Integration of genomic and ecological methods inform management of an undescribed, yet highly exploited, sardine species

Abstract

Assessing genetic diversity within species is key for conservation strategies in the context of human-induced biotic changes. This is important in marine systems, where many species remain undescribed while being overfished, and conflicts between resource-users and conservation agencies are common. Combining niche modelling with population genomics can contribute to resolving those conflicts by identifying management units and understanding how past climatic cycles resulted in current patterns of genetic diversity. We addressed these issues on an undescribed but already overexploited species of sardine of the genus Harengula. We find that the species distribution is determined by salinity and depth, with a continuous distribution along the Brazilian mainland and two disconnected oceanic archipelagos. Genomic data indicate that such biogeographic barriers are associated with two divergent intraspecific lineages. Changes in habitat availability during the last glacial cycle led to different demographic histories among stocks. One coastal population experienced a 3.6-fold expansion, whereas an island-associated population contracted 3-fold, relative to the size of the ancestral population. Our results indicate that the island population should be managed separately from the coastal population, and that a Marine Protected Area covering part of the island population distribution can support the viability of this lineage.

1. Introduction

The current biodiversity crisis is leading to the decrease in population size and the extinction of some species [1,2], threatening the balance of ecosystems [3,4]. The Convention on Biological Diversity (CBD) recognizes genetic diversity within species as one of the three main levels of biodiversity that must be monitored and preserved, along with species and ecosystem diversity [5]. This is because genetic diversity within species is the heritable variability upon which selection can act, and thus underlies the ability for a species to adapt and persist in changing environments [6]. Although the CBD set a target to maintain and restore genetic diversity within species until 2030 in the Kunming–Montreal Global Biodiversity Framework, monitoring changes of genetic diversity within species is challenging because it requires understanding how past and current environment affect the evolution of intraspecific lineages. This understanding is especially challenging in marine ecosystems, particularly in hyper-diverse tropical regions of the globe, where species are still being described, their relative abundances and connectivity are unknown, and yet some species are under strong human exploitation. Conflicts of interest between resource users and conservation managers start at a local scale owing to environmental policies that disregard scientific data and users' participation in management [7]. These local conflicts can, however, affect economy, food availability and ecosystem functioning at a broader spatial scale, demanding updated scientific information to support better conservation laws and practices. The integrations of genomic data [8–10] and ecological modelling [11,12] can potentially provide evidence-based suggestions for conservation plans and contribute to resolve such conflicts.

Determining the number of intraspecific lineages within a species of concern is an important task in conservation biology, as evolutionarily independent units often require independent management plans [13]. Likewise, assessing relative levels of genetic diversity within each intraspecific lineage (N_e or effective population size) is important because it reflects both the differences in current number of individuals (N or census size) and the demographic history of these lineages in response to past environmental change [14]. Both the number of intraspecific lineages and their genetic diversity depend on evolutionary processes such as genetic drift in isolated populations, selection in divergent environments and gene flow among them. Because such processes are highly dynamic over space and time, in order to aid conservation planning, the integration of genomic analyses and ecological niche modelling can provide fast and useful information regarding the drivers of population size and connectivity of exploited species. These processes are well understood in terrestrial systems [15–17], yet only recently have they begun to be studied in marine systems [18]. Despite the high potential dispersal ability of marine taxa, recent studies have shown that several oceanographic barriers, such as oceanic currents and depth, constrain gene flow [19], therefore enabling diversification within species. Understanding how such oceanographic barriers condition the number of management stocks, their current levels of genetic diversity, and genetic connectivity remain important tasks in conservation biology.

Marine systems support the livelihood of millions of people worldwide, increasing the conflict between local communities and conservation strategies. The application of genomics to fisheries management has contributed in resolving some of these conflicts by identifying the number of stocks within species of commercial interest, their levels of genetic connectivity, and relative measures of diversity [20]. Yet, such efforts have been mainly restricted to marine species from the Northern Hemisphere, where species-level diversity is lower relative to tropical regions. For example, genomic studies in the yellow-fin tuna and cod have helped identifying divergent intraspecific lineages as distinct fishery stocks with high genetic differentiation that require independent management [21,22]. Genomic data on species of conservation concern, such as cod, herring and European hake, have been used to identify the fishing origin of commercial products, and thus aid against illegal, unreported and unregulated fisheries [23]. A study in yellow-fin tuna indicates that the populations occurring in different oceans present significant and asymmetric gene flow [24], allowing the identification of populations that act as source or sink of migrants, another factor relevant for defining conservation priorities. Studies in a pipefish [19] and in the Arctic charr [25] have quantified changes in effective population size (N_e) of different stocks, potentially driven by both their past evolution and more contemporary exploitation by fisheries. Much less is known about species in tropical waters, even though such species play an indispensable role in the subsistence of local communities [26].

One such economically important group of fishes worldwide is the Clupeidae, a family that includes several species of forage fishes popularly known as sardines, herrings and shads, among others [27,28]. The genus Harengula Valenciennes 1847 comprises four taxonomically recognized species, three of which are reported as occurring in the western Atlantic. Whereas Harengula humeralis (Cuvier, 1829) is anatomically distinct and restricted mostly to the central Atlantic, Harengula clupeola (Cuvier, 1829) and Harengula jaguana Poey, 1865 are reported as having a much broader distribution, from the US coast to southern Brazil [27]. However, recent molecular studies indicate that specimens of Harengula off the Brazilian coast belong to a third mitochondrial lineage, which is more divergent (3.2% in the barcoding gene CO1) than H. clupeola and H. jaguana (1.8%) [29]. The lineage of Harengula likely endemic to the Brazilian Biogeographic Province, herein Harengula sp., is apparently allopatric relative to congeners and is the only member of the Clupeidae present in two oceanic archipelagos of the South Atlantic (Fernando de Noronha (FNO) and Trindade-Martin Vaz; [30]), suggesting a higher tolerance to ecological conditions than other overall similar species of sardines [31]. Although Harengula sp. remains unrecognized taxonomically, it shows signs of decline in population size [32] and likely requires specific management measures [33,34].

Harengula sp. is consumed as part of traditional dishes and is the main source of live bait in several regions of the Brazilian Exclusive Economic Zone [35]. Currently, there is no specific management strategy for Harengula sp. in Brazil, except in Marine Protected Areas (MPAs) and their buffer zones. The fishery of this species within the MPA of the FNO archipelago led to conflict between environmental managers, who established a no-take zone within the MPA in the 1980s, and local fishers, who claimed the right to fish in this area [36]. In an attempt to resolve this conflict, the Brazilian government lifted restrictions to exploit sardine in the no-take zone of FNO MPA, initially from 2020 to 2022, and later extended this lift of restrictions until 2024 [37]. Such conflicts have an impact in the local economy and fishery strategies of the island. To date there is no scientific evaluation of how the island population is divergent from the coastal population, what is their degree of genetic connectivity, or if there are temporal and spatial changes in genetic diversity (N_e).

Here, we addressed these issues by combining ecological niche models and population genomic methods in this highly exploited species of sardine that remains largely unknown by scientists and legislators. First, we used ecological niche models to establish predictions on how temporal and spatial changes in habitat suitability have constrained the distribution of Harengula sp. relative to other formally recognized species of the genus, and how these changes have affected divergence and gene flow within Harengula sp. Second, we used genomic data to test how those oceanographic barriers led to intraspecific divergence. Finally, we used demographic modelling to estimate the magnitude and direction of gene flow between independent lineages and changes in N_e. These results provide science-based information to support better management strategies focused on the sustainable exploitation of this species.

2. Material and methods

(a) . Habitat dynamics

To understand the environmental drivers of the distribution of Harengula sp., and how these have changed over time, we fitted an ecological niche model (ENM) in the current climate, and then projected it to the Last Glacial Maximum (LGM), approximately 21 thousand years ago. We used a machine-learning algorithm of maximum entropy (MaxEnt v. 3.4.4; [38]) in the dismo package 1.3–5 in R v. 4.0.5 [39,40]. MaxEnt default modelling parameters were employed (e.g. beta multiplier and feature classes) to perform the models, which were then evaluated according to area under the receiver operating characteristic (ROC) curve (AUC) [41,42].

We retrieved 227 georeferenced occurrence records (latitude and longitude) of ‘Harengula clupeola’ and ‘H. jaguana’ in the Brazilian Exclusive Economic Zone (EEZ) from online databases and from our fieldwork collection [43,44] (electronic supplementary material, table S1). Since a previous study showed that Harengula sp. is the only species occurring in the Brazilian EEZ [29], we re-classified these observations as Harengula sp. (figure 1). To increase accuracy in our database, we only considered georeferenced records that refer to specimens collected after 1945 [46] and that are available as vouchers in biological collections.

Figure 1.

Observation of Harengula sp. in the southwest Atlantic. Observations are marked by asterisks (occurrence record only), or by coloured circles for sampling sites included in the genomic analyses. Arrows indicate the direction of main oceanic currents. Ecoregions are marked as in Spalding et al. [45]. Sampling localities: FNO, Fernando de Noronha archipelago (oceanic island); CE, Ceará; RN, Rio Grande do Norte; PB, Paraíba; PE, Pernambuco; AL, Alagoas; BA, Bahia; ABR, Abrolhos (continental island); ES, Espírito Santo; RJ, Rio de Janeiro, SP, São Paulo; SC, Santa Catarina.

We downloaded four abiotic layers characterizing present and LGM climatic and geophysical aspects of the environment from MARSPEC, using the sdmpredictors package in R [40,47]. The layers used in the models were: bathymetry (or depth; m), slope (degrees), mean of sea surface salinity (psu) and mean of sea surface temperature (°C) [48,49]. We chose these variables because they (i) are of general relevance for species of the Clupeidae [27], (ii) show low correlation between them (Pearson's r < 0.8), and (iii) are available for both present and LGM scenarios. To better delimit the area of occurrence of this species and to avoid model overfitting due to differences between occurrence points and background points [50], we cropped the layers into provinces as defined by Spalding et al. [45]. See electronic supplementary material for details.

(b) . Molecular analyses

We collected 92 specimens from 12 sites representing most of the known distribution of Harengula sp., 10 of which are closely associated with the Brazilian coast, one is located at the coastal island of Abrolhos (ABR), and another is located in the oceanic archipelago of Fernando de Noronha (FNO) (figure 1; electronic supplementary material, table S1, licences SISBIO 76 053-1 and 74 802-1). Samples were either collected using a beach seine or bought in local fish markets. Muscle samples were preserved in 100% ethanol and sent for DNA extraction, sequencing and single nucleotide polymorphism (SNP) genotyping at Diversity Arrays Technology (DArT), following the automated DArTseq method, hereafter DArTseq. DArTseq is a cost-effective approach for SNP discovery using a reduced representation of the genome [51]. It is similar to the double digest restriction-site associated DNA sequencing (ddRAD) method as it uses one restriction enzyme that recognizes a common sequence motif, and another that recognizes a rarer motif. By linking sequencing adaptors to genomic DNA fragments that have both restriction sites and are within a target fragment size, we recover thousands of sequencing tags (or loci) scattered randomly along the genome, which contain one or more SNPs each [52]. This methodology does not require a reference genome for identifying thousands of SNPs in homologous sites across individuals and hence is widely applicable to undescribed species lacking such resources.

From the unfiltered dataset, we removed two individuals with an excessive amount of missing data (greater than 98%), and 166 loci that are putatively sex-linked (electronic supplementary material, figure S1). We further filtered SNPs considering the assumptions of each analysis (electronic supplementary material, table S2), using the package dartR in R [53]. For the diversity analyses, we generated a ‘linked-SNPs’ dataset, which contained all SNPs across loci, irrespective of the amount of missing data, and a ‘linked-SNPs-0MD’ dataset removing all SNPs containing missing data. For population structure analyses and divergence, we generated an ‘unlinked-best-SNPs’ dataset containing the most informative (i.e. higher frequency) SNP per locus. Finally, for the demographic analyses, we generated an ‘unlinked-random-SNPs' dataset, which retained a random SNP per locus, irrespective of its frequency. See electronic supplementary material, table S2 for details on the DArTseq service and data filtering details.

(c) . Population structure and genetic diversity

To estimate the number of intraspecific lineages of Harengula sp. we used two approaches that differ in their model assumptions. Because both approaches assume no linkage between SNPs within loci, we used the unlinked-best-SNPs dataset. First, we performed a non-model-based principal component analysis (PCA) to visualize the genetic similarities between individuals, using the dartR package with the file in the genlight format [53]. Second, we estimated the most likely number of clusters using the software ADMIXTURE v. 1.3.0 [54], which estimates individual ancestry by maximizing Hardy–Weinberg and linkage equilibria within K ancestral clusters. For this, we converted the genlight file to plink using the gl2plink function [53]. We considered models with K from 1 to 12 (total number of sampling sites), and used 10 independent runs for each K. The most likely number of clusters was estimated based on the lowest cross-validation error which indicates the predictive accuracy of each model [55].

To estimate the level of genetic divergence within Harengula sp., we used the same unlinked-best-SNPs dataset to calculate pairwise genetic differentiation (F_ST; [56]) between all sampling sites in R (function gl.fst.pop) [40,53]. Finally, we tested for isolation by distance within the coastal genetic cluster using the Mantel test in the dartR package (function gl.ibd), accessing its significance using 999 permutations [53,57,58].

To estimate genetic diversity within individuals, sampling sites and clusters, we used the linked-SNPs dataset and calculated nucleotide diversity based only on variant sites (π-SNP). In contrast with the traditional measure of nucleotide diversity (π; [59]), which computes the average number of nucleotide differences across sequences with variant and invariant sites, π-SNP specifically assesses nucleotide diversity in variant sites only. We estimated π-SNP using pixy v. 1.2.7.beta1 [60]. For this, we converted the genlight object to vcf using plink v. 3 and the gl2vcf function [53]. Using the linked-SNPs-0MD, we calculated inbreeding coefficient (F_IS), i.e. the level of heterozygosity of an individual relative to the heterozygosity observed in its cluster or population, using the basic.stats function from the hierfstat package [61] in R. It is important to note that diversity summary statistics assume that each group conforms to the expectations for Wright–Fisher population, and thus that allelic frequencies are not influenced by gene flow between populations nor by changes in population size. In contrast, the algorithm used in ADMIXTURE and in the following demographic analysis considers possible gene flow between clusters.

(d) . Demographic history

To test whether the observed genetic variation significantly deviates from the expected variation under neutrality (e.g. due to demographic change or to selection), we calculated Tajima's D [62] for each identified cluster, using DnaSP 6 [63] and the linked-SNPs-0MD dataset because this software does not take missing data into consideration.

To estimate the magnitude and direction of gene flow between population clusters, we used a diffusion approximation method implemented in δaδi v. 2.0.5 [64] and the δaδi-pipeline [65]. Because this method is based on the site frequency spectrum, we estimated it based on the dataset unlinked-random-SNPs, which contains 34% missing data, considering variant sites only. We converted the data to vcf using the function gl2vcf. We reduced the amount of missing data by projecting down the number of SNPs and individuals, choosing a projection that maintained a relatively large number of segregating sites and individuals. We tested three simpler demographic model scenarios where change in effective population occurs during population splitting: (i) ‘no migration’, with three parameters: effective size of population 1 (N_e1), size of population 2 (N_e2) and time since split (T); (ii) ‘symmetric migration’, with a fourth parameter reflecting migration (2N_em); and (iii) ‘asymmetric migration’, with two migration parameters in opposing directions (2N_em1 > 2 and 2N_em2 > 1). Additionally, we tested three more complex models allowing additional changes of effective population size after population splitting (T2): (iv) ‘symmetric migration with size change in one population’, with six parameters (N_e1a, N_e1b, N_e2, 2N_em, T1 and T2); (v) ‘symmetric migration with size change in both populations', with a seventh parameter describing change in the second population (N_e2b); and (vi) ‘asymmetric migration with size change in both populations’, with an eighth parameter describing different migration rates (2N_em1 > 2 and 2N_em2 > 1). Since mutation rate and generation time are unknown for this species, we were not able to estimate absolute parameter values. Therefore, all parameters were estimated relative to the effective population size of the ancestral population [64]. We selected the model with a combination of lowest Akaike information criterion (AIC) and higher parameters convergence, accounting for the different number of parameters in competing models [66,67]. To verify how well the selected model and estimated parameters fitted our data, we visually inspected the residuals between empirical and model site frequency spectra (SFSs) plotted using the δaδi-pipeline [64]. We considered the parameter values estimated under the model and replicate with lowest AIC score.

3. Results

(a) . Habitat dynamics

Our ecological niche model (ENM) for Harengula sp. shows good ability to discriminate between areas where the species occurs and areas where it does not occur (AUC 0.964, s.d. 0.017). Bathymetry and mean sea surface salinity showed the highest percentage contribution to model performance, of 73.5 and 23.7%, respectively (electronic supplementary material, figure S2 and table S3).

The ENM for the present (figure 2b) shows continuous, although heterogeneous, habitat suitability along ca 6100 km of the Brazilian continental shelf, bounded by the Amazon–Orinoco Barrier in the north and the La Plata River in the south. The suitable habitat at the oceanic archipelagos of FNO (included in the genetic sampling of this study) and of Trindade-Martin Vaz (not sampled here) are separated from the coast by approximately 400 and 1000 km of unsuitable habitat, respectively. Seamounts between these islands and the coast provide patchy suitable area. All the islands located on the continental shelf, such as Abrolhos (ABR), are connected to the coast by highly suitable habitat.

Figure 2.

Ecological niche model for the occurrence of the scaled sardine Harengula sp. in (a) the Last Glacial Maximum (LGM, approx. 21 ka) and (b) present climates. The magnified area in the upper inset shows suitability among seamounts connecting the oceanic archipelago of Fernando de Noronha (FNO) to the Brazilian coast in the LGM, and that in the lower inset highlights the increase in habitat suitability at the continental shelf Abrolhos Island (ABR). The contour of the models reflects marine provinces as described in Spalding et al. [45].

The ENM for the LGM (figure 2a) shows that the suitable habitat was narrower than in the present, although still continuous along the Brazilian coast. Habitat suitability at the oceanic archipelagos of FNO and Trindade-Martin Vaz was similar to the present model. Islands in the continental shelf had lower suitability than present and were exposed owing to the lower sea level.

(b) . Molecular analyses

Our unfiltered dataset comprised 91 individuals genotyped for 66 639 loci, with 98 313 binary SNPs (on average, two SNPs per locus), a total of 30.35% missing data and an average read depth across all markers of 20 reads per locus. Individuals had an average of 28.84% missing data, while SNPs had 69%. Five individuals had more than 40% missing data and were excluded from downstream analyses. After filtering for genotype call consistency (call rate) and locus quality (reproducibility) (electronic supplementary material, table S2), we kept 86 individuals genotyped for 62 295 loci, containing 88 639 SNPs and 29.5% missing data (linked-SNPs dataset). Further removal of missing data retained 9369 SNPs in the linked-SNPs-0MD dataset. The unlinked-best-SNPs and unlinked-random-SNPs datasets had monomorphic loci removed, resulting in each dataset consisting of 59 992 SNPs with 34% missing data.

(c) . Population structure and genetic diversity

The first dimensional axis of the PCA (figure 3b) explains 4% of the genetic variance and separates individuals from the coast of Brazil (hereafter ‘coastal’ population), including those from Abrolhos Island on the continental shelf, from individuals collected in the oceanic FNO archipelago (hereafter ‘island’ population). The second dimensional axis explains 1.8% of the variance and further divides the individuals from the coastal population latitudinally.

Figure 3.

Genomic divergence and gene flow between intraspecific lineages of the scaled sardine Harengula sp. (a) ADMIXTURE analyses assuming two ancestral clusters (model with highest likelihood). (b) Principal component analysis. (c) Competing demographic models depicting divergence between the coastal and island clusters. The coloured model is most likely to fit the data according to Akaike information criterion (electronic supplementary material, table S7), but parameters are not to scale. (d) Demographic parameters estimated using the best demographic model (symmetric migration). The width of the population boxes reflects effective population (N_e) relative to the ancestral N_e (anc). For explanation of abbreviations for site names, see figure 1.

The ADMIXTURE analysis shows the highest likelihood for the model assuming two ancestral clusters (K = 2, figure 3a), which separated all the island individuals from all the coastal individuals. Individuals from the island do not share ancestry with the coastal population. Conversely, individuals from the coast share some ancestry with the island population, with the population closest to the island CE showing up to 32.4% of island ancestry, and the furthest population (SC) showing no island ancestry.

Genetic differentiation (F_ST) between the island population and every locality from the coastal population was significant and above 0.17 (electronic supplementary material, table S4); F_ST between the island cluster and the coastal cluster was 0.14 (p < 0.05). Between coastal sites, F_ST was relatively low and was higher between the extremes of the costal distribution (CE–SC: 0.04, p < 0.05; electronic supplementary material, table S4). Consequently, there is significant isolation by distance in the coastal population (y = 0.0054 + 7.1×10^−09x, R² = 0.318, p = 0.001; figure 4a).

Figure 4.

Genetic diversity of Harengula sp. (a) Isolation by distance, considering mainland coastal sites only (excluding FNO), and (b) individual nucleotide diversity (π-SNP) grouped per sampling site (mainland coastal sites in orange and oceanic island in dark grey). For explanation of abbreviations for site names, see figure 1.

Estimates of individual nucleotide diversity estimated only using variant sites (π-SNP) ranged from 0.047 to 0.069 for the coast, and from 0.044 to 0.054 for the island (figure 4b). The sampling sites at the known extremes of the species distribution showed average levels of diversity. Overall π-SNP for the coast was 0.06 while in FNO it was 0.05 (figure 4b; electronic supplementary material, table S5). The inbreeding coefficient (F_IS) was 0.033 in FNO and ranged from 0.024 to 0.109 for coastal sampling sites (electronic supplementary material, table S6 and figure S3).

(d) . Demographic history

Values of Tajima's D were negative for every coastal locality (between −0.24 and −0.85; electronic supplementary material, table S5), as well as for the whole coastal cluster considered together (−1.50, p > 0.1). Tajima's D was also negative for the FNO population, yet closer to neutral expectations (−0.29, p > 0.1). These results are consistent with deviations from the assumptions from a Wright–Fisher population [68]. Some such deviations, such as change in population size and migration, are accounted for in the demographic models below.

After down-projecting the SNP data, we considered 19 905 SNPs, 62 individuals from the coastal populations and seven individuals from FNO. All models including gene flow have lower AIC score than the model without gene flow (electronic supplementary material, figure S4). The lowest AIC values were observed for both the ‘symmetric migration’ and ‘asymmetric migration’ models (electronic supplementary material, table S7; figure 3c). However, only the simpler model of ‘symmetric migration’ showed a complete convergence of the estimated parameters across replicates and presented low residuals (electronic supplementary material, figures S4 and S5). N_e estimation indicated that the population from the coast experienced an expansion after splitting (N_e1 = 3.597 relative to the ancestral population), while the island population experienced a contraction (N_e2 = 0.324; electronic supplementary material, table S7; figure 3d). The estimated symmetric migration (2N_em) is 2.785.

4. Discussion

(a) . How does habitat suitability generate intraspecific divergence?

Ecological niche models contribute to uncovering environmental corridors and locations of potential population settlement. For species with high dispersal ability, like fishes, these models identify potential corridors connecting distantly located habitats, as well as zones of environmental inadequacy isolating them, aiding the detection of population structure that can be tested with other tools, such as genomics. Our ecological models indicate that depth and salinity explain most the distribution of Harengula sp. (contribution is 73.5 and 23.7%, respectively; electronic supplementary material, table S3). The present-day model (figure 2b) shows a continuous suitable habitat covering the continental shelf of most of the Brazilian coast, whereas suitable habitats in the two oceanic archipelagos (i.e. Fernando de Noronha (FNO) and Trindade-Martin Vaz) are disconnected from the coast. During the LGM, chains of seamounts provided suitable habitat that connected the coast to the islands (inset; figure 2a); such a putative ecological corridor has low relative suitability in the present-day model. Studies on co-distributed species dependent on shallow waters, such as parrotfishes [69], octopuses [70], moray eels, mullets and other animals [30], highlight the importance of seamounts for the colonization of oceanic islands, during glacial periods of lower sea level.

Our results also reveal the effectiveness of oceanographic barriers in the evolution of Harengula sp. The population structure analyses indicate that Harengula sp. is structured into a coastal population, including the Abrolhos Island on the continental shelf, and into an island population at FNO (figure 3a,b). Some co-distributed coastal species likely originally from the Brazilian coast that colonized FNO also show this genetic divergence, as in the case of corals [71], lobster [72] and the reef-associated rockpool blenny [73]. However, other species that colonized FNO do not show this pattern of divergence, such as parrotfishes [74], octopuses [70] and snappers [75], indicating that species occurring in similar habitats can have different evolutionary responses to the same oceanographic barriers. Future studies can clarify if such idiosyncratic patterns of genetic divergence are related to dispersal traits, such as the duration of larval stage, type of reproduction and body size.

In accordance with our ENM showing continuous suitability along the 3700 km of the coastal range sampled here, we find low genetic differentiation between coastal localities (F_ST < 0.04; electronic supplementary material, table S4) and significant isolation by distance (figure 4a). These levels of genetic differentiation are similar to values found in other pelagic species with large dispersal capacity, such as the dog snapper (F_ST = 0.037) [75]. Thus, our results suggest that continuous habitat allows gene flow along the coastal range, but limited dispersal rate of this species also resulted in clinal divergence along the coast.

(b) . How does demographic history condition current levels of genetic variability?

We tested if changes in habitat suitability during glacial cycles are consistent with temporal changes in effective population size (N_e), a measure of genetic variability. Ecological and demographic models of Harengula sp. show congruent scenarios. Our ENMs suggest that the suitable habitats for Harengula sp. strongly increased at the coast but remained stable at the oceanic islands, including the sampled archipelago FNO (figure 2). Similar spatial patterns have been detected for multiple co-distributed coastal species [76,77], raising the hypothesis that marine species restricted to shallow habitats might have contrasting demographic histories at oceanic islands relative to the coast, a hypothesis that is supported by our demographic analyses of Harengula sp.

Our demographic modelling shows that a relatively simple model of split between coastal and island populations with instantaneous change in population size, followed by symmetric gene flow (i.e. ‘symmetric migration’ model with four parameters), is the best model explaining the demographic history of these populations and is a good representation of the observed genomic data (electronic supplementary material, figure S5). This model, along with other models allowing gene flow, clearly rejects a ‘no migration’ model (electronic supplementary material, figure S4), showing that, after the colonization of the oceanic island from the coast (i.e. population split), the two divergent sardine populations evolved in the presence of genetic connectivity across the oceanographic barriers described above. More complex models including asymmetric migration or a second change in population size do not show convergence in parameter estimates and AIC scores across replicates, and therefore could not be interpreted (electronic supplementary material, figure S4). This genetic connectivity is congruent with our ENM for Harengula sp., suggesting that seamount chains between the Brazilian coast and the oceanic archipelago of FNO might act as an effective ecological corridor between differentiated populations, particularly during the glacial periods, when the sea level was lower (figure 2).

Under the best model of ‘symmetric migration’, we find that the coastal population experienced an expansion to 3.6 times the size of the ancestral population, while the island population experienced a contraction to one-third of the ancestral population (electronic supplementary material, table S7). Such contrasting demographic histories recapitulate the results from our ENM, suggesting that the observed changes in genomic variability and effective population size (N_e) reflect long-term evolutionary processes associated with the glacial cycles rather than shorter-term processes associated with human induced over-exploitation. The magnitude of the gene flow (2N_em, or the effective number of gene migrations received by a population per generation) detected is also surprising. In the absence of selection, 2N_em values equal to or higher than 1 prevent populations from accumulating or maintaining divergence [68]. Therefore, finding higher values of 2N_em between populations that maintain genomic differentiation implies that selection, at least in part of the genome, must counteract gene flow [78]. In Harengula sp. we find that 2N_em is 2.7, suggesting that selection is maintaining divergence between coastal and island populations in the face of such high migration rates. Yet, it is unclear which forms of selection could maintain divergence between these populations. Interestingly, Harengula sp. is the only species among more than 400 species of the Clupeiformes that occurs in the archipelago of FNO [30], a situation that indicates that oceanic islands in Brazil are generally unsuitable for sardines and herrings. Future studies (of e.g. diet, morphology, physiology, development) are necessary for understanding which factors are associated with divergence and persistence of the oceanic lineage of Harengula sp. in such an unusual habitat. In addition, increasing the genomic data by using high-coverage whole-genome sequencing can provide insights about underlying genetic adaptation in the island population.

(c) . How does evolutionary genomics provide insights into conservation and management of fishery stocks?

Previous studies using catch data of Harengula sp. from the coast of Brazil have shown that the species already presents signs of overexploitation [32], suggesting the need for protection measures. Yet, it was unclear to what extent the coastal population is representative of the whole species range, making it difficult to establish recommendations for sustainable exploitation of this species in its entire distribution. Our study provides the first evidence that this undescribed, yet exploited, sardine species encompasses one unique coastal lineage, and at least another isolated lineage in the oceanic archipelago of FNO. It remains to be tested if there is a third lineage in the isolated oceanic archipelago of Trindade-Martin Vaz. Levels of genomic differentiation between lineages (F_ST = 0.14; electronic supplementary material, table S4) are similar to or above those distinguishing fishery stocks such as the European sardine (minimum F_ST between stocks > 0.05; [79]), the European hake (min F_ST > 0.058; [80]) and yellow-fin tuna (min F_ST > 0.15; [22]) estimated using similar data. The coastal and oceanic populations of Harengula sp. should therefore be managed as independent stocks, with strategies considering the particularities of each location, including the different fishing activities and their economic dependence on this natural resource. The FNO population, in particular, is likely locally adapted to environmental conditions that are exceptional for species of the Clupeidae. The ecological and social importance of the island lineage, in addition to its independent evolution, suggests that the no-take zone of the Marine Protected Area in FNO, which covers only part of the distribution of the island population, is important for supporting the viability of this population of Harengula sp. This reinforces the need for monitoring and management of this likely sensitive resource in order to maintain exploitation at sustainable levels in the long term, ensuring not only the livelihoods of the local fishing population, but also the proper functioning of the archipelago's marine ecosystem, which is highly dependent on these forage fishes. More generally, the framework presented here can be applied to other species in need of urgent management decisions but where ecological and genetic information remains limited. Such limitation is disproportionally affecting less developed areas of the world, particularly tropical regions of the Southern Hemisphere, where human dependence on natural resources is stronger, scientific resources are more limited, and a significant proportion of biodiversity remains to be described. Thus, the integration of ecological and genomic methods can offer a low cost and efficient solution to resolve the accelerating conflicts between management and local communities that depend on undescribed biodiversity.

Ethics

This work did not require ethical approval from a human subject or animal welfare committee.

Data accessibility

Data used in the ecological niche model and genomic analyses are available at the Dryad Digital Repository and can be accessed here: https://doi.org/10.5061/dryad.np5hqbzzg [81].

Supplementary material is available online [82].

Declaration of AI use

We have not used AI-assisted technologies in creating this article.

Authors' contributions

J.F.R.C.: conceptualization, data curation, formal analysis, investigation, visualization, writing—original draft, writing—review and editing; L.d.F.M.: data curation; F.D.D.: data curation, review and editing; P.H.C.: data curation; R.M.D.: data curation; S.M.Q.L.: conceptualization, resources, supervision, writing—review and editing; J.T.V.: conceptualization, formal analysis, funding acquisition, investigation, project administration, supervision, visualization, writing—original draft, writing—review and editing; R.J.P.: conceptualization, project administration, resources, supervision, visualization, writing—original draft, writing—review and editing.

All authors gave final approval for publication and agreed to be held accountable for the work performed herein.

Conflict of interest declaration

The authors have no conflict of interest to declare.

Funding

J.F.R.C. was financed in part by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior – Brasil (CAPES) – finance code 001. Financial support to F.D.D. was provided by CNPq (Conselho Nacional de Desenvolvimento Científico e Tecnológico – PROTAX 443302/2020). S.M.Q.L. receives a Conselho Nacional de Desenvolvimento Científico e Tecnologico (CNPq) productivity research grant (no. Proc 312066/2021-0). This study was developed in the context of the ‘Projeto MULTIPESCA – Ciência para a sustentabilidade da pesca, pescado e pescadores do Rio de Janeiro’, which received support from the ‘Marine and Fisheries Research Project’. The Marine and Fisheries Research Project is an offset measure established under a consent decree agreed between the company PRIO and the Federal Public Prosecutors' Office in Rio de Janeiro. It is implemented by FUNBIO.