Hoffman_human_bulk_analysis.Rmd

---
title: "Analysis of the Hoffman et al. human bulk RNA-seq data"
author: "Tobias Tekath"
date: "`r Sys.Date()`"
description: "Exemplified analysis of human bulk RNA-seq data."
output: 
  rmarkdown::html_vignette:
    df_print: paged
    toc: true
    toc_depth: 3
vignette: >
  %\VignetteIndexEntry{Analysis of the Hoffman et al. human bulk RNA-seq data}
  %\VignetteBuilder{knitr}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r setup, include = FALSE}
root_dir <- "/data/hoffman_bulk/dtu_results"
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  cache = FALSE, 
  eval = dir.exists(root_dir)
)
if(dir.exists(root_dir)){
  knitr::opts_knit$set(root.dir = root_dir)
  #remove previous results
  unlink(paste0(root_dir,"/*"),recursive=TRUE)
}
start_time <- Sys.time()
```


This vignette exemplifies the analysis of **bulk RNA-seq data with DTUrtle**. The data used in this vignette is publicly available as *Bioproject PRJNA594939* and the used *FASTQ*-files can be downloaded from [here](https://www.ebi.ac.uk/ena/browser/view/PRJNA594939). The corresponding publication from Hoffman et al. can be found [here](https://doi.org/10.1038/s42003-020-0837-0).

---

The following code shows an example of an DTUrtle workflow. Assume we have performed the pre-processing as described [here](https://tobitekath.github.io/DTUrtle/articles/Hoffman_human_bulk_preprocess.html) and the R working directory is a newly created folder called `dtu_results`.

---

## Setup

Load the DTUrtle package and set the BiocParallel parameter. It is recommended to perform computations in parallel, if possible. 

```{r message=TRUE, warning=TRUE}
library(DTUrtle)
#use up to 10 cores for computation
biocpar <- BiocParallel::MulticoreParam(10)
```

---

## Import and format data

We want to start by reading in our quantification counts, as well as a file specifying which transcript ID or name belongs to which gene ID or name.

### Importing and processing GTF annotation (tx2gene)

To get this transcript to gene (`tx2gene`) mapping, we will utilize the already present Gencode annotation file `gencode.v34.annotation.gtf`. The `import_gtf()` function utilizes the a `rtracklayer` package and returns a transcript-level filtered version of the available data.


```{r}
tx2gene <- import_gtf(gtf_file = "../gencode.v34.annotation.gtf")
```

```{r}
head(tx2gene, n=3)
```


There are a lot of columns present in the data frame, but at the moment we are mainly interested in the columns `gene_id`, `gene_name`, `transcript_id` and `transcript_name`. 

---

As we want to use gene and transcript names as the main identifiers in our analysis (so we can directly say: Gene x is differential), we should ensure that each gene / transcript name maps only to a single gene / transcript id.

For this we can use the DTUrtle function `one_to_one_mapping()`, which checks if there are identifiers, which relate to the same name. If this is the case, the names (**not** the identifiers) are slightly altered by appending a number. If *id_x* and *id_y* both have the name *ABC*, the *id_y* name is altered to *ABC_2* by default.  

```{r}
tx2gene$gene_name <- one_to_one_mapping(name = tx2gene$gene_name, id = tx2gene$gene_id)
tx2gene$transcript_name <- one_to_one_mapping(name = tx2gene$transcript_name, id = tx2gene$transcript_id)
```

We see that it was a good idea to ensure the one to one mapping, as many doublets have been corrected.

---

For the `run_drimseq()` `tx2gene` parameter, we need a data frame, where the first column specifies the transcript identifiers and the second column specifying the corresponding gene names. Rather than subsetting the data frame, a column reordering is proposed, so that additional data can still be used in further steps. DTUrtle makes sure to carry over additional data columns in the analysis steps. To reorder the columns of our tx2gene data frame, we can utilize the `move_columns_to_front()` functionality.

```{r}
tx2gene <- move_columns_to_front(df = tx2gene, columns = c("transcript_name", "gene_name"))
```

```{r}
head(tx2gene, n=5)
```


This concludes the tx2gene formatting.

---

### Reading in quantification data

The read-in of the quantification counts can be achieved with `import_counts()`, which uses the `tximport` package in the background. This function is able to parse the output of many different quantification tools. Advanced users might be able to tune parameters to parse arbitrary output files from currently not supported tools.

In the pre-processing vignette we quantified the counts with `Salmon`. The folder structure of the quantification results folder looks like this:

```{r}
list.files("../salmon/")
```

We will create a named files vector, pointing to the `quant.sf` file for each sample. The names help to differentiate the samples later on.

```{r}
files <- Sys.glob("../salmon/bulk_*/quant.sf")
names(files) <- gsub(".*/","",gsub("/quant.sf","",files))
```

The files object looks like this:
```{r, echo=FALSE}
files
```

The actual import will be performed with `import_counts()`. As we are analyzing bulk RNA-seq data, the raw counts should be scaled regarding transcript length and/or library size prior to analysis. DTUrtle will default to appropriate scaling schemes for your data. It is advised to provide a transcript to gene mapping for the `tx2gene` parameter, as then the `tximport` `dtuScaledTPM` scaling scheme can be applied. This schemes is especially designed for DTU analysis and scales by using the median transcript length among isoforms of a gene, and then the library size. 

The raw counts from `Salmon` are named in regard to the `transcript_id`, thus we will provide an appropriate tx2gene parameter here. Downstream we want to use the `transcript_name` rather than the `transcript_id`, so we change the row names of the counts:

```{r}
cts <- import_counts(files, type = "salmon", tx2gene=tx2gene[,c("transcript_id", "gene_name")])
rownames(cts) <- tx2gene$transcript_name[match(rownames(cts), tx2gene$transcript_id)]
```


```{r}
dim(cts)
```

There are ~227k features left for the six samples. Please note, that in contrast to the single-cell workflow utilizing `combine_to_matrix()`, these counts also include features with no expression. These will be filtered out by the DTUrtle filtering step in `run_drimseq()`.

---

### Sample metadata

Finally, we need a sample metadata data frame, specifying which sample belongs to which comparison group. This table is also convenient to store and carry over additional metadata.

If such a table is not already present, it can be easily prepared:

```{r}
pd <- data.frame("id"=colnames(cts), 
                 "group"=c(rep("Dex2hr",3), rep("EtOH",3)), 
                 stringsAsFactors = FALSE)
```

```{r}
head(pd, n=5)
```

---

## DTU analysis

We have prepared all necessary data to perform the differentially transcript usage (DTU) analysis. DTUrtle only needs two simple commands to perform it. Please be aware that these steps are the most compute intensive and, depending on your data, might take some time to complete. It is recommended to parallelize the computations with the `BBPARAM` parameter, if applicable.

First, we want to set-up and perform the statistical analysis with DRIMSeq, a DTU specialized statistical framework utilizing a Dirichlet-multinomial model. This can be done with the `run_drimseq()` command. We use the previously imported data as parameters, specifying which column in the cell metadata data frame contains ids and which the group information we want. We should also specify which of the groups should be compared (if there are more than two) and in which order. The order given in the `cond_levels` parameter also specifies the comparison formula.

```{r}
dturtle <- run_drimseq(counts = cts, tx2gene = tx2gene, pd=pd, id_col = "id",
                    cond_col = "group", cond_levels = c("Dex2hr", "EtOH"), filtering_strategy = "bulk", 
                    BPPARAM = biocpar)
```

As in all statistical procedures, it is of favor to perform as few tests as possible but as much tests as necessary, to maintain a high statistical power. This is achieved by filtering the data to remove inherently uninteresting items, for example very lowly expressed genes or features. DTUrtle includes a powerful and customizable filtering functionality for this task, which is an optimized version of the `dmFilter()` function of the DRIMSeq package. 

Above we used a predefined filtering strategy for bulk data, requiring that features contribute at least 5% of the total expression in at least 50% of the cells of the smallest group. Also the total gene expression must be 5 or more for at least 50% of the samples of the smallest group.
Additionally, all genes are filtered, which only have a single transcript left, as they can not be analyzed in DTU analysis. The filtering options can be extended or altered by the user.

```{r}
dturtle$used_filtering_options
```

The resulting `dturtle` object will be used as our main results object, storing all necessary and interesting data of the analysis. It is a simple and easy-accessible list, which can be easily extended / altered by the user. By default three different meta data tables are generated:

 - `meta_table_gene`: Contains gene level meta data.
 - `meta_table_tx`: Contains transcript level meta data.
 - `meta_table_sample`: Contains sample level meta data (as the created `pd` data frame)
 
These meta data tables are used in for visualization purposes and can be extended by the user.

```{r}
dturtle$meta_table_gene[1:5,1:5]
```


---

As proposed in [Love et al. (2018)](https://doi.org/10.12688/f1000research.15398.3), we will use a two-stage statistical testing procedure together with a post-hoc filtering on the standard deviations in proportions (`posthoc_and_stager()`). We will use *stageR* to determine genes, that show a overall significant change in transcript proportions. For these significant genes, we will try to pinpoint specific transcripts, which significantly drive this overall change. As a result, we will have a list of `significant genes` (genes showing the overall change) and a list of `significant transcripts` (one or more transcripts of the `significant genes`). Please note, that not every `significant gene` does have one or more `significant transcripts`. It is not always possible to attribute the overall change in proportions to single transcripts. These two list of significant items are computed and corrected against a **overall false discovery rate (OFDR)**.

Additionally, we will apply a post-hoc filtering scheme to improve the targeted OFDR control level. The filtering strategy will discard transcripts, which standard deviation of the proportion per cell/sample is below the specified threshold. For example by setting `posthoc=0.1`, we will exclude all transcripts, which proportional expression (in regard to the total gene expression) deviates by less than 0.1 between cells.
This filtering step should mostly exclude 'uninteresting' transcripts, which would not have been called as significant either way. 

```{r}
dturtle <- posthoc_and_stager(dturtle = dturtle, ofdr = 0.05, posthoc = 0.1)
```

The `dturtle` object now contains additional elements, including the lists of `significant genes` and `significant transcripts`.

```{r}
head(dturtle$sig_gene)
head(dturtle$sig_tx)
```

---

## (optional) DGE analysis

Alongside a DTU analysis, a DGE analysis might be of interest for your research question. DTUrtle offers a basic DGE calling workflow for bulk and single-cell RNA-seq data via DESeq2.

To utilize this workflow, we should re-scale our imported counts. The transcript-level count matrix has been scaled for the DTU analysis, but for a DGE analysis un-normalized counts should be used (as DESeq2 normalizes internally). We can simply re-import the count data, by providing the already defined `files` object to the DGE analysis specific function `import_dge_counts()`. This function will make sure that the counts are imported but not scaled and also summarizes to gene-level.

Our files object looks like this:

```{r}
head(files)
```

We re-import the files with the same parameters as before:

```{r}
cts_dge <- import_dge_counts(files, type="salmon", tx2gene=tx2gene[,c("transcript_id", "gene_name")])
```

With this data, we can perform the DGE analysis with DESeq2 with the DTUrtle function `run_deseq2()`. DESeq2 is one of the gold-standard tools for DGE calling in bulk RNA-seq data and showed very good performance for single-cell data in multiple benchmarks. The DESeq2 vignette recommends adjusting some parameters for DGE analysis of single-cell data - DTUrtle incorporates these recommendations and adjusts the specific parameters based on the provided `dge_calling_strategy()`.

After the DESeq2 DGE analysis, the estimated log2 fold changes are shrunken (preferably with `apeglm`) - to provide a secondary ranking variable beside the statistical significance. If a shrinkage with `apeglm` is performed, `run_deseq2()` defaults to also compute s-values rather than adjusted p-values. The underlying hypothesis for s-values and (standard) p-values differs slightly, with s-values hypothesis being presumably preferential in a real biological context. Far more information about all these topics can be found in the excellent [DESeq2 vignette](https://bioconductor.org/packages/release/bioc/vignettes/DESeq2/inst/doc/DESeq2.html) and the associated publications.

`run_deseq2()` will tell the user, if one or more advised packages are missing. It is strongly advised to follow these recommendations.

```{r}
dturtle$dge_analysis <- run_deseq2(counts = cts_dge, pd = pd, id_col = "id",
                                   cond_col = "group", cond_levels = c("Dex2hr", "EtOH"), 
                                   lfc_threshold = 1, sig_threshold = 0.01, dge_calling_strategy = "bulk",
                                   BPPARAM = biocpar)
```

We provided a log2 fold change threshold of 1 (on log2 scale), thus preferring an effect size of 2x or more.
The output of `run_deseq2()` is a list, containing various elements (see `run_deseq2()`'s description). This result list can easily added to an existing DTUrtle object, as shown above.

We can now identify genes which show both a DGE and DTU signal:

```{r}
dtu_dge_genes <- intersect(dturtle$sig_gene, dturtle$dge_analysis$results_sig$gene)
length(dtu_dge_genes)
```

---

## Result aggregation and visualization

The `DTUrtle` package contains multiple visualization options, enabling an in-depth inspection.

### DTU table creation

We will start by aggregating the analysis results to a data frame with `create_dtu_table()`. This function is highly flexible and allows aggregation of gene or transcript level metadata in various ways. By default some useful information are included in the *dtu table*, in this example we further specify to include the `seqnames` column of the gene level metadata (which contains chromosome information) as well as the maximal *expressed in* ratio from the transcript level metadata. 

```{r}
dturtle <- create_dtu_table(dturtle = dturtle, add_gene_metadata = list("chromosome"="seqnames"), 
                               add_tx_metadata = list("tx_expr_in_max" = c("exp_in", max)))
```


```{r}
head(dturtle$dtu_table, n=5)
```

The column definitions are as follows:

* "gene_ID": Gene name or identifier used for the analysis.
* "gene_qvalue": Multiple testing corrected p-value (a.k.a. q-value) comparing all transcripts together between the two groups ("gene level").
* "minimal_tx_qvalue": The minimal multiple testing corrected p-value from comparing all transcripts  individually between the two groups ("transcript level"). I.e. the q-value of the most significant transcript.
* "number_tx": The number of analyzed transcripts for the specific gene.
* "number_significant_tx": The number of significant transcripts from the 'transcript level' analysis.
* "max(Dex2hr-EtOH)": Maximal proportional difference between the two groups (Dex2hr vs EtOH). E.g. one transcript of 'SMOX' is ~94% more expressed in 'Dex2hr' cells compared to 'EtOH' cells.
* "chromosome": the chromosome the gene resides on.
* "tx_expr_in_max": The fraction of cells, the most expressed transcript is expressed in. "Expressed in" is defined as expression>0. 

This table is our basis for creating an **interactive HTML-table** of the results.

---

### Proportion barplot

As a first visualization option we will create a barplot of the proportions of each transcript per sample. We can use the function `plot_proportion_barplot()` for this task, which also adds the mean proportion fit per subgroup to the plot (by default as a red line).

As an example, we will create the plot for the gene *TSC22D3* (a Glucocorticoid-Induced Leucine Zipper), which is one of the significant genes found in the analysis. Additionally, it is identified in [Hoffman et al.](https://doi.org/10.1038/s42003-020-0837-0) as one of the Dexamethasone / Glucocorticoid Receptor target genes, which also shows differential expression already after 2 hours of Dex treatment.

We will optionally provide the gene_id for *TSC22D3*, which is stored in the column `gene_id.1` of `dturtle$meta_table_gene`.


```{r fig.height=6, fig.width=8, fig.align = "center", out.width="80%"}
temp <- plot_proportion_barplot(dturtle = dturtle, genes = "TSC22D3", meta_gene_id = "gene_id.1")
temp$TSC22D3
```

We see, that there are striking proportional differences for the three transcripts of *TSC22D3*. Due to *Dexamethasone* treatment, there seems to be a shift to almost exclusive expression of *TSC22D3-206*. This transcript has also been identified as a significant transcript in the analysis (as the name is marked in red). From the top annotation we can also see, that the mean total expression of *TSC22D3* is higher in the `Dex2hr` group (as expected) - but also the coefficient of variation (CV) is higher, indicating a higher variance between the samples.

For the interactive HTML-table we would need to save the images to disk (in the to-be-created sub folder "images" of the current working directory). There is also a convenience option, to directly add the file paths to the `dtu_table`. As multiple plots are created, we can provide a `BiocParallel` object to speed up the creation. If no specific genes are provided, all significant genes will be plotted.

```{r}
dturtle <- plot_proportion_barplot(dturtle = dturtle, 
                                   meta_gene_id = "gene_id.1", 
                                   savepath = "images", 
                                   add_to_table = "barplot",
                                   BPPARAM = biocpar)
```

```{r}
head(dturtle$dtu_table$barplot)
head(list.files("./images/"))
```


---

### Proportion heatmap

A different visualization option is a heatmap, where additional meta data can be displayed alongside the transcript proportions (`plot_proportion_pheatmap()`). This visualization uses the `pheatmap` package, the user can specify any of the available parameters to customize the results.


```{r fig.height=6, fig.width=8, fig.align = "center", out.width="80%"}
temp <- plot_proportion_pheatmap(dturtle = dturtle, genes = "TSC22D3", 
                                 include_expression = TRUE, treeheight_col=20)
temp$TSC22D3
```

By default, row and column annotations are added. This plot helps to examine the transcript composition of groups of cells. We can see a quite uniform expression pattern between the two conditions. From the column annotation we can see, that the *TSC22D3* total expression in the sample *bulk_D2hr_rep1* is quite lower than in the rest.

Again, we can save the plots to disk and add them to the `dtu_table`:

```{r}
dturtle <- plot_proportion_pheatmap(dturtle = dturtle,
                                    include_expression = TRUE,
                                    treeheight_col=20,
                                    savepath = "images", 
                                    add_to_table = "pheatmap",
                                    BPPARAM = biocpar)
```

```{r}
head(dturtle$dtu_table$pheatmap)
head(list.files("./images/"))
```

---

### Transcript overview

Until now, we looked at the different transcripts as abstract entities. Alongside proportional differences, the actual difference in the exon-intron structure of transcripts is of great importance for many research questions. This structure can be visualized with the `plot_transcripts_view()` functionality of *DTUrtle*.

This visualization is based on the *Gviz* package and needs a path to a *GTF* file (or a read-in object). In [Import and format data](#import-and-format-data) we already imported a GTF file. This was subset to transcript-level (via the `import_gtf()` function), thus this is **not** sufficient for the visualization. We can reuse the actual *GTF* file though, which should in general match with the one used for the `tx2gene` data frame.

As we have ensured the one_to_one mapping in [Import and format data](#import-and-format-data) and potentially renamed some genes, we should specify the `one_to_one` parameter in this call. 

```{r fig.height=6, fig.width=8, fig.align = "center", out.width="80%"}
plot_transcripts_view(dturtle = dturtle, 
                              genes = "TSC22D3", 
                              gtf = "../gencode.v34.annotation.gtf", 
                              genome = 'hg38',
                              one_to_one = TRUE)
```



This visualization shows the structure of the transcripts of *TSC22D3*. The significant transcript (*TSC22D3-206*) is quite different to the other mainly expressed transcript (*TSC22D3-203*). *TSC22D3-206* has a different first exon than *TSC22D3-203*, whose first exon is quite far away from the conserved transcript body. The arrows on the right side indicate the mean fitted proportional change in the comparison groups, thus showing a over-expression of *TSC22D3-206* in `Dex2hr` compared to `EtOH`.

The areas between exons indicate intron sequences, which have been compressed in this representation to highlight the exon structure. Only consensus introns are compressed to a defined minimal size. This can be turned off with `reduce_introns=FALSE` or alternatively reduced introns can be highlighted by setting a colorful `reduce_introns_fill`.

Analogous as before, we can save plots to disk and add them to the `dtu_table`:


```{r}
dturtle <- plot_transcripts_view(dturtle = dturtle, 
                                 gtf = "../gencode.v34.annotation.gtf", 
                                 genome = 'hg38', 
                                 one_to_one = TRUE,
                                 savepath = "images", 
                                 add_to_table = "transcript_view",
                                 BPPARAM = biocpar)
```

```{r}
head(dturtle$dtu_table$transcript_view)
head(list.files("./images/"))
```

---

### Visualize DTU table

The `dturtle$dtu_table` is now ready to be visualized as an interactive HTML-table. Please note, that it is optional to add any plots or additional columns to the table. Thus the visualization will work directly after calling `create_dtu_table()`.

The `dtu_table` object looks like this:

```{r}
head(dturtle$dtu_table)
```


Before creating the actual table, we can optionally define column formatter functions, which color the specified columns. The coloring might help with to quickly dissect the results.

*DTUrtle* come with some pre-defined column formatter functions (for p-values and percentages), other formatter functions from the `formattable` package can also be used. Advanced users might also define their own functions.

We create a named list, linking column names to formatter functions:

```{r}
column_formatter_list <- list(
      "gene_qvalue" = table_pval_tile("white", "orange", digits = 3),
      "minimal_tx_qvalue" = table_pval_tile("white", "orange", digits = 3),
      "number_tx" = formattable::color_tile('white', "lightblue"),
      "number_significant_tx" = formattable::color_tile('white', "lightblue"),
      "max(Dex2hr-EtOH)" = table_percentage_bar('lightgreen', "#FF9999", digits=2),
      "tx_expr_in_max" = table_percentage_bar('white', "lightblue", color_break = 0, digits=2))

```

This `column_formatter_list` is subsequently provided to `plot_dtu_table()`:

```{r}
plot_dtu_table(dturtle = dturtle, savepath = "my_results.html", 
               column_formatters = column_formatter_list)
```


> <div class="alert alert-info">
>
> <strong>Note:</strong> ️As seen above, the paths to the plots are relative. Please make sure that the saving directory in `plot_dtu_table()` is correctly set and the plots are reachable from that directory with the given path.
>
> <strong>The links in the following example are just for demonstration purposes and do not work!</strong> 
>
> </div>


<style>
.vscroll-plot {
    overflow: auto;
    resize: both;
}
</style>

<div class="vscroll-plot">
```{r, echo=FALSE}
plot_dtu_table(dturtle = dturtle, column_formatters = column_formatter_list, min_page_length=10)
```
</div>

---

For later use we can save the final DTUrtle object to disk:

```{r}
saveRDS(dturtle, "./dturtle_res.RDS")
```

---


## Session info

Computation time for this vignette:

```{r, echo=FALSE}
Sys.time()-start_time
```

---

```{r sessionInfo, echo=FALSE}
sessionInfo()
```