Vicennial metagenomic time series unveils evolutionary dynamics of giant viruses in a freshwater ecosystem Authors: Yumary M. Vasquez, Miguel F. Romero, Robert M. Bowers, Robin R. Rohwer, Katherine D. McMahon, Tanja Woyke, Frederik Schulz Published: 2026 ABSTRACT Giant viruses play crucial ecological roles in aquatic ecosystems, yet their evolutionary dynamics in response to environmental changes, particularly in freshwater environments, are not well understood. We analyzed a 20-year time series (2000-2019) of 471 co-assembled metagenomes from Lake Mendota (USA) to reconstruct 1512 giant virus metagenome-assembled genomes, providing insights into viral genome evolution. Viruses in the order Imitervirales dominate the virome, remaining consistent across seasons and years. Our findings reveal gene duplication (23% of genes) and horizontal gene transfer (29% of genes) as key drivers of genomic innovation. A co-occurrence network analysis indicates increased virus-host interactions following the introduction of an invasive predatory zooplankton in 2009, highlighting potential hosts in Bigyra, Perkinsea, and Euglenozoa. While single nucleotide polymorphism analysis shows predominantly purifying selection in viral genes, there is a significant increase in positively selected genes post-invasion, particularly those related to infection. Comparative evolutionary analyses reveal that giant viruses exhibit genome-wide substitution rates similar to co-occurring bacteria but significantly slower than smaller dsDNA phages, suggesting both stability and adaptability. Our study demonstrates that freshwater giant viruses employ various evolutionary strategies to respond to environmental change. These results underscore their significant yet often underappreciated role in freshwater ecosystem dynamics. Introduction Nucleocytoplasmic large DNA viruses (NCLDVs), commonly known as giant viruses and classified within the viral phylum Nucleocytoviricota, have genomes ranging from hundreds of kilobases to over 2.5 megabases, and infect diverse eukaryotes across marine, freshwater, and terrestrial environments, where they significantly impact ecological dynamics1,2. Many giant virus species maintain their own machinery for DNA repair, transcription, and even translation, challenging traditional boundaries between viral and cellular domains3–6. Despite their ecological significance and genomic complexity, how these viruses evolve in response to environmental change remains poorly understood, particularly in natural settings where host communities fluctuate over time. Giant virus genomes exhibit remarkable evolutionary plasticity, with genomic diversity emerging through a balance of gene acquisition and loss, a process termed the “genomic accordion”7. This expansion and contraction of genomes enables rapid adaptation to environmental pressures while maintaining core viral proteins. Multiple molecular mechanisms drive this genomic innovation: horizontal gene transfer from hosts and other microbes provides “ready-made” functional genes8–10, gene duplication creates opportunities for functional divergence11, and de novo gene emergence potentially generates entirely novel proteins12. Despite this genomic plasticity, giant viruses maintain an overall pattern of purifying selection that preserves essential functions12,13. However, our understanding of these evolutionary processes remains constrained to laboratory studies that offer mechanistic insights but examine only single host-virus systems over brief timescales14,15, while studies of environmental samples often lack the long-term temporal resolution needed to capture adaptation in progress16,17. Long-term ecological monitoring of natural ecosystems provides the ideal framework for studying how giant viruses evolve in response to environmental change17–19. Among such systems, Lake Mendota (Wisconsin, USA) offers exceptional advantages for studying viral adaptation. This eutrophic freshwater lake has been systematically monitored since 1981, establishing one of North America’s most comprehensive ecological time series20. Critically, Lake Mendota experienced a major ecological disruption in 2009 with the detection of the invasive predatory zooplankton spiny water flea (Bythotrephes cederströmii), though sediment evidence suggests earlier arrival21, followed by zebra mussel (Dreissena polymorpha) invasion in 201522. The spiny water flea invasion triggered substantial restructuring of the lake’s protistan communities, including documented declines in cryptophyte diversity and relative abundance23. Such protistan community shifts are particularly relevant for giant viruses, as these eukaryotes, along with other eukaryotic groups, represent their primary hosts. The Lake Mendota system thus serves as a natural experiment to investigate how giant viruses respond and adapt to rapid shifts in host community composition, a question with broad implications for understanding virus-host coevolution and ecosystem resilience in the face of environmental change. Here, we leverage a resource for studying viral evolution: a two-decade metagenomic time series from Lake Mendota comprising 471 samples collected between 2000 and 201924. Using a metagenomic co-assembly approach25, we reconstructed 1512 non-redundant giant virus genomes, a collection comparable in size to the 1885 bacterial genomes assembled from the same dataset. This framework enables us to address three fundamental questions about giant virus evolution: (1) How do mechanisms of genomic innovation (i.e., gene duplication, horizontal gene transfer, and de novo gene emergence) operate in natural settings? (2) How does ecosystem disruption by invasive species reshape giant virus-host relationships and the selective pressures acting on giant virus genomes? and (3) How do evolutionary rates of giant viruses compare to those of bacteria and smaller viruses in the same ecosystem? Our analyses reveal that the 2009 spiny water flea invasion serves as a natural experiment, allowing us to directly observe genomic adaptations including real-time gene gains and losses, shifts in host association networks, and accelerated selection on infection-related genes. By comparing substitution rates across microbial domains, we demonstrate that giant virus genomes demonstrate genome-wide rates matching co-occurring bacteria while adaptive rates more closely resemble other viruses. This work examines how giant viruses respond dynamically to ecological perturbation while maintaining long-term population stability, establishing their importance in freshwater ecosystem resilience. Results Community structure and temporal distribution of Lake Mendota’s giant viruses We recovered 1512 high-quality, non-redundant giant virus metagenome-assembled genomes (GVMAGs) from Lake Mendota’s 20-year time series (2000–2019), spanning six distinct seasonal periods24,25 (Fig. 1A). This viral genome count approaches the 1885 bacterial genomes assembled from the same dataset25, highlighting the substantial contribution of giant viruses to freshwater microbiome genome diversity. Taxonomic classification revealed that the majority of GVMAGs belonged to the order Imitervirales (n = 1354; 89.6% of all GVMAGs), a pattern consistent with observations from other ecosystems26–28. Within Imitervirales, Mesomimiviridae constituted the majority (n = 918; 60.7% of all GVMAGs), followed by Mimiviridae (n = 197) and Allomimiviridae (n = 84). Other orders, Pimascovirales (n = 115), Asfuvirales (n = 26), Pandoravirales (n = 8), and Algavirales (n = 9), were present but less abundant. Only PM_01 (order Pimascovirales, n = 83) exceeded 50 genomes among non-Imitervirales families (Supplementary Data S1).Fig. 1Giant virus population dynamics in Lake Mendota.A Distribution of metagenome samples from Lake Mendota across 20 years and 6 seasons. Open black circles denote the sampling time points for each metagenome sample. B Number of unique giant viruses with at least 1× coverage and 20% of the genome mapped (i.e., breadth) mapped per year (barplot) with the breakdown of composition by family (line plot). C Giant virus presence in each year and season broken down to family level. Years include the total number of giant virus MAGs (not unique) per year with at least a mean coverage of 1 and at least 20% breadth mapped per year. Season presence includes the log transformation of the mean abundance calculated using ANCOM-BC2 which takes into account the number of samples per year. Viral families are colored by their order. Giant virus population dynamics in Lake Mendota. A Distribution of metagenome samples from Lake Mendota across 20 years and 6 seasons. Open black circles denote the sampling time points for each metagenome sample. B Number of unique giant viruses with at least 1× coverage and 20% of the genome mapped (i.e., breadth) mapped per year (barplot) with the breakdown of composition by family (line plot). C Giant virus presence in each year and season broken down to family level. Years include the total number of giant virus MAGs (not unique) per year with at least a mean coverage of 1 and at least 20% breadth mapped per year. Season presence includes the log transformation of the mean abundance calculated using ANCOM-BC2 which takes into account the number of samples per year. Viral families are colored by their order. Tracking GVMAGs across the 20-year time series revealed distinctive temporal persistence patterns (Fig. 1B, C). Imitervirales maintained consistent presence throughout all 20 years, establishing this order as the stable foundation of Lake Mendota’s giant virus community. In contrast, other orders displayed intermittent detection: Algavirales appeared briefly in 2002, became undetectable for over a decade, then re-emerged during 2014–2017, while Asfuvirales was detected during 2000–2002 and later during 2008–2019. The single Algavirales GVMAG detected in 2002 was not observed again, suggesting three possible scenarios: adaptation of existing viral strains to changing conditions, emergence of new viral strains, or methodological limitations in years with lower sampling intensity (i.e., less deep sequencing). We also identified clear seasonal dynamics (Fig. 1C), with bias-corrected abundance analysis (ANCOM-BC2) revealing that giant virus abundance peaks during Spring, reaches its minimum in Early Summer, and gradually increases through Fall and Ice-on before reaching another Spring maximum. High ortholog conservation suggests common selective pressures shape giant virus genomes Lake Mendota giant virus genomes exhibit remarkably high genetic conservation despite their taxonomic diversity. Our ortholog analysis revealed that approximately 86% of genes are shared between multiple virus species, while 14% are singletons (species-specific paralogs or genes unassigned to orthogroups). At least half of each giant virus genome falls within an orthogroup (Supplementary Fig. S1). Only a single genome, of uncertain taxonomy but potentially belonging to Pandoravirales, showed lower conservation (41.7% genes in orthogroups). The degree of gene sharing varies significantly across viral families, Imitervirales members display the highest proportion of conserved genes, while Asfuvirales shows the lowest (Supplementary Fig. S2). While sampling bias partially explains this pattern, the abundant Imitervirales genomes (n = 1354) naturally share more orthogroups than the rarer Asfuvirales (n = 26). However, the overall level of gene sharing is comparably consistent29 (83% reported for 207 NCLDV genomes). This unusually high functional overlap likely reflects shared selective pressures within Lake Mendota’s relatively homogeneous environment, where similar eukaryotic host communities and physicochemical conditions impose convergent functional constraints29. We reclustered the predicted proteins from all Lake Mendota giant viruses together with those of 630 marine giant virus genomes from the Tara Oceans sequencing effort, yielding 31,222 orthogroups and 98,952 singletons30. More than half of the orthogroups (n = 17,348; 55.5%) and singletons (n = 72,032; 73%) are populated exclusively by proteins found in Lake Mendota genomes, whereas only 4967 orthogroups (15.9%) and 26,920 singletons (37%) are unique to the Tara Oceans genomes. The remaining 8907 orthogroups (28.5%) contain sequences from both environments. Thus, although almost one-third of orthogroups encode functions common to freshwater and marine viruses, the majority appear ecosystem-specific, reinforcing the idea that the ecological challenges present in Lake Mendota have led to a shared genetic toolkit. We recognize that metagenomic assemblies preferentially recover consensus sequences, which could inflate apparent conservation by under-representing highly divergent alleles. Duplication, acquisition, and real-time genome restructuring in giant viruses Gene gain, particularly through horizontal gene transfer (HGT) and gene duplication, is a major force driving the expansion of giant virus genomes11. Our analysis revealed that species-specific HGTs and gene duplications constitute a substantial portion of giant virus genetic material, accounting for approximately 29% and 23% of all genes across our dataset, respectively, with 8% of genes showing evidence of both HGT and duplication. At the order level, Imitervirales showed higher HGT rates (32%), while Pimascovirales genomes showed the highest duplication rates (25% of genes). Pandoravirales exhibited remarkably few horizontally acquired genes (3%) and the lowest duplicated genes (19%). These patterns become more pronounced at the family level, with certain families (e.g., Algavirales families AG_04 and AG_02) maintaining notably lower HGT and duplication rates compared to other viral lineages (Fig. 2A). Particularly interesting are the families IM_13 and IM_14, where recent HGTs (≥80% identity) constitute 18% and 17% of all horizontally acquired genes, respectively (Fig. 2A). Functionally, HGTs and duplications were non-randomly distributed. Replication and repair (L; n = 14,707) and transcription (K; n = 9969) genes, critical functions that aid viral reproduction capacity, show expansion in HGTs and duplications. Whether these genes are truly functional in all Lake Mendota GVMAGs, or simply represent genomic redundancy, remains unclear and will require future transcriptomic or proteomic studies. Nevertheless, our findings strongly suggest that HGTs and gene duplication serve as key mechanisms driving genome expansion in giant viruses and may significantly contribute to their ecological success and evolutionary diversification across aquatic environments.Fig. 2Horizontal gene transfer (HGT) and duplications across Nucleocytoviricota lineages.A Percentage of giant virus genes per viral family that are predicted to be the result of horizontal gene transfer or gene duplication. HGTs are further broken down to those that have a high percent identity to a non-viral homolog. B Gene gain in a Mesomimiviridae genome following the spiny water flea invasion in 2009 (See Supplementary Fig. S9 for information on years 2002−2008). Sparse and inconsistent mapping was observed in years prior to 2009 (2002–2008), with no reliable counts obtained; these data are not shown. BLASTp results indicate a homolog to the PhoH gene in bacteria. Total mapped reads per gene are indicated by the gradient colors. Viral families are colored by their order. Horizontal gene transfer (HGT) and duplications across Nucleocytoviricota lineages. A Percentage of giant virus genes per viral family that are predicted to be the result of horizontal gene transfer or gene duplication. HGTs are further broken down to those that have a high percent identity to a non-viral homolog. B Gene gain in a Mesomimiviridae genome following the spiny water flea invasion in 2009 (See Supplementary Fig. S9 for information on years 2002−2008). Sparse and inconsistent mapping was observed in years prior to 2009 (2002–2008), with no reliable counts obtained; these data are not shown. BLASTp results indicate a homolog to the PhoH gene in bacteria. Total mapped reads per gene are indicated by the gradient colors. Viral families are colored by their order. While prior research has highlighted the roles of gene duplication and horizontal gene transfer in giant viruses11, our study extends this knowledge by capturing the temporal variation in these processes, revealing patterns of adaptation and evolutionary change over two decades. Over this 20-year time series, we identified 17 putative gene gains and 13 losses across giant virus genomes, revealing a dynamic balance between genome expansion and contraction (Supplementary Data S2). Gene gains occurred predominantly in Imitervirales, with one in Pimascovirales. Several genes (n = 13) belonged to existing orthogroups but lacked paralogs within their own genomes, with three genes showing gene duplication within the genome. Three gained genes, encoding a nematode hypothetical protein, DNA-directed RNA polymerase, and phosphate starvation-inducible protein (phoH), had functions previously absent in recipient genomes but present in other Lake Mendota viruses, suggesting selective acquisition of ecologically relevant traits; in particular, phoH may aid energy homeostasis under nutrient limitation31 (Fig. 2B). We identified one gene (921 bp, 40.1% GC) lacking any homologs or duplication signature, potentially representing de novo gene emergence. This gene shows high structural confidence (pLDDT = 74.06), but minimal sequence identity (28%) to an uncharacterized protein from a marine bacterium32. The timing of gene gains showed potential correlation with ecological events, with six gains occurring in 2009–2010, coinciding with the spiny water flea invasion and subsequent protistan community restructuring. In contrast, losses, primarily in Mesomimiviridae, lacked temporal patterns. Most lost genes belonged to existing orthogroups and only three have identifiable functions: ribonuclease H, exocyst complex component sec10, and a gene annotated as “determination of stomach left/right asymmetry.” ORFan losses and genes with limited distribution among genomes suggest removal of non-essential or non-adaptive functions. One lost gene (297 bp, 50% GC) has moderate sequence similarity (83% in AlphaFold) to an uncharacterized protein from a Lake Croche (Canada) viral metagenome, albeit with low structural confidence (pLDDT = 54.5). Together, these patterns highlight the interplay of duplication, horizontal acquisition, and gene loss in shaping giant virus genomes over ecological timescales. Virus-host relationships restructure but persist following predator invasion The 2009 invasion of the spiny water flea (Bythotrephes cederströmii) in Lake Mendota created a natural experiment to investigate how giant viruses respond to ecosystem disruption through predicted shifts in host associations. To study these dynamics, we conducted a co-occurrence analysis between viral DNA polymerase B (PolB) sequences and eukaryotic 18S rRNA genes across our 20-year metagenomic time series. Our co-occurrence network comprised 7423 Nucleocytoviricota PolB sequences and 4561 eukaryotic 18S rRNA gene sequences, with positive associations dominating the network (96.5%) compared to negative associations (3.5%) (Supplementary Data S3). Network connectivity varied substantially across viral families: Mesomimiviridae showed exceptionally high connectivity (consistent with their genomic dominance in the lake), followed by Allomimiviridae, Mimiviridae, and PM_01 (Supplementary Data S3). Weighted connection score and phylogeny-aware enrichment analyses confirmed that Mesomimiviridae lacks strong preference for any single eukaryotic group (Fig. 3A; Supplementary Fig. S3), displaying instead a generalist strategy with connections across diverse protistan lineages29,33,34. Mesomimiviridae shows strong associations with previously unknown hosts such as Euglenozoa and Gastrotricha (Supplementary Fig. S3). Similarly, our co-occurrence network predicted associations with previously unknown hosts, such as Bigyra and Perkinsea across Nucleocytoviricota. The strong fungal associations observed across all viral families support recent metagenomic evidence for fungal infection by giant viruses35,36 (Supplementary Data S4), yet contrast with traditional isolation-based studies that have not recovered fungi as giant virus hosts.Fig. 3Putative eukaryotic hosts of giant viruses.A Weighted connection of Nucleocytoviricota-eukaryote pairs that takes into consideration the number of consecutive years where the association occurs. Stars indicate experimentally confirmed eukaryote-giant virus associations from published literature. B Network associations before and after the invasion of the spiny water flea in Lake Mendota. Associations present before and after invasion are colored in grey, associations present in and before 2009 colored in blue, and associations formed in 2010 and later are colored in red. Eukaryotic host subdivision (triangle) and giant virus family (circle) are colored by their group color as seen in A and sized by the number of edges (connections). Putative eukaryotic hosts of giant viruses. A Weighted connection of Nucleocytoviricota-eukaryote pairs that takes into consideration the number of consecutive years where the association occurs. Stars indicate experimentally confirmed eukaryote-giant virus associations from published literature. B Network associations before and after the invasion of the spiny water flea in Lake Mendota. Associations present before and after invasion are colored in grey, associations present in and before 2009 colored in blue, and associations formed in 2010 and later are colored in red. Eukaryotic host subdivision (triangle) and giant virus family (circle) are colored by their group color as seen in A and sized by the number of edges (connections). To investigate how the spiny water flea invasion reshaped virus-host dynamics, we compared association networks between pre-invasion (2000–2009) and post-invasion (2010–2019) periods, focusing on positive associations that represent potential infection relationships (Fig. 3B; Supplementary Fig. S4 and S5). First, we examined whether the invasion caused an abrupt structural shift in host-virus networks by analyzing annual network metrics (edges, density, modularity) and implementing formal breakpoint analyses. Results revealed no statistically significant sudden shifts in network structure, even though segmented regression identified a statistical breakpoint near 2009, indicating that ecological impacts developed progressively rather than through abrupt ecological reorganization (Supplementary Fig. S4–S6). Despite this gradual transition, we maintained 2009 as our analytical boundary for three reasons: (1) it corresponds to the first documented detection of the spiny water flea in Lake Mendota20, (2) it aligns with invasion ecology theory, which predicts that ecosystem effects emerge progressively as invader populations establish and native communities respond37, and (3) it enables comparison between two ecologically distinct states: before and after invasion establishment. This approach acknowledges the continuous nature of ecological change while facilitating statistical comparison between fundamentally different ecosystem configurations. Comparing pre-invasion (2000–2009) and post-invasion (2010–2019) networks revealed a restructuring of virus-host association patterns, with a marked increase in total network associations following invasion establishment (Supplementary Data S5). This finding raised an important methodological question: could the observed increase simply reflect improved detection power from deeper sequencing in later years? To address this concern, we controlled for sequencing depth and performed degree-preserving null models, network rarefaction, removal of highly connected nodes, and permutation tests (p = 0.046; see Supplementary Methods) to confirm that the probability of virus-host associations remained significantly higher in the post-invasion period. Finally, we used a bootstrapping approach that identified 72 high-confidence associations in the pre-invasion network versus 99 in the post-invasion network, a 38% increase (Supplementary Fig. S7). Of the high-confidence associations, 62 associations (86% of pre-invasion relationships) persisted across both periods, demonstrating stability in core virus-host relationships despite ecosystem perturbation. On the other hand, 46 high-confidence associations emerged in the post-invasion network (present in ≥95% of bootstrap iterations), particularly involving the eukaryotic group Dinoflagellata, whose abundance increased following the invasion23. This pattern suggests opportunistic viral exploitation of newly abundant potential hosts. Many moderate-confidence associations (bootstrap confidence 0.50–0.95) from the pre-invasion period strengthened post-invasion, while 10 high-confidence associations from the pre-invasion period fell to moderate confidence levels (0.557–0.937) rather than disappearing completely, indicating relationship shifts rather than complete host abandonment (Supplementary Fig. S8). Collectively, these findings reveal that giant viruses appear to maintain established infection networks while opportunistically expanding into newly available host niches created by community restructuring. This adaptive approach would enhance viral resilience to environmental change by balancing stability with opportunity, potentially explaining how giant viruses maintain long-term persistence while responding to ecosystem fluctuations. Genomic signatures of selection shift following ecosystem disruption To investigate real-time molecular adaptation in giant viruses, we analyzed fine-scale genetic variation across our 20-year time series, identifying 519 of 1512 GVMAGs (~34%) that met stringent SNP analysis criteria (≥95% read sequence identity, ≥10× sequencing depth across multiple timepoints). Across the 519 genomes, we observed a predominant signature of purifying (negative) selection (ratio of nonsynonymous to synonymous substitutions (pN/pS) < 1). This genome-wide pattern confirms previous findings that giant viruses primarily experience selection that preserves protein function13. Despite this predominant pattern of purifying selection, we detected positive selection in at least one gene (pN/pS >1) in 301 genomes. These adaptive signals were unevenly distributed taxonomically: Allomimiviridae showed the highest proportion of genomes with positive selection (30% of all Allomimiviridae genomes) (Supplementary Data S6). This variation suggests fundamentally different selective pressures across viral lineages. Seasonal patterns emerged more strongly, with SNP counts peaking during Ice-on (p < 0.001) and Spring (p = 0.009) periods relative to Early Summer, while declining significantly in Late Summer (p = 0.003). These seasonal fluctuations suggest that viral evolutionary rates may track seasonal dynamics in host availability. The balance between selection types remained stable throughout the time series, with negatively/neutrally selected genes consistently outnumbering positively selected genes (p < 0.001) and no significant temporal shifts in this balance (interaction terms: p > 0.05). While year-to-year SNP patterns showed limited variation, we hypothesized that the spiny water flea invasion might have triggered broader shifts in evolutionary dynamics. In fact, when examining viral families with SNP data from before and after the spiny water flea invasion, we observed consistent increases in the proportion of genes under selection post-invasion (Fig. 4). Comparing pre-invasion (2000–2009) and post-invasion (2010–2019) periods revealed a pattern: after normalizing for sequencing effort, SNP rates were approximately 33% higher in the post-invasion period (p = 0.022). Using a rarefied data approach to ensure identical sequencing depth between periods, we confirmed that the post-invasion increase remained robust (mean estimate = −0.52; 95% empirical interval = −0.62 to −0.43). Additionally, a change-point analysis supported 2009 as a biologically meaningful breakpoint, with an AIC sweep showing minima between 2009–2011, with all models in this range statistically equivalent (ΔAIC < 3). This pattern extends to only genes showing signatures of positive selection (defined by pN/pS >1, where pN and pS represent observed nonsynonymous and synonymous polymorphism rates). We found that positively selected genes were significantly more abundant in the post-invasion period (p < 0.001; rarefaction mean estimate = −0.39). To address potential discrepancies arising from sequencing coverage, we normalized SNP counts by the length of the viral genome (Kb) and the total metagenome size (Gb/Mbp). This normalization process confirmed that the patterns observed in our analysis were consistent. Specifically, when comparing the pre-invasion period (2000–2009) to the post-invasion period (2010–2019), normalized SNP rates remained higher in the post-invasion period (Supplementary Data S7).Fig. 4Microdiversity of giant virus families before (2000–2009; shaded grey) and after (2010–2019) spiny water flea invasion.Bar plot shows the percentage of genes with single nucleotide polymorphisms (SNPs) that are undergoing positive and negative selection prior and after invasion. Dot plot displays the presence of SNPs in genes of interest. If at least one gene per family and gene category shows positive selection, a black outline was applied. Genes with no dot indicate that the gene does not contain SNPs, it does not denote missing genes. A32: A32-like packaging ATPase; MCP: Major capsid protein; PolB: DNA polymerase family B; RNAPL: RNA polymerase large subunit; RNAPS: DNA-dependent RNA polymerase subunit; SFII: superfamily II helicase; TFIIB: Transcription Initiation factor TFIIB; TopoII: DNA topoisomerase 2; VLTF3: Poxvirus Late Transcription Factor VLTF3 like; PEPCK: phosphoenolpyruvate carboxykinase; 6PGD: 6-phosphogluconate dehydrogenase. Viral families are colored by their order (as seen in Figs. 1–3). Microdiversity of giant virus families before (2000–2009; shaded grey) and after (2010–2019) spiny water flea invasion. Bar plot shows the percentage of genes with single nucleotide polymorphisms (SNPs) that are undergoing positive and negative selection prior and after invasion. Dot plot displays the presence of SNPs in genes of interest. If at least one gene per family and gene category shows positive selection, a black outline was applied. Genes with no dot indicate that the gene does not contain SNPs, it does not denote missing genes. A32: A32-like packaging ATPase; MCP: Major capsid protein; PolB: DNA polymerase family B; RNAPL: RNA polymerase large subunit; RNAPS: DNA-dependent RNA polymerase subunit; SFII: superfamily II helicase; TFIIB: Transcription Initiation factor TFIIB; TopoII: DNA topoisomerase 2; VLTF3: Poxvirus Late Transcription Factor VLTF3 like; PEPCK: phosphoenolpyruvate carboxykinase; 6PGD: 6-phosphogluconate dehydrogenase. Viral families are colored by their order (as seen in Figs. 1–3). Despite testing for key environmental parameters (temperature, oxygen, and carbon chemistry variables), we found that invasion status remained a significant predictor of SNP rates independent of these factors (p < 0.001). However, dissolved inorganic carbon and total organic carbon showed positive associations with SNP counts (p = 0.004 and p = 0.001, respectively), while total inorganic carbon showed negative associations (p = 0.003). Environmental correlates of positively selected genes largely mirrored those of overall SNP counts: dissolved inorganic carbon and total organic carbon showed positive associations with genes under positive selection (p = 0.002 and p = 0.0002, respectively), while total inorganic carbon and dissolved organic carbon showed negative associations (p = 0.001 and p = 0.005, respectively). These relationships suggest a plausible mechanistic link through carbon availability: viral SNPs arise during replication, and elevated dissolved inorganic carbon and total organic carbon stimulate primary production38, increasing protist biomass and turnover. A larger, faster-growing host pool supports more frequent or larger viral outbreaks, creating additional replication cycles and therefore a larger pool of detectable mutations39. Collectively, these findings provide evidence that the spiny water flea invasion triggered accelerated adaptive evolution in giant virus genomes, potentially reflecting selection for novel host exploitation strategies or altered competitive dynamics in the restructured ecosystem. Giant viruses have bacterial-like genome stability and viral-like adaptive capability To place giant virus evolution in a broader evolutionary context, we conducted, to our knowledge, the first systematic comparison of evolutionary rates across three major groups from the same ecosystem, based on multi-year time-series data: giant viruses, bacteria, and smaller viruses (predominantly Caudoviricetes phages; Supplementary Data S8). Our analysis quantified both whole-genome single nucleotide substitution (SNS) rates and observed nonsynonymous substitutions (obsN) in protein-coding genes (Fig. 5). This approach revealed that giant viruses and bacteria showed similar whole-genome substitution rates, while smaller viruses evolved significantly faster than both groups. The similarity in evolutionary rates between giant viruses and bacteria is particularly interesting given their fundamentally different biological organization and genome sizes. Despite this broad similarity in overall substitution rates, a different pattern emerged when we focused specifically on genes under positive selection. In these adaptively evolving protein-coding genes, giant viruses showed nonsynonymous substitution rates that differed significantly from bacteria but not from other viruses. This finding reveals a fundamental duality in giant virus evolution: while their overall genomic conservation patterns match closer to bacteria, their adaptive evolution dynamics mirror those of other viruses.Fig. 5Evolutionary rates of bacteria, giant viruses and other viruses in Lake Mendota.A Rates of single nucleotide substitutions by genome length. Inset displays rates as violin plots and statistical tests to compare between genomes. B The number of nonsynonymous substitutions within protein coding genes that exhibit positive selection according to microdiversity analysis. Overall significance between groups was assessed using a Kruskal-Wallis test, followed by Dunn’s test for pairwise comparisons. A Holm correction was applied to adjust p-values for multiple comparisons. Red coloring refers to viral species, while black coloring refers to bacterial species. Evolutionary rates of bacteria, giant viruses and other viruses in Lake Mendota. A Rates of single nucleotide substitutions by genome length. Inset displays rates as violin plots and statistical tests to compare between genomes. B The number of nonsynonymous substitutions within protein coding genes that exhibit positive selection according to microdiversity analysis. Overall significance between groups was assessed using a Kruskal-Wallis test, followed by Dunn’s test for pairwise comparisons. A Holm correction was applied to adjust p-values for multiple comparisons. Red coloring refers to viral species, while black coloring refers to bacterial species. Discussion Giant viruses are increasingly recognized as important members of microbial communities, yet their evolutionary dynamics in natural settings have remained largely theoretical rather than empirical. While marine metagenomics has illuminated giant virus diversity and distribution, freshwater studies have generally been short-term snapshot studies lacking temporal depth. Our reconstruction of 1512 giant virus genomes from a single lake over two decades positions Lake Mendota as the currently most comprehensively characterized freshwater system for giant virus evolution. This dataset’s unique temporal resolution has enabled us to address fundamental questions about giant virus evolution and reveal that giant viruses employ multiple evolutionary strategies to maintain long-term persistence while adapting to ecological disruption. Our findings both confirm and extend previous observations of temporal patterns in giant virus communities17–19,28,40. Our 20-year time series reveals that distinct taxonomic groups show different persistence patterns: Imitervirales, particularly Mesomimiviridae, maintain continuous presence while other orders show intermittent detection across years. This consistent dominance of Mesomimiviridae suggests these viruses possess ecological advantages in freshwater systems, likely stemming from broader host ranges41. The seasonal pattern of giant viruses in Lake Mendota closely tracks the abundance cycles of eukaryotic hosts, with microeukaryotic taxa peaking in abundance during Spring, Ice-on, and Fall periods23, while the early summer minimum in giant viruses corresponds to the lake’s “Clearwater” phase, characterized by low phytoplankton abundance and biovolume42. The high gene conservation across our dataset (86% of genes in shared orthogroups) indicates strong functional convergence, in contrast to lower inter-order gene conservation in giant viruses across diverse environments43, suggesting that Lake Mendota’s giant viruses have converged on similar genetic solutions to common ecological challenges, creating a functionally cohesive viral community despite their taxonomic diversity29. Together, persistent dominant lineages and high functional conservation suggest that freshwater adaptation favors evolutionary stability rather than continuous turnover. By leveraging our time-series approach, we provide a direct observation of multiple genomic innovation mechanisms operating simultaneously in giant virus populations. Gene duplication emerged as a driver of genomic expansion (23% of all genes), enriched for functions in host infection and viral replication. Duplication events preferentially amplify core viral functions already prevalent across Nucleocytoviricota44,45, potentially enhancing infection efficiency. For example, we found duplications in genes involved in ubiquitination (K08770) thought to be involved in host protein degradation or in host signaling46,47. Similarly, horizontal gene transfer also proved to be a driver of genome innovation (29% of genes) and showed striking lineage-specific biases, aligning with previous studies documenting non-viral origins of many giant virus DNA repair genes48–50. The patterns observed in Lake Mendota giant viruses mirror observations in viral-viral gene transfer studies51, suggesting that lineage-specific factors may shape the likelihood for horizontal gene acquisition regardless of the gene’s origin. Our temporal resolution allowed us to directly observe 30 gene gain and loss events occurring in real time. These real-time observations advance our understanding of the “genomic accordion” model7, where expansion and contraction maintain large genomes while pruning non-essential functions. Genes lost were often redundant or poorly characterized functions, allowing essential functions to be preserved even as individual genes are lost52. The coexistence of duplication, horizontal transfer, and gene loss shows that adaptation relies on balancing innovation with core function preservation. Host association analyses linked Lake Mendota giant viruses to a diverse range of eukaryotic lineages, from established hosts such as Stramenopiles to putative hosts among Alveolata, Opisthokonta, and other eukaryotic supergroups33. Many predictions align with marine metagenomic studies34,53, suggesting conserved host range patterns across aquatic environments. However, our predictions also included some unknown hosts, such as Bigyra, Perkinsea, and Euglenozoa. Recent single-cell studies have confirmed some computationally predicted relationships, such as Mesomimiviridae infecting Ciliophora54, showing that computational host predictions provide valuable guidance for targeted isolation efforts. Beyond their taxonomic implications, these predicted virus-host relationship patterns may serve as sensitive indicators of ecosystem disturbance55, as evidenced by the emergence of multiple new associations with dinoflagellates following the spiny water flea invasion. The ability to identify such ecological reorganizations through virus-host network analysis demonstrates the potential of viral ecology as a tool for monitoring freshwater ecosystem health and resilience. The 2009 spiny water flea invasion created a natural experiment to investigate viral adaptation to ecosystem disruption. Our statistical analyses found that giant viruses responded to this perturbation through a retention of established host relationships coupled with opportunistic expansion into newly available host niches. Post-invasion networks showed significant increases in both the number and stability of predicted virus-host associations, with new connections forming with taxa that increased in abundance following invasion23. This ecological response was mirrored at the genomic level, where the number of single nucleotide polymorphisms and the proportion of genes under positive selection increased following invasion. The relationship between host interactions and viral evolution became particularly evident when examining specific cases of adaptation following ecosystem disturbance. For example, thioredoxin genes in Pimascovirales genomes showed positive selection specifically after the spiny water flea invasion. Thioredoxin likely mediates oxidative stress responses and is actively released by giant viruses during infection56,57, suggesting adaptation in viral control of cellular redox environments. The functional profile of positively selected genes suggests targeted adaptation of genes involved in host exploitation, with varying intensities of selection across viral families, highlighting the complex selective landscape even within a single ecosystem. The evolutionary flexibility of giant viruses likely enhances long-term persistence in a fluctuating environment, allowing exploitation of newly available host niches without sacrificing established infection networks. Additionally, by examining SNPs across multiple years, we demonstrate the shifting influences of evolutionary pressures, presenting a more dynamic understanding of how giant viruses adapt to changing environmental and ecological contexts. Our comparative evolutionary rate analyses position giant virus genome evolution between cellular life and other viruses (e.g., bacteriophages). Whole-genome substitution rates in giant viruses matched those of co-occurring bacteria, yet lagged the accelerated rates of smaller phages from the same ecosystem. Previous studies have hypothesized evolutionary rate similarities between giant viruses and cellular organisms58,59, but empirical evidence from natural systems has been lacking. Our findings suggest that genome size and gene composition, not intrinsic viral characteristics, are the primary determinants of evolutionary rates. When focusing specifically on adaptive evolution in positively selected genes, giant viruses showed nonsynonymous substitution patterns statistically similar to smaller viruses but significantly different from bacteria. This creates an evolutionary duality, giant viruses maintain bacterial-like genomic stability in core functions while retaining virus-like adaptive flexibility in genes directly involved in environmental response. The two-decade record from Lake Mendota increases our understanding of giant virus evolution, demonstrating that viruses are capable of multiple evolutionary responses to environmental change. Our findings reveal a multi-faceted adaptive strategy: continuous gene content innovation through duplication, horizontal gene transfer, and occasional de novo emergence; flexible host range expansion rather than replacement following ecosystem disruption; and a hybrid evolutionary mode that combines bacterial-like genomic stability with virus-like adaptive capacity. This evolutionary versatility likely explains how giant viruses have achieved ecological dominance across diverse aquatic environments. Beyond advancing viral evolution theory, these findings have significant implications for ecosystem modeling, as they demonstrate that viral communities respond to disturbance in ways that likely stabilize rather than amplify ecosystem perturbations. Our findings not only confirm existing hypotheses related to purifying selection and host range but also introduce perspectives on how these factors evolve over time, emphasizing the importance of considering long-term data in viral evolutionary studies. While our study provides a comprehensive, vicennial analysis of giant virus evolution in any ecosystem, important questions remain. Future work should employ single-virus genomics and multiomics to validate host predictions and determine the expression patterns of putatively adaptive genes. Comparative time-series analyses across multiple lakes with different disturbance histories will reveal whether the evolutionary patterns observed here represent universal strategies or lake-specific adaptations. Ultimately, integrating viral evolutionary dynamics into ecosystem models will be essential for understanding and predicting microbial community responses to environmental change. This time-series approach not only enhances our understanding of giant virus evolution but also serves as a model for future studies aiming to investigate the complex interactions between viruses and their ecosystems in a temporal framework. Methods Tymeflies dataset 471 water filter samples were collected from the eutrophic freshwater Lake Mendota in Madison, WI, USA. Samples span the years 2000–2019 and include multiple samples from different microbial seasons (i.e., Ice-on, Spring, Clearwater, Early Summer, Late Summer and Fall; defined by60). The collection of these sequences is designated as the Twenty Years of Metagenomes Exploring Freshwater Lake Interannual Eco/evo Shifts (TYMEFLIES) dataset24,61. Co-assembly with MetaHipMer was performed on the TYMEFLIES dataset25. Briefly, the DOE-JGI metagenome workflow v5.1.1162 was used. Viral contigs identified by geNomad v.1.7.463, as well as contigs with matches to Nucleocytoviricota hidden Markov models (HMMs) from GVClass v.1.064 were grouped into metagenome-assembled genomes (GVMAGs) using tetranucleotides and differential coverage with MetaBAT2 v.2.1565. Viral MAGs were then subjected to taxonomic classification and genome completeness using GVClass v.1.0 and CheckV v.1.0.364,66. MAGs that were not predicted as Nucleocytoviricota, or putative Nucleocytoviricota MAGs that contained a high number of contamination with GVClass were removed from further analysis, leaving 1,512 GVMAGs for further analysis. Average Nucleotide Identity (ANI) calculated using fastANI v.1.3467 to assess whether dereplication was necessary; no GVMAGs were clustered at a 96% ANI threshold, resulting in as distinct population genomes for downstream analyses. Because these genomes were recovered from a multi-year co-assembly, GVMAGs should be interpreted as metagenome-assembled population genomes rather than isolated genomes. For family-level analyses, only genomes with inferred family taxonomy were retained. Pandoravirales was excluded from family-level analyses as there was no family-level taxonomic classification identified by GVClass at the time of publication. To calculate relative abundance of GVMAGs, mapping was performed with bowtie2 v.2.5.368 and converted to bam with samtools v.1.19.269. Using inStrain v.1.9.070, a GVMAG was considered present in a sample if at least 20% of the viral genome had a sequencing depth of at least 1. This allowed us to visualize richness (the number of distinct GVMAGs) per year and the absolute number (abundance) of GVMAGs per year. To calculate abundance of GVMAGs per season, we used the bam output and raw read counts were aggregated at the family level. Only samples present with collection metadata (i.e., samples with Year, Month and Day recorded) were retained for analysis. Differential abundance analysis was performed using ANCOM-BC2 (Analysis of Compositions of Microbiomes with Bias Correction 2)71, which accounts for the compositional nature of microbiome data by modeling log-ratios of abundances rather than raw counts. ANCOM-BC2 tests for differential abundance of taxa while controlling for compositionality bias inherent in relative abundance data. The implementation included a fixed effect for seasonal periods to identify taxa that significantly change across seasons, and a random effect for collection year to account for temporal pseudoreplication in the sampling design. The ANCOM-BC2 analysis incorporated several critical parameters: a structural zero detection mechanism with a prevalence threshold of 10% (prv_cut = 0.10) to distinguish between true zeros and sampling zeros; a minimum library size filter (lib_cut = 1000) to remove low-quality samples; and a small effect size threshold (s0_perc = 0.05) to improve sensitivity. Multiple hypothesis testing correction was applied using Holm’s method to control family-wise error rate. Four complementary statistical tests were performed: (1) global tests to identify taxa that show any significant differences across all seasons, (2) pairwise tests to detect differences between specific season pairs, (3) Dunnett’s tests to compare each season against a reference season, and (4) trend tests to identify taxa showing monotonic patterns across the seasonal gradient. This comprehensive testing approach enabled detection of various patterns of temporal dynamics in microbial families. The resulting bias-corrected abundances were log10-transformed and averaged across samples within each seasonal category to generate a heatmap visualization matrix of family-level abundance patterns throughout the seasonal cycle. Phylogenetic tree building Phylogenetic trees of GVMAGs were reconstructed using nsgtree v.0.5 (https://github.com/NeLLi-team/nsgtree), which leverages MAFFT v.7.31072, IQ-TREE v.2.3.373, and GVOG7 hidden Markov models (HMMs). The marker genes included in the GVOG7 set include: superfamily II helicase (GVOGm0013), large RNA polymerase subunit (GVOGm0023), family B DNA Polymerase (GVOGm0054), TFIIB transcriptional factor (GVOGm0172), Topoisomerase family II (GVOGm0461), A32-like packaging ATPase (GVOGm0760), virus late transcription factor 3 (GVOGm0890). Tree visualization was carried out using ggtree in R74. Eukaryotic host prediction To assess host-giant virus co-occurrences, we used PolB as a marker gene for Nucleocytoviricota and the 18S rRNA gene as a marker gene for eukaryotes. For Nucleocytoviricota, we used HMMER v.3.4 and hmmsearch (-E 1e-5)75 with HMMs generated by Endo et al., 202053. Using the entire TYMEFLIES dataset, we recovered 194,957 PolB sequences, although this collection included sequences other than Nucleocytoviricota across 471 samples. The number of PolBs per sample ranged from 42 to 1342. We identified 61,426 Nucleocytoviricota PolBs using the MAFFT “addfragments” option72 and pplacer v.1.176 with the alignment and reference tree provided by Endo et al., 202053. Samples with less than 50 sum of raw abundances were removed to avoid low sequencing bias. For the 18S rRNA gene, we used Infernal v.1.1.5 and the cmsearch program77 to search the contig files for sequences matching to each of the following small subunit rRNA (SSU rRNA) models from the Rfam database78: eukaryotic 18S rRNA (RF01960), microsporidial 18S rRNA gene (RF02542), archaeal 16S rRNA gene (RF01959) and bacterial 16S rRNA gene (RF00177). In order to include long SSU rRNA sequences in the analysis, the --anytrunc option in cmsearch was used and the output parsed to find contigs in which sequential and multiple non overlapping alignments to the covariance model were found. We searched the retrieved SSU rRNA sequences against a database that contained all SSU rRNA sequences from both SILVA79 and PR2,80 databases using BLAST+ v.2.17.081. This workflow is provided here at https://github.com/NeLLi-team/ssuextract. Sequences with no hits to the database were added to a reference SSU rRNA tree to confirm their phylogenetic affiliation at the domain level.The reference tree was built as follows: SSU rRNA from the SILVA and PR2 databases were clustered at 75% identity using Usearch82 and the resulting centroids were aligned using L-INS-i in MAFFT72. Tree building was done with IQ-TREE73 with default parameters including the substitution model selection feature83. Sequences that branched within a cluster that belonged to a different domain were considered misannotations. We used identity-based sequence clustering at 97% and then at 85% to get operational taxonomic units of the SSU sequences, which had a blast best hit with 18S rRNA gene sequences and those with no blast hit but that branched within the 18SrRNA gene region of the reference tree, together with the sequences from SILVA, PR2, SSU gene sequences from eukaryotic genomes from NCBI, and long 18S rRNA gene amplicons from environmental samples84. After discarding singleton clusters we aligned the OTU representative 18S rRNA gene sequences concatenated with their respective 28S rRNA gene sequence, whenever they were extracted from the same contig, using the cmalign program with the RF01960 and RF02543 covariance models, respectively. Columns with more than 90% missing data were removed with trimAl85. We then built a tree using the IQ-TREE program with the GTR+F+R10 model and 1000 ultrafast bootstrap replicates using the aligned centroid sequences. We manually inspected the tree and built it again after discarding those sequences with terminal branches with length >0.5 and those that branched within clades dominated by sequences classified into a different eukaryotic supergroup. Finally, all environmental SSU OTU representative sequences were placed into the tree using the EPA-ng program86 and then taxonomy was assigned with the gappa software87 using the PR2 taxonomy. Using similar methods from Meng et al., 202134, we created a PolB and 18S rRNA gene matrix to perform network inference. We first applied a centered log-ratio (clr) transformation on each matrix separately, after adding a pseudo count of one to all values. We also filtered out low-abundance genes using a lower quartile filtering approach. Finally, we used a combined matrix as the input for Flashweave v.0.19.2 (sensitive = true, heterogeneous = false, alpha = 0.01)88. Following results from Flashweave, we recreated two Nucleocytoviricota PolB phylogenetic trees. The first to be used to classify PolBs at the family level, we used the GVOGm0054 marker and aligned PolB sequences with reference sequences using MAFFT and FastTree41. Any PolBs not within our families of interest (i.e., those included in Fig. 1C) were removed from visualization. The second tree was used to confirm relationships using the Taxon Interaction Mapper (TIM)89, however, these only included sequences with at least 500 amino acid length resulting in 3137 polB sequences. To estimate the strength of associations from the Flashweave output, we calculated an association score. First, we clustered Nucleocytoviricota PolB sequences using CD-HIT v.4.8.190, then used this clustering group to find consecutive associations and the max edge per Nucleocytoviricota-eukaryote pair. We only kept clusters that had at least three genomes. Our formula includes a decay factor using a decay rate from Hewson et al., 201291 for a Phycodnaviridae virus converted into decay per year. Additionally, we applied a small bonus factor of 1.1 for associations that were seen in consecutive years. We then took the max edge value at each year into consideration. The resulting formula:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sum }_{t=1}^{T}{d}^{t} * \max \left({E}_{{ij}}^{(t)}\right) * {B}^{{b}_{{ij}}(t)}$$\end{document}∑t=1Tdt*maxEij(t)*Bbij(t)where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d$$\end{document}d is the decay constant, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${E}_{{ij}}^{(t)}$$\end{document}Eij(t) is the edge between an Nucleocytoviricota and eukaryote in that year, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${B}^{{b}_{{ij}}(t)}$$\end{document}Bbij(t) is the bonus factor to the power of the number of current consecutive associations in that year. Using this weighted score, we found the median per Nucleocytoviricota-eukaryote pair (removing any pairs that only had one cluster). Finally, we visualized changes in associations in the Flashweave network output before and after the spiny water flea invasion in Lake Mendota. We implemented multiple analytical approaches to empirically determine when significant network restructuring occurred. First, we calculated annual network metrics after preprocessing our data consistently by aggregating edges to their maximum weights between each virus-host pair. To objectively identify potential abrupt shifts in network structure, we implemented a changepoint detection analysis using the Pruned Exact Linear Time (PELT) algorithm. This approach analyzes time series data to identify statistically significant points where the underlying distribution of the data changes, particularly in mean or variance. For each network metric (edges, density, and modularity), we constructed annual time series (2000–2018) and applied the PELT algorithm with a penalty parameter determined by the Bayesian Information Criterion to prevent overfitting. We removed 2019 from our analysis as it did not comprise a full calendar year of data. Our analysis did not detect statistically significant changepoints in any of the network metrics, suggesting that network restructuring occurred as a continuous process rather than through discrete shifts detectable at annual resolution. We then performed a sliding window analysis testing alternative division points between 2006 and 2012. For each potential breakpoint year, we divided our dataset into “before” and “after” periods and calculated the absolute difference in network density between periods after aggregating edges by maximum weight between each virus-host pair. To assess statistical significance, we implemented a permutation test (500 permutations) for each breakpoint. In each permutation, we randomly reassigned samples to “before” and “after” periods while maintaining the original sample sizes, recalculated network metrics, and computed the density difference. The p-value for each breakpoint represents the proportion of permutations that produced a density difference greater than or equal to the observed difference. This approach tests whether the observed network differences are statistically distinguishable from those arising by random temporal partitioning of the data. None of the tested breakpoints yielded statistically significant results, with p-values ranging from 0.162 (2009) to 0.456 (2011). This suggests that while visual inspection of network metrics indicates changes in community structure around the time of spiny water flea detection, these changes were not statistically significant to be distinguished from the background of temporal variability when analyzed at annual resolution. We caution that our network plot should be interpreted as contrasting two distinct ecological states: a pre-established and post-established invasion system. To assess the robustness of observed virus-host association changes while accounting for temporal differences in sampling intensity in the pre-invasion (2000–2009) and post-invasion (2010–2019) periods, we implemented a mixed‑effects logistic model that included yearly sequencing effort as a covariate confirmed that the probability that any virus–host pair formed an edge remained significantly higher in the post‑invasion period (Edge_presence ~ Period(pre/post) + log(Samples) + (1|Year)). We then used a bootstrapping approach with 1000 iterations to randomly choose samples from the period with the greater number of samples (i.e., post). For each iteration, we randomly sampled (with replacement) an equal number of samples from both time periods, constructed the network, and recorded which virus-host pairs were present. This generated a frequency distribution of associations, allowing us to calculate the proportion of bootstrap networks in which each association appeared. We considered associations appearing in ≥95% of bootstrap networks as high-confidence associations. This approach permitted statistical discrimination between robust ecological patterns and potential sampling artifacts, particularly for identifying associations that consistently appeared or disappeared between periods despite sampling variation. Additional network analyses were performed to ensure the robustness and statistical validity of observed shifts in virus-host associations. A bias-aware null model was used to test whether observed virus-host associations were significantly enriched or depleted compared to random expectations, we generated 5000 degree-preserving random graphs and calculated Z-scores and empirical p-values for each virus-host subdivision combination. A down-sampling (rarefaction) analysis was performed to account for sampling bias due to increased sequencing depth post-invasion. We randomly subsampled the post-invasion network to match the pre-invasion edge count, repeating this process 5000 times to ensure robustness of observed patterns. A leave-top-hubs-out robustness analysis was performed by systematically removing highly connected viral nodes to test network stability and assess whether observed structural patterns were robust to the loss of key taxa. Finally, a permutation test on year by shuffling labels of network edges to statistically confirm that structural changes in the network were not due to random temporal variation. Identification of gene duplication & HGTs To identify gene duplications, we used Orthofinder v.2.5.592 to find orthogroups and putative gene duplications in giant virus genomes, resulting in 29,649 orthogroups. Using OrthoFinder, we define an orthogroup as the set of genes descended from a single gene in the last common ancestor of the Lake Mendota genomes. Using DIAMOND and an all-versus-all sequence similarity search followed by clustering with the Markov Cluster Algorithm, we set boundaries between two orthogroups with the default MCL inflation parameter (n = 1.5). For horizontal gene transfer (HGT) identification we used BLASTp (settings: --evalue 1e-5 --id 30 --max-target-seqs 50) with diamond using a custom-made database from Lake Mendota bacterial single assembly bins as well as the NR database93. We identified HGTs as those that contained a percent of identical positions (pident) as 50%. Putative HGTs above 80% were labeled as recent HGTs as this most likely indicated a recently acquired viral gene. Finally, we considered putative de-novo genes as those that belonged to a single orthogroup (i.e., ORFan) and had no homolog to the NR database or Lake Mendota single assembly bacterial bins. It is important to note that these de-novo genes could have been transferred from other species not in this set of genomes, thus creating artifacts of ORFans. To annotate genes, we used eggNOG-mapper v.2.1.1294 with database v.5.0.2 and pfam_scan v.1.6 with a database from Nov 2019. Using results from inStrain, we considered genes as present if they contained at least one coverage and 50% breadth70. For every gene considered at absence in a sample, we checked if the surrounding genes were considered present. This was applied in order to exclude possible gene candidates that contained multiple genes on the same contig in close proximity not mapped due to insufficient mapping rather than actual gain/loss. Furthermore, we removed any genes left that contained higher mapping (>1) or those with high breadth (>50%) in the dataset prior to further analysis. We then used the years the respective giant virus was present in the data and the years the gene was present to determine if the gene could be a candidate for gene gain or gene loss. This method is very conservative as it only considered gene candidates if it passed multiple parameters. For gene gain we only considered genes that occurred at least 3 years after the first presence of the giant virus, had mapping on either side, and persisted until at least two years before the last sampling of the giant virus. For gene loss, we applied a similar process but for genes that occurred at least 3 years before the last presence of the giant virus, had mapping on either side, and persisted at least two years after the first sampling. For each putative gene gain and loss, mappings were manually checked to confirm true gain and loss. Evolutionary rates The bam files generated from the presence/absence method above were further used in Metapop v.0.0.60 for analysis of micro-diversity (settings: –id_min 95 –snp_scale both)95. Prodigal genes generated from GVClass v.1.064 were used in place of self-annotation by MetaPop95. For SNP calling, at least 70% of genome coverage and at least 10× average read depth was required. To consider the rates of substitution between bacterial, viral (primarily bacteriophages) and giant virus genomes, we used Instrain v.1.9.0 with default settings for profile70 and MetaPop for rates within coding sequences. Bacterial and viral genomes included were recovered following methods from Oliver et al., 202425. In our whole genome single nucleotide substitution comparison, we only considered genomes that contained at least a coverage of 5 in each sample and were present in at least 3 years. First, we divided the number of nucleotide substitutions by the length of the genome. We then calculated the mean of the previous value per genome and divided by the year difference between the first year and the last year of observed substitutions. We further filtered to remove genomes that did not have a difference of at least three years between the first year and the last year. In order to find rates within coding genes, we used prodigal-gv in geNomad to annotate other virus genomes63 and Prodigal v2.6.396 to annotate bacteria. For substitutions within genes, we only considered genomes that were present in at least three years, with at least three year difference between the first year and the last year and genes with a pNpS ratio above 1, as considering all genes was consistent with the inStrain results. We took the mean of the number of observed nonsynonymous substitutions divided by the gene length of each genome, and further divided this by the number of years between the first year and the last year of the genome. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Data Supplementary Data Reporting Summary Transparent Peer Review file Supplementary Information Description of Additional Supplementary Data Supplementary Data Reporting Summary Transparent Peer Review file Source data Source Data Source Data