Dispersal mechanisms for zebra mussels: population genetics supports clustered invasions over spread from hub lakes in Minnesota

Transcription

1 Biol Invasions ORIGINAL PAPER Dispersal mechanisms for zebra mussels: population genetics supports clustered invasions over spread from hub lakes in Minnesota Sophie Mallez. Michael McCartney Received: 8 September 2017 / Accepted: 17 March 2018 Ó Springer International Publishing AG, part of Springer Nature 2018 Abstract Invasion genetic studies show great promise for inferring sources, pathways and vectors of spread, but have to date focused on introductions from native ranges at continental scales and larger. Here we present genetic analyses of post-introduction (i.e. secondary) spread of zebra mussels at the scale of a US state, Minnesota (MN). We genotyped 9 microsatellite DNA loci in 2047 zebra mussels collected from 40 lakes and 4 rivers that geographically and chronologically span the MN invasion. We analyzed genetic variation, geographic differentiation, and genetic clustering of populations across water bodies, and performed invasion-scenario contrasts using Approximate Bayesian Computation. Our population genetic analyses revealed that the pattern of spread of zebra mussels to inland lakes did not conform to the hub-lake model that is often proposed for this species. Consistent with a stratified dispersal model, lake rich regions were colonized from afar, followed by spread within regions by short-distance Electronic supplementary material The online version of this article ( contains supplementary material, which is available to authorized users. S. Mallez (&) M. McCartney Department of Fisheries, Wildlife, and Conservation Biology, Minnesota Aquatic Invasive Species Research Center (MAISRC), University of Minnesota, St. Paul, MN 55108, USA smallez@umn.edu dispersal, but high-traffic hub lakes made no contribution. For the first time, we obtained evidence about the pattern of spread of zebra mussels to inland lakes that is not based on inference from analysis of boater movements; rather it derives from direct evaluation of genetic attributes of invasive populations. Local introductions overwhelmed more distant ones in three independently invaded regions, suggesting a need to understand the lake colonization process and identify the vectors responsible. Keywords Microsatellite Approximate Bayesian computation Dreissena polymorpha Genetic diversity Genetic structure Aquatic species Shortdistance spread Stratified dispersal Introduction Over twenty years ago, Vitousek et al. (1996) wrote that our mobile society is redistributing the species on the earth at a pace that challenges ecosystems, threatens human health and strains economies. In just a few decades, biological invasions have become one of the leading threats to biodiversity and drivers of global environmental changes (Mack et al. 2000; Ricciardi 2007; Sakai et al. 2001; Sala et al. 2000; Vitousek et al. 1996; Wilcove et al. 1998). These are complex phenomena, made up of three fundamental

2 S. Mallez, M. McCartney steps: introduction, establishment and proliferation (Mack et al. 2000; Sakai et al. 2001). Primary spread of invasive species i.e. the initial introduction of organisms from their native ranges into new, often distant areas has been the subject of studies in invasion genetics in nearly every case (e.g. Arca et al. ; Asunce et al. 2011; Boissin et al. 2012; Boucher et al. 2013; Fraimout et al. 2017; Guillemaud et al. ; Kelager et al. 2013; Miller et al. 2005; Perdereau et al. 2013; Wan et al. 2012). For this stage of primary spread of numerous species, invasion genetics has made considerable and growing contributions to our knowledge of sources and pathways, and constitutes a crucial step in our understanding of biological invasions (Estoup and Guillemaud 2010). Nevertheless, the introduction of an organism outside of its native range does not alone constitute a successful invasion. An invasion persists mainly due to the proliferation, numerically and spatially, of the organism in the area of introduction (Daehler 2001; Richardson et al. 2001). Post-introduction range expansion or secondary spread of invasive populations (Geller et al. 2010) governs the scale of impacts of an invasion and often drives the design of management and control measures at convenient spatial scales (e.g., within a country or US state) to be more realistically and easily undertaken (e.g., Rollins et al. 2009). Post-introduction range expansion has been the subject of spatial ecology theory, which has identified three non-exclusive mechanisms: (1) neighborhood diffusion (Okubo and Levin 2001) or stepping-stone dispersal (Kimura and Weiss 1964), in which individuals spread nearby, over short-distances; (2) longdistance dispersal, in which individuals are transported or disperse far away; and (3) stratified dispersal (Shigesada et al. 1995), in which individuals are transported to and establish new founding populations in distant regions, and then spread over short distances within these new regions. Secondary spread travels over geographic pathways and by means of vectors that are typically quite distinct from those followed by an organism when it first colonizes a new region, during primary spread. Empirical studies have demonstrated the important insights obtained from the investigation of secondary spread for several invasive species (raccoon, Fischer et al. 2017; red swamp crayfish, Huang et al. 2017; saltcedar, Lee et al. 2018; Asian tiger mosquito, Medley et al. ; harlequin ladybird, Veran et al. ; American mink, Zalewski et al. 2009). Indeed, determining which mechanism(s) is(are) involved in secondary spread and the degree of connectivity between outbreaks permit the identification of the key processes and vectors at play, and has direct implications for management efforts to contain existing outbreaks and prevent further spread (Huang et al. 2017; Medley et al. ; Zalewski et al. 2009). Benefiting from the opening of the St. Lawrence Seaway, zebra mussel [Dreissena polymorpha (Pallas)] primary spread occurred as the species reached the Laurentian Great Lakes in the first half of the 1980s in ballast water discharge from transoceanic ships (Carlton 2008; Hebert et al. 1989). Native to southern Russia/eastern Europe (estuaries of the Black, Caspian and Azov seas), zebra mussels have invaded waters throughout Europe (Karatayev et al. 1997, 2003) and North America (Benson ; Carlton 2008; Hebert et al. 1989). They have become one of the most widespread and damaging of the world s aquatic invasive species (Karatayev et al. 2007). Zebra mussels clog water intake pipes of industrial facilities (De Leon 2008), impact native freshwater mussel and fish populations (Karatayev et al. 1997; Lucy et al. ; McNickle et al. 2006; Raikow 2004; Strayer et al. 2004; Ward and Ricciardi ) and restructure aquatic food webs (Bootsma and Liao ; Higgins and Vander Zanden 2010; Mayer et al. ). Considerable research effort has focused on the process of secondary spread of the zebra mussel in both its European and North-American invaded ranges (Bobeldyk et al. 2005; Johnson et al. 2006; Johnson and Carlton 1996; Karatayev et al. 2003, ; Kraft et al. 2002). There are two ways for zebra mussels to disperse and spread: (1) through downstream, natural dispersal of larvae or mussels attached to floating vegetation or debris and (2) overland, through dispersal of larvae or mussels mediated by vectors that do not require waterway connections. While natural dispersal of larvae down streams that interconnect lakes accounts for a sizeable fraction of new inland lake invasions (Bobeldyk et al. 2005; Horvath et al. 1996; Johnson et al. 2006), mussels transported overland by human activities are responsible for many others including long-distance dispersal events that enable new geographic regions to be colonized. It is quite clear that human-mediated vectors are a necessary facilitator of secondary spread of zebra mussels, far more important than natural vectors (e.g. waterfowl)

3 Dispersal mechanisms for zebra mussels that play at best a minor role in their spread over land (Karatayev et al. 2003; Kraft et al. 2002), and that recreational boats play a key role in human-mediated spread (De Ventura et al. ; Johnson et al. 2001, 2006; Johnson and Padilla 1996; Padilla et al. 1996). In this work, we asked whether population genetic data could help account for geographic patterns, and whether these data can improve our understanding of zebra mussel inland lake invasions. In North America, secondary spread has been hypothesized to trace its origins both to chronic re-invasion from sources in the Great Lakes and large rivers (e.g. the Mississippi), as well as spread from important inland source lakes that act as hubs in invasion networks (e.g. Muirhead and MacIsaac 2005). Kraft et al. (2002) have also demonstrated that zebra mussel infested lakes are significantly clustered at small spatial scales (\ 50 km in Belarus and the US) and segregated at larger spatial scales ([ 120 km in Belarus and km in the US). Overland long-distance dispersal is certainly an important mechanism in the spread of zebra mussels at continental or world-wide scales but what is going on at smaller scales (state-wide or within states) is less clear. What are the contributions of chronic, independent events of introduction from distant sources compared to local spread from sources nearby? By inferring patterns of gene flow and reproductive connectivity among invasive outbreaks (Hamilton 2009), population genetics is able to identify shortand long-distance dispersal events, which can clarify invasion processes and provide direction to management. In particular, different vectors of transport may be involved in short vs. long-distance dispersal events, and so determining dispersal scale can provide clues about the vectors responsible for spread in a particular geographic region. Management responses might therefore be designed to target the right vector at the right scale (Geller et al. 2010) and be less timeconsuming, less expensive and more efficient (De Ventura et al. ). In this paper, we focused our study on Minnesota (MN), a US state where zebra mussels are actively spreading to new inland lakes (see Materials and Methods). We performed population genetic analyses to determine the pattern of spread of zebra mussels between inland lakes. We focused on potential invasion hub lakes as well as geographic clusters of infested lakes, since each are thought to be hallmarks of zebra mussel secondary invasions. Our results allow us to examine the contributions of short- versus longdistance dispersal events to the spread of this invader at the western front of its range expansion in North America. Materials and methods Pattern of invasion of zebra mussels in Minnesota Only seven states New York (NY), Ohio (OH), Michigan (MI), Indiana (IN), Illinois (IL), Wisconsin (WI), and Minnesota (MN) have [ 25 infested inland lakes as of. Minnesota data, updated through end-, were obtained from the MN Department of Natural Resources (K. Pennington, MN DNR, personal communication). For the 6 other states, we used the US Geological Survey (USGS) database of collection sites for Nonindigenous Aquatic Species ( current update for the end of, accessed 6/02/16). The earliest date with confirmed presence of zebra mussels was used as the date of first infestation, and the cumulative number of new lake infestations was plotted against the year of infestation (Fig. 1). In the 6 other states (aside from MN), the number of newly infested lakes per year has declined towards the present, with the possible exception of WI (where a stairstep pattern with no clear slowing of infestation rate is observed). In MN, the rate has not yet slowed. MN remains in an active stage of inland spread and thus is suitable for invasion genetic studies aimed at identifying routes of inland invasions and at aiding managers with prevention efforts. Sampling, DNA extraction and microsatellite loci genotyping Over 3 years (,, ), we sampled 40 lakes and 4 rivers infested with zebra mussels in Minnesota (two lakes in Wisconsin, see below), for a total of 69 geographical sites and 2047 individuals (Table 1, Fig. 2). Sites were selected to span the chronology of inland invasion in MN, and to represent all major geographic clusters of infested lakes in the state. Specifically, three lake-rich regions were densely sampled in Minnesota: the Alexandria region, the Brainerd region and the Detroit Lakes region, with 12,

4 S. Mallez, M. McCartney facilitate scoring (Table S1). PCR amplifications of these microsatellite loci were carried out in 10 ll volumes containing 1X Bioline MyTaq TM HS Mix, 1.6 lm of each primer with forward primers labeled with a fluorescent dye (6-FAM, VIC, PET or NED) and 10 ng of genomic DNA. The amplifications were carried out in a Bio-Rad T100 TM Thermal Cycler and included a 2 min denaturation step at 95 C, followed by 25 cycles of 30 s at 95 C and 4 min at 60.4 C (adapted from the Multiplex PCR protocol recommended by the Bioline MyTaq TM HS Mix). Genotype scoring was performed using an ABI 3730xl genetic analyzer (Applied Biosystems) with the 500 LIZ TM GeneScan TM size standard (Applied Biosystems) and GeneMarker TM software version (SoftGenetics LLC, State College, PA). Genetic diversity and genetic structure analyses Fig. 1 Pattern of invasion of zebra mussels in the USA. The cumulative number of infested inland lakes and reservoirs (rivers, streams, riverine lakes, flowages, and Great Lakes records were excluded) is plotted against the year of infestation (earliest date of discovery). Only the US states having more than 25 infested lakes are shown and are the following: Illinois (IL), Indiana (IN), Minnesota (MN), New-York (NY), Ohio (OH), Wisconsin (WI) and Michigan (MI) 10 and 7 lakes sampled (sample codes beginning with AL, BL or DL: Table 1, Fig. 2) respectively. These regions included both unconnected lakes at short overland distances from one another and chains of lakes, i.e. lakes joined by small streams, rivers, and channels (Fig. 2). Between 75 and 400 individual mussels were collected per location, stored in sealed bags on ice and immediately frozen at -20 C upon transport to the laboratory. DNA extractions were performed from mantle, foot, gill or adductor muscle tissues, with the DNeasy Ò blood and tissue kit (Qiagen, Valencia, California), using an elution volume of 200 ll, according to the manufacturer s instructions and including the optional RNase step. Each individual was genotyped at 9 microsatellite loci, obtained from the literature (Naish and Boulding 2011; Peñarrubia et al. ; Thomas et al. 2011) and optimized for multiplexing (Table S1). For some of the markers, the reverse primers were PIG-tailed (5 0 -GTTT added: Brownstein et al. 1996)to For each locus in each sample, we evaluated the presence of null alleles, large allele dropout or scoring errors with MICROCHECKER (Van Oosterhout et al. 2004) and we computed the frequencies of null alleles, if present, with FREENA (Chapuis and Estoup 2007). Results from these analyses showed that null alleles may be present for some of the loci and samples. However, null alleles were not observed at a given locus across all the samples or for a given sample across all loci. In addition, none of the samples or loci had more than 5% null alleles on average (data not shown). We therefore retained all nine loci and all the samples for further analyses. We determined the mean number of alleles (mean N a ), observed (H o ) and expected (H e ) heterozygosities per sample with GENETIX version 4.05 (Belkhir et al ) and mean allelic richness (mean A r ) with FSTAT version (Goudet 2002). We evaluated deviation from Hardy Weinberg equilibrium (HWE) and linkage equilibrium with GENEPOP version (Rousset 2008) and we quantified any inferred deviations from HWE by calculating the Weir and Cockerham estimate of F IS (Weir and Cockerham 1984) with FSTAT version (Goudet 2002). We took multiple testing (HWE tests) and the nonindependence of tests (linkage tests) into account by performing false discovery rate (FDR) correction (Benjamini and Hochberg 1995) and sequential Bonferroni adjustment (Sokal and Rohlf 1995), respectively.

5 Dispersal mechanisms for zebra mussels Table 1 Characteristics and population genetics summary statistics of each sample of zebra mussel used in this study Population codesite number Lake-rich region Lake/River Site State Latitude ( N) Longitude ( W) Sample date First obs. (years) Ns Mean Na Mean Ar He Ho F IS DHTW-6 Superior Thomas Wilson s Wreck MN /26/ DHPP-6 Superior Park Point MN /10/ MRBLD-8 Mississippi d Blanchard Dam MN /5/ MRHSL-13 Mississippi d Hok-Si-La MN /22/ MRLPA-14 Mississippi d Lake Pepin MN /22/ MRLPB-14 Mississippi d Lake Pepin MN /22/ MRNCA-15 Mississippi d West Newton Chute MRNCB-15 Mississippi d West Newton Chute MN /22/ MN /22/ LGV-17 Geneva WI /3/ MISSC-10 Mississippi d Junction Mississippi/St Croix WI /29/ SCBBB-10 St Croix d Black Bass Bar WI /28/ SCLSCB-10 St Croix d Lake St. Croix Beach MN /28/ * * SCP-10 St Croix d Prescott MN /29/ SCSCB-10 St Croix d St. Croix Bluff MN /29/ ZL-16 Zumbro MN * BLOW Brainerd Ossawinamakee MN /20/ BLPB Brainerd Pelican Brook d Ossawinamakee outlet MN /18/ MLBP-7 Mille Lacs Brown s Point MN /13/

6 S. Mallez, M. McCartney Table 1 continued Population codesite number Lake-rich region Lake/River Site State Latitude ( N) Longitude ( W) Sample date MLGR-7 Mille Lacs Garrison Reef MN /13/ MLH-7 Mille Lacs Hennepin Island MN /21/ MLKP-7 Mille Lacs Knox Point MN /13/ MLRR-7 Mille Lacs Rocky Reef MN /13/ MLSR-7 Mille Lacs Spider Reef MN /21/ MLTM-7 Mille Lacs Three Mile Reef MN /21/ BLRI Brainerd Rice MN /21/ ALC Alexandria Carlos MN /4/ ALGV Alexandria Geneva MN /17/ ALHD Alexandria LeHomme Dieu MN /9/ DLPEA Detroit Lakes DLPF Detroit Lakes DLPWA Detroit Lakes DLNLIZ Detroit Lakes DLSLIZ Detroit Lakes Pelican East Boat Access MN /4/ Pelican Fish Lake MN /8/ Pelican West Boat Access MN /8/ Lizzie North MN /26/ Lizzie South MN /26/ PL-5 Pike MN /31/ PRLA-12 Prior Grainwood Crossings Park MN /8/ PRLB-12 Prior Watzl s Beach MN /8/ First obs. (years) Ns Mean Na Mean Ar He Ho F IS

7 Dispersal mechanisms for zebra mussels Table 1 continued Population codesite number Lake-rich region Lake/River Site State Latitude ( N) Longitude ( W) Sample date PRLC-12 Prior Sand Point Park MN /8/ BL-9 Bass WI /21/ MTKCB-11 Minnetonka Crystal Bay MN /10/ MTKGB-11 Minnetonka Gray s Bay MN /10/ MTKSA-11 Minnetonka St Albans Bay MN /10/ ALBRO Alexandria Brophy MN /7/ ALD Alexandria Darling MN /17/ ALV Alexandria Victoria MN /9/ BLGA Brainerd Gull MN /25/ BLGB Brainerd Gull MN /21/ BLGC Brainerd Gull USACE Boat Access MN /25/ BLRO Brainerd Round MN /21/ ALIR Alexandria Irene MN /9/ ALCWD Alexandria Cowdrey MN /7/ DLPRA Detroit Lakes Prairie MN /26/ ALMT Alexandria Miltona MN /9/ BLP Brainerd Pelican MN /21/ DLCRY Detroit Lakes Crystal MN /26/ First obs. (years) Ns Mean Na Mean Ar He Ho F IS * * * * * *

8 S. Mallez, M. McCartney Table 1 continued Population codesite number Lake-rich region Lake/River Site State Latitude ( N) Longitude ( W) Sample date First obs. (years) Ns Mean Na Mean Ar He Ho F IS DLORW Detroit Lakes Orwell MN /26/ LW-2 Winnibigoshish MN /7/ ALSM Alexandria Mary MN /7/ ALMP Alexandria Maple MN /7/ BLC Brainerd Cross MN /22/ BLLH Brainerd Lower Hay MN /22/ DLNLID Detroit Lakes Lida MN /26/ SL-4 Sand MN /28/ XL-11 Christmas MN /3/ ALIDA Alexandria Ida MN /9/ BLNL Brainerd North Long MN /21/ BLGIL Brainerd Gilbert MN /22/ DLMEL Detroit Lakes Melissa MN /26/ CAL-1 Cass MN /6/ LSLC-3 Little Sand d Channel MN /6/ * * Samples are listed by date of first observation, earliest to most recent. Site numbers refer to Fig. 2. d indicates the rivers. Ns is the sample size (number of individuals). Mean Na is the mean number of alleles per sample over all loci, Mean Ar is the mean allelic richness per sample over all loci (based on 20 individuals), He and Ho are the expected and observed heterozygosities, respectively. FIS was calculated as described by Weir and Cockerham (1984). * HWE test is significant at P \ 0.05 after FDR correction (Benjamini and Hochberg 1995)

9 Dispersal mechanisms for zebra mussels Fig. 2 Location of the sampling sites of zebra mussels used in this study. The map in the top panel shows the region of MN sampled (with landmarks labeled) and corresponds to the green area in the inset US map. Numbers in red text correspond to the sampling site numbers (Table 1, note that only the water bodies sampled in MN are indicated on the map). The three most densely sampled lake-rich regions are labeled and highlighted with blue shading. The expanded maps of each region are shown in the lower panel of the figure. For more details about the sampling sites, see Table 1 Genetic structure and relationships between the different samples We evaluated genotypic differentiation between each pair of sample sites by the exact G test (Raymond and Rousset 1995) with GENEPOP version (Rousset 2008). Non-independent multiple tests were performed and so the sequential Bonferroni correction (Sokal and Rohlf 1995) was applied to adjust significance levels. We also computed Weir and Cockerham estimates of F ST (Weir and Cockerham 1984) correctedfornull alleles with FREENA (Chapuis and Estoup 2007). We evaluated isolation by distance (Rousset 1997) by testing the correlation between pairwise genetic (F ST / 1 - F ST ) and geographic distances with GENEPOP version (Rousset 2008). We then performed Bayesian clustering as implemented in STRUCTURE version (Pritchard et al.

10 S. Mallez, M. McCartney 2000) to study the genetic structure of zebra mussel populations at different levels: (1) within water bodies (6 lakes and 2 rivers were sampled at multiple, welldispersed locations, Table 1), (2) within lake-rich regions (i.e., Alexandria, Brainerd and Detroit Lakes regions, Fig. 2) and (3) globally (i.e., state-wide). An admixture model with correlated allele frequencies was used (Falush et al. 2003). The number of clusters tested, K, varied from 1 to the total number of samples? 2 for the within-waterbody analyses, from 1 to 4 or 6 for the lake-rich region analyses, and from 1 to 15 for the global analysis. Twenty independent runs for each value of K were carried out, each one involving a Markov Chain Monte Carlo (MCMC) procedure with 10 6 iterations, following a burn-in period of iterations. Default values were maintained for all the other parameters. We used the CLUMPAK server (Kopelman et al. ), which combines CLUMPP (Jakobsson and Rosenberg 2007) and DISTRUCT (Rosenberg 2004) software packages, to post-process the STRUCTURE outputs: (1) to determine the number of clusters in our dataset using both the DK of Evanno et al. (2005) and the posterior probability per K as presented in the STRUCTURE documentation, (2) to identify the most frequent clustering pattern for each value of K (across the 20 independent runs) and (3) to display the corresponding bar plots. For an alternative assessment of genetic clustering, we generated a neighbor-joining tree (Saitou and Nei 1987) with POPULATION version ( populations/) using Cavalli-Sforza and Edwards genetic distances (Cavalli-Sforza and Edwards 1967). Five thousand bootstrap replicates over loci were used to assess support for the consensus tree typology. Finally, we performed an analysis of molecular variance (AMOVA) with Arlequin version (Excoffier and Lischer 2010) to assess the distribution of the genetic variability among and within lake-rich regions. Approximate Bayesian computation (ABC) analyses The genetic structure analyses described previously revealed the existence of well-defined genetic clusters (see Results) that allowed us to further investigate two aspects of the zebra mussel invasion using ABC analyses (Beaumont et al. 2002): (1) the existence of hub lakes i.e. important inland source lakes, that act as hubs in the invasion network (e.g. Muirhead and MacIsaac 2005) and (2) the pattern of spread in lakerich regions to assess the respective role of short- and long-distance overland dispersal. Investigation of the hub lake hypothesis One of the main scenarios proposed to explain the secondary spread of zebra mussels in MN relies upon the existence of inland hub lakes serving as epicenters for subsequent invasion of satellite lakes (e.g. Bossenbroek et al. 2001; Kraft et al. 2002). In the state, potential hub lakes are commonly referred to as super-spreader lakes, designations based entirely upon high rates of trailered boat traffic and public accesses, which create the greatest numbers of pathways connecting outbound boats to next-visited destination lakes. MN DNR Watercraft Inspection Program data (A. Doll, MN DNR WIP coordinator, pers. comm.) identify Lake Mille Lacs (a 128,000-acre highly popular fishing lake located north of the Twin Cities Metro, invaded in 2005 and now heavily infested) and Prior Lake (a Metro-area lake heavily infested since 2009, with extensive residential development and high boater traffic from this home site to destination lakes throughout the state) as among the most important potential hub lakes in MN. For testing models involving each of these source lakes, we compared two different invasive scenarios (Figure S2): (1) a scenario of successive introductions, in which an unspecified water body infested Lake Mille Lacs (or Prior Lake), followed by Lake Mille Lacs (or Prior Lake) infesting another destination lake more recently, post-2005 (or post-2009), and (2) a scenario of independent introductions, in which Lake Mille Lacs (or Prior Lake) was not the source for the destination lake. We tested 35 and 28 lakes as destination lakes for Lake Mille Lacs or Prior Lake, respectively, using these scenario comparisons. Each of the ABC analyses were carried out using DIYABC version (Cornuet et al. 2010). For each analysis, one million datasets were simulated per competing scenario. The following summary statistics (SuSts) were used to compare and select scenarios: the mean number of alleles, the mean expected heterozygosity and the mean allelic size variance per population and per pair of populations, the mean ratio of the number of alleles to the range in allele size (Garza and

11 Dispersal mechanisms for zebra mussels Williamson 2001), the mean individual assignment likelihoods of population i to population j (Pascual et al. 2007) and F ST (Weir and Cockerham 1984). The posterior probabilities were estimated using a polychotomous logistic regression on the 1% of simulated datasets closest to the observed dataset (Cornuet et al. 2008, 2010). The LDA option, permitting a discriminant analysis on the SuSts prior to regression, was used (Cornuet et al. ; Estoup et al. 2012). To evaluate the robustness of our inferences, we performed the ABC analyses with two different sets of prior distributions (Table S2). The confidence in scenario choice was also evaluated by (1) verifying that the 95% confidence intervals of the posterior probabilities of competing scenarios did not overlap and (2) by computing the posterior error rate for every analysis. The posterior error rate was obtained by performing ABC analyses on 1000 simulated datasets, called pseudo-observed datasets or pods. These pods were simulated by drawing with replacement the scenario identity and the parameter values among the 500 sets of scenarios and parameters values that generated simulated datasets closest to the observed dataset. The posterior error rate corresponds to the proportion of ABC analyses that wrongly identified scenarios and thus helps us to test our ability to select the true scenario. (3) We finally checked the capacity of the selected scenario to generate datasets similar to the observed one (Cornuet et al. 2010). This step was carried out by simulating new datasets drawing the parameter values from the posterior distributions of the selected scenario. Each observed SuSt was then localized in the distribution formed by the simulated SuSts, which gave a rejection probability for each observed SuSt. Two new SuSts were added to the previous ones for this purpose: the shared allele distances between populations (Chakraborty and Jin 1993) and the dl 2 distance (Goldstein et al. 1995) between pairs of populations. A large number of probabilities is obtained during this step, we therefore adjusted the significance threshold with the FDR method (Benjamini and Hochberg 1995). Studies of clustered invasions Spatial clustering of inland lakes and the presence of waterway connections between them are factors that facilitate secondary spread and generate clustered zebra mussel infestations, as detailed above. Three regions of clustered lakes were sampled in our study Alexandria, Brainerd and Detroit Lakes. In the Alexandria region, different genetic groups were observed (see Results) and this, coupled with historical information, guided the ABC analyses. We explicitly investigated the existence of sub-structure by assessing the relationships between lakes from different inferred genetic groups (Lake Carlos and Lake Darling) and between lakes from the same inferred genetic group (Lake Carlos and Lake Irene). Lake Mary was also incorporated into this analysis due to its lack of clear membership in either genetic cluster. The Alexandria-region analysis examined 24 different scenarios that were simultaneously compared to one another (Figure S1). The standard genetic analyses did not point out any particular genetic pattern within the Brainerd and Detroit Lakes regions and so the ABC analyses performed in these cases were guided by invasion chronology. As the number of alternative scenarios increases exponentially with the number of samples included in the analyses, we considered only four lakes per region that reflected the chronology of the invasion Ossawinnamakee, Rice, Gull and Pelican for the Brainerd region, and Pelican, Prairie, Orwell and Melissa for the Detroit Lakes region and we compared all the possible scenarios in accordance with the dates of first observation. This generated 24 scenarios to compare for each region (Figure S1). All ABC analyses were carried out using DIYABC version (Cornuet et al. 2010) as described earlier, except for the computations of the posterior error rates. Posterior error rates were obtained by performing ABC analyses on 100 pods (instead of 1000) to reduce computation time from several weeks to several days. Results Standard population genetic analyses of microsatellite data High levels of genetic diversity were observed in zebra mussel populations, with a total number of alleles detected per sample site, across the nine loci, ranging from 42 to 102. The mean Na and mean Ar per sample site ranged from 4.76 to and from 4.44 to 9.88, respectively (Table 1). Observed and expected

12 S. Mallez, M. McCartney heterozygosities were also quite high, ranging from 0.55 to 0.75, and 0.59 to 0.75, respectively (Table 1). Most of the samples (57 out of 69) did not deviate from HWE and only 1 test out of 2473 showed significant linkage disequilibrium. Ten lakes Mille Lacs, Prior, Carlos, LeHomme Dieu, Geneva (Douglas County, MN), Irene, Miltona, Ida, Victoria and Mary exhibited significantly lower A r and He than the other water bodies studied (permutation test among groups with FSTAT, 1000 permutations, P \ for both tests of both estimates). No correlation was observed between A r and the lake surface area [Spearman s rank correlation, S = 30442, P [ 0.2; performed with R version (2017)]. A r was also not correlated with the year of discovery of zebra mussels in a lake (obtained from MN DNR records) as would be expected in a scenario of serial bottlenecking (e.g. Clegg et al. 2002) [Spearman s rank correlation, S = 20737, P [ 0.1; performed with R version (2017)]. Genetic structure and relationships within and between water bodies Within water bodies, no significant genetic differentiation was found among samples within the six lakes [Superior, Gull, Mille Lacs, Minnetonka, Pelican (Otter Tail County) and Prior] and two rivers (Mississippi and St. Croix) in which multiple sampling sites were genotyped. Accordingly, Bayesian clustering analysis inferred a single cluster within each water body, in every case (Appendix 1 in Supplementary Information). Between water bodies, significant genetic differences were revealed among 78% of the lake samples with low to moderate F ST values ranging from to (Table S3). No IBD pattern was revealed within the state (P = 0.17). Two lakes, Mille Lacs and Prior, were significantly differentiated from every other lake studied. Most samples representing the oldest zebra mussel infestations in the state, i.e. Lake Superior, the Mississippi and St. Croix Rivers, were not significantly differentiated from one another (except for a few samples from the Mississippi River and Lake Superior). The lake-rich regions showed contrasting patterns of differentiation. No significant differentiation was revealed within the Detroit Lakes region. For the Brainerd region, mainly, BLGIL was significantly differentiated from every other lake in the cluster and BLP was differentiated from all Gull Lake samples (BLGA, BLGB, and BLGC) and from BLRO, BLNL and BLGIL. In the Alexandria region, no significant differentiation was found between lakes within the (ALD, ALBRO and ALCWD) group and between lakes within the (ALIR, ALMT, ALIDA, ALC, ALHD, ALV, ALGV and ALMP) group but significant differences were found between these two groups. Moreover, ALSM was significantly different from every other lake in this region. State-wide, well-defined genetic clusters were identified by the Bayesian clustering analysis. The DK of Evanno et al. (2005) peaked at K = 4 and K = 9 clusters and the posterior probability was the highest for K = 9 clusters (Figure S3). A structure of K = 6 clusters appeared the more meaningful with two lakes Mille Lacs and Prior and the three geographic clusters Alexandria, Brainerd and Detroit Lakes standing out (Fig. 3a and S4). Interestingly and in accordance with the genetic differentiation tests, we observed some sub-structure within the Alexandria cluster, where two (or perhaps three) different genetic clusters were resolved (Fig. 3a). Lakes within the Brainerd and Detroit Lakes regions, however, each grouped into a single regional cluster in which no further substructure was present (Fig. 3a). These results were confirmed by running the STRUC- TURE software within each of the three geographic regions separately (Fig. 3b). Similar grouping results were obtained with the neighbor-joining tree (Figure S5). The analysis of molecular variance revealed that the great majority of the genetic variance was explained by the variation between individuals within populations (95.3%), in accordance with the high within-population genetic diversities observed. The proportion of variance between lake-rich regions (3.2%) was twice the value between populations within regions (1.5%). Lack of support for the hub lake hypothesis Each one of the 252 ABC analyses of hub origins gave conclusive results. The scenario of independent introductions where Lake Mille Lacs (or Prior Lake) was not the source for infestation of another inland lake that was infested later was selected in every analysis with high posterior probabilities, ranging from to 1.00 for the analyses of Lake Mille Lacs (35 Lakes,

13 Dispersal mechanisms for zebra mussels Fig. 3 Genetic structure of zebra mussel populations in MN. Bar plots of the coefficients of co-ancestry obtained in STRUCTURE analyses with several values of K for (A) the entire dataset and (B) samples within the lake-rich regions: (i) Brainerd, (ii) Alexandria and (iii) Detroit Lakes. Each vertical bar corresponds to one mussel (bar height = posterior probability of cluster membership) and each cluster is represented by a color. The tested populations and the major clusters that discriminate lakes are indicated below and above the figure, respectively (see Table 1 for population label codes). The most frequent clustering patterns per K value, identified with Clumpp and represented with Distruct, are shown (the minor ones, which do not exhibit important differences, are shown in Figure S4 for the entire dataset, when applicable) Table 2) and from to for the analyses of Prior Lake (28 lakes, Table 3). Non-overlapping 95% confidence intervals for the probabilities were obtained in every case and posterior error rates were globally low and ranged from 0 to (Tables 2, 3). These results were robust to changes in sets of priors and sets of population samples tested from the focal source (three different samples from Lake Mille Lacs and from Prior Lake were tested, Tables S4 and S5). Finally, the selected scenarios and their posterior distributions were able to simulate data very similar to that observed. Only one SuSt (out of 4032, in total across all the analyses) was in the extreme 5% tails of the distribution of simulated SuSts after correction (data not shown). Stratified dispersal to explain the spread of zebra mussels We performed ABC analyses within lake-rich regions to investigate the pattern of spread of zebra mussels within them. Different levels of support and robustness for the ABC analyses were observed for each of the three regions. For the Alexandria region, the ABC analysis was conclusive. Among the 24 scenarios being contrasted, Scenario 16 was selected with a moderately high posterior probability of and a 95% CI ( , Table 4) that did not overlap with any other scenario. Scenario 16 describes 3 independent events of introduction within the Alexandria lake-rich region and 1 subsequent successive introduction event: the 1st introduction into Lake Carlos (ALC),

14 S. Mallez, M. McCartney Table 2 Results of the scenario comparisons performed with ABC analyses to investigate the role of Lake Mille Lacs as secondary source for invasion Destination lake samples Introduction scenario Prior set 1 Prior set 2 Posterior probability [95% CI] PER Posterior probability [95% CI] PER ALBRO Independent [0.9882, ] [0.9701, ] Successive [0.0093, ] [0.0240, ] CAL Independent [0.9999, ] [0.9994, ] Successive [0.0000, ] [0.0004, ] BLC Independent [0.9994, ] [0.9999, ] Successive [0.0004, ] [0.0001, ] DLCRY Independent [0.9995, ] [0.9989, ] Successive [0.0003, ] [0.0008, ] ALCWD Independent [0.9991, ] [0.9971, ] Successive [0.0006, ] [0.0020, ] BLGIL Independent [0.9988, ] [0.9986, ] Successive [0.0008, ] [0.0009, ] ALGV Independent [0.9757, ] [0.8784, ] Successive [0.0199, ] [0.1085, ] BLLH Independent [0.9997, ] [0.9990, ] Successive [0.0002, ] [0.0006, ] ALIDA Independent [0.9985, ] [0.9458, ] Successive [0.0009, ] [0.0469, ] ALIR Independent [0.9963, ] [0.8931, ] Successive [0.0026, ] [0.0957, ] ALMT Independent [0.9930, ] [0.8971, ] Successive [0.0051, ] [0.0919, ] LSLC Independent [0.9993, ] [0.9927, ] Successive [0.0005, ] [0.0055, ] ALV Independent [0.9840, ] [0.8592, ] Successive [0.0128, ] [0.1274, ] LW Independent [0.9999, ] [0.9997, ] Successive [0.0000, ] [0.0002, ] DLMEL Independent [0.9980, ] [0.9724, ] Successive [0.0014, ] [0.0229, ] ALMP Independent [0.9946, ] [0.9072, ] Successive [0.0038, ] [0.0826, ] DLNLID Independent [0.9994, ] [0.9975, ] Successive [0.0004, ] [0.0018, ] DLNLIZ Independent [0.9908, ] [0.9671, ] Successive [0.0070, ] [0.0261, ] DLNL Independent [0.9999, ] [1.0000, ] Successive [0.0001, ] [0.0000, ] DLORW Independent [0.9991, ] [0.9985, ] Successive [0.0006, ] [0.0010, ] BLP Independent [0.9999, ] [0.9862, ] Successive [0.0001, ] 0.0 [0.0108, ]

15 Dispersal mechanisms for zebra mussels Table 2 continued Destination lake samples Introduction scenario Prior set 1 Prior set 2 Posterior probability [95% CI] PER Posterior probability [95% CI] PER DLPRA Independent [0.9969, ] [0.9940, ] Successive [0.0023, ] [0.0043, ] BLRO Independent [0.9898, ] [0.8756, ] Successive [0.0079, ] [0.1120, ] ALSM Independent [0.9988, ] [0.9675, ] Successive [0.0007, ] [0.0273, ] Results shown are for Lake Mille Lacs sample MLBP, see Table S4 for the results with 2 other samples. Posterior probabilities and their 95% confidence intervals (CI) were obtained using a polychotomous logistic regression (Cornuet et al. 2008, 2010) and the LDA option (Cornuet et al. ; Estoup et al. 2012). PER corresponds to Posterior Error Rate, see main text for details. The codes of the samples are used for simplicity, see Table 1 for details and full description of the samples. See Figure S2 for a description of the two scenarios contrasted which then infested Lake Irene (ALIR), a short distance upstream, the 2nd introduction into Lake Darling (ALD) and a 3rd introduction into Lake Mary (ALSM, Figs. 2 and 4a). The only other scenario obtaining a posterior probability higher than 0.05 was Scenario 14, with a probability of ( , Table 4). This scenario is very similar to Scenario 16, the only difference between them being the origin of Lake Mary (ALSM), independent from every other lake in Scenario 16 and coming from Lake Darling (ALD) in Scenario 14 (Fig. 4). These results were robust to changes in prior sets and in lake population samples tested (Table S6). The posterior error rates obtained for every analysis were quite high and ranged from 0.49 to However, more careful examination of posterior error rates revealed that, in all analyses, Scenario 16 was not frequently selected in place of another scenario (other than Scenario 14) and vice versa, suggesting that it was distinguishable and that the high PER value was mainly due to swapping between the other unfavored scenarios (Table S7). Finally, the selected scenario and its posterior distributions were able to simulate data very similar to the observed ones. None of the observed SuSts was in the extreme 5% tails of the distribution of simulated SuSts after correction (data not shown). For the two other regions Detroit Lakes and Brainerd, the posterior probabilities were low, ranging from to for the Detroit Lakes region and from to for the Brainerd region, with some overlapping 95% CIs (Table 4). None of the 24 scenarios was thus reliably selected in these regions, likely due to the lack of genetic differentiation and structure. However, the scenarios with the highest posterior probabilities (and overlapping 95% CIs) share some similarities: Pelican Lake (west access site: DLPWA) was the source for Prairie Lake (DLPRA), a short distance downstream, Orwell Reservoir (DLORW) was the source for Melissa Lake (DLMEL), a long distance upstream; and Lake Ossawinnamakee (BLOW) was the source for Pelican Lake (BLP), a short overland distance south, for the Detroit Lakes and the Brainerd regions respectively (Figs. 2 and S1). Discussion In this work, we studied the genetic features of zebra mussel populations at the scale of a US state. We provide the first large dataset and genetic analyses that are capable of examining dispersal of zebra mussels to individual water bodies and of resolving patterns of spread within the state at a regional scale smaller than most European countries. While doing so, we observed high within-population genetic variation and genetic structure strong enough to identify several well-delimited population clusters. Our genetic data did not support the existence of hub lakes to explain the spread of zebra mussels. They revealed instead that short-distance dispersal dominates in lake-rich regions, in accordance with the stratified dispersal mechanism.

16 S. Mallez, M. McCartney Table 3 Results of the scenario comparisons performed with ABC analyses to investigate the role of Prior Lake as secondary source for invasion Destination lake samples Introduction scenario Prior set 1 Prior set 2 Posterior probability [95% CI] PER Posterior probability [95% CI] PER ALD Independent [0.8306, ] [0.8648, ] Successive [0.1574, ] [0.1189, ] ALBRO Independent [0.7963, ] [0.7938, ] Successive [0.1897, ] [0.1894, ] CAL Independent [0.9987, ] [0.9996, ] Successive [0.0009, ] [0.0002, ] BLC Independent [0.9955, ] [0.9958, ] Successive [0.0034, ] [0.0029, ] DLCRY Independent [0.9373, ] [0.9676, ] Successive [0.0545, ] [0.0248, ] ALCWD Independent [0.9697, ] [0.9581, ] Successive [0.0254, ] [0.0329, ] BLGIL Independent [0.9562, ] [0.9535, ] Successive [0.0373, ] [0.0393, ] BLLH Independent [0.9821, ] [0.9899, ] Successive [0.0146, ] [0.0074, ] ALIDA Independent [0.9968, ] [0.9147, ] Successive [0.0023, ] [0.0755, ] ALIR Independent [0.9508, ] [0.7696, ] Successive [0.0424, ] [0.2159, ] ALMT Independent [0.8824, ] [0.7962, ] Successive [0.1061, ] [0.1896, ] LSLC Independent [0.9869, ] [0.9503, ] Successive [0.0104, ] [0.0425, ] ALV Independent [0.8439, ] [0.7415, ] Successive [0.1430, ] [0.2432, ] LW Independent [0.9913, ] [0.9963, ] Successive [0.0066, ] [0.0024, ] DLMEL Independent [0.9925, ] [0.9929, ] Successive [0.0057, ] [0.0051, ] ALMP Independent [0.9806, ] [0.8658, ] Successive [0.0159, ] [0.1215, ] DLNLID Independent [0.9627, ] [0.9701, ] Successive [0.0315, ] [0.0238, ] BLNL Independent [0.9980, ] [0.9975, ] Successive [0.0014, ] [0.0017, ] DLORW Independent [0.9635, ] [0.9696, ] Successive [0.0309, ] [0.0237, ] BLP Independent [0.9965, ] [0.9900, ] Successive [0.0026, ] [0.0068, ] DLPRA Independent [0.9291, ] [0.9475, ] Successive [0.0617, ] [0.0413, ]