A systematic comparison of data- and knowledge-driven approaches to disease subtype discovery

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Large-scale genomics data analysis typically uses dimensionality reduction and clustering (DR-CL).
• Biological knowledge-driven clustering (BK-CL) transforms gene expression to pathway-level data for improved disease grouping.
• BK-CL is underutilized due to a lack of systematic comparison tools and metrics favoring DR-CL methods.

Purpose of the Study:

• To develop a computational protocol for quantitative analysis of DR-CL and BK-CL clustering results.
• To propose a novel BK-CL method integrating prior disease gene knowledge, network diffusion, and gene set enrichment analysis.
• To facilitate systematic evaluation and comparison of BK-CL methods against established DR-CL approaches.

Main Methods:

• Developed a computational protocol for analyzing clustering results.
• Proposed a new BK-CL method using network diffusion and gene set enrichment.
• Conducted benchmarking studies comparing DR-CL and BK-CL on TCGA gene expression datasets.

Main Results:

• No single clustering approach dominated all evaluation metrics, highlighting the need for multi-objective evaluation.
• BK-CL methods, particularly the proposed approach, identified groupings with significant prognostic value.
• Results demonstrated BK-CL's ability to find actionable disease subtypes in breast and prostate cancers.

Conclusions:

• BK-CL offers a valuable alternative to DR-CL for analyzing large-scale genomics data.
• The proposed BK-CL method and evaluation protocol enable more robust and clinically relevant disease subgroup discovery.
• BK-CL demonstrates significant prognostic value, particularly in cancer genomics studies.