Dynamics of gut bacteriophage in diversity outbred mice studied over lifespan and during extreme caloric restriction

Abstract

Background

The majority of bacteria in the vertebrate gut harbor integrated bacterial viruses (“bacteriophages” or “phages”; integrated phage are termed “prophages”). To probe phage replication strategies in the mammalian gut microbiome, we investigated phage activity in a large longitudinal study of diversity outbred mice (913 animals) undergoing extreme dietary restriction with detailed phenotypic characterization across lifespan.

Results

We assembled 54,119 candidate DNA viral genomes from 2997 longitudinal metagenomes, forming 6462 viral operational taxonomic units (vOTUs). Over 85% of vOTUs annotated as novel. Viruses annotated predominantly as prophages in the Caudoviricetes class. We detected no eukaryotic DNA viruses, and none of the strictly lytic Crassvirales order that is abundant in human gut. The most prevalent phages had the widest predicted host ranges. The relative abundance of most phages was highly correlated to that of their inferred host bacteria, suggesting quiescent prophages dominate viral metagenomes, consistent with “piggyback-the-winner” dynamics. After accounting for close phage-bacterial covariation, we did identify a subset of phages changing in relative abundance and prevalence relative to their hosts in response to dietary restriction and aging. In particular, phages with larger genomes become less common in diets with restricted calories, potentially reflecting a higher fitness cost to their host. Generalist phages were enriched for a gene encoding a single-strand DNA binding protein which is reportedly involved in DNA repair and protection from nucleases encoded by host cells. Lytic phages became more common with aging, and we observed a reduction in phage richness with age, both findings previously observed in human cohorts.

Conclusion

These studies enrich our understanding of DNA phage dynamics in gut while emphasizing the predominance of “piggyback-the-winner” strategies.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40168-026-02362-4.

Introduction

The mammalian gut is home to vast populations of viruses [1–5]. These include viruses of the mammalian host and viruses infecting members of the host’s microbiome, including bacteria, archaea, and eukaryotes. Recent studies of the gut virome in human cohorts have disclosed communities often reaching 10⁸–10¹⁰ per gram of stool [3, 6–8], the large majority of which are viruses that infect bacteria, called bacteriophages or phages. Bacteriophages can modulate bacterial communities by predation, alter host metabolism, and induce bacterial or human immunity [6, 9, 10]. The largest portion of this community is commonly dominated by tailed phages of the class Caudoviricetes [3, 6, 11–13]. Extensive new tools for analyzing and enumerating viruses are now available, leading to an increase in studies linking changes in the virome to health outcomes or environmental conditions [11, 14, 15]. Fully modeling the mechanisms involved often turns on understanding the underlying phage population dynamics.

Phages can replicate via several strategies. Phages undergoing lytic replication infect bacterial cells, produce viral macromolecules, assemble new particles, and lyse their hosts, releasing large numbers of viral particles. In contrast, temperate phages infect bacterial cells and integrate their genomes into the bacterial genome to form prophages. During host genome replication, prophage genomes are copied as with any bacterial gene. Prophages can induce in the presence of a suitable signal such as DNA damage and excise their genomes and resume lytic growth, producing new phage particles and lysing cells. In a third option, an otherwise virulent phage can replicate in a temperate manner by copying its genome in the host as a plasmid, without producing viral particles [13, 16]. Lytic phages can modulate bacterial population structure by predation. Lytic phages are also of interest for potential use in phage therapy [17]. Lysogenic phages can be vectors for horizontal gene transfer, contributing virulence or resistance genes to host bacteria, thereby altering microbiome function without direct bacterial lysis [18].

Phage-host dynamics are believed to be driven in part by environmental factors. When resources are plentiful and bacterial abundance is high, lysogeny often dominates—a so-called “piggyback-the-winner” dynamic. Under these conditions, the fitness cost of carrying integrated prophages is minimal, and a phage can reliably replicate its genome integrated in a successful host [19, 20]. A recent study of the human gut microbiome quantified lytic viral particles and integrated prophages to demonstrate a low rate of lytic growth and minimal fitness costs of carrying prophages, suggesting that “piggyback-the-winner” dominated [21]. In contrast, in environments where resources are either very low in abundance or very high in abundance, a “kill-the-winner” lytic dynamic may be favored [19, 20].

Recent work has led to the development of diverse novel virus discovery tools, allowing analysis of the virome of new environments [22–24]. Murine systems offer an attractive model for studies of phage–bacterial dynamics. Both lytic and lysogenic phages have been previously identified in mice, but their relative contributions to the murine microbiome have not been fully clarified [25, 26]. Replication dynamics are only beginning to be investigated.

Here we characterize the DNA virome of diversity outbred mice [27] under dietary perturbations and natural aging using the Dietary Restriction in Diversity Outbred mice (DRiDO) cohort. This cohort consists of 913 mice with longitudinally sampled gut metagenomic data [28, 29], for a total of 2997 fecal DNA samples analyzed by shotgun metagenomics. Adult female mice were randomized into five dietary groups—either receiving ad libitum amounts of food, 20% caloric restriction, 40% caloric restriction, fasting for 1 day each week, or fasting for 2 days each week (Fig. 1A). Mice were sampled an average of four times throughout their lives. The bacterial microbiome of these mice has been extensively categorized [30], but the virome component remains unexplored.

Fig. 1.

Experimental design and creation of a mouse gut viral database. A Mice were randomized into one of five dietary conditions—ad libitum, 1- or 2-day fast per week, or 20% or 40% caloric restriction (CR). Metagenomic samples were collected every 6 months, starting just prior to randomization, for a total of 2997 samples. B Schematic diagram of viral database construction. C Distribution of viral genome lengths. D Proportion of phages shown by inferred replication strategy. “Suspected” virulent or temperate phages were annotated based on PhaBOX prediction, while phages that were actually observed as integrated prophages in our contigs were labeled as “Temperate.” E Distribution of vOTU prevalences

Prior work on the gut virome in response to dietary restriction is sparse. Early studies in human cohorts reported changes in the virome associated with diet [4, 12, 31]. Studies of mice fed a high-fat diet suggested that the virome is highly responsive to changes in caloric availability, but the microbes involved have commonly differed between studies [9, 32]. In aging, human studies have found a loss of viral richness in older populations, which has been seen across several human cohorts [6], with a shift towards a more lytic gut virome in elderly populations [33]. Little is known about the virome of ageing mice.

Previous work in the DRiDO cohort found that a majority of bacterial species changed in relative abundance in response to dietary restriction, while aging generally increased uniqueness in the microbiome, potentially representing stochastic acquisition and loss of new bacterial species over time [28]. These samples provide an opportunity to study the dynamics of the murine DNA virome in unprecedented detail.

Methods

Metagenomic data

All data from the DRiDO mouse metagenome study was downloaded from NCBI’s Sequence Read Archive [34]. Original mouse experiments were approved by the Institutional Animal Care and Use Committee at The Jackson Laboratory (protocol number 06005). Metadata was downloaded from github (https://github.com/levlitichev/DRiDO_microbiome). Information on sample processing and the DRiDO mouse cohort was described previously [28, 29]. In brief, female mice from 6 generations of diversity outbred mice were enrolled from March 2016 to November 2017. Mice were randomized into five dietary groups, starting at 6 months of age. The sample size was determined to detect a 10% change in mean lifespan between intervention groups with allowance for some loss of animals due to non-age-related events. Authors used female mice only due to concerns about male aggression. The AL feeding group was provided with unlimited access to food and water. The intermittent fasting (1D and 2D) mice were provided unlimited access to food and water. On Wednesday of each week at 15:00, the IF mice were placed into clean cages and food was withheld for the next 24 or 48 h for the 1D and 2D groups, respectively. Caloric restriction (20% and 40%) mice were provided with unlimited access to water and measured amounts of food daily at around 15:00, 2.75 g per mouse per day for 20% CR and 2.06 g per mouse per day for 40% CR. Mice were monitored daily and weighed weekly. Research staff regularly evaluated mice for prespecified clinical symptomology: palpable hypothermia, responsiveness to stimuli, ability to eat or drink, dermatitis, tumors, abdominal distention, mobility, eye conditions (such as corneal ulcers), malocclusion, trauma and wounds of aggression. Overall mortality for each group by cause is described in detail in the original manuscript [29].

Data preprocessing

We performed quality control of our sequence reads using the Snakemake pipeline Sunbeam [35] (v2.1.1). More specifically, we removed adaptors with cutadapt [36] (v3.1, forward and reverse adaptors = CTGTCTCTTATACACATCT), trimmed low-quality bases and discarded low-quality reads with trimmomatic [37] (v0.39, ILLUMINACLIP:trimmomatic/adapters/NexteraPE-PE.fa:2:30:10:8:true LEADING:3 TRAILING:3 SLIDINGWINDOW:4:15 MINLEN:36), discarded reads with many repetitive sequences with komplexity (https://github.com/eclarke/komplexity) and removed host reads (mean 9% of input reads) using bwa [38] (v0.7.17) against the mm10 genome.

Viral genome discovery

Short read sequences were assembled into contigs using Meta SPAdes (V.3.15.1) [39]. Contigs over 1000 bp were passed to Virsorter2 (v2.2.4) [24] (https://github.com/jiarong/VirSorter2) and Cenote-Taker3 (v3.3.1) [22] (https://github.com/mtisza1/Cenote-Taker3) to identify possible virus contigs. We did not filter by viral completeness score, as these scores are computed by comparison to well characterized viruses, and our contigs generally showed low similarity to references outside of the viral hallmark genes. Both methods use hidden Markov models specific for viral hallmark genes as previously described [22, 24]. Resulting contigs were then de-duplicated at 99% identity using cd-hit-est [40]. We ran CheckV (v1.0.1) [23] on deduplicated contigs, keeping all viral contigs with > 3 viral genes, and all circular viral contigs. Flanking host regions, and bacterial DNA present as a result of capsid packaging [41], were subsequently trimmed from viral contigs based on annotations from CheckV and Cenote-Taker3. The post-CheckV contigs were used as our final database of mouse viruses in subsequent analysis.

To define viral OTUs, contigs were aligned via BLASTN, and clusters were made using a 95% average nucleotide identity (across 85% alignment fraction) cutoff [42] and MCL clustering (https://micans.org/mcl/) [43].

Virus and bacteria quantification

Viral contigs and bacterial genomes from the Mouse Gastrointestinal Bacterial Catalog (MGBC (v1.0.0)) were added to a custom Kraken2 (v2.1.3) database [44]. Regions in bacterial genomes of the MGBC that were annotated as prophages by geNomad (v1.7.4) [45] were masked with Ns to prevent double counting viral reads. Taxa were quantified at the species level with Kraken2 default parameters, except the confidence threshold, which was raised to 0.1. Samples with fewer than 10,000 viral reads were discarded (n = 87). Viruses and bacteria were normalized separately to reads per 100,000 (to reflect the approximate average number of viral reads per sample) and log 10 transformed. Both viral and bacterial taxa were threshholded to only keep those with at least 0.1% relative abundance in at least 100 samples, resulting in 310 bacterial species and 754 viral OTUs.

K-mer based tools run the risk of inaccurate quantification if a large number of reads hit only a small section of a viral genome, resulting in false positives. To assess whether this was happening in our data, we aligned reads via BWA (0.7.18-r1243) [38] from a random subset of 100 samples to 10 phage genomes, spanning from the least abundant vOTU to the most. We observed that for the majority of non-zero counts, vOTUs detected by Kraken were also well covered across their whole genome when aligning reads via BWA (Supplementary Fig. S1).

Phage genome annotation

Phage OTUs were defined as temperate if at least one of the contigs in that species was detected as an integrated prophage via CheckV. Other vOTUs were defined as either “suspected virulent” or “suspected temperate” based on the prediction from PhaTYP (database version November 16, 2022) [46]. In some instances, highly similar contigs from the same vOTU differed in their predicted replication mode. In such cases, if more than 25% of the contigs in the vOTU were classified as temperate, the vOTU was considered temperate. This weighing toward temperate phages was based on the observation that some incorrect virulent calls appeared to stem from incompletely assembled genomes wherein integration genes were missing because of technical issues or low coverage.

Phage proteins were identified using geNomad, and clustered by structural similarity using FoldSeek (v 9.427df8a) easy-cluster with default parameters [45, 47]. Proteins were subsequently aligned via FoldSeek to all PDB sequences to identify possible functions [48].

Viral hosts were annotated in three ways. First, we identified prophages integrated into genomes from the Mouse Gastrointestinal Bacterial Catalog (MGBC) [49] using geNomad [45], and aligned these prophages to our own contigs. If a vOTU was within 95% average nucleotide identity and 85% alignment fraction of a prophage, then the host from which that prophage originated was considered a possible host for the vOTU. These methods were used, rather than an off-the-shelf pipeline, so that phages could be tied directly to the mouse-specific genomes in MGBC, many of which are not present in existing phage–host prediction databases.

Next, we identified CRISPR spacers in host genomes from MGBC using MinCED (https://github.com/ctSkennerton/minced (v0.4.2)) and aligned via BLASTN (v 2.14.0 +), with a percent identity threshold of 95%. Bacteria with at least 8 different CRISPR spacers aligned to a vOTU were considered possible hosts of this vOTU. We could assign possible hosts to 730 vOTUs with these methods, with an average of 22 host species via CRISPR and 8 via prophage similarity. At the genus level, we identified an average of 2.2 (sd = 2.8) host genera per vOTU. These sequence-based methods do not necessarily reflect the ground truth of which phages can infect which hosts, but do suggest possible host ranges based on a history of similar viruses infecting similar bacteria.

Finally, possible hosts were identified using correlation between relative abundances of phage and bacteria. When the R² of a linear fit between log10 bacterial reads and log10 phage reads was greater than 0.7, the bacterial species was considered a possible host for the virus species. In some phage-host pairs, a subset of samples appears to show strong correlation, while other groups of samples do not (Supplementary Fig. S2). We hypothesize that this occurs when a phage can infect a particular bacterial species, but the phage is not always present, or the phage has other possible hosts as well. In such instances, phage-bacteria relative abundance correlation is weak when considering all samples, but strong in this subset, providing evidence that this bacterium can serve as a host for this virus. To identify these instances, we clustered samples with DBScan (R version 4.1.2 parameters eps = 0.15, minPts = 50) [50], then considered a phage-bacteria pair a positive hit when any of the DBScan clusters had an R² greater than 0.7.

The seed contigs for each vOTU were aligned via BLASTN to all viral genomes from the IMG-VR (v2022-12–19 7.1) in order to assign taxonomy [51]. Our vOTUs were considered to be in the same species as a reference virus based on a cutoff of 95% identity and 85% alignment fraction. Viruses not found in this database were considered potentially novel viruses.

Linear mixed effects model

All univariate analyses were conducted using a linear mixed effects model unless otherwise specified. The lme4 R package was used with age, diet, and chronological time as fixed effects, and random effects batch and cohort, age, and mouse nested in that order [52]. Significance was identified by a likelihood ratio test using a null model with and without each fixed effect. Models for viral richness included bacterial richness as a fixed effect covariate.

PCoA and Procrustes analysis

Principal coordinate analysis was conducted in R using Bray–Curtis distances from the vegan R package [53]. Procrustes was used from the same package with default settings (scaling = TRUE, symmetrical = FALSE).

Virus presence/absence analysis

For highly correlated phage–host pairs, changes in vOTU relative abundance with diet, age, or chronological time commonly reflect changes in the host bacteria relative abundance. In order to identify virome-specific changes in these phage–bacteria pairs, we quantified phage presence or absence in samples where the putative host bacterium was present. Changes in phage prevalence that were associated with age, diet, or chronological time were identified by a generalized linear model, predicting presence/absence as a binary outcome. Diet, age, and time were fixed effects, while batch and cohort/cage/mouse were random effects. For this analysis, we analyzed only phage–bacteria pairs with a strong correlation as described above.

Sequence acquisition is a process of random sampling, and a phage can be missed if it is present at only low levels. When looking at host-prophage dynamics, this creates a false-negative problem, wherein samples with low abundance of bacteria may appear phage-negative because of random sampling, despite hosting a prophage or phage-plasmid. This is complicated by the fact that at equal abundance, a phage will generate fewer reads than a bacterial genome because its genome is substantially smaller. This necessitates a cutoff below which a bacterium is considered too low-abundance to accurately determine if it is carrying a given phage. Unfortunately, the noise associated with each phage-host pair was inconsistent, meaning that the threshold below which a phage might be missed was different for each phage. Because of this inconsistency, a single threshold for all phages could not be derived, even accounting for genome lengths. Therefore, we derived a relative abundance cutoff for each bacterial species, below which a sample was excluded from analysis. To do this, we fit a linear model lm (phage counts ~ bacterial counts) and calculated the variance of the residuals to obtain a 95% prediction interval. One such model was fit for each phage-bacteria pair (Supplementary Fig. S3). From this linear fit, a minimum threshold of bacterial counts was determined where the lower 95th percentile of the viral read counts became less than 5 reads. Below this point, even when a phage is present, it may not be detected. If a sample had fewer counts of a given bacterial species, it was excluded from the phage presence/absence analysis for that phage-bacteria pair.

Random forest predicting age and diet

To determine how independent changes to the virome and bacteriome affected dietary restriction or aging, we built random forest models using the R package randomForest [54] to predict diet or age from either data type. For each model, the top 100 taxa were included, and default parameters were used, except ntrees was increased to 1000.

Random forest predicting phage relative abundance

To account for possible non-linear dynamics between the bacteriome and the virome, we used a random forest regression to predict viral relative abundance from bacterial relative abundance. Samples were split into a 70% training set and a 30% test set, keeping all mice from the same cage in the same group to prevent overfitting.

To analyze changes in the virome with aging beyond what was expected from the bacteriome alone, residuals from this model were used and tested using a linear mixed effects model as described previously. Only predictions in the test set were used for this analysis. Only phages for which the model explained > 50% of the variance in the test set were evaluated.

To analyze changes associated with diet, a separate random forest model was used. As a training set, all pre-randomization samples were included, such that the model only saw ad libitum mice. The first post-randomization sample was then used as a test set. Residuals were tested for association with diet using a Wilcox non-parametric test. Only taxa for which the model explained > 50% of variance in the ad libitum samples were evaluated.

Genes associated with viral phenotypes

To identify genes enriched in vOTUs associated with each phenotype, phage proteins from each genome were identified using geNomad [45] and clustered using foldseek easy-cluster with default parameters [47]. Cluster seeds were searched against all protein structures in PDB using foldseek easy-search and the ProstT5 protein language model [55]. To identify genes associated with host range, we used a Fisher’s exact test. To identify genes changed with diet and aging, we grouped vOTUs’ response to each phenotype into “increasing,” “no change,” and “decreasing,” and tested for association with each gene cluster using an ordinal generalized linear model.

Heritability in the virome

To determine if variation in viral relative abundance was associated with host genotype, we used a linear mixed effects model with a kinship matrix describing the relatedness of every mouse [30]. This was implemented in R using the package lme4qtl [56].

Results

Mouse gut DNA virome characteristics

We sought to assess viral dynamics in the murine gut associated with diet and age, taking advantage of the 2997 stool metagenomic DNA sequence samples available from the DRIDO cross. We assembled 48 million total contigs from the DRIDO cohort stool metagenomes, then assessed the DNA virome fraction using Cenote-taker3, VirSorter2, and CheckV (see “Methods” section) [22–24]. Our virus discovery and de-replication pipeline ultimately yielded 54,119 viral contigs (Fig. 1B). We identified 6,462 viral operational taxonomic units (vOTUs) from these contigs, using 95% average nucleotide identity and 85% alignment fraction and MCL clustering. Of these, 5578 were novel to this study, based on alignment to the IMG-VR database and using this same identity and coverage threshold [51]. Median viral genome length was 8.5 kb (Fig. 1C). Genomes annotating as temperate or suspected temperate phages predominated (Fig. 1D).

Across all genomes, the most common viral genes were for tail proteins, with 10,876 proteins annotated as tail tube proteins, 9513 as the HK97-gp10 tail protein, and 8,725 as tail minor proteins. Proteins annotated as the RinA transcriptional activator were the most common non-structural protein, with 7424 copies across all genomes. Phage terminases, capsid proteins, and portal proteins were also among the most common proteins (see extended data, “Genomad protein annotations” Supplemental Table S1).

Over 99% of our viral contigs were assigned to the viral class Caudoviricetes, the tailed phages, suggesting that this family dominates among dsDNA viruses of the murine gut virome. We note that the Caudoviricetes are especially well-represented in reference databases used to train virus discovery tools, so less common viral types may be under-counted. Additionally, dsDNA metagenomic data will typically not capture genomes of ssDNA or RNA viruses [30]. The method may also under-count viral particles that are difficult to extract with the sample preparation and metagenomic methods used. We found no eukaryotic viruses in our data based on alignments of our viral genomes to references in the Integrated Microbial Genomes Virus (IMG/VR) database [51]. We found no evidence of CrAssphages, a highly abundant phage in human virome samples [57]. It is unclear whether our protocol failed to capture these phages, or whether Crassvirales are indeed absent in laboratory mice. In the below, we refer to our samples collectively as “viromes”, reflecting the methods used, but recognize that most or all are dsDNA phages.

The total inferred DNA phage virome accounted for an average of 1.18% (sd = 0.28%) of sequencing reads per metagenome. Phage richness was high, with a mean of 571 (sd = 53) vOTUs per sample. As expected, many vOTUs were low in prevalence, with most present in fewer than 7% of samples (Fig. 1E). Conversely, 38 vOTUs appear to be ubiquitous within our mouse cohorts, present in over 95% of samples. These vOTUs tend to be generalist phages, infecting multiple host species or even multiple genera, as identified by CRIPSR spacers and integrated prophages. There was a positive relationship between the number of possible hosts and prevalence (Spearman Rho = 0.17, p = 1e−6), which suggests that a generalist strategy may be beneficial for phage proliferation.

We identified two phage protein clusters associated with host range using a Fisher’s test and FDR < 0.05. One protein (closest PDB reference 6bhx) was found more often in viruses with a wider host range (> 1 host genus) and appears to be a novel single-stranded DNA binding protein. It is related to a previously described CRISPR-Cas evasion protein [58] (Fig. 2A, B) that is involved in DNA repair after cleavage by Cas enzymes, a host-agnostic strategy to survive bacterial nuclease-based defense systems. We identified several other phage proteins that co-occurred with this ssDNA binding protein, which showed a trend towards association with the generalist phenotype and suggest additional possible members of this pathway (Supplemental Table S2). Co-occurring proteins show similarity to a maintenance endonuclease and a reverse transcriptase, proteins possibly involved in DNA metabolism and repair. Conversely, a protein structurally similar to a bacterial transcriptional repressor (PDB reference 1y7y) was found more often in specialist phages (Fig. 2C, D); the basis for this association is unknown. Multiple sequence alignments of each protein cluster revealed higher average entropy in the specialist protein cluster than the generalist protein cluster (1.6 vs 1.85 Shannon entropy), consistent with the need for more specific adaptation to each host.

Fig. 2.

Examples of viral proteins associated with host range generalists (> one host genera) or specialists (only one host genus). A A protein cluster (internal ID:”out_16_dedupe_26503@Chunk_0_35″, closest PDB reference: 6bhx) found more often in the genomes of viruses with generalist host range (Fisher’s exact FDR < 0.05). B The core domain is predicted to be a single stranded DNA binding protein. The protein from the present work is shown in blue, with the close reference single-stranded binding protein in brown. C A cluster (internal ID: “vOTU1462643_dsDNAphage_2″, closest PDB reference: 1y7y) found more often in specialist viruses (Fisher’s exact FDR < 0.05). D The core domain shows similarity to an HTH-type transcriptional repressor. The protein from the present work is shown in blue, with the close reference HTH repressor in brown

The mouse virome showed a high diversity of bacterial hosts, with phages from an average of 96 different genera in each sample. Phages infecting the unnamed genera CAG-873 and COE1 were the most common, as well as phages infecting Eubacterium, Duncanella, Acetatifactor, and Lachnospiraceae (Supplementary Fig. S4).

Most contigs were inferred to be from temperate phages, meaning those capable of integrating into the bacterial genome. Twenty percent of phages were directly observed as integrated prophages, as identified by CheckV (see methods), while 53% of phages were annotated as likely temperate using the prediction tool PhaTYP [46] (Fig. 1C). The remaining phages were predicted to be virulent, meaning they carry out only a lytic replication cycle. Reads from lytic phages showed a relative abundance of 14.6% (mean value; sd = 7.5%).

Parallel structure in the mouse phage dsDNA virome and bacteriome

We observed parallels in structure between the dsDNA phage virome and the bacterial microbiome. Figure 3A compares results for the dsDNA phage virome and bacteriome as principal coordinate analyses based on Bray–Curtis distances derived from relative abundances (Fig. 3A). After Procrustes transformations, samples in the phage virome clustered more closely with their corresponding sample in the bacteriome than expected by chance (p < 0.001, Fig. 3B). This suggests potential influence of the virome and bacteriome on each other, and that they may be similarly influenced by factors such as diet and aging. Constructing principal coordinates using only temperate phages or only virulent phages showed that the phage virome-bacteriome Procrustes distance was shorter when only considering temperate phages (Fig. 3C). This is as expected if integrated prophages are present proportionally to the genomes of their bacterial hosts. Nevertheless, a high degree of overall similarity is still seen between the phage virome and bacterial microbiome even when analyzing only virulent phages (Supplementary Fig. S5, Procrustes p < 0.001).

Fig. 3.

Covariation of the mouse dsDNA phage virome and the bacterial microbiome. A Bray–Curtis PCoA plot of bacteriome and phage virome, shown after Procrustes transformation. The amount of variation captured on each axis is indicated. B The bacteriome-virome distance is shorter within the same samples than expected by chance (Procrustes p < 0.001). The x-axis shows the bacteria-phage distance, the y-axis shows relative frequency as density. C Relative bacteriome-phage virome distance is shorter when only considering predicted temperate phages compared to those predicted to be virulent. The y-axis summarizes the Procrustes distances. D Most phages have at least one host with which they are highly correlated, regardless of predicted replication style. The x-axis shows the predicted bacteria-phage distance, the y-axis shows the number of vOTUs. The surrounding graphs show three examples of the supporting data, with normalized bacterial abundance on the x-axis and the normalized phage abundance on the y-axis

We next quantified associations between each viral OTU and each bacterial species. Many phage genomes showed a nearly perfect correlation in relative abundance to one or more bacterial species (Fig. 3D). In some instances, the virus appeared ubiquitous when the bacteria were present. For example, one phage of Acetatifactor sp. was present in 99.3% of the samples in which this bacteria was present, with an R² of 0.94 between their relative abundances. This may suggest the virus has existed as a prophage in the host genome for a long time, undergoing prophage domestication, or providing some fitness advantage to the host, while retaining enough phage genes to be recognizably viral in origin [59]. Alternatively, the high prevalence could be due to founder effects and bottlenecking in a controlled mouse facility where a given species might only have a single strain present, also resulting in near-perfect correlation. In other instances, a high degree of correlation is observed in a subset of samples, but the phage is absent in others, suggesting a more recent association or loss (Supplementary Fig. S2). We saw no greater bacteria–phage relative abundance correlation among temperate phages compared to virulent ones (p = 0.2). This could partially be explained by technical noise in phage replication cycle prediction, but may also indicate that growth of lytic phages is as tightly controlled by host abundance as that of temperate phages [13].

The high degree of correlation we observed between the phage DNA virome and bacteriome suggests that the majority of variation in the virome could be explained by bacterial abundances alone. To test this, while accounting for phages that might have more than one bacterial host, we constructed a random forest classifier using the bacteriome to predict the relative abundance of each phage (Fig. 4A). For most phages (59%), bacterial relative abundance alone explained a majority of the variation. Prediction accuracy varied by replication mode, with temperate phages having lower absolute error than virulent phages (p = 2.7e−4, Fig. 4B). This suggests that the abundance of bacterial hosts is the dominant factor in shaping the murine DNA virome, particularly temperate phage abundance.

Fig. 4.

Predictability of phage relative abundance from bacterial relative abundance. To account for phages with multiple hosts, or unknown hosts, a random forest model was trained to predict phage relative abundance from all bacterial relative abundances. One model was trained for each phage. Samples were split into 70% training set and 30% test set, keeping samples from the same mice and same cages within the same set. A Histogram of percent of variance explained by each random forest model. Abundance of most phages are highly predictable based on bacterial abundances. B High quality temperate phage genomes (those observed as prophages via CheckV) have lower prediction error compared to virulent phages and those just predicted to be temperate by PhaBox

To assess the generality of these observations, we quantified phage relative abundance from a publicly available mock community engrafted into mice [60]. We identified prophages in bacterial genomes from this community using geNomad and quantified relative abundance using the bwa aligner (v 0.7.18) [38, 45]. Coverage of most phages showed a strong correlation to the coverage of bacterial hosts, with the majority of identified phages having an R² greater than 0.5 (Supplementary Fig. S6). This is in line with previous analysis of these data, suggesting very little viral induction [21], and reinforces our finding that the abundance of most phages is tightly controlled by host abundance in the mouse gut.

Alterations in the virome with dietary restriction

We next investigated how the DNA phage virome varies with dietary restriction. In the most extreme diet, 40% caloric restriction, more than half of all vOTUs significantly changed in relative abundance at an FDR cutoff of 0.01 (Fig. 5A). Across all restricted diets, more phages decreased in relative abundance than increased. This may indicate that resource scarcity creates few winners while negatively impacting most taxa, though the compositional nature of microbiome data only specifies relative and not absolute phage abundances.

Fig. 5.

Changes in the mouse dsDNA phage virome associated with dietary restriction. A Graph of the fraction of vOTUs that changed in relative abundance (y-axis) in response to each diet (x-axis) at FDR < 0.01. B Random forest predictions of diet from bacteriome, dsDNA phage virome, or both show similar accuracies. Actual diets are shown along the x-axis, diets predicted from the data are shown on the y-axis. The color code shows the fraction of correct calls. C Graph of the fraction of vOTUs that changed in prevalence in response to each diet at FDR < 0.01. D Example of a bacterial species for which one phage became less common while another became more common upon dietary restriction. The X axis indicates the diet group, the Y axis shows the percent of mice that carry Ligalactibicillus murinis that do or do not carry the phage. E Genome lengths of viruses changing in prevalence upon dietary restriction. Longer viruses tend to become less common, while shorter viruses become more common. F The fraction of viral reads coming from virulent OTUs was associated with dietary restriction, and highest in 2-day fasting mice (Kruskal–Wallis p = 1.6e−9). The x-axis shows the diet group, the y-axis shows the precent of reads inferred to be from virulent vOTUs

To determine the overlap between the virome and bacteriome response to diet, we used each data type independently to predict diet from the microbiome. Both showed similar accuracy (~ 60%), and similar accuracy as a model trained using both data types (Fig. 5B). This indicates that the response to diet is largely shared between bacteria and phages in the mouse microbiome, with negligible additional predictive power added from combining the data types. Subsetting the data to only virulent phages decreased this accuracy to 51%, primarily associated with an inability to distinguish 1-day fasting mice from ad-libitum (Supplementary Fig. S7). There was still no increase in predictive power from a model using both the virome and microbiome, again indicating minimal evidence of independent signals affecting their relative abundances.

To identify changes in the DNA phage virome that might be independent of changes in the bacteriome, we took advantage of the high correlation between phage and host abundances to identify pairs where the prevalence of the phage was variable. We tested for instances where the prevalence of a phage within a given host changed in association with diet (Fig. 5C). In one such example, Ligilactobacillus murinus carried vOTU_42999 about 50% of the time in ad libitum mice, but only 20% of the time in extreme dietary restriction. Another phage, vOTU_42713, became more common in L. murinus upon dietary restriction, increasing from 20 to 75% prevalence (Fig. 5D). In total, around 5–10% of phages showed such changes across the four types of dietary restriction (Fig. 5C). We hypothesize that these phages confer a fitness advantage or cost to the host during dietary restriction.

Within this group of differentially prevalent phages, we identified no obvious taxonomic units nor specific phage proteins that might explain their cost or benefit. Rather, phages found less often under caloric restriction tended to have longer genome lengths compared to those that became more common (Fig. 5E, Kruskal–Wallis p = 0.03). We hypothesize that when resources become scarce, longer phages are more of a fitness burden, perhaps resulting in bacteria with shorter phages outcompeting those with longer ones. We observed a trend toward longer phages decreasing in each of the other diets relative to ad-libitum feeding, but only the 40% caloric restriction group achieved statistical significance. Further studies in a model system are required to understand the mechanism of this association and to confirm the direction of causality. Lastly, we observed an increase in the percentage of phage reads coming from predicted lytic phages under dietary restriction, especially in the 2-day fasting group (Fig. 5F, Kruskal–Wallis p = 1.6e−9).

Alterations in the virome with aging

We measured a decrease in phage richness with aging (Fig. 6A, p = 2.15e−6, linear mixed effects model) as was observed previously in human cohorts [6]. We also measured a slight increase in the percentage of reads coming from lytic phages, averaging ~ 2% per year (Fig. 6B, p = 0.001, linear mixed effects model). Compared to the response to diet, fewer phages changed in relative abundance in response to aging (Fig. 6C).

Fig. 6.

Changes to the phage virome with aging. A Phage richness decreases with age (linear mixed effects model p = 2.15e−6). Ages is shown on the x-axis, viral richness on the y-axis. B Change in the percentage of lytic phages over time. Relative abundance of inferred lytic phages (y-axis) increased with age (x-axis), at a rate of 2% per year (linear mixed effects model p = 0.001). C Fractions of vOTUs with relative abundances associated with age or chronological time at FDR < 0.01. D Random forest prediction of age from bacteriome, dsDNA phage virome, or both shows similar accuracy. Actual age is shown on the x-axis, predicted age is shown on the y-axis. E Fraction of vOTUs that changed in prevalence with age and time. F Example of a bacterial species for which one phage became less common while another became more common over the murine lifespan. Age is shown on the x-axis, while the y-axis shows the percent of mice with Muribaculaceae_NOV.MGBC128991 that do or do not carry the indicated phage

Past work on the microbiome of aging mice showed that random drift in the mouse facility, represented by including chronological time in a linear model, could actually account for many apparently age-associated changes [28]. In this study, analysis of multiple successive cohorts of mice allows us to account for age and time separately. We found many more phage taxa changing in relative abundance over time—associated with facility drift—rather than from aging specifically (Fig. 6C). Nevertheless, the virome was still able to predict mouse age about as well as the bacteriome, and as well as both data types combined (Fig. 6D).

As with the effects of diet, we found instances where specific phages became more or less common with age (Fig. 6E). In a particularly striking example, one Muribaculaceae sp. associated virus, vOTU_1515, transitioned from 15% prevalence at five months to over 95% prevalence by old age, while an unrelated phage, vOTU_29605, was virtually eliminated from the same Muribaculaceae sp. over the murine lifespan (Fig. 6F). Such examples accounted for 17% of measured phages (Fig. 6E).

Associations with mouse host genetics

Prior work demonstrated that a substantial fraction of microbiome variation was explained by the genetics of the mouse host [30]. We thus investigated whether such associations could be found with the phage virome as well. Indeed, 27% of vOTUs were associated with murine kinship, meaning a mixed-effects model accounting for murine kinship explained vOTU relative abundance better than a model ignoring kinship (FDR < 0.01). This value is smaller than what was reported for the bacterial covariation with kinship, which was 65%. These findings are consistent with the idea that vOTU abundance is related to murine kinship primarily via bacterial host abundance.

Discussion

In this work, we described the dynamics of the mouse dsDNA phage virome in the DRIDO cross, a large cohort of 913 outbred mice studied over their lifespan and on different diets. We constructed a database of mouse dsDNA gut viruses, all of which were likely bacteriophages, and assigned bacterial hosts. We found that most observed phages were highly correlated to the relative abundance of their bacterial hosts, and changes in the virome largely paralleled those in the bacteriome. We observed changes in the virome in association with dietary changes and with aging. By controlling for the high degree of virus-host correlation, we were able to identify modest virome-specific changes. Our data indicate that replication strategies in the mouse gut virome are primarily driven by “piggyback-the-winner” dynamics, where most phages are relatively quiescent prophages, and responses to environmental changes are modest over the phages studied.

Our finding of low rates of prophage induction in mouse gut is consistent with parallel studies in the human gut as well. A recent meta-analysis of human gut viromes reported a low rate of viral induction (as measured by viral particle numbers in stool) despite a high abundance of phage genomes (as measured by whole-stool metagenomic sequencing) [21]. The same study replicated these results in germ-free mice colonized with a human mock community, suggesting the mouse gut may have similar selective pressures favoring temperate phages [16]. The present study extends this finding to extreme perturbation—even 40% caloric restriction, which resulted in considerable weight loss and frailty [29], did not drive notable induction for most species of phage.

We found that 5–10% of phages became more or less common upon dietary restriction, independent of changes in relative abundance or prevalence of their host bacteria. The strongest effects were seen with the most extreme diet (40% caloric restriction). Phages that became less common were enriched for larger genomes, suggesting that the fitness cost of a phage may be proportionate to its genome length when resources are scarce. This differs from the dominant view of phage fitness, wherein rates of induction are the main driver of fitness cost [61, 62]. Further work using phages identified via virome-enrichment methods would be valuable to rule out technical effects such as incomplete assembly leading to shorter genomes in dietary restriction. Conversely, phages that became more common in dietary restriction represent a diverse set of phages that warrant further exploration for the fitness advantage they may confer to their hosts.

We identified far fewer changes in the virome upon aging than upon dietary restriction, a result that mirrors the bacteriome of the same mice [29]. We found an increase in the percent of lytic phages with age, which has been previously reported in human populations [33]. Additionally, we found a slight but significant decrease in phage richness with age, mirroring data from human cohorts and non-human primates [6, 63]. We also observed that 10% of phages became relatively less common with age, while only 5% increased with age. This is consistent with a model wherein phages are lost at a faster rate than they are replaced, explaining the overall decrease in richness with time.

Murine host genetics was associated with phage relative abundance as well [30]. Given the tight correlation between many phages and their bacterial hosts, we expect that the influence of genetics is largely driven by influences on bacterial hosts, rather than direct interaction between the mouse and the phage virome. Computational constraints in logistic mixed-effects modeling obstructed more detailed analysis, and would be a useful area for methodological development.

We also identified a phage protein associated with predicted broad host range. The protein, a single-stranded DNA binding protein, has been implicated in DNA repair after CRISPR-mediated DNA cleavage. This was associated with a more generalist host strategy and suggests a potential mechanism of generalist phage anti-host defense: repairing breaks quickly to overcome defense systems. Future work is necessary to identify other components of such a system or phages that use similar anti-defense strategies via other proteins. In general, our finding that vOTUs with broad host range are highly prevalent in the mouse gut is fitting with recent work in multiple ecosystems documenting potential broad host range for some phages [64, 65].

This study had several limitations. We only assessed the DNA virome of the mouse gut, excluding all RNA viruses and ssDNA viruses. Future work is necessary to determine the replication strategies favored by these viruses [66]. We recovered almost exclusively phage in the Caudoviricetes class, potentially missing other less common viruses because of poor coverage in our sequence data and in relevant databases. It would also be useful to investigate whether laboratory-reared mice lack phages from the ubiquitous Crassvirales order, the most common phage in many human gut viromes [13], or whether they were not detected here for technical reasons. It also would be useful to carry out additional studies based on virus enrichment before sequence acquisition and studies based on absolute rather than relative quantification. Finally, it is unclear how much the virome composition and dynamics found here are specific to laboratory-reared female animals. It would be useful to repeat aspects of this study with “re-wilded” mouse models of both sexes.

In conclusion, we took advantage of a large dataset of mouse metagenomes to characterize dsDNA phages in the mouse gut. This work established what we hope will be a useful resource of murine phage genome sequences and introduced novel analytical approaches while emphasizing the dominant “piggyback-the-winner” dynamic in the gut.

Authors’ contributions

CM, LL, CAT and FDB designed the study. CM, LL, CAT, RGC and FDB collected, analyzed and interpreted the data. CM and FDB drafted the manuscript. CM, LL, CAT, RGC and FDB reviewed the manuscript. CM and LL conducted statistical analysis. CAT, RGC and FDB supervised the study.

Funding

This work was supported in part by The PennCHOP Microbiome Program (CM, LL, CT, RGC, FDB, KB), U19AI174998 (FDB, KB), R01LM014503 (FDB), P30AI045008 (FDB), U54AG089323 (RGC, FDB, KB), and F31HL170550 (CM). LL was supported by the Blavatnik Family Fellowship in Biomedical Research and T32HG000046. CAT is a Pew Biomedical Scholar and a Burroughs Wellcome Fund Investigator in the Pathogenesis of Infectious Diseases, and is supported by NIH DP2-AG-067492, NIH R01-DK-129691, NIH R01-NS-134976, NIH DP1-DK-140021, the Kenneth Rainin Foundation, a McKnight Brain Research Foundation Innovator Award, the Human Frontier Science Program (HFSP), and the Penn Institute on Aging.

Data availability

Sequences used in this study were downloaded from NCBI Sequence Read Archive (Bioproject Accession: PRJNA1054518). Code used is available from github. https://github.com/cmerenstein/drido_mouse_virome. Code from the original DRiDO analysis, and necessary metadata are available: https://github.com/levlitichev/DRiDO_microbiome

Extended data and large downloads are available from zenodo: 10.5281/zenodo.15312368.

Declarations

Ethics approval and consent to participate

All data analyzed in this paper has been previously published.

Consent for publication

No data is presented from an individual person.

Competing interests

The authors declare no competing interests.