Visualizing single-cell data with the neighbor embedding spectrum.

Area of Science:

• Computational Biology
• Machine Learning
• Data Visualization

Background:

• t-distributed Stochastic Neighbor Embedding (t-SNE) and Uniform Manifold Approximation and Projection (UMAP) are widely used for visualizing high-dimensional single-cell data.
• These methods, despite different formulations, are theoretically linked.
• A single parameter can interpolate between t-SNE and UMAP, revealing a spectrum of visualization techniques.

Approach:

• Investigated the theoretical connection between t-SNE and UMAP using machine learning principles.
• Introduced a parameter to create a spectrum of visualization methods interpolating between t-SNE and UMAP.
• Proposed visualizing this spectrum as an animation for enhanced data exploration.

Key Points:

• The spectrum shifts focus from local structures (t-SNE-like) to continuous structures (UMAP-like).
• This spectrum presents a trade-off in single-cell analysis: highlighting rare cell types versus continuous variations like developmental trajectories.
• Animation of the spectrum offers a more comprehensive understanding than static t-SNE or UMAP plots.

Conclusions:

• A unified view of t-SNE and UMAP through a parameter spectrum enhances single-cell data visualization.
• This approach provides flexibility to emphasize different data aspects, aiding biological interpretation.
• Animated visualization of the spectrum offers deeper insights into complex high-dimensional datasets.