Improving gut virome comparisons using predicted phage host information

Abstract

The human gut virome is predominantly made up of bacteriophages (phages), viruses that infect bacteria. Metagenomic studies have revealed that phages in the gut are highly individual specific and dynamic. These features make it challenging to perform meaningful cross-study comparisons. While several taxonomy frameworks exist to group phages and improve these comparisons, these strategies provide little insight into the potential effects phages have on their bacterial hosts. Here, we propose the use of predicted phage host families (PHFs) as a functionally relevant, qualitative unit of phage classification to improve these cross-study analyses. We first show that bioinformatic predictions of phage hosts are accurate at the host family level by measuring their concordance to Hi-C sequencing-based predictions in human and mouse fecal samples. Next, using phage host family predictions, we determined that PHFs reduce intra- and interindividual ecological distances compared to viral contigs in a previously published cohort of 10 healthy individuals, while simultaneously improving longitudinal virome stability. Lastly, by reanalyzing a previously published metagenomics data set with >1,000 samples, we determined that PHFs are prevalent across individuals and can aid in the detection of inflammatory bowel disease-specific virome signatures. Overall, our analyses support the use of predicted phage hosts in reducing between-sample distances and providing a biologically relevant framework for making between-sample virome comparisons.

Importance: The human gut virome consists mainly of bacteriophages (phages), which infect bacteria and show high individual specificity and variability, complicating cross-study comparisons. Furthermore, existing taxonomic frameworks offer limited insight into their interactions with bacterial hosts. In this study, we propose using predicted phage host families (PHFs) as a higher-level classification unit to enhance functional cross-study comparisons. We demonstrate that bioinformatic predictions of phage hosts align with Hi-C sequencing results at the host family level in human and mouse fecal samples. We further show that PHFs reduce ecological distances and improve virome stability over time. Additionally, reanalysis of a large metagenomics data set revealed that PHFs are widespread and can help identify disease-specific virome patterns, such as those linked to inflammatory bowel disease.

Keywords: bacteriophages; bioinformatics; gut microbiome; microbial interactions; virome.

Fig 1

Computationally predicted bacterial hosts for vOTUs are concordant with in situ associations to the bacterial family level. Agreement between iPHoP predicted host range and Hi-C assigned host range at various taxonomic ranks for 1,243 vOTUs. Additional comparisons were made when iPHoP predicted multiple hosts for a vOTU (see main text for details on the three comparisons).

Fig 2

PHFs reduce interindividual variation and increase intraindividual virome stability in a cohort of 10 healthy individuals. Data were analyzed from a previously published study of 10 healthy individuals (1). (A) Taxonomic bar plots of virome composition at the PHF level for each individual over time. Facet labels above the bar plots correspond to the subject IDs from the original study. (B) Ecological distances between samples with Bray-Curtis at the contig level, Bray-Curtis at the PHF level, and Weighted UniFrac at the PHF level. Interindividual and intraindividual comparisons are both shown. Significance was assessed using the Friedman test with the post-hoc Wilcoxon signed-rank test, using Bonferroni correction for multiple comparisons (***P < 0.001, ****P < 0.0001). (C) Virome stability, defined here as (1—ecological distance from the previous sample), was calculated for each individual using the Bray-Curtis distance metrics at the contig level and at the PHF level. Significance was assessed using the Wilcoxon signed-rank test (*P < 0.05, **P < 0.01, ***P < 0.001).

Fig 3

PHFs are prevalent and reduce intra- and interindividuality in a large human IBD cohort. Data were analyzed from the previously published HMP2 data set (13). Samples with low viral read counts (<1,500) were removed from analyses. In total, bulk metagenomes from 1,093 samples from 115 individuals (57 CD, 31 UC, 27 non-IBD controls) were included for downstream analyses. (A, B) Rank prevalence distributions of vOTUs (A) and PHFs (B) across individuals. In total, there were 3,870 distinct vOTUs and 74 distinct PHFs. The dotted red line indicates the rank at which features are more, or less than, 50% prevalent. (C) Mean relative abundance of features (PHFs vs vOTUs) that were present in more than 50% of individuals in the data set. (D) Bray-Curtis distance between samples according to interindividual or intraindividual comparisons. Significance was assessed using the Wilcoxon signed-rank test (****P ≤ 0.0001).

Fig 4

PHFs reveal disease-specific signatures of IBD. Data were analyzed from the previously published HMP2 data set (13). Samples with low viral read counts (< 1,500) were removed from analyses. (A) PCoA plots generated from Bray-Curtis distance matrices using vOTUs (left) and PHFs (right). Samples are color-coded according to the dysbiotic status identified in (13). (B) Differentially abundant PHFs based on dysbiosis status. Only individuals which had both a dysbiotic and non-dysbiotic sample were included. Only PHFs that were more than 50% prevalent across individuals were considered for these analyses. PHFs with an adjusted P value ≤ 0.05 and with a log₂ fold-change ≥1 or with a log₂ fold-change ≤ −1 were considered differentially abundant.