Visualization of SNPs with t-SNE

Abstract

Background: Single Nucleotide Polymorphisms (SNPs) are one of the largest sources of new data in biology. In most papers, SNPs between individuals are visualized with Principal Component Analysis (PCA), an older method for this purpose.

Principal findings: We compare PCA, an aging method for this purpose, with a newer method, t-Distributed Stochastic Neighbor Embedding (t-SNE) for the visualization of large SNP datasets. We also propose a set of key figures for evaluating these visualizations; in all of these t-SNE performs better.

Significance: To transform data PCA remains a reasonably good method, but for visualization it should be replaced by a method from the subfield of dimension reduction. To evaluate the performance of visualization, we propose key figures of cross-validation with machine learning methods, as well as indices of cluster validity.

Figure 1. SNP data transformed with PCA and t-SNE 1/2.

On the left is a PCA-plot with the first two components, on the right a t-SNE-plot of the very same data from each data source. Data sources: Panel (a) is from the 1001 genomes project, (b) from the RegMap panel and (c) from hapmap3 r2.

Figure 2. SNP data transformed with PCA and t-SNE 2/2.

On the left is a PCA-plot with the first two components, on the right a t-SNE-plot of the very same data from each data source. Data sources: Panel (a) from hapmap3 r3 (compare with Fig. 1c) and (b) from the Rice Haplotype Map Project (only wild type where the label information was available).