The second phase of the HMP, also called the Integrative Human Microbiome Project (iHMP), consisted of focused studies of three diseases – Inflammatory Bowel Disease (IBD), Type II Diabetes (T2D), and Multi-Omic Microbiome Study: Pregnancy Initiative (MOMS-PI) ⁵ . The overall goal of the project was to identify associations between human microbiome community census data and the three diseases. Each of the studies were structured for the specific disease and consisted of separate cohorts.

Metaviz ⁶ is a web-based interactive visualization tool for microbiome data analysis. The architecture consists of a JavaScript front-end suite of charts (based on D3.js and Canvas) and a navigation component that lets users select portions of taxonomic hierarchies to visualize and analyze. Metaviz supports two backend data stores – a graph database and the metavizr R/Bioconductor package. Metaviz is tightly integrated with the metagenomeSeq statistical testing package so differential abundance testing results can be viewed directly in a Metaviz session. We host an instance of Metaviz that we call the UMD Metagenome Browser ( http://metaviz.cbcb.umd.edu).

Visualization tools for large-scale sequencing consortium projects provide a mechanism to explore and interact with data from multiple studies. These applications help users analyze individual datasets and examine trends across the entire project. MAGI is a web-application that enables a user to examine data from TCGA data ⁷ . The Earth Microbiome Project provides an interactive visualization web-application to analyze its data ⁸ . EMPeror offers interactive 3D visualizations of PCA plots to show distances between microbiome samples ⁹ . QIIME packages a number of tools for static plotting of Principal Coordinate Analysis and stacked bar plots ¹⁰ . MetaPhlAn2 uses a visualization package called GraphPhlan to produce phylogenetic trees and other plots ¹¹ . The HMP2Data R/Bioconductor provides processed 16S sequencing data from the iHMP project in Bioconductor data structures ¹² . We implemented Metaviz using design patterns from Epiviz ¹³ , an interactive epigenetics visualization tool, that visualizes data from a variety of epigenetic sequencing projects. We show how we leverage the microbiome measurement-based design of Metaviz to implement interactive exploration and hypothesis-testing of the iHMP resource.

The HMP Data Access and Coordination Center maintains a data repository and web portal ( https://www.ihmpdcc.org). From this web portal, users can browse metadata for datasets, raw sequencing files, and processed files including taxonomic community profile abundance matrices. We implemented several mechanisms to interact with the HMP data resources through Metaviz ⁶ .

We loaded the 16S community profile abundance matrices for the samples from the IBD, T2D, and MOMS-PI studies as provided by the HMP2Data Bioconductor package ¹² into the UMD Metagenome Browser. A user can select each dataset from the application start screen. Figure 1 details the number of samples, with metadata to the extent available as of May 2020 from the HMP2Data package, from each project currently available in the UMD Metagenome Browser.

When browsing the samples available from the HMP Data Portal, a user can view an individual abundance matrix in Metaviz using the Metaviz tool link from the file description page. When the user clicks the link, a redirect occurs to the UMD Metagenome Browser with a new workspace containing a FacetZoom navigation utility and a heatmap for that sample. Figure 1A shows the direct link functionality for samples in the IBD dataset and resulting workspace in Metaviz ( Figure 1B).

In the HMP data portal, a user can select files with a shopping cart utility and download the selections as a manifest file. In the UMD Metagenome Browser, the user can upload the manifest file to create a Metaviz workspace on the fly for those samples. Currently, only files from the same project can be viewed in one workspace. Resolving taxonomic hierarchies across datasets in Metaviz is future work that could use a utility such as the metagenomeFeatures R/Bioconductor package ¹⁴ . Figure 1C shows the manifest file workflow for samples from the IBD dataset and resulting workspace in Metaviz ( Figure 1D).

For ease of use, we provide tutorials at https://epiviz.github.io/tutorials/metaviz/. As a community resource, we plan to update the Metaviz database within a month of Bioconductor releases of HMP2Data. We maintain links to the HMP Data Portal through the update of HMP2Data package URLs and provide default workspaces for the HMP2 datasets as well as those in HMP16SData R/Bioconductor package ( https://epiviz.github.io/metaviz-workspaces/). For generating data summaries, we recommend using the HMP2Data package with appropriate R libraries to summarize sample information and the interactive HMP Data Portal for data summaries over different samples or study attributes. We provide instructions in the metavizr vignette to handle visualizing data that would be added to an analysis session like protected variables from dbGAP.

In the IBD cohort of the iHMP dataset, investigators sequenced a subset of samples using whole metagenome and 16S sequencing. We developed functions in metavizr to compare 16S and whole metagenome data for individual samples. Using the taxonomic profiles of the IBD samples, we matched the taxonomic features discovered with both sequencing methods. With this subset of features, we generated a single taxonomic hierarchy then loaded the 16S and whole metagenome abundance measurements into a metavizr object. Figure 2 shows an example analysis with stacked plots and scatter plots that link to a single FacetZoom to compare the degree of consistency of the data across sequencing methods.

The IBD study consisted of two phases: a pilot, which we refer to in this work as the IBD Stool Pilot, and a larger phase that we call IBD iHMP. We use the taxonomic profiles for each phase available from the HMP2Data package and use the same taxonomic classification identifiers in the package. To upload project data on to the UMD Metagenome Browser, we extracted 16S count table and taxonomic annotation using the otu_table() and tax_table() methods of HMP2Data package. We then use metagenomeSeq and metavizr to import the count data along with taxonomy and sample metadata into a neo4j graph database ¹⁵ using the metavizr neo4j import functionality. We used Metaviz ⁶ for exploratory analysis and metagenomeSeq for confirmatory statistical testing. We examined the IBD Stool Pilot and IBD iHMP dataset separately.

The IBD Stool Pilot dataset contains 16S and whole metagenome sequencing results of stool samples from 41 Crohn’s disease (CD) subjects and 10 ulcerative colitis (UC) subjects. We focused our analysis on 16S sequencing and used Metaviz to visually identify taxa that showed a difference in abundance between CD and UC subjects. Figure 3 shows a typical visualization.

We also used metagenomeSeq to test the differential abundance of features aggregated to each level of the taxonomy using the fitFeatureModel method that is based on a zero-inflated log-normal linear model. As shown in Table 1, two species had an absolute log fold-change greater than 1 and adjusted (Benjamini-Hochberg) p-value less than 0.1. Visually inspecting the IBD Stool Pilot data by aggregating counts to each level of the taxonomy we found the following features appearing differentially abundant: “c__Betaproteobacteria”, “o__Burkholderiales”, “f__Ruminococcaceae”, “g__Lachnospira”, “g__[Ruminococcus]”, “g__Faecalibacterium”, “s__:589277”, “s__:333166”, “s__:564806”, “s__:369227”, “s__:358104”, “s__:369486”, “s__gnavus:360015”, “s__prausnitzii:851865”. These taxonomic features describe paths in the taxonomy of the Kingdom Bacteria that was derived from the SILVA ¹⁷ database. The documentation ( https://ibdmdb.org/tunnel/public/HMP2/16S/1806/products) for the abundance profiles used in this analysis denotes that this taxonomic string was generated with the sequence of an OTU derived with the UPARSE algorithm ¹⁸ that was mapped to the SILVA database. Among these features, "c__Betaproteobacteria" refers to Class Betaproteobacteria, "o__Burkholderiales" to Order Burkholderiales, while values towards the leaves of the taxonomy refer to entries in the SILVA database that have an identifier and a sequence but have not been provided formal names in the binomial nomenclature system. Comparing the visual analysis results and the metagenomeSeq differential abundance testing results in Table 1 shows that the taxonomic feature s__:369227 (member of the Lachnospiraceae family which are strictly anaerobic ¹⁹ ) was identified using both methods. Members of Lachnospiraceae are abundant in human intestinal tracts and have been linked specifically to production of butyric acid ¹⁹ . Also, colonization with a specific strain of Lachnospiraceae in obese mice has been linked to development of hyperglycemia ²⁰ . The second taxon, s__:363232, is a member of the genus Dorea which has recently been shown to be associated with diarrhea predominant irritable bowel syndrome ²¹ .

The IBD iHMP dataset consists of samples from subjects with CD, UC, and those without IBD (nonIBD). For these samples, we analyzed the 16S sequencing data of an ileum biopsy from the first visit for each subject, which yielded 72 samples with 32 from CD, 18 from CD, and 22 from nonIBD. We used metagenomeSeq to compute an F-statistic to determine if any taxonomic feature is associated with at least one group using the fitZig method (based on a zero-inflated Normal linear model on log-transformed counts appropriate for multi-category experiment designs). Figure 4 shows an example using Metaviz to visualize abundance profiles for phylum Fusobacteria, which was found to be differentially abundant across the three groups. Differential abundance of members of this phylum has previously been reported in studies of IBD ²² . Analysis code and results are available as Extended data ²³ .

The 16S abundance matrices for IBD, T2D and the MOMS-PI studies were downloaded from the HMP2Data Bioconductor package. These datasets are then loaded into the neo4j graph database using import methods available in the metavizr ²⁴ Bioconductor package. These import scripts are available at https://gist.github.com/jkanche/c57d8220a33b41e21c4f6769a7aef7e4.

Figshare: Differential Abundance Analysis - IBD ( Figure 4). https://doi.org/10.6084/m9.figshare.12404222.v2 ²³ .

This file contains differential abundance analysis code and results.

Extended data are available under the terms of the Creative Commons Attribution 4.0 International license (CC-BY 4.0).