Workflow for Microbiome Data Analysis: from raw reads to community analyses.
25 July 2017
Abstract
High-throughput sequencing of PCR-amplified taxonomic markers (like the 16S rRNA gene) has enabled a new level of analysis of complex bacterial communities known as microbiomes. Many tools exist to quantify and compare abundance levels or OTU composition of communities in different conditions. The sequencing reads have to be denoised and assigned to the closest taxa from a reference database. Common approaches use a notion of 97% similarity and normalize the data by subsampling to equalize library sizes. In this paper, we show that statistical models allow more accurate abundance estimates. By providing a complete workflow in R, we enable the user to do sophisticated downstream statistical analyses, whether parametric or nonparametric. We provide examples of using the R packages dada2, phyloseq, DESeq2, ggplot2, structSSI and vegan to filter, visualize and test microbiome data. We also provide examples of supervised analyses using random forests and nonparametric testing using community networks and the ggnetwork package.
Introduction
The microbiome is formed of the ecological communities of microorganisms that dominate the living world. Bacteria can now be identified through the use of next generation sequencing applied at several levels. Shotgun sequencing of all bacteria in a sample delivers knowledge of all the genes present. Here we will only be interested in the identification and quantification of individual taxa (or species) through a ‘fingerprint gene’ called 16s rRNA which is present in all bacteria. This gene presents several variable regions which can be used to identify the different taxa.
Previous standard workflows depended on clustering all 16s rRNA sequences (generated by next generation amplicon sequencing) that occur within a 97% radius of similarity and then assigning these to ‘OTUs’ from reference trees (Caporaso et al. 2010; Schloss et al. 2009). These approaches do not incorporate all the data, in particular sequence quality information and statistical information available on the reads were not incorporated into the assignments.
In contrast, the de novo read counts used here will be constructed through the incorporation of both the quality scores and sequence frequencies in a probabilistic noise model for nucleotide transitions. For more details on the algorithmic implementation of this step see (Benjamin J Callahan et al. 2016).
After filtering the sequences and removing the chimeræ, the data are compared to a standard database of bacteria and labeled. In this workflow, we have used the labeled sequences to build a de novo phylogenetic with the .
The key step in the sequence analysis is the manner in which reads are denoised and assembled into groups we have chosen to call ASVs (Amplicon Sequence Variants)(Callahan, McMurdie, and Holmes 2017) instead of the traditional OTUs(Operational Taxonomic Units).
A published (but essentially similar) version of this workflow, including reviewer reports and comments is available (Ben J Callahan et al. 2016), see F1000Research From Raw reads.
There are extensive documentation and tutorial pages available for dada2 and phyloseq. If you have questions about this workflow, please start by consulting the relevant github issues sites for dada2, phyloseq, if the answers are not available, please post to the issues pages or Bioconductor forum. Note the posting guide for crafting an optimal question for the support site.
This is a workflow for denoising, filtering, performing data transformations, visualization, supervised learning analyses, community network tests, hierarchical testing and linear models. We provide all the code and give several examples of different types of analyses and use-cases. There are often many different objectives in experiments involving microbiome data and we will only give a flavor for what could be possible once the data has been imported into R.
In addition, the code can be easily adapted to accommodate batch effects, covariates and multiple experimental factors.
This workflow is based on software packages from the open-source Bioconductor project (Huber et al. 2015) and some CRAN packages.
We provide all steps necessary from the denoising and identification of the reads input as raw sequences as fastq files to the comparative testing and multivariate analyses of the samples and analyses of the abundances according to multiple available covariates.
Methods
Amplicon bioinformatics: from raw reads to tables
This section demonstrates the “full stack” of amplicon bioinformatics: construction of the sample-by-sequence feature table from the raw reads, assignment of taxonomy, and creation of a phylogenetic tree relating the sample sequences.
First we load the necessary packages.
library("knitr")
library("BiocStyle")
.cran_packages <- c("ggplot2", "gridExtra")
.bioc_packages <- c("dada2", "phyloseq", "DECIPHER", "phangorn")
.inst <- .cran_packages %in% installed.packages()
if(any(!.inst)) {
install.packages(.cran_packages[!.inst])
}
.inst <- .bioc_packages %in% installed.packages()
if(any(!.inst)) {
source("http://bioconductor.org/biocLite.R")
biocLite(.bioc_packages[!.inst], ask = F)
}
# Load packages into session, and print package version
sapply(c(.cran_packages, .bioc_packages), require, character.only = TRUE)## ggplot2 gridExtra dada2 phyloseq DECIPHER phangorn
## TRUE TRUE TRUE TRUE TRUE TRUEset.seed(100)The data we will analyze here are highly-overlapping Illumina Miseq 2x250 amplicon sequences from the V4 region of the 16S gene (Kozich et al. 2013). These 360 fecal samples were collected from 12 mice longitudinally over the first year of life one mock community control. These were collected to investigate the development and stabilization of the murine microbiome (Schloss et al. 2012). These data are downloaded from the following data location and unzip. For now just consider them paired-end fastq files to be processed. Define the following path variable so that it points to the extracted directory on your machine:
miseq_path <- "./MiSeq_SOP" # CHANGE to the directory containing the fastq files after unzipping.
list.files(miseq_path)## [1] "F3D0_S188_L001_R1_001.fastq" "F3D0_S188_L001_R2_001.fastq"
## [3] "F3D1_S189_L001_R1_001.fastq" "F3D1_S189_L001_R2_001.fastq"
## [5] "F3D141_S207_L001_R1_001.fastq" "F3D141_S207_L001_R2_001.fastq"
## [7] "F3D142_S208_L001_R1_001.fastq" "F3D142_S208_L001_R2_001.fastq"
## [9] "F3D143_S209_L001_R1_001.fastq" "F3D143_S209_L001_R2_001.fastq"
## [11] "F3D144_S210_L001_R1_001.fastq" "F3D144_S210_L001_R2_001.fastq"
## [13] "F3D145_S211_L001_R1_001.fastq" "F3D145_S211_L001_R2_001.fastq"
## [15] "F3D146_S212_L001_R1_001.fastq" "F3D146_S212_L001_R2_001.fastq"
## [17] "F3D147_S213_L001_R1_001.fastq" "F3D147_S213_L001_R2_001.fastq"
## [19] "F3D148_S214_L001_R1_001.fastq" "F3D148_S214_L001_R2_001.fastq"
## [21] "F3D149_S215_L001_R1_001.fastq" "F3D149_S215_L001_R2_001.fastq"
## [23] "F3D150_S216_L001_R1_001.fastq" "F3D150_S216_L001_R2_001.fastq"
## [25] "F3D2_S190_L001_R1_001.fastq" "F3D2_S190_L001_R2_001.fastq"
## [27] "F3D3_S191_L001_R1_001.fastq" "F3D3_S191_L001_R2_001.fastq"
## [29] "F3D5_S193_L001_R1_001.fastq" "F3D5_S193_L001_R2_001.fastq"
## [31] "F3D6_S194_L001_R1_001.fastq" "F3D6_S194_L001_R2_001.fastq"
## [33] "F3D7_S195_L001_R1_001.fastq" "F3D7_S195_L001_R2_001.fastq"
## [35] "F3D8_S196_L001_R1_001.fastq" "F3D8_S196_L001_R2_001.fastq"
## [37] "F3D9_S197_L001_R1_001.fastq" "F3D9_S197_L001_R2_001.fastq"
## [39] "filtered" "HMP_MOCK.v35.fasta"
## [41] "Mock_S280_L001_R1_001.fastq" "Mock_S280_L001_R2_001.fastq"
## [43] "mouse.dpw.metadata" "mouse.time.design"
## [45] "stability.batch" "stability.files"Filter and Trim
We begin by filtering out low-quality sequencing reads and trimming the reads to a consistent length. While generally recommended filtering and trimming parameters serve as a starting point, no two datasets are identical and therefore it is always worth inspecting the quality of the data before proceeding.
First we read in the names of the fastq files, and perform some string manipulation to get lists of the forward and reverse fastq files in matched order:
# Sort ensures forward/reverse reads are in same order
fnFs <- sort(list.files(miseq_path, pattern="_R1_001.fastq"))
fnRs <- sort(list.files(miseq_path, pattern="_R2_001.fastq"))
# Extract sample names, assuming filenames have format: SAMPLENAME_XXX.fastq
sampleNames <- sapply(strsplit(fnFs, "_"), `[`, 1)
# Specify the full path to the fnFs and fnRs
fnFs <- file.path(miseq_path, fnFs)
fnRs <- file.path(miseq_path, fnRs)
fnFs[1:3]## [1] "./MiSeq_SOP/F3D0_S188_L001_R1_001.fastq" "./MiSeq_SOP/F3D1_S189_L001_R1_001.fastq"
## [3] "./MiSeq_SOP/F3D141_S207_L001_R1_001.fastq"fnRs[1:3]## [1] "./MiSeq_SOP/F3D0_S188_L001_R2_001.fastq" "./MiSeq_SOP/F3D1_S189_L001_R2_001.fastq"
## [3] "./MiSeq_SOP/F3D141_S207_L001_R2_001.fastq"Most Illumina sequencing data shows a trend of decreasing average quality towards the end of sequencing reads.
The first two forward reads:
plotQualityProfile(fnFs[1:2])
The first two reverse reads:
plotQualityProfile(fnRs[1:2])
Here, the forward reads maintain high quality throughout, while the quality of the reverse reads drops significantly at about position 160. Therefore, we choose to truncate the forward reads at position 245, and the reverse reads at position 160. We also choose to trim the first 10 nucleotides of each read based on empirical observations across many Illumina datasets that these base positions are particularly likely to contain pathological errors.
We define the filenames for the filtered fastq.gz files:
filt_path <- file.path(miseq_path, "filtered") # Place filtered files in filtered/ subdirectory
if(!file_test("-d", filt_path)) dir.create(filt_path)
filtFs <- file.path(filt_path, paste0(sampleNames, "_F_filt.fastq.gz"))
filtRs <- file.path(filt_path, paste0(sampleNames, "_R_filt.fastq.gz"))We combine these trimming parameters with standard filtering parameters, the most important being the enforcement of a maximum of 2 expected errors per-read (Edgar and Flyvbjerg 2015). Trimming and filtering is performed on paired reads jointly, i.e. both reads must pass the filter for the pair to pass.
Filter the forward and reverse reads:
out <- filterAndTrim(fnFs, filtFs, fnRs, filtRs, truncLen=c(240,160),
maxN=0, maxEE=c(2,2), truncQ=2, rm.phix=TRUE,
compress=TRUE, multithread=TRUE) # On Windows set multithread=FALSE
head(out)## reads.in reads.out
## F3D0_S188_L001_R1_001.fastq 7793 7113
## F3D1_S189_L001_R1_001.fastq 5869 5299
## F3D141_S207_L001_R1_001.fastq 5958 5463
## F3D142_S208_L001_R1_001.fastq 3183 2914
## F3D143_S209_L001_R1_001.fastq 3178 2941
## F3D144_S210_L001_R1_001.fastq 4827 4312Infer sequence variants
After filtering, the typical amplicon bioinformatics workflow clusters sequencing reads into operational taxonomic units (OTUs): groups of sequencing reads that differ by less than a fixed dissimilarity threshhold. Here we instead use the high-resolution DADA2 method to to infer amplicon sequence variants (ASVs) exactly, without imposing any arbitrary threshhold, and thereby resolving variants that differ by as little as one nucleotide (Benjamin J Callahan et al. 2016).
The sequence data is imported into R from demultiplexed fastq files (i.e. one fastq for each sample) and simultaneously dereplicated to remove redundancy. We name the resulting derep-class objects by their sample name.
Dereplication
Dereplication combines all identical sequencing reads into into “unique sequences” with a corresponding “abundance”: the number of reads with that unique sequence. Dereplication substantially reduces computation time by eliminating redundant comparisons.
derepFs <- derepFastq(filtFs, verbose=TRUE)
derepRs <- derepFastq(filtRs, verbose=TRUE)
# Name the derep-class objects by the sample names
names(derepFs) <- sampleNames
names(derepRs) <- sampleNamesThe DADA2 method relies on a parameterized model of substitution errors to distinguish sequencing errors from real biological variation. Because error rates can (and often do) vary substantially between sequencing runs and PCR protocols, the model parameters can be discovered from the data itself using a form of unsupervised learning in which sample inference is alternated with parameter estimation until both are jointly consistent.
Parameter learning is computationally intensive, as it requires multiple iterations of the sequence inference algorithm, and therefore it is often useful to estimate the error rates from a (sufficiently large) subset of the data.
errF <- learnErrors(filtFs, multithread=TRUE)## Initializing error rates to maximum possible estimate.
## Sample 1 - 7113 reads in 1979 unique sequences.
## Sample 2 - 5299 reads in 1639 unique sequences.
## Sample 3 - 5463 reads in 1477 unique sequences.
## Sample 4 - 2914 reads in 904 unique sequences.
## Sample 5 - 2941 reads in 939 unique sequences.
## Sample 6 - 4312 reads in 1267 unique sequences.
## Sample 7 - 6741 reads in 1756 unique sequences.
## Sample 8 - 4560 reads in 1438 unique sequences.
## Sample 9 - 15637 reads in 3590 unique sequences.
## Sample 10 - 11413 reads in 2762 unique sequences.
## Sample 11 - 12017 reads in 3021 unique sequences.
## Sample 12 - 5032 reads in 1566 unique sequences.
## Sample 13 - 18075 reads in 3707 unique sequences.
## Sample 14 - 6250 reads in 1479 unique sequences.
## Sample 15 - 4052 reads in 1195 unique sequences.
## Sample 16 - 7369 reads in 1832 unique sequences.
## Sample 17 - 4765 reads in 1183 unique sequences.
## Sample 18 - 4871 reads in 1382 unique sequences.
## Sample 19 - 6504 reads in 1709 unique sequences.
## Sample 20 - 4314 reads in 897 unique sequences.
## selfConsist step 2
## selfConsist step 3
## selfConsist step 4
## selfConsist step 5
##
##
## Convergence after 5 rounds.
## Total reads used: 139642errR <- learnErrors(filtRs, multithread=TRUE)## Initializing error rates to maximum possible estimate.
## Sample 1 - 7113 reads in 1660 unique sequences.
## Sample 2 - 5299 reads in 1349 unique sequences.
## Sample 3 - 5463 reads in 1335 unique sequences.
## Sample 4 - 2914 reads in 853 unique sequences.
## Sample 5 - 2941 reads in 880 unique sequences.
## Sample 6 - 4312 reads in 1286 unique sequences.
## Sample 7 - 6741 reads in 1803 unique sequences.
## Sample 8 - 4560 reads in 1265 unique sequences.
## Sample 9 - 15637 reads in 3414 unique sequences.
## Sample 10 - 11413 reads in 2522 unique sequences.
## Sample 11 - 12017 reads in 2771 unique sequences.
## Sample 12 - 5032 reads in 1415 unique sequences.
## Sample 13 - 18075 reads in 3290 unique sequences.
## Sample 14 - 6250 reads in 1390 unique sequences.
## Sample 15 - 4052 reads in 1134 unique sequences.
## Sample 16 - 7369 reads in 1635 unique sequences.
## Sample 17 - 4765 reads in 1084 unique sequences.
## Sample 18 - 4871 reads in 1161 unique sequences.
## Sample 19 - 6504 reads in 1502 unique sequences.
## Sample 20 - 4314 reads in 732 unique sequences.
## selfConsist step 2
## selfConsist step 3
## selfConsist step 4
## selfConsist step 5
## selfConsist step 6
##
##
## Convergence after 6 rounds.
## Total reads used: 139642plotErrors(errF)
plotErrors(errR)
Figure 1: Estimated Error rates (both forward and reverse)
In order to verify that the error rates have been reasonably well-estimated, we inspect the fit between the observed error rates (black points) and the fitted error rates (black lines) in Figure 1. These figures show the frequencies of each type of transition as a function of the quality.
The DADA2 sequence inference method can run in two different modes: Independent inference by sample (pool=FALSE), and inference from the pooled sequencing reads from all samples (pool=TRUE). Independent inference has the advantage that computation time is linear in the number of samples, and memory requirements are flat with the number of samples. This allows scaling out to datasets of almost unlimited size. Pooled inference is more computationally taxing, and can become intractable for datasets of tens of millions of reads. However, pooling improves the detection of rare variants that were seen just once or twice in an individual sample but many times across all samples. As this dataset is not particularly large, we perform pooled inference. As of version 1.2, multithreading can now be activated with the arguments multithread = TRUE, which substantially speeds this step.
dadaFs <- dada(derepFs, err=errF, multithread=TRUE)## Sample 1 - 7113 reads in 1979 unique sequences.
## Sample 2 - 5299 reads in 1639 unique sequences.
## Sample 3 - 5463 reads in 1477 unique sequences.
## Sample 4 - 2914 reads in 904 unique sequences.
## Sample 5 - 2941 reads in 939 unique sequences.
## Sample 6 - 4312 reads in 1267 unique sequences.
## Sample 7 - 6741 reads in 1756 unique sequences.
## Sample 8 - 4560 reads in 1438 unique sequences.
## Sample 9 - 15637 reads in 3590 unique sequences.
## Sample 10 - 11413 reads in 2762 unique sequences.
## Sample 11 - 12017 reads in 3021 unique sequences.
## Sample 12 - 5032 reads in 1566 unique sequences.
## Sample 13 - 18075 reads in 3707 unique sequences.
## Sample 14 - 6250 reads in 1479 unique sequences.
## Sample 15 - 4052 reads in 1195 unique sequences.
## Sample 16 - 7369 reads in 1832 unique sequences.
## Sample 17 - 4765 reads in 1183 unique sequences.
## Sample 18 - 4871 reads in 1382 unique sequences.
## Sample 19 - 6504 reads in 1709 unique sequences.
## Sample 20 - 4314 reads in 897 unique sequences.dadaRs <- dada(derepRs, err=errR, multithread=TRUE)## Sample 1 - 7113 reads in 1660 unique sequences.
## Sample 2 - 5299 reads in 1349 unique sequences.
## Sample 3 - 5463 reads in 1335 unique sequences.
## Sample 4 - 2914 reads in 853 unique sequences.
## Sample 5 - 2941 reads in 880 unique sequences.
## Sample 6 - 4312 reads in 1286 unique sequences.
## Sample 7 - 6741 reads in 1803 unique sequences.
## Sample 8 - 4560 reads in 1265 unique sequences.
## Sample 9 - 15637 reads in 3414 unique sequences.
## Sample 10 - 11413 reads in 2522 unique sequences.
## Sample 11 - 12017 reads in 2771 unique sequences.
## Sample 12 - 5032 reads in 1415 unique sequences.
## Sample 13 - 18075 reads in 3290 unique sequences.
## Sample 14 - 6250 reads in 1390 unique sequences.
## Sample 15 - 4052 reads in 1134 unique sequences.
## Sample 16 - 7369 reads in 1635 unique sequences.
## Sample 17 - 4765 reads in 1084 unique sequences.
## Sample 18 - 4871 reads in 1161 unique sequences.
## Sample 19 - 6504 reads in 1502 unique sequences.
## Sample 20 - 4314 reads in 732 unique sequences.Inspecting the dada-class object returned by dada:
dadaFs[[1]]## dada-class: object describing DADA2 denoising results
## 128 sample sequences were inferred from 1979 input unique sequences.
## Key parameters: OMEGA_A = 1e-40, BAND_SIZE = 16, USE_QUALS = TRUEThe DADA2 algorithm inferred 128 real sequence variants from the 1979 unique sequences in the first sample. The dada-class object contains multiple diagnostics about the quality of each inferred sequence variant(see help("dada-class") for some info).
The DADA2 sequence inference step removed (nearly) all substitution and indel errors from the data (Benjamin J Callahan et al. 2016). We now merge together the inferred forward and reverse sequences, removing paired sequences that do not perfectly overlap as a final control against residual errors.
Construct sequence table and remove chimeras
The DADA2 method produces a sequence table that is a higher-resolution analogue of the common “OTU table”, i.e. a sample by sequence feature table valued by the number of times each sequence was observed in each sample.
mergers <- mergePairs(dadaFs, derepFs, dadaRs, derepRs)seqtabAll <- makeSequenceTable(mergers[!grepl("Mock", names(mergers))])
table(nchar(getSequences(seqtabAll)))##
## 251 252 253 254 255
## 1 84 182 5 2Notably, chimeras have not yet been removed. The error model in the sequence inference algorithm does not include a chimera component, and therefore we expect this sequence table to include many chimeric sequences. We now remove chimeric sequences by comparing each inferred sequence to the others in the table, and removing those that can be reproduced by stitching together two more abundant sequences.
seqtabNoC <- removeBimeraDenovo(seqtabAll)Although exact numbers vary substantially by experimental condition, it is typical that chimeras comprise a substantial fraction of inferred sequence variants, but only a small fraction of all reads. That is what is observed here chimeras make up about 22% of the inferred sequence variants, but those variants account for only about 4% of the total sequence reads.
Assign taxonomy
One of the benefits of using well-classified marker loci like the 16S rRNA gene is the ability to taxonomically classify the sequence variants. The dada2 package implements the naive Bayesian classifier method for this purpose (Wang et al. 2007). This classifier compares sequence variants to a training set of classified sequences, and here we use the RDP v16 training set (Cole et al. 2009).
The dada2 tutorial website contains formatted training fastas for the RDP training set, GreenGenes clustered at 97% identity, and the Silva reference database available. For fungal taxonomy, the General Fasta release files from the UNITE ITS database can be used as is. To follow this workflow, download the rdp_train_set_16.fa.gz file, and place it in the directory with the fastq files.
fastaRef <- "./rdp_train_set_16.fa.gz"
taxTab <- assignTaxonomy(seqtabNoC, refFasta = fastaRef, multithread=TRUE)
unname(head(taxTab))## [,1] [,2] [,3] [,4] [,5] [,6]
## [1,] "Bacteria" "Bacteroidetes" "Bacteroidia" "Bacteroidales" "Porphyromonadaceae" NA
## [2,] "Bacteria" "Bacteroidetes" "Bacteroidia" "Bacteroidales" "Porphyromonadaceae" NA
## [3,] "Bacteria" "Bacteroidetes" "Bacteroidia" "Bacteroidales" "Porphyromonadaceae" NA
## [4,] "Bacteria" "Bacteroidetes" "Bacteroidia" "Bacteroidales" "Porphyromonadaceae" "Barnesiella"
## [5,] "Bacteria" "Bacteroidetes" "Bacteroidia" "Bacteroidales" "Bacteroidaceae" "Bacteroides"
## [6,] "Bacteria" "Bacteroidetes" "Bacteroidia" "Bacteroidales" "Porphyromonadaceae" NAConstruct phylogenetic tree
Phylogenetic relatedness is commonly used to inform downstream analyses, especially the calculation of phylogeny-aware distances between microbial communities. The DADA2 sequence inference method is reference-free, so we must construct the phylogenetic tree relating the inferred sequence variants de novo. We begin by performing a multiple-alignment using the DECIPHER R package (Wright 2015).
seqs <- getSequences(seqtabNoC)
names(seqs) <- seqs # This propagates to the tip labels of the tree
alignment <- AlignSeqs(DNAStringSet(seqs), anchor=NA,verbose=FALSE)The phangorn R package is then used to construct a phylogenetic tree. Here we first construct a neighbor-joining tree, and then fit a GTR+G+I (Generalized time-reversible with Gamma rate variation) maximum likelihood tree using the neighbor-joining tree as a starting point.
phangAlign <- phyDat(as(alignment, "matrix"), type="DNA")
dm <- dist.ml(phangAlign)
treeNJ <- NJ(dm) # Note, tip order != sequence order
fit = pml(treeNJ, data=phangAlign)
fitGTR <- update(fit, k=4, inv=0.2)
fitGTR <- optim.pml(fitGTR, model="GTR", optInv=TRUE, optGamma=TRUE,
rearrangement = "stochastic", control = pml.control(trace = 0))
detach("package:phangorn", unload=TRUE)Combine data into a phyloseq object
The package phyloseq organizes and synthesizes the different data types from a typical amplicon sequencing experiment into a single data object that can be easily manipulated. The last bit of information needed is the sample data contained in a .csv file. This can be downloaded from github:
samdf <- read.csv("https://raw.githubusercontent.com/spholmes/F1000_workflow/master/data/MIMARKS_Data_combined.csv",header=TRUE)
samdf$SampleID <- paste0(gsub("00", "", samdf$host_subject_id), "D", samdf$age-21)
samdf <- samdf[!duplicated(samdf$SampleID),] # Remove dupicate entries for reverse reads
rownames(seqtabAll) <- gsub("124", "125", rownames(seqtabAll)) # Fix discrepancy
all(rownames(seqtabAll) %in% samdf$SampleID) # TRUE## [1] TRUErownames(samdf) <- samdf$SampleID
keep.cols <- c("collection_date", "biome", "target_gene", "target_subfragment",
"host_common_name", "host_subject_id", "age", "sex", "body_product", "tot_mass",
"diet", "family_relationship", "genotype", "SampleID")
samdf <- samdf[rownames(seqtabAll), keep.cols]The full suite of data for this study – the sample-by-sequence feature table, the sample metadata, the sequence taxonomies, and the phylogenetic tree – can now be combined into a single object.
ps <- phyloseq(otu_table(seqtabNoC, taxa_are_rows=FALSE),
sample_data(samdf),
tax_table(taxTab),phy_tree(fitGTR$tree))
ps <- prune_samples(sample_names(ps) != "Mock", ps) # Remove mock sample
ps## phyloseq-class experiment-level object
## otu_table() OTU Table: [ 215 taxa and 19 samples ]
## sample_data() Sample Data: [ 19 samples by 14 sample variables ]
## tax_table() Taxonomy Table: [ 215 taxa by 6 taxonomic ranks ]
## phy_tree() Phylogenetic Tree: [ 215 tips and 213 internal nodes ]Using phyloseq
phyloseq(McMurdie and Holmes 2013) is an R package to import, store, analyze, and graphically display complex phylogenetic sequencing data that has already been clustered into Operational Taxonomic Units (OTUs) or more appropriately denoised, and it is most useful when there is also associated sample data, phylogeny, and/or taxonomic assignment of each taxa. leverages and builds upon many of the tools available in R for ecology and phylogenetic analysis (ape, vegan, ade4), while also using advanced/flexible graphic systems (ggplot2) to easily produce publication-quality graphics of complex phylogenetic data. The phyloseq package uses a specialized system of S4 data classes to store all related phylogenetic sequencing data as a single, self-consistent, self-describing experiment-level object, making it easier to share data and reproduce analyses. In general, phyloseq seeks to facilitate the use of R for efficient interactive and reproducible analysis of amplicon count data jointly with important sample covariates.
This tutorial shows a useful example workflow, but many more analyses are available to you in phyloseq, and R in general, than can fit in a single workflow. The phyloseq home page is a good place to begin browsing additional phyloseq documentation, as are the three vignettes included within the package, and linked directly at the phyloseq release page on Bioconductor.
Loading the data
Many use cases result in the need to import and combine different data into a phyloseq class object, this can be done using th import_biom function to read recent QIIME format files, older files can still be imported with import_qiime. More complete details can be found on the phyloseq FAQ page.
In the previous section the results of dada2 sequence processing were organized into a phyloseq object. We have actually run dada2 on a larger set of samples from the same data source. This object was also saved in R-native serialized RDS format. We will re-load this here for completeness as the initial object ps. If you have not downloaded the whole repository you can access the ps file though github:
ps_connect <-url("https://raw.githubusercontent.com/spholmes/F1000_workflow/master/data/ps.rds")
ps = readRDS(ps_connect)
ps## phyloseq-class experiment-level object
## otu_table() OTU Table: [ 389 taxa and 360 samples ]
## sample_data() Sample Data: [ 360 samples by 14 sample variables ]
## tax_table() Taxonomy Table: [ 389 taxa by 6 taxonomic ranks ]
## phy_tree() Phylogenetic Tree: [ 389 tips and 387 internal nodes ]Shiny-phyloseq
It can be beneficial to start the data exploration process interactively, this often saves time in detecting outliers and specific features of the data. Shiny-phyloseq (McMurdie and Holmes 2015) is an interactive web application that provides a graphical user interface to the phyloseq package. The object just loaded into the R session in this workflow is suitable for graphical exploration with Shiny-phyloseq.
Filtering
phyloseq provides useful tools for filtering, subsetting, and agglomerating taxa – a task that is often appropriate or even necessary for effective analysis of microbiome count data. In this subsection, we graphically explore the prevalence of taxa in the example dataset, and demonstrate how this can be used as a filtering criteria. One of the reasons to filter in this way is to avoid spending much time analyzing taxa that were seen only rarely among samples. This also turns out to be a useful filter of noise (taxa that are actually just artifacts of the data collection process), a step that should probably be considered essential for datasets constructed via heuristic OTU-clustering methods, which are notoriously prone to generating spurious taxa.
Taxonomic Filtering
In many biological settings, the set of all organisms from all samples are well-represented in the available taxonomic reference database. When (and only when) this is the case, it is reasonable or even advisable to filter taxonomic features for which a high-rank taxonomy could not be assigned. Such ambiguous features in this setting are almost always sequence artifacts that don’t exist in nature. It should be obvious that such a filter is not appropriate for samples from poorly characterized or novel specimens, at least until the possibility of taxonomic novelty can be satisfactorily rejected. Phylum is a useful taxonomic rank to consider using for this purpose, but others may work effectively for your data.
To begin, create a table of read counts for each Phylum present in the dataset.
# Show available ranks in the dataset
rank_names(ps)## [1] "Kingdom" "Phylum" "Class" "Order" "Family" "Genus"# Create table, number of features for each phyla
table(tax_table(ps)[, "Phylum"], exclude = NULL)##
## Actinobacteria Bacteroidetes Candidatus_Saccharibacteria
## 13 23 1
## Cyanobacteria/Chloroplast Deinococcus-Thermus Firmicutes
## 4 1 327
## Fusobacteria Proteobacteria Tenericutes
## 1 11 1
## Verrucomicrobia <NA>
## 1 6This shows a few phyla for which only one feature was observed. Those may be worth filtering, and we’ll check that next. First, notice that in this case, six features were annotated with a Phylum of NA. These features are probably artifacts in a dataset like this, and should be removed.
The following ensures that features with ambiguous phylum annotation are also removed. Note the flexibility in defining strings that should be considered ambiguous annotation.
ps <- subset_taxa(ps, !is.na(Phylum) & !Phylum %in% c("", "uncharacterized"))A useful next step is to explore feature prevalence in the dataset, which we will define here as the number of samples in which a taxon appears at least once.
# Compute prevalence of each feature, store as data.frame
prevdf = apply(X = otu_table(ps),
MARGIN = ifelse(taxa_are_rows(ps), yes = 1, no = 2),
FUN = function(x){sum(x > 0)})
# Add taxonomy and total read counts to this data.frame
prevdf = data.frame(Prevalence = prevdf,
TotalAbundance = taxa_sums(ps),
tax_table(ps))Are there phyla that are comprised of mostly low-prevalence features? Compute the total and average prevalences of the features in each phylum.
plyr::ddply(prevdf, "Phylum", function(df1){cbind(mean(df1$Prevalence),sum(df1$Prevalence))})## Phylum 1 2
## 1 Actinobacteria 120.15385 1562
## 2 Bacteroidetes 265.52174 6107
## 3 Candidatus_Saccharibacteria 280.00000 280
## 4 Cyanobacteria/Chloroplast 64.25000 257
## 5 Deinococcus-Thermus 52.00000 52
## 6 Firmicutes 179.24771 58614
## 7 Fusobacteria 2.00000 2
## 8 Proteobacteria 59.09091 650
## 9 Tenericutes 234.00000 234
## 10 Verrucomicrobia 104.00000 104Deinococcus-Thermus appeared in just over one percent of samples, and Fusobacteria appeared in just 2 samples total. In some cases it might be worthwhile to explore these two phyla in more detail despite this (though probably not Fusobacteria’s two samples). For the purposes of this example, though, they will be filtered from the dataset.
# Define phyla to filter
filterPhyla = c("Fusobacteria", "Deinococcus-Thermus")
# Filter entries with unidentified Phylum.
ps1 = subset_taxa(ps, !Phylum %in% filterPhyla)
ps1## phyloseq-class experiment-level object
## otu_table() OTU Table: [ 381 taxa and 360 samples ]
## sample_data() Sample Data: [ 360 samples by 14 sample variables ]
## tax_table() Taxonomy Table: [ 381 taxa by 6 taxonomic ranks ]
## phy_tree() Phylogenetic Tree: [ 381 tips and 379 internal nodes ]Prevalence Filtering
The previous filtering steps are considered supervised, because they relied on prior information that is external to this experiment (a taxonomic reference database). This next filtering step is completely unsupervised, relying only on the data in this experiment, and a parameter that we will choose after exploring the data. Thus, this filtering step can be applied even in settings where taxonomic annotation is unavailable or unreliable.
First, explore the relationship of prevalence and total read count for each feature. Sometimes this reveals outliers that should probably be removed, and also provides insight into the ranges of either feature that might be useful. This aspect depends quite a lot on the experimental design and goals of the downstream inference, so keep these in mind. It may even be the case that different types of downstream inference require different choices here. There is no reason to expect ahead of time that a single filtering workflow is appropriate for all analysis.
# Subset to the remaining phyla
prevdf1 = subset(prevdf, Phylum %in% get_taxa_unique(ps1, "Phylum"))
ggplot(prevdf1, aes(TotalAbundance, Prevalence / nsamples(ps),color=Phylum)) +
# Include a guess for parameter
geom_hline(yintercept = 0.05, alpha = 0.5, linetype = 2) + geom_point(size = 2, alpha = 0.7) +
scale_x_log10() + xlab("Total Abundance") + ylab("Prevalence [Frac. Samples]") +
facet_wrap(~Phylum) + theme(legend.position="none")
Figure 2: Taxa prevalence versus total counts
Each point in Figure 2 is a different taxa. Exploration of the data in this way is often useful for selecting filtering parameters, like the minimum prevalence criteria we will used to filter the data above.
Sometimes a natural separation in the dataset reveals itself, or at least, a conservative choice that is in a stable region for which small changes to the choice would have minor or no effect on the biological interpreation (stability). Here no natural separation is immediately evident, but it looks like we might reasonably define a prevalence threshold in a range of zero to ten percent or so. Take care that this choice does not introduce bias into a downstream analysis of association of differential abundance.
The following uses five percent of all samples as the prevalence threshold.
# Define prevalence threshold as 5% of total samples
prevalenceThreshold = 0.05 * nsamples(ps)
prevalenceThreshold## [1] 18# Execute prevalence filter, using `prune_taxa()` function
keepTaxa = rownames(prevdf1)[(prevdf1$Prevalence >= prevalenceThreshold)]
ps2 = prune_taxa(keepTaxa, ps)Agglomerate taxa
When there is known to be a lot of species or sub-species functional redundancy in a microbial community, it might be useful to agglomerate the data features corresponding to closely related taxa. Ideally we would know the functional redundancies perfectly ahead of time, in which case we would agglomerate taxa using those defined relationships and the function in phyloseq. That kind of exquisite functional data is usually not available, and different pairs of microbes will have different sets of overlapping functions, complicating the matter of defining appropriate grouping criteria.
While not necessarily the most useful or functionally-accurate criteria for grouping microbial features (sometimes far from accurate), taxonomic agglomeration has the advantage of being much easier to define ahead of time. This is because taxonomies are usually defined with a comparatively simple tree-like graph structure that has a fixed number of internal nodes, called “ranks”. This structure is simple enough for the phyloseq package to represent taxonomies as table of taxonomy labels. Taxonomic agglomeration groups all the “leaves” in the hierarchy that descend from the user-prescribed agglomerating rank, this is sometimes called ‘glomming’.
The following example code shows how one would combine all features that descend from the same genus.
# How many genera would be present after filtering?
length(get_taxa_unique(ps2, taxonomic.rank = "Genus"))## [1] 49ps3 = tax_glom(ps2, "Genus", NArm = TRUE)If taxonomy is not available or not reliable, tree-based agglomeration is a “taxonomy-free” alternative to combine data features corresponding to closely-related taxa. In this case, rather than taxonomic rank, the user specifies a tree height corresponding to the phylogenetic distance between features that should define their grouping. This is very similar to “OTU Clustering”, except that in many OTU Clustering algorithms the sequence distance being used does not have the same (or any) evolutionary definition.
h1 = 0.4
ps4 = tip_glom(ps2, h = h1)Here phyloseq’s plot_tree() function compare the original unfiltered data, the tree after taxonoic agglomeration, and the tree after phylogenetic agglomeration. These are stored as separate plot objects, then rendered together in one combined graphic using gridExtra::grid.arrange.
multiPlotTitleTextSize = 15
p2tree = plot_tree(ps2, method = "treeonly",
ladderize = "left",
title = "Before Agglomeration") +
theme(plot.title = element_text(size = multiPlotTitleTextSize))
p3tree = plot_tree(ps3, method = "treeonly",
ladderize = "left", title = "By Genus") +
theme(plot.title = element_text(size = multiPlotTitleTextSize))
p4tree = plot_tree(ps4, method = "treeonly",
ladderize = "left", title = "By Height") +
theme(plot.title = element_text(size = multiPlotTitleTextSize))# group plots together
grid.arrange(nrow = 1, p2tree, p3tree, p4tree)
Figure 3: Different types of agglomeration
Figure 3 shows the original tree on the left, taxonomic agglomeration at Genus rank in the middle and phylogenetic agglomeration at a fixed distance of 0.4 on the right.
Abundance value transformation
It is usually necessary to transform microbiome count data to account for differences in library size, variance, scale, etc. The phyloseq package provides a flexible interface for defining new functions to accomplish these transformations of the abundance values via the function transform_sample_counts(). The first argument to this function is the phyloseq object you want to transform, and the second argument is an R function that defines the transformation. The R function is applied sample-wise, expecting that the first unnamed argument is a vector of taxa counts in the same order as the phyloseq object. Additional arguments are passed on to the function specified in the second argument, providing an explicit means to include pre-computed values, previously defined parameters/thresholds, or any other object that might be appropriate for computing the transformed values of interest.
This example begins by defining a custom plot function, plot_abundance(), that uses phyloseq’s function to define a relative abundance graphic. We will use this to compare more easily differences in scale and distribution of the abundance values in our phyloseq object before and after transformation.
plot_abundance = function(physeq,title = "",
Facet = "Order", Color = "Phylum"){
# Arbitrary subset, based on Phylum, for plotting
p1f = subset_taxa(physeq, Phylum %in% c("Firmicutes"))
mphyseq = psmelt(p1f)
mphyseq <- subset(mphyseq, Abundance > 0)
ggplot(data = mphyseq, mapping = aes_string(x = "sex",y = "Abundance",
color = Color, fill = Color)) +
geom_violin(fill = NA) +
geom_point(size = 1, alpha = 0.3,
position = position_jitter(width = 0.3)) +
facet_wrap(facets = Facet) + scale_y_log10()+
theme(legend.position="none")
}The transformation in this case converts the counts from each sample into their frequencies, often referred to as proportions or relative abundances. This function is so simple that it is easiest to define it within the function call to transform_sample_counts().
# Transform to relative abundance. Save as new object.
ps3ra = transform_sample_counts(ps3, function(x){x / sum(x)})Now we plot the abundance values before and after transformation.
plotBefore = plot_abundance(ps3,"")
plotAfter = plot_abundance(ps3ra,"")
# Combine each plot into one graphic.
grid.arrange(nrow = 2, plotBefore, plotAfter)
Figure 4: Comparison of original abundances with transformed data
Figure 4 shows the comparison of original abundances (top panel) and relative abundances (lower).
Subset by taxonomy
Notice on the previous plot that Lactobacillales appears to be a taxonomic Order with bimodal abundance profile in the data. We can check for a taxonomic explanation of this pattern by plotting just that taxonomic subset of the data. For this, we subset with the function, and then specify a more precise taxonomic rank to the argument of the function that we defined above.
psOrd = subset_taxa(ps3ra, Order == "Lactobacillales")
plot_abundance(psOrd, Facet = "Genus", Color = NULL)
Figure 5: Violin plot of relative abundances of Lactobacillales
Figure 5 shows the relative abundances of Lactobacillales taxonomic Order, grouped by host sex and genera. Here it is clear that the apparent biomodal distribution of Lactobacillales on the previous plot was the result of a mixture of two different genera, with the typical Lactobacillus relative abundance much larger than Streptococcus.
At this stage in the workflow, after converting raw reads to interpretable species abundances, and after filtering and transforming these abundances to focus attention on scientifically meaningful quantities, we are in a position to consider more careful statistical analysis. R is an ideal environment for performing these analyses, as it has an active community of package developers building simple interfaces to sophisticated techniques. As a variety of methods are available, there is no need to commit to any rigid analysis strategy a priori. Further, the ability to easily call packages without reimplementing methods frees researchers to iterate rapidly through alternative analysis ideas. The advantage of performing this full workflow in R is that this transition from bioinformatics to statistics is effortless.
Let’s start by installing a few packages that are available for these complementary analyses:
.cran_packages <- c( "shiny","miniUI", "caret", "pls", "e1071", "ggplot2", "randomForest", "dplyr", "ggrepel", "nlme", "devtools",
"reshape2", "PMA", "structSSI", "ade4",
"ggnetwork", "intergraph", "scales")
.github_packages <- c("jfukuyama/phyloseqGraphTest")
.bioc_packages <- c("genefilter", "impute")
# Install CRAN packages (if not already installed)
.inst <- .cran_packages %in% installed.packages()
if (any(!.inst)){
install.packages(.cran_packages[!.inst],repos = "http://cran.rstudio.com/")
}
.inst <- .github_packages %in% installed.packages()
if (any(!.inst)){
devtools::install_github(.github_packages[!.inst])
}
.inst <- .bioc_packages %in% installed.packages()
if(any(!.inst)){
source("http://bioconductor.org/biocLite.R")
biocLite(.bioc_packages[!.inst])
}We back these claims by illustrating several analyses on the mouse data prepared above. We experiment with several flavors of exploratory ordination before shifting to more formal testing and modeling, explaining the settings in which the different points of view are most appropriate. Finally, we provide example analyses of multitable data, using a study in which both metabolomic and microbial abundance measurements were collected on the same samples, to demonstrate that the general workflow presented here can be adapted to the multitable setting.
Preprocessing
Before doing the multivariate projections, we will add a few columns to our sample data, which can then be used to annotate plots. From Figure 6, we see that the ages of the mice come in a couple of groups, and so we make a categorical variable corresponding to young, middle-aged, and old mice. We also record the total number of counts seen in each sample and log-transform the data as an approximate variance stabilizing transformation.
qplot(sample_data(ps)$age, geom = "histogram",binwidth=20) + xlab("age")