Cancer subtype identification by multi-omics clustering based on interpretable feature and latent subspace learning

Area of Science:

• Computational Biology
• Bioinformatics
• Cancer Research

Background:

• Multi-omics clustering is crucial for understanding cancer heterogeneity.
• Existing methods often struggle with noisy or redundant multi-omics data, leading to suboptimal cancer subtyping.
• Direct integration of heterogeneous omics features can limit clustering performance.

Purpose of the Study:

• To develop a novel multi-omics clustering method for improved cancer subtyping.
• To extract interpretable and discriminative features from multi-omics data prior to integration.
• To leverage clinical information via SHAP values for supervised feature extraction.

Main Methods:

• Feature extraction using SHAP (SHapley Additive exPlanations) values guided by clinical information.
• Integration of extracted multi-omics features using Singular Value Decomposition (SVD) into a latent subspace.
• Clustering using shared nearest neighbor-based spectral clustering on the integrated latent representation.

Main Results:

• The proposed method demonstrates superior performance in multi-omics cancer subtyping compared to state-of-the-art approaches.
• Utilizing clinical information with SHAP values significantly enhances clustering analysis performance.
• Enrichment analysis of identified gene signatures validates the biological relevance of the discovered cancer subtypes.

Conclusions:

• The novel multi-omics clustering approach effectively addresses limitations of existing methods for cancer subtyping.
• SHAP-based feature extraction using clinical data is a powerful strategy for improving multi-omics data analysis.
• The method provides a robust framework for identifying biologically meaningful cancer subtypes from integrated omics data.