Chemical patterns of colony membership and mother-offspring similarity in Antarctic fur seals are reproducible

Study site and fieldwork

Chemical samples were taken from six Antarctic fur seal breeding colonies on Bird Island, South Georgia (54°00′S, 38°02′W) during the peak of the 2016 breeding season (November–December; the previous study was conducted during the peak of the 2011 breeding season). A total of 50 mother-offspring pairs (including one pair of twins) were sampled from SSB and FWB as part of annual routine procedures of the long-term monitoring and survey program of the British Antarctic Survey (BAS). Additional samples were collected from a total of 60 pups from four colonies (15 samples each from Johnson Cove, Main Bay, Landing Beach and Natural Arch, Fig. 1). Here, pups were opportunistically sampled from areas of the beach that were easily accessible. Adult females and pups were captured and restrained on land using standard methodology (Gentry & Holt, 1982). Chemical samples were obtained by rubbing the cheek underneath the eye and behind the snout with sterile cotton wool swabs, which were stored individually at −20 °C in glass vials containing approximately 10 mL of 60%/40% (vol/vol) ethanol/water. All of the chemical samples were collected immediately after capture by the same team of experienced field scientists. The samples were frozen at the latest 1 h after collection and were stored for approximately 18 months prior to analysis.

GC-MS profiling and data alignment

We first took two mL of each sample and allowed the ethanol to evaporate at room temperature for a maximum of 12 h before resuspending in two mL dichloromethane (DCM). After a further evaporation step, in which the DCM was reduced to a final volume of approximately 100 µL, the samples were analyzed on a GC with a VF5-MS column (30 m × 0.25 mm inner diameter, 10 m guard column; Agilent Technologies, Santa Clara, CA, USA) connected to a mass spectrometer (GCMS-QP2020, Shimadzu, Kyoto, Japan). One µL of each sample was injected into a deactivated glass-wool-packed liner with an inlet temperature of 225 °C. A split ratio of 3.2 was used and the carrier gas (Helium) flow rate was held constant at 1.2 mL/min. The GC run started with 3 min at 60 °C and then ramped up in increments of 10 °C/min to reach a final temperature of 280 °C, which was maintained for 30 min. Mass spectra were taken in electron ionization mode with five scans per second in full scan mode (50–600 m/z). The resulting GC-MS data were then processed using OpenChrom (Wenig & Odermatt, 2010) for detection and correction of split peaks. Afterwards, we used GCalignR (Ottensmann et al., 2018; R Core Team, 2019) to align the resulting chromatograms by correcting minor shifts in retention times among samples and maximizing the number of shared components.

Data visualization and statistical analysis

Prior to data analyses, we excluded any compounds that were only observed in a single sample. We then used non-metric multidimensional scaling (NMDS) to visualize the chemical data. This approach reduces dimensionality so that each individual data point can be placed in a 2D scatterplot where ranked between-individual distances are preserved and individuals that are chemically more similar are closer together. NMDS was performed on a log(x+1) transformed relative abundance matrix comprising pairwise Bray–Curtis similarity values. We tested for differences among and between a priori defined groups (i.e., the breeding colonies and mother-offspring pairs) using a non-parametric permutational multivariate analysis of variance (PERMANOVA). PERMANOVA tests whether the centroids of pre-defined groups differ statistically for a chosen distance measure. It compares within-group to among-group variance components and assigns statistical significance based on random permutations of objects within groups. Each PERMANOVA was based on 99,999 permutations, although comparable results were also obtained with 9,999, 999 and 99 permutations. To determine whether differences between our pre-defined groups were attributable to compositional differences between groups rather than compositional differences within groups, we used the “betadisper” function in the vegan package in R to analyze the multivariate homogeneity of group dispersions (Oksanen et al., 2019). In addition, we performed pairwise PERMANOVAs for different groups within the model strata based on age and colony and Bonferroni corrected the resulting p-values.

Quantification of the explained variance

To facilitate a comparison of our effect sizes with those reported by Stoffel et al. (2015), we quantified the proportion of the total chemical variance attributable to colony membership and family ID in both studies. The scripts that Stoffel et al. (2015) used to align their data have now been embedded into GCalignR and are therefore consistent between the two studies. As different chemical datasets will have different optimal parameter settings for the alignment algorithm, we did not re-align or adjust the dataframe of Stoffel et al. (2015). Enforcing the same parameter settings as in the current study would almost certainly lead to a loss of data quality and result in artificially reduced effect sizes. To standardize effect size estimates between the studies, both chemical datasets were bootstrapped over individuals to generate 5,000 datasets per study, each comprising 15 mother-offspring pairs from SSB and 15 pairs from FWB (i.e., a total of 60 individuals). PERMANOVA was then implemented separately for each dataset and the resulting R² values were extracted for each of the predefined groups.

Data availability

The raw chemical data generated during this study are available via GitHub and the data of Stoffel et al. (2015) can be downloaded from https://github.com/mastoffel/seal_chemical_fingerprints. All of the code used to analyze the raw data are available as a PDF file written in Rmarkdown (see Supplemental Information). The full documented data analysis pipeline can be downloaded from our GitHub repository at https://github.com/tebbej/SealScent2020/.

Ethical statement

Samples were collected as part of the Polar Science for Planet Earth program of the British Antarctic Survey under the authorization of the Senior Executive and the Environment Officers of the Government of South Georgia and the South Sandwich Islands (permit no. 2016/013). Samples were collected and retained under Scientific Research Permits for the British Antarctic Survey field activities on South Georgia, and in accordance with the Convention on International Trade in Endangered Species of Wild Fauna and Flora. All field procedures were approved by the British Antarctic Survey Animal Welfare and Ethics Review Body (reference no. PEA6).

Reproducibility of chemical patterns

Multivariate statistical analysis of the relative proportions of each substance revealed highly significant differences between animals from SSB and FWB (PERMANOVA, p < 0.0001, Fig. 2A; Table 1A). A highly significant effect of mother-pup pair ID nested within colony (PERMANOVA, p < 0.0001, Fig. 2B; Table 1A) was also found, indicating that mothers and their pups are chemically more similar to one another than expected by chance. A test for multivariate homogeneity of group variances uncovered marginally significant differences among the groups (p = 0.026, Table 1A), which could potentially indicate the involvement of additional explanatory factors that were not accounted for in the model. We therefore investigated further by splitting the chemical data into four groups, corresponding to mothers and pups from SSB and FWB respectively. Performing PERMANOVAs for all possible pairwise combinations of these groups resulted in three important outcomes. First, all of the pairwise PERMANOVAs involving groups of animals from the two different colonies were highly significant after table-wide Bonferroni correction for multiple tests (Table S1). This indicates that colony membership is chemically encoded irrespective of whether individuals are mothers or pups. Second, both of the pairwise PERMANOVAs involving mothers and pups within colonies were non-significant after Bonferroni correction (Table S1). This suggests that mothers and their pups are chemically similar to one another, regardless of the colony in question. Finally, tests for the homogeneity of group variances were not significant for any of the pairwise group comparisons after Bonferroni correction (Table S2). This implies that our results are unlikely to be driven by differences in chemical variance among groups.

As p-values cannot be directly compared between studies with different sample sizes, we used the PERMANOVA framework to estimate the effect sizes of colony membership and mother-offspring similarity in both studies. To facilitate direct comparisons while also incorporating uncertainty due to chemical variation among individuals, both datasets were bootstrapped over individuals as described in the Materials and Methods. We found that effect size estimates for colony membership and mother-offspring similarity (maximum density R² values) differed by only few percent between the two studies (Fig. 3) and consistently fell within the range of 0.08 < R² < 0.15.

Generality of chemical patterns

To investigate whether chemical signatures are colony-specific in general, we analyzed chemical data from pups sampled from a total of six colonies around Bird Island. PERMANOVA uncovered chemical differences not only between SSB and FWB, but also more generally among colonies (Fig. S1). These differences were statistically significant both overall (p < 0.0001, Table 1B) and for the majority of pairwise comparisons after Bonferroni correction (Table 2).

Motivation and study design

A number of factors motivated the current replication attempt. First, the results of Stoffel et al. (2015) were based on a modest sample of Antarctic fur seal mother-offspring pairs sampled in a single season. We therefore wanted to safeguard against type I error while also testing for the repeatability of chemical patterns over time. Second, chance results can become highly influential (Kelly, 2006) and our original study already appears to have motivated comparable investigations in other pinniped species. For example, a recent study of Australian sea lions using a very similar experimental design also reported chemical differences between two breeding colonies, but chemical similarities were not found between mothers and their pups (Wierucka et al., 2019). Although it is not unreasonable to assume that different species might vary in how chemical information is encoded and used in mother-offspring recognition, this point of difference nevertheless encouraged us to revisit our original findings. Finally, being able to confirm and extend our original results strengthens the case for follow-up studies and reduces the risk of time and resources being wasted on chasing up false positives.

Although we acknowledge that no study of a wild population can ever be perfectly replicated (Nakagawa & Parker, 2015; Fidler et al., 2017), we believe that our replication study of chemical patterns in Antarctic fur seals is sufficiently close to that of Stoffel et al. (2015) in terms of both experimental design and implementation to be considered an exact replication. In practice, there were a handful of small differences between the two studies, but these were mainly a consequence of incremental improvements to our methodology and are unlikely to have had a major influence on the final outcome. For example, because replication studies often produce smaller effect sizes than original studies (Simonsohn, 2015; Open Science Collaboration, 2015), we attempted to enlarge our sample size of mother-offspring pairs as far as was practicable. We also improved the standardization and reproducibility of our chemical analysis pipeline by performing peak detection with open source software and by integrating the alignment algorithm of Stoffel et al. (2015) into an R package (Ottensmann et al., 2018). However, these small modifications appear to have been of little consequence as the effect sizes of colony membership and mother-offspring similarity did not differ systematically between the two studies.

Two further methodological differences were beyond our control. First, owing to the fact that the original and replication studies were carried out 5 years apart, the sampling was conducted by different teams of field biologists. However, we used carefully standardized field protocols in order to minimize any inadvertent experimental variation. Second, the GC-MS machine used by Stoffel et al. (2015) was subsequently replaced by a newer and more sensitive model. One might have expected this to result in more chemicals being detected in the replication study, which would be expected to provide greater power to detect chemical patterns. If anything, however, fewer chemicals in total were detected in the current study, possibly because of differences in the concentrations of samples or because we used different peak calling software and manually curated the resulting dataset to remove redundant split peaks. Regardless of the exact explanation, the overall similarity of the results of the two studies suggests that patterns of colony membership and mother-offspring similarity in Antarctic fur seals are robust to these minor sources of experimental variation. This robustness would be expected if chemical patterns are influenced by large numbers of compounds and therefore persist independently of minor methodological differences that may influence which subsets of peaks are detected and retained for analysis.

Replication outcomes

Successful replication can be defined either in the context of statistical significance (Rosenthal, 1991) or on the basis of a comparison of effect sizes (Goodman, Fanelli & Ioannidis, 2016; Piper et al., 2019). We not only tested for significance but also developed an approach based on PERMANOVA to evaluate the effect sizes of colony membership and mother-offspring similarity in both datasets. Specifically, we extracted R² values for the terms in question after bootstrapping both chemical datasets over individuals. This approach controlled for differences in sample size between the two studies while also providing a visual representation of the magnitude of uncertainty associated with the R² estimates. We not only found that the patterns reported by Stoffel et al. (2015) remained highly significant, but also that the effect size estimates of colony membership and mother-offspring similarity in the two studies were more or less similar, varying by at most a few percent. Elsewhere, in a study that attempted to replicate a hundred psychological studies (Open Science Collaboration, 2015), variation in the strength of the original evidence, such as p-values, was more predictive of replication success than other characteristics such as the experience or expertise of the original and replication teams. This is consistent with the outcome of the current replication exercise given the high statistical significance (p < 0.0001) of the patterns originally reported by Stoffel et al. (2015).

Generality of the colony membership pattern

We went a step beyond simply repeating our previous study by investigating whether chemical differences between SSB and FWB are specific to these two colonies, or whether chemical signatures are colony-specific in general. Unfortunately, it was not possible to sample mothers from locations other than SSB and FWB due to the difficulty of capturing adult females farther away from the BAS field station where fieldwork on seals is rarely if ever performed. However, the relative ease of capturing pups enabled us to gather a more representative collection of chemical samples from multiple breeding sites around Bird Island. After controlling for the false discovery rate, statistically significant chemical differences were detected in all but two out of 15 pairwise comparisons between colonies. This suggests not only that chemical patterns of colony membership are repeatable over time, but also that they can be generalized over space. Interestingly, we did not find a clear correspondence between chemical similarity and the geographical proximity of colonies. For example, Freshwater Beach and Main Bay were among the most chemically dissimilar colonies despite being only around 500 m apart, while Johnson Cove and Natural Arch were among the most chemically similar colonies despite being situated at the opposite extremes of Bird Island. The most probable explanation for this pattern is that chemical differences among colonies are predominantly shaped by as yet unknown environmental factors (see below).

Mechanisms encoding chemical information

Relatively little is currently known about the mechanisms by which colony membership and mother-offspring similarity are chemically encoded in Antarctic fur seals. We know that animals from SSB and FWB exhibit chemical differences despite a lack of genetic differentiation (Stoffel et al., 2015), which implies that environmental drivers play an important role. However, it remains unclear exactly what these drivers might be. Food is unlikely to be an important determinant of colony-specific chemical patterns because all of the breeding females around Bird Island feed predominantly on Antarctic krill (Boyd, Staniland & Martin, 2002). The underlying substrate is also relatively homogenous, with the vast majority of animals occupying cobblestone breeding beaches that show little in the way of obvious differences to the human eye. It is therefore more likely that colony-specific chemical phenotypes are influenced by differences in local conditions such as temperature, wind or solar radiation, either directly or via alterations to the skin microbiota (Grosser et al., 2019). A further possibility could be that chemical differences between colonies reflect differences in microbial communities shaped by social stress. For example, stressful conditions such as high densities of conspecifics can suppress microbial diversity (Bailey et al., 2011; Stothart et al., 2016; Noguera et al., 2018; Partrick et al., 2018; Zha et al., 2018). This is consistent with our data, as breeding females on SSB are present at higher density and have chronically elevated levels of the stress hormone cortisol (Meise et al., 2016), while skin microbial diversity is also lower in this colony (Grosser et al., 2019). Investigating the potential linkages between social stress, cortisol, microbial community structure and chemical phenotypes represents a promising avenue for future research.