Automated optimized parameters for T-distributed stochastic neighbor embedding improve visualization and analysis of large datasets

Anna C Belkina,Richard Halpert,Josef Spidlen,Christopher O Ciccolella,Rina Anno,Jennifer E Snyder-Cappione

doi:10.1038/s41467-019-13055-y

Abstract

Accurate and comprehensive extraction of information from high-dimensional single cell datasets necessitates faithful visualizations to assess biological populations. A state-of-the-art algorithm for non-linear dimension reduction, t-SNE, requires multiple heuristics and fails to produce clear representations of datasets when millions of cells are projected. We develop opt-SNE, an automated toolkit for t-SNE parameter selection that utilizes Kullback-Leibler divergence evaluation in real time to tailor the early exaggeration and overall number of gradient descent iterations in a dataset-specific manner. The precise calibration of early exaggeration together with opt-SNE adjustment of gradient descent learning rate dramatically improves computation time and enables high-quality visualization of large cytometry and transcriptomics datasets, overcoming limitations of analysis tools with hard-coded parameters that often produce poorly resolved or misleading maps of fluorescent and mass cytometry data. In summary, opt-SNE enables superior data resolution in t-SNE space and thereby more accurate data interpretation.

Highlights

Accurate and comprehensive extraction of information from high-dimensional single cell datasets necessitates faithful visualizations to assess biological populations
In order to determine the cause of the difference in cluster resolution between the “standard” and “extended” t-SNE runs, we examined the behavior of KullbackLeibler Divergence (KLD) (Kullback-Leibler divergence, see Methods) over the duration of t-SNE embeddings (Fig. 1c)
Once we found EE to be crucial for map optimization, we examined if the value of the EE factor α (EEF) can be tuned to improve the results of t-SNE

Summary

Introduction

Accurate and comprehensive extraction of information from high-dimensional single cell datasets necessitates faithful visualizations to assess biological populations. A state-of-theart algorithm for non-linear dimension reduction, t-SNE, requires multiple heuristics and fails to produce clear representations of datasets when millions of cells are projected. The precise calibration of early exaggeration together with opt-SNE adjustment of gradient descent learning rate dramatically improves computation time and enables high-quality visualization of large cytometry and transcriptomics datasets, overcoming limitations of analysis tools with hard-coded parameters that often produce poorly resolved or misleading maps of fluorescent and mass cytometry data. Multiple dimensionality reduction techniques have been applied to cytometry data with variable success Linear methods, such as PCA, are mostly unsuitable for cytometry data visualization as such techniques cannot faithfully present the non-linear relationships. Embedding (t-SNE) is a state-of-the-art dimensionality reduction algorithm for non-linear data representation that creates a lowdimensional distribution, or a ‘map’, of high-dimensional data[1,2]. This restrains tSNE’s utility for cytometry datasets that often include millions of observations (events) routinely collected for phenotypic analysis

Methods

Results

Conclusion