Gut virome remodeling after fecal microbiota transplantation is associated with virus–bacteria co-variation and metabolic response in obesity and type 2 diabetes

ABSTRACT

Fecal microbiota transplantation (FMT) modifies gut bacterial composition and improves metabolic outcomes in subjects with obesity and type 2 diabetes (ObT2), but bacterial changes alone do not fully explain the heterogeneity in clinical response. We performed virus-like particle (VLP)-enriched metagenomic sequencing coupled with bulk-DNA metagenomic sequencing on longitudinal fecal samples from a randomized controlled FMT trial in ObT2. Using random forest classification, post-FMT virome changes showed accuracy comparable to bacteriome changes (84.8% vs 81.4%) in distinguishing LDL responders from non-responders. Donor-derived viral contig engraftment was detected in 64.7% (11/17) of recipients receiving FMT alone and 58.8% (10/17) receiving FMT with lifestyle intervention (LSI), versus 5.6% (1/18) in the sham with LSI group. FMT was associated with increased virome diversity, compositional shifts toward the donor virome, and coordinated virus–bacteria associations, with most responsive viral contigs classified as Caudoviricetes. FMT-responsive viral contigs that were shared between the two FMT groups were predicted to infect Roseburia and Bifidobacterium, while contigs associated with clinical responses were predicted to infect Escherichia, Roseburia, Blautia, and Enterococcus. These findings suggest that the gut virome is a complementary ecological component associated with FMT responses and metabolic outcomes in ObT2.

Introduction

The gut microbiome is recognized as a key mediator of metabolic dysfunction in obesity and type 2 diabetes mellitus (T2DM), ^1-3 with microbiome-targeted therapies offering a promising therapeutic avenue. ^4-7 In our previous randomized controlled trial, repeated FMT promoted sustained engraftment of donor-derived microbiota and was associated with improvements in low-density lipoprotein (LDL) cholesterol and liver steatosis in obese individuals with T2DM. ⁸

However, the gut microbiome extends beyond bacteria. The gut virome, composed predominantly of bacteriophages that infect bacteria, is increasingly recognized as an active ecological agent capable of reshaping microbial community structure through lytic and lysogenic cycles. ^9-13 Our group previously demonstrated that T2DM are associated with reduced gut virome diversity and disrupted virus–bacteria interactions. ¹⁴ In the context of FMT, mounting evidence suggests that the gut virome is “not just a passenger” but may play active roles in treatment response. ^{15 , 16} Bacteriophage composition has been linked to microbial engraftment and treatment response following FMT. ^{17 , 18} A randomized controlled trial using sterile fecal filtrate transplantation showed that transfer of bacteriophages alone can transiently reshape gut microbiota and phage-microbe interactions. ¹⁹ Moreover, preclinical studies have further demonstrated that fecal virome transplantation (FVT) alone can alleviate obesity and T2DM-related metabolic features in mouse models, supporting a functional role of the gut virome. ^{20 , 21} More recently, human gut-derived virus-like particle (VLP) transplantation was shown to improve glucose tolerance while reshaping both the gut microbiota and virome in diet-induced obese mice, further strengthening the evidence that the gut virome may actively contribute to host metabolic regulation. ²²

Despite these advances, the contribution of the gut virome to FMT outcomes in human metabolic disease remains poorly defined. Existing human studies are limited in scale, and no randomized controlled trial has systematically characterized virome dynamics following FMT in obese T2DM populations. Building on our previous randomized controlled trial, we conducted VLP-enriched metagenomic sequencing on longitudinal fecal samples to characterize gut virome dynamics following repeated FMT. We hypothesized that: (1) repeated FMT is associated with sustained engraftment of donor-derived viral contigs; (2) FMT is associated with altered virus‒bacteria associations; and (3) specific viral contigs are associated with metabolic responses to FMT.

Methods

Study cohort

This study is a mechanistic extension of previously published RCT investigating the effects of FMT in obese subjects with type 2 diabetes (ObT2). ⁸ The original trial recruited adults aged 18‒70 y with a body mass index (BMI) ≥28 kg/m² and <45 kg/m² and a diagnosis of T2DM for at least 3 months. Key exclusion criteria included recent use of antibiotics or probiotics, severe comorbidities, and contraindications to endoscopic procedures, as previously described. Participants were randomly assigned to receive FMT with lifestyle intervention (LSI), FMT alone, or sham with LSI. Per protocol, participants were asked to maintain stable doses of medications during follow-up, and any medication changes were recorded. FMT solution was prepared from six healthy lean donors. Clinical follow-up and biospecimen collection were performed at baseline and longitudinally up to week 24. Detailed study procedures, donor screening, FMT preparation, and clinical assessments have been described previously. All patients gave written informed consent. The original trial was approved by the Joint Chinese University of Hong Kong New Territories East Cluster Clinical Research Ethics Committee (2016.136-T). The study was registered with ClinicalTrials.gov (NCT03127696).

Fecal VLP DNA extraction and sequencing

Archived stool samples collected during the original trial were used for VLP-enriched metagenomic sequencing. The detailed DNA extraction procedures can be found in our previous study. ¹⁴ Briefly, stool (100–200 mg) was added to 400 μL saline-magnesium buffer and vortexed for 10 minutes. Suspensions were centrifuged at 2000 g to remove debris and cells. The supernatant was passed through 0.45- and 0.22-μm filters. To remove residual bacterial and host cell membranes, samples were treated with lysozyme (1 mg/mL at 37 °C for 30 minutes), followed by chloroform (0.2 × volume at room temperature for 10 minutes). Non-virus-protected DNA was removed by treatment with 1 U Baseline-Zero DNase (Epicenter), followed by heat inactivation at 65 °C for 10 minutes.

To extract nucleic acids from VLP, samples were cleaved with 4% sodium dodecyl sulfate plus 38 mg/mL proteinase K at 56 °C and then treated with CTAB buffer and phenol-chloroform-isoamyl alcohol (pH 8.0). The aqueous portion was washed once with an equal volume of chloroform, followed by concentration using a DNA Clean & Concentrator kit (Zymo Research). VLP DNA was amplified for 2 hours using Phi29 polymerase before sequencing. Sequencing libraries were prepared from 1 μg DNA per sample using the NEBNext Ultra DNA Library Prep Kit (Illumina) and sequenced on an Illumina NovaSeq 6000 platform with paired-end 150 bp reads, generating a median of 49.8 million reads per sample.

Sequencing data processing, annotation, and quantification

Shotgun metagenomic reads were quality-filtered and decontaminated of human sequences using KneadData (https://bitbucket.org/biobakery/kneaddata/wiki/Home), which integrates Trimmomatic ²³ (v.0.39.1) and Bowtie2 ²⁴ (v.2.3.4.3). Reads mapping to the human reference genome (hg38) were discarded. To assess the purity of VLP-enriched libraries, reads were taxonomically classified using Kraken2 ²⁵ (v.2.1.3) against the PlusPF reference database (build 2026-06-26), with species-level abundances re-estimated using Bracken ²⁶ (v3.0.1). The proportions of bacterial, archaeal, fungal, viral, human, and unclassified reads were summarized to evaluate the extent of non-viral DNA contamination.

Paired-end VLP reads were assembled into contigs using MEGAHIT ²⁷ (v.1.0.3), yielding 349,698 contigs longer than 1000 bp. Contigs with an estimated completeness ≥50% and contamination ≤10% (medium quality or above; n = 3293) by CheckV ²⁸ (v1.0.1) were retained. The quality-controlled viral contigs were dereplicated at 99% average nucleotide identity (ANI) and 90% alignment fraction (AF) to generate a representative viral contigs set (n = 2588).

Representative viral contigs were taxonomically annotated using geNomad ²⁹ (v1.11.1; genomad_msa_v1.8) in end-to-end mode. Complementary protein-homology-based taxonomic annotations were obtained using DIAMOND ³⁰ (v2.1.10) searches against the RefSeq viral protein database (retrieved June 24, 2026), followed by taxonomic annotation using MEGAN ³¹ (v6.25.10). Putative host associations were inferred using iPHoP (v1.3.3), and viral lifestyles were predicted using PhaTYP ³² based on a database merged from “Sept_2021_pub_rw” and the MAGIC database (https://zenodo.org/records/13989345). ³³

To quantify the abundance of viral contigs, quality-filtered metagenomic reads from both bulk and VLP-enriched sequencing were competitively mapped against the representative viral contigs set using CoverM ³⁴ (v0.7.0) in genome mode with Strobealign (v0.17.0). Genome-level abundance was calculated as reads per kilobase million (RPKM). To avoid bias associated with Phi29 amplification of VLP libraries, RPKM values derived from bulk metagenomic reads were used and transformed into relative abundance for downstream analyzes.

Viral contigs detected in recipient post-FMT samples were classified into four categories based on their presence in the recipient baseline and the donor: (1) donor-derived: absent in recipient baseline but present in donor; (2) recipient-exclusive: present in recipient baseline but absent in donor; (3) co-existing: present in both recipient baseline and donor; and (4) unknown: absent in both recipient baseline and donor. The ≥20% threshold for donor-derived viral contigs was adapted from the microbiota engraftment framework of the original trial. ⁸

Machine learning classification of LDL response

To compare the classification performance of post-FMT changes in the gut virome versus the bacteriome for LDL response (≥0.5 mmol/L reduction at week 24), we performed random forest (RF) classification. RF was chosen for its ability to handle high-dimensional data, capture non-linear relationships, and provide interpretable variable importance measures. For both virome and bacteriome data, features present in at least 5% of samples with a mean relative abundance ≥0.01% were retained to remove low-abundance or rare taxa. Retained features were transformed using centered log-ratio (CLR), and Δ was calculated as the difference between week 24 and baseline (week 0). Recursive feature elimination (RFE) combined with RF was performed with 50 repeats to select stable features. Features selected in ≥50% of RFE repeats were defined as stable. Final RF classifiers were trained using these stable features with 500 trees per forest. Model performance was evaluated using stratified 5-fold cross-validation (CV), with the same folds applied to the virome and bacteriome models. Performance metrics included CV mean accuracy, Cohen's Kappa, and the area under the receiver operating characteristic curve (AUC), calculated from CV predictions to avoid overestimation. Sensitivity, specificity, and balanced accuracy were additionally calculated from cross-validated predictions to evaluate model performance under class imbalance. To assess whether differences in classification performance were influenced by feature-number asymmetry, an additional sensitivity analysis was performed by constructing both virome and bacteriome models using the same number of stable features selected by RFE. No class weighting or optimized classification threshold was used. Class imbalance was instead addressed using a resampling-based sensitivity analysis with the synthetic minority oversampling technique (SMOTE). SMOTE was applied only within the training portion of each stratified CV split, while the held-out validation folds retained the original class distribution. Variable importance was used to identify the top influential features for each model.

Statistical analysis

Statistical analyzes were performed in R (V.4.1.2). Machine learning analyzes were conducted using the caret, randomForest, pROC, and themis packages. Categorical variables, including the proportion of subjects with ≥20% donor-derived virome, were compared across the three groups using the Fisher‒Freeman‒Halton exact test, followed by pairwise Fisher's exact tests with Bonferroni correction for multiple comparisons if the overall test was significant. For longitudinal continuous variables, non-parametric repeated measures ANOVA was used to test group, time, and interaction effects. Metrics absent at baseline (e.g., donor-derived virome proportions) were analyzed using only post-FMT time points (weeks 4, 16, 24), whereas intrinsic ecological metrics (Shannon, Chao1) included baseline. Significant interactions were followed by Kruskal‒Wallis tests (between-group) or Friedman tests (within-group), with post hoc Mann‒Whitney U or Wilcoxon signed-rank tests as appropriate. All post hoc p-values were Bonferroni-corrected. Alpha diversity indices were calculated using the vegan package. Beta diversity was assessed using Bray-Curtis dissimilarity, and principal coordinates analysis (PCoA) was performed using the phyloseq package. Microbial relative abundance data were transformed using CLR transformation prior to downstream analyzes. Differential abundance analysis of viral contigs was performed using Multivariate Association with linear models (MaAsLin2). Viral contigs with a mean relative abundance of ≥0.01% combined with a prevalence of ≥5% were selected for MaAsLin2 analysis. Trans-kingdom correlations between viral contigs and bacterial species at week 24 were assessed using Spearman's rank correlation based on CLR-transformed data. Correlation networks were constructed using significant associations defined by |r| > 0.5 and p < 0.05. Hub nodes were identified based on node degree centrality. A two-sided p-value < 0.05 was considered statistically significant.

Results

A total of 61 subjects were randomly assigned to receive FMT with LSI, FMT alone, or sham with LSI. Among these, 52 subjects had paired VLP-enriched and bulk-DNA metagenomic data available at baseline and week 24, including 17 in the FMT with LSI group, 17 in the FMT alone group, and 18 in the sham with LSI group (Supplementary Figure 1). Baseline characteristics are summarized in Supplementary Table S1. No significant differences were observed between participants included in (n = 52) and excluded from (n = 9) the virome analysis (Supplementary Table S2). Of the 34 subjects who received FMT, 32 had longitudinal LDL measurements and were included in the RF model analysis. VLP-enriched sequencing and assembly yielded a representative set of 2588 viral contigs after quality filtering and dereplication. Among taxonomically classified reads in the VLP libraries, viral reads constituted the largest fraction overall (median 68.3%), followed by bacterial reads (median 29.5%), whereas fungal and archaeal reads were minimal (maximum 2.99% and 0.07%, respectively) and human-derived reads were low (median 0.08%). A median of 86.7% of VLP reads were recruited to the representative viral contigs catalog, indicating that the catalog captured a substantial proportion of the sequenced viral community in most samples. At the class level, Malgrandaviricetes comprised the largest proportion (50.3%), followed by Arfiviricetes (10.4%), Repensiviricetes (6.9%), and Caudoviricetes (5.1%), while 26.2% of contigs remained unclassified (Supplementary Figure 2). In contrast, Malgrandaviricetes contigs had a median relative abundance of only 0.7%, whereas Caudoviricetes accounted for 12.1% and were detected in all samples (Supplementary Figure 2). Comprehensive annotations and quality metrics of the contigs are provided in Supplementary File 1.

Post-FMT gut virome changes are associated with LDL response

Post-FMT remodeling of both the gut bacteriome and virome was associated with LDL response. Among the 32 FMT recipients (23 non-responders, 9 responders), changes in lipid-lowering and antidiabetic medications were examined as potential confounders and did not differ significantly between responders and non-responders (Supplementary Table S3). We identified 8 stable virome features and 40 stable bacteriome features through RFE with 50 repeats (selection frequency ≥50%). The virome-based RF classifier achieved a cross-validated accuracy of 84.8% (Kappa = 0.539, AUC = 0.916), while the bacteriome-based RF classifier achieved an accuracy of 81.4% (Kappa = 0.318, AUC = 0.899) (Figure 1). The virome model showed numerically higher balanced accuracy and comparable or higher sensitivity than the bacteriome model across the standard, feature-number-matched (Supplementary Figure 3), and SMOTE sensitivity analyzes (balanced accuracy: 0.756 vs. 0.667 in the standard analysis; 0.734 vs. 0.700 with SMOTE), while overall accuracy and AUC were broadly comparable (Supplementary Table S4). Feature importance analysis identified Caudoviricetes contig 278491 as the highest-ranked feature in the virome model (Figure 1a), while Bulleidia extructa was the highest-ranked feature in the bacteriome model (Figure 1b).

Figure 1.

Post-FMT gut virome changes are associated with LDL response. Random forest classifiers were trained on post-FMT changes in gut virome and bacteriome for LDL-response classification among 32 FMT recipients. (a) Stable features in the virome model. (b) Top 10 features in the bacteriome model. Model performance metrics (accuracy, Kappa, AUC) and the number of stable features are shown for both classifiers.

FMT is associated with sustained engraftment of donor-derived viral contigs independent of LSI

We next examined the engraftment of donor-derived viral contigs and changes in virome composition following FMT. At baseline, four viral contigs were enriched in donors and six viral contigs were enriched in recipients (MaAsLin2, p < 0.05; Figure 2a). Three of the four donor-enriched viral contigs and two of the six recipient-enriched viral contigs were classified as Caudoviricetes, whereas the remaining contigs were unclassified (Supplementary File 1). Notably, 64.7% of participants in the FMT alone group and 58.8% in the FMT with LSI group reached the 20% threshold, compared with only 5.6% in the sham with LSI group (overall p = 0.0005, Fisher‒Freeman‒Halton exact test). Post hoc pairwise comparisons with Bonferroni correction showed that both the FMT alone group (adjusted p = 0.0009) and the FMT with LSI group (adjusted p = 0.0028) had significantly higher proportions than the sham with LSI group, whereas no significant difference was observed between the two FMT groups (Figure 2b). Both FMT-treated groups had significantly higher proportions than the sham with LSI group at weeks 4, 16, and 24 (all adjusted p < 0.05; Figure 2c). No significant differences were observed between the two FMT groups at any time point (all adjusted p > 0.05) (Figure 2c). At week 24, donor-derived viral contigs accounted for a median of 21.9% of all detected viral contigs in the FMT alone group, followed by 21.3% in the FMT with LSI group and 7.1% in the sham with LSI group (Figure 2c). Within the FMT with LSI group, no significant difference was detected at week 24 between participants with fair-to-high (n = 7) and low (n = 10) compliance with LSI (Wilcoxon rank-sum test, p = 0.962; Supplementary Figure 4).

Figure 2.

FMT is associated with sustained engraftment of donor-derived viral contigs. (a) Baseline viral contigs enriched in donors vs. recipients with obesity and T2DM (MaAsLin2, p < 0.05). (b) Proportion of subjects achieving ≥20% donor-derived viral contigs. Fisher‒Freeman‒Halton exact test (overall p = 0.0005) followed by pairwise Fisher's exact tests with Bonferroni correction. (c) Proportions of donor-derived viral contigs at weeks 4, 16, and 24. Stratified between-group comparisons with Bonferroni correction were performed at each time point using Kruskal–Wallis tests followed by the Mann‒Whitney U tests.

Gut virome alpha diversity and compositional dynamics after FMT

Among the 46 participants with virome data available at all four time points, aligned rank transform analysis for repeated measures revealed a significant effect of time on Shannon diversity (p = 0.0009), whereas neither the group effect (p = 0.270) nor the group × time interaction (p = 0.261) was significant (Figure 3a). For Chao1 richness, the group × time interaction was significant (aligned rank transform analysis, p = 0.041), with a between-group difference detected at week 16 (Kruskal‒Wallis, p = 0.016). Post hoc comparisons showed that the FMT with LSI group had significantly higher Chao1 richness than the sham with LSI group (adjusted p = 0.024), whereas no other significant pairwise differences were observed (all adjusted p > 0.05) (Figure 3b). Within-group comparisons showed that both Shannon diversity and Chao1 richness were significantly increased in the FMT alone group at week 24, compared with baseline (both p = 0.003), whereas no significant changes were observed in other groups (Supplementary Figures 5 and 6). No significant temporal changes in beta diversity based on Bray‒Curtis dissimilarity to donors were detected in any group by Friedman tests, although PCoA suggested some compositional shifts towards the donors (Figure 3c and d).

Figure 3.

Gut virome alpha diversity and compositional dynamics after FMT. (a) Shannon diversity. Non-parametric repeated measures ANOVA revealed a significant effect of time (p = 0.0009), whereas neither the group effect (p = 0.270) nor the group × time interaction (p = 0.261) was significant. (b) Chao1 richness. Non-parametric repeated measures ANOVA identified a significant group × time interaction (p = 0.041). Stratified between-group comparisons showed a significant difference among groups at week 16 (Kruskal‒Wallis test, p = 0.016), with higher Chao1 richness in the FMT with LSI group than in the sham with LSI group (Bonferroni-adjusted p = 0.024); all other pairwise comparisons were not significant (adjusted p > 0.05). (c) Principal coordinates analysis (PCoA) of virome composition across treatment groups and time points. (d) Bray‒Curtis dissimilarity to lean donors.

At week 24, 62, 38, and 3 viral contigs were significantly altered compared with baseline in the FMT alone, FMT with LSI, and sham with LSI groups, respectively (MaAsLin2, p < 0.05; Supplementary Figures 7–9), comprising 19, 18, and 2 enriched and 43, 20, and 1 depleted contigs. Most altered contigs in the two FMT-treated groups were classified as Caudoviricetes. A total of 10 contigs were enriched in both FMT groups, of which 2 were predicted to infect Roseburia and Mediterraneibacter. Conversely, 15 viral contigs were depleted in both FMT groups, of which 2 were predicted to infect Bifidobacterium species (Figure 4a and b; Supplementary Tables S5 and S6).

Figure 4.

Coordinated remodeling of FMT-responsive viral contigs and virus-bacteria association networks following FMT. (a) Venn diagram showing the overlap of viral contigs significantly depleted at week 24 compared with baseline in the FMT alone and FMT with LSI groups (MaAsLin2, p < 0.05). (b) Venn diagram showing the overlap of viral contigs significantly enriched at week 24 compared with baseline in the FMT alone and FMT with LSI groups (MaAsLin2, p < 0.05). (c, d) Virus-bacteria association networks in the FMT alone group (c) and the FMT with LSI group (d). Each node represents a microbial taxon. Circles represent bacterial species, and inverted triangles represent viral contigs. Lines (edges) represent significant Spearman correlations between taxa. Red lines indicate positive correlations, and blue lines indicate negative correlations (|r| > 0.5, p < 0.05). Node size (degree) reflects the number of direct correlations for a given microbe, with larger nodes representing more highly connected hub taxa.

Trans-kingdom association networks following FMT

We next assessed the trans-kingdom virus–bacteria associations following FMT. In both FMT groups, viral hubs were predominantly classified as Caudoviricetes, with Anaerostipes hadrus and Faecalibacterium prausnitzii as shared bacterial hubs. The FMT alone group showed additional bacterial hubs including Anaerostipes caccae, Fusobacterium ulcerans, Lactobacillus spp., and Ruminococcus gnavus (Figure 4c). Several FMT-responsive Caudoviricetes contigs were associated with multiple FMT-responsive bacterial species (all |r| > 0.5, p < 0.05). For example, Caudoviricetes contig 148383, predicted to infect the butyrate-producing genus Roseburia, was inversely correlated with several FMT-responsive bacterial taxa, including Anaerostipes caccae and Lactobacillus spp.

The network was sparser in the FMT with LSI group (Figure 4d). Caudoviricetes contig 3655 was the major viral hub. It was positively correlated with Ruminococcus gnavus, Coprococcus catus, and Flavonifractor plautii, and inversely correlated with Roseburia hominis and Adlercreutzia equolifaciens. Notably, Faecalibacterium prausnitzii was positively associated with multiple Caudoviricetes contigs, placing this beneficial commensal at a central position in the network following FMT and LSI.

Specific viral contigs associated with clinical response to FMT with LSI

Finally, we explored viral contigs associated with clinical response in the FMT with LSI group, as subjects in this group exhibited the most pronounced improvements in serum LDL cholesterol and liver CAP score. ⁸ For LDL response, subjects were classified as responders (≥0.5 mmol/L reduction, n = 8) or non-responders (n = 9). At baseline, five viral contigs were enriched in responders and two in non-responders (MaAsLin2, p < 0.05; Supplementary Figure 10). Of these, one responder-enriched Caudoviricetes contig 288490 remained enriched at week 24 and was predicted to infect Escherichia. For liver CAP response, subjects were classified as responders (≥30 dB/m CAP reduction, n = 8) and non-responders (n = 9). At baseline, five viral contigs were enriched in responders and five in non-responders (MaAsLin2, all p < 0.05; Supplementary Figure 10). Several responder-enriched contigs were predicted to infect bacterial genera previously implicated in gut metabolic homeostasis, including Roseburia and Blautia, whereas one baseline non-responder-enriched Caudoviricetes contig 5905 was predicted to infect Enterococcus (Supplementary File 1).

Discussion

Using VLP-enriched metagenomic sequencing, this study provides a longitudinal characterization of gut virome dynamics following FMT in individuals with obesity and T2DM. We found that FMT was associated with sustained changes in virome composition, engraftment of donor-derived viral contigs, virus–bacteria association networks, and virome features associated with metabolic response. Together, these findings extend current evidence that the gut virome is a dynamic component of the intestinal microbial ecosystem following FMT and support its potential relevance to microbiome restructuring and metabolic response.

Independent analyzes of post-FMT virome and bacteriome remodeling showed that both microbial components were associated with metabolic response, with virome features also contributing to LDL response classification, consistent with the close ecological links between viral and bacterial communities. ^{35 , 36} Importantly, the virome signals observed here were supported by complementary sequencing and quantification approaches. VLP enrichment resulted in viral reads forming the largest fraction of taxonomically classified reads (median 68.3%), compared with bacterial reads (median 29.5%), while human, fungal, and archaeal reads remained minimal. To minimize potential bias from Phi29 amplification, VLP sequencing was used for viral contig discovery, whereas relative abundance was independently quantified using unamplified bulk metagenomic reads. At the class level, Malgrandaviricetes accounted for 50.3% of representative contigs but had a median relative abundance of only 0.7%, whereas Caudoviricetes represented only 5.1% of contigs but had a median relative abundance of 12.1%. This discrepancy highlights that contig representation does not necessarily reflect ecological abundance.

Donor-derived viral contig engraftment was detected following FMT and persisted through the 24-week follow-up. We previously reported engraftment of donor-derived bacteriophages following FMT in patients with Clostridioides difficile infection (CDI). ³⁷ The current findings extend this phenomenon to obesity, suggesting that viral engraftment following FMT may occur across distinct disease settings. Engraftment was comparable between the FMT groups, with no differences observed according to LSI compliance. This pattern mirrors the bacterial engraftment previously reported in the same cohort, ⁸ suggesting that viral and bacterial engraftment may follow similar dynamics. Although only the FMT alone group showed a significant within-group increase in virome alpha diversity, this should not be interpreted as attenuation of virome remodeling by LSI, given the small sample size and comparable donor-derived contigs between the FMT groups. Virome alpha diversity is highly heterogeneous across disease settings and may not represent a robust ecological signature, whereas changes in viral community composition and specific viral taxa are more consistent indicators of microbiome perturbation. ³⁸ Thus, donor-derived contigs engraftment, global virome diversity, and clinical improvement may capture distinct aspects of post-FMT ecological remodeling.

Caudoviricetes emerged as a prominent feature of post-FMT virome remodeling. Despite comprising only a small proportion of the viral catalog, this class was consistently identified among FMT-responsive contigs, virus‒bacteria association networks, and contigs associated with LDL and liver CAP responses. Caudoviricetes are abundant members of the human gut virome ³⁹ and infect diverse bacterial hosts, including genera associated with metabolic health, such as Prevotella, ^{40 , 41} Faecalibacterium, ⁴² and Roseburia. ⁴³ Alterations in Caudovirales-dominated phage communities have been reported in obesity and metabolic syndrome, ⁴⁴ and recent multicohort evidence showed increased abundance of Caudoviricetes in individuals with obesity. ⁴⁵ This observation is also consistent with our previous study showing enrichment of Caudovirales following FMT in patients with CDI. ³⁷

In line with this, predicted hosts of FMT-responsive Caudoviricetes contigs included members of the Lachnospiraceae family, such as Roseburia and Mediterraneibacter, as well as Bifidobacterium species. Trans-kingdom analyzes further linked these contigs with metabolically relevant bacteria, including Anaerostipes hadrus, Faecalibacterium prausnitzii, and Lactobacillus species, which have been associated with butyrate production, intestinal barrier integrity, insulin sensitivity, anti-inflammatory properties, and bile acid metabolism. ^46-51 Together with the association of several Caudoviricetes contigs with LDL and liver CAP responses, these findings highlight the potential relevance of Caudoviricetes to post-FMT virome remodeling in obesity, although their functional contribution to metabolic response requires mechanistic validation.

This study has several strengths, including its randomized, double-blind design, longitudinal follow-up, and integration of VLP-enriched and bulk metagenomic sequencing. However, several limitations should be acknowledged. First, the moderate sample size and the observational nature limit causal interpretation. Second, although Phi29-mediated amplification was used only for viral contig discovery, its preferential amplification of certain viral genomes, particularly small circular ssDNA viruses, may have biased contig recovery. ⁴⁴ Third, our analysis was restricted to DNA viruses and did not capture RNA viruses. Fourth, residual bacterial DNA cannot be entirely excluded despite decontamination. Finally, mechanistic studies are needed to establish how specific virus‒bacteria interactions influence metabolic responses to FMT.

In conclusion, our findings highlight the gut virome as an underappreciated component of post-FMT microbial remodeling in obesity and T2DM. Sustained donor-derived viral engraftment, virus‒bacteria associations, and virome features associated with clinical response support the relevance of the gut virome to FMT outcomes. These findings highlight an integrated virus‒bacteria‒host framework for understanding FMT and provide a foundation for future mechanistic studies.

Supplementary Material

Supplementary File 1_revised.xlsx

Supplementary File 1_revised.xlsx

Supplementary Table S1_revised.docx

Supplementary Table S1_revised.docx

Supplementary Table S2_revised.docx

Supplementary Table S2_revised.docx

Supplementary Table S6_revised.csv

Supplementary Table S6_revised.csv

Supplementary File 2_revised.pdf

Supplementary File 2_revised.pdf

Supplementary Figures and Legends_revised.pdf

Supplementary Figures and Legends_revised.pdf

Supplementary Table S5_revised.csv

Supplementary Table S5_revised.csv

Supplementary Table S4_revised.docx

Supplementary Table S4_revised.docx

Supplementary Material

Supplementary Table S3_revised.docx

Funding Statement

This study was supported by the InnoHK initiative of the Innovation and Technology Commission of the Hong Kong Special Administrative Region Government and Centre for Gut Microbiota Research, Faculty of Medicine, The Chinese University of Hong Kong.

Disclosure of potential conflicts of interest

FKLC and SCN are co-founders and shareholders of GenieBiome Holdings Limited and MicroSigx Diagnostic Holdings Limited. FKLC receives advisory fees from Comvita New Zealand Limited. SCN serves as an advisory board member for Pfizer, Ferring, Janssen and Abbvie. FKLC, SCN, ZX are named inventors of patent applications held by the CUHK and MagIC that cover the therapeutic and diagnostic use of microbiome and receive patent royalties through their affiliated institutions.

Data availability statement

The sequencing data of this article is available in the NCBI under Bioproject, PRJNA644456. All data and key analysis code needed to evaluate the findings are presented in the paper and/or the Supplementary Files 1 and 2.

Supplementary material

Supplemental data for this article can be accessed at https://doi.org/10.1080/19490976.2026.2746820.

Abbreviations

ObT2

obese subjects with type 2 diabetes

FMT

fecal microbiota transplantation

VLP

virus-like particle

LSI

lifestyle intervention

LDL

low-density lipoprotein

CAP

controlled attenuation parameter

T2DM

type 2 diabetes mellitus

BMI

body mass index

ANI

average nucleotide identity

AF

alignment fraction

RPKM

reads per kilobase million

CLR

centered log-ratio

PCoA

principal coordinates analysis

FVT

fecal virome transplantation

RF

Random Forest

RFE

Recursive feature elimination

CV

Cross-validation

AUC

Area under the receiver operating characteristic curve

SMOTE

Synthetic Minority Over-sampling Technique