Reading in the count data
files <- set_names(files, sampleinfo$SampleName) tx2gene <- read_tsv("references/tx2gene.tsv")## Rows: 119414 Columns: 2
-## ── Column specification ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
+## ── Column specification ─────────────────────────────────────────────────────────────────────────────────────────────────────
## Delimiter: "\t"
## chr (2): TxID, GeneID
##
@@ -1615,15 +379,6 @@ Reading in the count data
## ENSMUSG00000000056 1693.000 2260.046Save the txi object for use in later sessions.
saveRDS(txi, file = "salmon_outputs/txi.rds")Exercise 1
---We have loaded in the raw counts here. These are what we need for the differential expression analysis. For other investigations we might want counts normalised to library size.
-tximportallows us to import “transcript per million” (TPM) scaled counts instead.-
-- Create a new object called
-tpmthat contains length scaled TPM counts. You will need to add an extra argument to the command. Use the help page to determine how you need to change the code:?tximport.
A quick intro to dplyr
One of the most complex aspects of learning to work with data in R is getting to grips with subsetting and manipulating data tables. The package dplyr (Wickham et al. 2018) was developed to make this process more intuitive than it is using standard base R processes.
Create a raw counts matrix for data exploration
Filtering the genes
- -For many analysis methods it is advisable to filter out as many genes as possible before the analysis to decrease the impact of multiple testing correction on false discovery rates. This is normally done by filtering out genes with low numbers of reads and thus likely to be uninformative.
-With DESeq2 this is however not necessary as it applies independent filtering during the analysis. On the other hand, some filtering for genes that are very lowly expressed does reduce the size of the data matrix, meaning that less memory is required and processing steps are carried out faster. Furthermore, for the purposes of visualization it is important to remove the genes that are not expressed in order to avoid them dominating the patterns that we observe.
We will keep all genes where the total number of reads across all samples is greater than 5.
+Many, if not most, of the genes in our annotation will not have been detected at meaningful levels in our samples - very low counts are most likely technical noise rather than biology. For the purposes of visualization it is important to remove the genes that are not expressed in order to avoid them dominating the patterns that we observe.
+The level at which you filter at this stage will not effect the differential expression analysis. The cutoff used for filtering is a balance between removing noise and keeping biologically relevant information. A common approach is to remove genes that have less than a certain number of reads across all samples. The exact level is arbitrary and will depend to some extent on nature of the dataset (overall read depth per sample, number of samples, balance of read depth between samples etc). We will keep all genes where the total number of reads across all samples is greater than 5.
# check dimension of count matrix
dim(rawCounts)## [1] 35896 12Filtering the genes
# keeping outcome in vector of 'logicals' (ie TRUE or FALSE, or NA) keep <- rowSums(rawCounts) > 5 # summary of test outcome: number of genes in each class: -table(keep, useNA="always") +table(keep, useNA = "always")## keep
## FALSE TRUE <NA>
## 15805 20091 0Count distribution and Data transformations
Raw counts
Why not raw counts? Two issues:
-
-
- Raw counts range is very large -
- Variance increases with mean gene expression, this has impact on assessing the relationships. +
- The range of values in raw counts is very large with many small values and a few genes with very large values. This can make it difficult to see patterns in the data.
summary(filtCounts)## SRR7657878 SRR7657881 SRR7657880 SRR7657874
@@ -1702,80 +454,35 @@ Raw counts
## 3rd Qu.: 1296 3rd Qu.: 1215 3rd Qu.: 1392 3rd Qu.: 1424
## Max. :722648 Max. :652247 Max. :616071 Max. :625800# few outliers affect distribution visualization
-boxplot(filtCounts, main='Raw counts', las=2)
-
+
- Variance increases with mean gene expression, this has impact on assessing the relationships, e.g. by clustering. +
# Raw counts mean expression Vs standard Deviation (SD)
plot(rowMeans(filtCounts), rowSds(filtCounts),
- main='Raw counts: sd vs mean',
- xlim=c(0,10000),
- ylim=c(0,5000))
Data transformation
-To avoid problems posed by raw counts, they can be transformed. Several transformation methods exist to limit the dependence of variance on mean gene expression:
+To avoid problems posed by raw counts, they can be transformed. A simple log2 transformation can be used to overcome the issue of the range of values. Note, when using a log transformation, it is important to add a small “pseudocount” to the data to avoid taking the log of zero.
+logCounts <- log2(filtCounts + 1)
+boxplot(logCounts, main = 'Log2 counts', las = 2)
However, this transformation does not account for the variance-mean relationship. DESeq2 provides two additional functions for transforming the data:
-
-
- Simple log2 transformation -
- VST : variance stabilizing transformation -
- rlog : regularized log transformation +
VST: variance stabilizing transformation
+rlog: regularized log transformation
log2 transformation
-Because some genes are not expressed (detected) in some samples, their count are 0. As log2(0) returns -Inf in R which triggers errors by some functions, we add 1 to every count value to create ‘pseudocounts’. The lowest value then is 1, or 0 on the log2 scale (log2(1) = 0).
# Get log2 counts
-logcounts <- log2(filtCounts + 1)We will check the distribution of read counts using a boxplot and add some colour to see if there is any difference between sample groups.
-# make a colour vector
-statusCols <- case_when(sampleinfo$Status=="Infected" ~ "red",
- sampleinfo$Status=="Uninfected" ~ "orange")
-
-# Check distributions of samples using boxplots
-boxplot(logcounts,
- xlab="",
- ylab="Log2(Counts)",
- las=2,
- col=statusCols,
- main="Log2(Counts)")
-# Let's add a blue horizontal line that corresponds to the median
-abline(h=median(logcounts), col="blue")
From the boxplots we see that overall the density distributions of raw log-counts are not identical but still not very different. If a sample is really far above or below the blue horizontal line (overall median) we may need to investigate that sample further.
-# Log2 counts standard deviation (sd) vs mean expression
-plot(rowMeans(logcounts), rowSds(logcounts),
- main='Log2 Counts: sd vs mean')
In contrast to raw counts, with log2 transformed counts lowly expressed genes show higher variation.
-VST : variance stabilizing transformation
-Variance stabilizing transformation (VST) aims at generating a matrix of values for which variance is constant across the range of mean values, especially for low mean.
-The vst function computes the fitted dispersion-mean relation, derives the transformation to apply and accounts for library size.
vst_counts <- vst(filtCounts)
-
-# Check distributions of samples using boxplots
-boxplot(vst_counts,
- xlab="",
- ylab="VST counts",
- las=2,
- col=statusCols)
-# Let's add a blue horizontal line that corresponds to the median
-abline(h=median(vst_counts), col="blue")
# VST counts standard deviation (sd) vs mean expression
-plot(rowMeans(vst_counts), rowSds(vst_counts),
- main='VST counts: sd vs mean')
Exercise 2
----
-- Use the
-DESeq2functionrlogto transform the count data. This function also normalises for library size.- Plot the count distribution boxplots with this data
-
-How has this affected the count distributions?
As well as log2 transforming the data, both transformations produce data which has been normalized with respect to library size and deal with the mean-variance relationship. The effects of the two transformations are similar. rlog is preferred when there is a large difference in library size between samples, however, it is considerably slower than VST and is not recommended for large datasets. For more information on the differences between the two transformations see the paper and the DESeq2 vignette.
Our data set is small, so we will use rlog for the transformation.
rlogcounts <- rlog(filtCounts)## converting counts to integer modeboxplot(rlogcounts, main = 'rlog counts', las = 2)
We can see from this that the infected and uninfected samples cluster separately and that day 11 and day 33 samples cluster separately for infected samples, but not for uninfected samples.