H-tSNE

H-tSNE integrates hierarchical supervised label information into nonlinear dimensionality reduction to modify the manifold of methods such as t-SNE, UMAP, and PHATE and preserve hierarchical (parent-child) relationships among data points.

Key Features:

• Hierarchical Label Integration: Incorporates hierarchical supervised label information to emphasize parent-child relationships within datasets during dimensionality reduction.
• Variable Perturbation: Applies a variable perturbation of the high-dimensional space to adjust embeddings while maintaining core visualization characteristics.
• Preservation of Visualization Characteristics: Modifies the manifold without discarding the intrinsic properties of the underlying methods (t-SNE, UMAP, PHATE).
• Cross-Domain Applicability: Extends to multiple nonlinear dimensionality reduction methods and is applicable across diverse analytical domains.
• Semi-Supervised Learning Compatibility: Leverages partial label information when only a subset of data points are labeled to guide embedding structure.
• User-Defined Hierarchical Distancing Factor: Uses a user-specified hierarchical distancing factor to control the strength of hierarchical adjustments to the manifold.

Scientific Applications:

• Single-cell RNA sequencing analysis: Assists interpretation of single-cell RNA sequencing datasets by embedding hierarchical label relationships among cells.
• Cellular differentiation studies: Highlights parent-child lineage relationships relevant to cellular differentiation and developmental trajectories.
• Exploratory analysis of complex datasets: Enhances discovery of hierarchical patterns and structured class relationships in high-dimensional biological data.

Methodology:

Implements a variable perturbation of the high-dimensional space combined with a user-defined hierarchical distancing factor to adjust the manifold of existing nonlinear dimensionality reduction methods (t-SNE, UMAP, PHATE) so embeddings respect hierarchical supervised labels.