t-Distributed Stochastic Neighbor Embedding (t-SNE)

Metabolites and lipids multivariate statistical analysis in R

Practical applications of t-SNE (examples)

t-SNE is a non-linear dimensionality reduction technique designed to represent high-dimensional data (e.g., lipid or metabolite concentrations) in a lower-dimensional space (typically two or three dimensions) while preserving the similarities between neighboring data points, i.e., local structures. Its application in the field of -omics has gained increasing interest in recent years, especially in genomics and transcriptomics. t-SNE is more frequently featured in lipidomics and metabolomics studies, particularly those focused on single-cell -omics.

Here, please find examples of practical applications of this technique:

• D. Hornburg et al. Dynamic lipidome alterations associated with human health, disease and ageing. DOI: https://doi.org/10.1038/s42255-023-00880-1 - Fig. 2b (the authors utilized t-SNE to examine clustering based on the 100 most personalized lipids from 11 participants who provided at least 12 healthy samples).

• S. E. Hancock et al. FACS-assisted single-cell lipidome analysis of phosphatidylcholines and sphingomyelins in cells of different lineages. DOI: https://doi.org/10.1016/j.jlr.2023.100341 - Fig. 4A (the authors use t-SNE to examine clustering of single-cell lipidomics dataset consisting of C2C12 & HepG2 cells grown in both control (CON) and docosahexaenoic acid (DHA)-supplemented media).

• Z. Wang et al. Data-Driven Deciphering of Latent Lesions in Heterogeneous Tissue Using Function-Directed t-SNE of Mass Spectrometry Imaging Data. DOI: https://doi.org/10.1021/acs.analchem.2c02990 (t-SNE application for clustering of mass spectrometry imaging data).

• H. Tian et al. Multimodal mass spectrometry imaging identifies cell-type-specific metabolic and lipidomic variation in the mammalian liver. DOI: https://doi.org/10.1016/j.devcel.2024.01.025 (t-SNE in mass spectrometry imaging application - cell clustering based on lipid & metabolite profiles).

Required packages

The required packages for this section are uwot, Rtsne, scales, and ggrepel. These can be installed with the following command in the command window (Windows) / terminal (Mac):

# Installation of all required packages:
install.packages("Rtsne")
install.packages("scales")
install.packages("ggrepel")

# Activate libraries:
library(Rtsne)
library(scales)
library(ggrepel)

# Additionally, activate tidyverse:
library(tidyverse)

Loading data into R

Here, we will use the data set presented in the manuscript:

3MB

GOUT_CTRL_QC_Ales_data_31012025.xlsx

The lipidomics data set published by Kvasnička et al. in their manuscript Alterations in lipidome profiles distinguish early-onset hyperuricemia, gout, and the effect of urate-lowering treatment; Arthritis Research & Therapy (2023).

Always ensure you have set the appropriate working directory (wd). If you haven't done that yet, this is the first line of the code, followed by loading data into R. Read the data into R with the 'read_excel()' function from the readxl package we saw earlier in the GitBook. Set as a data.frame to make it easier to handle the data:

# Setting a working directory (wd)
setwd('D:/Data analysis')

# Loading data into R:
data <- readxl::read_excel("GOUT_CTRL_QC_Ales_data_31012025.xlsx")
data <- as.data.frame(data)
head(data)

Next, set the `Sample Name` column as row names:

# Set the row names the same as `Sample Name`
rownames(data) <- data$`Sample Name`
data$`Sample Name` <- NULL

Principal Component Analysis

Usually, before running t-SNE, PCA is performed on the high-dimensional data to reduce the dimensions (e.g., 30, 50), and then t-SNE is applied to the derived Principal Components:

The first step is to normalize the data such that the features (lipids) have zero mean and unit variance; this can easily be done with the 'scale function. By indexing the data frame with `data[,-1]`, we'll select all the data in the data frame except for the first column, which contains the labels. The data frame is re-annotated to data_normalized to make it clear that we are working with normalized data:

# Data normalization (Auto-scaling)
data_normalized <- data
data_normalized[, -1] <- scale(data[, -1])

Next, we use the PCA with the `prcomp()` function. We'll select the 50 first principal components, and we'll apply the PCA algorithm to the normalized data:

# Performing PCA:
n_components <- 50
pca_result <- prcomp(data_normalized[, -1], center = FALSE) # Note! OUR FEATURES WERE AUTO-SCALED EARLIER!
pca_features <- pca_result$x[, 1:n_components]

Performing t-SNE

Using the `Rtsne()` function from the Rtsne package, we'll apply the t-SNE algorithm on the 50 Principal Components:

# t-SNE in R
n_tsne_components <- 2
tsne_result <- Rtsne(pca_features, dims = n_tsne_components, perplexity = 30, verbose = TRUE)

The `Rtsne()` function returns the results as a matrix, so let's convert these results into a data frame for easier handling:

# Changing results output (matrix) into a data frame:
tsne_df <- data.frame(X = tsne_result$Y[,1], Y = tsne_result$Y[,2])
tsne_df$Label <- data_normalized$Label

We can now visualize the projection of the samples to the new feature space with a scatter plot:

# Obtaining t-SNE score plot:
tSNE <- ggplot(tsne_df, aes(x = X, y = Y, color = Label)) +
  geom_point(alpha = 0.7) +
  theme_minimal() +
  labs(
    title = "t-SNE Visualization",
    x = "t-SNE1", 
    y = "t-SNE2"  
  ) +
  scale_color_manual(values = c("Gout" = "red", "Control" = "blue")) +  
  theme(legend.title = element_blank()) 
 
# Exporting a high-quality publication-plot: 
# Call library:
library(ggimage)

# Score plot
## Generate a preview and optimize the plot presentation (tSNE scores plot):
ggpreview(plot = tSNE,               # The object that you want to preview.
          width = 300,               # Width in px.
          height = 300,              # Height in px.
          units = "px",              # Unit - of size - px.
          dpi = 300,                 # Sharpness.
          scale = 6)            # You may need to use a different scale.


## Save the plot in the working directory using ggsave (ggplot2 package - tidyverse):
ggsave(plot = tSNE,    # The R object to be saved.        
       device = "svg",           # Format.
       filename = "tSNE_gout_FINAL.svg",
       width = 300,
       height = 300,
       units = "px",
       dpi = 300,
       scale = 6)

We obtain the following t-SNE score plot (in gray, QC samples):

The t-SNE score plot for the selected lipidomics data set.

If one would like to obtain a score plot without QC samples, this can be done through the following block of code:

# Remove QC samples from the data set:
data_normalized_NO_QC <- 
  data_normalized %>%
  filter(Label != 'QC')

# Performing PCA:
n_components <- 50
pca_result_NO_QC <- prcomp(data_normalized_NO_QC[, -1], center = FALSE) # Note! OUR FEATURES WERE AUTO-SCALED EARLIER!
pca_features_NO_QC <- pca_result_NO_QC$x[, 1:n_components]

# t-SNE in R
n_tsne_components_NO_QC <- 2
tsne_result_NO_QC <- Rtsne(pca_features_NO_QC, dims = n_tsne_components_NO_QC, perplexity = 30, verbose = TRUE)

# Changing results output (matrix) into a data frame:
tsne_df_NO_QC <- data.frame(X = tsne_result_NO_QC$Y[,1], Y = tsne_result_NO_QC$Y[,2])
tsne_df_NO_QC$Label <- data_normalized_NO_QC$Label

# Obtaining t-SNE score plot:
tSNE_2 <- ggplot(tsne_df_NO_QC, aes(x = X, y = Y, color = Label)) +
  geom_point(alpha = 0.7) +
  theme_minimal() +
  labs(
    title = "t-SNE Visualization",
    x = "t-SNE1", 
    y = "t-SNE2"  
  ) +
  scale_color_manual(values = c("Gout" = "red", "Control" = "blue")) +  
  theme(legend.title = element_blank()) 

# Exporting high-quality score plot.
## Generate a preview and optimize the plot presentation (t-SNE score plot):
ggpreview(plot = tSNE_2,               # The object that you want to preview.
          width = 300,               # Width in px.
          height = 300,              # Height in px.
          units = "px",              # Unit - of size - px.
          dpi = 300,                 # Sharpness.
          scale = 6)            # You may need to use a different scale.


## Save the plot in the working directory using ggsave (ggplot2 package - tidyverse):
ggsave(plot = tSNE_2,    # The R object to be saved.        
       device = "svg",           # Format.
       filename = "tSNE_2_gout_FINAL.svg",
       width = 300,
       height = 300,
       units = "px",
       dpi = 300,
       scale = 6)

The updated version of the t-SNE score plot:

The updated version of the t-SNE score plot (without QC samples) for the selected lipidomics data set.