Fieldwork and genetic sample collection
From 2010 to 2017, we trapped harvest mice (Reithrodontomys spp.) at 27 locations, including multiple sites within each of the five designated marsh complexes within the U.S. Fish and Wildlife Service (USFWS) Suisun Bay Area Recovery Unit (Fig. 1; Supplemental Fig. 1). At each site, we used ~ 100 Sherman live traps (H.B. Sherman Traps, Tallahassee, FL), spaced at ~ 10 m intervals, with the layout depending on the wetland size and shape. We baited traps with mixed bird seed and ground walnut, added cotton bedding for warmth, set the traps at dusk, and checked them at dawn for three to four consecutive days. We plucked hair as a source of DNA. Prior to sampling from an individual, we physically wiped down the forceps with a clean tissue, sterilized the forceps in a 2% bleach solution, rinsed with water to remove the bleach, and dried the forceps with a second tissue (Statham et al. 2016). We stored the hair in 95–100% ethanol until DNA extraction. Animal trapping, handing, and genetic sampling procedures were approved by UC Davis Institutional Animal Care and Use Committee (IACUC No. 19686) and authorized by the appropriate state and federal agencies (SC-011578, TE-35000), and under a cooperative agreement between the California Department of Fish and Wildlife and the U.S. Fish and Wildlife Service.
Laboratory methods
We extracted DNA from 777 harvest mouse samples using one of two methods. The first method involved digesting samples in a hair lysis buffer (Statham et al. 2016) and purifying with a modified phenol chloroform method. For the second method, we used a DNeasy Blood & Tissue kit (Qiagen Ltd), modifying the digestion buffer to include 20 µl 1 M Dithiothreitol (DTT), 300 µl buffer ATL, and 20 µl (mg/ml) proteinase K. Each extraction set included a negative control to monitor for contamination. Both R. raviventris and the morphologically similar R. megalotis (western harvest mouse) are present throughout the study area (Statham et al. 2016; Sustaita et al. 2018). We PCR-amplified and sequenced a portion of the cytochrome b gene, and used the associated > 10% sequence divergence to discriminate between species with certainty (Statham et al. 2016).
Microsatellite amplification
All subsequent analyses were restricted to the R. raviventris samples. We screened all samples with 20 microsatellite loci: Rrav 1, 6, 8, 10, 13, 18, 21, 29, 36 (Reponen et al. 2014), R34 (Vázquez-Domínguez and Espindola 2013), and new loci Rrav 40, 43, 44, 46, 47, 49, 51, 57, 61, 62 (Supplemental Table 1). We discovered and developed the ten new loci using next generation sequencing (see the Supplemental Information for full details). We PCR-amplified loci in 4 multiplexes using the Qiagen Multiplex PCR Kit (Valencia, CA, USA) according to manufacturer’s guidelines (including Q-solution), and with the following thermal profile: 15 min at 95 °C, 33 cycles of 30 s at 94 °C, 1.5 min at 58 °C, and 1 min at 72 °C, and a 10-minute extension at 72 °C. We electrophoresed products on an ABI 3730 capillary sequencer and scored alleles relative to an internal size standard, Genescan 500 LIZ (Applied Biosystems), using STRand software (Locke et al. 2007). All hair samples were genotyped in duplicate.
Genetic analyses
Unless otherwise stated, all statistical analyses were conducted on sampling locations with ≥ 10 individuals. We tested for deviations from Hardy-Weinberg and linkage equilibrium using Genepop http://genepop.curtin.edu.au/) with default Markov chain parameters. We corrected for multiple tests using the sequential Bonferroni method (Rice 1989). We calculated the observed heterozygosity (Ho), expected heterozygosity (He), and average number of alleles per locus (A) in the Excel Microsatellite toolkit (Park 2001). We calculated allelic richness (Ar; the mean number of alleles across loci) and private allelic richness (Pr; the mean number of private alleles across loci) rarified for 10 diploid individuals in HP-Rare v1.1 (Kalinowski 2005). We estimated pairwise FST among sampling sites using Arlequin 3.5 (Excoffier and Lischer 2010). We calculated Nei’s DA genetic distance (Takezaki and Nei 1996), in the program Populations 1.2.32 (Langella 1999; http://bioinformatics.org/~tryphon/populations/), and then used the program to generate a neighbor-joining tree with 200 bootstrap replicates.
We investigated population subdivision across Suisun Bay using all R. raviventris individuals (including those from populations with < 10 individuals) in the Bayesian clustering program Structure 2.3.3 (Pritchard et al. 2000). We used the admixture model with correlated allele frequencies without a priori population assignment (Pritchard et al. 2000; Falush et al. 2003). Iterations were run assuming numbers of genetic clusters (K) ranging 1–10, with a burn-in of 50,000 steps followed by a run of 50,000 steps. Simulations were run five times at each value of K to assess consistency across runs. We followed the guidelines of Pritchard et al. (2000) and the Structure manual to infer K.
Genetic effective population size
We estimated the genetic effective population size (Ne) of R. raviventris populations using the bias-corrected linkage disequilibrium method (Waples and Do 2008) in the program NeEstimator 2.1 (Do et al. 2014). We used the random mating model, excluded low-frequency alleles at the 0.05 level, and calculated 95% confidence intervals. In this single sample method, we examined single year samples from each subpopulation identified in the program Structure.
Landscape genetic analyses
We examined how landscape variables restricted gene flow among R. raviventris populations using isolation by resistance modeling (McRae et al. 2008). We obtained spatial data for landscape variables from the U.S. Forest Service for cover types (Existing Vegetation: Region 5; https://data.fs.usda.gov/geodata/edw/datasets.php) and the Data Basin for elevation (90 m DEM of California; https://databasin.org/datasets/). We then clipped the layers to the study area in ArcMap 10.4 (Environmental Systems Research Institute (ESRI) 2015). We grouped cover types into the following broad categories: Barren, Cropland, Grassland, Urban (developed areas including roads), Water (including rivers, lakes, marine), Wetland, and Woodland (for further information see the Supplemental Information). R. raviventris are restricted to salt and brackish marshes that occur at sea level (Shellhammer 1982). The mean higher high water (MHHW) for Suisun is at 2 m and represents the highest elevation of salt marshes. Therefore, we separated our elevation layer into two categories ≤ 2 m and > 2 m.
Rather than assigning a priori resistance values to landscape characteristics, we chose to follow a causal modeling approach (Cushman et al. 2006) to test a range of values, allowing the data to determine the optimal weighting scheme. For each landscape variable we generated resistance surfaces where the variable of interest was weighted 2, 5, 10, 25, 50, 100, 200, and 500, while other variables were weighted 1 (Roffler et al. 2016). In the case of elevation, we varied the resistance weight of elevations > 2 m. We generated 65 models including a null model of Isolation by Distance (IBD). In the IBD model, all cell values in the resistance surface were weighted 1. This model is considered the appropriate surrogate for geographic distance for circuit theory analyses (Tucker et al. 2017). We estimated the resistance distance between all pairs of populations under each of the landscape resistance schemes using Circuitscape v 4.05 (McRae et al. 2008), and then evaluated each of these landscape distances against genetic distance (Nei’s DA). We tested the relationship at two different scales: the entire Suisun Bay dataset (hereby called Bay-Wide) and among the Northern Marshes. The Bay-Wide dataset consisted of 20 sites (each representing ≥ 10 individuals). The Northern Marshes consisted of 17 sites, excluding the two Contra Costa shore sites and Ryer Island. The Bay-Wide dataset allowed us to assess the causes of deeper subdivision, while the Northern Marshes dataset allowed an examination of more nuanced population subdivision among populations not separated by Suisun Bay.
We tested among resistance values for each landscape variable using maximum-likelihood population effects (MLPE) models (Clarke et al. 2002) estimated in the R package ‘lme4’ (Bates et al. 2015). We fitted sixty-five models for univariate analyses at both the Bay-Wide and Northern Marsh scales (Supplemental Table 2). Model support was assessed through the Information Theoretic approach using the Akaike Information Criterion of small sample sizes (AICc) (Burnham and Anderson 2002. The optimal resistance weight for each landscape variable was determined as the resistance value with the lowest AICc in univariate MLPE models (Shirk et al. 2018). Where AICc indicated similar support for two or more resistance values we used the resistance weight with the highest R2 between the resistance distances and genetic distances among sites (calculated using linear regression in Microsoft Excel 2019). We excluded landscape variables from subsequent multivariable analyses when the best supported resistance weight was no improvement over IBD (within ΔAICc < 2.0 of IBD).
Multivariable models were constructed for all possible combinations of landscape variables supported by the univariate analysis. We used MLPE for multivariable modeling. Landscape resistance distance matrices from Circuitscape incorporate both the effects of landscape variables and geographic distance between sample locations. Therefore, we subtracted the geographic distance (the null resistance surface) from each landscape resistance distance to isolate the impact of the landscape variable on landscape resistance (Tucker et al. 2017). Additionally, we included geographic distance in each model as an independent covariate. Finally, we excluded all models with significant multicollinearity (one or more variables with Variance Inflation Factor, VIF > 10) and uninformative landscape variables (β coefficient 95% confidence intervals included 0). Final candidate models included a null model of IBD and model support was assessed using AICc. We used the most parsimonious model (i.e., the model with the lowest AICc) for downstream multivariable resistance surface analyses in ArcMap. This resistance surface was then used to generate a cumulative current map of R. raviventris connectivity in the program Circuitscape.