www.nature.com/scientificreports OPEN Genetic connectivity of wolverines in western North America Casey C. Day1,33, Erin L. Landguth1,33, Michael A. Sawaya2, Anthony P Clevenger3, Robert A. Long4, Zachary A. Holden5, Jocelyn R. Akins6, Robert B. Anderson7, Keith B. Aubry5, Mirjam Barrueto8, Nichole L. Bjornlie9, Jeffrey P. Copeland10, Jason T. Fisher11, Anne Forshner12, Justin A. Gude13, Doris Hausleitner14, Nichole A. Heim15, Kimberly S. Heinemeyer16, Anne Hubbs17, Robert M. Inman18, Scott Jackson5, Michael Jokinen7, Nathan P. Kluge13, Andrea Kortello19, Deborah L. Lacroix20, Luke Lamar21, Lisa I. Larson12, Jeffrey C. Lewis22, Dave Lockman23, Michael K. Lucid9,10, Paula MacKay5, Audrey J. Magoun24, Michelle L. McLellan25, Katie M. Moriarty26, Cory E. Mosby10, Garth Mowat27, Clifford G. Nietvelt27, David Paetkau28, Eric C. Palm1, Kylie J.S. Paul29, Kristine L. Pilgrim5, Catherine M. Raley5, Michael K. Schwartz5, Matthew A. Scrafford30, John R. Squires5, Zachary J. Walker23, John S. Waller31, Richard D. Weir32 & Katherine A. Zeller5 Wolverine distribution contracted along the southern periphery of its range in North America during the 19th and 20th centuries due primarily to human influences. This history, along with low densities, sensitivity to climate change, and concerns about connectivity among fragmented habitats spurred the recent US federal listing of threatened status and special concern status in Canada. To help inform large scale landscape connectivity, we collected 882 genetic samples genotyped at 19 microsatellite loci. We employed multiple statistical models to assess the landscape factors (terrain complexity, human disturbance, forest configuration, and climate) associated with wolverine genetic connectivity across 2.2 million km2 of southwestern Canada and the northwestern contiguous United States. Genetic similarity (positive spatial autocorrelation) of wolverines was detected up to 555 km and a high-to-low gradient of genetic diversity occurred from north-to-south. Landscape genetics analyses confirmed that wolverine genetic connectivity has been negatively influenced by human disturbance at broad scales and positively influenced by forest cover and snow persistence at fine- and broad–scales, respectively. This information applied across large landscapes can be used to guide management actions with the goal of maintaining or restoring population connectivity. The wolverine (Gulo gulo) is the largest terrestrial member of the weasel family (Mustelidae) and is found in remote areas of the subarctic, alpine tundra, and boreal forests across North America, Europe, and Asia1. In 1University of Montana, 32 Campus Drive, Missoula, MT 59812, USA. 2Sinopah Wildlife Research Associates, Missoula, MT, USA. 3MSU Western Transportation Institute, Bozeman, MT, USA. 4Woodland Park Zoo, Phinney Ridge, WA, USA. 5USDA Forest Service, Missoula, WA, USA. 6Cascades Carnivore Project, Hood River, OR, USA. 7Alberta Conservation Association, Crowsnest Pass, Blairmore, AB, Canada. 8University of Calgary, Alberta, CA, Canada. 9Fish & Wildlife Service, US, USA. 10Idaho Department of Fish and Game, Boise, ID, USA. 11School of Environmental Studies, University of Victoria, Victoria, BC, Canada. 12Parks Canada, Lake Louise, AB, Canada. 13Montana Fish, Wildlife, and Parks, Helena, MT, USA. 14Selkirk College, Castlegar, BC, Canada. 15Alberta Environment and Parks, Alberta, AB, Canada. 16Braided Knowledge Consulting, Bozeman, MT, USA. 17Government of Alberta, Edmonton, AB, Canada. 18Wildlife Conservation Society, Newyork, NY, USA. 19Poisson Consulting, Nelson, Canada. 20Simon Fraser University, Burnaby, BC, Canada. 21Swan Valley Connections, Condon, MT, USA. 22Washington Department of Fish and Wildlife, Olympia, WA, USA. 23Wyoming Game & Fish Department, Cheyenne, WY, USA. 24Alaska Department of Fish and Game, Juneau, AK, USA. 25Biodiversity Pathways, Wildlife Science Center, Kelowna, BC, Canada. 26National Council for Air and Stream Improvement, Corvallis, OR, USA. 27Ministry of Forests, British Columbia, BC, Canada. 28Wildlife Genetics International, Nelson, BC, Canada. 29Center for Large Landscape Conservation, Bozeman, MT, USA. 30Wildlife Conservation Society, Toronto, ON, Canada. 31National Park Service, West Glacier, MT, USA. 32Ministry of Water, Land and Resource Stewardship, Victoria, BC, Canada. 33Casey C. Day, Erin L. Landguth contributed equally to this work. email: casey.day@mso.umt.edu; erin.landguth@mso.umt.edu Scientific Reports | (2024) 14:28248 | https://doi.org/10.1038/s41598-024-77956-9 1 www.nature.com/scientificreports/ North America, wolverines occur throughout Alaska and northwestern Canada, and as a small metapopulation occupying island-like, high-elevation montane regions of the northwestern United States. They naturally occur at low densities and were extirpated from their southern range in North America by about 19202. Wolverines have since recolonized portions of their historic range in the contiguous United States and Canada, inhabiting the Central Rocky Mountains in the mid-20th century3, whereas their more recent re-colonization into the Cascade Range in BC, Canada and Washington, USA is ongoing4. While wolverines are now widely distributed across suitable habitat in the coterminous northwestern United States5, densities are low and areas with suitable conditions are fragmented. Single individuals are also occasionally found in isolated mountain ranges in California, Utah, and Colorado6–8. In 2013, wolverines were proposed for listing as threatened by the US Fish and Wildlife Service under the Endangered Species Act due to habitat and range loss from climate warming, harvest, and small population sizes, but listing was found not warranted due to lack of information9. Subsequently, through a court-mandated reevaluation of this decision, wolverines in the contiguous US were listed as threatened10, while in 2014 wolverines were assessed as a species of special concern in Canada11. Despite their Holarctic distribution, wolverines may be experiencing population declines in some places, likely due to habitat loss in forested and montane ecosystems12. Furthermore, wolverines are a snow-adapted species, and models of species distribution, habitat selection, and landscape connectivity suggest that they may rely on cold microclimates and snowpack for niche space, predator avoidance, and denning, particularly at the southern extent of their range4,13–18. Regional habitat suitability analyses for wolverines have also identified terrain complexity, previous persecution, prey availability, and lack of human disturbance (built-up environments, human density, transportation and energy infrastructure, recreation, etc.) as potential predictors of wolverine habitat use19–26. A major management goal is to maintain connectivity among wolverine populations across large spatial extents to preserve genetic diversity and buffer against local extirpation10,17,27. An understanding of how landscape configuration and composition affects wolverine genetic connectivity can ultimately inform land-use planning efforts. Here, we used the largest wolverine genetic dataset ever assembled in North America, including samples from over 800 individuals collected across ~ 2.2 million km2 (Fig. 1), to develop models of genetic connectivity that identify landscape characteristics that promote or impede gene flow across the study area. We mapped the resulting models of landscape connectivity across a matrix of topographic, vegetative, climatic, and anthropogenic landscape variables (Table 1). We fit multiple models at variable scales and expanded on existing landscape genetics frameworks by employing validation techniques that included random and spatial crossvalidation as well as individual-based simulation modeling. Specifically, we tested support for our hypothesis that landscape features explain more variation in genetic (dis)similarity than the null model of geographic distance. For the most supported model, we produced a continuous genetic connectivity surface for wolverines as a function of landscape characteristics that promote or impede genetic connectivity in western Canada and the United States. Mapped connectivity surfaces in our study area help inform ongoing efforts to identify and prioritize land preservation (e.g., protected vs. non-protected areas) and mitigation (e.g., for highway crossings or forest rehabilitation), particularly in those areas that promote genetic connectivity for populations of wolverines. Results Multiple statistical approaches converged on the conclusion that wolverine genetic connectivity across western North America was positively associated with forest cover and negatively associated with human disturbance. Genetic connectivity was greater in the northern than the southern extent of wolverine range, as was expected based on available habitat, past genetic studies, recolonization history, and inclusion of the periphery of the species’ range28–30. Overall, our results suggest that large areas of forested, snowy habitat with low human disturbance have likely facilitated recent gene flow via wolverine dispersal across western North America (Fig. 2, Supplementary Data 3). Multiple modeling approaches and validation techniques were used to corroborate these results as follows, with full details found in Supplementary Information. Spatial genetics The relationships between genetic distance, geographic distance, and the logarithmic transformation of geographic distance are r = 0.290 and r = 0.321, respectively (Supplementary Fig. 1). A positive autocorrelation was observed up to 461 km with negative autocorrelation existing beyond 672 km (Supplementary Fig. 2). An interpolated value of 555 km as the threshold between positive and negative autocorrelation was chosen for subsequent analyses (i.e., Scale-dependent results). The spatial genetic diversity indices for estimates of allelic richness, inbreeding coefficients, and the effective number of breeding individuals within the 555 km spatial genetic neighborhoods are presented in Supplementary Fig. 3. Sampling density was disproportionately higher in the southern half of the study area, from southern British Columbia through northern Idaho and Montana (Supplementary Fig. 3A). We observed lower genetic diversity (allelic richness corrected by rarefaction; Supplementary Fig. 3B) and more inbreeding in the south than the north (Supplementary Fig. 3C). The northerncentral portion of the study area had the largest relative effective number of breeders, decreasing towards the south and west (Supplementary Fig. 3D). In southeastern British Columbia, researchers found that the wolverine population was divided into several small areas of higher density bisected by corridors or areas with low to zero density31. In Idaho and Montana, USA, the estimated effective populations size was previously reported to be 28–52 individuals13, which aligns with our estimates of 19–51 (Supplementary Fig. 3D purple). Model optimization and relative importance of landscape predictors In single-variable models of wolverine landscape genetic connectivity using maximum likelihood population effects analysis, the variables of human disturbance, forest cover, and snow days each outperformed the null Scientific Reports | (2024) 14:28248 | https://doi.org/10.1038/s41598-024-77956-9 2 www.nature.com/scientificreports/ Fig. 1. Study area. Extent of landscape connectivity analysis (green polygon) and locations of genetic samples (red points) for landscape genetics analysis of wolverines in North America. model of geographic distance (Table 2). The best-performing multi-variable model included the human disturbance PC (upper exponential transformation, 1000 km2 window) and forest cover (lower exponential transformation, 1 km2 window; Table 2). Study-wide connectivity maps are presented in Fig. 2 and show areas of predicted low to high genetic connectivity for wolverines in western North America. Consistency of top model Because factors affecting landscape connectivity may vary at different spatial scales32, by sex and/or by region, we tested the robustness of results from the full data set to analyze subsets of data for each of these factors. Scale effects. For finer-scale geographic distances ( < ~ 500 km), the same variables of forest cover (lower exp. transformation, 10 km2) and the human disturbance PC (lower exp. transformation, 1000 km2) were carried forward and outperformed all models when combined (Supplementary Table 1). For broader-scale geographic distances ( > ~ 500 km) the climate PCA produced the best-performing univariate model, but at the multivariate step the best-performing model included forest edge (lower exp. transformation, 10 km2) and human disturbance PC (lower exp. transformation, 1000 km2). Scientific Reports | (2024) 14:28248 | https://doi.org/10.1038/s41598-024-77956-9 3 www.nature.com/scientificreports/ Variable Category Source Climate PCA Hypo-thesis Description Snow Days Climate NDVI1 ● + Number of days with snow cover from NDVI NoData values (Fig. S5) SWE Climate CHELSA2 ● + 30-year average (1981–2010) of snow-water equivalent with permanent water bodies removed Temp Climate CHELSA2 ● - 30-year average (1981–2010) of annual maximum temperature Building density Dstrb Micro-soft3 ● - From all buildings in North America with point density tool in ARCGIS Human footprint 1 Dstrb ● - 8 anthropogenic variables: built-up environment, population density, electric power infrastructure, crop lands, pasture lands, roads, railways, navigable waterways Human footprint 2 Dstrb ● - 14 anthropogenic variables: urban, crop land, grazing, mining, energy production (oil and gas, renewable), roads, railways, power lines, electrical infrastructure, logging, human intrusion, reservoirs, air pollution HWY All Dstrb ● - Major Canadian and US Highways HWY 1 Dstrb - North or south of Trans-Canada Highway 1 Lights Dstrb NASA7 - Nighttime lights DEM Topo USGS8 + 30-m digital elevation model TPI Topo USGS8 + Topographic position index: From DEM using R terrain() in ‘raster’ TRI Topo USGS8 + Terrain ruggedness index: From DEM using R terrain() in ‘raster’ Forest cover For NALCMS9 + Proportion of cells in window of conifer or mixed forest type, 30-m For NALCMS9 + Proportion of 30-m cells in window where conifer/mixed forest was adjacent to non-forested habitat Forest edge NASA SEDAC4 Earth Systems Science Data5 Street Maps6 HF PCA Street Maps6 ● Table 1. Landscape variables. A summary of all variables by category rescaled to 1 km2 and used in the prediction of wolverine genetic connectivity. Because of correlation among variables in the same categories (r > 0.3), composite climate and human disturbance variables were also created using the indicated variables in a Principal Components Analysis (PCA). The Hypothesis column indicates whether an increase in a given variable is expected to result in an increase (+) or a decrease (-) in genetic connectivity. TPI = Topographic Position Index; TRI = Topographic Ruggedness Index; HF = Human footprint; Dstrb = Disturbance Category; Topo = Topography Category; For = Forest Category. 1https://modis.gsfc.nasa.gov. 2https://chelsa-climate. org. 3https://www.github.com/Microsoft/USBuildingFootprints. 4https://sedac.ciesin.columbia.edu/data/ set/wildareas-v3-2009-human-footprint. 5https://zenodo.org/record/3963013#.YAG7DuhKiUk. 6https:// openstreetmap.org. 7https://www.earthdata.nasa.gov/learn/backgrounders/nighttime-lights. 8https://www.usgs. gov/the-national-map-data-delivery/gis-data-download. 9http://www.cec.org/north-american-land-changemonitoring-system/. Sex Effects. Male connectivity returned the same top model as in the full wolverine genetic dataset (Supplementary Table 2) for forest cover and human disturbance. However, geographic distance narrowly outperformed all models for females, suggesting that distance is more influential than the landscape variables tested. Individual splits by sex did not reveal statistically significant differences in spatial autocorrelation with females showing a positive spatial autocorrelation up to 435 km, negative spatial autocorrelation beyond 672 km, and Mantel correlation tests for isolation-by-distance (r = 0.288 [0.276,0.302]). Males displayed a positive spatial autocorrelation up to 467 km, negative spatial autocorrelation beyond 695 km, and r = 0.287 (0.274,0.299). For male genetic connectivity model optimization, the same three categories as in the global model were carried forward to the multi-variable step that included snow (linear transformation, 1000 km2), forest cover (linear transformation, 1 km2), and human footprint (lower exp. transformation, 1,000 km3). For female genetic connectivity, the logarithmic transformation of geographic distance outperformed all univariate and multivariable models. We considered the second-best performing model of forest cover and human disturbance (Supplementary Table 2) to be competing with geographic distance (ΔAIC = 7 and higher R2 value). Northern and southern populations. Finally, for both the northern and southern regional tests, the models with the variable of forest cover alone outperformed all other models (Supplementary Table 3). The northern population had lower genetic structure than the southern population (Mantel correlation tests with geographic distance; r = 0.216 [0.206, 0.227] versus r = 0.272 [0.258, 0.285], respectively). We repeated the maximum likelihood population effects model selection procedure on the two different sets of observations, thus assessing the influence of regional variation in habitat and landscape features on model inferences. Only forest cover (linear transformation, 10 km2) outperformed geographic distance in the less genetic structured northern population (Supplementary Table 3). For the more structured southern population, variables from all four categories were carried forward to the second step, including forest cover (lower exp. transformation, 1 km2), topographic ruggedness index (TRI; lower exp. transformation, 100 km2), human disturbance (linear transformation, 1,000 km2), and snow days (lower exp. transformation, 10 km2). However, forest cover alone outperformed all multi-variable models (Supplementary Table 3). Scientific Reports | (2024) 14:28248 | https://doi.org/10.1038/s41598-024-77956-9 4 www.nature.com/scientificreports/ Fig. 2. Genetic connectivity. Prediction surface for top performing model showing low (black) to high (white) landscape resistance due to human disturbance and lack of forest cover (A) with a resistant kernel (B) and factorial least-cost paths (C) indicating low (dark red) to high (yellow) genetic connectivity areas (B) and leastcost path density (C). These surfaces are inclusive of areas outside of current wolverine distribution to provide an understanding of how much resistance these areas pose to potential wolverine dispersal and connectivity. Validating top genetic connectivity models All 10 folds of the cross-validation procedure selected the same top model with forest cover and human disturbance, including scale and transformation, as in the full wolverine genetic dataset. Repeating the maximum likelihood population effects model selection procedure for each independent landscape genetics simulation and comparing with the global ‘True’ model of forest cover and human disturbance PC resulted in a confusion matrix with an overall accuracy of 76.7%, sensitivity of 66.7%, and specificity of 81.0%. In summary, 6 of the 10 replicates selected the same top model as the global empirical model (human disturbance PC – upper exp. transformation, 1000 km2 and forest cover – lower exp. transformation, 1 km2), 3 of the 10 replicates selected global human footprint PC with snow days, and 1 of the 10 replicates selected the model with all three variables. No univariate model outperformed the multi-variable models for any of the 10 replicates. Assessing rigor of top genetic connectivity model with inter-model comparisons The primary difference between the least-cost transect analysis and maximum likelihood population effects results was the retention of snow days in the majority of least-cost transect analysis’s top performing feature selection models. The top least-cost transect analysis model included snow days, forest cover, and human disturbance in addition to geographic distance and reported a root mean squared error = 0.077, R2 = 0.197, and mean absolute error = 0.060 (Table 2). The least-cost transect analysis feature selection modeling on 21 iterations of the spatial-fold cross-validations are reported for each of the 5 feature selection models in with the gradient boosting machine algorithm reporting slightly lower RMSE than (in ascending order) our generalized linear Scientific Reports | (2024) 14:28248 | https://doi.org/10.1038/s41598-024-77956-9 5 www.nature.com/scientificreports/ Window (km2) Transformation marg. R2 cond. R2 Δ AICc Human disturbance PC1 1000 upper exp 0.174 0.338 3002 Snow days 1000 linear 0.157 0.281 3438 Climate PC1 1000 lower exp 0.155 0.282 3774 Human footprint 1 1000 upper exp 0.197 0.375 3983 Log of distance NA NA 0.126 0.224 4814 Forest cover 1 linear 0.171 0.294 4854 Human footprint 2 1000 lower exp 0.139 0.247 6288 SWE 1000 linear 0.135 0.242 7205 Distance NA NA 0.139 0.245 8060 Building density 1000 lower exp 0.138 0.244 8074 TRI 1000 upper exp 0.141 0.248 8158 NASA lights 1000 lower exp 0.138 0.244 8164 Elevation 1 upper exp 0.140 0.245 8240 TPI 1000 upper exp 0.138 0.245 8243 Forest edge 1 upper exp 0.138 0.244 8261 All highways NA NA 0.091 0.211 17,660 Highway 1 NA NA 0.057 0.158 28,962 Human disturbance PC1 + forest cover 0.171 0.32 0 Human disturbance PC1 + snow days + forest cover 0.157 0.29 1105 Human disturbance PC1 + snow days 0.169 0.31 1533 Snow days + forest cover 0.167 0.29 3014 Variables MLPE - Univariate MLPE – Multi-variable LCTA – Straight line transect (Step 1) Snow days + forest cover + geographic distance 0.194 Human disturbance PC1 + snow days + forest cover + geographic distance 0.197 LCTA – Least-cost path line (Step 2) Snow days + forest cover + geographic distance 0.1919 Human disturbance PC1 + snow days + forest cover + geographic distance 0.1920 Table 2. Model selection results for maximum likelihood population effects model regression and leastcost transect analysis of wolverine genetic distances and landscape resistance based on least-cost paths. For maximum likelihood population effects models, the univariate step evaluated only univariate models and selected variables for the multi-variable step based on whether they outperformed geographic distance alone. For each variable, four moving window sizes and three transformations were tested, and the best performing combination is presented. Variables for the multi-variable step were produced by taking a weighted average across the variable rasters and recalculating least-cost paths. Distance and log of distance represent the null model of isolation-by-distance. AIC is used for ranking the maximum likelihood population effects models across both univariate and multi-variable models, and not for evaluating relative performance among models. For least-cost transect analysis, a gradient boosting machine was used to select features important to describing genetic distance based on average values along least-cost paths. TPI = Topographic Position Index; TRI = Topographic Ruggedness Index; PC1 = principal component analysis (axis 1); LCTA = least-cost transect analysis; MLPE = maximum likelihood population effects. model, random forest model, generalized additive model, and linear model (Supplementary Fig. 4). The random forest model selected all variables 100% of the time during the training process and variable selection proportions for the other 4 models are shown in Supplementary Fig. 5. For the top performing algorithm (gradient boosting machine), two models were compared. The gradient boosting machine model with 3 variables (geographic distance, snow days, forest cover) reported a RMSE = 0.077, R2 = 0.194, and MAE = 0.061. The gradient boosting machine model with 4 variables (geographic distance, snow days, forest cover and human disturbance PC) reported a RMSE = 0.077, R2 = 0.197, and MAE = 0.060. Discussion Genetic connectivity of wolverines in our study area was positively associated with forest cover and negatively associated with human disturbance with snow days receiving mixed support depending on the analysis. These environmental factors eclipsed geographic distance, despite the vast area across the extent of North American wolverine populations. Other regional-scale studies on wolverine landscape genetics have offered concordant Scientific Reports | (2024) 14:28248 | https://doi.org/10.1038/s41598-024-77956-9 6 www.nature.com/scientificreports/ conclusions13,18, although our continental-scale research placed stronger emphasis on forest cover and provided insights for a much larger spatial extent, addressed multiple regions, and used newer modeling techniques. This analysis represents one of the largest assembled collaborative datasets in terms of both geographic extent and number of wolverine genetic samples in North America. Below, we discuss these findings in the context of each landscape variable, landscape genetics methodology, and implications for conservation planning. Human disturbance In our landscape genetics analysis, human disturbance was the strongest indicator of resistance to genetic connectivity (i.e., high resistance) for wolverines for all subsets of pairwise geographic distances using both modeling frameworks (maximum likelihood population effects models and least-cost transect analysis). This finding was also supported by genetic and habitat selection models at regional scales12. This result may suggest that protected areas with minimal recreation and development are important for promoting connectivity22. Of our variables tested describing human disturbance, those that consistently predicted genetic dissimilarity included composite variables with multiple factors such as road density, energy development, urban areas, and tourism. Non-composite variables describing more specific phenomena performed poorly, such as building density, nighttime lights (urbanization), and highways (movement barriers). Major highways correlated with population genetic structure in female wolverines at small spatial extents33 and induced avoidance behavior in terms of movement and habitat selection at the within-home range scale22,34,,35. It remains unclear how much of this result was due to effects of human disturbance on wolverine movements and the ecological changes that accompany development, including changes to prey, competitor, and predator communities. For example, humans tend to develop less rugged, less snowy areas in some regions, likely forcing a difference in habitat selection between humans and wolverine dispersers. However, while some cold and snowy areas of wolverine habitat are unoccupied by humans (e.g., high Rocky Mountains), others overlap with human recreation activities, are currently (e.g., arctic tundra), or were historically (e.g., Great Lakes region) occupied by both humans and wolverines. Furthermore, the landscape genetics analyses identified human disturbance as important when using the full data set, but not for either the north orsouth regions (Supplementary Table 3), suggesting that human disturbance may be more important to connectivity across larger distances. Overall, because of the differences from north to south, our results highlight benefits of conducting both focused regional and collaborative rangewide studies of wolverine’s genetic status and connectivity. Forest configuration (cover and edge habitat) Previous habitat and genetic studies have suggested that (non-Arctic) wolverines select areas with high forest cover36–39 or prefer forest edge habitat4,18,22,35,40. We found that increased forest cover was associated with greater wolverine genetic connectivity. The positive response of forest cover was evident in regional analyses, as models with forest cover alone outperformed all other models for both north and south regions (Supplementary Table 3). Similarly, forest cover was positively associated with male and female genetic distances and least-cost paths (Supplementary Table 2). Forest cover was also an important variable for predicting genetic connectivity through least-cost transect analysis. At the finer scale for the maximum likelihood population effects analysis ( < ~ 500 km), genetic connectivity was partially explained by forest cover, while genetic connectivity at the broader scale distances ( > ~ 500 km) resulted in the only set of observations explained at least partially by forest edge, which included edge due to factors such as alpine timberline and forest fragmentation. In some regions dispersing wolverines may be attracted to forest openings and edge, presumably for prey resources and navigation, especially during particular seasons35. At the scale of our research, our analyses suggest that coniferous and mixed forest positively facilitated connectivity among wolverine populations. We interpret these results as evidence that forest cover may have increased opportunities for dispersing individuals and movement that facilitates genetic connectivity. Snow persistence The extant wolverine fundamental niche has been generally characterized by cold, snowy areas14,16,42, though recent studies from North America41 and Scandinavia25 have questioned the importance of persistent snow cover for wolverine denning. Evidence that both snow persistence and snow depth promote wolverine genetic connectivity has been reported previously13,18 in studies that were limited to the southeastern periphery of the wolverine’s North American range where snow is more limited and ephemeral than at higher latitudes. Studies of habitat use within home ranges have reported similar patterns of selection for snow persistence21,24,27; however, habitat suitability models have often been unsuccessful at predicting genetic connectivity44,45. Our results provide some support for the hypothesis that snow persistence was important for wolverine genetic connectivity, even if this variable was not in our top maximum likelihood population effects model. Snow or snow-water equivalent was retained in all of the least-cost transect analyses modeling algorithms, and snow was retained more than forest cover when using the top-performing least-cost transect analysis algorithm (Supplementary Figs. 4, 5). This relationship was captured by the nonlinear machine learning models within the least-cost transect analysis algorithm but not the linear regression model of the maximum likelihood population effects method and may highlight a difference in capability between the two methods43. Snow days outperformed geographic distance in most maximum likelihood population effects univariate analyses (Supplementary Tables 1–3), emerged as the second-best model from the full dataset, and the final genetic connectivity prediction surfaces were highly correlated with the top model (Supplementary Fig. 6, Table 1). In addition, model results from the simulated wolverine populations (see Methods - Landscape resistance model validation tests) selected snow days in place of forest cover in 3/10 model runs, and snow days along with forest cover in 1/10 model runs, suggesting that fewer snow days was likely collinear with lack of forest cover, though resistance associated with fewer snow days Scientific Reports | (2024) 14:28248 | https://doi.org/10.1038/s41598-024-77956-9 7 www.nature.com/scientificreports/ was correlated with resistance due to human disturbance as well (Supplementary Table 4). Overall, our results showed that climate factors were not the only suite of variables influencing wolverine connectivity. Scale and shape of environmental variables Wildlife behavior and genetic processes respond to environmental features occurring at variable spatial scales32,46. Accounting for this scale-dependent variation increases our ability to accurately describe how animals select habitat at multiple scales (e.g., within and among home ranges, during dispersal), and how these decisions translate to gene flow across large distances47. Ultimately, genetic connectivity is the result of wolverines making fine-scale dispersal movement decisions that accumulate to produce larger dispersal distances, and over time these behaviors accumulate to produce large-scale population connectivity. Our multi-scale investigations revealed that wolverine connectivity was influenced simultaneously by both fine-scale vegetation factors and broad-scale influences of human impact and climatic factors. We performed four different analyses to understand how issues of the scale and shape of environmental variables affected our results. First, environmental resistance variables included in the maximum likelihood population effects analysis were sampled at four scales to identify the optimal scale at which to capture the resistance of the variable to gene flow. With this first approach, we found that most environmental resistance variables influenced genetic connectivity to a greater degree at the largest scale we measured (1,000 km2), although forest cover and terrain variables tended to perform better with the finer spatial scale (≤ 10 km2). Second, we tested multiple transformations (i.e., shape) to allow for nonlinear relationships between each variable and its resistance to gene flow, which improved model fit and accuracy. Third, because of the differences in genetic diversity and habitat characteristics from the northern to the southern extent of the study area, we conducted regional analyses on this north-south gradient to test for the spatial robustness of model results and account for non-stationary factors48. Finally, we conducted analyses on different subsets of geographic distance among wolverines, which also resulted in identification of different variables associated with genetic similarity. Despite the well-known effect of scale on habitat use, adoption of multi-scale analyses of landscape genetic connectivity has lagged18,47,49. In the accumulation of our analyses, our results demonstrate the benefit of evaluating multiple scales, particularly for landscape connectivity studies on wide-ranging species. For conservation and management, we interpret our results as demonstrating practical applications where we acknowledge the effects of scale on the variables that best explain genetic connectivity. Landscape genetics methodology, validation, and importance of inter-model comparison Our model validation and evaluation efforts50 served to strengthen the reliability of our methods and results, but have also produced insights about the potential relationships among variables such as forest cover and snow days. We used two different approaches as a means of inter-model comparison to test for model robustness and performance and improve confidence in our conclusions51. We have shown that when combined with machine learning approaches43,49,52, the least-cost transect analysis produced similar results to the pseudo-optimization routine for the maximum likelihood population effects modeling and provided additional information. For example, the best performing least-cost transect models retained snow days as positively associated with wolverine genetic connectivity, whereas the multi-variable maximum likelihood population effects models did not. It is possible that least-cost transect methods combined with machine learning are better equipped to account for interactive and non-linear effects43. As statistical models for landscape genetics continue to be refined and developed, it is critical that multiple methods are used to validate results and to glean insights based on differences among model outcomes. We also employed other means of validating and evaluating our results that strengthened the evidence in favor of our statistical approaches, such as cross-validation and simulation modeling. These measures, such as cross-validation and simulation modeling to test model performance, are needed in fields where methodology is rapidly developing and debate continues about the robustness of those methods to the diversity of applications in which they are applied53,54. Differences in results that emerge across models spur alternative hypotheses and guide future research that may elucidate reasons for the discrepancies. In the present case, for example, we expect there could be different interpretations with higher genetic resolutions, such as what could be gained from single nucleotide polymorphism panels instead of microsatellite markers55. Implications for transboundary conservation and management of wolverines In southwestern Canada and the northwestern contiguous United States, wolverines have large home-ranges (500–1,000 km2) that cross national, tribal, provincial, and state boundaries; consequently, effective conservation efforts will require cooperation among many levels of government. We provided the first landscape genetics analysis spanning the transition zone from the Canadian and Alaskan populations in the north to the southern populations in the coterminous United States. Our study enables consideration of suitable management actions at large spatial scales and can serve as a baseline against which to compare future genetic patterns as landscapes change. Our results indicated high levels of landscape genetic connectivity across the wolverine’s range in much of western North America, with a clear signal of high-to-low genetic diversity and connectivity from north-to-south, corroborating past literature28,56. The covariates that parsimoniously explained wolverine genetic connectivity reflect that gene flow occurred across large areas of southwestern Canada, with populations becoming more fragmented towards the southern extent of the range. This pattern may be partially due to the fact that wolverines were extirpated from and have recolonized multiple southern portions of the study area since the mid-20th century30. Species conservation for wolverines hinges largely on large-scale landscape management with connectivity increasing with forest cover, less human disturbance, and persistent snow cover57. Based on our results, minimizing anthropogenic landscape development or forest removal in regions between areas with predicted wolverine occurrence would promote gene flow. Understanding how our predicted least-cost corridors intersect with existing and planned land-use change and transportation infrastructure is the first step to provide scienceScientific Reports | (2024) 14:28248 | https://doi.org/10.1038/s41598-024-77956-9 8 www.nature.com/scientificreports/ based management actions and conservation planning within the wolverine’s range in North America. Additional work will be needed to understand how localized projects that affect forest cover and human disturbance could impact potential dispersal pathways of wolverines, particularly in areas where genetic connectivity is more fragmented. Regardless, the large spatial extent of wolverine habitat combined with long-distance dispersal behavior requires that cumulative effects of such projects need to be considered at multiple spatial scales. Future analyses should focus on utilizing the results of this genetic connectivity study to identify existing and potential key habitat linkages between wolverine core habitats. Our study builds on a limited but growing foundation regarding wolverine genetics. We achieved relevant and timely assessments from the collaboration of researchers, landowners, and managers within two countries, resulting in the first wholistic description of genetic relatedness and connectivity for North America’s wolverines. Methods Study system The 2,173,501 km2 study area encompassed most of the western Canadian provinces of British Columbia and Alberta, southeastern Alaska, and the currently occupied wolverine habitat in the contiguous United States in Montana, Idaho, Wyoming, and Washington (Fig. 1). This transboundary study area contains diverse wolverine habitat from montane ecosystems in the south to boreal forests in the north. DNA sample collection and analysis We collected 882 multi-locus genotypes (489 males, 393 females) using 19 loci for the landscape genetic analysis (data available upon request). These wolverine genetic data were extracted from hair snagging, live-capture, and trapper harvest from a group of > 40 wildlife researchers and managers between 2006 and 2016 across Alberta, British Columbia, Idaho, Montana, Washington, Wyoming, and Southeast Alaska (Fig. 1). All DNA extractions (using QIAGen58 DNA extraction kits) and genetic analyses were performed at Wildlife Genetics International (WGI) in Nelson, BC, and the National Genomics Center for Wildlife and Fish Conservation at the Rocky Mountain Research Station (RMRS) in Missoula, MT. The aggregated dataset contained 882 complete and unique genotypes (489 males, 393 females) for the landscape genetic analysis. Summary statistics for the 19-microsatellites from GenAlex59 included mean number of alleles = 6.3 with SE = 0.42, total alleles, Na = 118, and Fixation Index = 1 - observed heterozygosity / expected heterozygosity = 0.078 with SE = 0.008 (Supplementary Table 5), as well as tests for neutrality (Supplementary Fig. 9). Individual-based genetic differentiation To estimate genetic distance among individuals, we used the software Spagedi60 and the R package ‘gstudio’61 to calculate several metrics that were recommended in a review of genetic distance measures for landscape genetics62. These metrics included kinship coefficients63,64, relationship coefficients65–68, Rousset’s A69, proportion of shared alleles70, and Euclidean distance71. All metrics were highly correlated, r > 0.92, which can result in non-robust resistance surface optimization72. To simplify the various metrics into a single distance measure, we conducted principal components analysis across the pairwise matrices of genetic distance and used the first principal component as our measure of genetic distance for all landscape genetics analyses. We provide genetic distances for all subsets in Supplementary Data 1. Spatial genetic structure Mantel tests for Pearson’s correlation and spatial correlograms (‘ecodist’ package in R73) were used to assess the extent of spatial genetic structure and scale effects in the wolverine data set74. We used the ‘sGD’ package in R75 to estimate genetic diversity based on grouping individuals into overlapping genetic neighborhoods. We inferred a genetic neighborhood based on the correlograms that depict the spatial autocorrelation of genotypes76. This approach has the advantages of being able to capture spatially complex patterns of genetic diversity in clinal or landscape-driven populations while not having to dictate population boundaries a priori. Within each neighborhood, we calculated estimates corrected for sample size of allelic diversity77, inbreeding coefficient78 (or Fixation Index), and effective number of breeders79 (Wright’s neighborhood size) using NeEstimator80 within ‘sGD’, as well as sample size within the genetic neighborhood81. Landscape variables and curation We identified hypotheses for how landscape variables likely impede or promote genetic connectivity for wolverines in North America (Table 1). Selection of landscape variables was based on existing literature and an expert opinion working group of collaborators. We assigned the selected variables to one of four categories for how genetic connectivity will be affected, that is, hypothesized to be negatively affected by human disturbance, climate, topography, and vegetation. We curated a set of spatial datasets to represent each variable in terms of resistance to wolverine movement by using three transformations and four spatial scales of each resistance surface18 (described below; Table 1) to represent alternative sensitivity to landscape resistance and spatial scale. All variables can be obtained publicly (but see snow days Table 1 and Supplementary Fig. 7). To ensure consistency of variables across the transboundary study extent, raster datasets were only obtained from sources that covered both the United States and Canada. To minimize distortion of spatial data projected across such a large study extent, we reprojected all raster data to a custom azimuthal equidistant projection centered within the study area. Landscape resistance hypotheses Landscape resistance surfaces are characterized by geospatial rasterized cells in a digital layer where each cell is given a relative value reflecting the impediments or barriers to gene flow. We rescaled all variables to Scientific Reports | (2024) 14:28248 | https://doi.org/10.1038/s41598-024-77956-9 9 www.nature.com/scientificreports/ standardize environmental predictors (Table 1)18. First, we resampled all raster grids to 1 km2 resolution because this resolution was broad enough to make the data analysis computationally feasible, and is an appropriate minimum scale given the daily movements of wolverines, this resolution was unlikely to influence biological or computational outcomes. Then, we rescaled raster values between 0 and 1, representing the minimum and maximum resistance present within the study area (raster cells with highways were assigned a value of 1, with all other values 0). We used three transformations on these rescaled values43 to represent variable sensitivity to resistance following a weighted distance approach82: linear, upper exponential, and lower exponential (e.g., upper exponential transformations represent high sensitivity to changes at low values of resistance and vice versa18). Finally, because the scale at which animals respond to stimuli and the scale of dispersal habitat needed to survive a dispersal event can vary across variables, we used a moving window analysis to capture each variable at 4 different scales with windows of 1-, 10-, 100-, and 1,000 km2. Once the complete set of resistance surfaces was created, we used the costDistance() function in the ‘gdistance’ package in R83 to calculate effective distances between all pairs of individual wolverine observations. The resulting resistance or effective distance matrices were used as explanatory variables in the landscape genetics analyses (lower triangle of the full dataset; 882 × 882 matrix or 882 * 881 / 2 = 388,521 pairwise observations; Supplementary Data 2). No transformations or moving windows were run on the highway rasters, as these represented linear barriers. Linear mixed models for spatial genetics Statistical modeling techniques have emerged over the last few decades to test for individual (or population) measures of genetic connectivity relationships, and many criticisms have followed from subjectivity of resistance surfaces to methodology issues associated with spatial autocorrelation84,85. We used maximum-likelihood population-effects, a mixed modeling regression approach that treats the residual for each pairwise distance as the sum of two random population-level effects and an observation-level error86. Because the maximum likelihood population effects correlation structure allows modelling the non-independence of pairwise distances within a likelihood framework, compatible with model selection, it is particularly appealing for landscape genetic studies87. The fixed effects are the pairwise effective distances from each landscape resistance hypothesis, and the random effect utilizes a covariance structure to account for the correlation between pairwise data points. Maximum likelihood population effects modeling has outperformed multiple regression with distance matrices and causal modeling with Mantel tests53 while accounting for autocorrelation structure inherent in all distancebased analyses88. Prior to analysis, we evaluated all pairs of variables for multicollinearity (Supplementary Table 4) and excluded one variable out of any pair with a Pearsons’ correlation value > 0.789. Snow days and maximum annual temperature were the only two variables that exceeded this correlation value, and therefore we excluded maximum temperature. We used a two-step hierarchical model selection procedure to identify the best model for predicting genetic distances among pairs of wolverines (i.e., pseudo-optimization89). While optimization methods for maximum likelihood population effects models exist (e.g., ‘ResistanceGA’90), these are computationally demanding and were unable to accommodate the combination of landscape size and number of observations in our data set, as well as recent evidence pointing to unreliability of results72. In the first step of model selection, we ran univariate maximum likelihood population effects models for each combination of variable, transformation, and scale (e.g., snow days, linear transformation, 10 km2). The best scale-transformation combination for each category of variables (climate, human disturbance, topography, forest) that outperformed log-transformed geographic distance was carried forward to the second step. Distance-only models represent the null model of isolation-by-distance because they assume that landscape patterns have no effect on genetic connectivity. In the second step of model selection, we created a candidate set of multi-variable models by combining variables that were carried forward from the first step92. Rasters for each variable in a candidate model were combined into a single raster using weighted averaging, where the weight a variable received was equal to 1/n, with n equal to the rank of that variable from the univariate step. We averaged resistance across explanatory variables instead of using multiple regression on multiple distance matrices because least-cost paths associated with different environmental variables can take widely different routes across a landscape92. We tested all possible 2- and 3-variable combinations and included all 3 transformations for each variable to allow for the possibility of nonlinear associations among variables to improve performance under different transformations. We repeated the method used in step 1 to create effective distance matrices and conducted AIC model selection on the full set of univariate and multi-variable models. Model performance was judged primarily on Akaike’s information criterion (AIC), which is suitable for maximum likelihood population effects models that don’t use restricted maximum likelihood62,88,91. Landscape resistance model sensitivity tests Scale effects. Because factors affecting landscape connectivity may vary at different spatial scales32, we tested for scale effects in the full wolverine data set using the spatial autocorrelograms genetic neighborhood cutoff value18 (Supplementary Fig. 2; ~555 km). Based on these results, we partitioned the data set into two groups: one including only pairwise comparisons less than or equal to a geographic distance of 555 km (“fine scale”) and the other including only those pairs of individuals with geographic distances greater than 555 km (“broad scale”). We then repeated the maximum likelihood population effects procedure at the two geographic scales, thus assessing the influence of scale effects on model inferences. Sex effects. We split the wolverine genetic distance data set by male and female. We then repeated the maximum likelihood population effects procedure for the two data sets to assess the effect of sex on model results. Northern and southern populations. Wolverines in North America have greater genetic structure in the southern periphery of their range than in regions further north29. Two landscape genetics studies in the southeastern population have shown support for spring snow cover, snow depth, buildings, and terrain ruggedness13,18. Therefore, to identify regional differences in connectivity, as well as the ability to model features that may only Scientific Reports | (2024) 14:28248 | https://doi.org/10.1038/s41598-024-77956-9 10 www.nature.com/scientificreports/ serve as limiting factors in one region48 (e.g., snow days in the south), we conducted maximum likelihood population effects analyses on north and south data sets (“northern” n = 403 vs. “southern” n = 479 individuals). Observations were grouped as either north or south of Highway 1 in Canada, which we hypothesized was a potential barrier to female wolverine connectivity33. Landscape resistance model validation tests Random cross-validation of predicted genetic distance. To test for consistency in maximum likelihood population effects model selection results, we conducted 10-fold cross-validation by randomly withholding 20% of the individuals and training the model on the remaining 80%. We used the same candidate set of models that was used on the full data set, and recorded the number of times that maximum likelihood population effects analyses selected the same model as the full data set. Using the selected model from each training data set, we predicted genetic distances for the remaining 20% of observations and recorded the root mean squared error (RMSE). Individual-based genetic simulations. Landscape genetics models and methods have been subject to uncertaintines in their ability to have enough power to identify the correct processes driving gene flow95,99. To formally evalute our statistical approach and further corroborate results, we used simulations to stipulate the actual species-and individual-specific processes in action and to test the reliability with which MLPE was capable of identifying the correct resistance models i (pattern-process or pattern-oriented modeling93). We used the individual-based landscape genetics simulator (CDPOP v1.3.1594) to (1) test the ability of the resistance surface to reproduce empirical genetic patterns, and (2) test our ability to correctly attribute the causes of observed genetic structure with the statistical methods used here95. This approach provides a controlled simulated environment and helps support empirically derived results of observed genetic patterns. We therefore generated genetic data across individuals in a spatially distributed pseudo wolverine population over time as a function of the landscape resistance hypothesis produced from the empirical landscape genetics analyses that used the full data set described above. Using the empirical derived landscape resistance surface, we placed 1,000 spatially weighted points (i.e., low resistance cell values received a higher probability). We then calculated the pairwise effective distances between the 1,000 points with the landscape resistance surface using the costDistance() function in the ‘gdistance’ package in R83. These effective distance values were used to specify dispersal and mating probabilities based on an inverse-square distribution with 30% maximum movement of the total landscape’s effective distance (see Supplementary Fig. 8). This movement distribution was chosen through initial exploratory simulations varying both functional form and maximum movement, with the goal of calibrating the model to achieve realistic patterns of dispersal behavior and genetic diversity over time. Each simulation tracked 19 loci with an initial starting value of 30 alleles per locus randomly assigned to the first generation (the same number of total loci as our empirical data set) in a female without replacement and male with replacement mating structure for 300 non-overlapping generations. Each mate pair produced a number of offspring following a Poisson process such that population growth was kept constant. After ~ 25 generations, simulations with these parameters converged on a fixation index of 0.08 (the approximate population genetics structure of the empirical data set). We produced 10 replicates and within each replicate chose a generation time to analyze such that the maximum starting alleles of 570 decayed to 118 alleles (total alleles within empirical data set). The replicate generation times that matched both F = 0.08 and Na = 118 varied and occurred between 260 and 290 generations. Finally, each replicate with 1,000 individual genotypes was used to repeat the maximum likelihood population effects process described above and a confusion matrix (ConfusionMatrix() function in the ‘caret’ package in R96) was used to determine how many times the underlying landscape resistance surface was recovered (i.e., total accuracy, specificity, sensitivity). Landscape resistance inter-model comparisons Model comparison with least-cost transect analysis. Least-cost transect analysis or corridor-based approaches91 extract summarized values of covariates along straight-line or least-cost path lines that are then used to predict genetic distances. Optimization procedures97 with a maximum likelihood framework were introduced with more recent studies adding random forest models52,98. Applications of least-cost transect analyses were expanded to individual-based samples using gradient boosted regression models with spatial cross validation49. Machine learning approaches are beginning to show promising results for landscape genetics and may possibly improve predictive power99. However, inter-model comparisons ultimately provide improved reliability in prediction results, if many models converge to similar results. Here, we applied the individual-based, machine learning, least-cost transect analysis methodology49. First, we extracted the mean value along a 3 km buffered straight line (results were not sensitive to the size of buffer) between all pairwise locations for all environmental variables using the exactextract() function in the ‘exactextractr’ package in R100. The covariates extracted from these straight lines along with geographic distance (to minimize spurious results95) were then used as explanatory variables with genetic distance as the response variable. Prior to modeling, we used k-means clustering to assign observations to distinct spatial groups for use as cross-validation folds. We created groups of both k = 10 and k = 20 clusters, to evaluate the effect of different spatial structures on model selection results. Spatial clusters were assigned such that groups of individuals and all connections to other individuals were withheld at each fold (approximately 22% of the data). We then applied a forward feature selection algorithm in the ‘CAST’ package in R96 using 5 different model types or algorithms: gradient boosting machine101, random forests102 (‘ranger’ package in R103), generalized linear elastic-neural network model104, generalized additive model, and a linear model. For each model type, we performed 21 feature selection replicates using the groups of k = 10 or a unique combination of k = 20 clusters. The optimal predictor set for each forward feature selection replicate was retained and then summarized across model types. GBM errors were lowest among the 5 model types and we continued using only GBM. Using the model trained using straight line variables, we created a prediction surface using the ‘raster’ package in R105. We then used this Scientific Reports | (2024) 14:28248 | https://doi.org/10.1038/s41598-024-77956-9 11 www.nature.com/scientificreports/ straight line generated genetic connectivity prediction surface to create least-cost path lines using the paths() function in the ‘gdistance’ package in R. Landscape variables were then extracted along each path and used as predictors in a second stage of model training to determine the best final least-cost transect models. Study-wide connectivity We used the software UNICOR106 to map predicted genetic connectivity for the final resistance genetic connectivity surface. We used two connectivity algorithms to highlight probable dispersal corridors: (1) factorial least-cost paths (Dijkstra’s algorithm107) and (2) resistant kernels108. Given an input resistance surface, factorial least-cost paths calculate all potential routes for gene flow between pairs of source points and highlight optimal dispersal paths, while resistant kernels depict corridors or movement areas that are smoothed as a function of cumulative cost and corridor intensity109. For both factorial least-cost paths and resistant kernels, we used 1,000 randomly spatially weighted source points (higher probability of spatial location in lower resistance values) within the study area. For factorial least-cost paths, we did not specify a dispersal threshold in effective-distance units. For resistant kernels, we used the volume calculations (‘Kernel_volume’ = 5,000 and ‘Const_kernal_vol’ = FALSE) with ‘Transform_function’ = linear across an ‘Edge_Distance’ = 5,000. Data availability The datasets generated to replicate the landscape genetic analyses are available in the Supplementary Information as Supplementary Data 1-2. Raw genotypes with locations are available by request to author MAS at sawaya. mike@gmail.com with collaborator approval. Supplementary Data 3 provides files for visualization of the connectivity landscape. Code availability All code to replicate these analyses provided upon request. Received: 31 July 2024; Accepted: 28 October 2024 References 1. Pasitschniak-Arts, M. & Larivière, S. Gulo gulo. 1995. Mammalian species. 499:1 – 0. 2. Aubry, K. B., McKelvey, K. S. & Copeland, J. P. Distribution and broadscale habitat relations of the wolverine in the contiguous United States. J. Wildl. Manag. 71, 2147–2158 (2007). 3. Newby, F. E. & McDougal, J. J. Range extension of the wolverine in Montana. J. Mammal. 45, 485–488 (1964). 4. Aubry, K. B., Raley, C. M., Shirk, A. J., McKelvey, K. S. & Copeland, J. P. Climatic conditions limit wolverine distribution in the Cascade Range of southwestern North America. Can. J. Zool. 101, 95–113 (2023). 5. Lukacs, P. M. et al. Wolverine occupancy, spatial distribution, and monitoring design. J. Wildl. Manage. 84, 841–851 (2020). 6. Schwartz, M. K. et al. Inferring geographic isolation of wolverines in California using historical DNA. J. Wildl. Manag. 71, 2170– 2179 (2007). 7. Moriarty, K. M. et al. Wolverine confirmation in California after nearly a century: native or long-distance immigrant? Northwest. Sci. 83, 154–162 (2009). 8. Packila, M. L., Riley, M. D., Spence, R. S. & Inman, R. M. Long-distance wolverine dispersal from Wyoming to historic range in Colorado. Northwest. Sci. 91, 399–407 (2017). 9. USFWS. Endangered and threatened wildlife and plants; threatened status for distinct population segment of the North American wolverine occurring in the contiguous United States. Fed. Reg. 78, 7864–7890 (2013). 10. USFWS. Recovery outline for the contiguous United States distinct population segment of the North American wolverine (Gulo gulo luscus). 12 pp. Portland, OR, USA. (2023). 11. COSEWIC. COSEWIC assessment and status report on the wolverine Gulo gulo in Canada. Committee on the status of endangered wildlife species in Canada. Ottawa, 76 pp. (2014). 12. Fisher, J. T. et al. Wolverines (Gulo gulo) in a changing landscape and warming climate: A decadal synthesis of global conservation ecology research. Global Ecol. Conserv. 34, e02019 (2022). 13. Schwartz, M. K., Copeland, J. P., Anderson, N. J., Squires, J. R. & Inman, R. M. Wolverine gene flow across a narrow climatic niche. Ecology. 90, 3222–3232 (2009). 14. Copeland, J. P. et al. The bioclimatic envelope of the wolverine (Gulo gulo): do climatic constraints limit its geographic distribution? Can. J. Zool. 88, 233–246 (2010). 15. McKelvey, K. S. et al. Climate change predicted to shift wolverine distributions, connectivity, and dispersal corridors. Ecol. Appl. 21, 2882–2897 (2011). 16. Inman, R. M., Magoun, A. J., Persson, J. & J. Mattisson The wolverine’s niche: linking reproductive chronology, caching, competition, and climate. J. Mammal. 93, 634–644 (2012a). 17. Inman, R. M. et al. Developing priorities for metapopulation conservation at the landscape scale: Wolverines in the Western United States. Biol. Conserv. 166, 276–286 (2013). 18. Balkenhol, N. et al. Landscape genetics of wolverines (Gulo gulo): scale-dependent effects of bioclimatic, topographic, and anthropogenic variables. J. Mammal. 101, 790–803 (2020). 19. May, R., Landa, A., van Dijk, J., Linnell, J. D. C. & Andersen, R. Impact of infrastructure on habitat selection of wolverines Gulo gulo. Wildl. Biology. 12, 285–295 (2006). 20. Webb, S. M. et al. Distribution of female wolverines relative to snow cover, Alberta, Canada. J. Wildl. Manag. 80, 1461–1470 (2016). 21. Heim, N., Fisher, J. T., Clevenger, A., Paczkowski, J. & Volpe, J. Cumulative effects of climate and landscape change drive spatial distribution of Rocky Mountain wolverine (Gulo gulo L). Ecol. Evol. 7, 8903–8914 (2017). 22. Heinemeyer, K., Squires, J., Hebblewhite, M., O’Keefe, J. J. & Holbrook, J. D. J. Copeland. Wolverines in winter: indirect habitat loss and functional responses to backcountry recreation. Ecosphere 10:e02611. (2019). 23. Kortello, A., Hausleitner, D. & Mowat, G. Mechanisms influencing the winter distribution of wolverine Gulo gulo luscus in the southern Columbia Mountains, Canada. Wildl. Biology. 2019, 1–13 (2019). 24. Barrueto, M., Forshner, A., Whittington, J., Clevenger, A. P. & Musiani, M. Protection status, human disturbance, snow cover and trapping drive density of a declining wolverine population in the Canadian Rocky Mountains. Sci. Rep. 12, 17412. ​h​t​t​p​s​:​/​/​d​o​i​.o ​ ​r​g​ /​1​0​.​1​0​38​ ​/​s​4​1​5​9​8​-​02​ ​2​-​2​1​4​9​9​-​4​​​​ (2022). Scientific Reports | (2024) 14:28248 | https://doi.org/10.1038/s41598-024-77956-9 12 www.nature.com/scientificreports/ 25. Persson, J., Ordiz, A., Ladle, A., Andrén, H. & Aronsson, M. Recolonization following past persecution questions the importance of persistent snow cover as a range limiting factorfor wolverines. Glob. Change Biol. 29, 5802–5815. ​h​t​t​p​s​:​/​/​do ​ ​i​.​o​r​g​/​1​0​.1​ ​1​1​1​/​g​c​b​.​1​ 6​9​0​8​​​​ (2023). 26. Moqanaki, E., Milleret, C., Dupont, P., Brøseth, H. & Bischof, R. Wolverine density distribution reflects past persecution and current management in Scandinavia. Ecography 2023:e06689. (2023). 27. Carroll, K. A., Inman, R. M., Hansen, A. J., Lawrence, R. L. & Barnett, K. A framework for collaborative wolverine connectivity conservation. iScience. 24, 100882 (2021). 28. Kyle, C. J. & Strobeck, C. Genetic structure of North American wolverine (Gulo gulo) populations. Mol. Ecol. 10 (2), 337–347 (2001). 29. Kyle, C. J. & Strobeck, C. Connectivity of peripheral and core populations of North American wolverines. J. Mammal. 83 (4), 1141–1150 (2002). 30. McKelvey, K. S. et al. Recovery of wolverines in the Western United States: Recent Extirpation and Recolonization or Range Retraction and Expansion? J. Wildl. Manage. 78, 325–334 (2014). 31. Mowat, G. et al. The Sustainability of Wolverine Trapping Mortality in Southern Canada. Jour Wild Mgmt. 84, 213–226. ​h​t​t​p​s​:​/​/d ​ ​ o​i​.​o​r​g​/​1​0.​ ​1​0​0​2​/​j​w​m​g.​ ​2​1​7​8​7​​​​ (2020). 32. Galpern, P. & Manseau, M. Finding the functional grain: comparing methods for scaling resistance surfaces. Landscape Ecol. 28, 1269–1281 (2013). 33. Sawaya, M. A., Clevenger, A. P. & Schwartz, M. K. Demographic fragmentation of a protected wolverine population bisected by a major transportation corridor. Biol. Conserv. 236, 616–625 (2019). 34. Copeland, J. P. et al. Seasonal habitat associations of the wolverine in central Idaho. J. Wildl. Manag. 71, 2201–2212 (2007). 35. Scrafford, M. A., Avgar, T., Abercrombie, B., Tigner, J. & Boyce, M. S. Wolverine habitat selection in response to anthropogenic disturbance in the western Canadian boreal forest. For. Ecol. Manag. 395, 27–36 (2017). 36. Anni, K. K. et al. Does grey wolf presence affect habitat selection of wolverines? Annals Zool. Fennici. 50, 216–224 (2013). 37. Rauset, G. R., Mattisson, J., Andren, H., Chapron, G. & Persson, J. When species’ ranges meet: assessing differences in habitat selection between sympatric large carnivores. Oecologia. 172, 701–711 (2013). 38. Krebs, J., Lofroth, E. C. & Parfitt, I. A. Multiscale habitat use by wolverines in British Columbia, Canada. J. Wildl. Manag. 71, 2180–2192 (2007). 39. Webb, S. M. et al. Incorporating local ecological knowledge to explore wolverine distribution in Alberta, Canada. Wildl. Soc. Bull. 43, 414–424 (2019). 40. Inman, R. M. et al. J. A. Fredrick. 2012b. Spatial ecology of wolverines at the southern periphery of distribution. J. Wildl. Manag. 76:778–792 . 41. Jokinen, M. E., Webb, S. M., Manzer, D. L. & Anderson, R. B. Characteristics of Wolverine (Gulo gulo) dens in the lowland boreal forest of north-central Alberta. Can. Field-Naturalist. 133, 1–15 (2019). 42. Magoun, A. J., Robards, M. D., Packila, M. L. & Glass, T. W. Detecting snow at the den-site scale in wolverine denning habitat. J. Wildl. Manage. 41, 381–387 (2017). 43. Vanhove, M. & Launey, S. Estimating resistance surfaces using gradient forest and allelic frequencies (Molecular Ecology Resources in, 2024). https://doi.org/10.1111/1755-0998.13778 44. Mateo-Sánchez, M. et al. A comparative framework to infer landscape effects on population genetic structure: are habitat suitability models effective in explaining gene flow? Landscape Ecol. 8, 1405–1420 (2015). 45. Roffler, G. H. et al. Identification of landscape features influencing gene flow: How useful are habitat selection models? Evol. Appl. 9, 805–817 (2016). 46. Mayor, S. J., Schneider, D. C., Schaefer, J. A. & Mahoney, S. P. Habitat selection at multiple scales. Ecoscience. 16, 238–247 (2009). 47. McGarigal, K., Wan, H. Y., Zeller, K. A. & Timm, B. C. Multi-scale habitat selection modeling: a review and outlook. Landscape Ecol. 31, 1161–1175 (2016). 48. Short Bull, R. A. et al. Why replication is important in landscape genetics: American black bear in the Rocky Mountains. Mol. Ecol. 20, 1092–1107 (2010). 49. Palm, E. C. et al. Corridor-based approach with spatial cross-validation reveals scale-dependent effects of geographic distance, human footprint and canopy cover on grizzly bear genetic connectivity. Mol. Ecol. 32, 5211–5227. ​ht​ ​t​p​s​:​/​/​d​o​i.​ ​o​r​g​/​1​0​.​1​11​ ​1​/​m​e​c​.​1​ 7​0​9​8​​​​ (2023). 50. Augusiak, J., Van den Brink, P. J. & Grimm, V. Merging validation and evaluation of ecological models to ‘evaludation’: A review of terminology and a practical approach. Ecol. Model. 280, 117–128 (2014). 51. Grimm, V. Ecology needs to overcome siloed modeling. Trends Ecol. Evol. 38, 1122–1124 (2023). 52. Pless, E., Saarman, N. P., Powell, J. R., Caccone, A. & Amatulli, G. A machine-learning approach to map landscape connectivity in Aedes aegypti with genetic and environmental data. Proceedings of the National Academy of Science 118:e2003201118. (2021). 53. Shirk, A. J., Landguth, E. L. & Cushman, S. A. A comparison of regression methods for model selection in individual-based landscape genetic analysis. Mol. Ecol. Resour. 18, 55–67 (2018). 54. Beninde, J., Wittische, J. & Frantz, A. C. Quantifying uncertainty in inferences of landscape genetic resistance due to choice of individual-based genetic distance metric. Mol. Ecol. Resour. 24, e13831 (2023). 55. Lansink, G. et al. Potential for increased connectivity between differentiated wolverine populations. Biol. Conserv. 272, 109601 (2022). 56. Cegelski, C., Waits, L. P., Flagstad, O. & Kyle, C. J. Genetic diversity and population structure of wolverine (Gulo gulo) populations at the southern edge of their current distribution in North America with implications for genetic viability. Conserv. Genet. 7 (2), 197–211 (2006). 57. Charnley, S., Davis, E. J. & Schelhas, J. The Bipartisan Infrastructure Law and the forest service: insights for local job creation and equity from the American Recovery and Reinvestment Act. J. Forest. 121, 282–291 (2023). 58. Birnboim, H. C. A rapid alkaline extraction method for the isolation of plasmid DNA. Methods Enzymol. 100, 243 (1983). 59. Peakall, R. & Smouse, P. E. GENALEX 6: genetic analysis in Excel. Population software for teaching and research. Molecular Ecology Notes 6: 288–295. (2006). 60. Hardy, O. J. & Vekemans, X. SPAGeDi: a versatile computer program to analyse spatial genetic structure at the individual or population levels. Mol. Ecol. 2, 618–620 (2002). 61. Dyer, R. J. The gstudio package (Virginia Commonwealth University, 2012). 62. Shirk, A. J., Landguth, E. L. & Cushman, S. A. A comparison of individual-based genetic distance metrics for landscape genetics. Mol. Ecol. Resour. 17, 1308–1317 (2017). 63. Ritland, K. Estimators for pairwise relatedness and individual inbreeding coefficients. Genet. Res. 67, 175–185 (1996). 64. Loiselle, B. A. et al. Avoiding pitfalls of using species distribution models in conservation planning. Conserv. Biol. 17, 1591–1600 (2003). 65. Queller, D. C. & Goodnight, K. F. Estimating relatedness using genetic markers. Evolution. 43, 258–275 (1989). 66. Li, C. C., Weeks, D. E. & Chakravarti, A. Similarity of DNA fingerprints due to chance and relatedness. Hum. Hered. 43, 45–52 (1993). 67. Lynch, M. & Ritland, K. Estimation of pairwise relatedness with molecular markers. Genetics. 152, 1753–1766 (1999). 68. Wang, J. An estimator for pairwise relatedness using molecular markers. Genetics. 160, 1203–1215 (2002). 69. Rousset, F. Genetic differentiation between individuals. J. Evol. Biol. 13, 58–62 (2000). Scientific Reports | (2024) 14:28248 | https://doi.org/10.1038/s41598-024-77956-9 13 www.nature.com/scientificreports/ 70. Bowcock, A. M. et al. High resolution of human evolutionary trees with polymorphic microsatellites. Nature. 368, 455–457 (1994). 71. Wright, S. Isolation by distance. Heredity. 28, 114–138 (1943). 72. Schleimer, A. et al. Robustness of resistance surface optimisations: sampling schemes and genetic distance metrics affect inferences in landscape genetics. Landsc. Ecol. 38, 2861–2883. https://doi.org/10.1007/s10980-023-01752-5 (2023). 73. Goslee, S. C. & Urban, D. L. The ecodist package for dissimilarity-based analysis of ecological data. J. Stat. Softw. 22, 1–19. ​h​t​t​p​s:​ ​ /​/​d​o​i​.​o​r​g/​ ​1​0​.​1​8​6​3​7​/​j​s​s​​​​ (2007). 74. Legendre, P. & Fortin, M. Spatial pattern and ecological analysis. Vegetation. 80, 107–138 (1989). 75. Shirk, A. J. & Cushman, S. A. sGD: software for estimating spatially explicit indices of genetic diversity. Mol. Ecol. Resour. 11, 922–934. https://doi.org/10.1111/j.1755-0998.2011.03035.x (2011). 76. Waser, P. M. & Elliott, L. F. Dispersal and genetic structure in kangaroo rats. Evolution. 45, 935–943 (1991). 77. Mousadik, A. & Petit, R. J. High level of genetic differentiation for allelic richness among populations of the Argan tree [Argania spinosa (L.) Skeels] endemic to Morocco. Theor. Appl. Genet. 92, 832–839 (1996). 78. Wright, S. Coefficients of inbreeding and relationship. Am. Nat. 56, 330–338 (1922). 79. Wright, S. Isolation by distance under diverse systems of mating. Genetics. 31, 39 (1946). 80. Do, C. et al. NeEstimator V2: re-implementation of software for the estimation of contemporary effective population size (Ne) from genetic data. Mol. Ecol. Resour. 14, 209–214. https://doi.org/10.1111/1755-0998.12157 (2014). 81. Shirk, A. & Cushman, S. A. Spatially explicit estimation of Wright’s neighborhood size in continuous populations. Front. Ecol. Evol. https://doi.org/10.3389/fevo.2014.00062.83 (2014). 82. Singleton, P. H., Gaines, W. L. & Lehmkuhl, J. F. Landscape permeability for large carnivores in Washington: a Geographic Information System weighted-distance and least-cost corridor assessment. USDA Research Paper PNW-RP-549 (Pacific Northwest Research Station, 2002). 83. van Etten, J. R package gdistance: distances and routes on geographical grids. J. Stat. Softw. 76, 1–21 (2017). 84. Dyer, R. J. Is there such a thing as landscape genetics? Mol. Ecol. 24, 3518–3528. https://doi.org/10.1111/mec.13249 (2015). 85. Pope, N. S. & Peterman, W. radish: Fast gradient-based optimization of resistance surfaces. R package version (2020). 0.0.2. 86. Clarke, R. T., Rothery, P. & Raybould, A. F. Confidence limits for regression relationships between distance matrices: estimating gene flow with distance. J. Agricultural Biol. Environ. Stat. 7, 361–372 (2002). 87. Jaffé, R. et al. Beekeeping practices and geographic distance, not land use, drive gene flow across tropical bees. Mol. Ecol. 25, 5345–5358 (2016). 88. Row, J. R., Knick, S. T., Oyler-McCance, S. J., Lougheed, S. C. & Fedy, B. C. Developing approaches for linear mixed modelling in landscape genetics through landscape-directed dispersal simulations. Ecol. Evol. 7, 3751–3761 (2017). 89. Zeller, K. A., Wultsch, C., Welfelt, L. S., Beausoleil, R. A. & Landguth, E. L. Accounting for sex specific differences in gene flow and functional connectivity for cougars and implications for management. Landscape Ecol. 38, 223–237 (2023). 90. Peterman, W. E. ResistanceGA: An R package for the optimization of resistance surfaces using genetic algorithms. Methods Ecol. Evol. 9, 1638–1647 (2018). 91. van Strien, M. J., Keller, D. & Holderegger, R. A new analytical approach to landscape genetic modelling: least-cost transect analysis and linear mixed models. Mol. Ecol. 21, 4010–4023. https://doi.org/10.1111/j.1365-294X.2012.05687.x (2012). 92. Peterman, W. E. & Pope, N. S. The use and misuse of regression models in landscape genetic analyses. Mol. Ecol. 30, 37–47 (2021). 93. Grimm, V. & Railsback, S. F. Pattern-oriented modelling: a ‘multi-scope’ for predictive systems ecology. Philosophical Trans. Royal Soc. B. 367, 289–310 (2012). 94. Landguth, E. L. & Cushman, S. A. cdpop: A spatially explicit cost distance population genetics program. Mol. Ecol. Resour. 10, 156–161. https://doi.org/10.1111/j.1755-0998.2009.02719.x (2010). 95. Cushman, S. A. & Landguth, E. L. Spurious correlations and inference in landscape genetics. Mol. Ecol. 19, 3592–3602. ​h​t​t​p​s​:​//​ ​d​ o​i​.​o​r​g​/​10​ ​.​1​1​1​1​/​j​.​13​ ​6​5​-​2​9​4​X​.​20​ ​1​0​.​0​4​6​5​6​.​x​​​​ (2010). 96. Meyer, H. & Mila, C. M Ludwig. CAST: ‘caret’ Applications for spatial-temporal models. (2019). ​h​t​t​p​s​:/​ ​/​C​R​A​N​.​R​-​pr​ ​o​j​e​c​t​.​o​r​g/​ ​p​a​ c​k​a​g​e​=​C​A​S​T​​​​​​​ 97. Bouyer, J. et al. Mapping landscape friction to locate isolated tsetse populations that are candidates for elimination. Proc. Natl. Acad. Sci. 112, 14575–14580 (2015). 98. Bishop, A. P. et al. A machine learning approach to integrating genetic and ecological data in tsetse flies (Glossina pallidipes) for spatially explicit vector control planning. Evol. Appl. 14, 1762–1777 (2021). 99. Vanhove, M. & Launey, S. Estimating resistance surface on multilayered landscapes using gradient forest and allelic frequencies (Molecular Ecology Resources in, 2023). 100. Baston, D. Exactextractr: fast extraction from raster datasets using polygons. (2021). 101. Friedman, J. H. Greedy function approximation: a gradient boosting machine. Ann. Stat. 29, 1189–1232 (2001). 102. Breiman, L. Random forests. Mach. Learn. 45, 5–32 (2001). 103. Wright, M. N. & Ziegler, A. ranger: A fast implementation of random forests for high dimensional data in C + + and R. J. Stat. Softw. 77, 1–17. https://doi.org/10.18637/jss.v077.i01 (2017). 104. Friedman, J., Hastie, T. & Tibshirani, R. Regularization Paths for Generalized Linear Models via Coordinate Descent. J. Stat. Softw. 33, 1–22. https://doi.org/10.18637/jss.v033 (2010). https://www.jstatsoft.org/v33/i01/ 105. Hijmans, R. Raster: Geographic data analysis and modeling version 3.6. (2022). https://CRAN.R-project.org/package=raster 106. Landguth, E. L., Hand, B. K., Glassy, J., Cushman, S. A. & Sawaya, M. A. UNICOR: a species connectivity and corridor network simulator. Ecography. 35, 9–14 (2012). 107. Dijkstra, E. W. A note on two problems in connexion with graphs. – Numerische Math. 1, 269–271 (1959). 108. Compton, B. W., McGarigal, K. M., Cushman, S. A. & Gamble, L. R. A resistant-kernel model of connectivity for amphibians that breed in vernal pools. Conserv. Biol. 21, 788–799 (2007). 109. Cushman, S. A., Landguth, E. L. & Flather, C. H. Evaluating population connectivity for species of conservation concern in the American Great Plains. Biodivers. Conserv. 22, 2583–2605 (2013). Acknowledgements The authors thank Kim Hersey and Andrew Shirk for reviewing manuscript results. Computational resources from the University of Montana’s Computational Ecology Lab and Shared Computing Cluster contributed to this research (NSF award numbers 2018112 & 1925267). This project was supported by the United States Department of the Interior, Fish and Wildlife Service, Wildlife and Sport Fish Restoration Program, Cooperative Agreement Award F20AC00115 and by Yellowstone to Yukon Conservation Initiative. Genetic sampling was supported by Alberta Conservation Association, Alberta Environment and Parks, Alberta Trappers Association, Alpine Club of Canada, BC Fish and Wildlife Compensation Program, BC Ministry of Forests, Lands, and Natural Resource Operations and Rural Development, Bow Valley Naturalists, Columbia Basin Trust, Cross Foundation, Daishowa-Marubeni International, Disney Wildlife Conservation Fund, Habitat Conservation Trust Foundation, Liz Claiborne Art Ortenberg Foundation. National Geographic Society, Norcross Foundation, Parks Canada Scientific Reports | (2024) 14:28248 | https://doi.org/10.1038/s41598-024-77956-9 14 www.nature.com/scientificreports/ Agency, Shell Fueling Change Grant, US Fish and Wildlife Service Great Northern Landscape Conservation Cooperative Award F15AC01023, US Fish and Wildlife Service Competitive State Wildlife Grant Program Awards F12AP01101 and F15AP01072, Volgenau Foundation, Western Transportation Institute-MSU, Wilburforce Foundation, Wildsight, Woodland Park Zoo, Woodcock Foundation, Yellowstone to Yukon Conservation Initiative, and Zoo Boise Conservation Fund. This research was supported in part by the USDA Forest Service, Rocky Mountain Research Station. The findings and conclusions in this publication are those of the authors and should not be construed to represent any official agency or government determination or policy. Author contributions Conceptualization: ELL MAS APC Methodology: ELL CCD ZAH MASSoftware: ELL CCD ZAHValidation: CCD ELL ZAH MASFormal analysis: ELL CCD ZAH MASInvestigation: ELL CCD ZAH MASResources: AllData Curation: MAS ELL CCD ZAHFunding acquisition: ELL RMI JAG MAS APCWriting – original draft: ELL CCDWriting – review & editing: All. Funding acquisition: ELL RMI JAG MAS APC. Writing – original draft: ELL CCD. Writing – review & editing: All. Declarations Competing interests The authors declare no competing interests. Additional information Supplementary Information The online version contains supplementary material available at ​h​t​t​p​s​:​/​/d ​ ​o​i​.​o​r​g​/​1​ 0​.​1​0​3​8​/​s​4​1​59​ ​8​-​0​2​4​-​7​7​95​ ​6​-​9​​​​.​ ​​ Correspondence and requests for materials should be addressed to C.C.D. or E.L.L. Reprints and permissions information is available at www.nature.com/reprints. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. This is a U.S. Government work and not under copyright protection in the US; foreign copyright protection may apply 2024 Scientific Reports | (2024) 14:28248 | https://doi.org/10.1038/s41598-024-77956-9 15