Gaining Biological Insights through Supervised Data Visualization.

Area of Science:

• Computational Biology
• Data Visualization
• Machine Learning

Background:

• Unsupervised dimensionality reduction methods like PHATE, t-SNE, and UMAP often fail to capture biologically relevant structures tied to expert labels.
• Integrating expert knowledge is crucial for deeper insights into complex biological datasets.

Approach:

• Introduced RF-PHATE, a supervised dimensionality reduction technique leveraging random forests to capture feature-label relationships.
• RF-PHATE generates low-dimensional visualizations by extracting information from random forests, prioritizing relevant features and disregarding noise.
• The method is scalable to large datasets and applicable to both classification and regression tasks.

Key Points:

• Demonstrated RF-PHATE's utility in identifying a non-benign relapsing-remitting Multiple Sclerosis subgroup from longitudinal clinical and imaging data.
• Showcased RF-PHATE's proficiency in noisy environments by analyzing Raman spectral data to visualize antioxidant effects on lung cells.
• Applied RF-PHATE to COVID-19 patient data, aligning geometric structures with outcomes for hierarchical interpretability.

Conclusions:

• RF-PHATE effectively bridges expert insights with data visualization, facilitating knowledge generation.
• The method's adaptability, scalability, and noise tolerance suggest broad applicability in biological data analysis.