A Generalization Of T Sne And Umap To Single Cell Multimodal Omics

A generalization of t-SNE and UMAP to single-cell multimodal omics

Terms

• Single-cell sequencing: Sequencing technologies that measure genomic or molecular information from individual cells.
• Dimensionality reduction: The process of representing high-dimensional data in fewer dimensions while preserving important structure.
• t-SNE: A nonlinear dimensionality reduction algorithm commonly used to visualize high-dimensional datasets.
• UMAP: A dimensionality reduction algorithm that preserves both local and global structure in data.
• Multimodal omics: The simultaneous measurement and analysis of multiple molecular data types from the same biological sample.
• Modality: A distinct type of biological measurement from a sample.
• KL divergence: (Kullback–Leibler) A measure of how one probability distribution differs from another.
• Splatter: An R package for simulating realistic single-cell RNA sequencing datasets for benchmarking and evaluating analysis methods.

Background

Better techniques are leading to better resolutions for cells and RNA dynamics. But this results in higher dimensionality and then this data needs to be translated to 2-d for visualization. This is usually done with t-SNE representations. UMAP is also commonly used for this. the paper presents j-SNE and j-UMAP which preserves similarity across modalities.

Results and Discussion

$\mathcal{E}$ - embedding of cells with multiple measured modalities.

• $\alpha$ coefficient of the convex combination
• $\lambda$ regularization pattern

Conclusion

Higher levels of noise can be counteracted with regularization coefficients $\lambda$. Their j-SNE and j-UMAP can create an embedding that shows relationships overlooked by the t-SNE and UMAP alone.