Gaining biological insights through supervised data visualization.

Area of Science:

• Bioinformatics
• Data Visualization
• Machine Learning

Background:

• Unsupervised dimensionality reduction methods like t-SNE, UMAP, and Isomap often fail to capture biologically relevant structures aligned with specific analytical goals or expert annotations.
• Existing supervised visualization techniques have limitations in addressing this mismatch effectively.

Purpose of the Study:

• To introduce RF-PHATE, a novel supervised visualization approach designed to integrate expert knowledge for revealing label-relevant data structures.
• To enhance the interpretability of complex biological data by suppressing extraneous variation.

Main Methods:

• RF-PHATE employs random forests to learn feature-label relationships, translating this information into informative low-dimensional embeddings.
• The method is designed to handle large datasets and is applicable to both classification and regression tasks.

Main Results:

• Demonstrated utility across diverse case studies, including longitudinal multiple sclerosis data, Raman spectroscopy, COVID-19 patient outcomes, and RNA sequencing data.
• RF-PHATE successfully enhanced data interpretability, managed noise, and exposed meaningful biological structures.

Conclusions:

• RF-PHATE offers a powerful supervised approach for biological data exploration, effectively leveraging expert knowledge.
• The method shows broad potential for improving data analysis, discovery, and understanding across various biological domains.