---
title: "Exploring and updating FDR benchmarking results"
author:
- name: Patrick Kimes
affiliation: Dana-Farber Cancer Institute
- name: Keegan Korthauer
affiliation: Dana-Farber Cancer Institute
- name: Stephanie Hicks
affiliation: Johns Hopkins University
output:
BiocStyle::html_document:
toc_float: true
package: benchmarkfdrData2019
abstract: |
Instructions on how to re-create results from Korthauer and Kimes et al. (2019) on benchmarking methods that control the false discovery rate (FDR).
vignette: |
%\VignetteIndexEntry{Exploring and updating FDR benchmarking results}
%\VignetteEngine{knitr::rmarkdown}
---
```{r, echo=FALSE, results="hide", message=FALSE}
require(knitr)
opts_chunk$set(error=FALSE, message=FALSE, warning=FALSE)
```
```{r style, echo=FALSE, results='asis'}
BiocStyle::markdown()
```
# Introduction
This package provides a _R_ / _Bioconductor_ resource to
re-create plots and extend the analyses of
[Korthauer and Kimes et al. (2019)](https://www.biorxiv.org/content/10.1101/458786v1).
In this paper, methods controlling the False Discovery Rate (FDR) were applied
to a collection of simulated and biological data sets to generate the
benchmarking summaries provided with this package. Here, we give an example of
how to load summary objects, plot results, and apply a new method to the
dataset.
The package can be installed from R (version >= 3.6) using the
[`BiocManager`](https://cran.r-project.org/package=BiocManager) package,
available on CRAN.
```{r install-bioc, eval = FALSE}
if (!requireNamespace("BiocManager", quietly = TRUE))
install.packages("BiocManager")
BiocManager::install("benchmarkfdrData2019")
```
# Load Packages
```{r load-pkgs}
suppressPackageStartupMessages({
library(ExperimentHub)
library(benchmarkfdrData2019)
library(SummarizedBenchmark)
library(dplyr)
library(ggplot2)
library(rlang)
})
```
In addition to the `r Biocpkg("ExperimentHub")` and `r Biocpkg("benchmarkfdrData2019")`
packages, we also load the `r Biocpkg("SummarizedBenchmark")` package. Benchmarking
results made available with this package for all case studies and simulations described
in Korthauer and Kimes et al. (2019) were created using the `SummarizedBenchmark`
package and are stored as _SummarizedBenchmark_ objects.
However, note that the objects were generated using the [`fdrbenchmark`](https://github.com/areyesq89/SummarizedBenchmark/tree/fdrbenchmark)
branch of the corresponding `SummarizedBenchmark` GitHub repository, and do not include
all of the features described in newer versions of the package (e.g. available on _Bioconductor_).
In this vignette, we use the release version of the `SummarizedBenchmark` package
[available on Bioconductor](http://bioconductor.org/packages/SummarizedBenchmark/).
However, the `fdrbenchmark` version of the `SummarizedBenchmark` package can be
installed from GitHub, again, using the `BiocManager` package.
```{r, eval = FALSE}
BiocManager::install("areyesq89/SummarizedBenchmark", ref = "fdrbenchmark")
```
# Load Data
The data are available for downloaded from the _Bioconductor_
[ExperimentHub](http://bioconductor.org/packages/ExperimentHub)
web resource. The complete list of resources availble with the
`benchmarkfdrData2019` package can be queried using the following
command.
```{r}
hub <- ExperimentHub()
bfdrData <- query(hub, "benchmarkfdrData2019")
bfdrData
```
The above command only returns the metadata associated with each
data object available on ExperimentHub. Individual resources must
be retrieved from ExperimentHub before they can be loaded in R.
Here, we retrieve and load two resource objects to illustrate the
analyses that can be performed using the data available with this package.
First, we load the benchmark results for a ChIP-seq case study where differential
binding was tested using the `r Biocpkg("csaw")` package with region width used
as the independent covariate. The resource is stored with the title `"cbp-csaw-benchmark"`.
First, we determine the corresponding ExperimentHub ID for the resource.
```{r getid-chipres}
cbp_id <- bfdrData$ah_id[bfdrData$title == "cbp-csaw-benchmark"]
cbp_id
```
Using the ID, we can now access the metadata associated with the resource by
subsetting `bfdrData` using single brackets (`[`). Using double brackets (`[[`)
will retrieve the resource from the ExperimentHub server.
```{r load-chipres}
bfdrData[cbp_id]
chipres <- bfdrData[[cbp_id]]
chipres
```
Next, we load the benchmark results from a yeast in silico RNA-seq experiment
where differential expression was tested using DESeq2 with a strong (simulated)
independent and informative covariate. Unlike the ChIP-seq analysis above, with
the in silico experiment, we know ground truth, and therefore can evaluate FDR
control as well as the true positive rate (TPR) at nominal FDR significance
thresholds.
Since the in silico experiments were repeated 100 times, the data object is
a list of 100 _SummarizedBenchmark_ objects for each replication. The resource
is stored with the title `"yeast-results-de5"`. Here, we demonstrate an alternative approach
to retrieving the resource from the ExperimentHub server. Rather than subset `bfdrData`
using double brackets (`[[`), we retrieve the resource by calling the
resource name as a function (`yeast-results-de5()`). This functionality is available
for all resources available with this package (including the ChIP-seq resource
loaded above).
```{r load-yeastres}
yeast_id <- bfdrData$ah_id[bfdrData$title == "yeast-results-de5"]
bfdrData[yeast_id]
yeastres <- `yeast-results-de5`()
length(yeastres)
yeastres[[1]]
```
To be able to work with the latest release of the `SummarizedBenchmark` package,
we must fill in a missing slot of the _SummarizedBenchmark_ objects.
```{r fix-sbobjs}
chipres@BenchDesign <- BenchDesign()
yeastres <- lapply(yeastres, function(x) { x@BenchDesign <- BenchDesign(); x })
```
## SummarizedBenchmark Objects
The _SummarizedBenchmark_ objects include the original p-values, informative covariate,
and corrected significance values for the various methods compared in Korthauer and Kimes et al. (2019).
_SummarizedBenchmark_ objects are an extension of the Bioconductor _SummarizedExperiment_
class, with results organized as a rectangular "assay", with associated row and column
metadata. Here, the rows of the objects correspond to individual hypothesis tests and
the columns correspond to the approaches used for multiple testing correction.
We can take a look at the names of the methods included in the ChIP-seq results object.
```{r colnames}
colnames(chipres)
```
Notice that the results include the IHW and BL methods multiple times.
These `ihw-` and `bl-` columns correspond to separate runs of the
methods with different parameter settings. Briefly, the IHW method requires specifying an
alpha FDR threshold while running the method. Here, the method was run with alpha values of
`0.01, 0.02, .., 0.10`. The BL method was run with spline degrees of freedom `2, 3, 4, 5`.
The corrected significance returned by each method is included in the single assay, `"bench"`
(corresponding to the benchmarked results).
```{r show-assay}
dim(assay(chipres, "bench"))
head(assay(chipres, "bench"))
```
The ASH (`ashq`) results are `NA` as the method was not applied to the data.
# Exploratory Analysis
Given the multiple-testing-corrected results provided in the `"bench"` assay of the
_SummarizedBenchmark_ objects, we can take a look at several performance metrics
to compare the various methods. For the ChIP-seq case study, we can take a look at the
number of rejections at various significance cutoffs. With the in silico yeast
experiments, since truth is known, we can also look at FDR and TPR, as well as other
related metrics.
_SummarizedBenchmark_ objects include functionality to easily add and evaluate metrics
for data stored as assays. This is performed by first adding performance metrics
with `addPerformanceMetric`, followed by a call to `estimatePerformanceMetrics`.
While custom performance metrics can be defined by users, the package fortunately
includes several default metrics that can be added by name.
```{r default-metrics}
availableMetrics()
```
## ChIP-seq Case Study
We will add the `"rejections"` metric to the `"bench"` assay and compute the number
of rejections for each method at cutoffs between 0.01 and 0.10.
```{r chip-add-metrics}
chipres <- addPerformanceMetric(chipres,
evalMetric = "rejections",
assay = "bench")
```
Next, we compute the number of rejections and organize this as a tidy data.frame.
```{r chip-est-metrics}
chipdf <- estimatePerformanceMetrics(chipres,
alpha = seq(0.01, 0.10, by = .01),
tidy = TRUE)
dim(chipdf)
head(chipdf)
```
Each row in the data.frame corresponds to a `method + metric + cutoff` combination
(e.g. `"unadjusted" + "rejections" + "alpha = 0.01"`). This information is stored in
the `"label"`, `"performanceMetric"`, and `"alpha"` columns, with the corresponding
metric value in the `"value"` column. All other columns contain method metadata,
such as the package version, when the method was evaluated.
We will now clean up the IHW and BL methods which, as described above, include multiple
parameter settings.
```{r chip-clean-cols}
## subset IHW
chipdf <- dplyr:::filter(chipdf, !(grepl("ihw", label) & param.alpha != alpha))
chipdf <- dplyr:::mutate(chipdf, label = gsub("(ihw)-a\\d+", "\\1", label))
## subset BL
chipdf <- dplyr:::filter(chipdf, ! label %in% paste0("bl-df0", c(2, 4, 5)))
```
We only keep a subset of the columns and drop NA values.
```{r chip-subset-cols}
chipdf <- dplyr::select(chipdf, label, performanceMetric, alpha, value)
chipdf <- dplyr::filter(chipdf, !is.na(value))
head(chipdf)
```
We now plot the number of rejections.
```{r chip-plot-rejs, fig.width = 8, fig.height = 5}
ggplot(chipdf, aes(x = alpha, y = value, color = label)) +
geom_point() +
geom_line() +
scale_color_viridis_d("Method") +
scale_x_continuous(breaks = seq(0, 1, .01), limits = c(0, .11)) +
ylab("Rejections") +
theme_bw() +
ggtitle("Number of rejections across multiple-testing methods",
"ChIP-seq CBP differential analysis with informative covariate")
```
## Yeast in silico Data
We can similarly add performance metrics to each replication of the
yeast in silico experiment and aggregate across replicates. We
demonstrate this process using a subset of the 100 replications
in the interest of computational cost.
```{r yeast-subset}
yeastres10 <- yeastres[1:10]
```
As with the ChIP-seq results, we can add and evaluate performance
metrics using `addPerformanceMetric` and `estimatePerformanceMetrics`.
However, note that the yeast in silico results already include
several default performance metrics.
```{r yeast-already-metrics}
names(performanceMetrics(yeastres10[[1]])[["qvalue"]])
```
We can skip the process of adding performance metrics and just use
these metrics.
```{r yeast-eval-metrics}
yeastdf <- lapply(yeastres10, estimatePerformanceMetrics,
alpha = seq(0.01, 0.10, by = .01), tidy = TRUE)
```
Finally, we merge the 10 replications to a single data.frame.
```{r yeast-merge-reps}
yeastdf <- dplyr::bind_rows(yeastdf, .id = "rep")
```
As above, we clean IHW and BL results, remove `NA` values, and only keep
a subset of useful columns.
```{r yeast-clean-cols}
## subset IHW
yeastdf <- dplyr:::filter(yeastdf,
!(grepl("ihw", label) & param.alpha != alpha))
yeastdf <- dplyr:::mutate(yeastdf, label = gsub("(ihw)-a\\d+", "\\1", label))
## subset BL
yeastdf <- dplyr:::filter(yeastdf, ! label %in% paste0("bl-df0", c(2, 4, 5)))
yeastdf <- dplyr::select(yeastdf, rep, label, performanceMetric, alpha, value)
yeastdf <- dplyr::filter(yeastdf, !is.na(value))
head(yeastdf)
```
Finally, we summarize across replications for each method, for each metric,
at each nominal threshold.
```{r yeast-summarize}
yeastdf <- dplyr::group_by(yeastdf, label, performanceMetric, alpha)
yeastdf <- dplyr::summarize(yeastdf,
meanValue = mean(value),
seValue = sd(value) / sqrt(n()))
yeastdf <- dplyr::ungroup(yeastdf)
```
Now, we can plot the average and standard errors across replicates for each
method. Here, we will just plot FDR and TPR.
```{r yeast-plot-all, fig.width = 9, fig.height = 5}
yeastdf %>%
dplyr::filter(performanceMetric %in% c("FDR", "TPR")) %>%
ggplot(aes(x = alpha, y = meanValue,
color = label,
ymin = meanValue - seValue,
ymax = meanValue + seValue)) +
geom_point() +
geom_errorbar(width = .01 / 4, alpha = 1/4) +
geom_line(alpha = 1/2) +
scale_color_viridis_d("Method") +
scale_x_continuous(breaks = seq(0, 1, .01), limits = c(0, .11)) +
facet_wrap(~ performanceMetric, scales = 'free_y', nrow = 1) +
ylab("average across replicates") +
theme_bw() +
geom_abline(aes(intercept = i_, slope = s_), color = 'red', linetype = 2,
data = tibble(performanceMetric = 'FDR', i_ = 0, s_ = 1)) +
ggtitle("FDR and TPR across multiple-testing methods",
"yeast in silico experiment with informative covariate")
```
We have also included a red line to the FDR plot to assess whether methods
are appropriately controlling the FDR at the nominal thresholds.
# Adding Methods
The summary objects made available with this packages were constructed using an older version
of the `SummarizedBenchmark` package. Since then, functions have been added to the package
for updating benchmark comparisons and adding new methods to an existing _SummarizedBenchmark_
object. Unfortuntately, these functions cannot be used with the current objects made
available with this package.
However, the summary objects include both unadjusted p-values (as the `unadjusted` column) and
corresponding independent covariate values (as the `ind_covariate` rowData column). These values
can be used to apply new methods which only depend on these inputs.
```{r chip-show-data}
dat <- tibble(pval = assay(chipres)[, "unadjusted"],
covariate = rowData(chipres)$ind_covariate)
dat
```
To analyze this data using the `SummarizedBenchmark` package, we can
construct a new _BenchDesign_ object with any collection of new methods
to benchmark, and the above data as input. More details on the
_SummarizedBenchmark_ package and the _BenchDesign_ class can be found in
[the package vignettes](http://bioconductor.org/packages/SummarizedBenchmark/).
As an illustration of how a new method could be applied to this data, we will
re-apply the Benjamini-Hochberg correction and show that the results that we
obtain in this re-analysis match the results reported in the loaded
_SummarizedBenchmark_ object.
We construct a _BenchDesign_ with a single method, `newBH`, and the data given above.
```{r chip-new-benchdesign}
bh_method <- BDMethod(x = p.adjust,
params = rlang::quos(p = pval,
method = "BH"))
new_design <- BenchDesign(newBH = bh_method, data = dat)
new_design
```
To evaluate the benchmark experiment stored in the _BenchDesign_ object, we call
`buildBench`.
```{r chip-eval-buildBench}
new_res <- buildBench(new_design)
new_res
```
Now that we have a _SummarizedBenchmark_ object, we can evaluate performance
metrics as before.
```{r chip-new-metrics}
new_res <- addPerformanceMetric(new_res,
evalMetric = "rejections",
assay = "default")
new_df <- estimatePerformanceMetrics(new_res,
alpha = seq(0.01, 0.10, by = 0.01),
tidy = TRUE)
```
Finally, we subset on columns of interest and examine how many tests would be
rejected using BH in our re-analysis.
```{r chip-new-result}
new_df <- dplyr::select(new_df, label, value, performanceMetric, alpha)
new_df
```
We verify that this matches the number of rejections reported in the
results from above.
```{r chip-verify-bh}
dplyr::filter(chipdf, label == "bh")
```
Since the results are now just data.frame objects with similar columns,
they can be combined to compare new results with previous results.
# Session Information
```{r}
sessionInfo()
```