Interactive visualization of metric distortion in nonlinear data embeddings using the distortions package

Area of Science:

• Computational Biology
• Bioinformatics
• Data Visualization

Background:

• Nonlinear dimensionality reduction (NLDR) methods, such as Uniform Manifold Approximation and Projection (UMAP) and t-distributed stochastic neighbor embedding (t-SNE), are crucial for organizing high-dimensional genomics data.
• These methods reduce data complexity into lower dimensions for easier interpretation of biological structures like cell types or differentiation trajectories.
• However, NLDR techniques inevitably introduce distortions, which can lead to misinterpretations, such as inferring non-existent clusters in the original data.

Purpose of the Study:

• To address the challenges posed by distortions in NLDR for genomics data analysis.
• To develop and implement a software package named "distortions" for measuring and visualizing local distortions.
• To enhance the confidence and accuracy of researchers when using NLDR methods for biological data interpretation.

Main Methods:

• Implementation of a software package, "distortions", utilizing state-of-the-art algorithms for quantifying local distortions in dimensionality reduction.
• Development of intuitive and interactive visualization tools to display these measured distortions.
• Application of the package to case studies using both simulated and real-world genomics datasets.

Main Results:

• The "distortions" package effectively visualizes local distortions introduced by NLDR methods.
• Visualizations generated by the package can identify fragmented neighborhoods within the reduced data representations.
• The tool aids in optimizing hyperparameters for NLDR methods and facilitates informed method selection.

Conclusions:

• The "distortions" software package provides a valuable layer of information to assess the reliability of NLDR outputs in genomics.
• By highlighting data distortions, the package empowers researchers to use NLDR methods like UMAP and t-SNE with greater confidence.
• This approach mitigates the risk of drawing erroneous conclusions from the analysis of high-dimensional biological data.