Population Genetics to Population Genomics: Revisiting Multispecies Connectivity of the Hawaiian Archipelago ^†

Abstract

Understanding connectivity between populations is key to identifying hotspots of diversity, dispersal sinks and sources, and effective management units for natural resources. Multi-species connectivity seeks to overcome species-specific idiosyncrasies to identify shared patterns that are most critical to spatial management. The linear Hawaiian archipelago provides an excellent platform to assess multi-species connectivity patterns, with shared boundaries to gene flow identified across a majority of the 41 coral reef species surveyed to date. Here, we evaluate genome-scale data by comparing consistency and resolution to previous connectivity studies using far fewer loci. We used pool-seq to genotype 22,503–232,730 single nucleotide polymorphisms per species (625,215 SNPs total) from the same individuals published in previous studies of two fishes, two corals, and two lobsters. Additionally, one coral species (Pocillopora meandrina) without previous archipelago-wide population genetic data was included. With greater statistical power, most genetic differences between pairwise comparisons of islands were significant (250 of 308), consistent with the most recent larval dispersal models for the Hawaiian Archipelago. These data reveal significant differentiation at a finer scale than previously reported using single-marker studies, yet did not overturn any of the conclusions or management implications drawn from previous studies. We confirm that population genomic datasets are consistent with previously reported patterns of multispecies connectivity but add a finer layer of population resolution that is pertinent to management.

1. Introduction

Organisms use diverse dispersal strategies that have broad evolutionary consequences [1,2,3]. Connectivity describes how individuals among local populations distributed throughout a complex landscape are linked by dispersal and could potentially interbreed. Decreased connectivity typically reduces population size, increases stochasticity, reduces adaptative diversity, and ultimately increases extinction risk for each population [4,5,6]. Thus, connectivity is essential for maintaining natural populations and to implementing effective resource management and conservation efforts [7,8,9,10]. In the absence of modifying factors, connectivity is driven primarily by the capacity of organisms to disperse successfully among habitats [3,11,12]. Effective gene flow, however, requires more than simply arriving; dispersing individuals must survive, grow, and successfully reproduce for their genes to be to present in the next generation [6,13].

Marine species typically disperse much farther than terrestrial species [3,11,12]. While terrestrial systems present obvious physical barriers to dispersal (e.g., rivers and mountains), documenting barriers to dispersal where they are less obvious in the sea remains crucial for understanding population biology, conservation, and evolution. Barriers to dispersal are influenced by the life-history characteristics and life stage of the dispersing organism, potentially disrupting gene flow (e.g., adult physiological tolerances can overcome toxicants or stressors deadly to juveniles) [14,15,16]. Many marine species produce microscopic pelagic larvae that disperse vast distances relative to the sedentary adults, making quantitative estimates of connectivity a major challenge in marine science [17,18,19,20,21,22].

Direct tracking of tiny propagules (e.g., seeds, pollen, or aquatic larvae) is currently impossible, making genetic proxies widely appealing to infer connectivity of such species in nature [2,19,23,24,25]. Considerable debate remains regarding the accuracy of genetic proxies and the influences of organismal life history, marker choice, and history on quantitative estimates of connectivity [26,27,28,29,30,31,32]. For example, Weersing and Toonen [24] found that habitat, swimming ability, and pelagic larval duration were poor predictors of observed population genetic structure, and instead most of the variation was explained by marker class. Toonen et al. [33] collated patterns of genetic structure from 27 marine species throughout the Hawaiian archipelago, using primarily single mitochondrial markers supplemented with five microsatellite datasets, to identify shared barriers to connectivity across diverse taxa. Based on the best tools available at the time, these authors identified four shared barriers to dispersal among eight species of fishes, two species of corals, 15 species of non-coral invertebrates, and two species of marine mammals [33]. Subsequent regional structure identified among endemic fish assemblages throughout the archipelago were roughly concordant with patterns of multispecies connectivity reported by Friedlander et al. [34]. Regional oceanographic modelling by Wren et al. [35] confirmed limited connectivity between neighboring islands, predicting a sharp decline between more distant island pairs consistent with a stepping-stone model, and identified three channels as dispersal barriers (Figure 1).

Both oceanographic modeling and population genetic theory predict conformity to a stepping-stone model in a linear array of islands such as the Hawaiian Archipelago. Genetic connectivity studies by Toonen et al. [33] and Selkoe et al. [36] are consistent with the shared genetic breaks (Figure 1), but only ~10% of species show isolation-by-distance (IBD) [30]. Additional genetic work refined our understanding of genetic structure to match a stepping-stone model [25,36], and simulations confirmed that previous studies using F-statistics and mitochondrial markers lacked sufficient power to distinguish this pattern [30]. Using data from 41 species and a coalescent sampler with greater power (Migrate-n), Crandall et al. [30] show that the shared barriers to dispersal are overlaid on a general stepping-stone model for most species (~70%) across the archipelago. Subsequent work also refined the oceanographic modeling using finer-scale (1 km) nested models and incorporated biologically realistic life history and larval behaviors for 11 species spanning pelagic larval durations (PLD) from 3 to over 300 days [37]. Conklin et al. [37] showed considerable self-recruitment (20–96%) back to the natal island, even for species with the longest PLD. Incorporating larval swimming behavior reduced both median dispersal distance and variability, increasing natal recruitment to >45% for 9 of the 11 species [37]. Overall, these modeling efforts predict limited gene flow and island-level differentiation consistent with a stepping-stone model, because export to adjacent islands never exceeds local retention for any species [37].

Given previous studies, we test whether increased statistical power of genomics offers a better understanding of population structure in the Hawaiian Archipelago. Like computational advances, advances in molecular methods now provide substantially more data in a cost-effective manner. Restriction-enzyme associated digest (RAD) approaches enable scoring of many thousands of loci across the genome without the need for prior genomic information, which is particularly advantageous for non-model organisms [38,39]. The development of such tools and decreasing sequencing costs greatly increase our ability to generate population genomic data [40,41,42]. Does this technological shift invalidate previous work based on one or a few markers [43]? Assumptions that hold true for one or a few loci may not be acceptable when variance increases as a result of read depth or loci in the analysis. Given such tradeoffs, it is worthwhile to address two questions: (1) How robust are single-marker studies in a genomic era and does a multispecies approach to studying population connectivity buffer individual variation for both the markers and the species included? (2) Given the rapid technological advances in the field during the past decade, is it worthwhile to repeat previous work with increased power to re-evaluate management units?

Here, we employ a pooled RADseq (pool-seq) approach, in which the DNA of many individuals per sampling location are pooled equimolarly prior to sequencing to produce an allele frequency estimate for the site [44]. While pool-seq loses individual-level information, many population genetic analyses use only allele frequency information anyway, so this approach provides an accurate and cost-effective alternative to perform population genomics on non-model organisms [44,45,46]. We use this approach to address the questions posed above and test whether the increased power of many thousands of SNPs allows us to identify previously undetected population differentiation and whether increased power results in better guidance for management of natural resources.

2. Materials and Methods

2.1. Study Species

Seven coral reef species (6 surveyed previously) were evaluated for this study. These include two fishes (kole, Ctenochaetus strigosus (Cstr) and Mulloidichthys flavolineatus (Mfla)), two spiny lobsters (Panulirus marginatus (Pmar) and P. penicillatus (Ppen)), and three scleractinian corals (Montipora capitata (Mcap), the newly surveyed species Pocillopora meandrina (Pmea), and Porites lobata (Plob)). These species capture a broad range of life history strategies, including differences in geographic distribution, pelagic larval duration (PLD), and spawning type (

Table 1. Life history traits of study species in the Hawaiian Archipelago. PLD = pelagic larval duration (in days unless specified). Asterisks (*) indicate that congeneric data was used to estimate PLD.

Organism		Life History Trait
Species Name	Common Name	Geographic Range	PLD	Spawn Type
Ctenochaetus strigosus	Kole	HI endemic/ Johnston Atoll	31–58.5 * [47,48,49]	Broadcast [47,48,49]
	(Goldeneye surgeonfish)
Mulloidichthys flavolineatus	Weke ʻaʻa	Indo-Pacific	60.1 [47]	Broadcast [47]
	(Yellowstripe goatfish)
Panulirus marginatus	ʻUla poni	HI endemic	6–11 mo (wild) [50,51,52]	Benthic [52]
	(Hawaiian Spiny Lobster)
Panulirus penicillatus	ʻUla	Indo-Pacific	7–8 mo (wild) [52,53,54]	Benthic [52]
	(Green Spiny Lobster)		8.3–9.4 mo (lab) [54,55]
Montipora capitata	Koʻa	Pacific	3 [36]	Broadcast [36]
	(Rice Coral)
Pocillopora meandrina	Koʻa	Indo-Pacific	5–90 * [56]	Broadcast [56]
	(Cauliflower Coral)
Porites lobata	Pōhaku puna	Indo-Pacific	3 [36]	Broadcast [36]
	(Lobe Coral)

). Among these species are those with large geographic ranges, spanning Oceania and the Indian Ocean (C. strigosus), as well as an endemic restricted to the Hawaiian Archipelago (P. marginatus). Pelagic larval durations of these species range from the coral that lasts a few days (M. capitata) to a reef fish that last weeks (C. strigosus) to the spiny lobsters (Panulirus spp.) which spend up to ~270 days as pelagic phyllosoma larvae.

2.2. Sample Collection

Tissue samples for DNA analyses from Oʻahu, Lānaʻi, Molokaʻi, Kauaʻi, Niʻihau, and Lehua in the Main Hawaiian Islands (MHI) were collected in 2005 (under permit SAP-2005-1), and samples from eight Northwestern Hawaiian Island (NWHI) reefs (Kauō, Pūhāhonu, Kānemilohaʻi, Nihoa, Koʻanakoʻa, Pihemanu, Holoikauaua, and Mokupāpapa) were collected during 2006–2007 (under permit SAP-2006-1 and PMNM-2006-03). All samples were stored in 95% ethanol. Further collections of P. meandrina were made from Maui, Kauaʻi, and Hawaiʻi Island in 2016 (under permit SAP-2016-69) and from Molokaʻi in 2018 (under permit SAP-2018-03), and these samples were stored in saturated-salt-DMSO buffer and stored at room temperature [57]. Collections occurred at as many of the 16 primary islands and atolls as possible in the Hawaiian Archipelago, including the remote and tightly regulated Papahānaumokuākea Marine National Monument (

Table 2. Locations of all 16 Hawaiian Islands sampled.

Sampling Location	Site Code	Latitude (°N)	Longitude (°W)
Kure Atoll|° Mokupāpapa|* Hōlanikū	KURE	28.3925	−178.2936
Midway Island|° Pihemanu|* Kuaihelani	MID	28.2072	−177.3735
Pearl and Hermes Atoll|° Holoikauaua|* Manawai	PH	27.8333	−175.8333
Lisianksi Island|° Papaʻāpoho|* Kapou	LIS	26.0662	−173.9665
Laysan Island|° Kauō|* Kamole	LAY	25.7679	−171.7322
Maro Reef|° Koʻanakoʻa|* Kamokuokamohoaliʻi	MARO	25.415	−170.59
Gardner Pinnacles|° Pūhāhonu|* ʻŌnūnui, * ʻŌnūiki	GAR	24.9988	−167.9988
French Frigate Shoals|° Kānemilohaʻi|* Lalo	FFS	23.7489	−166.1461
Necker Island|°* Mokumanamana	NEC	23.5749	−164.7003
°* Nihoa	NIH	23.0605	−161.9218
°* Kauaʻi	KAU	22.0964	−159.5261
°* Oʻahu	OAHU	21.4389	−158.0001
°* Molokaʻi	MOL	21.1444	−157.0226
°* Lānaʻi	LAN	20.8166	−156.9273
°* Maui	MAUI	20.7984	156.3319
°* Hawaiʻi	HAW	19.5429	155.6659

° Hawaiian Lexicon Committee names (Contemporary Hawaiian); * Kaiʻaikawaha genealogy names (Ancient Hawaiian).

). Maps of sampling locations are found in Figures S2a–g.

Details of our non-lethal sampling protocols and tissue preservation have been published previously [58,59]. Briefly, tissue biopsy samples were taken in the field and stored in either SSD buffer [57] or >70% ethanol until processed for sequencing. Total numbers of samples used in these analyses are shown in

Table 3. Samples included per pool at each collection site for our suite of Hawaiian species. Species abbreviations: Ctenochaetus strigosus (Cstr), Montipora capitata (Mcap), Mulloidichthys flavolineatus (Mfla), Panulirus marginatus (Pmar), Panulirus penicillatus (Ppen), Pocillopora meandrina (Pmea), and Porites lobata (Plob).

Location	Species
	Cstr	Mcap	Mfla	Pmar	Ppen	Pmea	Plob	Site Total
Kure	38	51	27	51	—	—	49	216
Midway	32	40	31	41	—	—	69	213
Pearl and Hermes	33	44	44	49	26	—	73	269
Lisianski	—	50	—	47	21	—	—	118
Laysan	30	46	10	57	—	25	—	168
Maro Reef	—	47	8	56	—	—	51	162
Gardner Pinnacles	29	—	—	58	—	25	47	159
French Frigate Shoals	37	—	28	47	47	23	39	221
Necker	—	—	—	58	—	—	35	93
Nihoa	29	24	—	—	—	23	—	76
Kauaʻi	28	—	60	50	53	25	—	216
Oʻahu	40	57*(2)	57	54	—	96*(4)	53	357
Molokaʻi	25	—	—	—	21	250*(7)	—	296
Lānaʻi	38	—	—	—	38	—	—	76
Maui	—	47	25	34	25	25	—	156
Hawaiʻi	102	—	—	—	50	25	56	233
Species Total	461	406	290	602	281	517	472	3029

* multiple pools from same island (#).

. Due to difficulties in accessing and sampling remote sites, the number of species per island varied, resulting in 198 unique island-scale pairwise comparisons evaluated across all species.

2.3. Benchwork

DNA was extracted from tissues using either the DNeasy Blood & Tissue (Qiagen, Hilden, Germany) or E.Z.N.A. Tissue DNA (Omega Bio-Tek, Norcross, GA, USA) extraction kits, following manufacturer protocols, with modifications to elution outlined in the ezRAD protocol [60]. Extraction from coral samples involved the slight modifications of these protocols as outlined in Polato et al. [61] and Conception et al. [62]. Extracts were then assessed visually using gel electrophoresis of genomic DNA. DNA extracts were quantified using either the AccuClear Ultra High Sensitivity dsDNA (Biotium, Fremont, CA, USA) or Qubit dsDNA High Sensitivity (Invitrogen, Waltham, MA) Quantitation Kits to determine the concentration (ng/µL) of DNA for equimolar pooling by DNA molecular weight. For each collection site (island), all individuals were pooled into 2000 ng libraries containing equimolar DNA per individual to avoid disproportionate contributions from individuals to the library. When a given species had ample tissue samples at an island, they were pooled by collection site and are notated in Table 3. Additionally, pooling of Montipora capitata from Oʻahu was distinguished by morphotype (red vs. orange color morphs). Library preparation followed the ezRAD library preparation protocol [60,63]. Pooled libraries were sequenced on the Illumina MiSeq platform (V3 2 × 300 bp PE).

2.4. Bioinformatics

Sequence data are openly accessible via NCBI’s Sequence Read Archive (SRA) at BioProject number PRJNA1154437 (Accession numbers SRR30482393 through SRR30482402; BioSamples SAMN43416348 through SAMN43416357) and BioProject PRJNA1310660 (Accession numbers SRR35127775 through SRR35127713; BioSamples SAMN50778956 through SAMN50779022), while scripts used in analysis can be found on github (github.com/ebfreel/multispecies_connectivity). Raw reads were trimmed using Trim Galore! (v0.6.8) [64] to remove Illumina adapters on our paired end libraries, as well as trimming by default parameters (--quality Phred 20, -e error rate 0.1). The FastQC output was then visualized with MultiQC (v1.23) [65] to assess sequence quality scores, length distributions, and duplication levels and to ensure RAD tags and Illumina adapters were removed for the library to move forward. After quality control, SPAdes (v3.15.5) [66] was used to construct a de novo assembled reference genome from all populations for each species. Kmer values used were 71, 81, 91, 99, 121, 127, and default (which uses a suite). Default assembly consistently outperformed individually selected kmer values and were used for final assembly. Assembly statistics from QUAST (v5.2.0) [67] are reported in

Table 4. Table of SPAdes assembly statistics per species for the genomic survey of seven Hawaiian species. Species abbreviations are provided in Table 3.

Assembly	Species
	Cstr	Mcap	Mfla	Plob	Pmar	Pmea	Ppen
Contigs (>kbp)	62,561	95,400	43,139	76,769	199,186	116,459	62,312
Contigs (>5 kbp)	5	1950	8	217	219	16,900	303
Total length (>1 kbp)	86,419,269	180,076,159	56,199,729	116,992,871	287,573,493	361,961,856	83,887,656
Total length (>kbp)	32,937	12,845,954	63,455	1,364,810	1,382,369	156,903,027	2,056,972
Contigs	498,704	426,901	460,296	539,291	1,464,336	613,555	1,057,498
Largest contig	10,392	23,892	11,369	17,063	12,510	116,710	20,773
Total length	315,209,420	370,567,895	299,344,044	377,542,905	965,778,094	599,711,333	644,510,677
GC (%)	41	40	45	40	43	39	43
N50	679	958	635	691	702	1687	581
N75	505	560	507	507	510	593	485
L50	150,025	100,724	164,067	164,535	435,477	68,220	393,421
L75	286,906	232,813	297,063	326,836	844,717	234,790	699,183
N’s per 100 kbp	0	0	0	0	0	0	0

All statistics are based on contigs of size ≥200 bp, unless otherwise noted.

. Each de novo assembled genome resulted in over 400 k contigs used in further analysis.

Reads were then mapped back to the SPAdes assembled genome using bwa mem algorithm [68] with a matching score of 1, mismatch penalty of 4, and gap open penalty of 6, ensuring each pool was correctly identified using sample and read group assignments. Variant calling was done using a modified FreeBayes (v1.3.8) [69] script from dDocent (v2.9.4) [70]) (--min-alternate-count 2, --min-mapping-quality 30, --min-base-quality 30, --min-repeat-entropy 0, --min-coverage 50, --min-alternate-fraction 0.05).

SNP filtering, F_ST calculations, and visualizations were completed using assessPool (v2.0.0) (Freel et al. (in review)), which uses vcftools (v0.1.16) [71] and vcflib (v1.0.10) [72] to process SNPs, PoPoolation2 (v1.2.01) [73] and {poolfstat}(v3.0.0) [74] to calculate pairwise F_ST, and {tidyverse} (v2.0.0) [75] tools and {plotly} (v4.10.3) [76] for wrangling and visualizations. Filtering using vcffilter of vcflib included: (1) mapping quality ratio (“MQM/MQMR > 0.75 & MQMR < 1.25”), (2) mapping quality (“MQM > 39 & MQMR > 39”), (3) read balance (“RPR > 0 & RPL > 0”), and (4) depth to quality ratio (“QUAL/DP > 0.25”). Filtering through vcftools included (1) fraction of missing data (--max-missing 0.5), 2) minimum quality (--minQ 30), (3) minor allele count (--mac 2), (4) minimum mean depth (--minDP 10), and (5) maximum mean depth (--max-meanDP 500). Default assessPool parameters for F_ST calculations were used to generate Popoolation2 and {poolfstat} commands. Parameters used per bioinformatic step are summarized in Table S1.

Pairwise F_ST results of each locus were then aggregated into R for further analysis using assessPool. Output was grouped by minimum coverage, keeping only those with between 30–500x. Sites that were not significantly different (p > 0.05) from 0 as determined by PoPoolation2 Fisher’s exact test were treated as F_ST = 0. Because no such test is available in {poolfstat}, masking to F_ST = 0 was applied to the same loci for that output as well. This conservative approach allows us to include low-confidence SNPs without over-inflating our F_ST values. More importantly, by including more F_ST = 0 sites, when testing if all our loci significantly differ from 0 using a one-sided t-test, we were biasing away from detecting genetic differentiation, thus strengthening confidence in comparisons that yielded significant differentiation. Remaining SNP summaries are reported in

Table 5. Summary statistics of {poolfstat} loci included per species for the genomic survey of seven Hawaiian species.

Species	Loci Scored	SNPs > 30x	Mean FST (Stderr)	Mean Coverage
Ctenochaetus strigosus	139,016	22,503	0.0586 (2.5 × 10−4)	51.01
Montipora capitata	327,575	152,750	0.0275 (6.5 × 10−5)	63.2
Mulloidichthys flavolineatus	84,314	25,180	0.0066 (2.4 × 10−4)	44.21
Panulirus marginatus	127,447	43,842	0.0202 (9.8 × 10−5)	57.37
Panulirus penicillatus	120,875	61,819	0.0084 (1.9 × 10−4)	61.88
Pocillopora meandrina	262,600	86,391	0.0010 (1.3 × 10−4)	57.54
Porites lobata	373,879	232,730	0.0133 (5.7 × 10−5)	57.53
	1,435,706	625,215	0.0194	56.11

.

We tested for genetic differentiation in each paired comparison via a one-sided t-test (greater than µ = 0) of all SNP F_ST values, rejecting the null hypothesis of F_ST = 0 at p < 0.01. We then used the FDR() function in the {fuzzySim} (v4.9.9) R package to correct these p-values for a false discovery rate to account for multiple comparisons.

Our filtered VCF files were converted to genid and genlight objects using the vcf2genid() and vcfR2genlight() functions from the {vcfR} (v1.15.0) package [77] before being processed further. We generated de novo genetic clusters by k-means clustering using Bayesian information criterion (BIC) selection (Table S4) using the find.clusters() function from the {adegenet} (v2.1.7) R package [78,79]. De novo clusters from both the “diffNgroups” and “goodfit” criterion were retained.

We also conducted an analysis of molecular variance (AMOVA) test of our final filtered SNP sets using the amova() function in the {ade4} (v1.7-22) R package [78,80] (implemented using the {poppr} (v2.6.1) R package [81]), which uses allele frequency data as input to perform AMOVA. Since the pooled data do not contain individual genotypes, the within = F argument was added to refrain from calculating our data as haplotypes. All species were processed using a priori clusters defined by region (NWHI vs. MHI), subregion (according to shared barriers from previous genetic work), as well as our de novo clusters from the k-means clustering. Significance of the AMOVA were determined using the randtest.amova() function from the {ade4} R package which employs a Monte Carlo test of the components of covariance with permutation processes described by Excoffier et al. [82].

We tested for isolation-by-distance by computing the correlation between the geographic distance between islands and the observed genetic differentiation. Distance was determined with the haversine formula to calculate great-circle distance between coordinates of each site using the {geosphere} (v 1.5-18) package [83] in R and plotted against zero-corrected (i.e., negative F_ST values set to 0) pairwise F_ST values using the {ggplot2} (v4.0.0) package in R [84].

3. Results

Using 625,215 SNP loci across our seven study species (mean: 89,316 SNPs per species; range 22,503–232,730 per species), our pool-seq approach detected significant genetic differentiation between nearly 70 percent (284 of 409) of our pairwise comparisons. Overall, 375 of the 409 pairwise comparisons were between islands, where pools of individuals consisted of all individuals collected from the same island, regardless of exact collection location. However, we had sufficient samples to make pools per collection location for some species. When excluding multiple pool pairs within islands per species (i.e., Pmea MOL3-HAW and Pmea MOL4-HAW), 308 unique island comparisons were included among our seven species, with 250 (81%) of these showing significant differentiation. For Pocillopora meandrina, the 34 pairwise comparisons included 4 site-specific pools per island for Oʻahu (6 intra-island comparisons) and 7 for Molokaʻi (28 intra-island comparisons). Additionally, two Montipora capitata pools from Oʻahu were distinguished by morphotype (red vs. orange coloration) with their pools not genetically distinct from one another (F_ST = −0.0008, p = 0.959). This intra-island resolution for some species allowed us to compare inter-island population structure to intra-island structure. Unsurprisingly, these intra-island pairwise comparisons often showed less differentiation than among island comparisons with no significant comparisons between any of the Oʻahu pools for Pocillopora meandrina or Montipora capitata, nor for any of the Pocillopora meandrina within-Molokaʻi pairs.

Of the remaining non-significant pairwise comparisons, most involved Mulloidichthys flavolineatus, which had the lowest-quality libraries, fewest individuals per pool, and the fewest SNP loci. Additionally, 13 of 19 of these Mulloidichthys flavolineatus comparisons included pools from either Maro Reef or Laysan which contained only 8 and 10 individuals (respectively), far below the recommended minimum pool size [44,85,86]. Another consideration is that Mulloidichthys flavolineatus is among the most dispersive reef fishes in broader surveys across the Indo-Pacific [87]

The remaining non-significant comparisons included two Porites lobata pairs (adjacent GAR-FFS: F_ST = −0.0002, p = 0.33; KURE-MARO: F_ST = 0.0009, p = 0.101), 13 comparisons for Panulirus penicillatus, and 26 for Pocillopora meandrina. As with Mulloidichthys flavolineatus, these non-significant comparisons involved pools with the smallest sample sizes; for Panulirus penicillatus Molokaʻi and Lisianski with 21 individuals per pool, and all Pocillopora meandrina nonsignificant comparisons having at least one pool with 25 individuals or less per pool. These pool sizes are around the minimum recommended limit for pool-seq, with most authors recommending minimum pool sizes closer to double these values [73,86,88], highlighting the importance of sample size.

In each species, our de novo clustering resulted in an optimal number of genetic clusters (k) equal to number of pools minus two. Further AMOVA based on these de novo clusters explained the highest proportion of variation in every case when compared to clustering by region (NWHI vs. MHI) or subregions reported by Toonen et al. 2011 [33] (

Table 6. Per species results of analysis of molecular variance (AMOVA) under four clustering methods. Two a priori clusters based on region (NWHI vs. MHI) and subregion (as defined by Toonen et al. (2011) [33]), and two de novo clusters determined using either the “goodfit” or “diffNgroups” criterion of the find.clusters() function from the {adegenet} R package. Pool # represents the total number of sampling sites for the species. Reported p-values are from the randtest.amova() function in the {ade4} R package and significant values are bolded.

Species (Pool #)	Clustering Method	Degrees of Freedom	Between Variation	Within Variation	p-Value
	region	1	0.79%	99.21%	0.158
C. strigosus	subregion	5	−2.34%	102.34%	0.716
(n = 12)	goodfit k = 10	9	24.36%	75.64%	0.0001
	diffN k = 11	10	26.05	73.95%	0.016
	region	1	2.72%	97.28%	0.105
P. meandrina	subregion	3	5.37%	94.62%	0.003
(n = 16)	goodfit k = 14	13	21.74%	78.26%	0.0001
	diffN k = 15	14	22.55%	77.45%	0.01
	region	1	5.16%	94.84%	0.06
M. capitata	subregion	3	8.90%	91.10%	0.003
(n = 10)	goodfit k = 8	7	12.31%	87.69%	0.006
	diffN k = 9	8	15.14%	84.86%	0.0001
	region	1	0.35%	99.65%	0.236
P. marginatus	subregion	4	1.81%	98.19%	0.159
(n = 12)	goodfit k = 10	9	7.42%	92.58%	0.003
	diffN k = 11	10	8.59%	91.41%	0.016
	region	1	0.32%	99.68%	0.162
P. penicillatus	subregion	4	1.29%	98.71%	0.054
(n = 8)	goodfit k = 6	5	1.86%	98.14%	0.0001
	diffN k = 7	6	2.09%	97.90%	0.0001
	region	1	0.55%	99.45%	0.214
P. lobata	subregion	3	1.04%	98.96%	0.327
(n = 9)	goodfit k = 7	6	11.54%	88.46%	0.0001
	diffN k = 8	7	11.40%	88.61%	0.03
	region	1	0.05%	99.96%	0.439
M. flavolineatus	subregion	4	2.19%	97.81%	0.309
(n = 9)	goodfit k = 7	6	18.14%	81.86%	0.01
	diffN k = 8	7	20.72%	79.28%	0.0001

). The majority of island-by-island pairwise comparisons being significantly differentiated was consistent with a stepping-stone model of connectivity predicted by simulations of genetics [30] and larval dispersal [37] in the Hawaiian Archipelago.

We found no significant correlations between the magnitude of mean F_ST and geographic distance (IBD) in our analysis (Figure S4). Of the previously identified breaks (Figure 1), we found support for limited connectivity in all species included in our study, supported by significant allele frequency differences between islands. We did not find evidence of consistent long-distance connectivity beyond our intra-island comparisons with no significant IBD and clustering (k-means & AMOVA) consistently recovering the largest number of sites (islands) possible (see Supplementary Materials). Highest mean pairwise F_ST at 30× coverage was 0.088 (C. strigosus; Kure-Kauaʻi; p < 0.0001), with 108 of our comparisons being non-significant. Overall, the areas of the archipelago inferred to have shared genetic breaks are consistent with previous work, although with increased power, the majority of comparisons among islands for each species are now significant (Figure 2, Figures S5 and S6).

4. Discussion

Understanding the connectivity among populations distributed throughout a complex landscape is an essential consideration for effective management of natural resources [17,18,19,20,21,22]. It is not surprising that most connectivity work to date has examined one or a few key species [89,90,91]. However, patterns of connectivity rarely match, even among closely related and ecologically similar species [92], and this dilemma remains a considerable challenge to spatial planning when conservation interest [33,93,94]. Multi-species connectivity seeks to overcome species-specific idiosyncrasies to identify shared patterns that are most critical to spatial management [25,33,36]. Thus, to effectively manage ecosystems such as coral reefs, it is essential to consider multispecies patterns of connectivity [33,94]. The linear Hawaiian archipelago provides an excellent platform to assess multi-species connectivity patterns, because islands are discrete habitat patches and shared boundaries to gene flow have previously been identified across a majority of the 41 coral reef species. Incorporating a diverse suite of life-history characteristics ensures that management strategies capture underlying variability inherent in biological communities including the unstudied species as well as the ecologically or commercially important species targeted by management. Such multispecies studies have typically been compiled by performing many detailed studies of connectivity for individual species and overlaying the maps on one another (e.g., [33]) or by relatively coarse sampling of many species across a broad geographic range (e.g., [32,95]). In the case of Toonen et al. [33], 73 of 178 pairwise comparisons (41%) significantly differed when evaluating 27 species across 14 island channels (only island neighbors were included). Such previous connectivity work was limited to one or a few loci, but advances in sequencing capacity have allowed researchers to generate voluminous genomic data and fundamentally shifted our understanding of genetics. The ability to genotype hundreds of thousands of single nucleotide polymorphisms (SNPs) in a cost-effective manner has largely replaced the use of all other markers [6,41,96,97]. Thus, SNPs have rapidly gained popularity and become the marker of choice for connectivity studies [98,99,100], because many biallelic loci [26,27,28,101,102,103] provides substantially more statistical power than a handful of highly polymorphic loci [104,105,106,107] and simplifies interpretation [27]. Here we examined the power gains from using a cost-effective pooled RADseq approach based on many thousands of SNP loci for a subset of Hawaiian coral reef species in a previous multispecies connectivity study.

Pooled RADseq data provides a mechanism to minimize the costs of densely sampled multispecies connectivity studies and is expected to greatly increase statistical power beyond single-species and single-locus studies [44,108,109]. We showed that reanalysis of the same samples using a pool-seq approach provided the power to detect weak, but significant genetic differentiation between islands that were previously undetected in single-marker studies, even among species with extreme dispersal potential. These population genomic data are consistent with the most recent oceanographic modeling predictions that most successful larval recruitment will occur on the same island as the parents [110], resulting in limited gene flow among islands. Further, the results do not conflict with any previously identified shared breaks and connectivity patterns [25,33,36]. Instead, the increased power of this pool-seq approach provides statistical support for the island-by-island stepping-stone population genetic differentiation expected across the Hawaiian Archipelago that was postulated based on previous analyses [19,25,30,49].

The Hawaiian Archipelago is among the most studied of marine systems on the planet in terms of population genetic data as well as identifying shared genetic breaks where dispersal is limited across many species [19,25,30,36,111]. Rather than making direct numerical comparisons of our findings to those previously published using a variety of markers (microsatellites or mitochondrial DNA) and metrics of differentiation (F_ST, G_ST, and ɸ_ST), we focused on whether our findings from population genomic data are consistent with or divergent from patterns of genetic structure previously reported from single-marker studies. Using spiny lobsters as an example, these species have larvae specialized for open ocean dispersal (phyllosoma) with the longest pelagic larval durations (244 to 330 days [50,51,52,53,54,55]) of species studied here. While adults of spiny lobsters are known to undertake mass migrations, deep channels between the Main Hawaiian Islands limits dispersal among the Hawaiian Islands to occurring primarily via pelagic larvae even for highly motile species. Iacchei et al. [112] surveyed these same species of spiny lobsters using the mitochondrial cytochrome c oxidase subunit I (COI) gene and identified four pairs of islands with significant (p < 0.05) pairwise differentiation (F_ST/ɸ_ST). There was little population structure overall, but F_ST was generally elevated in the center of the archipelago for both species. However, after correcting for false discovery rate, no pairwise site comparisons were significantly differentiated. Modeled larval dispersal for these species finds a median dispersal distance of <200 km with at least 20% of larvae recruiting back to the natal island of origin [110]. Consistent with this dispersal prediction, we were able to detect significant differences in every pairwise comparison in the endemic Panulirus marginatus. As an endemic with a limited geographic distribution, this result is intuitive and potentially indicative of a more structured distribution as compared to its widely distributed congener. We still found little population structure for the broadly distributed Indo-Pacific congener Panulirus penicillatus, with 13 of the 28 comparisons being non-significant, but all included low sample size sites of Molokaʻi and Lisianski (both n = 21). In addition to potentially producing spurious results, small pool size is known to affect the significance of test results because the number of individuals in a pool is incorporated into both PoPoolation2 and {poolfstat} calculations of F_ST [44,85,86]. We also note that all three of the pairwise comparisons between the included islands of the Maui Nui island group (Maui, Molokaʻi, and Lānaʻi) were non-significant for P. penicillatus, again consistent with expectations from the larval dispersal modeling. These islands are separated by three relatively shallow and narrow channels: the 140 m deep, 14.2 km wide ʻAuʻau Channel between Maui and Lānaʻi; the 79 m, 15 km wide Kalohi Channel separating Lānaʻi and Molokaʻi; and the 258 m, 14.2 km Pailolo Channel separating Molokaʻi and Maui. For context, most of the other channels separating islands in the Hawaiian archipelago exceed 1000 m in depth and hundreds of kilometers in distance. Despite the increased power for statistical significance of population structure in spiny lobsters, the overall pattern of population genetic structure was generally similar, with low but elevated differentiation seen around the center of the archipelago for both species, as previously reported [112]. This pattern of center-archipelago elevated F_ST was also observed in Porites lobata, Ctenochaetus strigosus, and Mulloidichthys flavolineatus. Likewise, there was consistent differentiation towards the far northwestern end of the archipelago, but with consistently lower differentiation within comparisons involving only these furthest northwestern islands. No previously identified genetic breaks are challenged by our results, rather they are each reinforced with the additional power of our analyses and predicted island-scale differentiation that was previously undetected now has support.

In addition to the larval dispersal modeling [37,110], past genetic work in the region also lends support to these findings of island-by-island population structure [92,113]. For example, Coleman et al. [114] performed a parentage assignment study to understand larval dispersal of the convict tang (manini; Acanthurus triostegus) around the Main Hawaiian Islands. With a pelagic larval duration of 54–77 days, the potential for larval dispersal appears great. In contrast, the authors found most of the detected parent–offspring pairs on Oʻahu came from the same side of the island on which they recruited, with several cases of self-recruitment of juveniles back to the same embayment or even their natal reef [114]. Likewise, Christie et al. [115] showed that yellow tang (lauʻipala; Zebrasoma flavescens) on the Big Island of Hawaiʻi also stayed largely along the same coast of the island. These results are consistent with the bio-physical larval dispersal modeling study [110], and the new coral data we generated here for Pocillopora meandrina to validate that model. Conklin (2024) compared our genetic results to the connectivity matrix produced via Lagrangian dispersal simulations and found that F_ST correlates significantly with passive larval model predictions (r = 0.43, p = 0.006, Mantel test) [110]. Further, an active dispersal model that included realistic larval swimming behaviors resulted in a connectivity matrix with an even better fit (Mantel test, r = 0.71, p = 0.0001; Mantel test) to the pairwise genetic data [110]. These results are consistent with the overall pattern of population genomic structure we report here, with few migrants between populations to maintain genetic cohesion, but with significantly differentiated allele frequencies due to limited gene flow [30,33].

As with other studies of multispecies connectivity, there is variation in the patterns and magnitude of connectivity between islands and among species. The fish, corals, and invertebrates included in our analyses do not share obvious similarities among life history characteristics and include considerable variation in larval behaviors and pelagic dispersal potential [110]. However, despite those many differences, the important and consistent pattern that emerges is barriers to connectivity between islands for a broad suite of Hawaiian coral reef species. The strongest of these barriers have been reported in previous studies [25,33,36], but the pool-seq approach used here provides sufficient statistical power to yield a more cohesive image of significant genetic differentiation between every island sampled. Island-by-island (stepping-stone) population structure has long been suspected and is consistent with both the most recent larval dispersal models [110] and the theoretical expectation for equilibrium population genetic structure in a linear array of islands such as the Hawaiian Archipelago [30]. Previous population genetic work using traditional F-statistics lacked the power to consistently detect such differentiation, whereas use of a coalescent sampler detected stepping-stone population structure in most species, even with single-marker datasets [30]. Of the 37% of species in Crandall et al. [30] for which previous single locus genetic data were ambiguous or best fit one of the seven models other than stepping-stone dispersal (e.g., panmixia, regional grouping, or island model), we included three for this population genomic re-evaluation: Ctenochaetus strigosus, Mulloidichthys flavolineatus, and Panulirus marginatus. Here, we show that the additional power provided by many thousands of SNPs now supports island-by-island population differentiation in 100% of the species for which such data currently exist. Consistent with previous F_ST-based empirical evidence of population structure in the Hawaiian archipelago, as well as genetic simulation work from Crandall et al. [30], IBD is not detectable in our datasets. However, F-statistics are notoriously weak in detecting the underlying processes that produce IBD patterns [29,116,117], with only 10% of species surveyed in the Hawaiian archipelago and 33% globally detecting IBD in marine surveys [30,36]. Notably, IBD also relies on the assumption of an equilibrium between drift and mutation. We would expect sea level changes since the last glacial maximum to directly impact the drift-mutation equilibrium and while we note a very weak trend toward IBD in our genomic data, given large population sizes of marine organisms it might be undetectable with F_ST for tens to hundreds of thousands of generations [116].

5. Conclusions

Recovering the expected stepping-stone pattern of genetic differentiation for every species included in this study, regardless of their pelagic larval duration, indicates that each island should be considered a distinct management unit. Using these population genomic tools, we can estimate F_ST in many species simultaneously, each collected across many sites, and with thousands of loci, providing much greater statistical power than previously possible. While our findings are biologically insightful, their implications are perhaps more important in the context of management of marine resources. Widespread barriers to dispersal and connectivity, across multiple phyla, highlight the importance of sustainable local management because neighboring islands provide little buffer for recovery of depleted marine resources. With intra-island dispersal inferred from these findings to be low, recovery of overharvested resources is dependent on spawning adults on the same island. Our lowest significant value between any pair of sites compared was F_ST = 0.0016. To understand the biological significance of such low F_ST values, we look to genetic simulations and complementary coalescent analyses by Crandall et al. [30]. Crandall et al. estimated that approximately 100 effective migrants per generation (N_em) would be needed to yield an F_ST of 0.002, very similar to our lowest significant value. While this level of gene flow may preclude evolutionary partitions, it would take many generations of immigration to rebuild a population depleted by overharvest, disease, or natural catastrophe [118,119]. Thus, even if the effective population size is orders of magnitude lower than the census number [120,121,122], the management consequences of fewer than 100 individuals recruiting each generation emphasizes the importance of local management of populations.

Finally, our comparisons show the power and cost-effectiveness of pool-seq for multispecies studies of population connectivity. Combining the DNA from many individuals per site into a single sample allows accurate estimation of allele frequencies from which F_ST can be calculated at a fraction of the cost of performing many single-species studies. Here we evaluated seven species across an average of 10 islands per species throughout the Hawaiian Archipelago that included a total of 3029 individuals and 643,173 SNPs. Being able to run this as 70 sequencing libraries instead of over 3000 means this study cost less than any of the previous single-species studies alone. We corroborate the major findings of those single-species studies and confirmed that they are still useful and relevant, yet we also demonstrate finer-scale population structuring that is biologically meaningful but was previously undetected using single-locus approaches.