We draw together stomach contents data primarily collected from the North Atlantic shelf seas, with important contributions from the Baltic, Barents, and Norwegian seas (Figs.  and S1 in the Supplement). These data were sourced from a combination of previously published and unpublished data including DAPSTOM (an Integrated Database and Portal for Fish Stomach Records; Pinnegar, 2019), ICES Year of the Stomach (Daan, 1981; ICES, 1997), the north-east US continental shelf (Smith and Link, 2010), northern Spanish shelf (Arroyo et al., 2017), Gulf of Cadiz (Torres et al., 2013), and Swedish-, Icelandic-, Norwegian-, French- (Cachera et al., 2017; Timmerman et al., 2020; Travers-Trolet, 2017; Verin, 2018) and German-led surveys (e.g. FishNet, https://www.nationalpark-wattenmeer.de/wissensbeitrag/fishnet/, last access: 22 April 2021). We have included stomach contents data from outside the OSPAR area (i.e. north-east US continental shelf and Baltic Sea) to demonstrate the wider applicability of our approach to defining feeding guilds and because these data have been used to classify feeding guilds previously (Garrison and Link, 2000a). The full data collation contains observations from larval to adult predators (i.e. fish whose stomach contents have been sampled, ranging from < 1g to 351 kg), representing 14 196 unique interactions between 227 predator species and 2158 prey taxa (i.e. prey are defined as organisms found in stomach contents; 10.14466/CefasDataHub.149; Thompson et al., 2024). We provide a summary of data sources, spatial and temporal ranges, and sample distributions in Table . All data processing and subsequent analyses were conducted in R version 4.02 (R Core Team, 2020). Predator and prey taxonomy were processed using the taxize package (Chamberlain et al., 2020) and assigned to zooplankton, benthos, fish, nekton, and other functional groups after Webb and Vanhoorne (2020) using the worrms package (Chamberlain, 2019).

Prey count and biomass observations (wet weight in grams) are needed to estimate predator–prey mass ratios (PPMR), but these were available for only 56 % of the stomach contents data. Therefore, to make use of all the data (Fig. ; Table ) when assigning fish to feeding guilds, a linear mixed effect model of predator–prey body-mass scaling was constructed to estimate prey counts or biomass where one or both were available. The model was constructed only using data where the taxonomy for both predator and prey was resolved (i.e. to species and functional group respectively) and individual predator body mass, individual prey body mass, and prey counts were all available. Major axis regression following Brose et al. (2019) was not suitable because we needed to make predictions of individual prey body mass and minimize the squared residuals in the response (Legendre, 1998). The log₁₀-transformed individual prey mass (wet weight in grams) was modelled as the response variable, with log₁₀-transformed predator body mass (wet weight in grams) as a fixed effect. Random intercepts and slopes were included for predator taxa and prey functional group to account for potential variation in their relationships with predator body mass. Random intercepts were fit for datasets which follow different protocols to test for systematic differences in how data were generated (i.e. Spain, USA, and ICES, which each follow their own protocols for measuring prey biomass, and all others,` which represent a mixture of methods from across studies). Random intercepts were also fit for years, sites (Fig. S1; sites based on a 3 by 3 grid across the study region), and the number of stomachs sampled (i.e. 1 is from multiple stomachs and 2 is where stomach samples were pooled). We use a Student's t distribution to account for heavy tails in the distribution of the response. We use the following model: Log10(prey mass)i∼Student-t(αj[i],k[i],l[i],m[i],n[i],o[i]+β1j[i],k[i]Log10(predator mass),σ2,ν);αjβ1j∼Nμαjμβ1j,σαj2ραjβ1jρβ1jαjσβ1j2,for predator_taxa j=1,…,J;αkβ1k∼Nμαkμβ1k,σαk2ραkβ1kρβ1kαkσβ1k2,for prey_functional_group k=1,…,K;αl∼Nμαl,σαl2, for year l=1,…,L;αm∼Nμαm,σαm2, for data source m=1,…,M;αn∼Nμαn,σαn2, for site n=1,…,N; αo∼Nμαo,σαo2, for n stomachs pooled o=1,…,O, where Log10 transformed individual prey mass (wet weight in grams) is modelled following a Student's t distribution with mean μ, variance σ2, and degrees of freedom (Df) ν. The parameters α and β respectively represent intercepts and slopes that vary by grouping factors j to o. The Akaike information criterion (AIC) was used on nested models to assess the importance of all predictors. The full model had the lowest AIC by > 2 units, meaning all predictors were retained (Table S1 in the Supplement). Model diagnostic plots were performed using the R package DHARMa (Hartig, 2022).

The full model was then used to predict the mean individual body mass of prey functional groups for predator species of a given size. This enabled us to make use of many observations in DAPSTOM, for instance, which have recorded prey counts but no prey biomass. In such cases, we estimated the biomass of each prey taxa by multiplying the predicted mean individual body mass for their functional group by the observed prey count. Where prey counts were missing, e.g. much of the data from Smith and Link (2010; USA data in Fig. S1), we estimated these by dividing the observed biomass of each prey taxa by the predicted mean individual prey mass for their functional group. We provide R script and the data underlying our model, with an example showing how to predict mean individual prey size (i.e. generate fitted values) based on a list of predator taxa, predator body mass, and prey functional groups (https://github.com/MurraySAThompson/fish-feeding-traits-glmm, last access: 25 March 2024). Here we also demonstrate how to simulate data using the uncertainty measured by our model to help gauge its performance and because variability in individual prey masses is useful in food web research more broadly (Brose et al., 2019; Pomeranz et al., 2019; Scott et al., 2014). Our full model has temporal and spatial information as random effects because we were interested in developing general feeding traits for the study area, irrespective of spatial and temporal gradients. However, the significance of spatial and temporal random effects (Table S1) suggests future work exploring environmental change drivers of predator–prey body-mass scaling could be fruitful. All linear mixed effects models were fit using the glmmTMB R package (Brooks et al., 2017).

Predators were categorized by species and individual body mass. We use 20 equal size bins to categorize predator mass along a log⁡10 transformed gradient from 0.1 micrograms to 190 tonnes, capable of capturing organisms from plankton to blue whales (Table S2). Data for each species grouped into body-mass bins (henceforth species body-mass bins) were then estimated across all available stomach samples (Fig. ; Table ), with means calculated for percent prey functional group biomass, biomass-weighted PPMR (after Reum et al., 2019), and mean individual prey mass (see feeding guilds.csv; 10.14466/CefasDataHub.149). We used directly observed data where available and predictions (i.e. the fitted values) from our predator–prey body-mass scaling models where data were missing. Feeding guilds were assigned based on cluster analysis using the “ward D2” agglomeration method on Bray–Curtis dissimilarities between predator diets available in the R stats package (R Core Team, 2020).

We compared different methods of classifying feeding guilds where the dissimilarity matrix used in the cluster analysis was generated using one of the following: (1) the biomass of prey taxa (Garrison and Link, 2000a), (2) prey taxa occurrence (Thompson et al., 2020), or (3) via a novel method where dissimilarities are based on broad feeding traits (henceforth, the biomass, occurrence, and trait methods). Feeding traits were log⁡10 transformed mean individual prey mass (g), log⁡10 transformed mean biomass-weighted PPMR, and the mean percent biomass contribution to the stomach contents of zooplankton (including fish < 0.5 g), benthos, nekton (other than fish), and fish (all fish prey ≥ 0.5 g), with all variables rescaled to values of or between 0 and 1. We tested for differences between these methods of classifying feeding guilds by comparing them after re-sampling (n = 1000) subsets of the data (n = 30 unique stomach samples per predator). Predators with fewer than 30 samples were not classified into feeding guilds. Compositional change in predators between successively reclassified feeding guilds was used to determine the ability of each method to consistently classify similar predators in the same guild. Compositional change was measured using the distance to centroid following analysis of multivariate homogeneity of groups dispersions (Anderson, 2006). The method with the lowest mean distance to centroid was determined to have the most robust feeding guild classifications as determined using analysis of variance tests. First, we tested whether compositional change across feeding guilds was non-random for each method: distance to centroid (i.e. compositional change) was the response, with “Guild” and “Data” (i.e. a factor identifying each unique re-sampling event) as predictors. We then tested for significant differences between methods: distance to centroid (i.e. compositional change) was the response, with “Method”, “Guild”, and “Data” as predictors. Significant predictors were determined using the F test on nested models. Targeted tests for differences between the mean distance to centroid across methods were carried out using Tukey's all-pairwise comparisons that correct for multiple comparisons in the multcomp package (Hothorn et al., 2008).

The ability to classify common feeding guilds across ecosystems (e.g. sub-apex and apex predators) rather than area-specific guilds (e.g. a feeding guild unique to the North Sea) is another important quality for a feeding guild indicator to exhibit. We assessed how important spatial gradients were in our three different approaches to classifying feeding guilds. First, we generated latitudinal and longitudinal coordinate centroids for each predator by taking a mean across their stomach samples. We then took a mean across these predator centroids to generate a centroid for all the data and also took means across these predator centroids but grouped by feeding guild and method to generate method-specific guild centroids. Next, we measured the distance between the overall data centroid to the method-specific guild centroids using the geosphere package (Hijmans et al., 2021) and summed distances for each method. A large sum of distances for a method to the overall data centroid would indicate that feeding guilds were area-specific, largely made up of predators found close together, and thus spatial gradients would be important determinants of feeding guild structure. The method with the lowest sum of distances to the overall data centroid was deemed to be least affected by spatial gradients and thus preferred.

We provide a sensitivity analysis to determine if our modelled stomach contents data affected our conclusions about which approach to feeding guild classification was optimal. Using only observed data for prey weight and counts from DAPSTOM and ICES Year of the Stomach and data from the north-east US continental shelf (i.e. those that have published prey taxa information), we compare the ability of the different approaches to consistently classify similar predators in the same guild and classify common feeding guilds across ecosystems, as described above. Results are provided in the Supplement.

Four feeding guilds have been called for in OSPAR and MSFD guidance, i.e. planktivores, sub-apex demersal, sub-apex pelagic, and apex predators (Boschetti et al., 2021; Walmsley et al., 2016; see also https://oap.ospar.org/en/resource-catalogue/enumeration-tables/cemp-enumeration-tables/, last access: 25 March 2024), without consensus on how to categorize predators into these guilds. We use four feeding guilds here to help bridge this gap and so that we can elegantly capture a broad set of ecosystem components while exploring guild responses in biomass and species richness in the survey data. Changing the number of feeding guilds could be justified, depending on the question, and is straightforward to implement by taking a higher or lower split in the classification tree. We see this as a strength of our approach because feeding guilds are hierarchically structured much like how taxonomic or other trait information has been organized. We provide a table which details the branches for up to five feeding guilds so future assessments can choose which level of complexity suits their need. We also present axis scores from a non-metric multidimensional scaling analysis of the dissimilarities used in our cluster analysis, which provide a more nuanced understanding (i.e. bounded data as opposed to categorical) of different predator feeding traits in relation to others. Moreover, because it is a reproducible data-driven approach, new information can be systematically integrated to further resolve differences in (1) feeding traits and (2) feeding guild composition and (3) to test if changes in predator feeding traits provides evidence for spatially or temporally flexible classifications.

The new feeding guild classifications have been applied to processed otter trawl survey data for the north-east Atlantic shelf seas collected between 1997–2020 (Lynam and Ribeiro, 2022) to reveal spatial and temporal patterns in feeding guild responses (Fig. ). These survey data have been processed specifically to support state indicators, with observations for the biomass of species body-mass bins standardized to the area swept for each haul. Survey data corresponding with all our stomach content data, from the north of Norway, Icelandic waters, the Baltic Sea, and the eastern shelf seas of the USA, have not yet been standardized and processed in the same way, and hence we have not included them here. Extending this work to assess change in ecosystem structure and function across the study region covered by the stomach contents data (Fig. ) represents a key area for future development. We also provide the necessary R code (https://github.com/MurraySAThompson/fish-feeding-guild-classifcation, last access: 31 January 2025) so that our feeding guilds can be readily appended to new survey data when available and processed as required.

Compared with quarter 2 and 3 (April–September), data from quarters 1 and 4 (January–March and October–December respectively) typically have longer time series available over much of the study region and so were preferentially selected. Where data from quarters 1 or 4 were not available, otter trawl data from other quarters were used. Table S3 provides information on the surveys used and their spatial and temporal ranges and Fig. S2 depicts survey locations within OSPAR regions (e.g. Celtic Seas, North Sea). The temporal assessment covers 1997–2020 because the majority of the surveys considered have at least a near-complete time series covering that period. Longer time series do exist for some surveys but including these data would mean we are looking at long-term change for some areas but shorter-term change for others which could confound interpretation. Spatial and temporal change in feeding guild responses were determined for the Greater North Sea, Celtic Seas, Bay of Biscay and Iberian Coast, and the wider Atlantic. The assessment strata used here replicate those used for the OSPAR food web indicators: mean-maximum length and size-composition in fish communities (Lynam et al., 2022; Lynam and Piet, 2022).

Kendall's τ trend analysis was used to identify areas of significant temporal change in feeding guild responses based on the relationship between mean haul-level values of feeding biomass and species richness for each assessment strata and year. Kendall's τ scores of -1 to +1 represent a 100 % probability of a decreasing or increasing trend respectively. By using Kendall's τ, which is rank-based and non-parametric, we can detect correlations which may be non-linear. Stomach contents data, prey size predictions, and haul-level estimates of feeding guild biomass and their species richness, along with Kendall's τ correlation coefficients and p, have all been made available (10.14466/CefasDataHub.149; Thompson et al., 2024).

There were significant differences in the predator–prey body-mass scaling relationships between the different combinations of predators and prey functional groups (Fig. ). These results support our first hypothesis that predator species can have unique intra- and interspecific body-mass scaling relationships with different prey functional groups. Fish prey tended to be the biggest, meaning fish–fish interactions tended to have higher intercepts and slopes (Fig. m, lines a, c, d, and h) and thus the lowest mean PPMR, with predator species of the same size consuming relatively small benthic and zooplankton prey (Fig. m, lines b, e, f, g, I, j, k, and l). These models enabled us to estimate prey biomass and counts as well as predator–prey mass ratios across the different stomach contents datasets and species body-mass bins useful for feeding guild classifications.

Feeding guilds captured significant variation in the composition of predators for each cluster-based method (Table ), confirming our second hypothesis that multiple feeding guilds can be delineated from the analysis of feeding traits. The occurrence method had the most robust feeding guilds with the lowest compositional change in predators following re-sampling (mean distance to centroid: 0.13), followed by the trait (mean distance to centroid: 0.22) and then the biomass methods (mean distance to centroid: 0.34; randomly generated feeding guild mean distance to centroid ranged between 0.6–0.61; Fig. S3; Table S4). The trait method had the lowest sum of distances to the data centroid (2655 km), followed by the biomass (7034 km) and occurrence methods (8757 km; Fig. S4). The trait method was therefore preferred because it could identify multiple distinct feeding guilds even where we consider small subsets of predator stomach contents (n = 30 stomach samples) while being the least affected by spatial gradients in prey taxa composition. These results also confirm our third hypothesis that the effectiveness of reliably and robustly classifying predators into feeding guilds applicable across ecosystems varies due to whether classifications are based on the biomass of prey taxa, prey taxa occurrence, or broad feeding traits. Results from our sensitivity analysis using only directly observed prey count and weight information reveal that the trait approach had both the most robust feeding guilds and lowest sum of distances to the data centroid, providing further support for our decision to use it to assess changes in the survey data (Table S4; Fig. S3).

The four feeding guilds identified using the trait method have been named based on the percent biomass of prey functional groups as follows: planktivores, benthivores, bentho-piscivores, and piscivores (Fig. ). Differences between feeding guilds were related to predator size, which correlated positively with piscivory and negatively with planktivory (Fig. S5). Small body-mass classes of species often occur in the planktivore guild, moving to another guild as they increase in size, with multiple medium to larger body-mass classes of a species often in the same guild (see feeding guilds.csv; 10.14466/CefasDataHub.149). Typically, the biggest fish within and across feeding guilds had the highest PPMR (hence the sublinear relationship in Fig. m, where prey increased less than their predators per unit increase in individual body mass), yet piscivores were typically the biggest and had the lowest PPMR on average. This apparent contradiction is largely because small piscivores had some of the lowest PPMR values, whereas big planktivores and benthivores had some of the highest values (Fig. ).

When assigning feeding guilds in the survey data, we were able to classify 92 % of the biomass, which included 122 species body-mass bins. Many rare predators observed in the survey data (n = 366, representing 8 % of the surveyed biomass) remain unclassified due to insufficient stomach contents data (Table S5). The perspective of change in the survey data is therefore weighted towards predators contributing most to community biomass and ecosystem functioning. We found clear spatial structure and regions of contrasting temporal change in feeding guild biomasses and their species richness (Figs.  and ), confirming our fourth hypothesis. For instance, significant and spatially extensive temporal decreases in planktivore feeding guild biomass (i.e. lower in the food web) were evident in the Celtic Seas and Bay of Biscay, where the biomass of the bentho-piscivore and piscivore feeding guilds (i.e. higher in the food web) has increased (Fig. ). Benthivore biomass has increased in the southern North Sea, where there has been little change in other feeding guilds. Planktivore, bentho-piscivore and piscivore biomass all decreased in at least one assessment strata in the northern North Sea. Regions of temporal change in species richness were also different across feeding guilds (Fig. ). For instance, over large areas in the Celtic Seas, Bay of Biscay, and northern North Sea, where there was relatively limited change in planktivore species richness, the species richness of benthivores, bentho-piscivores, and piscivores all increased (see Fig. S6 for changes in unclassified biomass).