-specific variables.

Acknowledgements

The authors thank A Karpova, and W H Mehaffey for discussions and comments on the manuscript. This work was supported by the Howard Hughes Medical Institute, a PBBR award from the Sandler Family Foundation (MSB), and The Jane Coffin Childs Fund for Medical Research (DGM).

Funding Statement

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Contributor Information

David G Mets, Email: dmets@phy.ucsf.edu.

Michael S Brainard, Email: msb@phy.ucsf.edu.

Ronald L Calabrese, Emory University, United States.

Ronald L Calabrese, Emory University, United States.

Funding Information

This paper was supported by the following grants:

  • Jane Coffin Childs Memorial Fund for Medical Research Fellowship to DGM to David G Mets.

  • Howard Hughes Medical Institute to Michael S Brainard.

  • Sandler Foundation PBBR award to Michael S Brainard.

Additional information

Competing interests

No competing interests declared.

Author contributions

Conceptualization, Data curation, Software, Formal analysis, Funding acquisition, Investigation, Visualization, Methodology, Writing—original draft, Project administration, Writing—review and editing.

Conceptualization, Supervision, Funding acquisition, Writing—original draft, Writing—review and editing.

Ethics

Animal experimentation: All of the animals were handled according to approved institutional animal care and use committee (IACUC) protocols (#AN170723-02) of the University of California, San Francisco.

Additional files

Source code 1. Tutoring software.
elife-47216-code1.gz (937.9KB, gz)
DOI: 10.7554/eLife.47216.011

Data availability

All data generated or analysed during this study are included in the manuscript and supporting files.

References

  1. Asbury K, Plomin R. G Is for Genes: The Impact of Genetics on Education and Achievement. Wiley-Blackwell; 2013. [Google Scholar]
  2. Ballentine B. Vocal performance influences female response to male bird song: an experimental test. Behavioral Ecology. 2004;15:163–168. doi: 10.1093/beheco/arg090. [DOI] [Google Scholar]
  3. Branigan AR, McCallum KJ, Freese J. Variation in the heritability of educational attainment: an international Meta-Analysis. Social Forces. 2013;92:109–140. doi: 10.1093/sf/sot076. [DOI] [Google Scholar]
  4. Burkett ZD, Day NF, Peñagarikano O, Geschwind DH, White SA. VoICE: a semi-automated pipeline for standardizing vocal analysis across models. Scientific Reports. 2015;5:10237. doi: 10.1038/srep10237. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Chen Y, Matheson LE, Sakata JT. Mechanisms underlying the social enhancement of vocal learning in songbirds. PNAS. 2016;113:6641–6646. doi: 10.1073/pnas.1522306113. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Doupe AJ, Kuhl PK. Birdsong and human speech: common themes and mechanisms. Annual Review of Neuroscience. 1999;22:567–631. doi: 10.1146/annurev.neuro.22.1.567. [DOI] [PubMed] [Google Scholar]
  7. Fehér O, Wang H, Saar S, Mitra PP, Tchernichovski O. De novo establishment of wild-type song culture in the zebra finch. Nature. 2009;459:564–568. doi: 10.1038/nature07994. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Gardner TJ, Naef F, Nottebohm F. Freedom and rules: the acquisition and reprogramming of a bird's learned song. Science. 2005;308:1046–1049. doi: 10.1126/science.1108214. [DOI] [PubMed] [Google Scholar]
  9. Halldorsdottir T, Binder EB. Gene × environment interactions: from molecular mechanisms to behavior. Annual Review of Psychology. 2017;68:215–241. doi: 10.1146/annurev-psych-010416-044053. [DOI] [PubMed] [Google Scholar]
  10. Holm S. A simple sequentially rejective multiple test procedure. Scandinavian Journal of Statistics. 1979;6:65–70. [Google Scholar]
  11. Kuhl PK. Brain mechanisms in early language acquisition. Neuron. 2010;67:713–727. doi: 10.1016/j.neuron.2010.08.038. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Lachlan RF, Ratmann O, Nowicki S. Cultural conformity generates extremely stable traditions in bird song. Nature Communications. 2018;9:2417. doi: 10.1038/s41467-018-04728-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Lachlan RF, Slater PJB. The maintenance of vocal learning by gene–culture interaction: the cultural trap hypothesis. Proceedings of the Royal Society of London. Series B: Biological Sciences. 1999;266:701–706. doi: 10.1098/rspb.1999.0692. [DOI] [Google Scholar]
  14. Lahti DC, Moseley DL, Podos J. A tradeoff between performance and accuracy in bird song learning. Ethology. 2011;117:802–811. doi: 10.1111/j.1439-0310.2011.01930.x. [DOI] [Google Scholar]
  15. Lee JJ, Wedow R, Okbay A, Kong E, Maghzian O, Zacher M, Nguyen-Viet TA, Bowers P, Sidorenko J, Karlsson Linnér R, Fontana MA, Kundu T, Lee C, Li H, Li R, Royer R, Timshel PN, Walters RK, Willoughby EA, Yengo L, Alver M, Bao Y, Clark DW, Day FR, Furlotte NA, Joshi PK, Kemper KE, Kleinman A, Langenberg C, Mägi R, Trampush JW, Verma SS, Wu Y, Lam M, Zhao JH, Zheng Z, Boardman JD, Campbell H, Freese J, Harris KM, Hayward C, Herd P, Kumari M, Lencz T, Luan J, Malhotra AK, Metspalu A, Milani L, Ong KK, Perry JRB, Porteous DJ, Ritchie MD, Smart MC, Smith BH, Tung JY, Wareham NJ, Wilson JF, Beauchamp JP, Conley DC, Esko T, Lehrer SF, Magnusson PKE, Oskarsson S, Pers TH, Robinson MR, Thom K, Watson C, Chabris CF, Meyer MN, Laibson DI, Yang J, Johannesson M, Koellinger PD, Turley P, Visscher PM, Benjamin DJ, Cesarini D, 23andMe Research Team, COGENT (Cognitive Genomics Consortium), Social Science Genetic Association Consortium Gene discovery and polygenic prediction from a genome-wide association study of educational attainment in 1.1 million individuals. Nature Genetics. 2018;50:1112–1121. doi: 10.1038/s41588-018-0147-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Lynch A. The population memetics of birdsong. In: Kroodsma D. E, Miller E. H, editors. Ecology and Evolution of Acoustic Communication in Birds. Comstock Publishing Associates (Cornell University Press); 1996. pp. 187–197. [Google Scholar]
  17. Mandelblat-Cerf Y, Fee MS. An automated procedure for evaluating song imitation. PLOS ONE. 2014;9:e96484. doi: 10.1371/journal.pone.0096484. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Marler P, Peters S. Selective vocal learning in a sparrow. Science. 1977;198:519–521. doi: 10.1126/science.198.4316.519. [DOI] [PubMed] [Google Scholar]
  19. Marler P, Peters S. Developmental overproduction and selective attrition: new processes in the epigenesis of birdsong. Developmental Psychobiology. 1982;15:369–378. doi: 10.1002/dev.420150409. [DOI] [PubMed] [Google Scholar]
  20. Marler P, Peters S. The role of song phonology and syntax in vocal learning preferences in the song sparrow, Melospiza melodia. Ethology. 1988;77:125–149. doi: 10.1111/j.1439-0310.1988.tb00198.x. [DOI] [Google Scholar]
  21. Meredith W. A model for analyzing heritability in the presence of correlated genetic and environmental effects. Behavior Genetics. 1973;3:271–277. doi: 10.1007/BF01067604. [DOI] [PubMed] [Google Scholar]
  22. Mets DG, Brainard MS. Genetic variation interacts with experience to determine interindividual differences in learned song. PNAS. 2018a;115:421–426. doi: 10.1073/pnas.1713031115. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Mets DG, Brainard MS. An automated approach to the quantitation of vocalizations and vocal learning in the songbird. PLOS Computational Biology. 2018b;14:e1006437. doi: 10.1371/journal.pcbi.1006437. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Moser S, Zumbach J. Exploring the development and impact of learning styles: an empirical investigation based on explicit and implicit measures. Computers & Education. 2018;125:146–157. doi: 10.1016/j.compedu.2018.05.003. [DOI] [Google Scholar]
  25. Mundinger PC. Behaviour-genetic analysis of canary song: inter-strain differences in sensory learning, and epigenetic rules. Animal Behaviour. 1995;50:1491–1511. doi: 10.1016/0003-3472(95)80006-9. [DOI] [Google Scholar]
  26. Mundinger PC, Lahti DC. Quantitative integration of genetic factors in the learning and production of canary song. Proceedings of the Royal Society B: Biological Sciences. 2014;281:20132631. doi: 10.1098/rspb.2013.2631. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Murray JR, Varian-Ramos CW, Welch ZS, Saha MS. Embryological staging of the zebra finch, Taeniopygia guttata. Journal of Morphology. 2013;274:1090–1110. doi: 10.1002/jmor.20165. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Nowicki S, Searcy WA. Song and mate choice in birds: how the development of behavior helps Us understand function. The Auk. 2005;122:1. doi: 10.1642/0004-8038(2005)122[0001:SAMCIB]2.0.CO;2. [DOI] [Google Scholar]
  29. Okbay A, Beauchamp JP, Fontana MA, Lee JJ, Pers TH, Rietveld CA, Turley P, Chen GB, Emilsson V, Meddens SF, Oskarsson S, Pickrell JK, Thom K, Timshel P, de Vlaming R, Abdellaoui A, Ahluwalia TS, Bacelis J, Baumbach C, Bjornsdottir G, Brandsma JH, Pina Concas M, Derringer J, Furlotte NA, Galesloot TE, Girotto G, Gupta R, Hall LM, Harris SE, Hofer E, Horikoshi M, Huffman JE, Kaasik K, Kalafati IP, Karlsson R, Kong A, Lahti J, van der Lee SJ, deLeeuw C, Lind PA, Lindgren KO, Liu T, Mangino M, Marten J, Mihailov E, Miller MB, van der Most PJ, Oldmeadow C, Payton A, Pervjakova N, Peyrot WJ, Qian Y, Raitakari O, Rueedi R, Salvi E, Schmidt B, Schraut KE, Shi J, Smith AV, Poot RA, St Pourcain B, Teumer A, Thorleifsson G, Verweij N, Vuckovic D, Wellmann J, Westra HJ, Yang J, Zhao W, Zhu Z, Alizadeh BZ, Amin N, Bakshi A, Baumeister SE, Biino G, Bønnelykke K, Boyle PA, Campbell H, Cappuccio FP, Davies G, De Neve JE, Deloukas P, Demuth I, Ding J, Eibich P, Eisele L, Eklund N, Evans DM, Faul JD, Feitosa MF, Forstner AJ, Gandin I, Gunnarsson B, Halldórsson BV, Harris TB, Heath AC, Hocking LJ, Holliday EG, Homuth G, Horan MA, Hottenga JJ, de Jager PL, Joshi PK, Jugessur A, Kaakinen MA, Kähönen M, Kanoni S, Keltigangas-Järvinen L, Kiemeney LA, Kolcic I, Koskinen S, Kraja AT, Kroh M, Kutalik Z, Latvala A, Launer LJ, Lebreton MP, Levinson DF, Lichtenstein P, Lichtner P, Liewald DC, Loukola A, Madden PA, Mägi R, Mäki-Opas T, Marioni RE, Marques-Vidal P, Meddens GA, McMahon G, Meisinger C, Meitinger T, Milaneschi Y, Milani L, Montgomery GW, Myhre R, Nelson CP, Nyholt DR, Ollier WE, Palotie A, Paternoster L, Pedersen NL, Petrovic KE, Porteous DJ, Räikkönen K, Ring SM, Robino A, Rostapshova O, Rudan I, Rustichini A, Salomaa V, Sanders AR, Sarin AP, Schmidt H, Scott RJ, Smith BH, Smith JA, Staessen JA, Steinhagen-Thiessen E, Strauch K, Terracciano A, Tobin MD, Ulivi S, Vaccargiu S, Quaye L, van Rooij FJ, Venturini C, Vinkhuyzen AA, Völker U, Völzke H, Vonk JM, Vozzi D, Waage J, Ware EB, Willemsen G, Attia JR, Bennett DA, Berger K, Bertram L, Bisgaard H, Boomsma DI, Borecki IB, Bültmann U, Chabris CF, Cucca F, Cusi D, Deary IJ, Dedoussis GV, van Duijn CM, Eriksson JG, Franke B, Franke L, Gasparini P, Gejman PV, Gieger C, Grabe HJ, Gratten J, Groenen PJ, Gudnason V, van der Harst P, Hayward C, Hinds DA, Hoffmann W, Hyppönen E, Iacono WG, Jacobsson B, Järvelin MR, Jöckel KH, Kaprio J, Kardia SL, Lehtimäki T, Lehrer SF, Magnusson PK, Martin NG, McGue M, Metspalu A, Pendleton N, Penninx BW, Perola M, Pirastu N, Pirastu M, Polasek O, Posthuma D, Power C, Province MA, Samani NJ, Schlessinger D, Schmidt R, Sørensen TI, Spector TD, Stefansson K, Thorsteinsdottir U, Thurik AR, Timpson NJ, Tiemeier H, Tung JY, Uitterlinden AG, Vitart V, Vollenweider P, Weir DR, Wilson JF, Wright AF, Conley DC, Krueger RF, Davey Smith G, Hofman A, Laibson DI, Medland SE, Meyer MN, Yang J, Johannesson M, Visscher PM, Esko T, Koellinger PD, Cesarini D, Benjamin DJ, LifeLines Cohort Study Genome-wide association study identifies 74 loci associated with educational attainment. Nature. 2016;533:539–542. doi: 10.1038/nature17671. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Pashler H, McDaniel M, Rohrer D, Bjork R. Learning styles concepts and evidence. Psychol Sci Public Interes Suppl. 2009;9:105–119. doi: 10.1111/j.1539-6053.2009.01038.x. [DOI] [PubMed] [Google Scholar]
  31. Plomin R, DeFries JC, Loehlin JC. Genotype-environment interaction and correlation in the analysis of human behavior. Psychological Bulletin. 1977;84:309–322. doi: 10.1037/0033-2909.84.2.309. [DOI] [PubMed] [Google Scholar]
  32. Plomin R. Genotype-environment correlation in the era of DNA. Behavior Genetics. 2014;44:629–638. doi: 10.1007/s10519-014-9673-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Podos J. Motor constraints on vocal development in a songbird. Animal Behaviour. 1996;51:1061–1070. doi: 10.1006/anbe.1996.0107. [DOI] [Google Scholar]
  34. Podos J. A performance constraint on the evolution of trilled vocalizations in a songbird family (Passeriformes: emberizidae) Evolution. 1997;51:537–551. doi: 10.2307/2411126. [DOI] [PubMed] [Google Scholar]
  35. Podos J, Nowicki S, Peters S. Permissiveness in the learning and development of song syntax in swamp sparrows. Animal Behaviour. 1999;58:93–103. doi: 10.1006/anbe.1999.1140. [DOI] [PubMed] [Google Scholar]
  36. Podos J, Peters S, Nowicki S. Calibration of song learning targets during vocal ontogeny in swamp sparrows, Melospiza georgiana. Animal Behaviour. 2004;68:929–940. doi: 10.1016/j.anbehav.2003.11.018. [DOI] [Google Scholar]
  37. Rietveld CA, Medland SE, Derringer J, Yang J, Esko T, Martin NW, Westra HJ, Shakhbazov K, Abdellaoui A, Agrawal A, Albrecht E, Alizadeh BZ, Amin N, Barnard J, Baumeister SE, Benke KS, Bielak LF, Boatman JA, Boyle PA, Davies G, de Leeuw C, Eklund N, Evans DS, Ferhmann R, Fischer K, Gieger C, Gjessing HK, Hägg S, Harris JR, Hayward C, Holzapfel C, Ibrahim-Verbaas CA, Ingelsson E, Jacobsson B, Joshi PK, Jugessur A, Kaakinen M, Kanoni S, Karjalainen J, Kolcic I, Kristiansson K, Kutalik Z, Lahti J, Lee SH, Lin P, Lind PA, Liu Y, Lohman K, Loitfelder M, McMahon G, Vidal PM, Meirelles O, Milani L, Myhre R, Nuotio ML, Oldmeadow CJ, Petrovic KE, Peyrot WJ, Polasek O, Quaye L, Reinmaa E, Rice JP, Rizzi TS, Schmidt H, Schmidt R, Smith AV, Smith JA, Tanaka T, Terracciano A, van der Loos MJ, Vitart V, Völzke H, Wellmann J, Yu L, Zhao W, Allik J, Attia JR, Bandinelli S, Bastardot F, Beauchamp J, Bennett DA, Berger K, Bierut LJ, Boomsma DI, Bültmann U, Campbell H, Chabris CF, Cherkas L, Chung MK, Cucca F, de Andrade M, De Jager PL, De Neve JE, Deary IJ, Dedoussis GV, Deloukas P, Dimitriou M, Eiríksdóttir G, Elderson MF, Eriksson JG, Evans DM, Faul JD, Ferrucci L, Garcia ME, Grönberg H, Guðnason V, Hall P, Harris JM, Harris TB, Hastie ND, Heath AC, Hernandez DG, Hoffmann W, Hofman A, Holle R, Holliday EG, Hottenga JJ, Iacono WG, Illig T, Järvelin MR, Kähönen M, Kaprio J, Kirkpatrick RM, Kowgier M, Latvala A, Launer LJ, Lawlor DA, Lehtimäki T, Li J, Lichtenstein P, Lichtner P, Liewald DC, Madden PA, Magnusson PK, Mäkinen TE, Masala M, McGue M, Metspalu A, Mielck A, Miller MB, Montgomery GW, Mukherjee S, Nyholt DR, Oostra BA, Palmer LJ, Palotie A, Penninx BW, Perola M, Peyser PA, Preisig M, Räikkönen K, Raitakari OT, Realo A, Ring SM, Ripatti S, Rivadeneira F, Rudan I, Rustichini A, Salomaa V, Sarin AP, Schlessinger D, Scott RJ, Snieder H, St Pourcain B, Starr JM, Sul JH, Surakka I, Svento R, Teumer A, Tiemeier H, van Rooij FJ, Van Wagoner DR, Vartiainen E, Viikari J, Vollenweider P, Vonk JM, Waeber G, Weir DR, Wichmann HE, Widen E, Willemsen G, Wilson JF, Wright AF, Conley D, Davey-Smith G, Franke L, Groenen PJ, Hofman A, Johannesson M, Kardia SL, Krueger RF, Laibson D, Martin NG, Meyer MN, Posthuma D, Thurik AR, Timpson NJ, Uitterlinden AG, van Duijn CM, Visscher PM, Benjamin DJ, Cesarini D, Koellinger PD, LifeLines Cohort Study GWAS of 126,559 individuals identifies genetic variants associated with educational attainment. Science. 2013;340:1467–1471. doi: 10.1126/science.1235488. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Soha JA, Marler P. A species-specific acoustic cue for selective song learning in the white-crowned sparrow. Animal Behaviour. 2000;60:297–306. doi: 10.1006/anbe.2000.1499. [DOI] [PubMed] [Google Scholar]
  39. Tan YT, McPherson GE, Peretz I, Berkovic SF, Wilson SJ. The genetic basis of music ability. Frontiers in Psychology. 2014;5:658. doi: 10.3389/fpsyg.2014.00658. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Tchernichovski O, Lints T, Mitra PP, Nottebohm F. Vocal imitation in zebra finches is inversely related to model abundance. PNAS. 1999;96:12901–12904. doi: 10.1073/pnas.96.22.12901. [DOI] [PMC free article] [PubMed] [Google Scholar]
  41. Tchernichovski O, Nottebohm F, Ho CE, Pesaran B, Mitra PP. A procedure for an automated measurement of song similarity. Animal Behaviour. 2000;59:1167–1176. doi: 10.1006/anbe.1999.1416. [DOI] [PubMed] [Google Scholar]
  42. Thorpe WH. The process of Song-Learning in the chaffinch as studied by means of the sound spectrograph. Nature. 1954;173:465–469. doi: 10.1038/173465a0. [DOI] [Google Scholar]
  43. Thorpe WH. The learning of song patterns by birds, with especial reference to the song of the chaffinch Fringilla coelebs. Ibis. 1958;100:535–570. doi: 10.1111/j.1474-919X.1958.tb07960.x. [DOI] [Google Scholar]
  44. Vinkhuyzen AA, van der Sluis S, Posthuma D, Boomsma DI. The heritability of aptitude and exceptional talent across different domains in adolescents and young adults. Behavior Genetics. 2009;39:380–392. doi: 10.1007/s10519-009-9260-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  45. Yamasaki M, Tonosaki A. Developmental stages of the society Finch, Lonchura striata var. dornestica. (society Finch/developmental stages/morphogenesis/laboratory animal) Development, Growth and Differentiation. 1988;30:515–542. doi: 10.1111/j.1440-169X.1988.00515.x. [DOI] [PubMed] [Google Scholar]

Decision letter

Editor: Ronald L Calabrese1
Reviewed by: Marc F Schmidt2, Frank Johnson3

In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.

Thank you for submitting your article "Learning is enhanced by tailoring instruction to individual genetic differences" for consideration by eLife. Your article has been reviewed by three peer reviewers, and the evaluation has been overseen by Ronald Calabrese as the Senior and Reviewing Editor. The following individuals involved in review of your submission have agreed to reveal their identity: Marc F Schmidt (Reviewer #2); Frank Johnson (Reviewer #3).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

Summary:

In a previous study (PNAS 2018) the authors showed that genetic predispositions for vocal learning can be demonstrated in songbirds when birds are tutored using computerized instruction but that this difference can be eliminated when birds are tutored with live tutors. In this study the authors show that tailoring tutoring to the type of genetic predisposition can overcome potential deficits in learning. The authors further show that matching the tempo of the tutor song with the tempo of the tutee's father song (i.e. the genetically preferred tempo) greatly improves the tutee's overall ability of learning the acoustic features of the song. Tutoring juveniles with songs that are either of a faster or slower tempo than the father's song causes a decrease in overall copying performance when measured as a degree of acoustic similarity of individual syllables.

Essential revisions:

The results are exciting and of potential general interest, however, there were some concerns that must be addressed before publication. The written reviews are attached and are not completely consistent but do contain valuable insight into details of any revision plan. In consultation the reviewers reached the following consensus.

1) Given that authors recently shown that song tempo is heritable, and that other previous work showed that rhythm mismatch can decrease imitation (a 'calibration effect'), the current manuscript must more clearly articulate the novelty of the results. It seems reasonable to conclude that tutoring birds with songs that are faster in tempo than their predisposed (genetically determined) tempo, should have a negative effect on song learning. A similar effect was documented by Podos and Nowicki, 1999, who showed a calibration effect and 'broken syntax' (deviation from tutor song) when tutor song rhythm was too fast. In other words, similarity to tutor song decreases not necessarily because the bird 'failed' to imitate, but because the bird might have actively deviated from the tutor in order to optimize performance. Authors could test for such a scenario in their current data, and minimally they must discuss such an alternative scenario and clarify how their results are novel.

2) There were concerns about the statistical methods. First, in the operant training group authors performed more than one statistical test, but did not account for multiplicity, which pose a risk of false discovery. Further, in one of those tests (Figure 4B), authors report one-tailed t-test, which seems unjustified (one-tailed should only be used in cases where it is impossible to obtain the other tail, which is not the case here if we understand the measures correctly). Given that the one-tailed test gave p<0.02, and that multiple comparisons were not accounted for, the correct p-value should be 0.08.

3) The claims about genetic differences are based a common garden approach which is standard for determining heritability and therefore seems appropriate, and the overall design, was able to remove all possible environmental influences on learning. Nevertheless, there were concerns about the claims for genetic influence that are detailed in the comments of reviewer #3. The authors need not implement further genetic studies but two changes must be implemented to give more confidence in the conclusions. First a pedigree analysis of the birds used in the study should be provided in supplementary figures and claims about genetic difference and the degree of genetic variation in the population should be tempered in light of these pedigrees. Second, while in consultation it was agreed among the reviewers that the authors have done a good job controlling for maternal influence and thus showing that it is not necessary for the observe effects, the possibility of a maternal influence should be discussed.

Reviewer #1:

This study tested for heritable predispositions in song learning outcomes in Bengalese finches. Results show that chicks raised by foster parents imitated their father song less accurately compared to biological offspring. Interestingly, the decrease in song learning can be explained, to a large extent, by differences in tutor's song tempo, a result that authors confirmed in controlled operant song tutoring experiments.

Overall, this is an interesting and well executed study. However, given that authors have recently shown that song tempo is heritable, and that other previous studies showed that rhythm mismatch can decrease imitation (a 'calibration effect'), I found it difficult to evaluate the novelty of the results. Overall, it seems very reasonable to conclude that tutoring birds with songs that are faster in tempo than their predisposed (genetically determined) tempo, should have a negative effect on song learning. A similar effect was documented by Podos and Nowicki, who showed a calibration effect and 'broken syntax' (deviation from tutor song) when tutor song rhythm was too fast. In other words, similarity to tutor song decreases not necessarily because the bird 'failed' to imitate, but because the bird might have actively deviated from the tutor in order to optimize performance. Authors should test for such a scenario in their current data.

Finally, I had a few issues with the statistics: first, in the operant training group authors performed more than one statistical test, but did not account for multiplicity, which pose a risk of false discovery. Further, in one of those tests (Figure 4B), authors report one-tailed t-test, which seems unjustified (one-tailed should only be used in cases where it is impossible to obtain the other tail, which is not the case here if I understand the measures correctly). Given that the one-tailed test gave p<0.02, and that multiple comparisons were not accounted for, the correct p-value should be 0.08.

Reviewer #2:

In a previous study (PNAS 2018) the authors showed that genetic predispositions for vocal learning can be demonstrated in songbirds when birds are tutored using computerized instruction but that this difference can be eliminated when birds are tutored with live tutors.

In this study the authors show that tailoring tutoring to the type of genetic predisposition can overcome potential deficits in learning.

In this study the authors show that matching the tempo of the tutor song with the tempo of the tutee's father song (i.e. the genetically preferred tempo) greatly improves the tutee's overall ability of learning the acoustic features of the song. Tutoring juveniles with songs that are either of a faster or slower tempo than the father's song causes a decrease in overall copying performance when measured as a degree of acoustic similarity of individual syllables.

The paper is extremely well written and clear. The flow is logical and the findings have potentially profound implications on motor skill learning more generally by suggesting that tailored instruction can overcome potential genetic biases.

The authors arguments for a genetic predisposition in learning ability is based on a common garden approach which is standard for determining heritability and therefore seems appropriate. The overall experimental approach is elegant in that juvenile males are isolated from the father at a very early age (36 hours post hatch) significantly before the auditory system fully develops and therefore before they would get the opportunity to hear their fathers' song. In most of the experiments, juveniles are then raised with a "neutral" female before being tutored (at 45 days of age) using a computer-based system that is able to deliver songs at different tempos. The overall design, in my opinion, therefore is able to remove all possible environmental influences on learning and can as such make a strong claim regarding heritability influences.

I do not have any major criticisms of the paper

Reviewer #3:

This fascinating manuscript addresses the interaction of genetic predisposition and experience in determining adult behavioral phenotypes. However, in my view, needed data are lacking to justify the strong claims made about genetics in driving behavioral outcomes.

For example, in the absence of a pedigree chart it is difficult to assess the apparent father-to-son heritability of a genetic predisposition to learn songs of a specific temporal structure. I am puzzled by the absence of any consideration of a maternal contribution to a genetic predisposition to song imitation – perhaps I have misunderstood, but this seems like an issue that the authors should address or acknowledge. A pedigree chart could certainly help sort this out.

The role of genetic factors would also be strengthened by a multi-generational experiment showing the persistence of the genetic predisposition. While the manuscript's single-generation experiments are certainly compelling and clearly sufficient to demonstrate learning, a defining characteristic of a genetic influence is persistence across generations.

Because I am enthusiastic about the research direction described in the manuscript, I would be happy to be corrected if the authors believe that I am in error. Comments regarding specific sections of the manuscript are listed below.

"In agreement with extensive prior research, we found that there was a broad range in the quality of song learning across individuals in our genetically heterogeneous Bengalese finch colony."

  • Could the authors provide references for "extensive prior research" on individual variation in song learning? A number of references are cited in the preceding paragraph, but these papers refer primarily to species or strain differences.

  • Could the authors define what they mean by "genetically heterogenous?" Are there breeding records that establish the degree of genetic heterogeneity within the colony? This seems important, given the role that genetic predisposition is argued to play in the analysis of the findings.

"Prior work indicates that the broad variation in the quality of song learning (Figure 1C) could be influenced by both experiential and genetic factors."

• Seems like there are additional references that could be included here on the role of experience.

"We compared learning for birds that were home-reared and tutored by their genetic fathers, to learning for birds that were cross-fostered and tutored by genetically unrelated adults."

• What does "genetically unrelated" mean here? Were steps taken to insure a specific genetic distance between the adult pairs used in the home-reared and cross-fostered conditions?

Figure 2C: The effect shown in Figure 2C appears to be driven primarily by 3 of the 8 nests.

• Could the authors justify or explain use of a signed-rank test to determine significance, as well as use of median SD scores?

• Throughout the manuscript, reporting the results of statistical tests is reduced to the name of the test and a p-value. Not sure if this is journal preference but seems non-standard.

"Despite exposure to identical computer tutoring experience"

• In the Materials and methods subsection “Computer tutoring”, the authors describe their perch-activated tutoring setup. Because tutor-song playback was contingent on the tutee hopping on the perch, one wonders if all birds were equally active in terms of their perch hopping. Although the number of tutor song playbacks per day was limited to 30, were there days when some birds received fewer because of less perch hopping? It would be helpful if the authors provided some evidence to support the statement that all birds received identical computer tutoring experience.

"Thus, differences in parental behavior and in individual experience with the live tutor could not account for observed variation in the quality of learning."

• Seems like a nest-level analysis would be useful here. That is, the birds were raised to 35-40 days post-hatch by female foster parents. The text of the manuscript seems to assume that rearing by all female foster parents was of equivalent quality. A nest-level analysis of learning outcomes would address this assumption.

Figure 3B: Y-axis.

• Perhaps indicate "Mean Song Divergence (bits)" on the y-axis, since median is used in Figure 2C?

eLife. 2019 Sep 17;8:e47216. doi: 10.7554/eLife.47216.014

Author response

Essential revisions:

The results are exciting and of potential general interest, however, there were some concerns that must be addressed before publication. The written reviews are attached and are not completely consistent but do contain valuable insight into details of any revision plan. In consultation the reviewers reached the following consensus.

1) Given that authors have recently shown that song tempo is heritable, and that other previous work showed that rhythm mismatch can decrease imitation (a 'calibration effect'), the current manuscript must more clearly articulate the novelty of the results. It seems reasonable to conclude that tutoring birds with songs that are faster in tempo than their predisposed (genetically determined) tempo, should have a negative effect on song learning. A similar effect was documented by Podos and Nowicki, 1999, who showed a calibration effect and 'broken syntax' (deviation from tutor song) when tutor song rhythm was too fast. In other words, similarity to tutor song decreases not necessarily because the bird 'failed' to imitate, but because the bird might have actively deviated from the tutor in order to optimize performance. Authors could test for such a scenario in their current data, and minimally they must discuss such an alternative scenario and clarify how their results are novel.

As requested, we have carried out new analyses to test for the specific possibility raised here that poor copying of spectral structure might not reflect a failure to learn (because of a mismatch between genetic bias and tutor song tempo), but rather might reflect an active deviation from good spectral copying in order to optimize temporal aspects of song structure. The analyses that we carried out, detailed below and presented in revised text (Results, last paragraph) and Figure 4—figure supplement 3, argue against the possibility that the effects of genetic bias on learning that we observe could be explained by the kind of ‘trade-off’ or ‘calibration’ effects that are raised in Podos’s work:

As noted by the reviewers, work from Podos and colleagues raises the possibility that within a species there may be a trade-off between optimizing different features of song. For example, female birds might prefer both faster songs and songs with broader spectral bandwidth (and higher levels of spectral complexity), so that male birds have incentive to optimize both of these features. However, in this model, limitations on song production mechanisms create constraints that result in a trade-off between producing optimal values of these features. In support of this possibility, Podos, 1997, reported that within a given species, birds that produce the fastest songs generally have narrower spectral bandwidths, while birds that produce the slowest songs have the broadest spectral bandwidths; they further suggested that birds could 'calibrate' their songs at different positions along this tempo-spectral bandwidth axis, according to their individual capabilities, in order to optimize the learned song (Podos et al., 2004).

These observations from other species raise the possibility that birds in our study might not be failing to copy song spectral content (when tempo is mismatched to individual genetic bias), but, instead, might be actively “choosing” to produce worse spectral copying in order to optimize aspects of song tempo.

We therefore tested two aspects of song tempo that our birds might have been optimizing at the expense of copying song spectral content. First, we tested the possibility raised by Podos and colleagues (Lhati et al., 2011; Podos, 1996; Podos, 1997; Podos et al., 2004) that there is a tradeoff between producing better spectral content of syllables and producing faster songs (as faster songs in some species are construed as more attractive to females). According to this possibility, we would expect that better spectral copying would be associated with slower song tempos, and worse spectral copying with faster song tempos. We found no suggestion of such a relationship in our data (Figure 4—figure supplement 3A).

Second, we tested the possibility that there is a trade-off between better spectral copying and producing songs that match more closely to the tutor song tempo. According to this possibility, we would expect that better spectral copying would be associated with birds that produced a worse match to the tutor song tempo, and worse spectral copying would be associated with birds that produced a better match to the tutor song tempo. In contrast to such a trade-off between matching tutor song spectral content and tutor song tempo, we found the opposite result: birds with learned song tempos that most closely matched the tutor song tempo also produced better copying of spectral content (Figure 4—figure supplement 3B). This result argues against a ‘calibration’ or ‘trade-off’ effect and in favor of the interpretation that the capacity of individual birds to learn spectral content is enabled by a closer match of individual genetic predisposition to the tutor song tempo. These analyses are presented in the main text (Results, last paragraph) and Figure 4—figure supplement 3B.

In addition to addressing the specific issue of ‘calibration’ raised here, we have further elaborated in several places the broader significance of our observations regarding the interaction between individual genetic variation and experience in shaping learning and the “cultural transmission” of song and other complex phenotypes. We particularly emphasize the relevance of our findings to understanding how customization of instruction based on heritable differences across individuals can enhance learning. We also further discuss how the propensity to learn better from some songs (such as that of the father) over others, differs from a common presumption that there is “neutrality” of tutor song stimuli across a broad range of “species typical” songs. The concept of neutrality of tutor songs within a broad species-typical range has been incorporated into modelling of interactions of genes and experience in shaping cultural transmission and speciation (Lachlan, 1999; Lachlan et al., 2018; Lynch, 1996). A general idea that arises in this work is that the presence of a single, species-typical ‘template’ could prevent the drift of song structure over generations that might otherwise arise due to the effects of learning on individuals’ songs. That is, while learning can drive individual variation in song structure, for example across different families, species level genetic constraints ‘pull’ songs back towards species-typical values. In contrast to this prior work, our results show that within a species there may be no single “species-typical” template, and we have added a more explicit discussion of the possibility that individual and familial variation in genetic predispositions within a species could potentially contribute to divergence of song structure across subpopulations within a species (Discussion, second paragraph).

2) There were concerns about the statistical methods. First, in the operant training group authors performed more than one statistical test, but did not account for multiplicity, which pose a risk of false discovery. Further, in one of those tests (Figure 4B), authors report one-tailed t-test, which seems unjustified (one-tailed should only be used in cases where it is impossible to obtain the other tail, which is not the case here if we understand the measures correctly). Given that the one-tailed test gave p<0.02, and that multiple comparisons were not accounted for, the correct p-value should be 0.08.

We have revised our statistical testing in consultation with the biostatistics service of the UCSF Clinical and Translational Science Institute. In an expanded “Statistics” subsection of Materials and methods, we provide a more detailed description of statistical testing, and include additional information regarding directionality of hypothesis testing and corrections for multiple comparisons in the main text and figure legends.

We previously used one-tailed tests in instances where there was a directional hypothesis based on prior results. For example, we had hypothesized prior to the conduct of the cross-fostering experiments that home-reared birds would learn better than cross-fostered birds (rather than that they simply would be different, without regard to directionality). Similarly, for the computer tutoring experiments, we hypothesized that birds would learn better from tutor songs that had tempos matched to their genetic bias than from tutor songs that were unmatched. Our statistical consultancy endorsed the view that in cases such as this, where there is a directional hypothesis, one-tailed tests are appropriate. However, there is not universal agreement about what is the best practice, and we have therefore adopted the reviewer’s suggestion that we use more conservative two-tailed tests for the data in Figure 4B. Regarding multiple comparisons, we now use Holm-Bonferroni corrected p-values for statistical tests in all cases. Use of two-tailed tests versus one-tailed tests, and corrections for multiple comparisons, did not alter the statistical significance of any results of the study,

With respect to the results of Figure 4B: The p values for the one-tailed tests in Figure 4B that we previously reported as p < 0.02, had specific values of p = 0.006 (for comparison of F = T vs. F > T), p = 0.000004 (for comparison of F = T vs. F < T), and p = 0.0032 (for comparison of F < T vs. F>T). All three tests remain significant at p < 0.02 after switching to two-tailed tests and correcting for multiple comparisons using a Holm-Bonferroni correction.

3) The claims about genetic differences are based a common garden approach which is standard for determining heritability and therefore seems appropriate, and the overall design, was able to remove all possible environmental influences on learning. Nevertheless, there were concerns about the claims for genetic influence that are detailed in the comments of reviewer #3. The authors need not implement further genetic studies but two changes must be implemented to give more confidence in the conclusions. First a pedigree analysis of the birds used in the study should be provided in supplementary figures and claims about genetic difference and the degree of genetic variation in the population should be tempered in light of these pedigrees. Second, while in consultation it was agreed among the reviewers that the authors have done a good job controlling for maternal influence and thus showing that it is not necessary for the observe effects, the possibility of a maternal influence should be discussed.

Regarding ‘relatedness’:

Per the request of the reviewers, we have now included a pedigree in Figure 2—figure supplement 1 that illustrates what we know about the relatedness of birds used in the cross-foster study. For the 8 foster nests, we show the origins of both the resident male (tutor) and female traced back to 2001. These data illustrate that the relevant birds were dispersed across different lineages in our colony (in many cases derived from different vendors), but that in a number of instances, birds are more or less distantly related ‘cousins’, and we now provide this characterization in the Materials and methods (subsection “Cross-fostering”). We have also eliminated our too loose usage of the term ‘unrelated’ and instead referred to the relevant populations as ‘home-reared’ and ‘cross-fostered’ (Results, second and third paragraphs).

We also now note the important point that any relatedness between cross-fostered birds and tutors in these experiments would be expected to attenuate the effects that we report (rather than artifactually magnify them). That is, we are testing the hypothesis that shared genes between tutor and offspring, and correspondingly shared predispositions for producing songs with particular structure, enhance learning of the tutor’s song by his offspring, relative to the learning of the same tutor’s song by less closely related cross-fostered birds. To the extent that cross-fostered birds are more closely related to their foster tutors, we would expect that differences in the quality of learning for cross-fostered birds and genetic offspring would be smaller. Conversely, to the extent that we had studied cross-fostered birds that were more distantly related to their tutors (for example derived from other BF colonies) we would expect that our reported effects would be even larger (subsection “Cross-fostering”).

Regarding maternal influences:

With respect to the significance of our reported results, we note that this is a case where any maternal influences on heritable predispositions that we did not take into account would have tended to contribute noise to our estimates of heritability and would therefore have tended to ‘work against us’. Conversely, if we were able to improve our estimates of heritable predispositions by taking into account maternal contributions, we would correspondingly expect that we could explain even more variation in the quality of learning across individuals.

We now have added a brief discussion of these points: “It is noteworthy that our estimates of individual bias derived exclusively from the tempo of the father’s song; hence, a more accurate estimate of bias, incorporating other genetic factors such as maternal contributions, would potentially enable even more effective tailoring of instruction”.

Reviewer #1:

[…] Overall, this is an interesting and well executed study. However, given that authors recently shown that song tempo is heritable, and that other previous studied showed that rhythm mismatch can decrease imitation (a 'calibration effect'), I found it difficult to evaluate the novelty of the results. Overall, it seems very reasonable to conclude that tutoring birds with songs that are faster in tempo than their predisposed (genetically determined) tempo, should have a negative effect on song learning. A similar effect was documented by Podos and Nowicki, who showed a calibration effect and 'broken syntax' (deviation from tutor song) when tutor song rhythm was too fast. In other words, similarity to tutor song decreases not necessarily because the bird 'failed' to imitate, but because the bird might have actively deviated from the tutor in order to optimize performance. Authors should test for such a scenario in their current data.

As is articulated in our response to Essential revisions 1, we have now tested two possible 'calibration effects' and found no evidence for such phenomenon in our study. This suggests that the 'calibration effect' described previously in the work of Podos is supported by different mechanisms than the gene-experience mediated influences on learning that we describe here.

Finally, I had a few issues with the statistics: first, in the operant training group authors performed more than one statistical test, but did not account for multiplicity, which pose a risk of false discovery. Further, in one of those tests (Figure 4B), authors report one-tailed t-test, which seems unjustified (one-tailed should only be used in cases where it is impossible to obtain the other tail, which is not the case here if I understand the measures correctly). Given that the one-tailed test gave p<0.02, and that multiple comparisons were not accounted for, the correct p-value should be 0.08.

See response to Essential revisions 2, above.

Reviewer #3:

This fascinating manuscript addresses the interaction of genetic predisposition and experience in determining adult behavioral phenotypes. However, in my view, needed data are lacking to justify the strong claims made about genetics in driving behavioral outcomes.

For example, in the absence of a pedigree chart it is difficult to assess the apparent father-to-son heritability of a genetic predisposition to learn songs of a specific temporal structure. I am puzzled by the absence of any consideration of a maternal contribution to a genetic predisposition to song imitation – perhaps I have misunderstood, but this seems like an issue that the authors should address or acknowledge. A pedigree chart could certainly help sort this out.

Please see response to Essential revisions 3.

The role of genetic factors would also be strengthened by a multi-generational experiment showing the persistence of the genetic predisposition. While the manuscript's single-generation experiments are certainly compelling and clearly sufficient to demonstrate learning, a defining characteristic of a genetic influence is persistence across generations.

We appreciate that the reviewers, in consultation, decided that our ‘common garden’ approach was appropriate to support our conclusions. We provide below some additional discussion of the rationale for our experimental design.

Parent-offspring regression (as was used in our previous work (Mets and Brainard, PNAS, 2018) to assess heritability of song tempo) and other forms of single or double generation studies (such as twin studies or sibling-sibling comparisons) remain the most common methods for assessing heritability. As reviewer 3 points out, it is now also common to use a different approach based on a full pedigree. This approach, termed the ‘animal model’, uses a linear mixed-effects model and is especially useful in natural settings where environmental variables cannot be controlled experimentally. The estimates of heritability provided by the animal model are fundamentally derived from the same source as parent offspring regression, sib-sib correlation, or twin studies: the correlation between the genetic relatedness of individuals and the phenotype of interest. Pragmatically, there are two main reasons to use the animal model in place of a direct sibling-sibling or parent-offspring comparison. First, the animal model takes into account the relatedness of all individuals in a population, giving better statistical power than can be obtained in a similar population using only parent-offspring regression. Second, this approach allows statistical control over measured parameters (e.g. sex, water availability, food availability, mean temperature, etc.…) when these are not under experimental control. These two features of the animal model are particularly useful in wild populations where experimental control is impossible and the number of animals in a study may be limited. However, this approach has a limitation which is particularly problematic in the context of culturally learned behaviors generally, and song learning in particular. For phenotypes that are influenced by the phenotype of the parent through experience (trans-generational environmental effect) and where environmental influences are not experimentally controlled across the population, heritability can be over-estimated; an individual may behave in a way that is related to his cousin not because of the genetic similarity between those individuals, but because they have the same grandfather whose behavior was passed on to his offspring (and then their offspring) through cultural transmission. Therefore, without the ‘common garden’ approach used here, estimates of heritability for song phenotypes would be exaggerated.

Furthermore, the mixed-effects framework only allows statistical control of known and measured environmental variables whereas the 'common garden' approach controls for all environmental variables both known and unknown. Finally, the linear-mixed effects framework used in the ‘animal’ model only allows statistical control of environmental variables that exert a linear, additive influence on the phenotype of interest. We know from our previous work that some relevant environmental parameters have non-linear influences on song phenotypes (Mets and Brainard, 2018). We therefore felt the application of the animal model to our large pedigree could erroneously over-estimate the genetic influences on song and, instead, designed our experiments to directly control experiential factors.

Because I am enthusiastic about the research direction described in the manuscript, I would be happy to be corrected if the authors believe that I am in error. Comments regarding specific sections of the manuscript are listed below.

"In agreement with extensive prior research, we found that there was a broad range in the quality of song learning across individuals in our genetically heterogeneous Bengalese finch colony."

• Could the authors provide references for "extensive prior research" on individual variation in song learning? A number of references are cited in the preceding paragraph, but these papers refer primarily to species or strain differences.

We agree that there is a relatively small amount of research specifically addressing variation in the quality of song learning under these conditions and, since this is not a major point, we have simply removed the clause 'In agreement with extensive prior research'.

• Could the authors define what they mean by "genetically heterogenous?" Are there breeding records that establish the degree of genetic heterogeneity within the colony? This seems important, given the role that genetic predisposition is argued to play in the analysis of the findings.

We now provide a pedigree for the birds included in the cross-fostering experiments (Figure 2—figure supplement 1). While this does not include the entire colony, the pedigree illustrates the general nature of the genetic diversity in our colony. Please see Essential revisions 3 for further details.

"Prior work indicates that the broad variation in the quality of song learning (Figure 1C) could be influenced by both experiential and genetic factors."

• Seems like there are additional references that could be included here on the role of experience.

We have added several additional references.

"We compared learning for birds that were home-reared and tutored by their genetic fathers, to learning for birds that were cross-fostered and tutored by genetically unrelated adults."

• What does "genetically unrelated" mean here? Were steps taken to insure a specific genetic distance between the adult pairs used in the home-reared and cross-fostered conditions?

We have now provided a pedigree to clarify the relational structure of individuals in this experiment (Figure 2—figure supplement 1). We have also changed the wording throughout to remove the use of ‘unrelated’, and have simply noted that we are comparing ‘home-reared’ and ‘cross- fostered’ animals. Please see Essential revisions 3 for further details.

Figure 2C: The effect shown in Figure 2C appears to be driven primarily by 3 of the 8 nests.

• Could the authors justify or explain use of a signed-rank test to determine significance, as well as use of median SD scores?

There are several uncontrolled variables across nests that could potentially influence the measured quality and range of learning in a nest-specific fashion. These include, for example, the specific acoustic structure of the tutor song and how often it was produced. Such variation is likely to contribute to nest-specific differences in the magnitude or range of learning quality (SD scores). We therefore designed our experiments to enable paired comparisons between cross-fostered and home-reared birds within nests (in order to control for any nest-specific effects) and felt it appropriate to test the sign rather than the magnitude of differences between groups. We used median values to characterize learning within groups because the distribution of SD scores was γ, rather than Gaussian. We have confirmed that the significance of the results are unaltered if the mean is used instead of the median and/or if other statistical tests are used instead of a sign-rank test. These points are now noted in the second paragraph of the Results section and in the subsection “Statistics”.

• Throughout the manuscript, reporting the results of statistical tests is reduced to the name of the test and a p-value. Not sure if this is journal preference but seems non-standard.

We have expanded the Materials and methods section on statistics to provide further detail. Please see Essential revisions 2.

"Despite exposure to identical computer tutoring experience"

• In the Materials and methods subsection “Computer tutoring”, the authors describe their perch-activated tutoring setup. Because tutor-song playback was contingent on the tutee hopping on the perch, one wonders if all birds were equally active in terms of their perch hopping. Although the number of tutor song playbacks per day was limited to 30, were there days when some birds received fewer because of less perch hopping? It would be helpful if the authors provided some evidence to support the statement that all birds received identical computer tutoring experience.

We have now analyzed the perch hopping logs for each bird to provide a more quantitative assessment of the amount of tutoring that each bird received and to examine any potential effects on learning. We confirmed that all birds received a similar amount of tutoring, and that we could not detect any relationship between variation in amount of tutoring and variation in the quality of learning. This is summarized in the following addition to the description of computer tutoring: “It sometimes took birds a few days to begin consistently actuating song playbacks. Nevertheless, all birds elicited at least 90% of the available song playbacks during the tutoring period (median = 96%), and there was no relationship between the number of playbacks that an individual heard and the quality of song learning.”

"Thus, differences in parental behavior and in individual experience with the live tutor could not account for observed variation in the quality of learning."

• Seems like a nest-level analysis would be useful here. That is, the birds were raised to 35-40 days post-hatch by female foster parents. The text of the manuscript seems to assume that rearing by all female foster parents was of equivalent quality. A nest-level analysis of learning outcomes would address this assumption.

We have expanded the Materials and methods section covering female fostering to address this point. Specifically, we have added: “While the female foster mothers do not sing, they produce unlearned calls and influence other aspects of a hatchling’s experience, including feeding and social interactions, that could plausibly influence aspects of hatchling development and the quality of subsequent computer driven song learning. […] In order to prevent any ‘batch effects’, only a single egg/hatchling was raised at a time in any given foster nest.”

Figure 3B: Y-axis.

• Perhaps indicate "Mean Song Divergence (bits)" on the y-axis, since median is used in Figure 2C?

The previous figure and legend were confusing with respect to which SD scores reflected individual bird values versus group values. We have now added labels and legend to clarify that the SD values in 3B and 3C are values for individual birds, and that the summary values in 3D indicate both medians (gray bars) and means (red points).

Figure 3B reports both the mean (gray bars) and the median (red dots) SD for each of the three groups depicted. We have therefore left the label unchanged. We have modified the legend to clarify

Associated Data

    This section collects any data citations, data availability statements, or supplementary materials included in this article.

    Supplementary Materials

    Source code 1. Tutoring software.
    elife-47216-code1.gz (937.9KB, gz)
    DOI: 10.7554/eLife.47216.011

    Data Availability Statement

    All data generated or analysed during this study are included in the manuscript and supporting files.

    Articles from eLife are provided here courtesy of eLife Sciences Publications, Ltd

    RESOURCES