Partition: a surjective mapping approach for dimensionality reduction.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Genomic technologies generate vast, complex data with statistical and computational challenges like redundancy and noise.
• Existing dimensionality reduction (DR) methods often struggle with scalability to high dimensions or lead to significant information loss.

Purpose of the Study:

• To develop a novel, scalable DR method that mitigates information loss and noise in high-dimensional genomic data.
• To improve the detection of true associations between features and biological responses.

Main Methods:

• A new approach partitions data into feature subsets, summarizing each into a single new feature via a surjective mapping.
• Information loss constraints dictate the reduced dataset size.
• The method was evaluated using simulation studies and applied to real-world metastatic colorectal cancer gene expression data.

Main Results:

• The proposed DR method substantially increased the detection of true associations compared to principal components analysis, non-negative matrix factorization, and no DR.
• This improvement is attributed to reduced multiple-testing burdens and noise.
• Analysis of real data revealed more associations between gene expression and patient outcomes (progression-free survival, treatment response) in the reduced dataset.

Conclusions:

• The novel DR method offers a scalable and effective solution for analyzing high-dimensional genomic data.
• It enhances the discovery of true biological associations, with demonstrated utility in cancer research.
• The R package 'partition' is available for public use.