2.1 Fishing pressure and Indigenous knowledge

Currently, there is no mechanism in place for Indigenous fishers to report the number of fish caught in Mistassini Lake. Thus, to establish trends in fishing pressure, fish abundance, and body size, we conducted semi-directed interviews as in Fraser et al. (2006), during February and July 2018, with 17 elders and fisherman (30–79 years of age, with 13 respondents > 40 years) (see Table S1 for questions asked). Importantly, as per Tengö et al. (2017) elders and fishers were not chosen randomly. They were well-respected authorities on matters of fishing within the community, and there were few other community members with similar knowledge. Answers were not used for questions where respondents explicitly stated a lack of knowledge, as per Gagnon and Berteaux (2009), and the frequency of respondents for a given answer has been provided, using the total number of respondents for that question as the denominator. If the majority of respondents gave similar answers for frequency-based questions, we took this into consideration; because IK is rich in narrative, we also considered relevant pieces of information separately. In addition, we obtained census numbers for all people in the community close to the lake and the number of fish caught by non-Cree fishers for a subset of years (Table S2). Informed consent was obtained prior to each interview and for each of the interviewees (ethics certificate no. 30,008,247).

Rivers included in the study were Chalifour, Icon, and Perch in the south and Takwa in the north. Communicated incidentally by two IK respondents and by several Cree fishers and community members in 2017 and 2018, Takwa was perceived to be relatively unaffected by fishing until ~ 2015 when larger boat motors made access easier.

2.2 Fish sampling

Fish were sampled during spawning (after ice-off: mid-May in the south and early June in the north) at spawning rivers in 2002 and 2003 by Dupont et al. (2007) (historical), and in 2015, 2016, and 2017 (contemporary) by us (see Table 1 for sample sizes). Sampling was collaborative with subsistence fishers for 2015–2017. Walleye were captured via angling using the same lures and a combination of boats and shore fishing, from the same locations within rivers, for both historical and contemporary sampling (Table 1). Catch per unit effort was not available for historic samples or collaborative sampling and is therefore not included here for contemporary sampling. After capture, fish were immediately placed in freshwater baths with aerators. From each walleye, we collected total and fork length (TL ± 1 mm), wet mass (± 50 g), sex (M, F, U (unknown, either spawned out or premature)), and a tissue sample for genetics; otoliths were collected from a random subsample. Live walleye were returned to the water near the location of capture. Opercular bones but not otoliths were collected for aging for historic samples (Table 1), and this was done before this study, for Dupont et al. (2007); 2015 and 2017 otoliths were aged at the Wisconsin Cooperative Fishery Unit, US Geological Service, University of Wisconsin, Stevens Point, USA. Otoliths were aged by two experienced readers; if they disagreed on an age, they examined the structure together to agree upon one for that structure. No walleye were aged using both opercular bones and otoliths.

2.3 Body size at spawning and size-at-age

We modeled both body size (total length and mass) and size-at-age in this study because we had a far greater sample size for body size estimates than aged samples, and mass estimates had not been correlated with historic aged samples. Thus, evaluating body size allowed us to investigate changes in length and mass on a per-river, per-sex, and per-year basis.

2.3.1 Body size

We used multiple regressions and ANOVA in R (R Core Team, 2017) to test our prediction that body size of breeding adults had been reduced in southern rivers between 2002/03 and each 2015, 2016, and 2017. Year was set as a factor. Error was normally distributed for total length (TL), and mass was log-transformed to improve fit of the error term. Our full model for each TL and mass (Y_i) included the following.

To determine the best model, we used backward stepwise model selection and AIC (Akaike, 1974). Significance was detected at an alpha of 0.05, and all multicomparison p-values were adjusted using the false discovery rate (FDR) method for 64 planned contrasts (Tables S3 and S5) (Benjamini and Hochberg 1995).

There were insufficient samples collected across locations in 2002 and 2003 to use these years independently for body size analysis. Since no population genetic structure existed between 2002 and 2003 within rivers (Dupont et al., 2007), they were combined for all analyses, and denoted as 2002/03. In addition, in 2017 we were unable to collect any female walleye from Perch River, nor a sufficient number of female walleye in Icon River to be able to use them in length/mass models (see Table 1 for sample numbers).

2.3.2 Size-at-age

To test our prediction of reduction in size-at-age in the southern populations relative to the northern one through time, we used a Bayesian hierarchical regression model. We used Bayesian as opposed to frequentist modeling to account for possible bias due to sampling gear, as well as small and variable sample sizes for aging structures across rivers. In addition, only total length was modeled because mass data were not included in the historical dataset containing age information. We assumed walleye total length (TL) was normally distributed such that:

with shape parameters and representing the mean and standard deviation for walleye total length, respectively. Mean total length for the ith walleye was then modeled using linear regression:

We used vague normal priors for all coefficients and modeled hyperpriors for age and sex by river. Location (southern rivers versus northern river) and history (contemporary versus historic samples) were coded as categorical variables.

The Bayesian model was run using JAGS version 4.3.0 (Plummer, 2003) in R, using rjags and run.jags (Denwood, 2016). We described the posterior distribution for the model using four MCMC chains. Starting parameter values for each chain were jittered. Each chain took 20,000 samples of the posterior, thinned at a rate of 50. The adaption period was 1,000 iterations, and a burn-in rate of 50% was used, for a total chain length of 2,050,00. We evaluated MCMC chain convergence by visual inspection of trace plots to assess mixing. Additionally, we ensured that each parameter had effective sample sizes > 1,000 and that they passed the Gelman–Rubin diagnostic test with potential scale reduction factors (PSRF) <1.1 suggesting convergence on a common posterior mode (Gelman et al., 2013).

2.4 Sequencing

DNA was extracted using a modified Qiagen blood and tissue kit protocol (Qiagen Inc., Valencia, CA) (see Table 1 for sample sizes) and was sequenced using individual-based genotyping by sequencing (GBS). Libraries for Ion Proton GBS were prepared using the procedure described by Mascher, Wu, Amand, Stein, and Poland (2013) at IBIS, Université Laval, Québec, Canada, with modifications described in Abed, Légaré, and Pomerleau (2018). Libraries were prepared for sequencing using an Ion CHEF, Hi-Q reagents, and P1 V3 chips (Thermo Fisher), and the sequencing was performed for 300 flows. Enzymes used to cleave the DNA were rare cutter pst1 and frequent cutter msp1.

Single nucleotide polymorphisms (SNPs) were determined from raw sequence reads using the stacks pipeline v1.45 (Catchen, Hohenlohe, Bassham, Amores, & Cresko, 2013), and de novo sequence alignment, on the supercomputer Guillimin from McGill University, managed by Calcul Québec and Compute Canada. Preprocessing of fastq files was completed using fastQC (https://www.bioinformatics.babraham.ac.uk/projects/fastqc/) to assess read-quality before and after using cutadapt (Martin, 2011) to trim any remaining adapters and remove sequences < 50bp. Our stacks parameter optimization method was similar to Mastretta-Yanes, Arrigo, and Alvarez (2015), but we did not estimate error rate because we did not have enough positive controls to do so. Final stack parameters included default settings with the following custom options: within process_radtags, 80bp trim length; ustacks, SNP model, alpha = 0.1 for SNP calls, -m = 7; cstacks, -n = 3; rxstacks, log-likelihood cutoff = −30 for SNP calls; populations, log-likelihood cutoff of −30 for SNP calls, choose single SNP, maf = 0.01, -r = 0.8. We ran populations twice. First, we used the parameters listed here, and generated a blacklist of loci consisting of loci with F_IS < −0.3. We then reran populations with the same parameters listed here using this blacklist, with -p 6/8 (all years and rivers separately) and -p 4/6 (Icon–Perch merged); the numerator is the total number of populations required to have that locus, while the denominator is the total number of populations in the population map. All years and rivers were used initially to find the best-supported population structure and to discover any changes that were occurring between rivers and years. Once the appropriate population structure was determined, Icon–Perch were merged (see Population structure section of Results below for a full explanation). No negative controls produced stacks, and all positive controls assigned to the correct populations using discriminant analysis of principal components (DAPC) in Adegenet (Jombart, 2008; Jombart, Devillard, & Balloux, 2010). After quality trimming and filtering, an average of 8,457 (for -p-4/6) and 8,728 (for -p 6/8) high-quality SNPs were used to estimate population structure, genetic diversity, and effective population size (Ne). See Table 2 for a summary of the number of loci, SNPs, and sequencing coverage at each filtering stage.

2.5 Population structure

To test our prediction that harvesting in the southern populations would change genetic population structure, potentially homogenizing structure between southern rivers over time, we assessed population structure using DAPC, ADMIXTURE (Alexander, Novembre, & Lange, 2009), and genetic distance (F_ST) (using GenoDive and 999 permutations, Meirmans & Van Tienderen, 2004; Weir & Cockerham, 1984). The optimal number of principal components (PCs) to retain for DAPC was determined using the xval procedure, using n/3 (recommended by the manual) as the maximum number of PCs allowable, and 500 replicates. The population grouping that best fit the data was assessed using Bayesian information criterion (BIC) for DAPC, while for ADMIXTURE analysis we used cross-validation (CV) and 500 bootstrap replications. Both analyses were completed at least four times using different bootstrap values and numbers of replicates to ensure results were stable. Based on DAPC (when all rivers and years were considered separately), it was clear that 28 individuals, primarily sampled in Icon and Chalifour Rivers, were from different unsampled genetic source populations. We removed these individuals, reran the populations module of stacks, and conducted all subsequent analyses using this reduced dataset (i.e., for analyses where we show results for years and rivers considered separately and also where Icon and Perch were merged).

2.6 Removing loci potentially under selection

Global outlier loci (loci putatively under selection) were detected using PCAdapt (Luu, Bazin, & Blum, 2017), using the scree plot method to determine the best number of PCs (K) to retain (Jackson, 1993), and Mahalanobis distance with alpha = 0.1 to determine outliers. PCAdapt does not use predefined population structure, but instead ascertains structure based on PCA. The program detects outliers based on how they relate to the structure of populations on the PCA (i.e., the distance between a point and a distribution). After removing the outlier loci from the dataset, the effect of linkage disequilibrium (LD) on population structure was assessed by finding markers that were in LD (r² = 0.7) using plink v1.9 (Chang et al., 2015). We removed these markers (n = 507) and reanalyzed population structure with DAPC. Since linked loci had no effect on structure, they were retained for all subsequent analyses. Genetic diversity, F_ST, and Ne analyses were completed both including and excluding global outlier loci—while there was little difference between results including or excluding outlier loci for genetic diversity and Ne, the magnitude of F_ST was greater with outlier loci included, and the conclusion changed in one case for F_ST. Thus, we have included only results for neutral loci here for these two metrics.

2.7 Genetic diversity (H_E) and Ne

To test the prediction that genetic diversity (expected heterozygosity, H_E) and Ne were reduced over time in southern populations, genetic diversity estimates were obtained using the populations model of stacks, and per-generation Ne (5- to 7-year generation time) was estimated using the linkage disequilibrium method in NeEstimator v2.01 (Do, Waples, & Peel, 2014), requiring a minimum minor allele frequency of 0.02. Estimates were per river, for each of 2003 and 2015, with males and females combined (see genomic samples in Table 1). Ne estimates and confidence intervals were corrected for linkage by correcting for chromosome number according to Waples, Larson, and Waples (2016).

2.8 Signatures of selection

To test the prediction that signatures of selection would be most evident between timepoints within southern and not northern river(s) and that putatively selected loci would be associated with relevant biological processes, analyses to determine outlier loci were conducted with PCAdapt using the method described above. We chose to use PCadapt instead of other programs for outlier detection for two reasons. First, only two “populations” were sampled, and the authors of OutFlank do not recommend its use in these situations (Whitlock & Lotterhos, 2015). Second, the biological scenario in our populations is divergence and admixture. In this scenario, all other outlier programs that could be used with our data (SNPs not haplotypes) have between 20% and 40% false discovery rates (PCAdapt has 10%) (Luu et al., 2017). Analyses were conducted for sexes both combined and separately: (a) for all populations combined, (b) for southern rivers only, and (c) within each population. For all analyses, except when all populations and years were included, the populations module of stacks was rerun including only the populations and/or sexes that were being contrasted, specifying that loci had to be present in both populations (see Table S8 for sample sizes and numbers of loci in each analysis). For sex-based analysis of the full dataset including all rivers and years, loci were required to be present in 6 of 8 populations.

To determine possible functions for outlier loci, for all within-river historic–contemporary contrasts for which there were outlier loci, FASTA files were blasted, mapped, and annotated using blast2go (Götz et al., 2008). Default parameters were used with the following custom choices: proprietary cloudblast, fast-blast, UniProtKB/Swiss-Prot (Swissprot_v5) database, blast e-value 1.0E-5, 10 blast hits, and filtered GO by taxonomy taking only matches to animals (Metazoa).

3.1 Fishing pressure and Indigenous knowledge

Of 17 elders and fishers, most reported reductions in the size and number of walleye caught in the lake (15 and 14 respondents, respectively, or 94% and 93%, respectively, after respondents who did not know were removed) within the last 5–20 years (Table 3). Eleven respondents expressed concerns directly about overfishing, fishing during spawning, or taking of too many fish during spawning. There was no consistent change in the number of fish caught by non-Cree fishers between 1997 and 2015 (virtually the same in 1997, 2011, and 2015, but 54% higher in 2003) (Table S2), but the community of Mistissini (3,724 people in 2016) grew by ~ 29% between 1997 and 2011, and the population and number of households in the community increased by ~ 50% between 1997 and 2016 (Table S2). In addition, while there are currently no data on the number of fishers in the community nor the proportion of the population that fishes, the majority of fishers (16/17) fished in the southern area of concern in the lake, more than double than in all other areas of the lake except Takwa River (where 9 of 17 fishers fished) (Table 3).

3.2 Body size at spawning and size-at-age

3.2.1 Body size

Our prediction that body size would be reduced in southern populations within a 1- to 2.5-generation period was supported. The main effects, year, river, and sex all had a significant effect on total length and mass (Table 4), and the best-fit models for each TL and mass each contained a three-way interaction between year, sex, and river. AIC was > 10 better for the full model for TL and > 6 better for mass. Both regression models were significant (TL, R² adj = 0.418, F_{27, 1,486} = 41.24, p < .001; mass R² adj = 0.3907, F_{27, 1,481} = 36.81). Mean TL and mass decreased significantly for both sexes between 2002/2003 and each of 2015, 2016, and 2017 in the southern rivers (TL 7%–21% and mass 22%–47% reductions), except for Perch River males in two contrasts and Chalifour females in two contrasts. While there was no significant change between 2002/2003 and each of 2016 and 2017 years for females in Chalifour, the trend in decline remained clear (Figure 2); lack of significance may relate to low female sample size in this river (Table 1). Indeed, the trend for Chalifour females was particularly evident when contrasted to Takwa (the reference northern river), where mean sizes of fish were consistent across all sampling years. Finally, a sex bias for more males than females being captured at spawning sites was consistent for all sampling years and for all rivers, including Takwa (Figure S1).

TABLE 4. Analysis of variance table for best fit (full) model for each walleye total length (TL) and mass, with response log(mass). Years are each 2002/2003, 2015, 2016, and 2017. Rivers are Chalifour, Icon, Perch, and Takwa

	df	Sum Sq	Mean Sq	F value	p
TL
Year	3	656,921.887	218,973.962	111.058	.000
River	3	393,310.528	131,103.509	66.492	.000
Sex	1	860,350.946	860,350.946	436.347	.000
Year:River	8	199,492.923	24,936.615	12.647	.000
Year:Sex	3	9,581.311	3,193.770	1.620	.183
River:Sex	3	36,772.902	12,257.634	6.217	.000
Year:River:Sex	6	39,265.708	6,544.285	3.319	.003
Mass
Year	3	30.776	10.259	91.847	.000
River	3	23.854	7.951	71.189	.000
Sex	1	38.923	38.923	348.478	.000
Year:River	8	11.978	1.497	13.405	.000
Year:Sex	3	1.025	0.342	3.060	.027
River:Sex	3	2.461	0.820	7.345	.000
Year:River:Sex	6	1.996	0.333	2.979	.007

3.2.2 Size-at-age

Our prediction that fish in the southern rivers would be smaller for their age through time was supported, although the reduction in size was small (see Table 5 for posterior means and 95% credible intervals). Overall, fish were larger in the southern than the northern river(s), possibly due to a longer (warmer) growing season leading to larger size (though we do not have thermal data). Over all rivers, fish were 29.4 mm larger contemporaneously than they were historically, and males were 48.2 mm smaller than females on average. Lastly, and the term that tested our hypothesis, fish in the southern rivers were 13.7 mm smaller relative to fish in the north in contemporary relative to historical samples.

TABLE 5. Posterior means (PM) and 95% credible intervals (CIs) for walleye size-at-age model for total length (in mm)

Parameter	Parameter	PM (mm)	95% CI	Percent of the posterior mass below 0
	Intercept	463.8	418.7–506.4	0%
	Age	10.8	8.6–12.9	0%
	Location	27.8	−9.7–80.4	6%
	History	−29.4	−50.0–−9.3	100%
	Sex	−48.2	−65.6–−32.6	100%
	History x Location	−13.7	−37.1–9.6	87%

Note

• Age is the effect of each year on length, location is south versus north, and history is contemporary versus historical.

All parameter estimates passed convergence checks. Each parameter had effective samples sizes > 1,000 and passed the Gelman–Rubin diagnostic test with potential scale reduction factors (PSRF) <1.1, suggesting convergence on a common posterior mode (Gelman et al., 2013).

3.3 Population structure

Our prediction that population structure would change over time, possibly including homogenization of structure between southern rivers, was supported for Icon and Perch Rivers in two of the three analyses. Specifically, ADMIXTURE and F_ST analyses supported this prediction, while DAPC did not. Using both DAPC and ADMIXTURE 3 populations best described the data (k = 3,053.162 and CV = 0.421, respectively) (Figure 1), but the difference between CV 2 and CV 4 was small for ADMIXTURE (CV of 2 = 0.423 and CV of 4 = 0.428). In the 3-population scenario, Icon and Perch grouped as a metapopulation and Chalifour and Takwa Rivers were independent, with 2002/2003 and 2015 samples grouping together for each river. In a 4-population scenario, many Perch 2003 individuals showed a substantial fraction of loci that were different from the Icon–Perch group. Genetic differentiation by F_ST mirrored what was evident in the K = 4 scenario. F_ST showed weak differentiation (p < .05) between Perch and Icon in 2002/2003 (Table 6), and then merged as a single metapopulation in 2015. At a within-population level, Chalifour River was differentiated between timepoints, Icon did not diverge between timepoints, Perch diverged marginally between timepoints, and Takwa did not diverge between timepoints. Given that k = 3 was identified as the best structure overall, subsequent genetic diversity and Ne analyses were conducted using a metapopulation structure for Icon–Perch, with Chalifour and Takwa identified separately, but years were defined as separate populations for temporal analysis.

3.4 Genetic diversity and Ne

Our prediction that genetic diversity and Ne would be reduced between historical and contemporary timepoints was not supported. Genetic diversity fell within a tight range for all populations over all years, ranging from 0.21 to 0.23 (Figure 3a). Confidence Intervals (CIs) overlapped between timepoints for H_E in Chalifour and Icon–Perch, and there was a 4.9% loss in H_E in Takwa, though the reduced H_E still fell within the range of southern populations. Point estimates of Ne ranged from 1741 to 3,146 individuals across all populations, with the lowest and highest values being in the northern population. Ne CIs also overlapped between timepoints in Chalifour and Icon–Perch Rivers, and the data suggested a doubling in Ne over time in Takwa River (Figure 3b). There were likely insufficient samples sequenced per population to accurately detect a difference between time periods in populations consisting of thousands of individuals, and reliable detection of changes to Ne often requires ~ ten generations to have passed (Nunziata & Weisrock, 2018; Waples & Do, 2010); however, these results demonstrate that all populations remained large (i.e., Ne in the thousands).

3.5 Signatures of selection

Our prediction that signatures of selection would be present between historic and contemporary timepoints within southern rivers but not the northern river, with putatively selected loci related to growth, body size, and/or maturation, was supported. Eleven to 263 loci were outliers (0.17%–2.83%) depending on the scenario tested (Figure 4a-b). Outliers were found in the global PCA that included all rivers and both timepoints, for each F/M together, and for F & M individually. Removing these outliers did not change the population structure when all rivers (i.e., including Takwa) and years were assessed. On the contrary, when the southern rivers were analyzed as a unit (i.e., the south historic versus contemporary) and on their own (i.e., each Chalifour and Icon–Perch historic versus contemporary years), for F/M combined and for each F & M separately, population structure existed between the two timepoints and removing outlier loci usually collapsed the population structure. Further, there was no population structure between timepoints in Takwa (and thus no outlier loci) (Figure 4b, Table S8, and Figures S2–S16). In sum, while there were outliers separating north from south, outlier SNPs disproportionately contributed to the PCs between years in the southern subset. In addition, the greatest proportion of outlier SNPs found in each historic/contemporary PCA in the south maintained population structure between years (Figure 4c). Lastly, parallel outliers existed between the southern rivers (Figure 4d); of note, more outliers were in common between Icon–Perch F and Chalifour M (18 outliers) than between Icon–Perch F and Chalifour F (0) or Chalifour M and F (3).

Sex-specific analyses provided greater resolution than with sexes combined (Figures 4c and d). For example, outlier SNPs maintained population structure between timepoints in more of the southern rivers in sex-specific rather than combined analyses (Table S8, Figures S2–S7). In addition, there was no population structure in Icon–Perch males, but there was in females (Figure 4c and Figures S12–S13). Lastly, the number of outliers in common between rivers differed between the sexes (Figure 4d), though this may be due in part to the number of individuals sequenced.

Blasts were conducted for southern (F/M combined and separately), Chalifour (F/M combined and separately), and Icon–Perch F (no outliers were found for Icon–Perch M or Takwa) Rivers. Southern F/M combined and M had no blast hits. Otherwise, between 2 and 6 alleles were annotated for each blast. Because annotations were completed against all mapped Metazoa, at level 2 go annotation, functional annotations included many different biological processes, molecular functions, and cellular components. Three relevant processes indicated were growth, metabolism, and developmental process (Table S9).

4.1 Conclusions and management implications

We have presented coupled IK, phenotypic, and genetic evidence consistent with genetic changes associated with harvesting within 1–2.5 generations in wild walleye populations. Links with fishing pressure are not conclusive, but this study sets a precedent for the time frame needed for investigating concerns regarding harvest-induced evolution in fisheries. Furthermore, sex-specific dynamics for both body size and genomics herein highlight the importance of collecting sex-specific data; if we had looked at both sexes together, the stronger signal in females (i.e., possible changes in fecundity) would likely have been muted, providing inaccurate information for fisheries managers.

Our study illustrates the benefits of integrating life history and genomic methods for conservation in order to understand the factors affecting population change (Bernatchez et al., 2017), of interweaving these with IK, and of iterative population monitoring practices (Flanagan, Forester, Latch, Aitken, & Hoban, 2018); that is, this study would not have been possible without the historic data. Considerations for Cree management could include that observed phenotypic and genetic changes may cause reduced productivity (Allendorf et al., 2008; Hutchings, 2005) and that genomic changes are clearly nascent here. Depending on the severity of harvest (which is not precisely known in this case) and life history (Audzijonyte & Kuparinen, 2016), fisheries-induced changes may be reversed in 9 generations (Conover, Duffy, & Hice, 2009 or less (Feiner et al., 2015) if fishing is halted; the ability of populations to recover depends on their trophic position, age-specific fecundity, and survival at each life stage (Audzijonyte & Kuparinen, 2016) (Box 1).