Feature-preserving manifold approximation and projection to analyze single-cell data.

Area of Science:

• Computational biology
• Single-cell genomics
• Bioinformatics

Background:

• Single-cell data analysis requires methods to understand cellular heterogeneity and dynamics.
• Current manifold learning techniques like UMAP and t-SNE excel at visualizing cell clusters but often lose gene-level information.
• Preserving gene expression patterns is crucial for deeper biological insights.

Purpose of the Study:

• To introduce FeatureMAP, a novel framework for feature-preserving manifold approximation and projection.
• To enhance the visualization of single-cell data by retaining both clustering structures and gene-level information.
• To enable new analyses of gene contribution, variation trajectories, and cell states.

Main Methods:

• FeatureMAP integrates Uniform Manifold Approximation and Projection (UMAP) with Principal Component Analysis (PCA).
• It employs pairwise tangent space embedding to preserve feature variation.
• Key analytic concepts include gene contribution, gene variation trajectory, and core/transition states, derived from topological properties.

Main Results:

• FeatureMAP successfully retains clustering structures and gene variation in a low-dimensional representation.
• The framework enables the identification of core and transition cell states using topological features.
• Differential Gene Variation (DGV) analysis highlights regulatory genes driving cell state transitions, as shown in pancreatic development and T-cell exhaustion data.

Conclusions:

• FeatureMAP offers an improved approach to single-cell data visualization and analysis.
• It facilitates the discovery of regulatory genes critical for understanding cell state dynamics.
• The method enhances the analysis of developmental trajectories and cellular responses.