In martinenge/RNAseqFunctions: RNA sequencing functions for the Enge lab

Introduction

library(RNAseqFunctions)
library(tidyverse)

The vignette is designed to show the steps for running t-SNE and the associated plotting functions available in the package. The "counts" dataset included with the package will be used for demonstration purposes.

First we prepare the dataset by normalizing the counts to counts per million.

counts.cpm <- cpm(pro.counts)

Feature selection

It is a reasonable expectation that the majority of genes expressed in a two cell types will not display a distinct expression pattern. Instead, it will only be a subset of all the expressed genes that are cell type specific. In addition, many genes are correlated in their expression patterns. Therefore, when distinguishing cell types or cell states from transcriptional profiles, it is not useful to include many genes that exhibit the same expression profile.

To reduce the gene expression space we use feature (gene) selection. This can be done in many different ways. Within the package there are functions that support three different types of feature selection:

1. Select features based on max expression.
2. Select features based on variance.
3. Select features based on modelled coefficient of variation.

We typically use feature selection based on max expression upstream of t-SNE since it has shown reasonable performance accross a variety of datasets. All of the feature selection functions take a counts per million matrix and a number of features argument as input and return the index of the selected features.

selected_features <- nTopMax(counts.cpm, 2000)

We can see some of the features selected using the following code:

head(rownames(counts.cpm)[selected_features])

We can extract the features from the counts matrix in the following way:

selected <- counts.cpm[selected_features, ]

We can see that the feature selected dataset is now only includes 2000 features/genes and the origional 81 samples.

dim(selected)

Sample distance metrics

t-SNE can be run on a matrix of counts or counts per million, although, it is typically faster and gives better performance if is is provided with a distance metric between samples. For this metric we typically use 1 - Pearson's correlation which can be calculated in using the following function:

p.dist <- pearsonsCor(selected)

p.dist is a matrix (specifically, a lower triangle matrix) that describes the 1 - Pearsons correlation between each of the samples.

t-SNE

We now provide the distance matrix to the t-SNE algorithm to calculate the sample representation in t-SNE space.

tsne <- runTsne(p.dist, perplexity = 2)

The results include the sample names (rownames) and the placment of each sample along the x (column 1) and y (column 2) axis in t-SNE space.

Plotting t-SNE results

There are multiple plotting options for the t-SNE output. First, we can just plot the t-SNE results without the addition of any other variables to view the separations.

plotTsne(tsne, log2cpm(counts.cpm))

We can also add marker gene expression onto the t-SNE plot to see which population(s) of cells express a specific marker.

plotTsne(tsne, log2cpm(counts.cpm), "CD74")

We can also provide additional marker genes (typically one per cell type) to visualize their expression.

plotTsne(tsne, log2cpm(counts.cpm), c("CD74", "ANXA3", "ACTG2"))

Finally, it may be desired to get back the data used for plotting in order to modify or customize a plot. This can be achieved using the following function:

p <- plotTsne(tsne, log2cpm(counts.cpm), c("CD74", "ANXA3", "ACTG2"))
plotData(p)

martinenge/RNAseqFunctions documentation built on May 28, 2019, 3:10 p.m.