Partition_R

Partition_R performs dimensionality reduction of high-dimensional genomic datasets by partitioning correlated features into subsets and summarizing each subset with a surjective mapping to preserve a user-specified minimum level of information.

Key Features:

• Direct-Measure-Reduce: Implements the Direct-Measure-Reduce approach for partitioning and summarizing features in dimensionality reduction.
• Agglomerative partitioning framework: Uses agglomerative partitioning to group related variables into subsets.
• Surjective Mapping: Ensures every original variable maps to exactly one reduced variable to preserve interpretability and limit information loss.
• Customizable Framework: Allows specification of which variables to reduce, how to measure information loss, and how to aggregate subset features.
• Scalable Solution: Scales to large-dimensional datasets while controlling information loss relative to many traditional DR methods.
• Noise and multiple-testing mitigation: Reduces redundancy and extraneous noise and mitigates multiple-testing challenges to enhance detection of true associations.

Scientific Applications:

• Simulation studies: Demonstrated increases in the number of true associations detected when multiple related features associate with a response, compared to principal components analysis and non-negative matrix factorization.
• Metastatic colorectal cancer gene expression analysis: Applied to metastatic colorectal cancer tumors, it linked more gene expression features to progression-free survival and treatment response than analyses of the full untransformed data.

Methodology:

Partitions the dataset into subsets of related features and reduces each subset to a single new feature via a surjective mapping under constraints that limit information loss within an agglomerative Direct-Measure-Reduce framework.