Results
The Celtic Sea is characterized by a “slow-fast” continuum of life history, from small, short-lived taxa producing small offspring to large, long-lived taxa with large offspring (Pianka 1970; Beukhofet al. 2019b). Large, long-lived species with low reproductive output are typically the most sensitive to fishing pressure (Winemiller & Rose 1992; Le Quesne & Jennings 2012; Wiedmann et al. 2014). The most sensitive taxa to fishing are mostly elasmobranchs: sharks, spurdog Squalus acanthias , tope shark Galeorhinus galeusand smooth hound Mustelus sp. followed by rays, cuckoo rayLeucoraja naevus , thornback ray Raja clavata , blonde rayRaja brachyura and small-eyed ray Raja microocellata . Some large fishes also show high values of sensitivity such as European conger Conger conger and ling Molva molva (Fig. 1, axis 1).
Highest values of eigenvector centrality (hereafter called centrality for simplicity) characterize highly connected taxa linked to taxa that are themselves highly connected. In the Celtic Sea, these are large piscivorous fishes, namely whiting Merlangius merlangus , megrimLepidorhombus whiffiagonis , cod Gadus morhua , hakeMerluccius merluccius , turbot Scophthalmus maximus , and squid Loligo sp. . In our case, the most central species are not the most sensitive (Fig. 2). Notwithstanding this observation, taxa at high trophic levels tend to be more sensitive to fishing and more central than other species. Indeed, sensitivity tends to increase toward the top of the network (higher values of δ15N) and centrality increases with trophic levels (Appendix, Fig. S1).
Vulnerable taxa are defined as both sensitive and exposed to fishing. In the Celtic Sea, we found no highly vulnerable taxon, i.e. no taxa in the top right corner (Fig. 3). However, some taxa had medium-high values of vulnerability: cod, edible crab Cancer pagurus , smooth-hound, and to a lesser extent hake, anglerfish Lophius piscatorius , European conger, European plaice Pleuronectes platessa , blackbellied anglerfish Lophius budegassa and ling (Fig. 3). In addition, three of these vulnerable taxa (cod, hake and anglerfish) have high values of centrality. These taxa, despite being central, are not accounting for a large proportion of the total biomass (Fig. 3), which suggests that whether these taxa are affected or favored by an external factor (i.e. environmental conditions or human pressures), it would have a low impact on the total biomass of the Celtic Sea.
Simulating scenarios of species extinction sequences, we found that connectance (defined as the number of realized interactions in the network divided by the potential ones) is decreasing the fastest when the taxa are sequentially removed according to their number of preys (Preys removal sequence) and their influence (Centrality removal sequence) (Fig. 4A.). These removal scenarios are also responsible for the fastest collapse of the network (the remaining taxa are not linked together) after simulating the extinction of respectively 60% and 75% of the taxa of the network. These scenarios of taxa extinctions lead to a network with a lower connectance than if the taxa were deleted following a random sequence (Fig. 4A.). For these two removal sequences, values of modularity show the largest increase and values of nestedness the largest decrease (Fig. S2). Sequentially removing taxa with the highest number of predators (Predators removal sequence) provokes a less steady decrease of the connectance, but still with values lower than the model of random extinctions. The network collapses after the removal of 75% of the taxa. Conversely, the removal of only the 7% of the taxa that are the most exposed to fishing (Exposure removal sequence) leads to an increase in connectance, with values higher than the random model. Removing the taxa most sensitive to fishing (Sensitivity removal sequence) does not lead to variations in connectance different from the random model and causes the later collapse of the network, after removing 93% of the taxa (Fig. 4A.).
The removal of the first 7% of the most exposed taxa to fishing (Exposure removal sequence) causes the largest number of secondary extinctions (Fig. 4B.). Then, the simulated extinctions of taxa with the largest number of predators (Predators removal sequence) leads to the highest and fastest rate of secondary extinctions, higher than the null model, after the removal of 19% of the taxa. Removing taxa from the most to the least central (Centrality removal sequence) produces secondary extinctions comparable to the random model (Fig. 4B.). Finally, removing the taxa with the largest number of preys (Preys removal sequence) and the most sensitive taxa (Sensitivity removal sequence) leads to the lowest number of accumulated secondary extinctions, even lower than the null model (Fig. 4B.).
A network is the most robust to node loss when the removal of taxa (primary removal) does not lead to secondary extinctions. The R50 robustness (Dunne et al. 2004) is defined as the proportion of taxa that has to be removed to reach the loss of ≥50% of the taxa in the original network. The larger the R50 is (maximum value of 50%), the more robust the network is. Here, the Sensitivity and Preys removal sequences lead to the most robust network (R50=50%), followed by the random model (46%), the Centrality removal sequence (46%), the exposure removal sequence (45%) and the Predator removal sequence (39%).