STAT 540 - 2024 Analysis Assignment

YOUR NAME

Reminder: When answering the questions, it is not sufficient to provide only code. Add explanatory text to interpret your results throughout.
Total: 30 POINTS

Background

In this assignment, we will explore the complexity of Acute Myeloid Leukemia (AML), an aggressive type of blood cancer, in multiomics data integration. One of the biggest outstanding challenges in AML study is that a wide variety of cancer-causing mutations has been frequently spotted in genes involved in a hematopoietic stem cell (HSC) differentiation process. Such mutations can dramatically alter developmental/differentiation trajectories, increase heterogeneity among cancer cells, and obfuscate downstream analysis.

We will use a comprehensive data set, “BeatAML2,” which contains the raw and processed gene expression data used in Bottomly et al. as well as other clinical, genetic, and drug response variables. We will investigate how multiple types of high-dimensional sequencing results can be combined in real-world settings. Here, we will focus on data wrangling, exploratory data analysis, and differential gene expression analysis, but many other scientific questions can be explored, especially linking genetic and drug response variables.

Load libraries

Note: only a few libraries are provided for the initial steps; add additional libraries as needed for the rest of the analyses.

library(tidyverse)
library(readxl)
library(edgeR)
library(limma)

We may use some handy functions over and over again.

## Avoid overwriting
if.needed <- function(.file, .code) {
    if(!all(file.exists(unlist(.file)))){ .code }
    stopifnot(all(file.exists(unlist(.file))))
}

## A quick function to concatenate two strings
`%&%` <- function(a,b) paste0(a,b)

## Feel free to define your own housekeeping functions

A primer for data management

Some helper code to read the data in is provided.

url <- "https://github.com/biodev/beataml2.0_data/raw/main/beataml_waves1to4_counts_dbgap.txt"
data.file <- "data/beataml_raw_counts.txt"
dir.create("data", recursive=T, showWarnings=F)

## Download data if needed; avoid overwriting
if.needed(data.file, download.file(url, destfile = data.file))

## Download clinical information if needed
url <- "https://github.com/biodev/beataml2.0_data/raw/main/beataml_wv1to4_clinical.xlsx"
clin.file <- "data/beataml_clinical.xlsx"
if.needed(clin.file, download.file(url, destfile = clin.file))

You might build your own SummarizedExperiment object, but we do not need to make one for this project. We will preprocess the raw data files and feed them into a DGElist object, which will be repeatedly used. We can import the count matrix stored in “data/beataml_raw_counts.txt” file, using a general purpose utility function, such as read.table or fread in data.tabble package.

count.df <-
    data.table::fread(data.file, header = T, sep = "\t") %>%
    as.data.frame()

Note: read.table needed additional tuning since some fields contain rather unconventional characters. Instead of using read.table directly, here we read the data using fread and transformed the data.table type into data.frame.

This study also comes with clinical information in Microsoft Excel format. A function, read_xlsx, gives us a convenient and programmatic way to read the necessary information. If you want to live up to reproducible data science, it is highly encouraged to avoid arbitrary/manual operations in your analysis pipeline.

.cols <- c(
    "dbgap_rnaseq_sample",           # unique sample ID
    "cohort",                        #
    "currentStage",                  #
    "consensus_sex",                 # demographic
    "vitalStatus",                   # survival information
    "overallSurvival",               #
    "plateletCount",                 # some lab tests
    "wbcCount",                      #
    "%.Basophils.in.PB",             # cell type fractions
    "%.Blasts.in.BM",                #
    "%.Blasts.in.PB",                #
    "%.Eosinophils.in.PB",           #
    "%.Immature.Granulocytes.in.PB", #
    "%.Lymphocytes.in.PB",           #
    "%.Monocytes.in.PB",             #
    "%.Neutrophils.in.PB",           #
    "%.Nucleated.RBCs.in.PB")

## Read data from .xlsx file;
## this will be a Tibble object
meta.df <-
    read_xlsx(clin.file, sheet="summary") %>%
    dplyr::select(all_of(.cols))

colnames(meta.df) <- gsub("[%][.]", "", colnames(meta.df))

In practice, many columns, which are supposed to be quantitative, may come with corrupted, non-numeric values. Looking through every bit of such columns will be laborious and may cause another type of unwanted errors/formats. Let’s deal with such painful steps programmatically. We will deal with it in a tidyverse (or dplyr) way or data.table way. First, we can clean up the annotation data by applying the take.numeric function “dplyr::across” all the columns except for the categorical ones.

take.numeric <- function(x) {
    str_extract(x, pattern="\\d+") %>% as.numeric()
}

meta.df <- meta.df %>%
    mutate(across(
        -c(dbgap_rnaseq_sample, cohort, currentStage, consensus_sex, vitalStatus),    # excluding these columns
        ~. %>% take.numeric  # apply str_extract and as.numeric to all the other columns
    ))

Side note: You may lapply the take.numeric function within the following data.table list.

.dt <- as.data.table(meta.df)

non.num.cols <- c("dbgap_rnaseq_sample",
                  "cohort",
                  "currentStage",
                  "consensus_sex",
                  "vitalStatus")

meta.clean.dt <-
    cbind(.dt[, .SD, .SDcols = non.num.cols],
          .dt[,
              lapply(.SD, take.numeric),
              .SDcols = ! non.num.cols])

Question 1: Getting familiar with the data (5 POINTS)

A. Investigate the data matrix (1 pt).

How many rows and columns do you see in the data (count.df), and what do they correspond to (genes/features or samples/individuals)?

Number of rows:

# YOUR CODE HERE

Number of columns:

# YOUR CODE HERE

B. What are the features (1 pt)?

One of the columns, biotype, indicates the biological type of each row (feature). Can you count how many types exist and how many of each type we have? If possible, sort the rows in the descending order of frequency.

HINT: You can summarize these as a table.

# YOUR CODE HERE

D. Match multiple data types (1 pt).

Subset the rows in the metadata meta.df and the columns in the count data count.df to include only those samples for which we have gene counts for downstream analysis. HINT: Keep in mind that not every sample in the data is present in the metadata table, and vice versa. This is one of the most frequently error-prone steps in data analysis.

Create a new data matrix matched.count.df by only retaining samples with matched clinical information (0.5 pt). Also, report the size of the resulting data matrix.

# YOUR CODE HERE

Also create annotation information matched.meta.dt restricting on the matched samples (0.5 pt).

# YOUR CODE HERE

E. Add additional information (cell type scores) (1 pt)

Tumour microenvironments shaped by cellular environments help understand the disease development and progression mechanisms. However, cell type fraction information is frequently missed in many samples. This study provides another data file of inferred cell type scores to compensate for the missing cell type fractions.

## Download inferred cell type scores
url <- "https://github.com/biodev/beataml2.0_data/raw/main/beataml2_manuscript_vg_cts.xlsx"
cts.file <- "data/beataml_celltypes.xlsx"
if.needed(cts.file, download.file(url, destfile = cts.file))
cts.df <- read_xlsx(cts.file)

Update the meta data (matched.meta.df) by adding cell type scores (variables ending with “-like”), matching by dbgap_rnaseq_sample column. Show the top five rows after the update (0.5 pt).

# YOUR CODE HERE

Show that there is a strong correspondence between the observe and estimated cell type fractions by displaying a correlation heatmap. Provide your own interpretations (0.5 pt). HINT: Deal with missing values with use="pairwise.complete.obs" option when computing cor matrix; use pheatmap for visualization.

# YOUR CODE HERE

ADD YOUR INTERPRETATION.

F. Validate data preprocessing steps (1 pt).

How can you check whether we have successfully matched dbgap_rnaseq_sample in meta.matched.dt with the colnames of matched.count.dt?

# YOUR CODE HERE

Add your comment.

G. Build a wrapper class object (1 pt).

Finally, build a DGEList that will be used across this project. Hint: use as.matrix and/or as.data.frame if needed. The first four columns in the count data matrix contain gene information. Think about why we matched the rows of matched.meta.df with the columns of matched.count.df data matrix.

# YOUR CODE HERE

Question 2: Gene-level quality control (5 POINTS)

Should we keep all the genes for downstream analysis? It is often recommended to remove genes that contain very little information. But how do we know which genes can be safely removed or not?

A. Compute gene-level coefficient of variation (CV) values in CPM (1 pt).

Assign “0” CV values to zero mean/variance genes.
Hint: cpm(dge) is a matrix (gene by sample).
Compute the mean $\mu$ and standard deviation $\sigma$ for each row (0.5 pt):

# YOUR CODE HERE

Compute the coefficient of variation of genes (0.5 pt).

# YOUR CODE HERE

B. Explore gene-level statistics (1 pt).

Considering that CPM values below one () are technically missing observations, compute the frequency of missing observations for each gene and investigate its relationship with the log1p-transformed coefficient of variation (CV) and the log1p-transformed mean values graphically.

Compute the missing frequency, and show its relationship with CV and $\mu$ (0.5 pt).

# YOUR CODE HERE

Why is it desired to compare the missing frequency scores against the gene-level CV values rather than variance (0.5 pt)?

ADD YOUR INTERPRETATION.

C. Justify why we choose some cutoff value (1 pt)

On top of each plot, draw vertical lines on the same plot at 75%; and discuss about why retaining genes with 25% observation rate could make sense to you.

# YOUR CODE HERE

D. Determine a data-driven decision rule (1 pt).

Can we come up with a similar decision rule in a data-driven way? Construct a simple feature matrix only including the missing fraction values (gene x 1). Perform kmeans clustering with three clusters and 100 random restarts, and show how 0.75 cutoff aligns with the clustering results.

Hint: Colour the points based on the cluster membership.

## Fix the random seed to make sure that we get the same results
set.seed(1)

Perform kmeans clustering (0.5 pt):

# YOUR CODE HERE

Demonstrate the clustering results in the same type of plots (0.5 pt).

# YOUR CODE HERE

ADD YOUR THOUGHTS.

E. Removing less-informative genes (1 pt)

Remove lowly expressed genes by retaining genes that have CPM 1 in at least 25% of samples and report how many genes are left after filter.

Apply the filtering step (0.5 pt):

# YOUR CODE HERE

Report the remaining number of genes (0.5 pt):

# YOUR CODE HERE

Question 3: Assessing overall distributions (5 POINTS)

A. TMM normalization (1 pt).

The expression values are raw counts.

Calculate TMM normalization factors (and add them to your DGEList object) (0.5 pt).

# YOUR CODE HERE

Show the distribution (density, histogram, or both) of library sizes (0.5 pt).

# YOUR CODE HERE

B. Visualizing sample-specific distributions (1 pt).

Examine the distribution of gene expression on the scale of log2 CPM on the first 20 samples using box plots (with samples on the x-axis and expression on the y-axis).

Hint 1: Add a small pseudo count of 1 before taking the log2 transformation to avoid taking log of zero - you can do this by setting log = TRUE and prior.count = 1 in the cpm function.
Hint 2: To get the data in a format amenable to plotting with ggplot, some data manipulation steps are required. Take a look at the pivot_longer function in the tidyr package to first get the data in tidy format (with one row per gene and sample combination).

# YOUR CODE HERE

C. Compare distributions across samples (1 pt).

Examine the distribution of gene expression in units of log2 CPM across all samples using overlapping density plots (with expression on the x-axis and density on the y-axis; with one line per sample and lines coloured by sample).

# YOUR CODE HERE

D. Compare distributions across samples and gene types (1 pt).

Let’s compare the distributions visually while restricting the analysis to specific gene types. We investigate the first 20 samples and partition genes into two groups, protein-coding and non-protein-coding, for simplicity. What is your interpretation?

# YOUR CODE HERE

ADD YOUR THOUGHTS.

E. Focus on protein-coding genes (1 pt)

How many protein-coding genes are left after filtering? Can you create a table that counts each type of genes after the filtering (0.5 pt)?

# YOUR CODE HERE

Since protein-coding genes constitute a majority of features and are highly expressed, let’s take yet another subset of data by only retaining protein-coding genes (0.5 pt).

# YOUR CODE HERE

Question 4: Exploratory data analysis (EDA) of the samples (5 POINTS)

A. How do the samples correlate with one another (2 pt)?

Examine the correlation between samples using a heatmap with samples on the x-axis and the y-axis and colour indicating a correlation between each pair of samples. Again, use the log2 transformed CPM values. Display other variables, such as “cohort,” “consensus_sex,” “vital status,” “overall survival,” “platelet count,” and “wbcCount” along with the correlation matrix.

HINT: You may suppress sample names in pheatmap (show_rownames=F, show_colnames=F).

# YOUR CODE HERE

B. Display the observed cell type fractions (1 pt)

Show a different set of annotation variables in the same correlation heatmap. Display the observed cell type fractions (variables ending with “in.PB”). HINT: use dplyr::ends_with for your convenience.

# YOUR CODE HERE

C. Display the estimated cell type fractions (1 pt)

Add the estimated cell type scores (variables ending with “like”) in the same heatmap. For better visualization, apply sigmoid transformation on your annotation matrix. A sigmoid function converts logit into probabilistic scores.

# YOUR CODE HERE

D. Interpretation (1 pt)

Provide your interpretation of these results.

ADD YOUR THOUGHTS.

Question 5: Find cell-type-specific genes (5 POINTS)

A. Set up differential gene expression analysis (1 pt)

First, set up a model matrix with cell type estimation scores (variables ending with “like”) as covariates. Then calculate variance weights with voom, and generate the mean-variance trend plot (1 pt). HINT: Filter out samples with “NA” values in your design matrix. Include all the cell type scores in your design matrix.

# YOUR CODE HERE

B. Fit linear models (1 pt)

Use limma (lmFit and eBayes) to fit the linear model with the model matrix you just created.

# YOUR CODE HERE

C. Top genes (1 pt)

Print the 10 top-ranked genes by adjusted p-value for the “GMP.like” coefficient using topTable.

# YOUR CODE HERE

D. Quantify the number of genes differentially expressed (1 pt)

Using the linear model defined above, determine the number of genes differentially expressed in different cell types at a FWER (use adjust.method = "holm" in topTable) less than 0.01. Count the number of DEGs for each cell type and print a table.

HINT: Count how many significantly associated genes for each coefficient. Use dplyr::count with dplyr::group_by.

# YOUR CODE HERE

E. Interpretation of the DEG results (1 pt)

As explored in the previous question (Q.4), many genes fluctuate with cell type scores. It is not surprising to see many genes are strongly associated with more than one cell types.

Count how many cell types each gene is significantly associated with (under the same FWER cutoff). Show your results by drawing histogram (0.5 pt). HINT: use group_by(stable_id).

# YOUR CODE HERE

Can you further break down the number of genes while distinguishing the sign of logFC (0.5 pt)?

# YOUR CODE HERE

Question 6: Understand cell-type-specific cancer mechanisms (5 POINTS)

A. Differential expression analysis for good and poor outcomes (2 pt)

Of the deceased patients (vitalStatus == "Dead"), we may define two patient groups according to their survival time, using the median value as a threshold value. More precisely, we have

good prognosis: a patient deceased with the survival time greater than or equal to the median;
poor prognosis: a patient deceased with the survival time less than the median.

First, create a new dge by subsetting on the deceased individuals (0.5 pt).

# YOUR CODE HERE

How many samples do have in each prognosis group (0.5 pt)?

# YOUR CODE HERE

Set up the model matrix with cell type estimation scores and the new prognosis variable. Then calculate variance weights with voom, and generate the mean-variance trend plot (1 pt). HINT: Filter out samples with “NA” values in your design matrix. Make sure that the prognosis variable is of factors.

# YOUR CODE HERE

Use limma (lmFit and eBayes) to fit the linear model with the model matrix you just created. Print the 10 top-ranked genes by adjusted p-value for the “progpos” coefficient using topTable (0.5 pt).

# YOUR CODE HERE

How many genes are differentially expressed between the poor and good prognosis groups at FDR cutoff 10% (0.5 pt)?

# YOUR CODE HERE

B. Identify cell-type-specific differentially expressed genes (2 pt)

Build a new model matrix that includes all the marginal cell-type-scores, and interaction terms between the prognosis and the cell type scores; as usual, do weight adjustment via voom (1 pt).

# YOUR CODE HERE

Use limma (lmFit and eBayes) to fit the linear model with the model matrix you just created. Print the 10 top-ranked genes by adjusted p-value for the Progenitor.like:progpoor using topTable (0.5 pt).

# YOUR CODE HERE

Report how many interaction terms are significantly associated at FDR 10%. Break down these numbers into cell types (0.5 pt).

# YOUR CODE HERE

C. Visualize top interaction DEGs (1 POINT)

Show condition-specific scatter plots for these top genes (x-axis: the cell type scores; y-axis: gene expression). Show all the significant genes and the relevant cell types. HINT: Create a long-form data frame/tibble after subsetting top genes.

# YOUR CODE HERE

Files

analysis_assignment.md

Latest commit

History

analysis_assignment.md

File metadata and controls

STAT 540 - 2024 Analysis Assignment

Background

Load libraries

A primer for data management

Question 1: Getting familiar with the data (5 POINTS)

A. Investigate the data matrix (1 pt).

B. What are the features (1 pt)?

D. Match multiple data types (1 pt).

E. Add additional information (cell type scores) (1 pt)

F. Validate data preprocessing steps (1 pt).

G. Build a wrapper class object (1 pt).

Question 2: Gene-level quality control (5 POINTS)

A. Compute gene-level coefficient of variation (CV) values in CPM (1 pt).

B. Explore gene-level statistics (1 pt).

C. Justify why we choose some cutoff value (1 pt)

D. Determine a data-driven decision rule (1 pt).

E. Removing less-informative genes (1 pt)

Question 3: Assessing overall distributions (5 POINTS)

A. TMM normalization (1 pt).

B. Visualizing sample-specific distributions (1 pt).

C. Compare distributions across samples (1 pt).

D. Compare distributions across samples and gene types (1 pt).

E. Focus on protein-coding genes (1 pt)

Question 4: Exploratory data analysis (EDA) of the samples (5 POINTS)

A. How do the samples correlate with one another (2 pt)?

B. Display the observed cell type fractions (1 pt)

C. Display the estimated cell type fractions (1 pt)

D. Interpretation (1 pt)

Question 5: Find cell-type-specific genes (5 POINTS)

A. Set up differential gene expression analysis (1 pt)

B. Fit linear models (1 pt)

C. Top genes (1 pt)

D. Quantify the number of genes differentially expressed (1 pt)

E. Interpretation of the DEG results (1 pt)

Question 6: Understand cell-type-specific cancer mechanisms (5 POINTS)

A. Differential expression analysis for good and poor outcomes (2 pt)

B. Identify cell-type-specific differentially expressed genes (2 pt)

C. Visualize top interaction DEGs (1 POINT)