Robust correlation estimation and UMAP assisted topological analysis of omics data for disease subtyping

Area of Science:

• Computational biology and bioinformatics
• Genomics and precision medicine
• Cancer research and subtyping

Background:

• Identifying disease subtypes from gene expression data is crucial for precision medicine.
• High dimensionality and sparsity of omics data challenge conventional clustering algorithms.
• Existing methods struggle to discover clinically relevant and statistically significant patient subgroups.

Purpose of the Study:

• To develop a robust computational pipeline for discovering clinically relevant disease subtypes.
• To identify patient subgroups with distinct survival patterns for improved prognostic accuracy.
• To enhance clustering results for omics data by addressing dimensionality and sparsity issues.

Main Methods:

• Utilized a pipeline combining robust correlation estimation, Uniform Manifold Approximation and Projection (UMAP), and Mapper for non-linear dimensionality reduction.
• Developed a novel method to improve the robustness of gene expression correlation matrices for better clustering.
• Applied the pipeline to five cancer datasets from The Cancer Genome Atlas (TCGA) for validation.

Main Results:

• The proposed method significantly improved the separation of survival curves between discovered patient subgroups compared to state-of-the-art methods (e.g., NEMO, RSC-OTRI, SNF).
• Demonstrated a substantial increase in survival curve separability, exemplified by a 221-day separation in the Glioblastoma Multiforme (GBM) dataset.
• Showcased that robust correlation estimation enhances subgroup separability, even when using single omics profiles.

Conclusions:

• The developed pipeline offers a robust approach for disease subtyping with improved clinical relevance and prognostic value.
• This methodology outperforms existing techniques in separating survival curves, highlighting its potential for precision oncology.
• Pathway over-representation analysis confirmed distinct biological underpinnings for each identified subtype.