PARE: A framework for removal of confounding effects from any distance-based dimension reduction method

Area of Science:

• Computational biology
• Data science
• Bioinformatics

Background:

• High-dimensional data analysis often uses dimension reduction techniques like t-SNE and UMAP.
• These methods excel at revealing complex biological patterns but struggle to isolate effects of interest from confounding variables.
• Unwanted effects, such as batch effects or technical variability, can obscure true biological signals.

Purpose of the Study:

• To introduce a novel framework, partial embedding (PARE), for removing confounding effects from dimension reduction methods.
• To develop and apply partial t-SNE and partial UMAP using the PARE framework.
• To demonstrate the utility of PARE in enhancing the visualization of biological patterns in genomic and neuroimaging data.

Main Methods:

• The partial embedding (PARE) framework was developed to enable confounder removal from any distance-based dimension reduction technique.
• Partial t-SNE and partial UMAP were implemented based on the PARE framework.
• These novel methods were applied to analyze single-cell sequencing (genomic) and neuroimaging datasets.

Main Results:

• The PARE framework successfully removed batch effects in single-cell sequencing data.
• Partial t-SNE and partial UMAP effectively separated clinical and technical variability in neuroimaging measures.
• Lower-dimensional visualizations revealed biological patterns of interest more clearly after confounder removal.

Conclusions:

• The partial embedding (PARE) framework provides a versatile extension to existing dimension reduction methods.
• PARE enables the effective removal of confounding effects, improving the interpretability of high-dimensional biological data.
• This approach enhances the ability to highlight true biological patterns in fields like genomics and neuroimaging.