Metagenomic analysis reveals how multiple stressors disrupt virus–host interactions in multi-trophic freshwater mesocosms

Abstract

Virus–host interactions are vital to microbiome ecology and evolution, yet their responses to environmental stressors under global change remain poorly understood. We perform a 10-month outdoor mesocosm experiment simulating multi-trophic freshwater shallow lake ecosystems. Using a fully factorial design comprising eight treatments with six replicates each, we assess the individual and combined effects of climate warming, nutrient loading, and pesticide loading on DNA viral communities and their interactions with microbial hosts. Metagenomic sequencing recovers 12,359 viral OTUs and 1628 unique prokaryotic metagenome-assembled genomes. Our analysis shows that combined nutrient and pesticide loading causes significant disruption by synergistically reducing viral alpha diversity while altering beta diversity and predator-prey linkages. Stressors lead to the simplification of virus-bacteria cross-kingdom networks, with nutrient-pesticide combinations exerting the strongest influence, although warming impacts diminish in the presence of pesticides. Stressor-driven changes also affect the abundance and composition of viral auxiliary metabolic genes, leading to complex shifts in virus-mediated metabolic pathways under multiple stress conditions. These findings underscore the importance of understanding the regulatory role of viruses on microbial communities to effectively address the challenges posed by global change.

The impact of global change-related stressors on microbial communities is poorly studied. Here, the authors use mesocosm experiments to show that combined environmental stressors such as nutrient and pesticide loading disrupt freshwater viral communities and virus-host interactions even when single stressors do not.

Introduction

Viruses, as both regulators and predators in the microbial world, are numerically dominant and play crucial roles in host adaptation, biodiversity, horizontal gene transfer, biogeochemical cycling, ecological processes, and overall ecosystem functionality^1–6. In marine ecosystems, viruses are estimated to release at least 145 Gt of carbon annually by lysing prokaryotic cells in tropical and subtropical oceans⁷, suggested to contribute to the daily mortality of approximately 20-40% of microbial hosts⁸. A recent study in freshwater ecosystems has also demonstrated that carbon cycling in these environments relies heavily on virus-host interactions⁹. Despite the significant ecological roles viruses play, our understanding of their adaptive strategies in response to environmental changes—especially within the context of global climate change—remains limited. This knowledge gap presents challenges for developing effective ecosystem conservation and restoration strategies.

The influence of environmental stressors on virus-host interactions can occur through various potential pathways. Since bacteria are the most abundant cellular organisms, most viruses are likely to be bacteriophages (i.e., phages—viruses that infect bacteria), which are highly abundant and play critical roles in bacterial mortality and biogeochemical cycling¹. Two key ecological hypotheses describe the distinct interactions between phages and their hosts, corresponding to the two primary lifestyles of phages: the kill the winner¹⁰ and piggyback the winner^11,12 models. In the kill the winner scenario, phages predominantly exhibit a lytic (virulent) lifestyle, preying on fast-growing bacteria and thus freeing up space and nutrient resources for slower-growing bacteria, akin to a predator-prey interaction. In contrast, the piggyback the winner hypothesis involves phages adopting a temperate lifestyle, in which their genomes are integrated into the host and replicated during cell division without immediately lysing the host cell. This lysogenic state persists until internal or external signals trigger the lytic cycle¹³. Under stressful environmental conditions, virulent phages promote nutrient cycling, drive community evolution, and maintain host diversity and ecosystem stability, while temperate phages may modulate bacterial metabolism or traits to enhance host survival, fostering a mutualistic relationship that aids adaptation to adverse environments^14,15. Additionally, phages possess the ability to encode auxiliary metabolic genes (AMGs), which can influence their hosts and contribute to microbial-driven biogeochemical cycles^16,17. For instance, research in soil ecosystems has shown that AMGs may help host bacteria adapt to pollutants like pesticides and heavy metals^18,19. Similarly, studies from aquatic environments have highlighted the significant role of AMGs in biogeochemical processes^17,20.

Environmental stressors rarely occur in isolation, and the complex, often unpredictable interactions between multiple stressors can further complicate virus-host interactions. In recent decades, climate scientists have observed an unprecedented acceleration of global warming^21,22. Under a high greenhouse gas emissions scenario (SSP5-8.5), global surface temperatures are projected to rise by 6.6–14.1 °C by 2300 compared to the period between 1850-1900²². Evidence from a five-year study of experimental warming in permafrost soils suggests that viruses may influence microbial community responses to warming through functional genes and infection strategies²³. Alongside climate change, global eutrophication is expected to intensify during the 21st century²⁴. Furthermore, increased pesticide use is anticipated as a response to rising insect pest populations and metabolic rates driven by warming, which threatens crop yields²⁵. These stressors pose a growing threat to ecosystems in the context of future climate scenarios, underscoring the urgency of investigating their ecological impacts. Laboratory microcosm experiments have shown that nutrient availability can influence phage lysis rates, with lysis being enhanced in nutrient-rich environments²⁶. Pesticides, meanwhile, present substantial threats to microbial communities²⁷, though viruses may mediate host responses and influence resistance to pesticide contamination^18,28. In addition, pesticides could indirectly disrupt viral communities in freshwater ecosystems through cascade effects on the food web, specifically by targeting protozoa or other plankton^29,30. The combined effects of multiple stressors can potentially reshape the microbial world. A deeper understanding of the regulatory roles of viruses may offer new insights into evaluating the effects of these stressors and elucidating the mechanisms underlying their effects. Despite the growing evidence, our understanding of how environmental stressors affect viruses remains fragmented and incomplete, particularly regarding the joint effects of multiple stressors on virus-host interactions.

In this work, we conduct a replicated, factorial outdoor mesocosm experiment to explore how climate warming, nutrient loading, and pesticide loading individually and collectively impact viral communities and virus-host interactions. Our investigation focuses on the response strategies of DNA virus-host interactions under multiple stressors, including (1) viral and host community structures, (2) virus-host linkages, (3) virus-host cross-kingdom network structures, and (4) variations in viral AMGs. These analyses aim to deepen our understanding of how these globally pervasive environmental stressors affect biogeochemical cycles and ecosystem functions, ultimately informing strategies to mitigate the impacts of environmental change. Here, we show that multiple environmental stressors impose complex and non-additive effects on viral response strategies and virus–host interactions. Notably, combined stressors—such as nutrient and pesticide loading—induce significant alterations in viral community composition, even when individual stressors exert limited influence. These results emphasize the necessity of accounting for interactive effects among stressors and highlight their potential to restructure microbial communities in freshwater ecosystems.

Results

To realistically simulate natural freshwater ecosystems, we constructed a complex, multi-trophic mesocosm and applied environmental stressors that reflect real-world variation patterns (see Methods section and Fig. 1). After a continuous 10-month experiment, we obtained 12,359 unique DNA viral operational taxonomic units (vOTUs) and 1628 unique prokaryotic metagenome-assembled genomes (MAGs) through metagenomic sequencing. Approximately 99% of the identified viruses belong to the class Caudoviricetes (Supplementary Figs. 1 and 2). These double-stranded DNA-tailed phages are widely distributed in aquatic ecosystems^31–34 and are also frequently reported in soil studies^19,23. The MAGs were taxonomically classified as bacteria, with the dominant phyla being Actinomycetota, Bacteroidota, Cyanobacteriota, and Pseudomonadota (Supplementary Fig. 3). Subsequently, we examined the impact of environmental stressors on viral communities and their interactions with bacteria.

Fig. 1. The experimental design of this study.

A Photo of the mesocosm taken in June 2021. Letter C represents controls, W represents warming, E represents nutrient loading, P represents imidacloprid loading, and the letter combinations represent environmental stressor combinations. Each group has six replicates. B Warming treatment design: The temperature was maintained 3.5 °C above ambient, with the occurrence of multiple heatwaves. C Nutrient and D imidacloprid loading design: Nutrient (nitrogen and phosphorus) and imidacloprid doses were adjusted over time to reflect seasonal variations in agricultural practices and precipitation intensities in the middle and lower reaches of the Yangtze River throughout the year. Source data are provided as a Source Data file.

Responses of viral and host community structures

Intersection analyses showed that 1733 vOTUs were shared among all treatments, representing the largest observed intersection (Fig. 2A). A total of 980 vOTUs were shared across six treatments, excluding the EP (nutrient-insecticide co-loading) and WEP (warming plus nutrient-insecticide co-loading) groups. In contrast, 909 vOTUs were uniquely associated with individual treatments, with the highest number found in the C (control) group (171 vOTUs), followed by the E (nutrient loading) group (161 vOTUs). The distribution of vOTUs shared across different numbers of treatments was relatively uniform, ranging from 10.4% to 16.4% (Fig. 2A). Analysis of the relative abundance of broadly shared vOTUs (i.e., those present across all eight treatments) revealed that interactions involving EP or WEP were the primary drivers influencing their relative abundance (Fig. 2B; Supplementary Table 1). Specifically, the WEP group exhibited a significantly higher relative abundance of these shared vOTUs compared to the C group, whereas differences among the other groups were relatively minor. Regarding MAGs, a total of 284 were shared among all treatments (Supplementary Fig. 4). Additionally, 119 MAGs were shared across six treatments, excluding the EP and WEP groups, while 105 were shared among seven treatments, excluding the C group. The number of MAGs uniquely associated with individual treatments was relatively low, totaling only 80 (Supplementary Fig. 5A). The EP group had the most unique MAGs among all treatments, with 16 identified. Analysis of the relative abundance of broadly shared MAGs (i.e., those present in all eight treatments) revealed a decreased relative abundance in the EP group compared to the others (Supplementary Fig. 5B; Supplementary Table 1).

Fig. 2. Effects of multiple stressors on the viral community.

A An UpSet plot shows the intersection sizes of vOTUs, with each vertical bar indicating vOTUs uniquely detected in specific treatment combinations. Presence/absence data were aggregated across replicates, and presence indicates vOTUs found in at least one sample per group. Highlighted bars represent broadly shared vOTUs (i.e., those present across all eight groups). Only the top 50 vertical bars are shown; for the complete set of groups, please refer to Supplementary Fig. 8. The accompanying donut chart depicts the distribution of vOTUs shared among varying numbers of groups, with annotated counts and proportions. B The relative abundance of shared vOTUs. Data are shown with 6 biological replicates. The center line in the boxplot denotes the median, with the box edges representing the 25th and 75th percentiles (IQR). The whiskers extend to ±1.5×IQR. C Relative abundance of temperate viruses. Temperate viruses were identified via bioinformatic prediction methods. D Effect of stressors on viral richness. In C, D, data are presented for each group (n = 6 biological replicates), with error bars representing the mean ± 95% confidence interval. In B–D, the two-sided p-values from the emmeans tests were adjusted using the Benjamini-Hochberg method. Different lowercase letters between groups indicate significant differences at p < 0.05. Exact p-values are provided in Supplementary Table 14. The letter C on the x-axis represents controls, W represents warming, E represents nutrient loading, P represents imidacloprid loading, and the letter combinations represent stressor combinations. E Shifts in viral community composition under multiple stressors. Analysis of similarities (Anosim) and permutational multivariate analysis of variance (see Supplementary Table 3) with a two-sided 999-permutation test were used to assess significant differences (n = 6 biological replicates). Source data are provided as a Source Data file.

Temperate viruses were identified via bioinformatic predictions based on the detection of proviral integration sites, integrase genes, and characteristic patterns of protein composition and associations (see “Methods” for details). Under nutrient loading conditions, the proportion of temperate viruses increased slightly (Fig. 2C; Supplementary Table 1). In scenarios involving simultaneous nutrient (nitrogen and phosphorus) and imidacloprid (a common insecticide) loading, the proportion of temperate viruses displayed considerable variability, ranging from over 50% to below 30% (Fig. 2C).

The results indicated that viral alpha diversity was significantly disrupted by the combined effects of nutrient and imidacloprid loading; however, neither factor alone had a notable impact (Fig. 2D and Supplementary Fig. 6; Supplementary Table 2). Under their combined influence, both the richness and Shannon indices demonstrated significant reductions. Furthermore, the Simpson and Pielou indices revealed increased intragroup variability compared to other groups (Supplementary Fig. 6B, C). Additionally, warming did not influence the combined effect of nutrient and imidacloprid loading on alpha diversity, as changes observed in the EP group were consistent with those in the WEP group (Fig. 2D and Supplementary Fig. 6).

Warming, nutrient loading, and imidacloprid loading all resulted in significant shifts in viral community structure, which in turn altered viral beta diversity (Fig. 2E, Supplementary Table 3). Multiple stressors significantly affect viral beta diversity, with notable combined effects observed between nutrient and imidacloprid loading (Supplementary Table 3). Our findings indicate that the alpha diversity of both viruses and bacteria did not exhibit a significant response to warming (Fig. 2D, Supplementary Figs. 6 and 7A–D; Supplementary Tables 2 and 4). Notably, viral beta diversity responded more sensitively than alpha diversity, reflecting trends observed in bacterial beta diversity (Fig. 2E and Supplementary Fig. 7E; Supplementary Table 5).

Responses of virus-host linkages

Using a combination of bioinformatic methods for virus-host prediction (see “Methods” for details), we identified 3455 vOTUs-MAGs linkages, comprising 2031 vOTUs and 772 MAGs across 17 host phyla (Fig. 3A; Supplementary Data 1). Our findings indicate that multiple stressors can disrupt predator-prey linkages between viruses and hosts (Fig. 3B), with these effects varying according to viral lifestyle (Supplementary Fig. 9; Supplementary Table 6). A significantly elevated log₁₀-transformed virus-to-host ratio (VHR; calculated as the abundance ratio of viruses with identified putative hosts to their corresponding hosts) was observed in the P (imidacloprid loading), EP, WP, and WEP groups compared to the control group (Fig. 3B). For virulent viruses, similar trends were observed within the overall viral community (Supplementary Fig. 9), likely attributable to a higher proportion of virulent viruses within the abundance composition (Fig. 2C). For temperate viruses, the combined effects of nutrient and imidacloprid loading significantly disrupted predator-prey linkages; however, this effect was less pronounced under warming conditions (Supplementary Fig. 9; Supplementary Table 6).

Fig. 3. Effects of multiple stressors on virus–host linkages.

A Phylogenetic tree of MAGs at the phylum level recovered from metagenomes. Green circles indicate MAG lineages predicted to be infected by viruses, with the number of identified vOTUs shown inside the circles. The numbers in parentheses represent the number of MAGs associated with identified viruses relative to the total number of MAGs. B Effects of multiple stressors on the log₁₀ virus-host abundance ratio (VHR) of viruses. The ratio was calculated for viruses with identified putative hosts using bioinformatic prediction methods. The number of putative virus–host species pairs identified across six biological replicates for each treatment is shown above the plot. The center line in the half-boxplot denotes the median, with the box edges representing the 25th and 75th percentiles (IQR). The whiskers extend to ±1.5×IQR. The two-sided p-values from the emmeans tests were adjusted using the Benjamini-Hochberg method. Different lowercase letters between groups indicate significant differences at p < 0.05. Exact p-values are provided in Supplementary Table 15. The letter C on the x-axis represents controls, W represents warming, E represents nutrient loading, P represents imidacloprid loading, and the letter combinations represent environmental stressor combinations. C The correlation between viral and host relative abundances at the host family level, where cell size is positively correlated with the Pearson correlation coefficient (r). Asterisks indicate significance levels from a two-tailed t-test: * for p < 0.05, ** for p < 0.01, and *** for p < 0.001. The figure was divided at the phylum level based on the host, with the y-axis representing the host at the family level. Exact p values and source data are provided as a Source Data file.

Variability in correlations at the host family level (Fig. 3C) suggests that the effects of multiple stressors on virus-host interactions may differ among taxonomic groups. For instance, warming reduced correlations in lineage-specific interactions within families such as UBA1268 (Planctomycetota) and Nostocaceae (Cyanobacteriota), while enhancing correlations in families such as UBA5976 and Nanopelagicaceae (Actinomycetota) relative to the control group. The influence of multiple stressors introduces complexity to lineage-specific interactions (Fig. 3C), highlighting the necessity to consider how these stressors differentially impact viral control mechanisms within microbial communities.

Responses of virus–host network

The interactions of multiple stressors-warming, nutrient loading, and imidacloprid loading-significantly altered the cross-kingdom co-occurrence network of viruses and bacteria, leading to significant simplification of the network (Fig. 4A; Supplementary Table 7). These stressors generally decreased the network’s complexity and stability, with the exception of nutrient loading alone (Fig. 4B-I; Supplementary Table 8). Nutrient loading increased the network’s complexity (e.g., average K, in Fig. 4D) and enhanced its robustness (Fig. 4I).

Fig. 4. Effects of multiple stressors on the virus-bacteria cross-kingdom co-occurrence network.

A Network nodes are colored by virus and bacteria, while edges are colored to represent positive and negative interactions. The size of the nodes is positively correlated with their degree. B–H Changes in subnetwork properties across different groups are illustrated. Data are shown for 6 biological replicates, with error bars representing the mean ± 95% confidence interval. I Variations in the robustness of the cross-kingdom co-occurrence network are presented. Data are shown with 100 repetitions of the simulation. The center line in the boxplot denotes the median, with the box edges representing the 25th and 75th percentiles (IQR). The whiskers extend to ±1.5×IQR. The two-sided p values from the emmeans tests (B–H) and the pairwise Wilcoxon test (I) were adjusted using the Benjamini-Hochberg method. Different lowercase letters between groups indicate significant differences at p < 0.05. Exact p-values are provided in Supplementary Table 16. The letter C on the x-axis represents controls, W represents warming, E represents nutrient loading, P represents imidacloprid loading, and the letter combinations represent environmental stressor combinations. Source data are provided as a Source Data file.

When applied independently, both warming and imidacloprid loading significantly reduced the network’s complexity, including a decrease in the number of nodes, the number of links, and the average K (Fig. 4B–D). These stressors also increased average path length, the number of connected components and relative modularity (Fig. 4E–G). Additionally, they decreased network stability, as indicated by increased vulnerability and reduced robustness (Fig. 4H, I). However, the interaction between warming and imidacloprid resulted in antagonistic effects, alleviating the impact of each stressor on network complexity and stability when applied individually (Fig. 4B–I; Supplementary Tables 7 and 8). While nutrient loading notably enhanced the complexity and robustness of the network, significant disruption was observed when combined with imidacloprid loading. A simplified and less robust network was observed in the EP or WEP groups (Fig. 4A–I).

Responses of viral auxiliary metabolic genes

Through multiple screening phases (see “Methods”), we identified 2996 AMGs linked to potential metabolic functions in the KEGG database, corresponding to 100 KOs, the majority of which are involved in metabolic activities (Fig. 5, Supplementary Figs. 10A and 11; Supplementary Data 2). The richness of AMGs significantly declined under the combined effects of nutrient and imidacloprid loading (Supplementary Fig. 10B), while the composition of these genes underwent notable changes due to the stressors (Supplementary Fig. 10C). Each stressor—warming, nutrient loading, and imidacloprid loading—independently caused significant shifts in AMG composition, with a pronounced interaction between nutrient and imidacloprid loading that further amplified these changes (Supplementary Table 9).

Fig. 5. Effects of multiple stressors on viral auxiliary metabolic genes (AMGs).

A Variation in AMGs across different groups (n = 6 biological replicates). Only AMGs in treatment groups with significant differences from the control group and potential contributions to metabolism are shown. Data are presented as the mean abundance for each group. The abundance of AMGs was calculated based on the abundance of viral contigs containing AMGs (see Methods). The corresponding KEGG level 2 metabolic pathways for each gene are marked with red dots in the lower panel. Asterisks mark genes with Kruskal–Wallis p  <  0.05 and LDA score > 2. All analyses used two-sided p-values. Black asterisks (*) indicate genes with higher abundance in the treatment group compared to the control, while red asterisks indicate genes with lower abundance. Exact p-values are provided in Supplementary Table 10. The letter C on the y-axis represents controls, W represents warming, E represents nutrient loading, P represents imidacloprid loading, and the letter combinations represent environmental stressor combinations. B The roles of these AMGs in major KEGG level 3 metabolic pathways are shown, as along with the downstream and upstream compounds or pathways. Genes are shown in red font, while the names of metabolic pathways and compounds are displayed in black. Different metabolic pathways are distinguished by color blocks. Source data are provided as a Source Data file.

Stressors can alter the abundance of genes linked to various metabolic pathways, thereby affecting the regulation of host metabolism by viruses, with effects differing based on the specific type of stressor (Fig. 5A, Supplementary Table 10). The presence of multiple stressors leads to more complex changes in the abundance of viral AMGs (Fig. 5A). Under combined nutrient and imidacloprid conditions, the changes in AMG abundance diverged from those observed with each stressor alone. Climate warming amplifies these variations, resulting in increased complexity. For example, under WEP conditions, some auxiliary metabolic genes (e.g., acpP, cobS, deoC, EARS, queC, wbpD) exhibit smaller abundance differences compared to the control group relative to EP conditions, while differences in other genes (e.g., argH, dut, glnA, kdsB) are further amplified. Additionally, the direction of abundance change for certain genes (e.g., pyrB) is reversed. Furthermore, environmental stress not only directly affects the abundance of auxiliary metabolic genes and specific pathways but also indirectly influences genes regulating upstream and downstream pathways, potentially altering substance formation within these pathways and thus impacting overall metabolic activity (Fig. 5B).

Discussion

In the context of ecosystems subjected to multiple stressors, our understanding of how these stressors individually or collectively influence viral communities and their interactions with other microbial communities remains limited. This limitation hinders our ability to effectively manage and protect ecosystems in the context of climate change. Our findings indicate that the combined loadings of environmental pollutants, such as nutrient and insecticide loading, significantly disrupt DNA viral communities and their interactions with hosts. Climate warming reduces the complexity and stability of virus-host interaction networks; however, it exhibits an antagonistic effect when combined with insecticide loading. The impact of multiple stressors results in diverse changes in viral AMGs, potentially influencing biogeochemical cycles. To effectively address the impacts of climate change and environmental stressors, there is an urgent need to investigate viruses and their interactions with other microbes within complex ecosystems.

For the viral and host community structures we investigated, we observed that combinations of environmental stressors—such as the interaction between nutrient and insecticide loading (EP), or the three-way interaction involving warming (W), nutrient loading (E), and insecticide loading (P)—significantly altered both viral and host community structures. These stressor combinations significantly affected viral species’ composition and community diversity. Notably, only 14.0% of viral operational taxonomic units (vOTUs) were shared across all treatment groups, indicating strong treatment-specific niche preferences. This is consistent with the well-known ecological dependence of viruses on both host availability and local environmental conditions^35,36. Similarly, bacterial communities showed limited overlap across treatments, with only 17.4% of metagenome-assembled genomes (MAGs) common to all conditions. Together, these findings suggest that both viral and bacterial communities are highly sensitive to environmental stressors. The vOTUs shared across multiple treatments can be regarded as habitat generalists³⁷, which are viral taxa capable of maintaining detectable abundance and occurrence across a range of environmental stress conditions. These generalist viruses may possess higher environmental tolerance or enhanced mechanisms for persistence under disturbance^38,39, as evidenced by their significantly elevated relative abundance in the WEP treatment group. Such responses highlight their potential ecological advantage in complex, multi-stressor environments.

Previous studies in soil suggest that viral alpha diversity may exhibit lower sensitivity to warming than that of other prokaryotes^23,40,41. In contrast, viral communities often respond more strongly to environmental pollutants, such as heavy metal contamination¹⁹, which is consistent with our observations under insecticide and nutrient enrichment treatments. In our study, viral beta diversity exhibited more pronounced changes than alpha diversity, a trend also observed in bacterial communities. This divergence implies that different ecological mechanisms may govern alpha and beta diversities within viral and bacterial communities. Moreover, viral beta diversity was closely associated with changes in prokaryotic beta diversity, reflecting a strong coupling between viral and host community structures. This aligns with findings from soil systems, where viral beta diversity is significantly correlated with both abiotic factors and the beta diversity of prokaryotic (bacteria and archaea) communities⁴². Our findings from freshwater ecosystems reveal that viral community diversity closely tracks changes in prokaryotic beta diversity, suggesting a consistent regulatory role of biotic and abiotic factors across ecosystems. These results emphasize the complexity of virus-prokaryote interactions and underscore the need for further research on multiple stressors to elucidate the mechanisms driving these variations across broader spatial and temporal scales.

An increase in the abundance of temperate viruses was observed in the EP condition, alongside a rise in intra-group variation. The increased abundance of temperate viruses under nutrient and imidacloprid loading potentially reflects a viral strategy to evade environmental stress¹⁴. Under nutrient loading conditions, this increase may also be attributed to the enhanced density and growth of host microorganisms resulting from nutrient loading^30,43. In environments characterized by high host density and rapid growth, viruses tend to adopt a lysogenic lifestyle, consistent with the piggyback the winner hypothesis—a phenomenon observed in coral reef viromes^11,12. Although adopting a temperate lifestyle may facilitate evasion of adverse conditions imposed by multiple stressors¹⁴, the inconsistency of this response strategy and the diverse patterns complicate our understanding of ecosystem responses.

As for predator-prey linkages, imidacloprid loading alone significantly influenced the linkages between viruses and their hosts, likely through cascading effects on consumers, such as protists or copepods^44,45, across the food web. In addition to these cascading effects, even the mere presence of protists—at low abundance and with minimal predation pressure—can elevate the energy and resource demands of virus-infected cells (e.g., cyanobacterial virocells), inducing metabolic reprogramming and altering exometabolite release⁴⁶. These indirect effects may modulate the progression of viral infections and their ecological ramifications. Although not direct targets, freshwater microbes are nevertheless affected by pesticide exposure. Consequently, studying environmental pollutants within complex, multi-trophic systems is crucial to better capturing their broader ecological consequences, highlighting the urgency of incorporating viruses and other microbes into food web studies⁵. The combination of nutrient and imidacloprid loading resulted in an increased proportion of viruses exhibiting temperate strategies, which theoretically should decrease log₁₀(VHR) due to increased lysogeny^12,47. However, our study revealed a significant decline in species richness under the EP condition. Environmental stressors filtered viral species, with species homogenization resulting in stronger top-down controls on specific microbial groups. Warming also caused a significant decrease in log₁₀(VHR) for temperate viruses but had no significant effect on virulent viruses. These specific responses may introduce unexpected variations in biogeochemical cycles as a result of climate change. While adopting lysogeny can serve as an adaptive response to stressful conditions¹⁴, the combined effect of stressors, such as nutrient and imidacloprid loading, may exceed the adaptive capacity of lysogenic viruses, indicating potentially severe consequences of multiple environmental stressors. Given that different bacterial communities perform specialized functions, a deeper understanding of how stressors affect virus-bacteria interactions is essential for comprehending their broader ecological implications. Such insights may also enhance our understanding of prokaryotic succession within these microbial ecosystems.

Different environmental stressors and their combinations variably impact network complexity and stability. Nutrient loading increases network complexity (e.g., average K) and stability, likely by stimulating the growth and reproduction of basal organisms (e.g., bacteria)^30,43, which may enhance resource availability for organisms at other trophic levels (e.g., viruses) and foster interactions among them. In resource-rich environments, frequent interactions, such as predation, competition, and symbiosis, contribute to a more complex and stable ecological network, supporting multi-trophic interdependence^30,48. In contrast, warming and imidacloprid loading independently simplify and destabilize the network. Viral infection processes in microbial food webs are often temperature-dependent⁴⁹, and warming may destabilize certain virus-host relationships, reducing connections within the virus-bacteria network and disrupting its structure. Additionally, warming alters the metabolic rates of prokaryotic microbes^50,51, shifts resource use strategies, and changes community composition, which may destabilize specific virus-bacteria associations and decrease network complexity and stability. The impact of imidacloprid loading on network complexity and stability likely arises from its effects on invertebrate zooplankton, inducing trophic cascades that influence viral and bacterial interactions^29,44,45.

The combined effects of multiple stressors led to complex changes in network parameters, distinct from the impacts observed for each stressor alone. Interaction between warming and imidacloprid slightly increased network complexity and stability, possibly due to co-adaptive species responses, where exposure to one stressor enhances adaptation to additional stressors^52–54. Alternatively, warming-induced external pesticide degradation might contribute to this effect⁵⁵. In contrast, combined nutrient and imidacloprid loading simplified the network, reducing nodes and links, which corresponded with decreased diversity. Nutrients and imidacloprid exert bottom-up and top-down effects, respectively, on our complex multi-trophic ecosystem, potentially causing greater disturbances when both are present. These results align with findings that certain stressors interact non-additively^56,57. Previous research shows that environmental stressors can both destabilize and stabilize microbial networks, though most studies focus on bacteria, fungi, and cercozoans, with virus-bacteria interactions largely unexplored^58–61. Stressors may alter microbial hierarchies and cross-kingdom network linkages, depending on specific microbial domain combinations^48,61. Understanding virus-bacteria interactions across kingdoms is critical to addressing climate change threats. Our findings indicate that the effects of stressors on the virus-bacteria network depend on both the type of stressors and their interactions. While warming and imidacloprid each pose significant threats to the cross-kingdom network, nutrient loading increased network complexity and robustness. However, combined nutrient and imidacloprid loading resulted in a simpler network with fewer nodes and links, raising concerns about the combined effects of environmental pollutants on biogeochemical cycles and aquatic ecosystems. Additionally, the potential antagonistic effects of warming and imidacloprid loading under certain conditions present challenges for ecosystem management, underscoring the need for a more comprehensive understanding of the implications of multiple stressors for effective ecosystem management^62,63.

For viral auxiliary metabolic genes (AMGs), we found that their abundance was influenced by multiple environmental stressors. Previous research suggests that virus-associated AMGs may contribute to biogeochemical cycles, including those of carbon, nitrogen, phosphorus, and sulfur^17,64. Our study provides evidence that key ecological processes are affected by environmental stressors through the metabolic regulation of hosts in response to viruses. Genes involved in carbon cycle pathways¹⁸, such as the pentose phosphate pathway (deoC, TALDO1, hxlA, hxlB, PGD), methane metabolism (AGXT, hxlA, hxlB), and glyoxylate and dicarboxylate metabolism (glnA, AGXT), were significantly affected by stressors, particularly under conditions of multiple stressors (Fig. 5). For instance, in the control group, the abundance of TALDO1, PGD, hxlA, and hxlB were notably higher than in the EP, WP, and WEP groups, indicating that the ability of viruses to modulate the microbial carbon cycle is compromised by the combined effects of multiple stressors. It is important to note that the findings of this study are primarily based on metagenomic sequencing, which reveals the presence of viral AMGs but does not provide information on their actual expression. While previous studies have provided evidence that AMGs are actively involved in the host’s metabolic activities^16,65, our analysis of partial metatranscriptomic data also demonstrates that a significant number of viral sequences recovered from the metagenome are active (Supplementary Table 11). Furthermore, these AMGs are abundantly expressed in multiple metabolic pathways, indicating their regulatory role in the host’s metabolic processes (Supplementary Discussion; Supplementary Fig. 12). Future studies incorporating transcriptomics will further explore the expression differences of AMGs under multiple stressors, providing a more comprehensive understanding of their functional relevance.

Some genes are implicated in multiple metabolic processes, suggesting that the metabolic changes resulting from variations in these auxiliary metabolic genes require further experimental validation in future studies to fully understand their true impacts. However, it is crucial to note that environmental stressors can influence multiple viral AMGs upstream and downstream of metabolic pathways, potentially leading to significant ecological risks in environments exposed to multiple stressors. Therefore, it is urgent to incorporate virus-mediated auxiliary metabolic regulation mechanisms into the assessment of ecosystem impacts under multiple stressors.

Although viruses play a vital role in regulating the structure and function of microbial communities, their responses to multiple environmental stressors remain underexplored. This knowledge gap hinders our capacity to effectively address the challenges posed by global climate change and various environmental stressors. Our research, conducted in near-natural ecosystems and accounting for the temporal variability of stressors, offers significant insights into the responses of microbial communities to these challenges. Our findings indicate that viral communities and their interactions with hosts are profoundly affected by environmental stressors. Climate warming altered the structure of viral communities and the regulation of auxiliary metabolic processes, resulting in simplified virus-host networks that increase vulnerability and decrease robustness. While nutrient loading and pesticide application alone did not significantly disrupt viral diversity, predator-prey linkages, or co-occurrence networks, their combined effects severely disrupted viral communities and virus-host interactions. Furthermore, the interaction between warming and pesticide exposure mitigated the individual impacts of each under specific conditions. This evidence underscores the intricate complexity of ecosystem responses to multiple stressors. Viral auxiliary metabolic genes, known to regulate diverse metabolic pathways, showed stressor-induced shifts in their abundance across pathways. These findings emphasize the importance of considering viruses and their interactions with other organisms when evaluating the ecological impacts of multiple stressors. The combined effects of nutrient and pesticide loading observed in our study suggest that interactions between environmental pollutants may lead to greater disruptions in microbial communities, thereby interfering with essential biogeochemical cycling processes. However, pesticide loading demonstrates an antagonistic effect when combined with climate warming. In the conservation of shallow lakes, it is crucial to consider the cumulative effects of multiple pollutants and their potential interactions with climate change.

Our study highlights the significant interactive effects of multiple environmental stressors on viral and bacterial community structures, as well as their ecological linkages. These important effects would likely have been overlooked if the stressors had been examined in isolation. Given that most ecosystems are simultaneously exposed to various anthropogenic pressures, it is essential to consider their combined impacts—both within freshwater systems and across broader ecological contexts. Several limitations of the current study suggest directions for future research. First, our analysis primarily relied on metagenomic data, focusing on DNA viruses, particularly double-stranded DNA viruses. The ecological roles of single-stranded DNA and RNA viruses remain poorly understood. Future research may integrate viromics and metatranscriptomics, while also refining methods to enrich free viruses, thereby achieving more comprehensive taxonomic and functional characterization. Metatranscriptomic data are especially valuable for verifying the expression of auxiliary metabolic genes. The combination of metagenomics and metatranscriptomics will enhance our ability to assess the functional consequences of environmental stressors on microbial communities. Second, because our pollutant addition simulation was designed to closely mimic natural environmental conditions by incorporating fluctuating inputs that reflect temporal variations in pollutant loading, the study employed end-point sampling. However, incorporating temporal sampling would provide greater insight into ecologically relevant successional dynamics and transient microbial community shifts. Lastly, although this study focused on virus–bacteria interactions, other microbial taxa (e.g., fungi, archaea, and protists) may also influence virus–host interactions. Including these groups in future investigations will help to develop a more holistic understanding of microbiome responses to multiple stressors.

Methods

Ethics statement

This experiment was conducted in compliance with all relevant ethical regulations and was approved and supervised by the Animal Ethical and Welfare Committee, Institute of Hydrobiology, Chinese Academy of Sciences.

Experiment design

The mesocosm system used for this experiment was located at Huazhong Agricultural University in Wuhan, Central China (30°29′N; 114°22′E). It consisted of 48 insulated cylindrical polyethylene tanks (1.5 m in diameter and 1.4 m in height) equipped with an automatic dynamic temperature control system⁶⁶. Starting in early February 2021, the mesocosms were gradually filled with sediment, water, and various aquatic organisms to mimic the shallow lake ecosystem. Each mesocosm was half-filled with a 10 cm layer of muddy sediment collected from Lake Liangzi (30°11′3″N, 114°37′59″E). The sediment was homogenized and sieved through a 5 × 5 mm metal mesh to remove large debris, macrophyte seeds, and mollusks. To closely mimic the ecosystem conditions of natural lakes, fish, shrimp, snails, plankton, and submerged macrophytes were introduced. Each mesocosm received four bitterlings (Rhodeus sinensis, ~3 cm), four crucian carp (Carassius auratus, ~4 cm), five freshwater shrimps (Macrobrachium nipponense, ~4 cm), fourteen snails (Radix swinhoei, 1-2.5 cm), and twenty snails (Bellamya aeruginosa, ~2.5 cm). Additionally, 50 g each of Potamogeton crispus and Hydrilla verticillata were planted in the sediment to establish submerged macrophytes. Ten liters of lake water from nearby Lake Nanhu (30°28′57″N, 114°22′34″E) were added to each mesocosm to inoculate plankton and other microbes. These species, commonly found in lakes in the middle and lower reaches of the Yangtze River Basin, were introduced at densities and biomass levels consistent with those occurring in nature^67–70.

The experiment spanned from February to November 2021, with environmental stressors applied starting on April 8, 2021, after a two-month acclimation period. The study employed a fully factorial design involving three stressors: warming (W), nutrient loading (E), and the insecticide imidacloprid (P), with controls (C). This design resulted in eight treatments, each with six replicates. To minimize potential grouping biases, mesocosms were randomly assigned to one of the eight treatment groups prior to the acclimation period. All mesocosms were then acclimated under identical ambient conditions for two months to ensure they evolved in the same direction. On April 7, 2021, we systematically assessed the environmental conditions of each mesocosm. Dissolved oxygen, conductivity, and pH were measured in situ using a HACH HQD portable meter (HQ60d, HACH, USA). Nutrient concentrations (total nitrogen, total phosphorus, nitrate and phosphate) and phytoplankton chlorophyll-a content in depth-integrated water were quantified using a spectrophotometer⁷¹. Periphytic algae chlorophyll-a content was determined by suspending a polyethylene board (rough surface, 10 × 10 cm) in each mesocosm, positioned 60 cm below the water surface. Additionally, we calculated the plant volume infested (PVI), an indicator of submerged macrophyte community biomass⁷². No noticeable differences were observed in these parameters across treatment groups (Supplementary Table 12), confirming that all systems were in a comparable and consistent state prior to treatment application.

The warming treatment elevated the water temperature by +3.5 °C above ambient levels and included multiple heatwave events throughout the experiment (Fig. 1B and Supplementary Fig. 13). The frequency and intensity of these heatwaves were based on model projections from historical meteorological data for the middle and lower Yangtze River Basin and are expected to occur in this region by the end of the century due to ongoing climate warming²¹. The nutrient loading treatment involved the biweekly addition of nitrogen (N) and phosphorus (P) at a 10:1 mass ratio, achieved by dissolving NaNO₃ and KH₂PO₄ in demineralized water (Fig. 1C). The insecticide treatment involved biweekly applications of imidacloprid (70% active ingredient, PD20120072, Bayer, Germany) to simulate agricultural pesticide loading (Fig. 1D). Imidacloprid, a common insecticide belonging to the neonicotinoid class, is one of the most widely used agricultural pesticides globally and poses a higher risk to freshwater ecosystems⁷³. The intensity of both nutrient and pesticide applications was adjusted to reflect agricultural activities and precipitation patterns in the region^74,75. During the experiment, the average nutrient loading concentrations were 0.90 mg L⁻¹ for nitrogen (ranging from 0.25 to 1.6 mg L⁻¹) and 0.09 mg L⁻¹ for phosphorus (ranging from 0.025 to 0.16 mg L⁻¹). The average insecticide loading concentration was 32.67 µg L⁻¹ (ranging from 10 to 50 µg L⁻¹), consistent with levels typically observed in natural water bodies in agricultural regions worldwide^76,77.

Sampling collection

At the end of the experiment in November 2021, ten depth-integrated water samples were taken from each mesocosm using a transparent plexiglas tube (70 mm in diameter, 1 m in length). These samples were combined and thoroughly mixed in a cleaned, sterile water bucket. The mixed water samples were then filtered through 0.22 μm membrane filters (Merck millipore, Germany) using a peristaltic pump. The filters were stored at −80 °C prior to DNA extraction. A single water sample from each mesocosm was used for metagenomic sequencing, resulting in a total of 48 samples.

DNA extraction and metagenomic sequencing

Total genomic DNA was extracted from water samples using the CretMag^TM Power Soil DNA Kit (CretBiotech, China) following the manufacturer’s protocol. Concentration and purity of extracted DNA was determined with TBS-380 and NanoDrop2000, respectively. The quality of the DNA extracts was assessed by electrophoresis on a 1% agarose gel. The DNA was then fragmented to an average size of approximately 400 bp using a Covaris M220 (Gene Company Limited, China) to prepare for paired-end library construction. The paired-end library was built using the NEXTflex^TM Rapid DNA-Seq Kit (Bioo Scientific, Austin, TX, USA), where adapters containing the full complement of sequencing primer hybridization sites were ligated to the blunt-end of fragments. For each sample, 200 ng of DNA was used as input for library preparation, including 6 cycles of PCR amplification. Paired-end sequencing was subsequently carried out on the DNBSEQ-T7 platform at Wefind Biotechnology Co., Ltd. (Wuhan, China) using the DNBSEQ-T7RS Reagent kit (FCL PE150) V2.0, in accordance with the manufacturer’s instructions (https://www.mgi-tech.com/products/reagents_info/43/). For detailed information on the sequencing data, please refer to Supplementary Table 13.

Sequence quality control, genome assembly and binning

The raw reads obtained from metagenomic sequencing underwent initial trimming and contaminant removal using the Read_qc module (with default parameters) in metaWRAP v1.3.2⁷⁸. Afterward, the clean reads from each water sample were separately assembled into contigs through the metaWRAP assembly module, which employs the MegaHit⁷⁹ with default parameters. Subsequently, these contigs were binned into species-level metagenome-assembled genomes (MAGs) using the metaWRAP Binning module. This process utilized three distinct binning algorithms, namely CONCOCT⁸⁰, MaxBin2⁸¹, and metaBAT2⁸². The results of the multiple binning predictions were refined and consolidated into a more accurate bin set with the metaWRAP Binning refinement module, applying thresholds of at least 70% completeness and less than 10% contamination (-c 70 -x 10)^19,83. Remaining MAGs were dereplicated using dRep v3.4.5 with the parameters -S_algorithm ANImf -sa 0.99 -nc 0.1 -clusterAlg single⁸⁴.

For taxonomic classification, the classify_wf workflow in GTDB-TK v2.3.2 was used with the parameters -skip_ani_screen –extension fa⁸⁵, referencing the GTDB database release 214. A phylogenetic tree was then generated using the infer command of GTDB-TK with default parameters. The tree was subsequently visualized and annotated using iTOL v6⁸⁶.

Viral contig identification, taxonomic classification and lifestyle prediction

Viral contigs (primarily DNA viruses) were recovered from the metagenomic assemblies using a combination of tools, including DeepVirFinder⁸⁷, VirSorter2⁸⁸, VIBRANT⁸⁹, and CheckV⁹⁰, with strict quality thresholds. First, metagenomic contigs of at least 10,000 bp in length were identified through DeepVirFinder v1.0 (-l 10000) and VirSorter2 v2.2.4 (-min-length 10000 -keep-original-seq). Contigs were retained if they achieved a score ≥ 0.9 and p ≤ 0.01 in DeepVirFinder, or if they had a score ≥ 0.9 in VirSorter2, or a score ≥ 0.7 with hallmark genes present²³. Next, the selected contigs underwent further classification using VIBRANT v1.2.1 with the -virome parameter, where only high- and medium-quality contigs were kept⁶. These contigs were assessed (Supplementary Fig. 14) and trimmed using CheckV v1.0.1 with default parameters to eliminate potential host regions, discarding those containing host genes without any viral genes detected. Clean viral contigs were clustered into species-level viral operational taxonomic units (vOTUs) at 95% nucleotide identity using CD-HIT v4.8.1⁹¹ with the parameters -c 0.95 and -aS 0.8.

Taxonomic classification of viral sequences was initially performed using the end-to-end command in geNomad v1.7.1 with default parameters⁹², yielding a relatively high classification rate at the class level but lower resolution at the family level (Supplementary Fig. 1). To improve annotation, vContact3 v3.0.0b65 with the v220 database (https://bitbucket.org/MAVERICLab/vcontact3) was subsequently employed (-db-domain prokaryotes -db-version 220), which also resulted in a higher annotation rate at the class level but limited success at the family level (Supplementary Fig. 2).

The viral lifestyle was predicted using a combination of CheckV⁹⁰, VIBRANT⁸⁹, and PhaTYP⁹³. Specifically, vOTUs identified as proviruses by CheckV, labeled as lysogenic by VIBRANT, or predicted to be temperate (score of ≥ 0.8) by PhaTYP v0.3.0 with default parameters were categorized as temperate viruses. CheckV and VIBRANT detected temperate viruses by identifying contigs with proviral integration sites or integrase genes⁹⁴, while PhaTYP used machine learning to identify characteristic patterns of protein composition and associations to predict temperate lifestyles⁹³. A virus was classified as temperate if any of the tools identified it as lysogenic, while those not identified as temperate were classified as potential virulent viruses. Analysis of a subset of metatranscriptomic sequencing data confirmed the transcriptional activity of lysogeny-related genes in temperate viruses, with many of these genes showing active expression (Supplementary Methods, Supplementary Disscussion, and Supplementary Table 11).

Abundance quantification of vOTUs and MAGs

The abundances of vOTUs and MAGs were determined using transcripts per million (TPM) via the coverM tool (v0.6.1; https://github.com/wwood/CoverM)⁹⁵. For this, quality-trimmed metagenomic reads were mapped to dereplicated vOTUs and MAGs, employing the contig and genome modules, respectively. Alignments with a read identity of ≤ 95% or an aligned percentage of ≤ 75% were discarded to ensure quality (-min-read-percent-identity 95 -min-read-aligned-percent 75)⁹⁶. A threshold of 10 was set as a confidence filter (-min-covered-fraction 10), and values below this threshold were excluded from the analysis⁹⁷. The relative abundance of each vOTU and MAG was expressed as their proportion of the total TPM count within each sample.

Host prediction of viruses

Four approaches were employed to establish potential links between vOTUs and their prokaryotic host genomes^96–98: (i) CRISPR spacer matches, (ii) tRNA sequence similarity, (iii) nucleotide homology, and (iv) oligonucleotide frequency (ONF) comparison. For the CRISPR-based method, MAGs were screened for CRISPR spacers using CRT v1.2 with the parameters -minNR 4 -maxRL 55 -maxSL 70^97,99. Virus-host associations were subsequently identified with SpacePHARER v5.c2e680a through spacer matches¹⁰⁰. In the tRNA approach, tRNA sequences from both vOTUs and MAGs were detected using tRNAscan-SE v2.0.12 with the -G -Q parameters¹⁰¹, followed by BLASTn v2.15.0 comparisons, where results with 100% identity and query coverage were considered¹⁰². For nucleotide sequence homology, BLASTn was used to compare vOTUs and MAGs, with virus-host pairings confirmed if the matches had an e-value ≤ 10⁻⁵, nucleotide identity ≥ 90%, and alignment length ≥ 2000 bp⁹⁷. Lastly, for ONF analysis, VirHostMatcher v1.0.0 with default parameters was applied to compute distances between vOTUs and MAGs based on ONF profiles, identifying potential linkages for d₂* values ≤ 0.2^98,103,104. To investigate the potential impacts of multiple stressors on virus-host predator-prey linkages, the virus-to-host ratio (VHR) was calculated using the linked virus-host pairs, with values log₁₀-transformed for analysis⁹⁷.

Identification of viral auxiliary metabolic genes and functional annonation

All vOTUs were annotated using the DRAM-v workflow¹⁰⁵. First, vOTUs were processed through VirSorter2 to generate affi-contigs.tab files, which were then used as input for DRAM-v.py annotate with the parameters -use_uniref -min_contig_size 0 in DRAM-v v1.5.0. Since some sequences identified by DeepVirFinder were not classified as viral by VirSorter2, we applied the -min-score 0 parameter⁹⁷. Putative auxiliary metabolic genes (AMGs) were then summarized using DRAM-v.py distill (default parameters) with the following criteria: viral genes flagged for metabolism (-M) and assigned confidence scores between 1 and 3¹⁰⁵. For further refinement, we manually curated candidate AMGs using these methods: (i) genomic context of the candidate AMGs was assessed based on CheckV and VirSorter2 annotations to confirm the presence of viral hallmark or viral-like genes upstream or downstream of the AMG¹⁰⁶; (ii) illegitimate AMGs linked to nucleotide metabolism, ribosomal proteins, transcriptional/translational regulators, glycoside hydrolases, peptidases, glycosyl transferases, adenylyltransferases, methyltransferases, and structural proteins were excluded, as these are involved in viral replication, invasion, and structure, as noted in prior studies^16,107–109. The finalized AMGs were then annotated using the KEGG database¹¹⁰ to evaluate their roles in various metabolic pathways.

To ensure AMG abundances were derived solely from viruses and not confounded by host-derived sequences, the sum of abundances of viral contigs containing AMGs was used to represent AMG abundances¹¹¹. To enable competitive mapping, microbial contigs from MAGs containing AMG counterpart genes—defined by identical KEGG Orthology annotations—were also incorporated into the reference database⁶⁵. AMGs were grouped at the gene level for further analyses, including diversity, composition, and abundance. Gene-specific abundances were determined by aggregating the abundances of all AMGs associated with each gene.

Statistical analysis

All statistical analyses were conducted using R software v4.1.3. Alpha diversity metrics, including richness, Shannon’s index, Simpson’s index, and Pielou’s index, as well as beta diversity, were evaluated using the vegan package v2.6-4¹¹². Non-metric multidimensional scaling (NMDS) based on Bray-Curtis distances was performed using the metaMDS function in vegan, and significance was tested through analysis of similarity and permutational multivariate analysis of variance with 999 permutations. Correlations between log₁₀-transformed virus and host abundances within each group were calculated using the corr.test function from the psych package v2.3.3¹¹³. To visualize the shared vOTUs or MAGs across treatments, we utilized an Upset plot, implemented using the ggupset v0.4.1¹¹⁴ and ggplot2 v3.5.1¹¹⁵ packages.

Cross-kingdom co-occurrence networks constructed using the corAndPvalue function from the WGCNA package v1.72.1¹¹⁶. Only species (MAGs and vOTUs) present in at least four replicates per treatment were included in the analysis⁵⁹. Significant correlations were defined by a Spearman’s correlation coefficient > 0.6 and a FDR-corrected p-value < 0.01¹¹⁷. Network visualization was performed using the igraph package v1.3.4¹¹⁸ and Gephi v0.10.1. Nodes represent species that exhibit significant interspecies interactions within a given treatment, with edges representing these relationships. Subnetworks were extracted by retaining nodes (and corresponding edges) from individual mesocosms using the induced_subgraph function in igraph¹¹⁹. Network topological properties—including nodes, links (edges), average K, average path length, number of connected components, and relative modularity—were calculated using the net_properties.4 function in the ggClusterNet package v0.1.0¹²⁰. An increased values of nodes, links, average K, and shorter average path lengths indicate greater potential complexity within a microbiome network^{58,59,121,122}. Increased modularity or a greater number of connected components signifies a shift toward a high-modularity structure in microbial networks, which aids in mitigating the effects of species extinction within the network^60,123–126. Network stability metrics (i.e., robustness and vulnerability) were assessed using previously established methods⁵⁹, with the corresponding R scripts available at https://github.com/Mengting-Maggie-Yuan/warming-network-complexity-stability. Network robustness was defined as the proportion of species remaining after the random removal of nodes. To assess this, each network underwent 100 simulations in which 50% of its nodes were randomly removed⁵⁹. Robustness values were then calculated for each simulation, and differences in robustness across networks were compared. Vulnerability of each node measures the relative contribution of the node to the global efficiency. Network-level vulnerability is defined by the highest individual node vulnerability⁵⁹. Global efficiency was calculated as the average efficiency across all node pairs. Variations in network parameters for both the overall network and its subnetworks in response to stressors were found to be consistent (Fig. 4B–H, Supplementary Table 7).

Alpha diversity, the relative abundance of broadly shared vOTUs or MAGs (i.e., those found in all eight treatments), log₁₀ (VHR), and network properties were analyzed using generalized linear models (GLMs) with Gaussian distribution, with warming, nutrients, and pesticides (individually and in combination) as fixed factors. Residual normality was checked using quantile–quantile plots and Shapiro-Wilk tests¹²⁷. Post hoc pairwise comparisons were performed using the emmeans package v1.8.6¹²⁸. If data failed to meet normality assumptions (e.g., Simpson’s index, robustness), Aligned Rank Transform (ART) was applied using the ARTool package v0.11.1¹²⁹, followed by post hoc pairwise comparisons via the rstatix package v0.7.2 using the pairwise_wilcox_test function¹³⁰. The Benjamini-Hochberg method was used to adjust p-values to control the false discovery rate¹³¹. Statistical significance was set at p < 0.05. Differences in gene abundance were assessed using linear discriminant analysis (LDA) effect size in LEfSe v1.1.01 with default settings¹³², and LEfSe was applied to identify genes differentially represented between treatments and controls.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

T.W., P.Z., H.Z., H.W., M.Z., and J.X. designed and performed the research; T.W., M.Z., and J.X. analyzed the data; T.W., M.Z. and J.X. wrote the original manuscript; and T.W., P.Z., K.A., M.Z., and J.X. contributed to revisions. All authors reviewed and approved the final version.

Peer review

Peer review information

Nature Communications thanks Amy Zimmerman, and the other, anonymous, reviewer for their contribution to the peer review of this work. A peer review file is available.

Data availability

All raw sequencing data generated in this study have been deposited in the NCBI Sequence Read Archive (SRA) database under project ID PRJNA1189868. The primary data supporting the results of this study are publicly available on Figshare¹³³ [10.6084/m9.figshare.29040674] and KNB [10.5063/F1SJ1J3F]. Source data are provided with this paper.

Code availability

The code and programs used for analyses are described in the Methods, and are also available at 10.6084/m9.figshare.29040674.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-025-63162-2.