Detailed documentation of RNA-Seq downstream¶

Here we describe in detail the downstream analysis of RNA-Seq data performed in workflows/rnaseq/downstream/rnaseq.Rmd.

This page has one section per named chunk. For example, if rnaseq.Rmd has the following code:

```{r load_libraries}
library(DESeq2)
library(dplyr)
```

Then you can find the corresponding documentation on this page under the load_libraries heading.

The file ci/ensure_docs.py double-checks to make sure all chunks are documented and all documentation corresponds to a chunk, as part of the testing framework.

`global_options`¶

This chunk sets global rmarkdown options. Some of the lines provide a mechanism for all cached chunks to have their cache invalidated if either of the filenames’ modification times have changed. For example, the following argument to knitr::opts_chunk$set will inject the option cache.extra_file_dep_1 into all chunks:

cache.extra_file_dep_1=file.info('../config/sampletable.tsv')$mtime,

Therefore if the sample table of the RNA-Seq workflow has been updated since the last time the analysis script was run, the modification time (mtime) will be changed, so the cache.extra_file_dep_1 value will be different than it was before for every chunk, and so every chunk will be re-run.

`lcdbwf`¶

Loads the lcdbwf R package, stored in ../../../lib/lcdbwf/R. This chunk is not cached and fully reloads the package each time using devtools::load_all, so any changes to the code in that package will show up when this file is rendered. Documentation is also automatically re-generated.

Note that throughout this RMarkdown file, functions from this package will use the lcdbwf:: namespace prefix to be explicit about where that function is coming from.

`config`¶

This chunk loads the config files config.yaml and text.yaml. This chunk is not cached, so any changes in the config files automatically show up here.

See the config.yaml file for configuring the code.

See the text.yaml file for editing the explanatory text.

This chunk also configures the parallelization options in a chunk that is not cached.

`libraries`¶

Standard loading of the library dependencies used throughout the code.

`coldata_setup`¶

This chunk loads the sample table.

Use this chunk to add additional columns to your sampletable.

The only requirement is that the rownames correspond to sample names.

`dds_initial`¶

Constructs the initial dds object and the variance stabilized counts.

Note that the design used is ~1 and the call to varianceStabilizingTransformation uses blind=TRUE, so you don’t need to change anything here.

Also note that the entire config object is passed to the make_dds function, which reads options like whether to strip version numbers off of gene IDs or whether (and how) to collapse technical replicates.

`print_coldata`¶

Simply prints the colData for reference – excluding columns that might be in there that would clutter the output.

`sample_heatmap`¶

This chunk creates a clustered heatmap of sample distances. The columns specified in the config file’s “covariates_for_plots” item will show up as colors along the right side.

`pca`¶

Creates PCA plots, one tab per entry in the config file’s “covariates_for_plots” item. Each plot has the same points, but the colors differ. This can help assess the experimental design and set expectations on how many differentially expressed genes one may find.

These are interactive plots, and hovering over a point indicates the sample.

`sizefactors`¶

This chunk makes diagnostic plots. In general, we expect sizeFactors to correlate with total read count. When it doesn’t, it can indicate that a small number of genes are very highly expressed.

`dds_list`¶

This chunk sets up the dds objects to be used in the results section below for differential expression detection.

You may need different dds objects for testing different models, or perhaps removing outlier samples. If you have technical replicates you might need to combine them, and you might need to remove gene version identifiers. You might want to use salmon instead of featureCounts. These would need to be done for each dds, requiring code duplication.

After working on many complex and/or messy experimental designs, we have settled on the approach of a named list of dds objects, where later code refers to these objects by their name in the list.

The results chunk below expects such a list.

The simplest example is the following where we create a single dds and put it into a list.

dds <- DESeqFromCombinedFeatureCounts(
   '../data/rnaseq_aggregation/featurecounts.txt',
   sampletable=colData,
   design=~group)
dds <- DESeq(dds, parallel=parallel)

dds.list <- list(main=dds)

Now imagine a case where we want to remove a replicate that we think is an outlier, but we still want to compare it to the results when it is included. Let’s say we also need to collapse the technical replicates. Such code would look like this:

# The long way...
#
# First object with all replicates
dds1 <- DESeqFromCombinedFeatureCounts(
   '../data/rnaseq_aggregation/featurecounts.txt',
   sampletable=colData,
   design=~group)
dds1 <- collapseReplicates(dds1, 'biorep')
dds1 <- DESeq(dds1, parallel=parallel)

# Similar to above, but remove replicate 4
dds2 <- DESeqFromCombinedFeatureCounts(
   '../data/rnaseq_aggregation/featurecounts.txt',
   sampletable=colData %>% filter(replicate!='rep4'),
   design=~group,
   # need subset_counts=TRUE if we want to automatically
   # subset the featureCounts to match the filtered colData
   # we provided.
   subset_counts=TRUE
   )
dds2 <- collapseReplicates(dds, 'biorep')
dds2 <- DESeq(dds2, parallel=parallel)

Based on our experience, as we add more dds objects the code gets more error-prone. So for more complex use-cases, we have a function lcdbwf::make_dds.

Here is how the code above would look using this method:

lst <- list(

   main=list(sampletable=colData, design=~group),

   no.rep.4=list(
      sampletable=colData %>% filter(replicate!='rep4'),
      design=~group,
      subset.counts=TRUE))
)

dds.list <- map(lst, lcdbwf::make_dds, config=config, parallel=config$parallel$parallel)

That is, first we create a list of lists (lst), and then we used map() to apply the make_dds function to all items in the list. The collapsing of replicates and other dds-creation configuration like stripping dotted version names is determined by the config object which is passed along.

See the help for lcdbwf::make_dds for more details.

This chunk becomes a dependency of all of the results chunks below.

`dds_diagnostics`¶

If configured, this chunk will run the diagnostics on the dds objects and show tabbed reports on each dds.

`results_*`¶

Note

This is where most of the customization needs to happen for each project.

This is actually a series of chunks where the bulk of the differential expression analysis takes place.

For simple cases, you probably just need one of these. But for complex experimental designs where you end up doing lots of contrasts, it can get time consuming to run them every time you change the RMarkdown file.

The end result of these chunks is a single list containing DESeq2 results objects and associated metadata in (sub)lists. Each of these sublists has:

res, the results object
dds, the string name in names(dds.list)
label, a “nice” label which is used for headings and other output
additional optional arguments that are passed along to DESeq2::results() and/or DESeq2::lfcshrink()

To continue our example from above, we might want to run the same contrast on all samples (the “main” dds) and after removing replicate 4 (the “no.rep.4” dds). To illustrate how additional arguments are used, let’s imagine we also want to use ashr as the shrinkage method for the second contrast.

Use the lcdbwf::make_results() function for this. This function is a loose wrapper around DESeq2::results() and DESeq2::lfcshrink() that adds some extra convenience when working with lists of dds objects, including the detection of parallelization as set up in the config object. See the help for lcdbwf::make_results() for more details.

To take advantage of this infrastructure, we put each of those contrasts into its own chunk according to the following rules:

the chunk name must start with results_
the chunk is cached
the chunk depends on the 'dds_list' chunk
the variable name starts with contr_[index]_, and the rest of the variable name will be used as the name in the list. index is a string, containing 1 or more alphanumeric characters, and will be used as a sorting index for contrasts when generating output files. Note that the index string cannot contain “_”.

Our example would look like the following. Note that we’re showing the chunks here because that will be come meaningful in a moment. They are shown as comments here just to get the syntax highlighting to look OK.

# ```{r results_01, dependson='dds_list', cache=TRUE}
contr_1_ko.vs.wt <- lcdbwf::make_results(
  dds_name='main',
  label='Using all samples',
  contrast=c('genotype', 'KO', 'WT')
)
# ```

# ```{r results_02, dependson='dds_list', cache=TRUE}
contr_2a_no.rep.4 <- lcdbwf::make_results(
  dds_name='no.rep.4',
  label='Removing replicate 4 and using ashr for shrinkage',
  contrast=c('genotype', 'KO', 'WT'),
  type='ashr'
)
# ```

When combined with the assemble_variables chunk described below, this allows us to:

retain caching at the level of individual contrasts
combine all results into a single list used in later code while still respecting dependencies
reduce the bookkeeping overhead

In our experience this scales well with very complex experimental designs with lots of contrasts. For more information on creating complex contrasts, see Specifying contrasts. For more information on how these results are collected, see assemble_variables.

`assemble_variables`¶

If we had put all results() calls into the same chunk and cached that, then a change anywhere in that chunk would invalidate the cache which would cause all results to be regenerated. With many contrasts, this can get quite time-consuming. An alternative would be to put each results() call into its own chunk. But then we would need to keep track of dependencies and ensure those dependencies were specified in downstream chunks.

For example, if you add a new results chunk and cache it, but forget to add that chunk as a dependency in a later chunk, that later chunk will be inconsistent and may even be missing the new results. Keeping track of this can be error-prone.

Our solution is to set up the contrasts according to the rules described above. By following those rules, the following becomes possible:

we can detect all chunks creating results by looking for results_ in the chunk name and automatically inject these into dependencies of future chunks.
we can detect all results objects created by looking for variables starting with contr_[index]_
we can assemble all results objects into a list, and name each item in the list according to its variable name (minus the contr_[index]_).
we can alter the order of contrasts by simply modifying the index string in a single chunk. For example, if we have three contrasts contr_1_ko.vs.wt, contr_2_no.rep.4, and contr_3_no.rep.3, we can change the order of contrasts simply by modifying one index string (ex: change contr_3_no.rep.3 to contr_1a_no.rep.3).

The assemble_variables chunk does all of this. The end result of this chunk is a list of lists that is used by functions in the lcdbwf R package for downstream work. For more details, see res_list.

For each contrast (that is, each entry in res_list) the below chunks will automatically create a DE results section including:

a tabbed section using the label as a header
summary table
MA plot
counts plots of top 3 up- and down-regulated genes
p-value distribution
exported results tables with links

Specifying contrasts¶

Contrasts can be specified in three different ways.

Note

In these examples, “control” and “treatment” are factor levels in the “group” factor (which was in the colData), and the dds object was created with the design ~group:

A character vector to the contrast parameter.

This should be a three element vector:
- the name of a factor in the design formula
- name of the numerator for the fold change
- the name of the denominator for the fold change. E.g.,
```
res <- results(dds, contrast=c('group', 'treatment', 'control')
```
That is, the control must be last.
name parameter for results() function call or coef parameter for lfcShrink() call

name or coef should be one of the values returned by resultsNames(dds) that corresponds to the precomputed results. E.g.
```
resultsNames(dds)
# [1] "Intercept"  "group_treatment_vs_control"

res <- results(dds, name='group_treatment_vs_control')
```
A numeric contrast vector with one element for each element in the resultsNames() function call. This is useful for arbitrary comparisons in multi-factor designs with a grouping variable.
```
resultsNames(dds)
# [1] "Intercept"  "group_treatment_vs_control"

res <- results(dds, contrast=c(0, 1))
```

The most general way to specify contrasts¶

The most general way to specify contrasts is with a numeric vector (third option above).

Here is a worked example, using a two-factor experiment.

group encodes all combinations of a two-factor experiment, so we construct a sampletable that looks like the following (here, showing 2 replicates per group):

sample   genotype   condition   group
      A          I           IA
      A          I           IA
      B          I           IB
      B          I           IB
      A          II          IIA
      A          II          IIA
      B          II          IIB
      B          II          IIB

We can make arbitrary comparisons by fitting an ‘intercept-less’ model, e.g. design=~group + 0, and numeric contrast vectors:

dds <- DESeqDataSetFromCombinedFeatureCounts(
    '../data/rnaseq_aggregation/featurecounts.txt',
    sampletable=colData,
    # NOTE: the design is now different
    design=~group + 0
)
dds <- DESeq(dds)

Check resultsNames:

resultsNames(dds)
# [1] "groupIA"  "groupIB"  "groupIIA"  "groupIIB"

So any numeric vectors we provide must be 4 items long. Here is how we can make various contrasts with this experimental design. In each example, the coefficients are indicated above the resultsNames to make it easier to see.

To compare IA and IB (that is, the genotype effect only in condition I):

#     1          -1         0           0
# "groupIA"  "groupIB"  "groupIIA"  "groupIIB"

res <- results(dds, contrast=c(1, -1, 0, 0)

Effect of genotype B (that is, disregard information about condition):

#     1          -1         1           -1
# "groupIA"  "groupIB"  "groupIIA"  "groupIIB"

res <- results(dds, contrast=c(1, -1, 1, -1)

Effect of condition II (that is, disregard information about genotype):

#     1           1         -1          -1
# "groupIA"  "groupIB"  "groupIIA"  "groupIIB"

res <- results(dds, contrast=c(1, 1, -1, -1)

Interaction term, that is, (IA vs IB) vs (IIA vs IIB). This is effectively (IA - IB) - (IIA - IIB), which in turn becomes IA - IB - IIA + IIB:

#     1          -1         -1          1
# "groupIA"  "groupIB"  "groupIIA"  "groupIIB"

res <- results(dds, contrast=c(1, -1, -1, 1)

`summary`¶

This chunk prints a high-level overview of all the contrasts.

`reportresults`¶

This is the section that creates multiple, tabbed outputs for each of the contrasts in the res_list.

If the config specifies results diagnostics (config$toggle$results_diagnostics is TRUE), then this chunk will also run the diagnostics. You can select just the ones you want diagnostics on using the config$plotting$diagnostics_results_names config option.

`upsetplots`¶

This chunk produces UpSet plots comparing the contrasts.

`excel`¶

This chunk outputs an Excel spreadsheet with one contrast per sheet. Normalized counts for each sample, from the respective dds object used for the contrast, are also included on each sheet.

`write_output`¶

TSVs for each contrast’s results are written to disk.

`combined_rds`¶

A single object is written as an .Rds file. This can then be used for downstream visualization or it can be used as input to the functional enrichment RMarkdown document.

`sessioninfo`¶

The output of sessionInfo records the versions of packages used in the analysis.

Glossary¶

colData¶: The metadata describing the samples. This is originally defined in the sampletable for the entire lcdb-wf run, is imported into rnaseq.Rmd, and may be subsequently modified.
dds¶: DESeq data set object. Typically this is incrementally added to, as in the DESeq2 vignette.
vsd¶: The variance-stabilized transformed version of the counts. Used for PCA, clustered heatmaps, and gene patterns.
res_list¶: A list, with one item per contrast. Each of those items in turn is a list of objects that together compose the contrast (dds name, results object, and label). This list-of-lists, which we call res_list for short, is used by functions in the lcdbwf R package for more downstream work, like gene patterns, functional enrichment, and Shiny apps.

lcdb-wf

Navigation

Related Topics

Detailed documentation of RNA-Seq downstream¶

`global_options`¶

`lcdbwf`¶

`config`¶

`libraries`¶

`coldata_setup`¶

`dds_initial`¶

`print_coldata`¶

`sample_heatmap`¶

`pca`¶

`sizefactors`¶

`dds_list`¶

`dds_diagnostics`¶

`results_*`¶

`assemble_variables`¶

Specifying contrasts¶

The most general way to specify contrasts¶

`summary`¶

`reportresults`¶

`upsetplots`¶

`excel`¶

`write_output`¶

`combined_rds`¶

`sessioninfo`¶

Glossary¶

Detailed documentation of RNA-Seq downstream¶

global_options¶

lcdbwf¶

config¶

libraries¶

coldata_setup¶

dds_initial¶

print_coldata¶

sample_heatmap¶

pca¶

sizefactors¶

dds_list¶

dds_diagnostics¶

results_*¶

assemble_variables¶

Specifying contrasts¶

The most general way to specify contrasts¶

summary¶

reportresults¶

upsetplots¶

excel¶

write_output¶

combined_rds¶

sessioninfo¶

Glossary¶

`global_options`¶

`lcdbwf`¶

`config`¶

`libraries`¶

`coldata_setup`¶

`dds_initial`¶

`print_coldata`¶

`sample_heatmap`¶

`pca`¶

`sizefactors`¶

`dds_list`¶

`dds_diagnostics`¶

`results_*`¶

`assemble_variables`¶

`summary`¶

`reportresults`¶

`upsetplots`¶

`excel`¶

`write_output`¶

`combined_rds`¶

`sessioninfo`¶