Capturing discrete latent structures: choose LDs over PCs.

Area of Science:

• Bioinformatics
• Computational Biology
• Data Science

Background:

• High-dimensional biological data is common, requiring effective dimensionality reduction.
• Principal Component Analysis (PCA) assumes classes align with maximum variance, failing when this assumption is unmet.
• Visualization techniques like t-SNE and UMAP are computationally expensive and may not preserve discriminatory information.

Purpose of the Study:

• To develop a dimensionality reduction technique that effectively captures discrete latent structures in high-dimensional data.
• To overcome the limitations of PCA and visualization methods in identifying underlying data structures.
• To enable post hoc analysis for identifying features that define latent structures.

Main Methods:

• Proposed iterative Discriminant Analysis (iDA), a novel dimensionality reduction technique.
• iDA utilizes linear transformations to optimally separate latent clusters.
• The method facilitates post hoc analysis to identify defining features of latent structures.

Main Results:

• iDA produces embeddings that carry discriminatory information.
• The technique effectively separates latent clusters using linear transformations.
• Enables identification of features contributing to latent structure separation.

Conclusions:

• iDA offers an effective alternative for dimensionality reduction in high-dimensional biological data.
• The method addresses limitations of PCA and non-linear visualization techniques.
• iDA's linear transformations allow for interpretable feature identification crucial for biological insights.