In stephaniehicks/benchmarkfdrData2019: Data and Benchmarking Results from Korthauer and Kimes et al. (2019)

knitr::opts_chunk$set(echo = TRUE)

Summary

Here we download and analyze the Schubert diarrhea dataset (Schubert et al., 2014). The dataset includes 154 healthy nondiarrheal, 93 Clostridium difficile infection (CDI) associated diarrheal, and 89 non-CDI associated diarrheal stool samples. We'll download the processed OTU tables from Zenodo and unzip them in the data/cdi_schubert_results folder.

Diarrhea comes with a broad community re-structuring, so we expect there to be many truly differentially abundant OTUs in this dataset.

Workspace Setup

library(dplyr)
library(ggplot2)
library(magrittr)
library(SummarizedBenchmark)

## load helper functions
for (f in list.files("../R", "\\.(r|R)$", full.names = TRUE)) {
    source(f)
}

## project data/results folders
datdir <- "data"
resdir <- "results"
sbdir <- "../../results/microbiome"
dir.create(datdir, showWarnings = FALSE)
dir.create(resdir, showWarnings = FALSE)
dir.create(sbdir, showWarnings = FALSE)

cdi_data <- file.path(datdir, "cdi_schubert_results.tar.gz")
cdi_dir <- file.path(datdir, "cdi_schubert_results")
dir.create(cdi_dir, showWarnings = FALSE)

otu_result_file <- file.path(resdir, "schubert-otus-results.rds")
otu_bench_file <- file.path(sbdir, "schubert-otus-benchmark.rds")
otu_bench_file_abun <- file.path(sbdir, "schubert-otus-abun-benchmark.rds")
otu_bench_file_uninf <- file.path(sbdir, "schubert-otus-uninf-benchmark.rds")

Data Preparation

if (!file.exists(cdi_data)) {
    download.file("https://zenodo.org/record/840333/files/cdi_schubert_results.tar.gz",
                  destfile = cdi_data)
}

if (!file.exists(file.path(cdi_dir, "cdi_schubert.metadata.txt"))) {
    untar(cdi_data, exdir = datdir)
}

Next, we'll read in the unzipped OTU table and metadata files into R.

## load OTU table and metadata
otu <- read.table(file.path(cdi_dir, "RDP",
                            "cdi_schubert.otu_table.100.denovo.rdp_assigned"))
meta <- read.csv(file.path(cdi_dir, "cdi_schubert.metadata.txt"), sep = '\t')

# Keep only samples with the right DiseaseState metadata
meta <- filter(meta, DiseaseState %in% c("H", "nonCDI", "CDI"))

# Keep only samples with both metadata and 16S data
keep_samples <- intersect(colnames(otu), meta$sample_id)
otu <- otu[, keep_samples]
meta <- filter(meta, sample_id %in% keep_samples)

Since we'll be using OTU-wise covariates, we shouldn't need to perform any filtering/cleaning of the OTUs, apart from removing any that are all zeros. (This may happen after removing shallow samples, I think.) We still apply a minimum threshold of 10 reads per OTU across all samples. After removing these shallow OTUs, we also get rid of any samples with too few reads. We define the minimum number of reads per OTU in min_otu, and the minimum number of reads per sample in min_sample.

After we've removed any shallow OTUs and samples, we'll convert the OTU table to relative abundances.

min_otu <- 10
minp_otu <- 0.01
min_sample <- 100

## Remove OTUs w/ <= min reads, w/ <= min prop, samples w/ <= min reads
otu <- otu[rowSums(otu) > min_otu, ]
otu <- otu[rowSums(otu > 0) / ncol(otu) > minp_otu, ]
otu <- otu[, colSums(otu) > min_sample]

## Update metadata with new samples
meta <- dplyr::filter(meta, sample_id %in% colnames(otu))

## Remove empty OTUs
otu <- otu[rowSums(otu) > 0, ]

## Convert to relative abundance
abun_otu <- t(t(otu) / rowSums(t(otu)))

## Add pseudo counts
zeroabun <- 0
abun_otu <- abun_otu + zeroabun

Data Analysis

Differential Testing

Next, we need to calculate the pvalues, effect size, and standard error for each OTU. Here, we'll compare diarrhea vs. healthy. Diarrhea will include both CDI and nonCDI patients. We'll put these results into a dataframe, and label the columns with the standardized names for downstream use (pval, SE, effect_size, test_statistic). The test statistic is the one returned by wilcox.test().

Note that the effect here is calculated as logfold change of mean abundance in controls relative to cases (i.e. log(mean_abun[controls]/mean_abun[cases]))

While we're at it, we'll also calculate the mean abundance and ubiquity (detection rate) of each OTU. Later, we can assign their values to a new column called ind_covariate for use in downstream steps.

if (!file.exists(otu_result_file)) {
    res <- test_microbiome(abundance = abun_otu, shift = zeroabun,
                           is_case = meta$DiseaseState %in% c("CDI", "nonCDI"))
    saveRDS(res, file = otu_result_file)
} else {
    res <- readRDS(otu_result_file)
}

Add random (uninformative) covariate.

set.seed(9226)
res$rand_covar <- rnorm(nrow(res))

Finally, let's try to add phylogeny as covariates. Here we'll have columns for each separate taxonomic level.

res <- tidyr::separate(res, otu, 
                       c("kingdom", "phylum", "class", "order",
                         "family", "genus", "species", "denovo"), 
                       sep = ";", remove = FALSE)

Covariate Diagnostics

Here we look to see if the covariates do indeed look informative.

Ubiquity

strat_hist(res, pvalue="pval", covariate="ubiquity", maxy=20, binwidth=0.05)

rank_scatter(res, pvalue="pval", covariate="ubiquity")

Mean Abundance (across non-zero samples)

strat_hist(res, pvalue="pval", covariate="mean_abun_present", maxy=17, binwidth=0.05)

rank_scatter(res, pvalue="pval", covariate="mean_abun_present")

Phylogeny

Let's look at phylum-level stratification first. A priori, I might expect Proteobacteria to be enriched for low p-values? But I don't know if that's super legit, and Eric doesn't seem to think that phylogeny will be informative at all...

ggplot(res, aes(x=pval)) +
    geom_histogram() +
    facet_wrap(~phylum, scales = "free")

Random

strat_hist(res, pvalue="pval", covariate="rand_covar", maxy=17, binwidth=0.05)

rank_scatter(res, pvalue="pval", covariate="rand_covar")

Multiple-Testing Correction - ubiquity

Let's use ubiquity as our ind_covariate.

res <- dplyr::mutate(res, ind_covariate = ubiquity)

First, we'll create an object of BenchDesign class to hold the data and add the benchmark methods to the BenchDesign object. We remove ASH from the comparison.

Then, we'll construct the SummarizedBenchmark object, which will run the functions specified in each method (these are actually sourced in from the helper scripts).

if (!file.exists(otu_bench_file)) {
    bd <- initializeBenchDesign()
    bd <- dropBMethod(bd, "ashq")
    sb <- buildBench(bd, data = res, ftCols = "ind_covariate")
    metadata(sb)$data_download_link <-
                   "https://zenodo.org/record/840333/files/cdi_schubert_results.tar.gz"
    saveRDS(sb, file = otu_bench_file)
} else {
    sb <- readRDS(otu_bench_file)
}

Benchmark Metrics - ubiquity

Next, we'll add the default performance metric for q-value assays. First, we have to rename the assay to 'qvalue'.

assayNames(sb) <- "qvalue"
sb <- addDefaultMetrics(sb)

Now, we'll plot the results.

rejections_scatter(sb, as_fraction=FALSE, supplementary=FALSE)
rejection_scatter_bins(sb, covariate="ind_covariate", supplementary=FALSE)
plotFDRMethodsOverlap(sb, alpha=0.1, supplementary=FALSE, order.by="freq", nsets=100 )

covariateLinePlot(sb, alpha = 0.05, covname = "ind_covariate")

Hm, now the code runs. However, there are clearly still some issues: - ashs rejects all hypotheses (all q-values are essentially 0). - lfdr and scott-empirical are all NaN (I think this is likely related to the df error)

methods <- c("lfdr", "ihw-a10", "bl-df03", "qvalue", "bh", "bonf")
plotCovariateBoxplots(sb, alpha = 0.1, nsets = 6, methods = methods)

assays(sb)[["qvalue"]]["ashs"] %>% max()
sum(is.na(assays(sb)[["qvalue"]]["lfdr"]))
sum(is.na(assays(sb)[["qvalue"]]["scott-empirical"]))
sum(is.na(assays(sb)[["qvalue"]]["scott-theoretical"]))

Plotting methods are giving errors for some reason. Let's try to use Alejandro's code instead.

Multiple-Testing Correction - mean abundance

Let's use mean_abun_present as our ind_covariate.

res <- dplyr::mutate(res, ind_covariate = mean_abun_present)

First, we'll create an object of BenchDesign class to hold the data and add the benchmark methods to the BenchDesign object. We remove ASH from the comparison.

Then, we'll construct the SummarizedBenchmark object, which will run the functions specified in each method (these are actually sourced in from the helper scripts).

if (!file.exists(otu_bench_file_abun)) {
    bd <- initializeBenchDesign()
    bd <- dropBMethod(bd, "ashq")
    sb <- buildBench(bd, data = res, ftCols = "ind_covariate")
    metadata(sb)$data_download_link <-
                   "https://zenodo.org/record/840333/files/cdi_schubert_results.tar.gz"
    saveRDS(sb, file = otu_bench_file_abun)
} else {
    sb <- readRDS(otu_bench_file_abun)
}

Benchmark Metrics (mean abun)

Next, we'll add the default performance metric for q-value assays. First, we have to rename the assay to 'qvalue'.

assayNames(sb) <- "qvalue"
sb <- addDefaultMetrics(sb)

Now, we'll plot the results.

rejections_scatter(sb, as_fraction=FALSE, supplementary=FALSE)
rejection_scatter_bins(sb, covariate="ind_covariate", supplementary=FALSE)
plotFDRMethodsOverlap(sb, alpha=0.1, supplementary=FALSE, order.by="freq", nsets=100 )

covariateLinePlot(sb, alpha = 0.05, covname = "ind_covariate")

Multiple-Testing Correction - random

Let's use rand_covar as our ind_covariate.

res <- dplyr::mutate(res, ind_covariate = rand_covar)

First, we'll create an object of BenchDesign class to hold the data and add the benchmark methods to the BenchDesign object. We remove ASH from the comparison.

Then, we'll construct the SummarizedBenchmark object, which will run the functions specified in each method (these are actually sourced in from the helper scripts).

if (!file.exists(otu_bench_file_uninf)) {
    bd <- initializeBenchDesign()
    bd <- dropBMethod(bd, "ashq")
    sb <- buildBench(bd, data = res, ftCols = "ind_covariate")
    metadata(sb)$data_download_link <-
                   "https://zenodo.org/record/840333/files/cdi_schubert_results.tar.gz"
    saveRDS(sb, file = otu_bench_file_uninf)
} else {
    sb <- readRDS(otu_bench_file_uninf)
}

Benchmark Metrics - random

Next, we'll add the default performance metric for q-value assays. First, we have to rename the assay to 'qvalue'.

assayNames(sb) <- "qvalue"
sb <- addDefaultMetrics(sb)

Now, we'll plot the results.

rejections_scatter(sb, as_fraction=FALSE, supplementary=FALSE)
rejection_scatter_bins(sb, covariate="ind_covariate", supplementary=FALSE)
plotFDRMethodsOverlap(sb, alpha=0.1, supplementary=FALSE, order.by="freq", nsets=100 )

covariateLinePlot(sb, alpha = 0.05, covname = "ind_covariate")

Ubiquity/Abundance comparison

Here we compare the method ranks for the different comparisons at alpha = 0.10.

plotMethodRanks(c(otu_bench_file, otu_bench_file_abun, otu_bench_file_uninf), 
                colLabels = c("OTU-ubiquity","OTU-abun", "OTU-uninf"), 
                alpha = 0.10, xlab = "Comparison")

Session Info

sessionInfo()

stephaniehicks/benchmarkfdrData2019 documentation built on June 20, 2021, 10 a.m.