Food web rewiring drives long-term compositional differences and late-disturbance interactions at the community level

Our study shows significant differences in community composition caused by the nutrients’ enrichment in all three sampling days (permutational multivariate ANOVA [PERMANOVA], predisturbance: F_1,23 = 1.70, P = 0.022; maximum effect: F_1,23 = 3.03, P = 0.003; recovery: F_1,23 = 2.80, P = 0.001; SI Appendix, Table S2). Both the herbicide and the insecticide shifted the community composition in the maximum-effect phase (PERMANOVA, herbicide: F_1,23 = 2.73, P = 0.003; insecticide: F_1,23 = 6.75, P = 0.001; SI Appendix, Table S2) and in the recovery phase (PERMANOVA, herbicide: F_1,23 = 1.88, P = 0.017; insecticide: F_1,23 = 2.45, P = 0.002; SI Appendix, Table S2). Interactions between pesticides were only significant in the recovery phase under noneutrophic conditions (PERMANOVA, F_1,23 = 2.26, P = 0.006; SI Appendix, Table S2). Generally, we did not observe compositional recovery (Fig. 1). Communities undergoing different treatments were still well distinguished from the control in the multivariate space in the recovery phase, suggesting that different treatments promoted different long-term multivariate composition (2).

Species evenness was not significantly affected by the treatments (P > 0.05), whereas the Shannon’s index was significantly reduced by the insecticide in the maximum-effect phase (ANOVA, F_1,16 = 6.05, P = 0.021; SI Appendix, Table S2). Species richness was significantly reduced by the insecticide in the maximum-effect phase (ANOVA, F_1,16 = 5.89, P = 0.027; SI Appendix, Table S2) and by the nutrients in the maximum-effect phase (ANOVA, F_1,16 = 9.12, P = 0.008; SI Appendix, Table S2) and recovery phase (ANOVA, F_1,16 = 26.49, P < 0.001; SI Appendix, Table S2). Richness was also reduced by the simultaneous application of the herbicide and the insecticide under noneutrophic conditions in the recovery phase (ANOVA, F_1,16 = 8.496, P = 0.010; SI Appendix, Table S2).

Total biomass was significantly increased by nutrients enrichment in the predisturbance phase (ANOVA, F_1,16 = 2.24, P = 0.039; SI Appendix, Table S2) and in the maximum-effect phase (ANOVA, F_1,16 = 7.12, P = 0.016; SI Appendix, Table S2), while in the maximum-effect phase, the herbicide decreased the total community biomass (ANOVA, F_1,16 = 5.21, P = 0.038; SI Appendix, Table S2).

Unweighted topology-based food-web properties (number of links, link density, number of basal species, path length, generality, and vulnerability) were generally negatively influenced by nutrients enrichment in the maximum-effect and recovery phases (SI Appendix, Table S3). In the maximum-effect phase, also the insecticide significantly affected several food-web properties (SI Appendix, Table S3), while the herbicide did not. In the recovery phase, neither the insecticide nor the herbicide modified these metrics. Disturbances interactions on food-web properties appeared only in the recovery phase between the herbicide and the insecticide under noneutrophic conditions. The pesticide mixture significantly decreased the number of links (ANOVA, F_1,16 = 5.01, P = 0.039), link density (ANOVA, F_1,16 = 3.12, P = 0.036), and generality (ANOVA, F_1,16 = 6.32, P = 0.023) and increased the number of basal species (ANOVA, F_1,16 = 6.63, P = 0.020).

Node-weighted properties were never modified by the insecticide, whereas the herbicide significantly decreased connectance in the maximum-effect phase (ANOVA, F_1,16 = 7.63, P = 0.037; SI Appendix, Table S4). Nutrients’ enrichment decreased connectance in the predisturbance phase (ANOVA, F_1,16 = 1.41, P = 0.036) and generality and vulnerability in the recovery phase (ANOVA, F_1,16 = 9.56, P = 0.006 and F_1,16 = 15.75, P = 0.001, respectively). Significant interactions between disturbances were reported only in the recovery phase, when the mixture of pesticides (herbicide × insecticide) significantly decreased connectance (ANOVA, F_1,16 = 5.14, P = 0.037).

Link-weighted connectance was increased by the herbicide and nutrients as single treatments in the maximum-effect phase (ANOVA, F_1,16 = 11.18, P = 0.006 and F_1,16 = 5.74, P = 0.029, respectively; SI Appendix, Table S5) and by the herbicide as a single treatment in the recovery phase (ANOVA, F_1,16 = 4.46, P = 0.036, respectively). Conversely, the insecticide and the mixture of the pesticides in the recovery phase decreased link-weighted connectance (ANOVA, insecticide: F_1,16 = 3.32, P = 0.048 and insecticide × herbicide: F_1,16 = 8.12 P = 0.011). Link-weighted generality was increased by the nutrients in the predisturbance and recovery phase (ANOVA, F_1,16 = 6.89, P = 0.018 and F_1,16 = 6.80, P = 0.019). Linked-weighted generality was also significantly decreased by insecticide in the maximum-effect phase (ANOVA, F_1,16 = 11.18, P = 0.006) and increased by the herbicide, but decreased by the combination of the pesticides, in the recovery phase (ANOVA, F_1,16 = 4.32, P = 0.046 and F_1,16 = 4.94, P = 0.040, respectively).

We also found a significant correlation between the compositional log response ratio (LRR) and the interaction-strength LRR of all sampling days (Fig. 2).

In this study, we assessed the value of ecological network analyses to better understand the differences between univariate and multivariate compositional descriptors of community composition and to identify ecological processes that may be responsible for long-term dissimilarity and late interactions between disturbances at the community level. Our study shows that changes in multivariate community composition were linked to changes in community biomass or diversity indexes after the disturbances’ application (day 15, maximum-effect phase). After 50 d (recovery phase), all univariate descriptors of composition and function did not differ anymore from the control, indicating potential recovery. Yet, the analysis of multivariate composition showed that communities treated with a pulse application of insecticide or herbicide as a single disturbance were still significantly different from the control. Long-term compositional dissimilarities at the community level after pulse disturbance were reported in another study using aquatic communities (15) and have been observed in a wide range of ecosystems (2). Analyzing the food-web properties and interactions-strength dynamics, we found a coherent explanation for this apparent mismatch between univariate and multivariate descriptors of community composition, which has been reported extensively in the literature (4, 27, 28).

In the maximum-effect phase, topological properties were strongly influenced by the insecticide. Such properties are based on presence–absence interactions and thus heavily affected by the number of species (i.e., richness), which was indeed significantly reduced by the insecticide in the maximum-effect phase. Node-weighted properties (based on biomass) in the maximum-effect phase were significantly influenced by the herbicide. Consistently, nutrients enrichment and the herbicide were the only treatments modifying total community biomass in the same phase. In the recovery phase, when the effects of the pesticides were expected to be recovered, not a single property of the unweighted or of the node-weighted food webs was affected by the pesticides in isolation, and no univariate descriptor of community composition was significantly impacted.

Yet, looking at the link-weighted food-web properties, we found that connectance and generality were significantly modified by both pesticides applied as single disturbance in the recovery phase. An increase in link-weighted connectance suggests that several nodes/species have relatively high and similar energy flows (i.e., species-interactions strength), which increases link-weighted food-web complexity (23). Oppositely, a decrease in link-weighted connectance indicates an uneven distribution of energy flows, which may be clustered in a specific portion of the food web and disconnected from the whole network. In our model ecosystems, the herbicide increased link-weighted connectance, showing an even distribution of energy fluxes across the feeding guilds (Fig. 2B). This even distribution of interactions strength was also mirrored by the increase in link-weighted generality, which indicates an increased number of preys for each predator. Conversely, the insecticide caused a rewiring of the interactions strength that promoted the compartmentalization of energy fluxes (significant reduction in link-weighted connectance), reducing the outgoing energy fluxes from the macrozooplankton compartment, while expanding the outgoing fluxes from basal species and benthic organisms (Fig. 2B). Subsequently, this increased the incoming flow to high-trophic predators feeding on benthic organisms (Fig. 3).

Overall, the analysis of single pulse disturbance on our experimental systems shows that, after pulse disturbance, communities may change in their univariate and multivariate composition, as well as in their functional parameters. However, biomass and diversity may be recovered, but the rewiring of interactions strength prevents communities to regain their concise composition. This was further sustained by the significant correlation between the relative changes in multivariate composition and the relative changes in the interactions’ strength. Indeed, large compositional dissimilarities from the control were positively correlated with large differences in the interactions’ strength, suggesting that changes in the intensity of species interactions may shape community composition. We propose that such rewiring of the interactions' strength is linked to the different sensitivities of the taxa to a specific disturbance. Our systems were populated by several species with different sensitivities to the disturbances applied. The initial decrease in univariate descriptors of composition was mainly caused by the decline of sensitive species (25, 29–31). Later, through recolonization (particularly of sensitive Ephemeroptera taxa, such as Cloeon sp. and Caenidae) or the development of resistance forms (e.g., ephippia and eggs in the case of Cladocera), the total number of species could be regained. Similarly, total biomass was recovered by compensation dynamics, where tolerant species increased their abundance at the expenses of declined sensitive species (32). Nonetheless, the predisturbance relative abundances of all taxa were not recovered, as indicated by the multivariate analysis of composition. We suggest that such community dissimilarity was driven by the tolerant species that persisted in the environment for the whole experimental time, causing the rewiring of the interactions strength in the food web highlighted by our analysis. Thus, compositional recovery was stonewalled by what could be called the “postdisturbance priority effect.” The priority effect (33) suggests that the order and timing of species immigration during community assembly may affect species abundances and cause historical contingency in the structure and function of the community. We build on this concept and suggest that, after a disturbance large enough to cause the decline of some populations within a community, a similar process drives the rewiring of interactions strength. The niche pre-emption and niche modification (33) carried out by the tolerant species that persisted in the community hampered the recovery of declined population through competitive exclusion and/or strong predation pressure (34). This was particularly evident in our systems, where the insecticide almost completely eradicated all Cladocera (e.g., Daphnia sp., Ceriodaphnia, Simocephalus, Alona, and Alonella) and Ephemeroptera (Cleon sp. and Caenidae) taxa in the maximum-effect phase (SI Appendix, Fig. S1). Yet, the Ephemeroptera taxa could recolonize the mesocosms successfully, whereas the Cladocera taxa regained an abundance level comparable to the control’s one only under eutrophic condition. Noneutrophic systems treated with the insecticide showed a dominance turnover in favor of the Cyclopoid copepods from the maximum-effect phase onward. Probably, in eutrophic systems, the zooplankton suffered less from competition, as the nutrients treatment assured higher energy provision after the insecticide pulse that led to an increased overall zooplankton biomass in the recovery phase (Fig. 4A). In turn, this decreased the phytoplankton biomass, owing to the increased grazing pressure. Similarly, the herbicide treatment strongly reduced the abundance of the Chlorophyta after its application. This promoted a compensatory dynamic that increased Dinophytes abundance. Subsequently, Dinophytes competitively excluded some of the Chlorophyta, which never regained the predisturbance dominance (SI Appendix, Fig. S1).

We can conclude that interactions strength was rewired, which shifted network connectance and modified further single-population dynamics without modifying topological properties of the food web or univariate descriptors of community composition and function. Although these mechanisms have been long considered in theoretical ecology, in this work, such rewiring of species interactions is quantitatively assessed with empirical data and used to explain the reported delayed (or completely missing) compositional recovery of freshwater communities after pulse disturbance.

Nutrients enrichment caused a significant shift in several unweighted and node-weighted properties throughout the experiment. Eutrophic systems, despite showing a simplification of the food web, which was reflected by the changes in the architecture of the network and by univariate descriptors, were characterized by a general increase in biomass (Fig. 4A) and interaction strength in the long term (Fig. 3). Eutrophication is known to increase the energy fluxes due to higher resources availability (35). Consequently, eutrophication increases the bottom-up energy supply, which may lead to the “paradox of enrichment” (36) and to a following decline in richness (37). However, in the context of this study, it should be noted that nutrients were applied before and throughout the experiment as a press perturbation; thus, significant effects for the whole experimental time were expected. The nutrients treatment was designed to look for possible interactions between pulse (pesticides) and press (eutrophication) disturbances, which were not identified in this study.

Pulse-disturbance interactions on multivariate community composition only appeared in the recovery phase in the herbicide–insecticide treatment under noneutrophic conditions. It is unlikely that this compositional change was driven by the sensitivity of the different taxa toward the disturbances, as in the recovery phase, the concentration of the pesticides was too low to be harmful for aquatic organisms (25, 38). The analysis of the link-weighted properties highlighted that in the recovery phase, the connectance and the generality of the food web were decreased by the mixture of the pesticides. Fig. 2B shows that the outgoing energy fluxes in this treatment were dominated (>80%) by the basal species, whereas the outgoing fluxes from higher trophic levels were drastically reduced, with no outgoing fluxes from the high-trophic predators. At the same time, Fig. 4B shows that the per capita average interaction strength (sum of ingoing and outgoing flows divided by the number of nodes/species) for top predators was extremely low compared to the other treatments, where high-trophic predators exerted strong interactions. Although both strong and weak interactions are essential in promoting the stability of ecological networks (39, 40), the disruption of strong trophic interactions has been previously shown to produce drastic cascading effects on food webs, which lead to changes in community structure (41). The loss of strong interactions from high-trophic predators, despite being well reflected by the changes in connectance and generality, also resulted in significant changes of the node-weighted properties. Node-weighted connectance was decreased by the pesticides mixture, driven by a reduction of high-trophic predators and macrozooplankton biomass (Fig. 4A). Probably, the reduced grazing pressure derived from the pesticide-driven decrease in macrozooplankton biomass triggered an expansion at the base at the food web, which was reflected as an increased phytoplankton biomass (Fig. 4A), resulting from decreased top-down control by the zooplankton and as an increase in diversity of basal species (SI Appendix, Table S3). Despite the increased number of basal species, the mixture of pesticides reduced the total richness of the network in the recovery phase (SI Appendix, Table S2). Such reduction was then translated into significant changes in the topological properties of the food web (reduction in the number of links, link density, and generality; SI Appendix, Table S3). Particularly, the large changes in the relative abundances of macroinvertebrate composition (SI Appendix, Fig. S1) coupled with the reduced biomass of the zooplankton might have reduced the availability of preys for the high-trophic predators. High-order consumers are known to suffer disproportionately from instability at lower trophic levels (42). The changes at a lower trophic level thus may have led to a strong biomass reduction of the high-trophic predators.

That the loss of high-trophic predators may trigger rapid changes in community structure (41) and may lead to secondary extinctions (43) was already known. However, here, we show that the rewiring of the trophic interactions’ strength, which led to a loss of strong interactions with high-trophic predators without causing their extinction, is sufficient to trigger a complete change in the food web (that is, topological as well as interactions strength).

In this study, we investigated the differences between functional and compositional dynamics under different disturbance regimes. We show that functional and multivariate compositional responses to disturbance are substantially different. Importantly, we demonstrate that community biomass can be recovered, despite significant changes in community composition, supporting the idea that functional redundancy can ensure a similar level of functioning, even if species composition changes (1, 2). In line with other empirical studies (3, 4), our work suggests that under a changing environment are actually shifts in composition that allow maintaining biomass stable. Indeed, in field or semifield experiments (i.e., high richness), sensitivity-driven species turnover and changes in dominance after disturbance allow maintaining (or recovering) ecosystem functioning through species asynchrony (5, 6) (SI Appendix, Fig. S1). In this context, the eutrophication treatment deserves a particular mention. The press nutrients enrichment determined a reduction of species richness, but a general increase in biomass production. Nutrients’ addition favors the growth of primary producers with high nutritional demands and fast uptake strategy. In turn, the fast-growing species cause an increase in competition for other resources (e.g., light or space), which determines the decline of the other slower-growing species. Thus, under eutrophic conditions, functioning levels can increase, despite changes in composition and a diversity decline (4, 7).

Recently, energy fluxes going from one trophic compartment to another have been proven to be a sensitive and meaningful way to measure ecosystem functioning in multitrophic systems (44–46) and have been used to quantify the effect of biodiversity loss on ecosystem functions (47). However, how different measurements of functional and compositional stability (48) relate to the temporal variation in fluxes has not yet been investigated. Further investigation in this direction may give rise to new hypotheses and shed light on the relationship between functional and compositional aspects of stability (49, 50). Regarding the frequent disturbance-driven compositional shift we found, our study suggests that the rewiring of the interactions’ strength may be driving long-term compositional dissimilarity at the community level under both single and multiple disturbances. Interestingly, the only significant interaction between pulse disturbances on the community composition appeared at a late stage and was associated with a complete change in the ecological network, which produced an expansion of the bottom of the food web and a decrease in high-order consumers. We identify interactions strength (or energy fluxes) rewiring as the driving mechanism of the restructuring in the food-web architecture. Consequently, it appears that changes in trophic structure and energy fluxes are key elements to understand—and potentially explain—the mechanisms driving the late appearance of interactions between disturbances at high levels of biological organization (14, 51, 52).

We do recognize, however, that our approach has some limitations. First, this study includes only negative trophic interactions. Nontrophic and positive interactions have been shown to play a role in determining the overall effects of interacting disturbances (19) and in shaping food webs (53). Additionally, although our work involves a relatively large number of species and links for experimental systems, it is still incomparable with real-world ecosystems, where food webs are larger and more complex. Future research that may succeed in including the analysis of nontrophic (competition or facilitation) and positive trophic interactions (mutualism) in food webs containing a larger number of high-trophic predators (i.e., vertebrates) has the potential to eviscerate the mechanisms driving communities’ responses to disturbance at higher spatial resolution, as well as those causing the large compositional turnover reported at the global scale (54).

We are aware that the analysis of ecological networks poses great challenges, as the construction of food webs may be time- and resource-consuming. Moreover, resolving complex food webs at a high spatial scale is extremely laborious, if not impossible. Nonetheless, the application of food-web analyses to model ecosystems exposed to different types and gradients of disturbance may unveil mechanisms that explain ecological responses in natural ecosystems (55, 56). Communities and ecosystems are the primary target of management and policy makers (57, 58); thus, understanding the processes causing (unexpected) long-term interactions between disturbances is the first step toward increasing our predictive capacity of ecosystems response to multiple anthropogenic perturbations.

The dataset used in this study was obtained from an outdoor mesocosm experiment performed between April and July of 2019 at the facilities of the Institutos Madrileño de Estudios Avanzados (IMDEA) Water Institute (Alcalá de Henares, Spain). A full factorial design was used with eight different treatments applied in a randomized fashion and three replicates per treatment. The treatments included: the insecticide chlorpyrifos (1 µg/L and absent), the herbicide diuron (18 µg/L and absent), nutrients (750 µg/L N and 75 µg/L P, not added), and their binary and tertiary interactions. The nutrients treatment started 1 mo before the application of the pesticides. Since then, N and P were applied twice a week for the whole duration of the experiment. The freshwater community tested in the experiment was composed of phytoplankton, zooplankton, and macroinvertebrates, which were sampled one time before and several times after the pesticide application. However, for the purpose of this study, we focus on days −5 (predisturbance condition), 15 (expected maximum effect), and 50 (expected recovery) relative to the pesticides’ application. We expected recovery after 50 d based on previous mesocosm studies evaluating pulse exposure to these compounds (59, 60). Details on the experimental systems and the sampling techniques used in the experiment may be found in ref. 61.

We used publications, personal observations, and the database built by Gray et al. (2015) (62), which represents the largest standardized collection of trophic links for freshwater organisms, to establish the trophic interactions among the species of the experimental mesocosms. For some phytoplankton and microzooplankton species, references on feeding preference were not found, limiting the possibility to add these species to the network. In such cases, we aggregated the species for which feeding information was missing together with the closest taxonomically and/or size-wise species for which feeding information was available. A detailed list of the aggregations performed as well as on links’ references may be found in SI Appendix, Tables S1 and S2. The resulting interaction list comprised 208 links and 48 trophic species (24 phytoplanktonic basal species, 12 zooplanktonic species, 9 benthonic macroinvertebrates, and 3 high-trophic predators; Fig. 5). A total of 72 networks were built (8 treatments including control × 3 d × 3 replicates), considering the biomass for each species. Information on biomass calculations can be found in SI Appendix.

Two types of networks were defined to characterize the food webs for each mesocosm: qualitative or unweighted, formed by the presence or absence of interactions, and quantitative or weighted, where interactions’ strength was also considered. For the first, we calculated the following unweighted food-web properties: number of links (L), link density (L/S, where S is the number of species), connectance (C), generality, vulnerability, number of basal and top species, mean trophic level, maximum trophic level, omnivory, path length, and clustering coefficient. We used the R package multiweb [Saravia et al., 2019 (63)] to obtain these properties. As for the weighted food webs, we used a bioenergetic approach to estimate interactions’ strength, which enabled us to calculate weighted properties (23). Basically, this approach considers allometric scaling laws to quantify individual metabolic rates that are dependent on body mass and temperature, which, together with losses to predation and assimilation efficiencies, are used to quantify incoming fluxes (due to consumption) and outgoing fluxes (predation) for each species. Such an approach assumes the system’s equilibrium or steady state, implying that the total amount of energy lost by a species, either by consumption or physiological processes, is exactly compensated by the metabolized energy it gains from consumption. This means that each species loss to predation and metabolism is balanced by its energetic gains, which are defined as incoming fluxes multiplied by assimilation efficiencies (64). For instance, to calculate energy flux from a food chain (i.e., species A is eaten by species B, which is consumed by species C), we start from the top predator (species C in the food chain) and assess how much energy this species needs. Given the assumption of the system’s equilibrium, this energy needs to come out from the top predator’s prey (species B). As such, it is assumed that the energy demand of species at a given trophic level must be met by the energy intake of that level. Due to ecological efficiencies (i.e., consumers cannot use all consumed energy to produce biomass; some is lost because of respiration and excretion), to fulfill its energy demands, every species needs to consume more energy from the next-lower trophic level. The energy consumed from this lower level is then treated as energy loss for the species at this level. This loss to consumption is then added to the energy demand of the prey species itself (species B) to represent the joint energy loss of this species that needs to be compensated by the next-lower level (species A), and so on. Therefore, top predators (species without predators) only have metabolic losses, while their prey, the herbivores, have both losses due to metabolism and consumption (64). The interaction-strength calculations in a food chain or food web begin from the top predator. In this way, the calculation of the interaction strength might be biased when there is a direct effect of the experimental treatments on the influx of energy. However, none of the treatments applied here (insecticide chlorpyrifos [1 µg/L and absent], herbicide diuron [18 µg/L and absent], and nutrients [750 µg/L N and 75 µg/L P, not added]) have been reported to directly affect the top predators of our food web (Notonecta sp., Anisoptera, and Zygoptera are the high-trophic predators; Fig. 5) (65, 66). Following Kortsch et al. (2021) (23), we calculated four node-weighted and three link-weighted food-web properties, where species’ biomass and estimated energy fluxes were used, respectively. We calculated node-weighted connectance, generality, vulnerability, and mean trophic level, whereas as link-weighted descriptors of the food web, we calculated connectance, vulnerability, and generality. We used the R package fluxweb to estimate food-web fluxes (67). Table 1 provides an overview of the definition and ecological relevance of the food-web properties assessed in this study.

We analyzed the effects of single disturbances and their interactions on community composition using a PERMANOVA with 999 permutations (75), based on Bray–Curtis distances, obtained from log-transformed abundance data. PERMANOVA was performed by using the function “adonis2” from the R package vegan (76). To examine patterns in the whole community changes, we performed a nonmetric multidimensional scaling (NMDS) based on Bray–Curtis distances on abundance data for three sampling dates, representing the predisturbance condition (day −5 relative to the pesticides application), the moment of expected maximum effects (day 15 relative to the pesticides application), and the moment of expected recovery (day 50 relative to the pesticides application).

As univariate descriptors of community composition, we calculated species richness, Shannon’s index, and community evenness using the R packages “BiodiversityR” (77) and “lawstat” (78). Also, we studied how the function of the community was affected by the treatments using total community biomass. Differences among indexes and biomass values caused by treatments were analyzed by means of a three-way ANOVA followed by a Tukey post hoc test in case of significant effects of the treatments. To test for significant effects of the treatments on all the different food-web properties calculated, we employed again three-way ANOVAs followed by a Tukey post hoc test. To check ANOVA’s assumptions, we used the function “check_model” of the R package performance (79). If the assumptions were not met, we log- or sqrt-transformed the data to better fit the assumptions. To test for a correlation between the relative change in community composition and the relative change in interactions strength, we first calculated the LRR of the Bray–Curtis similarity of the community composition between each treatment and the controls: $\mathit{LRR} \mathit{compostion}=ln\left(\frac{{BC}_t}{{BC}_c}\right)$. Then, we calculated the LRR of the interaction’s strength between each treatment and the controls: $\mathit{LRR} \mathit{interaction} \mathit{strength}=ln\left(\frac{{IS}_t}{{IS}_c}\right)$. Finally, we calculated the nonparametric Spearman’s rank correlation between the absolute values of compositional LRR and interaction-strength LRR. We used the absolute values of the LRRs, as we were interested in assessing whether changes in interactions strength are well reflected by changes in community multivariate composition. Thus, both negative and positive deviations from the control’s interactions strength might be translated into compositional changes. All statistical analyses and food-web construction were conducted by using the R software (version 4.0.3).