Compressed Representation of Extreme Learning Machine with Self-Diffusion Graph Denoising Applied for Dissecting...

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Molecular heterogeneity is prevalent in biological systems like cancers and cell populations.
• Gene expression clustering is vital for dissecting heterogeneity but faces challenges with high dimensionality, noise, and redundancy.

Purpose of the Study:

• To introduce Extreme Learning Machine Self-Diffusion (ELMSD), a novel method for dissecting molecular heterogeneity.
• To enhance the efficiency and accuracy of clustering gene expression profiles.

Main Methods:

• ELMSD utilizes an autoencoder extreme learning machine for compressed feature representation.
• A self-diffusion graph denoising framework is integrated to improve sample-to-sample similarity.
• The method involves learning compressed representations followed by iterative graph diffusion for enhanced clustering.

Main Results:

• ELMSD was validated on simulation, single-cell, and cancer datasets.
• The approach demonstrated superior performance compared to existing state-of-the-art clustering methods.
• Identified cancer subtypes and cell types showed strong clinical relevance and biological interpretability.

Conclusions:

• ELMSD effectively addresses challenges in clustering high-dimensional, noisy gene expression data.
• The method enhances downstream clustering analysis, facilitating molecular property investigation.
• ELMSD offers a robust tool for dissecting molecular heterogeneity with significant biological and clinical implications.