Multiscale Hyperbolic Embedding for Cell Hierarchies in Large-Scale Bioinformatics Data.

Area of Science:

• Computational Biology
• Data Visualization
• Machine Learning

Background:

• Large datasets present challenges for visualization and interpretation.
• Hyperbolic embedding excels at capturing hierarchical structures but struggles with scalability.
• Existing methods lack efficient scaling for massive datasets.

Purpose of the Study:

• Introduce MuH-MDS, a multiscale hyperbolic multidimensional scaling algorithm.
• Address limitations of fixed curvature and scalability in current hyperbolic embedding techniques.
• Enhance analysis of large-scale biological datasets.

Main Methods:

• Developed MuH-MDS, a novel multiscale hyperbolic multidimensional scaling algorithm.
• Utilized "adiabatic" approximation from physics for optimizing local positions.
• Implemented a method capable of handling datasets with over 80,000 samples.

Main Results:

• MuH-MDS demonstrates a 10^3 improvement in computing time compared to previous methods.
• Successfully analyzed a large-scale C. elegans embryogenesis scRNA-seq dataset (>80,000 samples).
• Uncovered intrinsic hierarchical structure, improving pseudotime inference and lineage analysis over UMAP and other methods.

Conclusions:

• MuH-MDS offers a scalable and computationally efficient solution for hyperbolic embedding.
• Preserves global hierarchy and metric accuracy, unlike methods focusing solely on local structure.
• Provides superior performance for large-scale biological data analysis, particularly in lineage and pseudotime inference.