t-SNE (t-Distributed Stochastic Neighbor Embedding)

Overview

t-SNE is a nonlinear dimensionality-reduction algorithm that maps high-dimensional data (e.g., gene expression profiles, single-cell counts) to two or three dimensions for visualization. It minimizes the Kullback-Leibler divergence between a Student-t distribution in the low-dimensional embedding and a Gaussian distribution over pairwise distances in the high-dimensional space. A perplexity hyperparameter controls the effective number of neighbors considered. t-SNE preserves local structure well but does not reliably preserve global cluster distances; results depend on the random seed, perplexity, and number of iterations.

Used by

• Used to visualize transcriptome relationships among 28 metastatic neuroendocrine neoplasms (pog570_bcgsc_2020) alongside consensus hierarchical clustering; confirmed that POG NENs cluster together and away from TCGA primary tumours in a 1,553-gene discriminator embedding PMID:24326773.
• t-SNE on the 12,454 most-variable methylation probes (s.d. > 0.25) across 740 Group 3/4 medulloblastoma cases resolved eight molecular subtypes (I–VIII), each with distinctive driver-event enrichment; subtype structure stabilised at ≥500 samples PMID:28726821

Notes

• t-SNE should not be used for quantitative distance comparisons between clusters; UMAP is a common alternative with better global structure preservation.
• Typically applied after initial PCA or top-variable-gene selection to reduce noise and compute time.
• Reference implementation: Rtsne (R) or scikit-learn/MulticoreTSNE (Python).

Sources

This page was processed by crosslinker on 2026-05-09. - PMID:28726821

This page was processed by wiki-cli on 2026-05-15.