Gaussian mixture copulas for high-dimensional clustering and dependency-based subtyping.

Area of Science:

• Computational biology
• Statistical genetics
• Bioinformatics

Background:

• Patient subtyping is crucial for precision medicine, enabling tailored therapies to reduce mortality.
• Model-based clustering, like Gaussian mixture models, is widely used but limited by assumptions of identical marginal distributions.
• Existing methods struggle with high-dimensional data and complex dependencies, hindering accurate subtype identification.

Purpose of the Study:

• To develop a novel, scalable, and interpretable method for high-dimensional patient subtyping.
• To address the limitations of current clustering techniques in modeling complex dependencies and non-Gaussian data.
• To improve the accuracy and clinical relevance of patient subtyping for precision medicine.

Main Methods:

• Utilized the statistical framework of copulas to decouple marginal distribution modeling from dependency modeling.
• Developed HD-GMCM (High-Dimensional Gaussian Mixture Copula Models), the first copula-based clustering method designed for high-dimensional data.
• Employed Gaussian mixture copulas for robustness to outliers and modeling of non-Gaussian data.

Main Results:

• HD-GMCM successfully fits high-dimensional data, outperforming state-of-the-art model-based clustering methods.
• Demonstrated improved clustering performance on real-world high-dimensional gene-expression and clinical datasets.
• Identified clinically meaningful patient subtypes in lung cancer data (TCGA) based on interpretable dependencies and survival rates.

Conclusions:

• HD-GMCM offers a significant advancement in clustering for high-dimensional data, particularly in biomedical applications.
• The method's ability to model complex dependencies provides a new avenue for characterizing patient subtypes.
• HD-GMCM facilitates more accurate patient subtyping, leading to better-tailored therapies and improved clinical outcomes in precision medicine.