Introduction to snifter
In snifter: R wrapper for the python openTSNE library

knitr::opts_chunk$set(
    error = FALSE,
    warning=FALSE,
    message=FALSE,
    collapse = TRUE,
    comment = "#>"
)
library("BiocStyle")

Introduction

snifter provides an R wrapper for the openTSNE implementation of fast interpolated t-SNE (FI-tSNE). It is based on r Biocpkg("basilisk") and r CRANpkg("reticulate"). This vignette aims to provide a brief overview of typical use when applied to scRNAseq data, but it does not provide a comprehensive guide to the available options in the package.

It is highly advisable to review the documentation in snifter and the openTSNE documentation to gain a full understanding of the available options.

Setting up the data

We will illustrate the use of snifter using data from r Biocpkg("scRNAseq") and single cell utility functions provided by r Biocpkg("scuttle"), r Biocpkg("scater") and r Biocpkg("scran") - first we load these libraries and set a random seed to ensure the t-SNE visualisation is reproducible (note: it is good practice to ensure that a t-SNE embedding is robust by running the algorithm multiple times).

library("snifter")
library("scRNAseq")
library("scran")
library("scuttle")
library("scater")
library("ggplot2")
theme_set(theme_bw())
set.seed(42)

Before running t-SNE, we first load data generated by Zeisel et al. from r Biocpkg("scRNAseq"). We filter this data to remove genes expressed only in a small number of cells, estimate normalisation factors using r Biocpkg("scran") and generate 20 principal components. We will use these principal components to generate the t-SNE embedding later.

data <- ZeiselBrainData()
data <- data[rowMeans(counts(data) != 0) > 0.05, ]
data <- computeSumFactors(data, cluster = quickCluster(data))
data <- logNormCounts(data)
data <- runPCA(data, ncomponents = 20)
## Convert this to a factor to use as colouring variable later
data$level1class <- factor(data$level1class)

Running t-SNE

The main functionality of the package lies in the fitsne function. This function returns a matrix of t-SNE co-ordinates. In this case, we pass in the 20 principal components computed based on the log-normalised counts. We colour points based on the discrete cell types identified by the authors.

mat <- reducedDim(data)
fit <- fitsne(mat, random_state = 42L)
ggplot() +
    aes(fit[, 1], fit[, 2], colour = data$level1class) +
    geom_point(pch = 19) +
    scale_colour_discrete(name = "Cell type") +
    labs(x = "t-SNE 1", y = "t-SNE 2")

Projecting new data into an existing embedding

The openTNSE package, and by extension snifter, also allows the embedding of new data into an existing t-SNE embedding. Here, we will split the data into "training" and "test" sets. Following this, we generate a t-SNE embedding using the training data, and project the test data into this embedding.

test_ind <- sample(nrow(mat), nrow(mat) / 2)
train_ind <- setdiff(seq_len(nrow(mat)), test_ind)
train_mat <- mat[train_ind, ]
test_mat <- mat[test_ind, ]

train_label <- data$level1class[train_ind]
test_label <- data$level1class[test_ind]

embedding <- fitsne(train_mat, random_state = 42L)

Once we have generated the embedding, we can now project the unseen test data into this t-SNE embedding.

new_coords <- project(embedding, new = test_mat, old = train_mat)
ggplot() +
    geom_point(
        aes(embedding[, 1], embedding[, 2],
            colour = train_label,
            shape = "Train"
        )
    ) +
    geom_point(
        aes(new_coords[, 1], new_coords[, 2], 
            colour = test_label,
            shape = "Test"
        )
    ) +
    scale_colour_discrete(name = "Cell type") +
    scale_shape_discrete(name = NULL) +
    labs(x = "t-SNE 1", y = "t-SNE 2")