Introduction to PhILR
2022-04-26
Contents
- 1 Introduction
- 2 Overview of PhILR Analysis
- 3 Loading and Preprocessing Dataset
- 4 Data preparation: TreeSE
- 5 Transform Data using PhILR
- 6 Ordination in PhILR Space
- 7 Visualization with TreeSE
- 8 Identify Balances that Distinguish Human/Non-Human
- 9 Name Balances
- 10 Visualize Results
- 11 Use Balances for Dimension Reduction
- 12 Package versions
- References
1 Introduction
PhILR is short for “Phylogenetic Isometric Log-Ratio Transform” (Silverman et al. 2017). This package provides functions for the analysis of compositional data (e.g., data representing proportions of different variables/parts). Specifically this package allows analysis of compositional data where the parts can be related through a phylogenetic tree (as is common in microbiota survey data) and makes available the Isometric Log Ratio transform built from the phylogenetic tree and utilizing a weighted reference measure (Egozcue and Pawlowsky-Glahn 2016).
2 Overview of PhILR Analysis
The goal of PhILR is to transform compositional data into an orthogonal unconstrained space (real space) with phylogenetic / evolutionary interpretation while preserving all information contained in the original composition. Unlike in the original compositional space, in the transformed real space, standard statistical tools may be applied. For a given set of samples consisting of measurements of taxa, we transform data into a new space of samples and orthonormal coordinates termed ‘balances’. Each balance is associated with a single internal node of a phylogenetic tree with the taxa as leaves. The balance represents the log-ratio of the geometric mean abundance of the two groups of taxa that descend from the given internal node. More details on this method can be found in Silverman et al. (2017) (Link).
The analysis uses abundance table and a phylogenetic tree. These can
be provided as separate data objects, or embedded in standard
R/Bioconductor data containers. The philr R package supports two
alternative data containers for microbiome data, TreeSE (Huang et al. 2021)
and phyloseq (McMurdie and Holmes 2013).
3 Loading and Preprocessing Dataset
We demonstrate PhILR analysis by using the Global Patterns dataset that was originally published by Caporaso et al. (2011).
Let us first load necessary libraries.
library(philr); packageVersion("philr")## [1] '1.22.0'library(ape); packageVersion("ape")## [1] '5.6.2'library(ggplot2); packageVersion("ggplot2")## [1] '3.3.5'4 Data preparation: TreeSE
We show the GlobalPatterns example workflow as initially outlined in (McMurdie and Holmes 2013).
We retrieve the example data in TreeSummarizedExperiment (TreeSE)
data format in this vignette (Huang et al. 2021), and then show example also
for the phyloseq format. The TreeSE version
for the GlobalPatterns data is provided with the mia
package (Lahti et al. 2020).
Let us load the data.
library(mia); packageVersion("mia")## [1] '1.4.0'library(dplyr); packageVersion("dplyr")## [1] '1.0.8'data(GlobalPatterns, package = "mia")4.1 Filter Extremely Low-Abundance OTUs
Taxa that were not seen with more than 3 counts in at least 20% of samples are filtered. Subsequently, those with a coefficient of variation ≤ 3 are filtered. Finally we add a pseudocount of 1 to the remaining OTUs to avoid calculating log-ratios involving zeros. Alternatively other replacement methods (multiplicative replacement etc…) may be used instead if desired; the subsequent taxa weighting procedure we will describe complements a variety of zero replacement methods.
## Select prevalent taxa
tse <- GlobalPatterns %>% subsetByPrevalentTaxa(
detection = 3,
prevalence = 20/100,
as_relative = FALSE)
## Pick taxa that have notable abundance variation across sammples
variable.taxa <- apply(assays(tse)$counts, 1, function(x) sd(x)/mean(x) > 3.0)
tse <- tse[variable.taxa,]
# Collapse the tree!
# Otherwise the original tree with all nodes is kept
# (including those that were filtered out from rowData)
tree <- ape::keep.tip(phy = rowTree(tse), tip = rowLinks(tse)$nodeNum)
rowTree(tse) <- tree
## Add a new assay with a pseudocount
assays(tse)$counts.shifted <- assays(tse)$counts + 1 We have now removed the filtered taxa from the OTU table, pruned the phylogenetic tree, and subset the taxa table. Here is the result of those filtering steps.
## class: TreeSummarizedExperiment
## dim: 1248 26
## metadata(0):
## assays(2): counts counts.shifted
## rownames(1248): 540305 108964 ... 516119 145149
## rowData names(7): Kingdom Phylum ... Genus Species
## colnames(26): CL3 CC1 ... Even2 Even3
## colData names(7): X.SampleID Primer ... SampleType Description
## reducedDimNames(0):
## mainExpName: NULL
## altExpNames(0):
## rowLinks: a LinkDataFrame (1248 rows)
## rowTree: 1 phylo tree(s) (1248 leaves)
## colLinks: NULL
## colTree: NULL4.2 Process Phylogenetic Tree
Next we check that the tree is rooted and binary (all multichotomies have been resolved).
library(ape); packageVersion("ape")## [1] '5.6.2'is.rooted(tree) # Is the tree Rooted?## [1] TRUEis.binary(tree) # All multichotomies resolved?## [1] TRUENote that if the tree is not binary, the function multi2di from the
ape package can be used to replace multichotomies with a series of
dichotomies with one (or several) branch(es) of zero length.
Once this is done, we name the internal nodes of the tree so they are
easier to work with. We prefix the node number with n and thus the
root is named n1.
tree <- makeNodeLabel(tree, method="number", prefix='n')
# Add the modified tree back to the (`TreeSE`) data object
rowTree(tse) <- treeWe note that the tree is already rooted with Archea as the outgroup
and no multichotomies are present. This uses the function
name.balance from the philr package. This function uses a simple
voting scheme to find a consensus naming for the two clades that
descend from a given balance. Specifically for a balance named x/y,
x refers to the consensus name of the clade in the numerator of the
log-ratio and y refers to the denominator.
# Extract taxonomy table from the TreeSE object
tax <- rowData(tse)[,taxonomyRanks(tse)]
# Get name balances
name.balance(tree, tax, 'n1')## [1] "Kingdom_Archaea/Kingdom_Bacteria"4.3 Investigate Dataset Components
Finally we transpose the OTU table (philr uses the conventions of
the compositions package for compositional data analysis in R, taxa
are columns, samples are rows). Then we will take a look at part of
the dataset in more detail.
otu.table <- t(as(assays(tse)$counts.shifted, "matrix"))
tree <- rowTree(tse)
metadata <- colData(tse)
tax <- rowData(tse)[,taxonomyRanks(tse)]
otu.table[1:2,1:2] # OTU Table## 540305 108964
## CL3 1 1
## CC1 1 2tree # Phylogenetic Tree##
## Phylogenetic tree with 1248 tips and 1247 internal nodes.
##
## Tip labels:
## 540305, 108964, 175045, 546313, 54107, 71074, ...
## Node labels:
## n1, n2, n3, n4, n5, n6, ...
##
## Rooted; includes branch lengths.head(metadata,2) # Metadata## DataFrame with 2 rows and 7 columns
## X.SampleID Primer Final_Barcode Barcode_truncated_plus_T
## <factor> <factor> <factor> <factor>
## CL3 CL3 ILBC_01 AACGCA TGCGTT
## CC1 CC1 ILBC_02 AACTCG CGAGTT
## Barcode_full_length SampleType Description
## <factor> <factor> <factor>
## CL3 CTAGCGTGCGT Soil Calhoun South Carolina Pine soil, pH 4.9
## CC1 CATCGACGAGT Soil Cedar Creek Minnesota, grassland, pH 6.1head(tax,2) # taxonomy table## DataFrame with 2 rows and 7 columns
## Kingdom Phylum Class Order Family
## <character> <character> <character> <character> <character>
## 540305 Archaea Crenarchaeota Thaumarchaeota Cenarchaeales Cenarchaeaceae
## 108964 Archaea Crenarchaeota Thaumarchaeota Cenarchaeales Cenarchaeaceae
## Genus Species
## <character> <character>
## 540305 NA NA
## 108964 Nitrosopumilus pIVWA5A new variable distinguishing human/non-human:
human.samples <- factor(colData(tse)$SampleType %in% c("Feces", "Mock", "Skin", "Tongue"))5 Transform Data using PhILR
The function philr::philr() implements a user friendly wrapper for the key
steps in the philr transform.
- Convert the phylogenetic tree to its sequential binary partition (SBP) representation
using the function
philr::phylo2sbp() - Calculate the weighting of the taxa (aka parts) or use the user specified weights
- Built the contrast matrix from the SBP and taxa weights using the function
philr::buildilrBasep() - Convert OTU table to relative abundance (using
philr::miniclo()) and ‘shift’ dataset using the weightings (Egozcue and Pawlowsky-Glahn 2016) using the functionphilr::shiftp(). - Transform the data to PhILR space using the function
philr::ilrp() - (Optional) Weight the resulting PhILR space using phylogenetic distance. These
weights are either provided by the user or can be calculated by the function
philr::calculate.blw().
Note: The preprocessed OTU table should be passed to the function
philr::philr() before it is closed (normalized) to relative
abundances, as some of the preset weightings of the taxa use the
original count data to down weight low abundance taxa.
Here we will use the same weightings as we used in the main paper.
You can run philr with the abundance table and phylogenetic tree.
gp.philr <- philr(otu.table, tree,
part.weights='enorm.x.gm.counts',
ilr.weights='blw.sqrt')
gp.philr[1:5,1:5]## n1 n2 n3 n4 n5
## CL3 -1.3638521 1.9756259 2.6111996 -3.3174292 0.08335109
## CC1 -0.9441168 3.9054807 2.9804522 -4.7771598 -0.05334306
## SV1 5.8436901 5.9067782 6.7315081 -8.8020849 0.08335109
## M31Fcsw -3.9010427 -0.1816618 -0.5432099 0.1705271 0.08335109
## M11Fcsw -5.4554073 0.5398249 -0.5647474 0.5551616 -0.02389182Alternatively, you can provide the data directly in TreeSE format.
gp.philr <- philr(tse, abund_values = "counts.shifted",
part.weights='enorm.x.gm.counts',
ilr.weights='blw.sqrt')Alternatively, you can provide the data in phyloseq format. For
simplicity, let us just convert the TreeSE object to phyloseq
object to give a brief example.
pseq <- makePhyloseqFromTreeSummarizedExperiment(tse, abund_values="counts.shifted")
gp.philr <- philr(pseq,
part.weights='enorm.x.gm.counts',
ilr.weights='blw.sqrt')After running philr the transformed data is represented in terms of
balances and since each balance is associated with a single internal
node of the tree, we denote the balances using the same names we
assigned to the internal nodes (e.g., n1).
6 Ordination in PhILR Space
Euclidean distance in PhILR space can be used for ordination analysis. Let us first calculate distances and then calculate standard MDS ordination.
# Distances between samples based on philr transformed data
gp.dist <- dist(gp.philr, method="euclidean")
# Calculate MDS for the distance matrix
d <- as.data.frame(cmdscale(gp.dist))
colnames(d) <- paste0("PC", 1:2)7 Visualization with TreeSE
Let us next visualize the ordination. This example employs standard
tools for ordination and visualization that can be used regardless of
the preferred data container. Note that the phyloseq and TreeSE frameworks
may provide access to additional ordination and visualization methods.
# Add some metadata for the visualization
d$SampleType <- factor(metadata$SampleType)
# Create a plot
ggplot(data = d,
aes(x=PC1, y=PC2, color=SampleType)) +
geom_point() +
labs(title = "Euclidean distances with phILR")
8 Identify Balances that Distinguish Human/Non-Human
More than just ordination analysis, PhILR provides an entire
coordinate system in which standard multivariate tools can be
used. Here we will make use of sparse logistic regression (from the
glmnet package) to identify a small number of balances that best
distinguish human from non-human samples.
Now we will fit a sparse logistic regression model (logistic regression with \(l_1\) penalty)
library(glmnet); packageVersion('glmnet')## [1] '4.1.4'glmmod <- glmnet(gp.philr, human.samples, alpha=1, family="binomial")We will use a hard-threshold for the \(l_1\) penalty of \(\lambda = 0.2526\) which we choose so that the resulting number of non-zero coefficients is \(\approx 5\) (for easy of visualization in this tutorial).
top.coords <- as.matrix(coefficients(glmmod, s=0.2526))
top.coords <- rownames(top.coords)[which(top.coords != 0)]
(top.coords <- top.coords[2:length(top.coords)]) # remove the intercept as a coordinate## [1] "n16" "n106" "n122" "n188" "n730"9 Name Balances
To find the taxonomic labels that correspond to these balances we can
use the function philr::name.balance(). This funciton uses a simple
voting scheme to name the two descendent clades of a given balance
separately. For a given clade, the taxonomy table is subset to only
contain taxa from that clade. Starting at the finest taxonomic rank
(e.g., species) the subset taxonomy table is checked to see if any
label (e.g., species name) represents ≥ threshold (default 95%) of the
table entries at that taxonomic rank. If no consensus identifier is
found, the table is checked at the next-most specific taxonomic rank
(etc…).
tc.names <- sapply(top.coords, function(x) name.balance(tree, tax, x))
tc.names## n16
## "Kingdom_Bacteria/Phylum_Firmicutes"
## n106
## "Order_Actinomycetales/Order_Actinomycetales"
## n122
## "Kingdom_Bacteria/Phylum_Cyanobacteria"
## n188
## "Genus_Campylobacter/Phylum_Proteobacteria"
## n730
## "Order_Bacteroidales/Order_Bacteroidales"We can also get more information on what goes into the naming by viewing the votes directly.
votes <- name.balance(tree, tax, 'n730', return.votes = c('up', 'down'))
votes[[c('up.votes', 'Family')]] # Numerator at Family Level## votes
## Porphyromonadaceae
## 1votes[[c('down.votes', 'Family')]] # Denominator at Family Level## votes
## Bacteroidaceae Porphyromonadaceae Prevotellaceae Rikenellaceae
## 12 9 10 510 Visualize Results
library(ggtree); packageVersion("ggtree")## [1] '3.4.0'library(dplyr); packageVersion('dplyr')## [1] '1.0.8'Above we found the top 5 coordinates (balances) that distinguish
whether a sample is from a human or non-human source. Now using the
ggtree (Yu et al. 2016) package we can visualize these balances on the tree
using the geom_balance object. To use these functions we need to
know the acctual node number (not just the names we have given) of
these balances on the tree. To convert between node number and name,
we have added the functions philr::name.to.nn() and
philr::nn.to.name(). In addition, it is important that we know
which clade of the balance is in the numerator (+) and which is in the
denominator (-) of the log-ratio. To help us keep track we have
created the function philr::annotate_balance() which allows us to
easily label these two clades.
tc.nn <- name.to.nn(tree, top.coords)
tc.colors <- c('#a6cee3', '#1f78b4', '#b2df8a', '#33a02c', '#fb9a99')
p <- ggtree(tree, layout='fan') +
geom_balance(node=tc.nn[1], fill=tc.colors[1], alpha=0.6) +
geom_balance(node=tc.nn[2], fill=tc.colors[2], alpha=0.6) +
geom_balance(node=tc.nn[3], fill=tc.colors[3], alpha=0.6) +
geom_balance(node=tc.nn[4], fill=tc.colors[4], alpha=0.6) +
geom_balance(node=tc.nn[5], fill=tc.colors[5], alpha=0.6)
p <- annotate_balance(tree, 'n16', p=p, labels = c('n16+', 'n16-'),
offset.text=0.15, bar=FALSE)
annotate_balance(tree, 'n730', p=p, labels = c('n730+', 'n730-'),
offset.text=0.15, bar=FALSE)
We can also view the distribution of these 5 balances for human/non-human sources.
In order to plot with ggplot2 we first need to convert the PhILR transformed
data to long format. We have included a function philr::convert_to_long() for
this purpose.
gp.philr.long <- convert_to_long(gp.philr, human.samples) %>%
filter(coord %in% top.coords)## Warning: `gather_()` was deprecated in tidyr 1.2.0.
## Please use `gather()` instead.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_lifecycle_warnings()` to see where this warning was generated.ggplot(gp.philr.long, aes(x=labels, y=value)) +
geom_boxplot(fill='lightgrey') +
facet_grid(.~coord, scales='free_x') +
labs(x = 'Human', y = 'Balance Value') +
theme_bw()
11 Use Balances for Dimension Reduction
Lets just look at balance n16 vs. balance n730 (the ones we annotated in the above tree).
library(tidyr); packageVersion('tidyr')## [1] '1.2.0'gp.philr.long %>%
dplyr::rename(Human=labels) %>%
dplyr::filter(coord %in% c('n16', 'n730')) %>%
tidyr::spread(coord, value) %>%
ggplot(aes(x=n16, y=n730, color=Human)) +
geom_point(size=4) +
labs(x = tc.names['n16'], y = tc.names['n730']) +
theme_bw()
12 Package versions
sessionInfo()## R version 4.2.0 RC (2022-04-19 r82224)
## Platform: x86_64-pc-linux-gnu (64-bit)
## Running under: Ubuntu 20.04.4 LTS
##
## Matrix products: default
## BLAS: /home/biocbuild/bbs-3.15-bioc/R/lib/libRblas.so
## LAPACK: /home/biocbuild/bbs-3.15-bioc/R/lib/libRlapack.so
##
## locale:
## [1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C
## [3] LC_TIME=en_GB LC_COLLATE=C
## [5] LC_MONETARY=en_US.UTF-8 LC_MESSAGES=en_US.UTF-8
## [7] LC_PAPER=en_US.UTF-8 LC_NAME=C
## [9] LC_ADDRESS=C LC_TELEPHONE=C
## [11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C
##
## attached base packages:
## [1] stats4 stats graphics grDevices utils datasets methods
## [8] base
##
## other attached packages:
## [1] tidyr_1.2.0 ggtree_3.4.0
## [3] glmnet_4.1-4 Matrix_1.4-1
## [5] dplyr_1.0.8 mia_1.4.0
## [7] MultiAssayExperiment_1.22.0 TreeSummarizedExperiment_2.4.0
## [9] Biostrings_2.64.0 XVector_0.36.0
## [11] SingleCellExperiment_1.18.0 SummarizedExperiment_1.26.0
## [13] Biobase_2.56.0 GenomicRanges_1.48.0
## [15] GenomeInfoDb_1.32.0 IRanges_2.30.0
## [17] S4Vectors_0.34.0 BiocGenerics_0.42.0
## [19] MatrixGenerics_1.8.0 matrixStats_0.62.0
## [21] ggplot2_3.3.5 ape_5.6-2
## [23] philr_1.22.0 BiocStyle_2.24.0
##
## loaded via a namespace (and not attached):
## [1] fastmatch_1.1-3 plyr_1.8.7
## [3] igraph_1.3.1 lazyeval_0.2.2
## [5] splines_4.2.0 BiocParallel_1.30.0
## [7] scater_1.24.0 digest_0.6.29
## [9] foreach_1.5.2 yulab.utils_0.0.4
## [11] htmltools_0.5.2 magick_2.7.3
## [13] viridis_0.6.2 fansi_1.0.3
## [15] magrittr_2.0.3 memoise_2.0.1
## [17] ScaledMatrix_1.4.0 cluster_2.1.3
## [19] DECIPHER_2.24.0 colorspace_2.0-3
## [21] blob_1.2.3 ggrepel_0.9.1
## [23] xfun_0.30 crayon_1.5.1
## [25] RCurl_1.98-1.6 jsonlite_1.8.0
## [27] survival_3.3-1 phangorn_2.8.1
## [29] iterators_1.0.14 glue_1.6.2
## [31] gtable_0.3.0 zlibbioc_1.42.0
## [33] DelayedArray_0.22.0 phyloseq_1.40.0
## [35] BiocSingular_1.12.0 Rhdf5lib_1.18.0
## [37] shape_1.4.6 scales_1.2.0
## [39] DBI_1.1.2 Rcpp_1.0.8.3
## [41] viridisLite_0.4.0 decontam_1.16.0
## [43] gridGraphics_0.5-1 tidytree_0.3.9
## [45] bit_4.0.4 rsvd_1.0.5
## [47] ellipsis_0.3.2 pkgconfig_2.0.3
## [49] farver_2.1.0 scuttle_1.6.0
## [51] sass_0.4.1 utf8_1.2.2
## [53] ggplotify_0.1.0 tidyselect_1.1.2
## [55] labeling_0.4.2 rlang_1.0.2
## [57] reshape2_1.4.4 munsell_0.5.0
## [59] tools_4.2.0 cachem_1.0.6
## [61] cli_3.3.0 DirichletMultinomial_1.38.0
## [63] generics_0.1.2 RSQLite_2.2.12
## [65] ade4_1.7-19 evaluate_0.15
## [67] biomformat_1.24.0 stringr_1.4.0
## [69] fastmap_1.1.0 yaml_2.3.5
## [71] knitr_1.38 bit64_4.0.5
## [73] purrr_0.3.4 nlme_3.1-157
## [75] sparseMatrixStats_1.8.0 aplot_0.1.3
## [77] compiler_4.2.0 beeswarm_0.4.0
## [79] treeio_1.20.0 tibble_3.1.6
## [81] bslib_0.3.1 stringi_1.7.6
## [83] highr_0.9 lattice_0.20-45
## [85] vegan_2.6-2 permute_0.9-7
## [87] multtest_2.52.0 vctrs_0.4.1
## [89] pillar_1.7.0 lifecycle_1.0.1
## [91] rhdf5filters_1.8.0 BiocManager_1.30.17
## [93] jquerylib_0.1.4 BiocNeighbors_1.14.0
## [95] data.table_1.14.2 bitops_1.0-7
## [97] irlba_2.3.5 patchwork_1.1.1
## [99] R6_2.5.1 bookdown_0.26
## [101] gridExtra_2.3 vipor_0.4.5
## [103] codetools_0.2-18 MASS_7.3-57
## [105] assertthat_0.2.1 rhdf5_2.40.0
## [107] withr_2.5.0 GenomeInfoDbData_1.2.8
## [109] mgcv_1.8-40 parallel_4.2.0
## [111] quadprog_1.5-8 grid_4.2.0
## [113] ggfun_0.0.6 beachmat_2.12.0
## [115] rmarkdown_2.14 DelayedMatrixStats_1.18.0
## [117] ggbeeswarm_0.6.0References
Caporaso, J Gregory, Christian L Lauber, William A Walters, Donna Berg-Lyons, Catherine A Lozupone, Peter J Turnbaugh, Noah Fierer, and Rob Knight. 2011. “Global Patterns of 16S rRNA Diversity at a Depth of Millions of Sequences Per Sample.” Journal Article. Proceedings of the National Academy of Sciences 108: 4516–22.
Egozcue, J. J., and V. Pawlowsky-Glahn. 2016. “Changing the Reference Measure in the Simlex and Its Weightings Effects.” Journal Article. Austrian Journal of Statistics 45 (4): 25–44.
Huang, Ruizhu, Charlotte Soneson, Felix G.M. Ernst, Kevin C. Rue-Albrecht, Guangchuang Yu, Stephanie C. Hicks, and Mark D. Robinson. 2021. “TreeSummarizedExperiment: A S4 Class for Data with Hierarchical Structure [Version 2].” F1000Research 9: 1246. https://doi.org/10.12688/f1000research.26669.2.
Lahti, L, FGM Ernst, SA Shetty, T Borman, T Huang, DJ Braccia, and HC Bravo. 2020. “Upgrading the R/Bioconductor Ecosystem for Microbiome Research [Version 1].” F1000Research 9: 1464. https://doi.org/10.7490/f1000research.1118447.1.
McMurdie, P. J., and S. Holmes. 2013. “Phyloseq: An R Package for Reproducible Interactive Analysis and Graphics of Microbiome Census Data.” Journal Article. PLoS One 8 (4): e61217. https://doi.org/10.1371/journal.pone.0061217.
Silverman, Justin D, Alex D Washburne, Sayan Mukherjee, and Lawrence A David. 2017. “A Phylogenetic Transform Enhances Analysis of Compositional Microbiota Data.” eLife 6. https://doi.org/10.7554/eLife.21887.
Yu, Guangchuang, David K Smith, Huachen Zhu, Yi Guan, and Tommy Tsan‐Yuk Lam. 2016. “Ggtree: An R Package for Visualization and Annotation of Phylogenetic Trees with Their Covariates and Other Associated Data.” Journal Article. Methods in Ecology and Evolution. https://doi.org/10.1111/2041-210X.12628.