Samples

Altogether, we considered 77 mussel samples (total sample size N = 4304, individual sample size N = 18–173) representing five geographical contact zones between M. edulis and M. trossulus: the Gulf of Maine in the northwestern Atlantic (12 samples, N = 428), Loch Etive in northern Scotland (2 populations, N = 160), western Baltic Sea (8 samples, N = 601), Bergen city area in western Norway (5 samples, N = 365) and the coasts of the Kola Peninsula in northern Russia: 24 samples from the White Sea (N = 1105) and 26 samples from the Barents Sea (N = 1645) (Fig 1). Detailed information about samples and sampling localities is provided in the S1 Table.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Map of the study area and variation in shell morphotype frequencies.

The bottom panel (maps G-K) shows five geographical contact zones between M. edulis and M. trossulus, maps in the upper panel (A—F)—other studied areas. Pins depict sampling sites. Pie diagrams depict proportions of T-morphotypes (black sector) and E-morphotypes (white sector) in M. trossulus (diagrams with a red border) and M. edulis (those with a blue border) in combined samples from particular regions. If the data on salinity in sampling localities are available and considered in the analyses, it is indicated by the color of pins (light green–brackish, dark green–saline, white–salinity is unknown) and the proportions of the T-morphotypes in combined samples from brackish and saline localities are presented separately in diagrams placed on light and dark green background, respectively. Source data are given in S1 Table and S2 Table. Inkscape 0.92 [26] was used for producing the map.

https://doi.org/10.1371/journal.pone.0249587.g001

Samples in Gulf of Maine were permitted by Maine Department of Marine Resources (Special License Numbers ME 2014-41-00 and 2015-51-01) and Fisheries and Oceans Canada 2015 (Licence 341852). No special permits were required for the field studies carried out in other regions.

The Barents Sea samples were taken in the Kola Bay and at the open oceanic coast of the eastern Murman. Based on the salinity in the sampling localities, they were classified into brackish (salinity 5–30 ppt) and saline (>30 ppt). The first group consisted of nine samples from the freshened top of the Kola Bay and three samples from the open coast. The second group consisted of eight samples from the mouth of the Kola Bay and six samples from the open coast (Fig 1).

As for the samples from the other contact zones, all American samples and two out of five Norwegian samples were from saline habitats, while all the others were from brackish habitats. Salinity conditions in the sampling localities were either taken from the literature [27–32] or, in case of the few American and the Barents Sea open coast localities, predicted based on the presence or absence of large rivers nearby.

In addition to the samples taken in the five contact zones, we identified the morphotypes in 27 samples (total sample size N = 912, individual sample size N = 12–76) of supposedly pure blue mussel species from distant localities: M. trossulus from Passamaquoddy Bay and M. edulis from the Gulf of Saint Lawrence in eastern Canada, M. trossulus from the northern Baltic Sea, from Puget Sound in eastern Pacific and from multiple areas of western Pacific, M. edulis from southwestern Greenland, from the Long Island Sound and Cape Cod in the eastern USA, and from saline and brackish localities in Europe and in the southwestern Barents Sea (Fig 1, S2 Table). To note, the Baltic M. trossulus is more strongly introgressed by M. edulis alleles than other populations [33]. Information about the species identity of regional populations and salinity conditions in sampling localities was taken from the literature. Taxonomic affinities of mussels from Canada and from western Norway, where both species could be expected, were confirmed genetically (see S2 Table for details).

Genetic characters

A part of the material, from the contact zones, was genotyped in previous studies [3, 16, 31]. The other samples were collected and genotyped specially for this study (see S1 Table). The Gulf of Maine mussels, both from previous and new material, were genotyped using 109 260 SNPs (single nucleotide markers) as described in [31] (including some material from [34]). The mussels from the other areas were genotyped using various sets of allozyme loci, which, as a rule, included the Est-D, Gpi, Pgm and Odh loci. These four loci were involved in the initial diagnosis of M. edulis and M. trossulus and in description of all the contact zones under consideration [6, 19, 35]. They individually show 70–95% allele frequency differences between the species [6], and, being less affected by introgression than most of the conventionally used PCR-based “diagnostic” markers [1, 36], are reliable markers for species identification everywhere. The other new samples were genotyped by Est-D, Gpi, Pgm and Odh as in [16]. For seven samples from the Barents Sea the data on only three loci (Est-D, Gpi and Pgm) were available (see S1 Table). The SNP data set and each of the four regional 4-locus allozyme sets (from the Baltic, Norway, Scotland and Russia) were analyzed separately using STRUCTURE or fastSTRUCTURE software ([37], settings as in [16] and [31]). The obtained q-values are defined as estimates of the proportion of M. trossulus genes in individual genotypes (proportion of M. edulis genes is therefore 1-q). The material from Russia was also analyzed by three loci (all but Odh) to show that the exclusion of Odh did not affect the inference. The mussels were further classified into two categories by their q-values: genotypes dominated by M. trossulus genes (q-value > 0.5) and those by M. edulis genes (q-value ≤ 0.5). For ease of presentation, these categories will be referred to as “M. trossulus” and “M. edulis” genotypes although each includes both purebreds and hybrids.

Morphological characters

Data on the White Sea samples were taken from [16] and the other samples were processed accordingly. We measured the maximum length of each shell to the nearest 0.1 mm with electronic calipers and investigated the inner surface of the valves under a dissecting stereo-microscope. A mussel was classified as a T- or an E-morphotype based on, respectively, presence or absence of an uninterrupted strip of the prismatic layer under the ligament on the inner side of the shell.

Predictive values

For each sample we calculated the frequencies of M. trossulus genotypes (Ptros) and of T-morphotypes (PT) and four indices reflecting the strength of association between genotypes and morphotypes: P(T|tros)–the proportion of T-morphotypes among M. trossulus, P(E|edu)–the proportion of E-morphotypes among M. edulis (for practical reasons we used P(T|edu) = 1- P(E|edu), the proportion of T-morphotypes among M. edulis), P(tros|T)–the proportion of M. trossulus among T-morphotypes, P(edu|E)–the proportion of M. edulis among E-morphotypes. P(tros|T) and P(edu|E) are the key indices because they show, respectively, how likely it is that a randomly taken T-morphotype mussel is M. trossulus and a randomly taken E-morphotype mussel, M. edulis.

Here we would like to offer an analogy between the indices used in our study and those used in clinical medicine for evaluating the performance of diagnostic tests. If we consider M. edulis as a “healthy” mussel and M. trossulus as a “sick” mussel (which is not so far-fetching considering the threat presented by M. trossulus to the Scottish aquaculture [19]), then our terms have the following medical equivalents [24]: Ptros is prevalence, P(T|tros) is sensitivity, P(E|edu) is specificity, P(tros|T) is positive predictive value and P(edu|E) is negative predictive value of the morphotype test.

As with clinical tests where disease markers are not 100% sensitive, the positive and negative predictive values will depend on the prevalence, i.e. the frequency of the species (or disease) of interest in the material [24]. With increasing Ptros, P(tros|T) will gradually increase from 0 in pure populations of M. edulis to 1 in pure populations of M. trossulus, while P(edu|E) will show an opposite relationship. For the test to be meaningful, predictive values should be >0.5 since a predictive value of 0.5 indicates a random association between the genotype and the morphotype. Assuming that sensitivity and specificity do not depend on the prevalence (though this assumption may be violated, see below), predictive values could be directly predicted basing on the Ptros in a sample and the known sensitivity and specificity using the formulas: (1) (2)

The prevalence (Ptros) in turn could be predicted based on P(T|edu), P(T|tros) and PT in a sample: (3)

Statistical analyses

After the examination of the associations between morphotypes and individual q-values within sample sets representing different contact zones, the following analyses were made. Firstly, we studied variation of PT, P(T|tros), P(T|edu), P(tros|T), P(edu|E) as functions of Ptros within and between sample sets representing A) the White Sea (sample set WS) and the Barents sea coasts of the Kola Peninsula and saline (set BH) and brackish (set BL) water localities in the Barents Sea (Section “Associations between morphotypes and genotypes around the Kola Peninsula”), B) different geographical contact zones between species. Whenever possible, formulas describing empirical relationships between Ptros and PT and between positive (P(tros|T)) and negative (P(edu|E)) predictive values and Ptros were derived on the basis of regression analysis (Section “Associations between morphotypes and genotypes around the Atlantic”). Secondly, we analyzed genotype-specific associations between morphotypes and the shell size in order to verify the hypothesis that morphological variation under consideration is not related to mussel size (Section “Associations between morphotypes and shell size”). Finally, we tested how well Ptros, P(edu|E) and P(tros|T) could be predicted using formulas Eqs 1–3 and the data on the morphotype proportions among species (P(T|tros), P(T|edu)) in a few (minimum two, see below) genotyped samples. We concede that the assumption that sensitivity and specificity do not depend on the prevalence can be violated in the morphotype test, as it often is in clinical tests [38]. Therefore we focused on finding out which samples were better suited for prediction on the basis of Eqs 1–3: the most mixed samples (Ptros~0.5) or the combination of the two most pure samples of each species. The samples identified in this way as best suited for prediction can be used as “calibrating” ones (Section “Prediction of taxonomic structure of populations and predictive values of the morphotype test based on probability calculators”).

All statistical analyses were performed with functions of R3.6.1 statistical programming language [R_2019]. We used generalized linear (mixed) models, GL(M)Ms, with binomial distribution and a logit link-function. All GLM models were constructed with glm() function from the package “stats” [R_2019] whereas GLMM were fitted with glmer() function from the package “lme4” [39]. The validity of each model was checked by visual analysis of residual plots and the assessment of the overdispersion presence.

The goodness of fit for the models was assessed by means of pseudo-R² [40] using the function r.squaredGLMM() from the package “MuMIn” [41]. To assess the role of random factors in GLMM, we compared marginal and conditional pseudoR² [40]. After the model parameters were estimated, we visualized them by means of regression lines with corresponding 95% confidence intervals.

Associations between morphotypes and genotypes around the Kola Peninsula. The following three regression models were fitted for the data:

1. Model 1: Morphotype proportions (PT) as a function of taxonomic structure of mussel populations (Ptros). All mussels with a T-morphotype were coded as 1 and all mussels with an E-morphotype were coded as 0. These data were used as a dependent variable, which was regressed against Ptros (continuous predictor) and Set (discrete predictor with three levels) and interaction between them.
2. Model 2: Morphotype proportions among species (P(T|tros), P(T|edu)) as a function of taxonomic structure of populations (Ptros). The dependent variable was coded as in Model 1 and modeled as a function of Ptros, Set, Species (a discrete predictor with two levels) and interaction between terms. The sample was included into the model as a random factor influencing the model intercept.
3. Model 3: Correctness of species identification (P(tros|T) and P(edu|E)) as a function of taxonomic structure of populations. The dependent variable was coded as 1 if M. trossulus was represented by a T-morphotype or M. edulis was represented by an E-morphotype and as 0 in the other cases. The set of predictors for the model was as follows: Ptros, Morphotype (discrete predictor with two levels), Set and interaction between terms. The sample was included into the model as a random factor influencing the model intercept.
   To check whether it is possible to pool some of the geographical sets to construct a more general model without losing information, we constructed three complex data sets with different pairing combinations of WS, BL and BH: (WSBL) and BH; (WSBH) and BL; (BLBH) and WS. We did not consider a full combination of sets since in such a case the factor “Set” would be discarded from the model. We applied the structure of Model 3 to these new recombined datasets. Then we compared AICs of these new models with AIC of Model 3 based on non-pooled data. If AIC of a new model was less than the AIC of the initial one, we considered this as a basis for pooling of the corresponding sample sets.
   Associations between morphotypes and genotypes around the Atlantic. Five sample sets were considered, representing the Gulf of Maine (GOM), the Baltic Sea (BALT), western Norway (NORW), saline Barents Sea (BH) and the White Sea combined with the brackish Barents Sea (WSBL, sets WS and BL were pooled since there pooling did not affect the inference, see Results). Scotland (SCOT) was not included in regression analyses because it was represented by only two samples. Three models were constructed:
4. Model 4. Taxonomic structure (Ptros) as a function of morphotype frequencies in populations (PT). The dependent variable was coded as in Model 1 and modeled as a function of PT (continuous predictor), Set and interaction between them. We modeled Ptros vs. PT but not vice versa, as in the previous analysis, in order to use this model as a reference for the “Ptros by PT calculator” (see below).
5. Model 5. Morphotype proportions among species (P(T|tros), P(T|edu)) as a function of taxonomic structure of populations (Ptros). The model was constructed analogously to Model 2.
6. Model 6. Correctness of species identification (P(tros|T) and P(edu|E)) as a function of taxonomic structure of populations (Ptros). The model was constructed analogously to Model 3.

Associations between morphotypes and shell size. To check the possible association of morphotypes with size we undertook the following two analyses. Firstly, we constructed a set logistic regression models for each available species-specific genotype (i.e. M. edulis or M. trossulus) from each sample. The probability of the presence of the T-morphotype was a dependent variable and mussel size was a predictor in these models. Only cases where slope-terms of the models were statistically significant (p < 0.05) after Hochberg’s correction for multiple testing [42] were considered. Secondly, we checked the presence of any patterns in residuals from Model 6 (i.e. the main model designed to predict the probability of correct identification of an individual mussel by its morphotype) as a function of mussel size.

Geographical variation in mussel morphotypes

The studied binary morphological character was previously defined as the “presence/absence of a distinct uninterrupted dark prismatic strip under the ligament”, based on material from the White Sea only [16, 22]. In the broader material of this study, encompassing different geographical zones, the E-morphotypes looked the same everywhere and conformed to the previous description: the strip was absent, and the nacreous layer totally or partially covered the space under the ligament nympha (S1A and S1C Fig). However, T-morphotypes showed some variation previously unrecorded in the White Sea. Firstly, most of the populations contained, though rarely, shells in which the nacreous-free strip of the prismatic layer was quite narrow and looked like a stria rather than a strip. Secondly, in all T-morphotypes from the Gulf of Maine and in rare T-morphotypes from the other populations the strip was not dark but pale, as the prismatic layer itself. Such T-morphotypes were difficult to notice by an unaided eye, but could be unambiguously identified with a dissecting microscope, by the presence of a pronounced scar defining the boundary of the nacreous layer under the ligament nympha (S1 Fig).

Therefore, we amend the description of the T/E morphotype character as follows: the presence/absence of an uninterrupted strip of the prismatic layer under the ligament nympha clearly recognizable by a scar separating the strip from the nacreous layer of the rest of the shell. This description was applicable to all the mussel samples examined in this study.

Variation in morphotype frequencies between M. edulis and M. trossulus within and between contact zones revealed in the study is illustrated in Fig 1, where the estimates of P(T|edu) and P(T|tros) (i.e. the proportions of T-morphotypes among M. edulis and M. trossulus, respectively) in pooled samples from different sets are provided. P(T|edu) was 0.53 in the saline Barents Sea (BH) and less than 10% in all the other sets. In its turn, P(T|tros) was 0.17 in BALT, 0.42 in NORW, 0.49 in the GOM and more than 0.75 in WSBL and SCOT. P(T|tros) estimates in Norway and the Gulf of Maine were much affected by the few outlier samples (see below). If we discard these samples, P(T|tros) will make up 0.54 in Norway and 0.71 in the Gulf of Maine.

Fig 1 also shows the morphotype frequencies in putatively pure populations of species sampled at a distance from the contact zones. Within the ancestral range of M. trossulus in the Pacific, the populations were nearly monomorphic for the T-morphotype. In the Passamaquoddy Bay P(T|tros) was 0.81, i.e. close to that in most of M. trossulus populations in the Gulf of Maine. All reference M. edulis populations from temperate areas (Long Island Sound and Cape Cod in western Atlantic, Northern and Norwegian Seas in Europe) were nearly monomorphic for the E-morphotype. At the northeast extreme of the species range in eastern Atlantic, in the southwestern Barents Sea, P(T|edu) varied considerably between the samples, in particular between the samples from brackish (range 0–3%) and saline (0.35–0.70%) localities (see S2 Table), as it did along the Barents sea coast of the Kola Peninsula. Increased P(T|edu) was also recorded in two northernmost samples from western Atlantic, Greenland (0.66) and the Gulf of Saint Lawrence (0.73).

Genotypic structure of samples

Distributions of individual q-values in the samples ordinated by the proportion of M. trossulus (Ptros) and the frequency distributions of q-values at 10% intervals (E- and T-morphotypes are indicated) in different contact zones are provided in the S2 Fig and Fig 2, respectively. Distributions were pronouncedly bimodal in GOM, SCOT, WS, BL and BH, with most of the mussels having low (q<0.2) or high (q>0.8) q-values (putative purebreds of M. edulis and M. trossulus, respectively) and less than 15% having intermediate values (putative hybrids). However, the distributions were more flattened in NORW and BALT (about 30% and 40% of intermediates, respectively). PT positively correlated with q-values in each set (Fig 2, S2A Fig). Putative hybrids of T-morphotypes tended to be more common in M. trossulus-dominated populations (high Ptros) than in M. edulis-dominated ones (S2B Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Frequency distributions of individual q-values in pooled samples from contact zones between M. edulis and M. trossulus.

Numbers of individuals are plotted on the ordinates, with q-values at 10% intervals as abscises. Red and blue bars indicate T- and E-morphotypes, correspondingly.

https://doi.org/10.1371/journal.pone.0249587.g002

Relationships between PT estimates in the subsamples of mussels with q<0.2 (putative purebreds of M. edulis) and q<0.5 (genotypes dominated by M. edulis genes), as well as between subsamples with q>0.8 and q>0.5 (putative purebreds of M. trossulus and genotypes dominated by its genes) in the samples were close to 1:1 (S3 Fig).

Based on these results, two caveats should be made before embarking upon the analyses based on comparison of “M. trossulus” (q>0.5) and “M. edulis” (q<0.5). (1) These categories should be interpreted as admixed genotypes dominated by genes of one of the species in NORW and BALT and as purebreds in the other contact zones. (2) Morphotype frequencies in hybrids are intermediate between those in parental species, yet in not considering hybrids as a separate class we only slightly bias interspecific differences.

Associations between morphotypes and genotypes around the Kola Peninsula

Variation patterns of PT, P(T|tros), P(T|edu), P(tros|T), P(edu|E) as functions of Ptros in samples from the White Sea (WS), the brackish Barents Sea (BL) and the saline Barents Sea (BH) are visualized in Fig 3. The results of the regression analysis are summarized in S3 Table.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Variation of PT, P(T|tros), P(T|edu), P(tros|T), P(edu|E) as functions of Ptros in the White Sea (WS), brackish Barents Sea (BL) and saline Barents Sea (BH).

Points–empirical estimates, their size is proportional to sample size (see S1 Table). Lines–regression model predictions, grey filling– 95% confidence intervals of regressions. (A) Proportions of T-morphotypes (PT) (Model 1). (B). Proportions of T-morphotypes among M. trossulus (P(T|tros), filled points) and M. edulis (P(T|edu), empty points) (Model 2). (C) Frequencies of M. trossulus among T-morphotypes (P(tros|T), filled points) and of M. edulis among E-morphotypes (P(edu|E), empty points) (Model 4). Vertical lines on B and C connect subsamples of M. trossulus and M. edulis from the same samples.

https://doi.org/10.1371/journal.pone.0249587.g003

A significant positive association between the proportions of M. trossulus (Ptros) and the proportions of T-morphotypes (PT) in samples was revealed for all the three sample sets (Model 1, S3 Table, Fig 3). For WS and BL, the data points were generally scattered around the Y = X line, while the regression lines approached it closely, indicating a high proportionality between Ptros and PT and an almost straightforward relationship between these values. For BH, the data points were scattered above the Y = X line and the regression line lay higher than the regression lines constructed for WS and BL. This means that in samples with a similar taxonomic structure, the frequencies of T-morphotypes were always higher in the saline localities in the Barents Sea than in the White Sea and the brackish localities in the Barents Sea.

The analysis of the frequencies of T-morphotypes in subsamples of M. edulis (P(T|edu)) and M. trossulus (P(T|tros)) against proportions of M. trossulus in samples (Ptros) revealed the following patterns (Model 2, S3 Table, Fig 3). There was a universal tendency towards a higher frequency of T-morphotypes among M. trossulus than among M. edulis. This tendency was quite strong in WS and BL (expected differences in morphotype frequencies between species about 0.65 for Ptros = 0.5). In BH it was rather weak (expected differences 0.18 for Ptros = 0.5) due to an increased P(T|edu) but significant (confidential intervals for Ptros = 0.5 did not overlap, Fig 2). A positive correlation of P(T|tros) and P(T|edu) with Ptros was found in all the three subsets. This means that with the increasing contribution of M. trossulus to the samples the frequencies of T-morphotypes increased both among M. edulis and among M. trossulus.

The probability of correct identification of M. trossulus by the T-morphotype (the frequency of M. trossulus among T-morphotypes (P(tros|T)) expectedly increased with the increasing Ptros, while the probability of correct identification of M. edulis by the E-morphotype (P(edu|E)) demonstrated an opposite pattern (Model 3, S3 Table, Fig 3). In the M. trossulus-dominated populations, P(tros|T) tended to one (any mussel with a T-morphotype is 100% M. trossulus), while P(edu|E) tended to zero (any mussel with an E-morphotype is 100% M. trossulus), and vice versa. In the well-mixed samples (Ptros = 0.5) the predictive values for both species were about 0.75–0.85 in WS and BL but only 0.60–0.70 in BH (Fig 3). It means that the morphotype test has a much lower predictive value in the saline Barents Sea than in the brackish Barents Sea and in the White Sea (the predictive value of 0.5 means a random association between the genotype and the morphotype). It is evident from Fig 3 that a low predictive value of the test in BH is mainly due to a generally high frequencies of T-morphotypes in M. edulis P(T|edu). The statistical analysis indicates that both P(tros|T) and P(edu|E) predicted by the model were smaller in BH than in WS and BL.

For each of the GLMM models considered (Model 2 and 3), marginal and conditional pseudoR² were close to each other (S3 Table). This indicates that the role of the random factor (Sample) as regulator of models was weak, i.e. the reproducibility of the results in different populations was satisfactory.

In comparisons between sets, the regression coefficients did not differ statistically for WS and BL, while BH was always different from WS (S3 Table). To assess the possibility of pooling the data sets, we compared the AIC of Model 3 (AIC = 2175.1) with AICs of three other models based on differently pooled WS, BL and BH sets. The model based on pooled WS and BL (WSBL) and BH showed the lowest AIC (AIC = 2172.7). Therefore, in the following analyses we will consider two sets, WSBL and BH.

Associations between morphotypes and genotypes around the Atlantic

The patterns of Ptros variation against PT and the patterns of P(T|tros), P(T|edu), P(tros|T) and P(edu|E) variation against Ptros in samples from different geographical zones are visualized in Fig 4. The results of the regression analysis are summarized in S3 Table. The Scottish material was not included in the regression analyses. Re-analyses of the data from the White and the Barents Sea (WSBL and BH sets) together with the data from other regions revealed the same patterns as those described above. Again, in all the cases when mixed models were used (Model 5, Model 6, S3 Table) the marginal and conditional pseudoR2 were close to each other (S3 Table) indicating a weak role of the random factor (Set) as regulator of models, i.e. a satisfactory reproducibility of the results from population to population in all the regions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Predictive power of the morphotype test in different contact zones.

(A) Dependence of proportion of M. trossulus (Ptros) on proportion of T-morphotypes (PT). Dotted lines are empirical regressions (Model 4). Solid gray lines–predictions of “Ptros by PT calculator” (Eq 3). Solid black lines represent Y = X dependence. (B) Probability to find a mussel with a T-morphotype among M. edulis (P(T|edu)) (empty points), and M. trossulus (P(T|tros)) (filled points) as a function of Ptros. Lines are empirical regressions (Model 5). (C) Probability of correct species identification by the morphotype test: M. trossulus by T-morphotype, P(tros|T) (filled points) and M. edulis by E-morphotype, P(edu|E) (empty points) as a function of Ptros. Dotted lines are empirical regressions (Model 6). Sold lines–predictions of “genotype by morphotype calculator” for M. trossulus (Eq 1, red line) and M. edulis (Eq 2, blue line). On each graph, dots represent the observed proportions in samples, and shaded areas around regression lines– 95% CI of regressions.

https://doi.org/10.1371/journal.pone.0249587.g004

The proportion of M. trossulus in samples (Ptros) was positively correlated with the proportion of T-morphotypes (PT) in the other sets, as it did in the samples from the White and the Barents Sea. This tendency was significant for all the sets (Fig 4; Model 4, S3 Table). Otherwise, the patterns of variation were different for different sets. For GOM, the regression line stretched above the Y = X line but close to it, indicating the proportionality between PT and Ptros. For BALT, the regression slope was very steep, and the regression line rapidly diverged from the Y = X line. This was due to the fact that the PT range in BALT was, unlike the situation in the other sets, very narrow (0–0.4) as compared with the Ptros range (~0–1), and the small surplus of T-morphotypes in the samples was accompanied by a strong increase in the M. trossulus prevalence. A similar tendency was observed in the scanty material from NORW. Both SCOT samples fell on the Y = X line. Noteworthy are a few “outlier” samples from GOM and NORW, in which PT was close to zero but Ptros was high.

While frequencies of T-morphotypes in M. edulis (P(T|edu)) were low everywhere but in BH, frequencies of T-morphotypes in M. trossulus (P(T|tros)) demonstrated a strong variation among sets and a noticeable variation within some sets (Fig 4; Model 5; S3 Table). Similarly to WSBL, most M. trossulus had T-morphotypes in GOM and SCOT but not in BALT and NORW. For Ptros = 0.5, expected differences in the morphotype frequencies between the species were about 0.44 for GOM, 0.06 for BALT and 0.24 for NORW. A significant positive dependence of the frequencies of T-morphotype on Ptros among conspecific genotypes, which was so prominent in the White and the Barents Sea, was recorded elsewhere only in BALT for P(T|tros) (S3 Table).

The pattern of dependence of P(tros|T) and P(edu|E) on Ptros in GOM, BALT and NORW (Model 6. Fig 4, S3 Table) was the same as in the samples from the Kola Peninsula (Model 3. Fig 3, S3 Table): P(tros|T) increased with the increasing Ptros, while P(edu|E) showed an opposite tendency. To simplify and formalize the comparison, we provide the predictions of Model 6 for equally mixed populations (Ptros = 0.5) together with their 95% confidence intervals in Table 1, where actual proportions of M. trossulus among T-morphotypes (P(T|tros)) and M. edulis among E-morphotypes (P(T|edu)) in pooled samples from the respected sets are also provided.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Proportions of M. trossulus among T-morphotypes (P(tros|T)) and proportions of M. edulis among E-morphotypes (P(edu|E)) in pooled samples (direct count) and in equally mixed samples (predictions by the regression Model 6) in different sample sets.

https://doi.org/10.1371/journal.pone.0249587.t001

For equally mixed populations the predictive values of P(edu|E) in BALT did not differ significantly from 0.5, which corresponds to an equal probability of correct and incorrect identification. At the same time, the probabilities of correct identification of M. trossulus by the T-morphotype in GOM, BALT and NORW were quite high (for the range of Ptros≥0.5). In general, the highest predictive values for both species were revealed in WSBL.

Using the coefficients of the regression models Model 4 and Model 6 (S3 Table), we constructed a set of formulas predicting the taxonomic structure (Ptros) and the probability of correct species identification (P(tros|T), P(edu|E)) using the morphotype test (Table 2). These formulas were further used for the comparison of predictions made with these regression models and the predictions proposed by Eqs 1, 2 and 3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Formulas used for taxonomic and individual assignment using morphotype tests in different sample sets accordingly to the regression model coefficients represented in S3 Table.

https://doi.org/10.1371/journal.pone.0249587.t002

Associations between morphotypes and shell size

There was no clear statistical relationship between the size and the morphotype of conspecific mussels. At the level of individual samples, the probability of finding a T-morphotype increased with the mussel size (a positive slope-term of the regression) in 16 out of 34 informative comparisons (when species-specific genotypes were both present and polymorphic for morphotypes) for M. edulis and in 17 out of 43 comparisons for M. trossulus. The slope-terms of the regression models were individually significant (p<0.05) in four cases for M. edulis and in four cases for M. trossulus, but only in one case when the correction for multiple testing was applied (sample Berg05, see S4 Table). We also checked for the presence of any patterns in residuals from Model 6 as a function of mussel size but none was found.

Prediction of taxonomic structure of populations and predictive values of the morphotype test based on probability calculators.

We applied Eqs 1 and 2 (“genotype by morphotype calculator”) and Eq 3 (“Ptros by PT calculator”) using as an input for assessment of equations parameters (P(T|tros), P(T|edu)) the data on all possible pairs of samples from WSBL and compared the values predicted by these equations with those predicted by regression Models 6 and 4, respectively (Table 2).

Fig 5 illustrates the goodness of correspondence of the two predictions depending on the genetic constitution of the paired samples as expressed by the Delta index. The best predictions of Ptros were obtained when the most dissimilar samples consisting of nearly pure M. edulis and M. trossulus (Delta >0.75) were taken for Eq 3 calibration. The best predictions of P(edu|E) and P(tros|T) values were obtained when the most mixed samples (Ptros of both samples close to 0.5; range of Delta 0.25–0.5) were taken for Eqs 1 and 2 calibration.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Correspondence between “Ptros by PT calculator” (Eq 3, left graph) and “genotype by morphotype calculator” predictions (Eqs 1 and 2, right graph) and regression Model 6 and Model 4, respectively.

Each point corresponds to a unique pair combination of samples from WSBL set. OX axis reflects dissimilarity of genetic structure in each pair (Delta) (for pure conspecific samples Delta takes a value of zero, for equally mixed samples– 0.5, for two pure heterospecific samples– 1). OY: goodness of correspondence between assessment of predictive values by equations and regression models.

https://doi.org/10.1371/journal.pone.0249587.g005

Therefore, in order to predict Ptros using “Ptros by PT calculator” one should use the most dissimilar samples to assess P(T|edu) and P(T|tros) as calculator parameters. In order to predict P(edu|E) and P(tros|T) using “genotype by morphotype calculator” one should assess the parameters using the most mixed samples. However, the Ptros as input in Eqs 1 and 2 should be calculated as in the previous case i.e. using the most dissimilar samples and Eq 3. To illustrate the approach (see Fig 4) for WSBL, BH, GOM and BALT we used pooled sets of samples with Ptros <0.1 and >0.8 to calibrate “Ptros by PT calculator” and 0.45 < Ptros < 0.65 to calibrate “genotype by morphotype calculator” as described above (samples included are indicated in S1 Table). We used pooled but not individual samples to avoid basements due to small sample size. The given ranges of Ptros were used because of the lack of M. trossulus-dominated samples in most sets. For NORW and SCOT we pooled all the samples because of the lack of data.

Visual inspection of Fig 4A revealed a nearly ideal correspondence between regression lines and predictions of “Ptros by PT calculator” in the case of WSBS and GOM. In the case of SCOT, where only two samples were available, the line derived from Eq 3 approached the Y = X line. A rather close though not ideal correspondence was observed in the case of BALT, deviation being due to a very high slope term of the regression. Ptros was slightly underestimated by the calculator in this case. The worst correspondence between Eq 3 and Model 4 was observed in the case of NORW and BH. In BH Ptros was severely overestimated by the calculator, which was opposite to the situation in BALT. In NORW both regression and predictions of calculator were severely affected by the outlier sample.

As for the “genotype by morphotype calculator”, the predictions generally were in good correspondence with the regression lines (calculator’s lines were within 95% CI of regressions). Deviations were observed for P(tros|T) predictions in WSBL for Ptros<0.25 and P(edu|E) in GOM for Ptros>0.6 i.e. in the Ptros ranges corresponding to a small prevalence of the species.

An exercise with the “lazy Ptros by PT calculator”, in which the highest and the lowest PT in samples from regional sets are used as P(T|tros) and P(T|edu) parameters of Eq 3, had the following results (S4 Fig). In WS, BL and GOM correspondence between the observed and the predicted Ptros values in samples was generally good. In BH, Ptros was slightly overestimated by the calculator due to the absence of pure M. trossulus samples in the data and the formal choice of a numerically small (N = 22; see S1 Table) sample with the highest PT but not the highest Ptros as the “calibrating” one. For BALT and NORW discrepancies were much stronger, the reasons being the same as in case of “Ptros by PT calculator” (see above).

Variation of morphotype frequencies among conspecific populations

Some variation in the morphotype frequencies was observed among putatively pure conspecific populations sampled at a distance from the contact zones. Samples of pure M. edulis from the temperate seas (i.e. all except those from the eastern Barents Sea and Greenland) were nearly monomorphic for the E-morphotype, while the northern samples were more polymorphic and diverse. In turn, the reference populations of M. trossulus from the northwestern and northeastern Pacific (Washington) were nearly monomorphic for the T-morphotype. Nevertheless we cannot necessarily exclude the possibility of geographic variation in M. trossulus in its ancestral Pacific range or confirm that the T-morphotype is the “ancestral” state for this species. Zolotarev [47] identified morphotypes in small samples of genotyped mussels (from [7]) and found elevated frequencies of the E-morphotype in M. trossulus from Oregon (northeastern Pacific).

In M. trossulus the morphotype frequencies varied between the contact zones, and elevated frequencies of E-morphotypes were found in Norway and, especially, in the Baltic Sea. The variation within contact zones was mostly due to the few “outlier” samples from the Gulf of Maine and Norway. On the contrary, M. edulis showed little variation between zones, the T-morphotype being universally rare. In a notable exception the T-morphotype frequency was clearly elevated (up to 40%) in samples from saline localities (> 30 ppt) in Kola Bay and surroundings. Similar salinity-related variation was found in M. edulis from the more eastern areas of the Barents Sea, at some distance from the contact zone between these species along the Kola Peninsula coast.

Finally, an analysis of the abundant material from the White and the Barents Sea demonstrated how the morphotype frequencies varied with the taxonomic composition of the mixed populations. The frequencies of the T-morphotype increased both among M. edulis and among M. trossulus genotypes with the increasing prevalence of M. trossulus in the samples.

Unusual features of M. trossulus from Norway and the Baltic Sea

M. trossulus from the Baltic Sea and Norway were characterized by extremely high frequencies of the E-morphotype. Two hypotheses, which are not mutually exclusive, can be offered to explain this phenomenon. One hypothesis links the morphotypes to species-specific genes that can introgress between species as a result of extensive hybridization and backcrossing. The Baltic M. trossulus is stronger introgressed by M. edulis genes than any other Atlantic population [3, 33]. Due to its mixed genetic nature, the Baltic mussel is sometimes considered as a unique M. edulis x M. trossulus hybrid swarm, which is fundamentally different from the “oceanic” M. trossulus [3]. While the genetic data from Norway are limited, it is evident that the local Norwegian M. trossulus populations can be strongly introgressed by M. edulis genes too [48].

In the second hypothesis, the frequency of the T-morphotype in M. trossulus is reduced under the influence of some environmental factors, both micro- and macro-geographical. We suspect that the nearly zero frequencies of the T-morphotype in the “outlier” samples (one from Norway, almost from the same place as the other Bergen samples, and two from Cobscook Bay in the Gulf of Maine [CBCP, CBSC in S1 Table]) could be explained by the impact of some cryptic local factors, though a more prosaic explanation such as the mislabeling of mussels in the collections cannot be entirely ruled out.

Salinity-related variation in M. edulis

While local factors putatively affecting morphotype frequencies in M. trossulus remained cryptic, in the Barents Sea we managed to identify one such factor governing morphotype frequencies in M. edulis: salinity or a factor/factors linked to salinity. The eastern part of the Barents Sea, where this variation was evident, is also the coldest. The border between the more temperate populations of M. edulis with “normal” (high) frequencies of the E-morphotype and the more Arctic populations with lower frequencies of the E-morphotype in oceanic habitats runs somewhere between North Cape and the Kola Bay (Fig 1). This area has mean annual, summer and winter sea surface temperatures of about 6°C, 9°C and 4°C, respectively (http://esimo.oceanography.ru/).

The functional significance of the morphological character underlying the E- and the T- morphotype (the presence/absence of the nacreous layer under the ligament) is unclear. However, we suspect that the morphotypes might differ in conspecifics by the degree of development of the nacreous layer itself and thus in the thickness and strength of the shell. The nacreous shell layer is mechanically the strongest [49].

Can we expect the shell thickness and structure to differ in mussels from saline (oceanic) and brackish (estuarine) environments in the Arctic? Apart from the low temperatures, the Arctic Sea is characterized by a reduced concentration of calcium carbonates [50] and, seasonally, by low concentrations of planktonic algae, which the mussels feed on [51]. Estuarine habitats are generally characterized by the lowest saturation of carbonates but the highest concentrations of food (seston), which is due to the riverine discharge [52]. This is exemplified by the highest biomasses of Mytilus in estuaries in the Barents Sea [53] and elsewhere [12]. Mussels need both calcium carbonates and energy for shell growth and maintenance. In estuaries, the nacreous layer of the mussel shell is prone to dissolution and corrosion [54] but the mussels can still keep their shells strong if the food is sufficient [52, 54]. If the food is limited, the energy is likely to be allocated to the maintenance of the somatic mass rather than the conservation of the shell ([54] and references therein).

Our hypothesis explaining the assumed differences in the degree of the nacreous layer development between M. edulis from the brackish and the saline localities in the Arctic is that in the estuaries the mussels allocate more energy for shell maintenance thus keeping their nacreous layer thick while in less corrosive but more famished oceanic habitats they allocate more energy for somatic growth keeping their nacreous layer thin. As a result, the majority of M. edulis from the saline localities has the undeveloped nacreous layer. It is noteworthy that in the areas where M. edulis demonstrated salinity-related variation, the morphotype frequencies in M. trossulus varied negligibly. This could be attributed to a generally lower shell plasticity in “oceanic” (non-Baltic) M. trossulus than in M. edulis in response to the environmental stressors ([55], see [22] for more discussion).

Noteworthy, reduced frequencies of the E-morphotype were revealed not only in the eastern Barents Sea but also in northernmost populations of M. edulis from Greenland and the Gulf of Saint Lawrence in western Atlantic (Fig 1). This indicates that this is an Arctic phenomenon. Unfortunately, the salinity in the sampling localities in the latter two areas is unknown.

Variation with the taxonomic structure

A positive correlation of the T-morphotype frequencies both in M. edulis and M. trossulus with the prevalence of M. trossulus in the representative data sets from the White and the Barents Sea was to be expected, bearing in mind that M. edulis and M. trossulus genotypes are defined by the dominance of the conspecific genes in multilocus genotypes. Hence both genotypes included purebreds as well as some hybrids that are intermediate in morphotype frequencies between purebred M. edulis and M. trossulus but usually closer to species dominating the population [[16]; this study]. This means that in our analyses “M. edulis genotypes” from M. trossulus-dominated populations included mostly hybrids with an increased frequency of the T-morphotype as compared to the “M. edulis genotypes” in M. edulis-dominated populations. In turn, “M. trossulus genotypes” from M. edulis-dominated populations included mostly hybrids with a decreased frequency of the T-morphotype as compared to such genotypes in M. trossulus-dominated populations. This is the cause of the observed unidirectional variation in the morphotype frequencies among M. edulis and M. trossulus genotypes with the changing taxonomic structure of populations. To note, the variation of sensitivity and specificity of clinical diagnostic tests with the changing disease prevalence is often observed [56]. For instance, a patient population with a higher disease prevalence may include more severely diseased patients, and the test would consequently perform better [56].

Applications of the mussel morphotype test

The morphotype test can be universally applied as an alternative to genotyping in three fields. Firstly, it can be used for monitoring the species composition of commercial and wild populations, in particular those used in the “mussel watch” contaminant monitoring programs, because deviations of the morphotype frequencies may be a warning sign of the taxonomic change. Secondly, it may prove useful for mapping the species distributions. Detailed mapping is likely to require a great number of samples because the distribution of the species in contact zones is usually highly mosaic (see [16] and references therein). Thirdly, the morphotype test can be used when only dead mussel shells are available, e.g. for interpretations of the taxonomic structure of natural history collections or samples of dead shells left behind by some mussel predators.

Uses and abuses of single marker taxonomic tests

Traditional species identification relies on diagnostic (fixed) morphologic traits of the organism, usually included in the diagnosis. In the terms of the probability theory, it means that the probability of an individual with a species-specific diagnostic marker being a representative of the species in question is equal to one: P(species|trait) = 1. However, in practice the probability can be lower for two reasons. First, because of scoring errors related to the researcher’s skills or the defective condition of the specimen. Second, for the ambiguity in the diagnosticity of a trait. It is generally impossible to determine whether diagnostic characters are indeed fixed if the sample size is finite [57]. Hence, in practice, for diagnostic markers P(species|trait) ≤ 1.

Some taxa, however, lack diagnostic characters and have to be identified on the basis of semi-diagnostic ones. This is the case with the blue mussels [7]. In case of semi-diagnostic traits, the researchers do not identify the species of a given individual but assess the probability of its assignment to one or another species. For these traits, P(species|trait) < 1. Similarly, dealing with population assessment we assess the probabilities of finding the representatives of one or another species in a sample but not the true proportion. The most critical point is that P(species|trait) is not constant but varies, yet in predictable manner, with the prevalence of a species in a range [0;1].

A correct application of tests based on semi-diagnostic markers, such as clinical diagnostic tests, ultimately requires a “reference standard” used for verification of the index test results [24]. In our case study of the blue mussels, we used as references the groups of multilocus genotypes (from 4 to 171 645 loci depending on the geographical sample set) defined by the dominance of alleles characteristic of one or the other species. These groups did not represent true species. They included hybrids, some of which (e.g. first- and second generation hybrids) were assigned into groups randomly. To note, multilocus genotyping is seldom employed for identification of cryptic mussel species. Most studies rely on singular or few “standard” diagnostic PCR-based markers, usually nuclear Me15/16 and ITS and mitochondrial COI or 16S markers [11]. Offering the morphotype test as a rough but cost-efficient alternative to genotyping, we have to assess its reliability as compared to single- and few locus tests. It has been long known that the efficiency of “diagnostic” markers for discrimination between M. edulis and M. trossulus is different in contact zones in western Atlantic (i.e. the Gulf of Maine) and the Baltic Sea [1]. In the Northwest Atlantic the species are nearly fixed for alternative alleles at Me15/16, ITS and mitochondrial markers, while in the Baltic Sea intraspecific differences at these loci are 20%, 32% and 0%, respectively, due to a introgression of M. edulis genes into the local M. trossulus genome [1]. For comparison, the differences in morphotype frequencies between species in the Gulf of Maine and the Baltic Sea are 44% and 15%. As far as we know, the efficiency of taxonomic tests based on singular or a few “standard” loci has not been carefully evaluated for other M. edulis–M. trossulus contact zones, though some attempts have been made (see [3, 58]). The recent population genomic studies of hybridizing Mytilus species indicate that these species can altogether lack fixed genetic differences due to ubiquitous introgression and that loci can introgress in unpredictable manner in different contact zones [33, 59]. On these grounds, the conventional approach to mussel species identification based on singular molecular markers has been criticized [59]. We do not claim that the morphotype test would fare better than most single-locus taxonomic tests in any contact zone between M. edulis and M. trossulus. At the same time, we would like to point out that the performance of these genetic tests has not been evaluated in most contact zones, while that of the morphotype test has been.

A situation when one has to rely on a singular “informal” semi-diagnostic character to distinguish morphologically such old evolutionary lineages as M. edulis and M. trossulus is certainly uncommon in taxonomy. At the same time, it is not unique. There are other taxa, which lack fixed diagnostic morphological characters and are identified by semi-diagnostic traits, individual or complex such as like the coordinates of multifactorial analysis. These taxa include subspecies defined according to the 75% rule [60], cryptic species with statistical differentiation [61] and hybridizing species that secondarily lost fixed differences due to introgressive hybridization [62]. Therefore, we hope that our experience of dealing with a non-fixed taxonomic character would be interesting not only to our colleagues working with blue mussels but also to the researchers who study sympatric taxa with vague morphologies and semi-isolated gene pools.