Carbohydrate-active enzymes of soil prophages enhance global carbon cycling potential Authors: Hanpeng Liao, Chaofan Ai, Chen Liu, Hongbo Zhang, Dong Zhang, Pengfa Li, Xiang Tang, Xiaolong Liang, Ville-Petri Friman, Manuel Delgado-Baquerizo, Shungui Zhou Published: 2026 ABSTRACT Recent work suggests that soil-borne viruses play an important role in controlling carbon (C) cycling and stocks. However, the contribution of individual prophage (i.e., temperate phages residing within bacterial hosts during lysogenic cycle) to C degradation remains largely undocumented at global scale. Here, we generated a global gene catalog of prophage-encoded carbohydrate-active enzyme (pCAZymes), including 20,131 soil bacterial genomes, 3548 metagenomes, and 951 metatranscriptomes derived from pre-existing databases. The catalog includes 4708 pCAZymes associated with the degradation of lignocellulose, lignin, and pectin, with 21 lytic polysaccharide monooxygenase genes newly identified in phages. Our findings reveal that prophages have potential to accelerate labile soil C degradation by encoding pCAZymes that cooperate with their bacterial hosts. Using machine learning models, we predict a 13 ± 0.7% increase in the C metabolic potential driven by soil prophages by 2100 under a high-emission scenario (SSP585). In vitro experiments demonstrated that the transcriptional activity of pCAZyme genes is regulated by environmental temperature. Soil microcosm experiments further confirmed that pCAZymes can enhance host-mediated organic C mineralization by increasing degradative enzyme activity. This study reveals previously overlooked ecological functions of prophages in global soil C transformation, with important implications for the global climate and C cycling. Introduction Soils are the largest terrestrial C pool, containing an estimated 2500 ± 120 Gt of C—about 3.3 times more than the atmosphere1. Soil C stocks and pools provide multiple ecosystem services, including climate change regulation and soil fertility, which together sustain approximately 95% of global food production2. The transformation of C within soil systems is primarily driven by the soil microbiome, with bacteria and fungi serving as important contributors through two major pathways3. First, microorganisms secrete large amounts of extracellular enzymes (e.g., cellulases, hemicellulases, and chitinases) to hydrolyze lignocellulosic polymers into bioavailable sugars and resulting in significant soil organic matter decomposition. Second, these sugars are assimilated to support microbial growth; upon cell death, microbial biomass is transformed into soil organic C pools4. Microorganisms account for >90% of litter decomposition and biomass mineralization through secretion of carbohydrate-active enzymes (CAZymes)5,6. CAZymes are grouped into glycoside hydrolases (GHs), glycosyltransferases (GTs), carbohydrate esterases (CEs), polysaccharide lyases (PLs), auxiliary activities (AAs), and carbohydrate-binding modules (CBMs) that together regulate the efficiency of labile C degradation and conversion into microbial C. Microbial residues (e.g., fungal chitin, bacterial peptidoglycan) contribute 50–80% of stable soil organic matter7. While the role of bacteria for C turnover has been studied extensively via the microbial C pump6,8, the role of viruses in shaping soil organic matter degradation and C losses is less clear9–12, which limits our capacity to predict soil C cycling globally. Bacteriophages (phages) are the most abundant biological entities on Earth13, with soils harboring 10⁷–10¹⁰ viral particles per gram14. Based on their life cycles, phages are generally classified into two main categories: virulent phages and temperate phages. Virulent phages replicate via the lytic cycle, where they replicate and lyse their hosts to release viral progeny. In contrast, temperate phages integrate into bacterial genomes as prophages and can be transmitted vertically during the lysogenic cycle, which is often activated into lytic cycle under stress (e.g., environmental stress and DNA damage)15. More than half of bacterial genomes harbor prophages in the NCBI sequenced bacterial genome databases, underscoring their ubiquity and importance for the ecological functions16. While virulent phages are well known to contribute to C-cycling by releasing bacterial C through lysis17 (a process commonly referred to as the “viral shunt”), the effects of prophages on soil C turnover remain much less understood18. Prophages may influence soil C fluxes through host lysis but also by carrying auxiliary metabolic genes (AMGs) associated with C turnover19. In marine ecosystems, viruses lyse bacterial hosts and release approximately 145 Gt of C annually, accounting for 20–40% of prokaryotic biomass C20. Similar processes likely occur in soils as soil harbors abundant bacterial populations; viral lysis releases dissolved organic matter, thereby alleviating microbial nutrient limitation14. Moreover, prophage-encoded C-associated AMGs can enhance host C metabolism during the lysogenic cycle21. Although previous studies have inferred links between soil viruses and C metabolism based on the presence of AMGs, direct experimental evidence demonstrates that a single temperate phage can causally contribute to soil C transformation remains lacking. While the ability of temperate phages to alternate between lysogeny and lysis may confer more complex roles in C cycling than that of obligate virulent phages18. Although previous studies have suggested potential roles of viruses in nutrient transformation based on local farmland soil samples10,19,22, a systematic understanding of their functions in global soils remains lacking. First, the diversity and geographic distribution of CAZyme genes encoded by prophages have not yet been comprehensively characterized at the global scale. Second, direct experimental evidence demonstrating whether these AMGs actively participate in soil organic C transformation is still limited. Finally, the response dynamics of prophage-mediated C metabolism potential under future climate change scenarios remains largely unexplored. Here, we combined global genomic, metagenomic, and metatranscriptomics surveys with in vitro experiments to advance our understanding on the contribution of prophages to soil C cycling. We first integrated 20,131 publicly available bacterial genomes and metagenome-assembled genomes (MAGs) to construct the global database of prophage-encoded CAZymes (pCAZymes). Using global soil metagenomic and metatranscriptomic datasets, we quantified the C metabolic potential of prophages based on the community composition of prophages and their hosts, and the gene abundance of pCAZyme in global soils, and found that the profiles of pCAZyme genes are associated with global fast-turnover soil C and total C content. In addition, we conducted soil microcosm experiments using a temperate phage strain engineered to carry pCAZyme genes as an inoculant to evaluate whether these AMGs can directly influence soil C transformation. Predictive models under future climate scenarios were further used to explore how prophage-associated C metabolic potential might respond to climate change and impacts soil C stocks. Under a fossil fuel-dependent projection (SSP585), machine learning modeling suggest a 13 ± 0.7% increase in prophage-mediated soil organic C metabolism by 2100, likely accelerating soil C loss. This work provides the global-scale catalog of pCAZymes with the potential to regulate soil C cycling under climate change and offers direct experimental evidence that AMGs in prophage can influence soil C transformation. Results and discussion Construction of a global catalog of prophage-encoded CAZymes genes We first integrated 20,131 high-quality bacterial genomes—including 11,176 MAGs—from public repositories from global terrestrial habitats, spanning 42 bacterial phyla and 1273 genera (Supplementary Data 1). Pseudomonadota (64.1%), Bacillota (13.6%), Bacteroidota (12.8%), and Actinomycetota (7.0%) were dominant bacteria, together accounting for 97.5% of all genomes (Fig. 1A). We detected 51,611 putative prophages in 13,562 lysogens, yielding 21,626 non-redundant prophages clustered at 95% sequence similarity (Supplementary Data 2). Prophage genome sizes ranged from 5 to 445 kb, with 72% genomes classified as medium-high quality (Fig. S1), and Hendrixvirinae (21.7%), Peduoviridae (16.0%), and Peduoviridae-like (12.6%) were the most abundant taxa (Fig. 1B). Prophage host ranges spanned 42 bacterial phyla and 1149 genera, with Pseudomonadota, Bacillota, Actinomycetota, and Bacteroidota being the dominant hosts and together accounting for 96.8% of genomes (Fig. 1C and Supplementary Data 1). The number of prophages within lysogenic bacteria ranged from 1 to 23, with an average of 1.6 prophages per lysogen (Fig. 1C). These results suggest that prophages are widespread in soil bacterial communities and that their distributions are strongly shaped by host taxonomic lineages.Fig. 1Summary of the prophage-encoded CAZyme genes in the database.A Number of dominant bacterial genomes categorized by phylum. B Number of dominant prophage genomes categorized by family. C The proportion of dominant lysogenic bacteria categorized by phylum, with the red solid line indicating the average number of prophages per genome within each phylum and the red dashed line denoting the overall average across all phyla. D Number of CAZyme genes encoded by the top 10 prophages, with the red dashed line showing the average number of CAZyme genes encoded across all prophages. E Number of different type CAZyme genes in prophage genomes across samples. F Number of pCAZyme genes related to cellulose-, hemicellulose-, lignin-, and pectin-degradation in the database. Summary of the prophage-encoded CAZyme genes in the database. A Number of dominant bacterial genomes categorized by phylum. B Number of dominant prophage genomes categorized by family. C The proportion of dominant lysogenic bacteria categorized by phylum, with the red solid line indicating the average number of prophages per genome within each phylum and the red dashed line denoting the overall average across all phyla. D Number of CAZyme genes encoded by the top 10 prophages, with the red dashed line showing the average number of CAZyme genes encoded across all prophages. E Number of different type CAZyme genes in prophage genomes across samples. F Number of pCAZyme genes related to cellulose-, hemicellulose-, lignin-, and pectin-degradation in the database. To evaluate prophage potential contribution to soil C turnover, we annotated prophage-encoded AMGs related to carbohydrate-active enzymes using the CAZy database23. In total, 85.8% (18,564/21,626) of prophages encoded at least one CAZyme gene, yielding 64,678 pCAZyme genes, with an average of 3.5 genes per prophage (Supplementary Data 3). A key challenge in AMGs identification is distinguishing true viral genes from misclassified host functional genes24. To minimize this risk, the presence of viral structural or hallmark genes flanking both ends of an AMG is widely considered a robust criterion for reliable identification24. In our dataset, 89.3% of identified AMGs were flanked by viral genes at both ends, while 10.7% were flanked at only one end (Fig. S2). More than 99% of the identified AMGs were associated with viral genomic regions, strongly supporting their viral origin and indicating a very low probability of host gene contamination24. However, the identification of AMGs remains challenging, as no standardized procedure currently guarantees complete accuracy24. On average, temperate phages carried more CAZyme-related AMGs than obligate virulent phages (5.2 versus 1.7 per genome, Fig. S3A). This may be because temperate phages generally have longer genomes than virulent phages (p = 3.2e-56, Fig. S3B), allowing them to carry a greater number of genes25. Many virome studies rely on viral particle extraction, which preferentially captures free virions and may therefore bias detection toward actively virulent viruses26,27. In contrast, our genome-resolved approach enables the identification of prophages integrated within host genomes, thereby revealing hidden components in soil viral communities. As integrative elements, temperate phages often retain more mobile genetic components that enhance host adaptation, thereby forming a mutually beneficial symbiotic relationship with their hosts25,28. Supporting this, the number of CAZyme-related AMGs in prophages was positive significantly correlated with the CAZyme gene content of their hosts (Fig. S4, R2 = 0.0169, p < 2.2e-16), suggesting that the pCAZyme genes may acquire from hosts via horizontal gene transfer. To assess whether pCAZymes are likely to be secreted extracellularly via signal peptides, we used SignalP to predict the presence of signal peptide sequences29. Signal peptides were identified in 5% (3261 of 64,678) of pCAZymes (Fig. S5A, B), suggesting most CAZymes are likely intracellular. Despite this, intracellular hydrolases can still contribute to soil C transformation through released into soil environment upon host lysis30. The pCAZyme content varied considerably across viral taxa: Triavirus (5.0 genes per prophage), Azeredovirinae (4.4), Peduoviridae (4.1), Sepvirinae (4.0), and Hendrixvirinae (4.0) harbored significantly higher number of genes than average (1.6 genes per prophage, Fig. 1D), indicating lineage-specific enrichment of C metabolism genes through host interactions. Functionally, pCAZymes spanned six enzyme categories, including GHs, GTs, CBMs, PLs, CEs, and AAs. Among these, the majority (77%) belonged to the GHs category, which is associated with plant biomass degradation, followed by the GTs category (10%, Fig. 1E). All pCAZyme categories previously reported in prokaryotes could be detected in prophages31, suggesting that prophages are a previously underappreciated reservoir of CAZymes. We next focused on pCAZymes involved in lignocellulose degradation, a critical pathway for soil organic C turnover. We identified 7.3% pCAZyme genes (4706 of 64,678) linked to cellulase, hemicellulase, lignin oxidase, and pectinase across 125 CAZyme families (Fig. 1F and Supplementary Data 4). Specifically, 2575 pCAZyme genes associated with cellulose degradation were detected across 12 GH families, 3 AA families, and 2 CBM families (Supplementary Data 5). These included diverse cellulolytic enzymes such as GH51 endo-beta-1,4-glucosidase, GH3 beta-glucosidase, and GH9 cellobiohydrolase. Notably, 12 prophages also carried 21 AA10 lytic polysaccharide monooxygenases (LPMOs)32. LPMOs are recently identified key enzymes capable of oxidatively cleaving crystalline cellulose in fungi and bacteria33,34. However, their occurrence and functional roles in viruses have not been previously reported. To the best of our knowledge, this study provides the first report of LPMO-encoding genes in viral genomes. For hemicellulose degradation, a total of 1748 pCAZyme genes were identified across 22 GH families and 9 CE families, which were dominated by GH113 β-mannanase, GH35 beta-galactosidase, and GH36 alpha-galactosidase (Supplementary Data 5). Additionally, 97 pCAZyme genes related to lignin degradation were identified across 5 AA families, including AA2 manganese peroxidases, AA5 aryl alcohol oxidases, and AA6 benzoquinone reductases, indicative of their role in oxidative–reductive lignin decomposition (Supplementary Data 5). Together, these results reveal that prophages encode a broad repertoire of pCAZymes—particularly GHs and AAs—that may facilitate the breakdown of plant litter decomposition. Because pCAZymes (e.g., cellulase, hemicellulase, and pectinase) are mainly associated with the degradation of utilizable C, this suggests that prophages may primarily contribute to the transformation of labile C in soil. Prophage-encoded CAZyme genes are widely distributed and expressed in global soils Based on 3548 soil metagenomes, we investigated the distribution of pCAZyme genes across different terrestrial ecosystems globally (agricultural soil (n = 802), artificial surfaces (n = 406), forests (n = 594), grasslands (n = 725), Fig. S6, Supplementary Data 6 and 7). Overall, 65.6% of pCAZyme genes (42,429 of 64,678) could be detected across all soil metagenomic samples, and an average of 149 pCAZyme genes were detected per metagenomic sample. Across all metagenomes, the detection rates of pCAZymes genes varied among different functional categories (Fig. 2A). For example, the genes within GHs, GTs, and CBMs had over 95% detection rates on average across all samples, whereas CEs, PLs, and AAs exhibited relatively lower detection rates (31 to 71%). GHs enzymes involved in cellulose and hemicellulose degradation—such as GH3 β-glucosidase—was particularly prevalent in global soils with 99.5% detection rate (Fig. S7). The β-glucosidase catalyzes the final step of cellulose degradation to release the glucose that is widely distributed within soil bacteria and fungi, playing a crucial role in soil C turnover31. Moreover, the distribution of pCAZyme genes in global soils exhibited clear soil type specificity (Fig. 2B). For example, the Shannon diversity index of pCAZyme genes was strongly influenced by soil type and agricultural and forest soils exhibited significantly higher Shannon diversity than other soil types (p = 5.55e-55, Fig. 2B), whereas wetlands showed the lowest gene diversity. Forest soils also displayed the highest total gene abundance of pCAZymes (p = 1.13e-43, Fig. S8), dominated by GHs and CEs families, while wetlands had the lowest total gene abundance (Fig. S9). This may be attributed to the relatively lower lignocellulose degradation rate in wetland compared to terrestrial soils35. The total abundance of pCAZyme genes and Shannon diversity was significantly correlated with the latitude (R² = 0.025, p < 2.2e-16, n = 3544 and R² = 0.019, p = 5.36e-16, n = 3544, respectively; Fig. 2C), suggesting that latitude could be a factor structuring their global distribution. At the continental scale, soils from Southeast Asia, Central North America, Northern South America, and Southern Africa exhibited higher total pCAZyme gene abundances (Fig. 2D). This finding is consistent with previous research on the global distribution of soil viruses, in which latitude has been identified as the dominant influencing factor36.Fig. 2Global soil biogeography of prophage-encoded CAZyme genes.A The detection rate of pCAZymes across all soil samples. B Shannon diversity index of pCAZyme genes across different soil types, including agricultural land (n = 802), grassland (n = 725), forest (n = 594), wetland (n = 488), and artificial surfaces (n = 406). Differences among multiple groups were assessed using the Kruskal–Wallis H test (p = 5.55e-55), followed by pairwise two-sided Wilcoxon rank-sum tests. Different lowercase letters denote statistically significant differences at p < 0.05. Box plots show the 25th–75th percentiles, whiskers indicate the minimum and maximum values, and the midline represents the median. C The relationship between latitudinal distribution of the Shannon index (left panel) and gene abundance (right panel) of pCAZymes across global soils. The colors represent the annual mean air temperatures (MAT) of sampled locations. The line shows the second-order polynomial fit based on ordinary least squares regression (Two-sided F-tests, Left panel, R2 = 0.025, F = 46.44, p < 2.2e-16, AIC = 11497.63, n = 3544. Right panel, R2 = 0.0191, F = 35.51, p < 5.361e-16, AIC = 87971.67, n = 3544). Shaded areas represent the 95% confidence intervals. Letter n represents the number of samples. D Global soil biogeography of pCAZyme genes abundance across all soil samples. Map image is the intellectual property of Esri and is used herein under license. Copyright © 2025 Esri and its licensors. All rights reserved. Data sources: Esri, TomTom, Garmin, FAO, NOAA, USGS. E A three-dimensional structure diagram of a representative phage-encoded auxiliary metabolic genes associated with CAZymes including GH3 β-glucosidase, GH113 β-mannanases, GH36 α-galactosidases, AA10 lytic polysaccharide monooxygenases. F A schematic diagram of a representative phage-encoded auxiliary metabolic genes associated with CAZymes. G The gene abundance of pCAZymes related to cellulases, hemicellulases, and ligninases across different soil types. Global soil biogeography of prophage-encoded CAZyme genes. A The detection rate of pCAZymes across all soil samples. B Shannon diversity index of pCAZyme genes across different soil types, including agricultural land (n = 802), grassland (n = 725), forest (n = 594), wetland (n = 488), and artificial surfaces (n = 406). Differences among multiple groups were assessed using the Kruskal–Wallis H test (p = 5.55e-55), followed by pairwise two-sided Wilcoxon rank-sum tests. Different lowercase letters denote statistically significant differences at p < 0.05. Box plots show the 25th–75th percentiles, whiskers indicate the minimum and maximum values, and the midline represents the median. C The relationship between latitudinal distribution of the Shannon index (left panel) and gene abundance (right panel) of pCAZymes across global soils. The colors represent the annual mean air temperatures (MAT) of sampled locations. The line shows the second-order polynomial fit based on ordinary least squares regression (Two-sided F-tests, Left panel, R2 = 0.025, F = 46.44, p < 2.2e-16, AIC = 11497.63, n = 3544. Right panel, R2 = 0.0191, F = 35.51, p < 5.361e-16, AIC = 87971.67, n = 3544). Shaded areas represent the 95% confidence intervals. Letter n represents the number of samples. D Global soil biogeography of pCAZyme genes abundance across all soil samples. Map image is the intellectual property of Esri and is used herein under license. Copyright © 2025 Esri and its licensors. All rights reserved. Data sources: Esri, TomTom, Garmin, FAO, NOAA, USGS. E A three-dimensional structure diagram of a representative phage-encoded auxiliary metabolic genes associated with CAZymes including GH3 β-glucosidase, GH113 β-mannanases, GH36 α-galactosidases, AA10 lytic polysaccharide monooxygenases. F A schematic diagram of a representative phage-encoded auxiliary metabolic genes associated with CAZymes. G The gene abundance of pCAZymes related to cellulases, hemicellulases, and ligninases across different soil types. We further investigated the gene abundance of pCAZymes associated with plant biomass degradation in soils globally. Overall, plant biomass-degrading pCAZymes (targeting cellulose, hemicellulose, lignin, and pectin) accounted for 31% of the total abundance of all genes (Fig. S10A). Among these plant biomass-degrading pCAZymes, cellulose-degrading enzymes were predominant, representing 95% of the total abundances (Fig. S10B). Across all samples, GHs—a class of enzyme primarily involved in cellulose and hemicellulose degradation37—exhibited the highest abundances (accounting for 98% total abundance), whereas other enzymes such as AAs and PLs, primarily involved in lignin degradation, accounted for only 2% (Fig. S11A). This likely reflects differences in substrate bioavailability: cellulose is more abundant and readily degradable than lignin in soils, driving microorganisms to evolve a greater diversity of GHs38. In particular, GH3 β-glucosidases has highest abundance in global soils, comprising 92% of the total abundance of cellulase genes (Fig. S10C). Prophage-encoded LPMO genes were particularly abundant in agricultural and wetland soils (Fig. 2E–G and Fig. S11B), indicative of a more prominent role for LPMO enzymes in these ecosystems. Compared to cellulases, prophages encode a lower number of hemicellulase genes; nevertheless, these genes remain detectable at appreciable abundances in soils. Notably, prophage-encoded GH113 β-mannanases and GH36 α-galactosidases were the most abundant in soils globally (Fig. 2G). Soil type shaped plant biomass-degrading pCAZyme genes distribution, with agricultural and forest soils exhibiting higher gene abundances compared to others (Fig. S12). This is consistent with previous studies that soil type strongly influences the composition of bacteria and virus communities36,39. To identify potential underlying factors that shape the distribution of pCAZymes in soils, we examined their relationships with prophage and host community α-diversity (Shannon index) and environmental factors. Principal coordinate analysis (PCoA) analysis (based on Bray-Curtis distances) revealed that the profile of pCAZyme genes was significantly influenced by temperature, precipitation, and soil properties including soil cation exchange capacity (CEC) and pH, moisture, and total nitrogen and C content (all p < 0.05). Especially, the soil C content was the important predictor of global pCAZyme genes distribution (Fig. 3A). The Shannon diversity of pCAZyme genes in global soils was influenced by soil properties and climatic environmental factors (Fig. S13). Since prophages act as carriers of pCAZymes, we further assessed the factors influencing the global distribution of prophages. The Shannon diversity of prophage communities in global soils was shaped by both soil physicochemical properties and environmental factors (Fig. S14). Furthermore, the total gene abundance of pCAZymes (R2 = 0.056, p = 1.3e-46) and the diversity of prophages (R2 = 0.11, p = 7.7e-92) in global soils were both significantly correlated with host Shannon diversity (Fig. 3D), indicating that phage host communities are an important factor shaping the distribution of pCAZyme genes. Moreover, the total abundance of pCAZyme genes was negatively correlated with total C content (R² = 0.017, p = 8.2e-15) but positively correlated with labile C content (R² = 0.002, p = 0.015; Fig. 3B, C), suggesting that pCAZymes may primarily contribute to the metabolism of labile C in soils. Plant-derived polysaccharide substrates in soil can be rapidly decomposed by microbial hydrolytic enzymes, leading to CO₂ production and a reduction in total soil C content40,41. However, this interpretation is based on correlative evidence and remains speculative, requiring further experimental validation.Fig. 3The relationship between pCAZyme genes and environmental factors, including soil physicochemical parameters.A The PCoA ordination (based on envfit analysis, 999 permutations) shows the relationship between total abundance of pCAZyme and environmental factors. B The heatmap shows the correlation between dominant prophage and the gene abundance of pCAZymes and environmental factors. Statistical significance was determined using two-sided Pearson correlation test (*p < 0.05, **p < 0.01, and ***p < 0.001). C The relationship between the soil fast turnover carbon and total soil carbon content and the gene abundance of pCAZymes based on two-sided Pearson correlation (upper panel R2 = 0.002, p = 0.015; lower panel Pearson R2 = 0.017, p = 8.2e-15). D The relationship between the host bacterial Shannon index and the Shannon index of prophages (R2 = 0.11, p = 7.7e-92) and the abundance of pCAZyme genes (R2 = 0.056, p = 1.3e-46) based on two-sided Pearson correlation. E Transcriptional activity proportion of different types of pCAZyme genes across global soil. F The transcriptional activity of pCAZymes across different soil types (agricultural land (n = 258), grassland (n = 72), forest (n = 127), wetland (n = 296), and artificial surfaces (n = 7)). Differences among groups were assessed using Kruskal–Wallis and pairwise two-sided Wilcoxon rank-sum tests (p = 9.07e-7), different lowercase letters indicate significant differences (p < 0.05). Box plots show the 25th–75th percentiles, whiskers indicate the minimum and maximum values, and the midline represents the median. G The relationship between latitudinal distribution and gene transcriptional activity of pCAZyme genes across global soils. The Colors represent the annual mean air temperatures (MAT) of sampled locations. The line shows the second-order polynomial fit based on ordinary least squares regression (Two-sided F-tests, R2 = 0.2151, F = 135, p < 2.2e-16, AIC = 24743.67, n = 947). In (C), (D), and (G) panels, shaded regions in all linear correlation plots denote the 95% confidence intervals. The abbreviations in (A) and (B) panels are following: soil temperature (Ts), near-surface air temperature (Tas), precipitation (Pr), soil moisture (Mrso), total soil nitrogen content (Nsoil), total soil C content (Csoil), fast turnover C content (Csoilfast) and slow turnover C content (Csoilslow), and cation exchange capacity (Cec), prophage Shannon index (Psi), bacterial Shannon index (Bsi), total abundance of pCAZymes genes (Tpg) and Shannon index of pCAZymes genes (Sipg). The relationship between pCAZyme genes and environmental factors, including soil physicochemical parameters. A The PCoA ordination (based on envfit analysis, 999 permutations) shows the relationship between total abundance of pCAZyme and environmental factors. B The heatmap shows the correlation between dominant prophage and the gene abundance of pCAZymes and environmental factors. Statistical significance was determined using two-sided Pearson correlation test (*p < 0.05, **p < 0.01, and ***p < 0.001). C The relationship between the soil fast turnover carbon and total soil carbon content and the gene abundance of pCAZymes based on two-sided Pearson correlation (upper panel R2 = 0.002, p = 0.015; lower panel Pearson R2 = 0.017, p = 8.2e-15). D The relationship between the host bacterial Shannon index and the Shannon index of prophages (R2 = 0.11, p = 7.7e-92) and the abundance of pCAZyme genes (R2 = 0.056, p = 1.3e-46) based on two-sided Pearson correlation. E Transcriptional activity proportion of different types of pCAZyme genes across global soil. F The transcriptional activity of pCAZymes across different soil types (agricultural land (n = 258), grassland (n = 72), forest (n = 127), wetland (n = 296), and artificial surfaces (n = 7)). Differences among groups were assessed using Kruskal–Wallis and pairwise two-sided Wilcoxon rank-sum tests (p = 9.07e-7), different lowercase letters indicate significant differences (p < 0.05). Box plots show the 25th–75th percentiles, whiskers indicate the minimum and maximum values, and the midline represents the median. G The relationship between latitudinal distribution and gene transcriptional activity of pCAZyme genes across global soils. The Colors represent the annual mean air temperatures (MAT) of sampled locations. The line shows the second-order polynomial fit based on ordinary least squares regression (Two-sided F-tests, R2 = 0.2151, F = 135, p < 2.2e-16, AIC = 24743.67, n = 947). In (C), (D), and (G) panels, shaded regions in all linear correlation plots denote the 95% confidence intervals. The abbreviations in (A) and (B) panels are following: soil temperature (Ts), near-surface air temperature (Tas), precipitation (Pr), soil moisture (Mrso), total soil nitrogen content (Nsoil), total soil C content (Csoil), fast turnover C content (Csoilfast) and slow turnover C content (Csoilslow), and cation exchange capacity (Cec), prophage Shannon index (Psi), bacterial Shannon index (Bsi), total abundance of pCAZymes genes (Tpg) and Shannon index of pCAZymes genes (Sipg). To evaluate the potential activity of pCAZymes in soils, we mapped all pCAZyme sequences against 951 globally distributed soil metatranscriptomes (Supplementary Data 8 and 9). Compared with metagenomic data, global soil metatranscriptomic datasets are relatively limited, primarily from Asia, North America, and Europe. Across all samples, 13.8% of pCAZyme genes (8931 of 64,712) exhibited transcriptional activity. The proportion of transcriptionally active genes was substantially lower than their detection rate in metagenomic datasets, indicating that many genes were either not expressed or remained below the detection threshold. This is expected, as previous studies have shown that the AMGs within prophage genomes are selectively expression42, and the reason for their selective expression in soil remains unknown. Among different pCAZyme categories, GTs displayed the highest transcriptional activity (55%), followed by GHs (37%), while other enzymes showed relatively lower activity around 7% (Fig. 3E). The transcriptional activity of GTs is higher than that of GHs, possibly because GTs are more involved in fundamental metabolism processes, whereas GHs primarily function in the degradation of polysaccharide substrates43. We found that the total transcriptional activity of pCAZyme genes accounts for approximately 5% of the corresponding bacterial hosts in global soils (Fig. S15). Although this proportion appears relatively low, viruses are typically two orders of magnitude more abundant than bacteria in soils14, suggesting that their cumulative contribution should not be overlooked. Transcriptional activity varied significantly among soil types, with agricultural and forest soil showing the highest activity and grasslands the lowest (p = 9.07e-07, Fig. 3F). Geographically, pCAZyme transcriptional activity was greatest in East and Southeast Asia, followed by Western Europe, with lower expression observed in other regions (Fig. S16). Notably, total transcriptional activity was significantly correlated with latitude (R² = 0.2151, p = 2.2e-16, n = 947), with mid- and low-latitude soils showing markedly higher activity than high-latitude soils (Fig. 3G). For pCAZymes associated with lignocellulose degradation, 74% of pCAZyme genes exhibited transcriptional activity in soils globally (Fig. S17), with GH3 β-glucosidase and GH35 β-galactosidase being the most highly expressed pCAZymes. Although this study incorporated as much publicly available soil metagenomic and metatranscriptomic data as possible, the majority of samples originated from North America (56.0%), Asia (18.9%), and Europe (11.4%), limiting their global representativeness. Furthermore, variations in sampling strategies and time across different samples—such as differences in soil sampling depth—may influence the observed composition and activity of soil microbial communities. These results demonstrate that pCAZymes are widely distributed and transcriptionally active in soils globally, and that their global distribution is strongly shaped by soil properties and the community composition of prophages and their hosts. Prophages hold potential for global soil C loss via accelerating labile C turnover by encoding diverse pCAZymes We next classified all available soil samples with metagenomic data based on their prophage-mediated C degradation potential to identify which samples may be more vulnerable to prophage-driven C loss. As the prophage hosts in the soil could impact the soil C transformation, we integrated pCAZyme gene abundances and the diversity of phages and their hosts, to comprehensively evaluate the soil C transformation potential (CTP) driven by prophages. Across 3548 global soil metagenomes, PCoA revealed that all samples could be clearly separated into three distinct groups based on above parameters: CTP-1, CTP-2, and CTP-3 (Fig. 4A). 82.0% soil samples (n = 2912) were assigned to CTP-2 group, while only 18% samples were assigned to CTP-3 (n = 189) and CTP-1 (n = 447) groups. Total pCAZyme gene abundance differed significantly between CTP groups (p = 1.16e-231), with CTP-3 showing relatively higher abundance compared to the two other groups (Fig. 4B). There were also significant differences in enzyme composition of pCAZymes between the groups, while GHs (55.6–59.3%) and GTs (33.3–41.9%) were the dominant enzymes in three CTP groups (Fig. S18). For example, the genes associated with cellulose and hemicellulose degradation were 12.7 times higher in CTP-3 compared to CTP-1 group (p = 1.15e-14 and 2.11e-28, Fig. S19). This likely explains CTP-3 stronger carbohydrate-degrading capacity compared to other enzyme types44. Moreover, the composition and diversity of temperate phages were significantly different between CTP groups (p = 1.91e-276, Fig. 4C), with CTP-3 group displaying the highest Shannon diversity of temperate phages and their bacterial hosts compared to the other two groups (Fig. 4C and Fig. S20). Geographical distribution analysis revealed that the CTP-2 group was the most prevalent across all soil types (76.6–88.7%), while the CTP-1 group was enriched in agricultural soils (11.7%) and the CTP-3 group was primarily distributed in wetlands (21.5%) and grasslands (17.5%, Fig. S21). The CTP-3 group was primarily distributed in mid-to-high latitudes of the Northern Hemisphere, such as Southeast Asia and Central North America (Fig. S22). Overall, these results indicate that pCAZyme gene abundances, along with the community composition of prophage and their hosts, can jointly serve as effective indicators for assessing the potential of phage-mediated C transformation across soil types globally.Fig. 4Prophages are associated with global soil carbon loss via accelerating labile carbon turnover.A PCoA shows the distinct patterns of carbon transformation potential (CTP) of soil prophage across all samples. The global soils were classified into three clusters (CTP-1, CTP-2, and CTP-3) following dimensionality reduction followed by clustering analysis based on 18 variables. The point colors represent the categories of soil prophage carbon metabolic potential, while the shapes denote the type of soil samples. The box plot shows the differences in PC1 (p = 1.931e-285) and PC2 (p = 4.13e-54) for different types of carbon transformation potential of prophages. B Total abundance of pCAZyme genes in different CTP groups across global soils (p = 1.16e-231). C The Shannon index of prophages in different CTP groups across global soils (p = 1.91e-276). D The soil carbon transformation potential of prophages index in different CTP groups across global soils (p = 9.03e-291). E The global distribution of the prophage-driven carbon transformation potential index across global soils. Map image is the intellectual property of Esri and is used herein under license. Copyright © 2025 Esri and its licensors. All rights reserved. Data sources: Esri, TomTom, Garmin, FAO, NOAA, USGS. F Negative relationship between the prophage-driven carbon transformation potential index and soil total carbon content across global soils (R2 = 0.0484, p = 1.6e-38). G Positive relationship between the prophage-driven carbon transformation potential index and the soil fast turnover carbon content across global soils (R2 = 0.003969, p = 0.00017). In (A–D), n = 447 for CTP-1, n = 2911 for CTP-2, and n = 189 for CTP-3, where n represents the number of independent samples. Differences among groups were assessed using Kruskal–Wallis and pairwise two-sided Wilcoxon rank-sum tests; different lowercase letters indicate significant differences (p < 0.05). Box plots show the 25th–75th percentiles, whiskers indicate the minimum and maximum values, and the midline represents the median. In (F) and (G), the solid lines represent linear regression fits, and the shaded areas indicate the 95% confidence intervals. Prophages are associated with global soil carbon loss via accelerating labile carbon turnover. A PCoA shows the distinct patterns of carbon transformation potential (CTP) of soil prophage across all samples. The global soils were classified into three clusters (CTP-1, CTP-2, and CTP-3) following dimensionality reduction followed by clustering analysis based on 18 variables. The point colors represent the categories of soil prophage carbon metabolic potential, while the shapes denote the type of soil samples. The box plot shows the differences in PC1 (p = 1.931e-285) and PC2 (p = 4.13e-54) for different types of carbon transformation potential of prophages. B Total abundance of pCAZyme genes in different CTP groups across global soils (p = 1.16e-231). C The Shannon index of prophages in different CTP groups across global soils (p = 1.91e-276). D The soil carbon transformation potential of prophages index in different CTP groups across global soils (p = 9.03e-291). E The global distribution of the prophage-driven carbon transformation potential index across global soils. Map image is the intellectual property of Esri and is used herein under license. Copyright © 2025 Esri and its licensors. All rights reserved. Data sources: Esri, TomTom, Garmin, FAO, NOAA, USGS. F Negative relationship between the prophage-driven carbon transformation potential index and soil total carbon content across global soils (R2 = 0.0484, p = 1.6e-38). G Positive relationship between the prophage-driven carbon transformation potential index and the soil fast turnover carbon content across global soils (R2 = 0.003969, p = 0.00017). In (A–D), n = 447 for CTP-1, n = 2911 for CTP-2, and n = 189 for CTP-3, where n represents the number of independent samples. Differences among groups were assessed using Kruskal–Wallis and pairwise two-sided Wilcoxon rank-sum tests; different lowercase letters indicate significant differences (p < 0.05). Box plots show the 25th–75th percentiles, whiskers indicate the minimum and maximum values, and the midline represents the median. In (F) and (G), the solid lines represent linear regression fits, and the shaded areas indicate the 95% confidence intervals. Based on the variance contributions of PC1 axis in the PCoA analysis, we developed a quantitative metric (“C transformation potential index”, CTPI for short, see details in “Methods”) to assess the soil C transformation potential mediated by prophages in each soil sample. The CTPI differed significantly among the three CTP groups (CTP-3 > CTP-2 > CTP-1, p = 9.03e-291; Fig. 4D), indicating that CTPI may serve as an indicator of the phage-driven C metabolic potential in soils. The CTPI values were significantly correlated with latitude (R2 = 0.01477, p = 1.304e-12, n = 3544, Fig. S23), with high CTPI samples mainly distributed in Asia, southern North America, and southern Europe (Fig. 4E). Given previous results that GHs constitute the main component of pCAZymes, we focused on the relationship between GHs gene abundance and CTPI. The abundance of GHs genes in the CTP-3 group was significantly higher than in the CTP-1 and CTP-2 groups (p = 2.17e-228, Fig.S24A). Moreover, GHs gene abundance was significantly correlated with CTPI values (R = 0.5929, p < 2.2e-16; Fig. S24B), indicating that GHs enzymes are key contributors to soil C turnover. Compared to CTP-1 and CTP-2 groups, CTP-3 group soils exhibited significantly lower total C content (p = 5.22e-45, Fig. S25), likely due to their higher CTPI, which may promote greater C loss and microbial degradation of lignocellulose-derived soil C. Supporting this, CTPI in global soils was significantly negatively correlated with soil total C content (R2 = 0.0484, p < 1.6e-38, Fig. 4F), but positively correlated with labile C content (R2 = 0.003969, p = 1.7e-4, Fig. 4G) and atmospheric CO₂ concentrations measured at sampling points (R2 = 0.002401, p = 0.0071, Fig. S26A) with no correlation with recalcitrant C content (R2 = 0.000225, p = 0.36, Fig. S26B). Although this index integrates multiple factors, including genes, viruses, and hosts, the soil ecosystem is highly complex, and many additional influences—such as plant dynamics and human activities—are not fully accounted for. It should also be noted that CTPI is a model-based composite metric that reflects relative functional potential rather than a directly measured metabolic rate. These findings suggest that by encoding diverse pCAZymes prophages likely accelerate labile C turnover, promoting microbial conversion of plant biomass-derived C to CO₂ and thereby enhancing total soil C loss. Prophages could also affect C cycling by entering the lytic cycle and releasing bacterial biomass into environment through lysis10, which could be inferred as negative correlations between phage and bacterial abundances. We performed Pearson correlation analyses between the abundances of 14,626 prophages and 9417 bacterial hosts in global soils (Supplementary Data 10). Most prophages (63.2%) and their hosts—spanning 116 viral families and 28 bacterial phyla—showed positive abundance correlations (all p < 0.05), whereas 36.7% exhibited no significant correlation (Fig. S27). Only 17 virus–host pairs (0.1%) exhibited significant negative correlations (all p < 0.05), suggesting that prophages predominantly persist in the lysogenic stage and only rarely shift to lytic cycle in global soils. This result is consistent with previous reports that temperate phages typically persist predominantly in the lysogenic state in other environments such as ocean29 and the human gut45. Moreover, 80% of phage–host abundance ratios were consistently below 1 (Fig. S28), further supporting the dominance of lysogenic cycle of temperate phages in global soils14. We also found that temperate phages and their host α-diversity (Richness and Shannon index) were positively correlated (R2 = 0.1089, p = 2.2e-16 for Shannon, R2 = 0.3481, p < 2.2e-16 for Richness, Fig. S29). These analyses suggest that prophages may primarily contribute to soil C transformation by augmenting host metabolic capacity through the provision of CAZyme genes, rather than by releasing microbial residual C via lysis. Predicting the future potential of prophage-driven C transformation in response to climate change To understand the potential significance of pCAZymes for C cycling in response to climate change, we modeled the change dynamic of pCAZyme genes and prophage-driven soil C transformation potential in the year 2100 under two future climate scenarios: Shared Socioeconomic Pathway 126 (SSP126, sustainable scenario) and fossil-fuel development pathway (SSP585, non-sustainable scenario). Our modeling projections showed that the total gene abundances of pCAZymes in global soils would significantly decrease by 1.9% and 1.8% (p = 1.85e-225) in the year 2100 under the SSP126 and SSP585 scenarios, respectively (no significant difference between two scenarios; Fig. 5A). For the composition of different CAZyme categories, gene abundance of AAs and PLs enzymes were most strongly affected, with their abundances markedly increasing under both scenarios—39.5% and 20.1% under SSP126, and 31.9% and 19.7% under SSP585, respectively (p = 9.47e-310 for AAs; p = 1.63e-101 for PLs, Fig. S30). The dominant GHs enzyme exhibited relatively less but significant variation under two climate scenarios (both <5%, p = 1.85e-135), with no significant difference between two scenarios (p = 0.16). The abundances of CEs and CBMs genes were significantly reduced under both climate scenarios (p = 7.8e-20 for CEs; p = 6.61e-50 for CBMs, Fig. 5B and Fig. S30), with the decline in CEs being more pronounced under SSP585. We observed a substantial decrease (>90%) in the abundance of cellulase- and hemicellulase-related genes under both the SSP126 and SSP585 scenarios (Fig. S31). In contrast, genes associated with lignin and pectin degradation increased by approximately 95% (p = 3.84e-127 for ligninase and p = 2.81e-172 for pectinase, Fig. S31). Notably, compared with SSP126, the abundances of lignin- and pectin-degrading enzyme genes exhibited greater variation under the SSP585 scenario. Although the Shannon diversity of prophages remained relatively stable without significant change (p = 0.071, Fig. S32), the community composition is projected to shift significantly under climate change. By 2100, the abundances of four dominant phages taxa—Hendrixvirinae, Casadabanvirus-like, Peduoviridae, and Sepvirinae-like—are predicted to decline significantly (all p < 0.05), whereas Microviridae is expected to increase markedly (p = 3.90e-128), with a more pronounced increase under the SSP585 (Fig. S33). Previous studies have shown that future climate change can alter the diversity and stability of soil bacterial communities46; accordingly, bacteriophages—viruses that specifically infect bacteria—are also expected to undergo corresponding changes. In parallel, the functional gene content carried by bacteriophages may shift in response to environmental variation. Prophage host community diversity (Shannon index) decreased significantly under both climate scenarios (p = 4.29e-28), with a less decrease under the SSP585 (Fig. 5C). Taken together, these results suggest that the gene abundance of pCAZymes in global soils may significantly decline, accompanied by clear shifts in prophage and host community composition under non-sustainable development climate scenario.Fig. 5Climate change under fossil-fuel-dependent scenarios may exacerbate prophage-mediated soil total carbon loss.Predicted change in A total abundance of pCAZyme genes (p = 1.85e-225), B gene abundance of enzyme CEs (p = 7.82e-20), and C bacterial Shannon index (p = 4.29e-28) at present and in 2100 under SSP126 (sustainable scenario) and fossil-fuel development SSP585 scenarios (non-sustainable scenario). D The Sankey diagram illustrates predicted shifts in the composition of different prophage-driven carbon transformation potential (CTP) groups by 2100 under the SSP126 scenario (left panel) and SSP585 scenario (right panel), relative to current climate conditions. E The soil prophage carbon transformation potential index at present and in 2100 under SSP126 and SSP585 scenarios (p = 6.72e-83). Predictive models were cross-validated (CV) with common Pearson correlation test using 70% samples as a model training dataset and 30% as a validation dataset. All climate variables were derived from WorldClim2 using a 5 min (~10 km) resolution. The future climate data were derived from different CMIP6 downscaled global change models (GCMs; see detailed information in “Methods”). In (A–C) and (E), n = 3548 for present, n = 3548 for SSP126 and n = 3548 for SSP585, where n represents the number of independent samples. Differences among groups were assessed using Kruskal–Wallis and pairwise two-sided Wilcoxon rank-sum tests; different lowercase letters indicate significant differences (p < 0.05). Box plots show the 25th–75th percentiles, whiskers indicate the minimum and maximum values, and the midline represents the median. Climate change under fossil-fuel-dependent scenarios may exacerbate prophage-mediated soil total carbon loss. Predicted change in A total abundance of pCAZyme genes (p = 1.85e-225), B gene abundance of enzyme CEs (p = 7.82e-20), and C bacterial Shannon index (p = 4.29e-28) at present and in 2100 under SSP126 (sustainable scenario) and fossil-fuel development SSP585 scenarios (non-sustainable scenario). D The Sankey diagram illustrates predicted shifts in the composition of different prophage-driven carbon transformation potential (CTP) groups by 2100 under the SSP126 scenario (left panel) and SSP585 scenario (right panel), relative to current climate conditions. E The soil prophage carbon transformation potential index at present and in 2100 under SSP126 and SSP585 scenarios (p = 6.72e-83). Predictive models were cross-validated (CV) with common Pearson correlation test using 70% samples as a model training dataset and 30% as a validation dataset. All climate variables were derived from WorldClim2 using a 5 min (~10 km) resolution. The future climate data were derived from different CMIP6 downscaled global change models (GCMs; see detailed information in “Methods”). In (A–C) and (E), n = 3548 for present, n = 3548 for SSP126 and n = 3548 for SSP585, where n represents the number of independent samples. Differences among groups were assessed using Kruskal–Wallis and pairwise two-sided Wilcoxon rank-sum tests; different lowercase letters indicate significant differences (p < 0.05). Box plots show the 25th–75th percentiles, whiskers indicate the minimum and maximum values, and the midline represents the median. We further investigated the response of prophage-driven soil C transformation potential in global soils under two future climate scenarios. The composition of soil prophage-driven CTP groups was clearly impacted by both climate change (Fig. 5D). For example, the dominant CTP-2 group decreased substantially, from 81% to 46% and 53% under SSP126 and SSP585, respectively, whereas the proportion of CTP-3 increased from 6% to approximately 20% in both scenarios. Overall, the non-sustainable SSP585 scenario exerted a more pronounced impact on the composition of high CTP groups (CTP-2 and CTP-3), suggesting that intensified climate change may substantially reshape prophage-mediated C transformation potential in soils. These changes were most pronounced in grassland and wetland soils (58.2–76.8%; Fig. S34), whereas agricultural and artificial soils were comparatively less responsive. This pattern may be associated with differences in microbial community composition and the influence of human activities across soil types36. Consequently, the prophage-driven soil CTPI showed significant but distinct changes under both climate scenarios (p = 6.72e-83), decreasing by 3 ± 0.2% under SSP126 but increasing by 13 ± 0.7% under the SSP585 scenario (Fig. 5E). Under the SSP585 scenario, the increase in prophage-mediated C transformation potential was mainly observed in agricultural soils and artificial surfaces, while grassland and wetland soils showed a decline (Fig. S35). As a result, soil C change dynamics also diverged between two climate scenarios. By 2100 under SSP585, the labile C increased by 4.8%, while total soil C declined by 2.0% (p = 2.54e-50 for labile C, p = 0.0486 for total soil C; Fig. S36). In contrast, SSP126 scenario predicted an 8.9% decline in labile C but a 1.3% increase in total soil C (Fig. S36). These results suggest that non-sustainable development might accelerate soil labile C turnover and total C loss by altering the prophage-driven C transformation potential. These findings are consistent with previous studies on the impact of climate change on the abundance of plant-beneficial bacteria in global soils47. Climate change can modify soil properties, particularly temperature, which directly affects microbial metabolic activity and stability46. However, it should be noted that, given the considerable uncertainties in future climate projections and soil physicochemical data, machine learning predictions should be interpreted as indicative rather than definitive. Experimental evaluation of the pCAZymes and their activity in response to simulated warming To experimentally verify the role of pCAZymes in lignocellulose hydrolysis, four representative pCAZyme genes—GH2 β-galactosidase GH3 β-glucosidase, GH53 endo-1,4-β-galactosidase, and AA10 lytic polysaccharide monooxygenase—from the pCAZyme database, were cloned into Escherichia coli DH5α for heterologous expression. The degradation activity of all enzymes toward their respective substrates was experimentally demonstrated in the Escherichia coli DH5α (Fig. S37). These genes were inserted (GH3, GH53 and AA10) into the genome of the temperate phage y1 (located on soil-isolated lysogenic strain Shigella flexneri B3 stored in our lab), and their enzyme activities were further verified (Fig. 6A). Compared with the control group without gene insertion, the enzyme activity of the prophage host increased significantly by more than 1.3-fold after acquiring pCAZyme genes (Fig. 6B). In particular, insertion of the AA10 LPMO gene in the prophage host increased LPMO activity from 0 to 44.0 U g⁻¹ protein h⁻¹. This is because lysogenic strain S. flexneri B3 lacks the LPMO genes (Fig. 6B).Fig. 6Soil microcosm experiments demonstrate that pCAZymes participate in carbon turnover and influence soil carbon mineralization.A Schematic diagram of construction of prophage genome carrying AMGs (upper panel) and soil microcosm experimental design (lower panel). Soil microcosm experiments were performed using three independent biological replicates for each gene. B Enzymatic activities of three pCAZyme genes (GH3 β-glucosidase, GH53 endo-β-1,4-galactosidase, and AA10 lytic polysaccharide monooxygenase) expressed in the lysogenic phage y1 integrated on Shigella flexneri B3. Three biological replicates were included for each gene. C–E Comparison of soil carbon mineralization rate, water-soluble organic carbon content, and soil enzyme activity between treatment and control groups (including acidic and alkaline soils) after 5 days of incubation in the soil microcosm experiment. GH3, GH53, and AA10 denote GH3 β-glucosidase, GH53 endo-β-1,4-galactosidase, and AA10 lytic polysaccharide monooxygenase, respectively. “-pCAZyme” represents the control group without pCAZyme insertion in lysogenic phage y1 (n = 3), whereas “+pCAZyme” represents the treatment group with pCAZyme insertion (n = 3). In (B–E), box plots show the 25th–75th percentiles, whiskers indicate the minimum and maximum values, and the midline represents the median. Statistical significance was assessed using two-sided Student’s t-tests at p < 0.05. Soil microcosm experiments demonstrate that pCAZymes participate in carbon turnover and influence soil carbon mineralization. A Schematic diagram of construction of prophage genome carrying AMGs (upper panel) and soil microcosm experimental design (lower panel). Soil microcosm experiments were performed using three independent biological replicates for each gene. B Enzymatic activities of three pCAZyme genes (GH3 β-glucosidase, GH53 endo-β-1,4-galactosidase, and AA10 lytic polysaccharide monooxygenase) expressed in the lysogenic phage y1 integrated on Shigella flexneri B3. Three biological replicates were included for each gene. C–E Comparison of soil carbon mineralization rate, water-soluble organic carbon content, and soil enzyme activity between treatment and control groups (including acidic and alkaline soils) after 5 days of incubation in the soil microcosm experiment. GH3, GH53, and AA10 denote GH3 β-glucosidase, GH53 endo-β-1,4-galactosidase, and AA10 lytic polysaccharide monooxygenase, respectively. “-pCAZyme” represents the control group without pCAZyme insertion in lysogenic phage y1 (n = 3), whereas “+pCAZyme” represents the treatment group with pCAZyme insertion (n = 3). In (B–E), box plots show the 25th–75th percentiles, whiskers indicate the minimum and maximum values, and the midline represents the median. Statistical significance was assessed using two-sided Student’s t-tests at p < 0.05. To directly determine whether pCAZyme genes in prophage enable the host to participate in soil C transformation, lysogenic strain S. flexneri B3 carrying the above genes in prophages were inoculated into two agricultural soils (acidic and alkaline), while the same strain lacking corresponding genes served as a control. Soil microcosm incubation experiments were conducted to measure changes in C mineralization rate (CMR), organic C content, and hydrolytic enzyme activities (Fig. 6C–E). In acidic soils, CMR in the treatment group were 1.3 ± 0.3-fold higher than those in the control group (p < 0.05), whereas in alkaline soils they were 1.6 ± 0.3-fold higher (p < 0.05). Moreover, the content of soluble organic C was significantly higher in the treatment group than in the control group in alkaline soils (p < 0.05), whereas no significant change in acidic soils. Consistent with these results, the activities of LPMO enzyme, β-glucosidase, β-galactosidase were significantly higher in the treatment groups in both soils (β-glucosidase activity: 1.2-fold increase in acidic soil and 1.4-fold increase in alkaline soil, p < 0.05; β-galactosidase activity: 1.1-fold increase in both acidic and alkaline soils, p < 0.05). Overall, these findings indicate that prophage-encoded CAZymes can significantly enhance organic C transformation processes in soils. Because pCAZymes primarily assist host C transformation during the lysogenic cycle, any shift in their life strategy could directly alter their C transformation capacity48. To assess the impact of warming on prophage induction activity, we conducted in vitro experiments using six inducible prophages from four bacterial strains (stored in our lab), incubated at 20 °C, 25 °C, and 30 °C for 12 hours. qPCR results showed that increasing environmental temperature significantly increased the activity of prophage induction and the expression level of pCAZyme genes within lysogenic hosts (Figs. S38 and S39). This result is consistent with previous research findings that temperature fluctuations associated with future climate change could modulate the interactions between phages and their bacterial hosts49,50, thereby influencing prophage-mediated soil C metabolic functions. In summary, this study provides the global catalog of pCAZymes, revealing their hidden role in soil C transformation. By integrating global metagenomic and metatranscriptomic data, we uncovered distinct diversity, biogeography, and transcriptional activity of prophages and their pCAZyme genes in global soils. Machine learning further predicted that prophage-mediated C transformation potential is sensitive to future warming. Based on direct experimental evidence, pCAZymes represent a previously overlooked “microbial dark matter” in the soil C cycling. Although this study integrates multi-omics and experimental analyses to provide the insights into the role of soil viruses in C transformation, several limitations remain. First, the limited availability of large-scale, high-resolution soil physicochemical data, together with incomplete global coverage of publicly available metagenomic datasets, means that many factors influencing pCAZymes and their C metabolic potential may not be fully considered. Second, C transformation at the soil ecosystem level is inherently complex and regulated by multiple interacting factors, including plants, human activities, and soil texture. While our microcosm experiments partially support the functional potential of representative pCAZymes, further validation under natural field conditions is required. Finally, although prophage genomes encode numerous hydrolases, these enzymes cannot be independently synthesized and secreted outside the host; thus, the role of prophages in soil nutrient transformation is more likely to act as an extension of host metabolism rather than as an independent driver. Methods Bacterial genome data collection A total of 98,701 bacterial genomes (> 1 Mbp, without plasmid) isolated from global habitats worldwide were retrieved from NCBI (January 2023). Genome quality was assessed using CheckM2 (v1.1.0) with default parameters51. After filtering for ≥90% completeness and <5% contamination, 8955 high-quality genomes were retained for next analysis. To include a large number of unculturable bacteria from metagenomes, we obtained additional high-quality metagenome-assembled genomes (MAGs; completeness >90%, contamination <5%) from terrestrial habitats via the Global Microbial Gene Catalog (8455 genomes)52, Genomes from Earth’s Microbiomes (2064 genomes)53, and the SMAG catalog (657 genomes)54. Finally, a total of 20,131 high-quality bacterial genomes were compiled for analysis, consisting of 44.4% sequenced bacterial genomes and 55.6% MAGs (Supplementary Data 1). Taxonomic classification was performed with GTDB-Tk (“classify_wf” workflow) based on default parameters55. The dataset covers 46 bacterial phyla and 1283 genera, dominated by Pseudomonadota (64.0%), Bacillota (13.6%), Bacteroidota (12.8%), and Actinomycetota (7.0%), totaling 97.4% of genomes. Detection of putative prophages in bacterial genomes and taxonomic classification Putative prophage sequences were identified using a combination of two complementary tools, as described previously25. First, geNomad (v1.7.4), a deep neural network-based viral predictor, detected viral sequences >5000 bp in bacterial genomes (default parameters)56. Second, CheckV (v1.0.1, database v1.5)57 was used to remove non-viral sequences, to trim host regions, and assess viral sequence quality, retaining sequences with conserved viral signature genes or homology to known phage elements. Genomes with at least one potential prophage were classified as lysogens. This approach identified 51,611 putative prophages from 20,131 bacterial genomes. After removing duplicates, sequences were clustered into viral operational taxonomic units (vOTUs) using mmseqs2 (--cov-mode 0, >95% identity, <85% coverage)58, yielding 21,626 non-redundant prophage vOTUs from 13,533 bacterial genomes (Supplementary Data 2). Prophage quality assessment by CheckV showed 1.4% complete, 47.8% high-quality, 23.4% medium-quality, and 27.3% low-quality sequences (Fig. S1 and Supplementary Data 2). Phage taxonomy was assigned with PhaGCN259, a graph convolutional network tool, using a minimum length of 5000 bp and referencing the ICTV database (release v4)60. Identification of putative auxiliary metabolic genes associated with C metabolism To assess prophage C metabolic potential, AMGs associated with CAZymes from the 21,626 prophages were analyzed using DRAM-v with manual curation61. Open reading frames (ORFs) in prophage were predicted using Prodigal-gv (-p meta)62, optimized for viruses and alternative genetic codes. Viral proteins were clustered with CD-HIT (v4.8.1)63 at 60% identity and 80% coverage to generate a non-redundant protein set. CAZyme genes were identified by aligning these proteins against the dbCAN2 database60 using DIAMOND64 (E-value ≤ 1e-5, coverage ≥ 75%). We identified 64,678 putative pCAZymes from a total of 21,626 prophages. Robust identification of potential viral AMGs from metagenomic data remains a major challenge in the field24. A credible potential AMG must meet two key criteria65: (1) it should be located between viral genes, with both its upstream and downstream regions flanked by viral hallmark or viral-like genes; and (2) it should be involved in a cellular metabolic pathway. Therefore, pCAZymes were considered high-confidence viral auxiliary metabolic genes when viral-like genes were present in their upstream or downstream regions24. To this end, we used VirSorter2 to detect viral-like genes within 20 genes upstream or downstream of each of the pCAZyme gene66. In addition, we used Pfam annotations to manually curate pCAZyme genes, excluding those associated with viral structural components or infection-related functions (e.g., tail fibers, baseplates, capsids, lyases, and other related proteins). For comparison, we applied the same approach to quantify CAZyme genes in 654,143 lytic phages from the IMG/VR database (IMG_VR_2022-09-20_7.1)67, and compared the number of CAZyme genes between prophages and lytic phages. The three-dimensional structure of the gene was visualized using SWISS-MODEL (https://swissmodel.expasy.org/interactive). SignalP 6.0 was used to predict the presence of signal peptides in pCAZymes29, in order to assess their potential for active extracellular secretion. Soil metagenomes and environmental data collection We compiled a global dataset of 3548 soil metagenomes to evaluate prophage distribution and function across biogeographic contexts. Metagenomes were retrieved from NCBI using keywords (“soil,” “land,” “sand,” “metagenomic”) and manually curated to exclude low-depth (< 500 Mbp), non-soil, or host-contaminated samples. Only samples with accurate collection time were retained to enable temporal analyses. The metagenomic datasets span eight continents, with the majority of samples originating from North America (56.0%), Asia (18.9%), and Europe (11.4%). These datasets primarily represent five major soil types: agricultural land, grassland, forest, wetland, and artificial surfaces. As the data were compiled from multiple publicly available sources, sampling methodologies varied among studies, leading to differences in sampling depth and collection time. Notably, most samples lacked detailed depth information (79.1%); among those with available metadata, the majority were collected from 0 to 40 cm (76.3%). All metagenomic samples included in this study exceeded 2 Gb in sequencing depth (Supplementary Data 6). Available metadata indicates that sampling depth is predominantly between 0 and 40 cm. Sequence quality was assessed via FastQC (v0.12.1)68 and low-quality reads were trimmed with fastp (v0.24.0)55 based on default parameters. To integrate soil physicochemical properties, historical and projected climate-driven soil variables (1850–2100, monthly averages) were extracted from CMIP6 models (BCC-CSM2-MR, CMCC-ESM2, CMCC-CM2-SR5, https://pcmdi.llnl.gov/CMIP6) and ISRIC SoilGrids, https://isric.org/explore/soilgrids). Variables included soil temperature (Ts), soil air temperature (Tas, 2 m above the ground), precipitation (Pr), soil moisture (Mrso), total soil nitrogen content (Nsoil), total soil C content (Csoil), labile C content (Csoilfast) and recalcitrant C content (Csoilslow), soil pH and CEC. Data were matched by sampling year and month or, if unavailable, within ±1 month with flags for deviations. Samples with missing environmental data (< 2%) were excluded from downstream analyses. Based on available metadata, sampling depth information was reported for 34% of the 3548 metagenomic samples, of which 94.2% were collected from 0 to 40 cm. For metatranscriptomic datasets, 88% of the 951 samples included depth information, with 99% falling within the 0–40 cm (Supplementary Data 6 and 8). Although sampling depth may influence soil physicochemical and microbial characteristics, its impact on global-scale model analyses is relatively limited, as horizontal variability in soil properties far exceeds vertical variation across soil depths. Moreover, all soil physicochemical variables were standardized by depth (expressed in kg m⁻², e.g., Csoilfast and Csoil), reducing depth-related bias. Abundance of prophage-encoded CAZyme genes and prophages in soils To assess prophage distribution and abundance of pCAZyme genes, 21,660 non-redundant prophage sequences and their encoded pCAZymes was mapped against 3548 global soil metagenomes using the CoverM pipeline (v0.61, https://github.com/wwood/CoverM) in contig mode as described previously69. Briefly, to calculate the relative abundance of each gene, quality-controlled reads from each metagenome were mapped to the set of all gene sequences with CoverM pipeline using the “TPM” calculation method (Transcripts Per Kilobase Per Million Mapped Reads). TPM is recommended for relative abundance comparisons with metagenomic datasets25, because TPM normalizes the data based on both sequence depth (per million reads) and sequence length (in kilobases). The analysis workflow for pCAZyme gene abundance, through TPM normalization and relatively stringent mapping thresholds, effectively reduces the impact of sequencing depth differences across samples and minimizes potential biases arising from non-specific and multi-mapping reads (Supplementary Data 7). For details, clean reads were mapped to prophage sequences with stringent criteria (--percentage_id 0.95, --percentage_aln 0.75) to generate BAM files. Coverage profiles were computed using filtered BAMs (--trim-min 0.10, --trim-max 0.90, --min-read-percent-identity 0.95, --min-read-aligned-percent 0.75, -m mean). The abundances of prophages and their bacterial hosts in soil samples were determined using the same approach as for pCAZyme genes using viral hallmark gene and 16S rRNA gene sequence as inputs, respectively. 16S rRNA gene sequences were extracted from bacterial genomes using Barrnap (v0.9, https://github.com/tseemann/barrnap) with default settings. Global maps of the geographical distribution of pCAZyme genes were constructed based on total abundance of all genes in each metagenomic sample. To achieve this, Empirical Bayesian Kriging was implemented in ArcGIS Desktop as described previously25 (ArcGIS Pro v3.0.2). In this approach, the variogram model was first estimated from the input data. For each data location, new values were simulated, and the simulated data were subsequently used to generate semivariogram models iteratively, following Bayesian principles70. Analysis of transcriptional activity of pCAZyme genes using metatranscriptomic data To assess the transcriptional activity of pCAZyme genes in soil samples, we used 951 soil metatranscriptome datasets collected from around world available in NCBI databases (Supplementary Data 8). The transcriptional activity of genes in the 951 metatranscriptome datasets were quantified using the CoverM pipeline in the same way as the previous metagenome19 (v0.61, https://github.com/wwood/CoverM). RNA reads were quality filtered via Trimmomatic (v 0.39) using the following parameter (score >30 and length >36 bases)19. Moreover, SortMeRNA (v4.3.4)71 was used to remove non-coding RNA sequences (tRNA, tmRNA, 5S, 16S, 18S, 23S, and 28S rRNA sequences) from the metatranscriptomic reads. The remaining mRNA reads were mapped back to pCAZyme sequences to identify gene expression activity based on the average coverage of transcripts per sample using CoverM pipeline using the ‘TPM’ calculation method72. During the mapping, we set the threshold very high (read identity >95% and alignment percentage >95%) to reduce the likelihood of false mapping errors (parameters: --percentage_id 0.95 --percentage_aln 0.95). The total activity of all pCAZyme genes in each soil sample was calculated by summing the activity of all genes in the sample (Supplementary Data 9). The pCAZyme gene was deemed as active when the ‘TPM’ values were larger than 0 according to a previously published study65. In environmental metagenomics, TPM > 0 is commonly adopted as a permissive threshold to capture low-abundance but potentially functional genes, given the prevalence of low-activity microorganisms in natural systems73–75. Geographic distribution of pCAZyme genes based on metagenomics We used the sampled metagenome dataset to generate a global map of pCAZymes distributions geographically described by Liao et al.25. Specifically, spatial distributions and abundance of pCAZyme genes were calculated using inverse distance weighting (IDW) method following previously developed procedures76 based on ArcGIS Desktop (ArcGIS Pro v3.0.2). Maps were created using ArcGIS® software by Esri (Copyright © Esri, used herein under license; www.esri.com), with basemap data layers credited to Esri, TomTom, Garmin, FAO, NOAA, and USGS. The IDW method is a deterministic interpolation technique that estimates values at unsampled locations as a weighted average of surrounding measured points, where the weights are inversely proportional to the distance raised to a power. This approach assumes that points closer to each other are more similar than those farther apart. To ensure consistency in land surface representation, the interpolation results were masked using the ESA GlobCover dataset (https://due.esrin.esa.int/page_globcover.php), which provides global land cover classification at a 300 m spatial resolution. The GlobCover mask was applied to exclude non-terrestrial areas (e.g., oceans, lakes, and ice) and to retain only valid terrestrial regions for subsequent spatial analysis. Links between microbial profiles and environmental factors To investigate the relationships among genes, microbial abundances, and soil physicochemical and environmental parameters, we performed PCoA analysis and Mantel tests. In total, we collected all environmental factors (including Ts, Tas, Pr, Mrso, Nsoil, Csoil, Csoilfast and Csoilslow, soil pH CEC) from 3548 soil metagenomes in further analysis. The Bray‒Curtis distances of the taxonomic and functional traits of the microbial community in the global soil samples were calculated based on the abundances of the taxa and genes, respectively, using the ‘vegan’ R package. Determining the soil C transformation potential driven by prophage in soil metagenomes To evaluate the potential of prophage-driven C transformation in soil metagenomes, we integrated pCAZyme gene abundances, prophage and host abundances, and their community diversity and composition across global soils (a total of 18 variables) using PCoA analysis. Soil is a highly complex system, encompassing phages and their hosts, which themselves harbor a large number of CAZyme genes. Therefore, we integrated these factors to construct a comprehensive evaluation index to evaluate the potential of prophage-driven C transformation. The 18 variables comprised: (i) Ecological metrics for dominant hosts—top five taxa abundances and Shannon diversity index; (ii) Ecological metrics for dominant temperate phages—top five taxa abundances and Shannon diversity index; (iii) Functional profiles of different pCAZymes across six enzyme classes: GHs, CBMs, GTs, PLs, CEs, and AAs. Variables were log-transformed when appropriate and standardized using Z-score transformation in R (package “scale”) to mitigate effects of different scales. The framework of the matrix captures the indirect functional coupling between prophage-encoded functional genes and host-mediated soil processes using a niche-based approach—i.e., a prophage’s ecological influence is considered significant only when prophage, functional gene, and its host are simultaneously abundant in soil. Prior to hierarchical clustering, PCoA analysis was performed with scikit-learn (v1.1.3) to reduce dimensionality and improve clustering performance. Only components explaining ≥95% of cumulative variance (15 PCs) were retained. Hierarchical clustering was employed for agglomerative clustering with Ward’s linkage on a Euclidean distance matrix. The eigenvector (loading) matrix was computed to evaluate correlations between original variables and PCs, recording both individual and cumulative variance contributions. The optimal cluster number (range 2–10) was selected based on silhouette score maximization. Finally, the ‘ward’ clustering algorithm with Euclidean distance was chosen to generate three C transformation potential clusters (CTP-1, CTP-2, and CTP-3), which were defined as phage-driven C transformation potential in this study. We then analyzed the detailed global distribution and microbial profiles for each C transformation potential group. Determining the prophage-driven C transformation potential index in each sample To identify soil samples that may be more susceptible to phage-driven C loss, we developed a composite metric termed the soil CTPI. CTPI is a comprehensive metric for assessing the potential of virus-mediated C metabolism in soils, rather than relying solely on pCAZyme gene abundance. This integrative approach is motivated by previous findings that viral communities, bacterial communities, and their interactions collectively influence soil C transformation processes. Importantly, CTPI should be interpreted as a relative indicator of potential functional capacity rather than a direct proxy for realized C turnover. This index quantifies the C metabolic potential of prophages in each soil sample based on the clustering results described above. Given that the first principal component (PC1) captured the largest variation (80%), the loading scores of 18 contributing factors on PC1 were weighted by their standardized abundance or Shannon diversity values for each sample. The sum of these weighted scores yielded an integrated prophage metabolic potential index, representing the overall prophage-mediated C transformation capacity across soils (see formula below). This approach combines statistical significance (via PCA loadings) with ecological relevance (via abundance and diversity), providing a robust framework to assess prophage functional contributions to global soil C cycling.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${CTPI}={\sum }_{i=1}^{n}{A}_{i}\times \frac{{Z}_{i}}{\mathop{\sum }_{i}^{n}\left|{Z}_{i}\right|}\,$$\end{document}CTPI=∑i=1nAi×Zi∑inZiwhere Ai represents the standardized values of either factor abundance or Shannon diversity index in each soil sample, and Zi represents the loading score of each factor. Prediction of prophage-driven soil C transformation potential under future climate change To evaluate how soil C transformation potential driven by prophage may respond to future climate change, we developed an integrated machine learning framework to predict pCAZyme gene profiles and their C transformation potential using eight environmental factors. It should be noted that this framework is not intended to provide precise quantitative predictions, but rather to estimate the potential magnitude and direction of responses under projected climate change scenarios. Given that all soil property indicators were significantly associated with both CAZyme gene expression and phage diversity, we incorporated these physicochemical parameters as predictors in our model. The predictive features included Tas, Ts, Pr, Mrso, Nsoil, Csoil, Csoilslow, and Csoilfast. Variables with strong autocorrelation, such as CEC and pH, were excluded. The response variables consisted of 18 microbial indicators, including the five dominant bacterial hosts and prophages, their Shannon diversity indices, and the abundance of different types of pCAZyme genes. To identify the optimal algorithm and hyperparameters, we constructed regression models using Random Forest (RF), Gradient Boosted Machines (GBM), Elastic Net regression, and K-Nearest Neighbor approaches, each tested with multiple hyperparameter combinations as described previously77. Tenfold cross-validation was used to enhance performance and minimize overfitting. For model training, 70% of samples were used as the training set and 30% as the test set. The optimal model for each gene abundance was selected based on predictive performance (R²) evaluated on both training and test datasets. The final ensemble model comprised 16 RF and 2 GBM algorithms, achieving test-set R² values between 0.24 and 0.693 (see Supplementary Data 11 for details), indicating moderate explanatory power and associated uncertainty in predictive performance. Notably, these tree-based algorithms inherently capture complex nonlinear relationships and interactions among predictors, allowing the model to account for potential nonlinear responses between soil viral gene abundance and environmental variables. This ensemble model was then used to project 18 microbial indicators under two CMIP6 climate scenarios for 2100: SSP126 (sustainable development) and SSP585 (fossil fuel-dependent). The predicted microbial indicators were subsequently classified into three predefined CTP groups to assess soil prophage-mediated C transformation responses under future climate conditions. All models were implemented using the scikit-learn Python package (v1.1.3). To evaluate model reliability, we quantified the uncertainty analysis and sensitivity analysis of model prediction78,79. Model uncertainty was firstly evaluated for the 18 predictive models used to estimate CTPI under SSP126 and SSP585 scenarios. Because the optimal model varied by target, we adopted a model-specific strategy. For both random forest (RF) and gradient boosting machine (GBM), uncertainty was quantified as the coefficient of variation (CV) of predictions derived from multiple model realizations. For RF, predictions from individual trees (n = 300) were used. For GBM, multiple predictions were generated via repeated model training under different random seeds or cross-validation splits. The uncertainty for each sample was then calculated as: Uncertainty was calculated as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{C}}}}V\left(x\right)=\frac{\sqrt{\frac{1}{N}{\sum }_{i=1}^{N}{\left({f}_{i}\left(x\right)-\bar{f}\left(x\right)\right)}^{2}}}{\bar{f}\left(x\right)}$$\end{document}CVx=1N∑i=1Nfix−f¯x2f¯xwhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f}_{i}(x)$$\end{document}fi(x) represents the prediction from the ii-th model realization. Uncertainty was further summarized using the CV. Among the 18 models evaluated, 14 exhibited CV values between 0 and 0.4, indicating relatively high reliability of the predictions80. For the final CTPI model under the SSP126 and SSP585, we further evaluated uncertainty by comparing predictions derived from soil physicochemical and environmental data generated by three CMIP6 climate models (BCC-CSM2-MR, CMCC-ESM2, and CMCC-CM2-SR5). The resulting CVs were 5.5% and 5.4% under the SSP126 and SSP585 scenarios, respectively, indicating limited variability and good consistency across model inputs. For sensitivity analysis, we applied partial dependence plots (PDPs) to the machine learning models to characterize the response of CTPI to changes in soil environmental variables81. Based on PDP-derived results, we quantified the variation in model predictions associated with each variable, allowing assessment of their relative importance. Cloning, expression, and enzymatic activity essays of pCAZymes To validate the functional role of prophages in lignocellulose degradation, four putative pCAZymes involved in cellulose and hemicellulose metabolism were selected from three lysogenic strains isolated from soils stored in our laboratory (Bacillus velezensis BV2, Klebsiella ornithinolytica HP3, and Paenibacillus amylolyticus JI6) and heterologously expressed in Escherichia coli DH5α as described previously19. These genes included GH2 β-D-galactosidase (no signal peptide), GH3 β-glucosidase (no signal peptide), GH53 endo-1,4-β-galactosidase (no signal peptide), and AA10 LPMO (containing a signal peptide) (Supplementary Data 12). The genes encoding the putative pCAZymes were synthesized and cloned into the plasmid pUC57 (carrying ampicillin resistance and an intrinsic promoter, eliminating the need for induction, Tsingke Biotech, Beijing, China). Recombinant plasmids were introduced into chemically competent E. coli DH5α cells. Transformants were selected on LB agar plates supplemented with ampicillin (100 μg/mL), and positive single colonies were propagated in LB medium until reaching an OD600 of 0.8. The presence of recombinant plasmids was confirmed by PCR using specific primers (Supplementary Data 12). An empty pUC57 vector lacking pCAZyme genes insert was used as a control. Intracellular and extracellular enzymatic activities of the pCAZymes were determined using an Enzyme Activity Assay Kit, following the kit instructions (TIANGEN Biotech, Beijing, China). E. coli cells were cultured in LB medium with 100 mM ampicillin for 10 h until reaching an OD600 of approximately 0.6. Briefly, the cultures were disrupted by ultrasonication (200 W; 3 s burst followed by 10 s pause; repeated for 30 cycles) on ice, followed by centrifugation at 15,000 × g for 10 min at 4 °C. The supernatants were collected and kept on ice until use to assess enzymatic activity. Enzyme activities of GHs hydrolases (GH2 β-D-galactosidase, GH3 β-glucosidase and GH53 endo-1,4-β-galactosidase) were quantified using a microplate reader at a wavelength of 400 nm after 1 h of incubation at 37 °C using containing 20 mM phosphate buffer (pH 5.0) and 0.5% of various polysaccharide substrates as described previously19 (pNP-β-,4-galactobioside for GH2 β-D-galactosidase; pNP-β-Glc for GH3 β-glucosidase; o-Nitrophenyl-β-D-galactopyranoside for GH53 β-galactosidase). For each substrate, a product-based standard curve was established to convert absorbance values into concentrations. Specifically, p-nitrophenol (pNP) standards (0, 25, 50, 100, 150, and 200 μmol·L⁻¹) were used for assays utilizing p-nitrophenyl substrates (p-nitrophenyl-β-D-glucopyranoside and p-nitrophenyl-β-1,4-galactobioside), whereas o-nitrophenol (oNP) standards (0–200 μmol·L⁻¹) were applied for the assay using the o-nitrophenyl substrate (o-nitrophenyl-β-D-galactopyranoside). Standards, blanks, and samples were prepared in the same buffer and measured in triplicate under identical conditions. Absorbance at 400 nm was plotted against known product concentrations to generate linear calibration curves for pNP and oNP, which were subsequently used to convert sample absorbance to μmol of product. Enzyme activity was expressed as units per gram of protein, with one unit (U) defined as the amount of enzyme required to release 1 μmol of product per minute under the assay conditions. The activity of AA10 LPMO (polysaccharide monooxygenase) was determined using the widely adopted 2,6-DMP colorimetric assay82. AA10 LPMO exhibits peroxidase activity in the presence of H₂O₂, oxidizing 2,6-dimethylphenol (2,6-DMP) to a colored product, which was measured at 469 nm. Briefly, 860 μL of buffer, 100 μL of 2,6-DMP stock solution, and 20 μL of H₂O₂ stock solution were mixed and incubated at 30 °C for 15 min, followed by the addition of 20 μL of the sample. After reacting at 30 °C for 5 minutes, absorbance was measured at 469 nm. One unit (U) of LPMO enzyme activity was defined as the amount of product generated per hour (1 μmol). Construction of prophage mutants carrying pCAZyme genes and validation of their role in soil C transformation To investigate the role of pCAZyme genes in soil C transformation, three representative genes (GH3 β-glucosidase, GH53 endo-1,4-β-galactosidase, and AA10 LPMO) were inserted into the temperate phage y1, which does not naturally encode any CAZyme genes. Functional differences in soil incubation experiments between the engineered strain with CAZyme and the wild type (without gene insertion) were subsequently compared. Phage y1 is a temperate phage integrated into Shigella flexneri B3 (isolated from soil in our laboratory) and can be induced into the lytic cycle by mitomycin treatment. Gene insertion was performed as follows83: (1) A non-structural region of the prophage genome was selected as the insertion site. Homologous arms (500–600 bp) flanking the target region were designed, amplified, and fused with a kanamycin resistance cassette containing FRT sites to generate a repair template (primers listed in Supplementary Data 13). (2) S. flexneri B3 cells harboring plasmid pKD46 (Tsingke Biotech, Beijing, China) were revived and prepared as electrocompetent cells following established protocols. (3) Approximately 300–500 ng of repair template DNA was mixed with 100 μL of competent cells, transferred to an electroporation cuvette, incubated on ice for 10 min, and electroporated (0.22 kV, 60 ms pulse duration, 5 pulses). Cells were recovered in 800 μL LB medium for 1 h, centrifuged (5000 × g, 5 min), plated on LB agar containing 100 mg/L kanamycin, and incubated overnight at 37 °C. Positive clones were purified and verified by PCR. Using this approach, mutant prophages carrying pCAZyme genes and corresponding wild-type controls were obtained. The soil microcosm experiment was conducted to investigate the role of prophage-carrying CAZyme genes in soil C transformation. Because prophage y1 remains in the lysogenic state under non-induced conditions, host bacteria carrying target gene were directly inoculated into soil to assess the functional impact of pCAZymes on C transformation. Two soil types were used: acidic agricultural soil (Fuzhou, Fujian Province, 26.0833° N, 119.2333° E) and alkaline forest soil (from Wuhan, Hubei Province, 30.59° N, 114.31° E), both collected from 0 to 20 cm depth. Fresh soil samples were cleared of debris, sieved (2 mm), and homogenized prior to use. For each microcosm, 20 g of soil was placed into a 200 mL headspace bottle and pre-incubated in the dark at room temperature for one week to allow microbial stabilization. Subsequently, lysogenic strain S. flexneri B3 cells, previously washed with PBS to minimize external C inputs, were inoculated into the soil at the 10⁸ CFU g⁻¹. Soil moisture content was adjusted to 30%, and samples were incubated at room temperature for 5 days. The experiment included two soil types and four treatments (3 three genes treatment and 1 control), each with five biological replicates (total n = 40). To ensure adequate aeration, bottles were sealed with a breathable membrane. Prior to gas sampling, bottles were sealed for 6 h, and CO₂ concentrations in the headspace were measured every 24 h using gas chromatography (GC-2014C, Shimadzu, Japan). Based on previously established methods84, the soil C mineralization rate (CMR) was calculated from the soil CO₂ emission rate using the following equation:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{CMR}}}}=({{{{\rm{M}}}}}_{({{{{\rm{CO}}}}}_{2})}\times 12/44)/(1000\times {{{{\rm{W}}}}}_{({{{\rm{soil}}}})}\times {{{\rm{T}}}})$$\end{document}CMR=(M(CO2)×12/44)/(1000×W(soil)×T)where M₍CO₂₎ is the cumulative CO₂ emission (μg), W₍soil₎ is the dry soil weight (kg), and T is the incubation time (d). At the end of incubation, soil water-soluble C and enzyme activities were determined. For water-soluble C, 5 g of soil was mixed with water (2.5:1, v/w), shaken at 150 rpm for 1 h, allowed to settle for 30 min, and filtered (0.45 μm). Soluble total C content was measured using a TOC analyzer (Shimadzu TOC-L CPH, Kyoto, Japan). For enzyme activity, 1 g of soil was extracted with 2.5 mL PBS buffer, centrifuged (15,000 × g, 10 min, 4 °C), and the supernatant was used for colorimetric assays as described above. Experimental validation of prophage induction and their encoded CAZyme genes under different temperatures To experimentally validate the response of prophage induction to elevated temperatures simulating climate change scenarios, four sequenced bacterial isolates harboring prophages were randomly selected for in vitro assays. To assess the impact of warming on shifts from lysogenic to lytic phage life cycle, we conducted in vitro experiments using six inducible prophages from four bacterial strains (all temperate phage and bacteria listed in Supplementary Data 14) as previously described methods25, incubated at different temperatures (20 °C, 25 °C, 30 °C) for 12 h. Briefly, strains were grown in fresh LB at 30 °C overnight to the OD600 = 0.8. After centrifugation at 5000 × g, the bacterial cells were resuspended in phosphate-buffer for prophage induction. Mitomycin C was added to cultures at a final concentration of 0.5 μM as the inducer, and the cells were incubated statically for 10 h at 20 °C, 25 °C, and 30 °C at the dark, yielding a total of 36 cultures (4 strains × 3 temperature gradients × 3 biological replicates). As viral genomic DNA enclosed within protein capsids is resistant to nuclease degradation, DNase treatment can be used to distinguish between free bacterial genomic DNA released from lysed host cells and phage DNA. All induced cultures were passed through sterile 0.22-μm filters, and the filtrates were treated with DNase for 1 h to degrade free bacterial genomic DNA, leaving intact phage DNA protected within virions. The complete removal of free genomic DNA was confirmed by PCR targeting the 16S rRNA gene (primers: 27F, AGAGTTTGATCCTGGCTCAG; 1492R, TACCTTGTTACGACTT). Subsequently, samples were treated with proteinase K at 56 °C for 1 h to inactivate DNase and release phage DNA, followed by incubation at 70 °C for 10 min to inactivate proteinase K, leaving viral DNA as the sole nucleic acid in the samples. Quantification of each prophage induction profile was performed using qPCR with specific primers targeting the six prophages (Supplementary Data 14). Finally, the release of viral particles was calculated based on the number of targeted viral genes as described previously25. The induction rate of prophages entering the lytic cycle was used to estimate the average number of viral particles produced per bacterial cell. To experimentally verify whether the expression levels of phage-carried CAZyme genes are regulated by environmental temperature, three lysogenic bacterial strains preserved in this study (each harboring two prophages) were selected for in vitro temperature stimulation experiments. All strains were first cultured overnight in LB broth at 30 °C to an OD₆₀₀ of 0.8. A 100 μL inoculum was then transferred into 1.5 mL of fresh LB broth and incubated at 30 °C for 12 h. Subsequently, the cultures were subjected to different temperature treatments (20 °C, 25 °C, and 30 °C) for 1 h. After incubation, the cells were immediately centrifuged at 4500 × g for 5 min to collect the pellets, which were rapidly frozen in liquid nitrogen for total RNA extraction. Total RNA was extracted using the RNAsimple Total RNA Kit (TIANGEN Biotech, Beijing, China). RNA integrity and concentration were assessed by gel electrophoresis and spectrophotometry. A total of 200 ng RNA per sample was reverse-transcribed into cDNA using the HiScript III RT SuperMix for qPCR (Vazyme Biotech, Nanjing, China) following the manufacturer’s protocol. qPCR was then performed to determine the relative expression levels of pCAZyme genes, with primer sequences listed in Supplementary Data 15. Each temperature treatment was performed with a minimum of three biological replicates. To account for potential differences in bacterial growth rates under different temperature treatments, the relative expression levels of target genes were normalized to the expression level of the 16S rRNA gene. Statistical analysis Statistical analyses were conducted using the R platform85. Mantel tests and both linear and nonlinear regressions were used to evaluate direct effects of environmental factors on soil viral diversity and structure86. Two approaches were used to investigate the global diversity patterns of gene and microbial community profiles: generalized linear models (GLMs) and ordinary least squares regression. PERMANOVA (Adonis function with 999 permutations) combined with PCoA was performed through the “vegan” and “ggplot2” packages to examine correlations between soil physicochemical factors and prophages, their hosts, and gene abundance. Statistical analyses were performed based on data distribution. For comparisons among multiple groups, parametric tests (one-way ANOVA followed by Tukey’s HSD post-hoc test) were applied for normally distributed data, while non-parametric alternatives (Kruskal–Wallis H test followed by pairwise Wilcoxon rank-sum tests) were employed otherwise. For direct comparisons between only two groups, the Student’s t-test or Wilcoxon rank-sum test was used depending on data normality. All statistical tests with p < 0.05 were considered statistically significant. Pearson correlation analysis was performed, and a two-sided correlation test was used to assess statistical significance with p < 0.05. The dimensionality-reduced clustering results of the 18 metrics were visualized using python’s matplotlib and seaborn libraries. Pearson correlation analysis was employed to evaluate the C metabolic potential of soil prophages. Sampling site information was mapped globally using the “ggplot2” and “maps” packages, while various plots including bar charts, box plots, heatmaps, violin plots, stacked bar plots, and pie charts were generated via “ggplot2” package for comprehensive data visualization. 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 Files Supplementary Data 1-15 Reporting Summary Transparent Peer Review file Supplementary Information Description of Additional Supplementary Files Supplementary Data 1-15 Reporting Summary Transparent Peer Review file Source data Source Data Source Data