Population genomics traces zebra mussel invasions along the expanding western North American front in Minnesota, USA to their source waters

Abstract

Zebra mussels (ZMs) continue to transform aquatic ecosystems, threaten native species, and accrue high economic costs in the Northern Hemisphere. In 2024, Minnesota (MN) in the western Great Lakes became the state with the most ZM-infested lakes in the USA. We used population genomics to examine recent ancestry and infer the source water bodies from which ZMs colonized > 1/3 of MN lakes infested from 2003 to 2018 (when spread accelerated), then compared results to traffic between lakes of boaters and anglers, the suspect invasion vectors. Lake Superior, the Upper Mississippi River, Lake Erie, and several inland lakes were our top-ranked inferred sources. We traced ZMs in 51 of 58 infested lakes to source water bodies in-state, but most were not from alleged “superspreaders” (sensu epidemiology). Mille Lacs Lake, a popular angling destination and boater network hub, was a notable exception as the inferred source for three lakes. In three MN lake-rich regions in which spread continues to be concentrated, invasions from nearby were common, and sources were most often not high boat-traffic lakes, suggesting that vectors other than trailered boats need further evaluation. Geographic expansion of our population genomic dataset could provide genomic surveillance and guide prevention of the continuing spread of ZMs.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-27300-6.

Introduction

Zebra mussels (ZMs) Dreissena polymorpha Pallas 1771 and quagga mussels D. rostriformis bugensis Andrusov 1897 have transformed freshwaters more than any other aquatic invasive species (AIS)^1–4. These highly destructive bivalves continue to expand across the Northern Hemisphere, challenging invasion scientists and managers to understand and prevent spread. ZM invasions occurred in two phases in Eurasia and North America, the first a rapid spread following the construction of navigable passageways^5–8. In Eurasia, this began 200 years ago with canals connecting native-range rivers on the Black, Azov, and Caspian Seas to Baltic and North Sea rivers. Mussels attached to ships and cargo then rapidly colonized central and western Europe. The second phase, the invasion of waterbodies unconnected to navigable waters, was in the 19th and early twentieth centuries, so few details are clear. In Belarus, commercial fisherman moved nets between unconnected lakes, spreading mussels in the mid twentieth century⁹, while recreational watercraft caused overland spread in Spain¹⁰, Ireland⁵, and Switzerland¹¹.

In North America, the first phase followed the opening of the St. Lawrence Seaway in 1959. Here the likely vector was transatlantic cargo ships that discharged ballast water containing the veliger larvae that colonized Lakes Erie and St. Clair in the mid 1980’s^12,13. Then mussels spread to the remaining Laurentian Great Lakes, and to the Mississippi and other large river systems in just five to ten years, occupying much of the extent of their present North American range¹⁴. Meanwhile, secondary invasions of inland lakes continue to emerge, caused in part by larval dispersal. The dreissenid veliger is a planktonic feeding larva that during its 2 to 5-week lifespan before settlement^15,16 can potentially travel long distances within lakes and streams. Studies finding that lakes downstream have been infested more often than upstream or unconnected lakes show that downstream dispersal of veligers plays an important role^17,18, but most range expansion is attributed to overland transport by recreational boats^19–23.

Slowing or blocking spread by boaters is therefore the highest priority for North American management. The strategies differ by region. In the western USA, the small number of infested waters has favored watercraft inspection and decontamination programs designed to block contaminated watercraft from crossing state lines or launching on lakes that are not yet infested²⁴. In contrast, in the most infested regions of the country (surrounding the Great Lakes), prevention programs are stretched so thin by the geographic extent and number of infested lakes, that management prioritizes blocking invasions at their sources, whereas inspections on lakes that are not yet infested are much less common.

Prevention strategy is a formidable challenge in the Great Lake state of Minnesota (MN), where the North American ZM invasion is most rapidly expanding at its western front. The “land of 10,000 lakes,” with > 13,000 water bodies > 0.04 km² surface area, become the state with the most infested lakes in the USA in 2024 (Supplementary Fig. S1). Minnesota has responded with aggressive prevention. Together with local government, the MN Department of Natural Resources (MN DNR) in 2024 inspected > 450,000 and decontaminated > 4,100 watercraft departing ZM-infested waters²⁵. Determining where to position personnel across a state that is third in the USA in registered boats and first in boat ownership per capita²⁶ is a daunting task. Minnesota DNR uses boat traffic, prevalence of higher-risk vessels, and water body infestation status to distribute inspectors²⁷ who screen watercraft outbound from the busiest infested lakes.

Analysis of boater and angler traffic can assist managers tasked with these decisions on how to distribute prevention effort. Earlier on, models of human transportation behavior were adapted to study the pattern and risk of spread of zebra mussels between lakes. In so called “gravity models,” boat traffic and hence the likelihood of ZM transport were formulated to increase with factors (e.g. number of boat ramps, lake size) that create “attraction” between pairs of lakes and to decrease with distance—with analogies to gravitational force. Gravity models yielded a slower rate of invasion of new lakes that more closely matched true rates and made better predictions about which lakes were colonized than did earlier attempts^20,28,29. More recently, data from smartphone apps has been tapped to build angler network models^30,31, with the advantage being access to a vastly larger set of records than watercraft inspections and boat registrations provide.

Analysis of boater traffic tracks possible vectors, and not the organisms themselves. Invasion genetics and genomics, on the other hand, can trace introductions of invasive species to source populations, and while these studies are often at continental and global scales^32–36, increasingly they are providing finer spatial resolution^37–39. ZMs are managed in the USA by states or sometimes regions (e.g. Pacific Northwest), so prevention could benefit most from invasion tracking to source waterbodies. Tracking boats and genetically tracking mussels, however, each has its limitations, so this study combines the two complementary approaches. Our findings paint a picture that differs somewhat from expectations. Large, high-traffic inland lakes that are suspect AIS superspreaders infested few lakes in the first ten years as the MN invasion accelerated. Instead, MN sites on Lake Superior, the Mississippi River and several unsuspected inland lakes were the inferred sources for most infestations and spread between neighboring lakes in lake-rich regions predominated.

Results

MN ZM invasion chronology

ZM invasions of MN lakes rose sharply after 2009 (Supplementary Fig. S1). From 2014 to present, MN added the most newly infested lakes each year across the seven Great Lake states (Illinois, Indiana, MI, MN, New York, Ohio, and Wisconsin) that continue to accrue most US invasions. In 2024, 87% of MN infested lakes were in the six regions in which we focused sampling (Supplementary Tables S1, S2; Fig. 1, Supplementary Figs S2, S3). By 2018, when sampling ended, ZMs were found in 155 MN inland lakes, and as has been shown elsewhere^8,40, larger lakes were infested more often than smaller ones (Supplementary Fig. S4). As a result, only about 1% of 13,933 lakes, but 16.4% of total lake surface area (including 8 of the 19 largest lakes, > 40 km²) was ZM-infested.

Fig. 1.

Map of Minnesota regions infested with ZMs. Regions contain counties with a concentration of infested lakes. Regions are demarcated by county borders and in some cases, by separations between watersheds. For example, the Alexandria Lakes region contains Douglas and Pope counties and, in Todd County (the northeast section) that area within the Long Prairie River watershed. Some sampled, putative source waters are shown in red text. Numbered locations along the Mississippi and St. Croix Rivers denote sampling sites. The Northern-most reaches of the Mississippi cannot be viewed at this resolution. See Supplementary Fig. S2 and S3 to view its course through the Brainerd Lakes and Mississippi Headwaters regions. The outline map was downloaded from MN Geospatial Commons (https://gisdata.mn.gov) and modified using Adobe Illustrator (version 29.8.1, https://www.adobe.com).

Genotyping-by-sequencing metrics

Genotyping-by-sequencing (GBS) yielded 21,702 variants, of which 13,809 were biallelic single nucleotide polymorphisms (SNPs). A total of 4199 variants across 1900 mussels were retained after filtering using VCFtools. A median number of 509,144 reads were generated per genotyped mussel. A total of 98.0% of mussels were sequenced with > 100,000 reads and 51.9% with > 500,000 reads, and the average depth of coverage per variant per individual was approximately 250X. Fastq read quality scores were 30.5 on average, and > 90% of reads aligned to the ZM reference genome in 93% of mussel samples. In Plink, pruning of markers in linkage disequilibrium and filtering missing data retained 1890 mussels and 5775 SNPs and yielded nearly identical metrics of quality, depth of coverage, and alignment rates. The genotyping rate for the LD-pruned, filtered dataset was 0.936, compared to the genotyping rate of 0.753 for the unfiltered data set.

SNP diversity within waterbodies

As expected, pairwise differences (π per number of nucleotide sites) was far less sample size dependent than S (number of segregating sites) or Watterson’s θ_W (Supplementary Fig. S5). Across waterbodies that must span magnitudes in population size, we found no significant correlation between lake size and π (Kendall’s τ = − 0.0758, P = 0.38) and π in populations in the smallest and largest lakes overlapped greatly (Supplementary Fig. S6).

In all comparisons within larger, highly infested lakes that were sampled at multiple locations, values of F_ST (a measure of the proportion of genetic diversity attributed to allele frequency differences between populations) between sites within lakes were close to zero (Supplementary Table S3). The only exception was Pelican Lake in Otter Tail County, in which mussels in the main lake were significantly differentiated from mussels in the connected southern portion (called Fish Lake), but at levels less than between unconnected lakes. Spatial coverage was broadest in Mille Lacs, where we found no within-lake substructure. In contrast, we found differentiated populations coexisting in two of the Great Lakes (Supplementary Table S4). In Lake Superior, we found the population from Duluth/Superior Harbor (hereafter, “Duluth Harbor”) to be genetically distinct. Duluth Harbor showed moderate differentiation from all other Lake Superior populations, including the Superior, Duluth (main lake) population, collected just 1.7 km from the harbor inlet (F_ST = 0.060). In Lake MI, mussels from the southwest shore population near Milwaukee, Wisconsin (WI) were moderately divergent (F_ST = 0.027) from the southeast shore population near New Buffalo, MI (Supplementary Table S4).

Genetic structure within and between infested waters

In our Discriminant Analysis of Principal Components (DAPC), Bayesian Information Criterion values, plotted against the number of genetic clusters (K: Fig. 2) dropped and leveled out at K ≥ 4. A-score optimization favored 44 principal components, explaining about 40% of the cumulative variance, and discriminant analysis yielded three distinct clusters (Fig. 2). Cluster one contained populations in Lakes Erie, St. Clair, Huron and MI (New Buffalo), cluster two contained the MI (Milwaukee), the Illinois River, and all Lake Superior populations except Duluth Harbor, and cluster three contained Duluth Harbor and the Mississippi and St. Croix River populations.

Fig. 2.

DAPC results obtained using a-score optimization, Great Lakes and Upper Mississippi River. (A) Discriminant function axis (DA) DA2 scores (Y-axis) plotted against DA1 scores (X-axis). (B) DA3 scores plotted against DA1 scores. K = 4 clusters, number of retained principal components = 44. Mussels were not assigned to populations a priori. (C) Plot of Bayesian Information Criterion values against the number of K clusters.

Ancestry (Q)-plots from Admixture revealed patterns of genetic structure between lakes in each region that emerged with increasing K (Supplementary Fig. S7 is an example). Cross-validation error dropped steadily from K = 2 to 12, reaching a plateau at K = 9 (Supplementary Fig. S8). Q-plots at K = 9 (Fig. 2) revealed strong differentiation among and within different lake regions of MN. Genetic structure was notable within lake-rich regions (Figs. 1 and 3)—even between adjacent connected lakes, in some cases (e.g. Supplementary Fig. S9). Differentiation in MN was much greater than we found between large, popular recreational lakes in WI (Supplementary Fig. S10) well dispersed across the state (Supplementary Fig. S11) and infested from 1995 through 2017. Eight of nine WI lakes had a genetic signature like Lake MI (Milwaukee), Illinois River, and main Lake Superior populations but unlike populations in the Mississippi (which borders WI to the west).

Fig. 3.

Q-plots from Admixture analysis, MN inland lakes. Waterbody labels correspond to lakes (e.g. Rice = Rice Lake) except where indicated. Each stacked bar represents results for a single mussel. The height of each of the colored bar segments equals the ancestry proportions assigned to each of nine genetic clusters, and the legend labels the clusters.

Groups of genetically similar lake populations that were distinct from other lake populations were present in each of our six MN lake regions (Figs. 1 and 3). Genetic cluster four (yellow bars), for example, composed about 60% of the ancestry of mussels from five of the Detroit Lakes and five of the Brainerd Lakes, while contributions from this cluster were small in other lakes in both regions. One group of the Alexandria Lakes shared similar ancestry fractions from clusters one (dark blue) and four, while another group shared ancestry fractions from clusters one and five (orange) that were similar, but more variable. We also found region-specific genetic signatures, in which genetic clusters that were uncommon elsewhere were spread across most or all lakes in a region. For example, cluster eight in Detroit Lakes (brown) and cluster nine in Brainerd (grey) showed similar frequencies in all mussels from all lakes. Cluster one was rare in the Great Lakes and Mississippi, and across MN—except for Alexandria, where it was frequent in most mussels from 15 of 18 lakes. These patterns suggest scenarios in which mussels that founded populations were spread to multiple lakes in each region.

Other patterns pointed to longer-distance colonization events. Cluster seven (black) contributed > 95% to the ancestry of most mussels from Mille Lacs Lake, but only Lakes Clearwater, Stella, and Washington (located > 100 km from Mille Lacs) displayed this cluster, where its contribution was > 80% (Fig. 3). Genetic cluster four-derived ancestry ranged from 60 to 80% in all mussels collected from each of the Great Lakes [except Superior and one Lake MI population: (Fig. 4)], and this cluster contributed > 50% to the ancestry of mussels from 16 MN lakes, tracing their possible origins to the Great Lakes, or intermediate inland waters. These and other results guided formulation of scenarios tested with DIYABC-RF, the results from which we turn to next.

Fig. 4.

Q-plots from Admixture analysis of populations from the Mississippi River and tributaries, and the Great Lakes. Each stacked bar represents results for a single mussel. The height of each of the colored bar segments equals the ancestry proportions assigned to each of nine genetic clusters, and the legend labels the clusters.

Results from invasion scenario testing

The Great Lakes were the chosen invasion sources for 22 of 58 MN lakes (Supplementary Tables S5, S6; Fig. 5). Superior was the most often favored Great Lake, in fact the leading source that we tested. We inferred that 16 lakes (with posterior P ranging from 0.715 to 0.979, mean = 0.835) were infested from the Superior (Duluth) population (12 with admixture) and five from Duluth Harbor (0.827–0.937, mean = 0.877; none with admixture). Superior invasions were found in five of our six lake regions, and it was the favored source for the invasions of high traffic lakes Minnewaska (0.839) and Minnetonka (0.724; both with admixture) and several other lakes in Alexandria and the Twin Cities Metro. Duluth Harbor was the favored source for high-traffic lakes Mille Lacs (0.877) and Pelican (0.937), and for Pike (0.827), a small lake just 15 km from the harbor. Erie was favored in the infestation of six lakes (0.743 – 0.962, mean = 0.857; four without and two with admixture). In the four scenario tests in which Erie was favored as the sole source, the tests contrasted each of the other Great Lake populations (Huron, MI New Buffalo, and St. Clair) with which Erie shared similar ancestry profiles, so Erie was clearly the favored source. MI (Milwaukee) was chosen for two lakes (0.779, 0.754) with admixture, and St. Clair for one lake (0.931) without admixture.

Fig. 5.

Long-distance invasions of MN lakes. (A) Inferred source populations in the Great Lakes and Upper Mississippi River. Lakes Superior, Erie, Michigan, and St. Clair (in order of frequency) were the chosen sources for 22 of the tested MN inland lakes, and the Upper Mississippi River and tributaries tallied another 8.5 invasions. Lake Ontario was the only Great Lake not sampled. (B) Populations in MN waterbodies were the inferred sources for invasion of 51 out of 58 tested lakes. Populations in Lake Superior (purple arrows), the Mississippi River (orange arrows), and inland Lakes Pleasant (yellow arrows) and Mille Lacs (green arrows) were inferred in fifteen, seven, four, and three infestations, respectively. Maps were created using personal R scripts.

A large majority (87.3%) of invasions were assigned to sources within MN (Supplementary Tables S5, S6; Fig. 5). This is true, even though for all, but two of the 58 destination lakes with conclusive test results (Carlos and Pike), the scenarios included one, and more often two or more source water bodies outside MN. Note that in-state tallies included Duluth Harbor, Superior (Duluth), and the Upper Mississippi River (UMR). Treating Superior and the UMR as in-state assumes that mussels came from populations in MN, but a lack of geographic differentiation leaves the precise origins of these invasions ambiguous. We genotyped Superior populations from several sites in WI (Apostle Islands) and MI (Isle Royale National Park). Q-plots across all Superior sites were very similar (Fig. 4) and F_ST values were near zero (Supplementary Table S4) confirming a lack of differentiation between each of these and the main lake population off Duluth (Duluth Harbor is well differentiated). Mississippi River sites we sampled came only from MN and WI border waters and spanned about 340 river km distance, and we found no differentiation over this scale (Fig. 4), so we may not be able to distinguish Mississippi populations within MN from those outside the state.

The Mississippi River (Figs. 4 and 5) was the sole source for three invasions: Ossawinamakee (0.980) and Gull (0.996) in Brainerd Lakes, and Carlos (0.996) in Alexandria Lakes. The Mississippi also contributed to seven two-source invasions—four in the North Central Hardwoods (0.715–0.779, 0.747), and three in the Twin Cities Metro. The latter were the connected lakes Pleasant, Sucker and Vadnais (each = 0.797) comprising the Vadnais Chain, each infested in 2007. Ossawinamakee Lake was the first natural inland lake infested in MN (2003) and from it spread genotypes of UMR origin across north-central MN, while in west-central MN, Lake Carlos near Alexandria (infested in 2009) played a similar role.

We traced about half (46.6%) of in-state invasions to MN inland lakes (Supplementary Table S5, Figs. 5 and 6). Top ranked source lakes were Lizzie (0.828, 0.806, 0.816, 0.805), Minnewaska (0.816, 0.970, 0.783, 0.879), Rice (0.970, 0.948, 0.966, 0.855), and Pleasant (0.843, 0.779, 0.974, 0.773), each of which tallied four lakes (without admixture) in the Detroit Lakes, Alexandria, Brainerd and multiple regions, respectively. Next were Geneva (0.770, 0.812, 0.819) and Mille Lacs (0.980, 0.996, 0.989), each with three infested lakes, also without admixture. Carlos tallied three (0.804, 0.900, 0.839), and Pelican (0.897, 0.805) and Ossawinamakee (0.743, 0.810) two lakes, each with admixture. Finally, Cass (0.947) and Mary (0.841) were credited with one infestation each, without admixture. Spread within our three clustered-lakes regions to lakes nearby (Fig. 6) was a recurrent pattern.

Fig. 6.

Invasions from nearby lakes in three lake-rich regions. (A) Detroit Lakes region. Lizzie was the inferred, sole source for four nearby lakes; Pelican (two-source, admixed with Lake Superior) was the inferred source for two. (B) Alexandria region. Several infestations from nearby lakes were inferred. Carlos and Geneva were the sources for three lakes each (with admixture in each Carlos infestation; the Geneva infestations were sole source). Minnewaska was the sole source for four lakes. (C) Brainerd region. Rice was the inferred source for four lakes, all sole source. Upstream connecting waterways and/or navigable channels could account for only one of the invaded lakes shown in this figure: Prairie in Detroit Lakes; all others must have been overland. Maps were created using personal R scripts, then were modified using Adobe Illustrator (version 29.8.1, https://www.adobe.com).

Source lakes inferred in spread to multiple lakes within MN regions did not consistently rank high for boater and angler traffic. Lizzie in Detroit Lakes was chosen as the source for four lakes but is absent from the lists of lakes ranked for number of inspections and for what we refer to as “traffic” (inspections per hour) at access points (Supplementary Tables S7, S8). In contrast, Pelican, the next lake connected by its outlet river upstream from Lizzie, was ranked in the top 10, both for effort and traffic but Pelican was inferred in only two lakes with admixture (the equivalent of one lake infestation). Rice was the source for four lakes in Brainerd but was absent from any of our lists that estimate usage, whereas no lakes in MN were traced to invasion from Gull. This highly visited lake was ranked fourth in inspections, 13th in traffic (Supplementary Tables S7, S8) and was in the top five of all MN lakes, ranked for connectivity in boater travel networks⁴¹. Gull was also 28th of 100 lakes ranked by connectivity based on independent analysis of angler travel data³⁰ (Supplementary Table S9).

Chosen source lakes in Alexandria more closely fit expectations from measures of recreational use. The inferred source for four lakes infested in Alexandria was Minnewaska. This is a moderate to high-traffic lake, 17th in number of inspections (Supplementary Table S7) and 2nd in inspections/hour (SupplementaryTable S8), and it is the only Alexandria Lake in the top 100 (85th) for connectivity based on Fishbrain data³⁰ (SupplementaryTable S9). Geneva infested three lakes nearby (without admixture) and has moderate traffic: 42nd in inspections and 48th in inspections/hour. Carlos infested three neighboring lakes (with admixture in each case). It ranked 14th in inspections and 49th in inspections/hour. Finally, Mary infested one lake and is absent from all our lists that rank usage.

Gull, Minnetonka, Prior and Mille Lacs are high-traffic recreational lakes, and each has been implicated as a potential superspreader, yet we inferred just three infestations from these lakes, all from Mille Lacs (Supplementary Tables S5, S6, S10). Particularly striking was no infestations from Lake Minnetonka, the focus for the highest prevention effort in the state due to high traffic. Its spread potential was corroborated by network analyses that ranked it the number one lake, most connected to lakes in MN by boaters⁴¹ and the number one lake connected by angler traffic to lakes in MN and throughout the USA³⁰. These same two analyses ranked Mille Lacs Lake number two and number six in MN, respectively [it was 8th for inspections and 55th for traffic (Supplementary Tables S7, S8)]. Mille Lacs invasions of Clearwater, Stella, and Washington would require transport of mussels > 110 km over land. High confidence in scenario testing (Supplementary Tables S5, S6) was supported by Admixture Q-plots from Mille Lacs that were nearly identical to plots from these three destination lakes (Fig. 3).

Discussion

We have presented here, for the first time, an analysis of genomic variation in ZM populations in MN inland lakes, the Great Lakes, the UMR and tributaries. Using thousands of SNP markers, we uncovered sufficient differentiation between populations to allow robust testing of alternative invasion scenarios. We found three well-differentiated lineages that either descend from separate transatlantic transport events from populations previously differentiated in Eurasia, or from colonizing populations that diverged thereafter in North America, and we identified putative descendants of those lineages in MN lakes. Altogether, with moderate to high confidence, we inferred the source populations for over a third of all MN lakes infested prior to 2018 and reconstructed a portion of this invasion history in some detail.

The Lake Superior population from near Duluth, MN was our leading invasion source. ZM populations in Superior tend to be smaller and more fragmented than in other Great Lakes^42,43 and are restricted to nearshore locations, possibly due to dissolved calcium concentrations that are too low to allow populations to persist away from riverine inputs and land runoff⁴³. Nevertheless, propagule pressure may still be high because trailered boat traffic connects Superior extensively to other MN lakes. Data from 2018 showed it to be the MN water body ranked 12th for inspections (Supplementary Table S7), and three Superior public accesses were in the state’s top 50 for inspections per hour (Supplementary Table S8). During interviews conducted from 2014 to 2017, MN watercraft inspectors asked boaters entering and departing infested lakes to identify the last lake(s) visited, and the next lake(s) they planned to visit, and these data were used to generate networks of connections between 9182 MN lakes by simulation. Superior was ranked 7th in MN for “betweenness centrality;” a measure of the lake’s importance in maintaining connectivity in the network⁴¹.

Most of the scenarios favoring Lake Superior origins involved two source populations, Superior admixed with a second water body. Admixture scenarios were constructed to test cases in which lakes showed large contributions from genetic clusters falling into two distinct lineages (from DAPC analysis: Fig.2). Admixture events could reflect single or multiple introductions from two or more differentiated sources, however their frequency may be overestimated, if other source waters derived from admixed origins were the true intermediate sources and were not sampled. Pleasant, Rice and Minnewaska are each possible examples of intermediate lakes, each of which we inferred to derive from two-source invasions. If these possible intermediate lakes had not been fortuitously sampled and analyzed, the four lakes we inferred to be infested from Pleasant, the four from Rice, and the four from Minnewaska (single source each) may instead have been inferred to have two-source origins.

ZMs may have colonized the Mississippi River in 1989 by traveling from the Great Lakes down the Chicago Sanitary and Ship Canal. In 1991, they are proposed to have spread downstream in the Illinois River and upstream in the UMR, where they were first discovered in MN reaches in 1991⁴⁴. Lake Ossawinamakee was the first infested natural lake in MN (2003), at which time the UMR was one of the very few state waters that could have been its source⁴⁵, and our testing inferred that it was. Altogether, the UMR was our second-ranked inferred invasion source, the favored origin of mussels colonizing lakes that became sources themselves—Ossawinamakee, Rice, the Vadnais Chain, Carlos, and others. Our results imply that, even with considerable prevention effort, the UMR has been an important source for MN invasions.

Our third ranked invasion source was Duluth Harbor. Scenario testing deduced that it colonized two high-traffic lakes infested early in the MN invasion chronology: Mille Lacs (in 2005) and Pelican (in 2009). The population genomic distinctiveness of Duluth Harbor was shared by mussels from two Lake Superior sites. The first is off Park Point within the Harbor, and the second is from the Thomas Wilson shipwreck, a short distance away, near the outlet of the St. Louis River, the largest tributary entering the Lake’s western arm. Lake Superior tributaries transport salts and nutrients of riverine origin to nearshore sites^46,47, so we suspect that the genomic similarity between Duluth Harbor and the shipwreck population derives from delivery of larvae born within the estuary and the harbor out to the wreck, in the St. Louis River plume.

Overall, the results from this study show a mixed level of concordance with expectations from recreational use. Since our estimates of traffic are based in part on watercraft inspection effort, one explanation for discord would be an effective watercraft inspection and decontamination program that fails to prevent infestation of lakes on which effort is less—a distinct possibility worth further evaluation. Nearly 200 lakes have been confirmed since sampling for this study ended. Several are located within our designated regions of high spread, and some are near high-traffic lakes. An updated study would provide new insights.

On the other hand, the multiple invasions traced to low traffic lakes suggest that we consider alternative spread pathways and mechanisms. Most trailered boats outbound from MN access ramps have been in the lake < 24 h prior to take out²⁷. This is too little time for adult or juvenile mussels to attach, and any settled larvae would not survive aerial exposure during overland transport. Instead, these transient boats may move veliger larvae in residual water left in compartments (live wells, bilge, and ballast tanks) after the user has drained the vessel at take out, or they may transport mussels on vegetation caught on trailers or motors²³. Resident boats (such as those in marinas), in contrast, spend weeks to months immersed before take-out, allowing attachment to hulls and other surfaces^23,48. Desiccation resistance, survivorship and likelihood of producing offspring then yield an establishment risk per transport event for adult mussels > juveniles > larvae ^23,49. And so, while traffic at MN inspection stations is predominantly transient boats, some unknown (but important) fraction could be resident boats from the Great Lakes, large rivers, or inland lakes.

Other high-risk vectors of possible relevance include “water-related equipment” (docks, boat lifts, and swim rafts for example). In 2011, the MN DNR intercepted two boat lifts infested with ZMs⁵⁰. One was linked to the infestation of Lake Irene in the Alexandria region. The owner reported that the lift had been installed from Lake Le Homme Dieu, 15 km away. We contrasted scenarios in which Le Homme Dieu was the source, to scenarios with other local as well as distant sources for Irene. The favored source was Geneva, the next connected lake upstream of Le Homme Dieu. In the second incident, a boat lift was transported from Lizzie to Rose (within the Detroit Lakes region). It was heavily fouled, and numerous mussels were found on the lakebed after the lift was removed in 2011. While we could not sample Rose (its population was successfully suppressed with molluscicide soon after this), we did identify four lakes infested from Lizzie each year from 2011 to 2014. Therefore, the practice of moving boat lifts between lakes could account for these infestations from Lizzie, and others. While MN statute presently requires storage out of water for 21 days prior to movement to a new water body (the rule did not exist in 2011), water related equipment transport remains a potentially important route by which mussels are spread, despite this rule.

In total, we subjected four putative hub lakes to extended testing and only Mille Lacs was inferred to be an invasion source. Due to an intensive program of invasive species monitoring, the Mille Lacs population was discovered when it was very small (2005), and it grew slowly but exponentially to massive densities by 2009⁵¹. Boat traffic leaving Mille Lacs is highly connected to other MN lakes^30,41. As a large (519 km²) and heavily infested waterbody, we expected it to have invaded several lakes nearby. Instead, we detected no Mille Lacs invasions in either Brainerd or the Mississippi Headwaters, the two closest regions. The three Mille Lacs-infested lakes are instead in the North Central Hardwoods, ≥ 110 km away. Lake Minnetonka is a large (57.5 km²) lake in the Twin Cities Metro that has generated much notoriety as a likely superspreader. By 2010, three invasive aquatic plant species as well as ZMs were discovered here⁴⁵. The authors of the network analysis of Fishbrain mobile app data projected Minnetonka to be the number one source in MN for ZMs and Eurasian watermilfoil for destination waters across the USA³⁰. It is possible that in the future, recently infested lakes not sampled in this study may reveal some Minnetonka invasions, but we have not detected any to date.

We stress that our favored invasion scenarios can be evaluated only in comparison to the tested alternatives, which are limited to waterbodies in our database. With expanded sampling throughout MN and surrounding regions, we expect some changes to waterbodies that we have inferred to be invasion sources. The outcome that results in exclusion (e.g. of a putative hub as the source for a tested destination lake) will not change, but the putative hub may be the inferred source for other more recent invasions.

In conclusion, this first population genomic study of an active front of the North American ZM invasion creates a genomic surveillance database that can be updated with newly infested waters and easily expanded. Continuing to grow this database has strong potential to advance our understanding of both the past and present-day spread of ZMs in North America (and maybe elsewhere). Our hope is that it can provide information in a timely manner to managers and help them develop strategies to combat this still rapidly growing invasion.

Materials and methods

Sampling from geographic populations

We collected samples from 69 ZM-infested waterbodies within MN (63 inland lakes, five rivers and streams, and one riverine lake), selected from the following regions of ongoing spread in the state: Alexandria Lakes, Brainerd Lakes, Detroit Lakes, Mississippi Headwaters, North Central Hardwoods, and Twin Cities Metro (Fig. 1). Together, lakes in these regions totaled 88.6% of inland lakes invaded as of October 2024, including the period of most rapid spread (2009–2024) and 93.5% prior to 2018 when sampling was discontinued (Supplementary Tables S1, S2). We targeted infested lakes with high levels of trailered boat traffic and collected mussels from 20 of the top 25 water bodies ranked by number of inspections and hours of inspection effort (Supplementary Table S3). We sampled putative sources with massive ZM populations from additional water bodies within MN (Lake Superior in MN, UMR in MN, St. Croix River, and Lake Mille Lacs: Fig. 1) and outside the state (Lake Erie, Lake St. Clair, Lake Huron, Lake MI, Illinois River: Supplementary Fig. S12, Supplementary Table S11). Each of the listed Great Lakes is a nationally ranked hub in networks of between-lake connections that were constructed using data from Fishbrain, a fisherman’s mobile app³⁰. We hand-collected 1893 individuals from 109 geographic populations across a total of 82 waterbodies (Supplementary Table S11) and dissected and froze mantle, foot, and gonad tissues. We extracted genomic DNA using the DNeasy kit (Qiagen, Germantown MD, USA). Extracts from mussels collected from sites in the Great Lakes in which zebra and quagga mussels co-occur were screened using a PCR–RFLP assay that distinguishes the two species⁵² to ensure that only ZMs were processed further for GBS.

Library preparation and sequencing

We constructed GBS libraries^53–55 and sequenced them at the University of Minnesota Genomics Center (Minneapolis MN, USA). We digested genomic DNA (200 ng) at 37 °C for 2 h with SbfI-HF [20 Units, New England Biolabs (NEB): Ipswich, MA] in NEB CutSmart buffer and heat-inactivated for 20 min at 80 °C. Next, we ligated the digested DNA to TGCA-overhang adaptors containing priming sites for Illumina multiplexing [Integrated DNA Technologies (IDT): Coralville IA, USA] at 0.1 μM final concentration each adapter, and incubated at 22 °C for 1 h, followed by heat inactivation of T4 ligase (400 Units, NEB) at 65 °C for 20 min. After a solid phase reversible immobilization bead purification⁵⁶ we amplified half the volume of the adapter-ligated DNA fragments in Next High-Fidelity 2X PCR Master Mix Taq (NEB) with unique barcoded primers (IDT) containing complementary flow-cell adaptor sequences in a final concentration of 0.5 uM for each primer (forward indexing primer: AATGATACGGCGACCACCGAGATCTACACXXXXXXXXTCGTCGGCAGCGTC and reverse indexing primer: CAAGCAGAAGACGGCATACGAGATXXXXXXXXGTCTCGTGGGCTCGG; “X” represent sample-specific barcodes) using the following cycling conditions: initial denaturation at 98 °C for 30 s followed by 18 cycles of 98 °C for 10 s, 55 °C for 30 s, 72 °C for 30 s with a final extension step at 72 °C for 5 min. We quantified the purified libraries with the Quant-IT PicoGreen dsDNA Kit (ThermoFisher: Waltham MA, USA), pooled by mass, and removed the adaptor dimers using SPRI bead purification reagent. We assessed the final library fragment size distribution using a Bioanalyzer 2100 High Sensitivity Chip (Agilent Technologies: Santa Clara CA, USA). We used a library dilution of 8 pM for single end 1X100 bp sequencing on an Illumina (San Diego CA, USA) NextSeq 550 instrument, with a targeted average depth of coverage of > 100X per variant per individual mussel genotyped.

Alignment and variant call analysis

We performed read-trimming, alignment and variant call analyses automatically using an in-house pipeline. We demultiplexed (i.e., each read was associated to the corresponding individual based on barcodes) and trimmed (i.e., the Illumina adaptors and barcodes were removed), then discarded reads without perfect barcodes and/or reads shorter than 50 bp. Next, we aligned the trimmed sequences against the ZM reference genome one individual at a time (individuals aligned independently from one another) via Burrows-Wheeler transformation using BOWTIE2 version 2.2.4, with parameters set to the default values. After this, we called single nucleotide polymorphism (SNP) markers using Freebayes version 1.1.0⁵⁴ with default parameters. Freebayes is a Bayesian multi-sample variant caller, used to jointly call variants across all samples simultaneously.

Filtering and identification of SNP loci

The unfiltered variant call format (VCF) file from Freebayes retained 21,702 variants and 1916 individual mussels at a genotyping rate of 0.753. The genotyping rate = the proportion of successfully called genotypes across all individuals and variants in the dataset and is calculated by dividing the number of non-missing genotypes by the total possible genotypes (number of individuals × number of variants). The VCF generated by Freebayes was filtered using VCFtools to remove variants with minor allele frequencies (MAF) of < 1%, variants with frequencies of missing genotypes ≥ 5% across individuals, and individuals with > 50% missing genotype data across variants.

A separate round of filtering was then performed to prepare input files for downstream analyses. In Plink, we filtered to retain only biallelic SNPs. Pruning of pairs of SNPs in linkage disequilibrium (LD) was then performed in Plink using the independent-pairwise option in 200 kb sliding windows, with a window shift after each step of one variant count, and a cutoff of ≥ 0.5 for the squared correlation coefficient between alleles at pairs of loci (r²). We then filtered to remove SNPs that failed to genotype in > 20% of individuals, and individuals that failed to genotype at > 50% of loci. The filtered file retained 5775 SNPs and 1890 samples (mussels), with a total genotyping rate of 0.936. This LD-pruned filtered file was then converted to input files for analyses in SambaR, Admixture, and DIYABC-RF, below.

Analysis of genomic diversity and population structure

We imported the data into R and stored it in a genlight object with the function 'read.PLINK’ of the R package adegenet-2.1.4^59,60. Using the ‘calcdiversity’ function in the R package SambaR⁶¹ we estimated the following genomic diversity measures: S = number of segregating sites, MAF = minor allele frequency (at segregating sites), MLH = multilocus heterozygosity, π = average number of pairwise nucleotide differences per site per population, and θ_W = Watterson’s⁶² estimate of the genomic diversity estimator θ (Supplementary Table S12). We also performed a Discriminant Analysis of Principal Components (DAPC)^63–65 in SambaR using the function ‘dapc’ of the R package adegenet-2.1.4, both with and without prior population assignment. In the latter case, we used the k-means method, with the value of K populations chosen as the lowest point on the graph of the Bayesian information criterion (BIC) value plotted against successive values of K. We examined the DAPC plots for consistency over a broad range for the number of retained principal components (PCs) as suggested⁶⁵. We used the a-score method (as implemented in adegenet) to optimize the number of retained PCs. The a-score measures the difference in the proportion of successful reassignments of individuals to groups, compared to their reassignment to random groups, as a function of the number of PCs retained in the analysis⁶³.

We also examined population structure using Admixture v.1.3⁶⁶ with default options (i.e. block relaxation for optimization, quasi-Newton convergence with q = 3 secant conditions for acceleration, and termination when the log-likelihood increases by less than ε = 10^–4 between iterations). We ran 10 replicate runs, simultaneously on all 1890 mussel samples genotyped at all 5775 SNP sites, each run with the number of K genetic clusters varied from K = 1 through K = 12, and we estimated cross-validation error across the entire dataset of 1890 mussels to evaluate the best value for K. We aligned the clustering results obtained from replicate runs at K = 2 through K = 9 (the latter with the lowest estimated cross-validation error in Admixture) and generated the summary ancestry “Q-plots” using Clumpak⁶⁷.

Contrasts of invasion scenarios

We used DIYABC-RF^68,69 to evaluate competing scenarios for invasion of Minnesota waterbodies. This program employs simulation-based Approximate Bayesian Computation (ABC) to construct scenarios of biological invasions, with their complex histories of colonization, population bottlenecks, divergence, and admixture events, combined with supervised machine learning (SML) methods to select among competing scenarios or to estimate parameters. Random Forest is an efficient SML approach in cases such as this one, in which large panels of individuals from numerous populations, genotyped at thousands of SNP loci are the data upon which scenario choice is based.

DIYABC-RF invasion model formulation and scenario testing

Our invasion scenarios tested in DIYABC-RF (see Fig. 7 for an example) are alternative descriptions of the history of colonization of water bodies by zebra mussels. This history takes the form of evolutionary models that consider four categories of events, interspersed by time periods of independent evolution. First is population divergence. This occurs after infestation of a new waterbody and is represented as a splitting event on a population genealogy in which the source population separates into two descendant populations, one that remains in the source waterbody and the other that has colonized the destination waterbody. No gene flow is permitted between source and destination waterbody after colonization. The three other categories of events depicted in the scenarios are changes in population size, admixture events, and sampling from populations.

Fig. 7.

Example invasion scenario: The Ossawinamakee Lake invasion. The highly favored invasion scenario five (Supplementary Table S6) is shown, with Ossawinamakee infested from the Mississippi River. Invasion dates are below waterbody labels. Tracing back in time proceeds from the bottom (t₀ = when samples were collected: these are coded to coincide with year of sample collection, so they differ) to the top of the figure; the top is the time at which the two basal lineages split from the common ancestor of all populations and gene lineages. Populations and gene lineages “merge,” going back in time, at times marked with thin horizontal black lines that are labeled with coalescent times (t_ij). Colored wider bars denote time periods of independent evolution of populations at constant effective size N_e (legend), and colored narrow bars denote population bottlenecks with smaller, variable N_e. Bottlenecks start after a new lake is colonized at t_ij and end at t_ij–db (minus because it is time towards present; time increases towards the past), where db = the bottleneck duration. Note that t_ij values are written with the lineage listed first being the one that persists, going back in time, preceding the coalescent event. Note also that time is not to scale. For example, colonization of the Great Lakes Erie, Huron, Michigan, and St. Clair took just five years, but these events are spread out to allow them to be viewed on the tree.

The ABC part of DIYABC-RF performs coalescent simulations on the data (genotypes at all SNP loci from each mussel) to generate datasets known as training sets—these are the data sets that the Random Forest module uses to “learn” relationships between the simulated molecular data and the simulated genealogies. Coalescent models are formulated by ordering events back in time, from present day back to the time when all lineages have coalesced into a single common ancestral lineage. A splitting event, running backwards in time, is represented as a “merge” event—in which the gene lineages in the colonized lake and in the source-lake coalesce to form ancestral lineages.

Due to the very large number of possible scenarios that could be formulated to account for invasion of any given waterbody, we sorted the 60 sampled ZM infested lakes in MN into six geographic regions that in 2018 included 93.5% of all lakes invaded in the state (Supplementary Table S2). We used a single, final criterion to determine whether a scenario test could be used to make inferences about invasion sources, one that assesses the compatibility of the priors and formulated scenarios with the observed dataset. The authors of the DIYABC-RF package describe a projection of the data from the simulations (i.e. the training set) and the observed data onto the first two Linear Discriminant Analysis (LDA) axes.^68,69 Combinations of scenarios and priors that do not yield an LDA plot in which the observed data falls within the vicinity of the clouds of points from the training set are unreliable, and no inferences should be drawn in such cases. Several exploratory scenario testing attempts were discarded because they failed this LDA-plot criterion, and two lakes (Maple in Alexandria and Little Sand in the Mississippi Headwaters) never produced satisfactory results . We used invasion dates and population genomic affinities between 58 “destination lakes” and putative source waters to reduce the list of alternative invasion scenarios to a manageable number, prior to analysis. For judging genomic affinities, we used results from DAPC (Fig. 2) to sort waterbodies into lineages, and plots of ancestry fractions in mussels from all populations generated with Admixture (“Q-plots” in Figs. 3 and 4) to choose the lakes and rivers that were the most likely candidate sources for the invasion of destination lakes.

As their different histories required, the steps taken in scenario testing were specific to each of the six geographic regions, but we utilized a common structure for all. First, we tested the earliest-infested (founding) lakes in a region. Next (to determine whether founding lakes became sources for later invasions), we used the favored scenarios for these founding lake invasions to construct the genealogical framework for testing more recently invaded lakes. This process often required scenario testing to be revised several times in each region as new results were obtained. Finally, we also tested some lakes in MN as potential invasion hubs (Supplementary Table S11). We selected them based on MN Watercraft Inspection Program data that showed these lakes to rank among the highest for number of inspections and number of inspections per hour of effort, and based on connectivity in boater travel networks constructed from these data (Supplementary Tables S7-S9). For priors, other settings, and examples of scenarios tested see Supplementary Information, pp. 69–82.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

M.A.M. conceived the research, performed all analyses (except for analysis of raw GBS data), wrote the original draft and archived data; S.M. collected samples and performed all laboratory work to generate raw GBS data; S.A., J.G. and D.M.G. developed the GBS assay for ZMs; J.G. extracted and analyzed raw GBS and SNP data; J.G., V.H.H.E. and D.M.G. revised and edited the original draft; M.A.M. and D.M.G obtained funding; V.H.H.E. created data visualizations and archived data.

Data availability

All data generated and analyzed during this study are included in this published article and its Supplementary Information files. Raw sequence reads have been deposited on the NCBI Sequence Read Archive under BioProject number PRJNA1254680. Input and output files from DIYABC-RF have been deposited on Zenodo: [10.5281/zenodo.17180521].

Declarations

Competing interests

The authors declare no competing interests.