MCluster-VAEs: An end-to-end variational deep learning-based clustering method for subtype discovery using...

Area of Science:

• Computational biology
• Genomics
• Machine learning in oncology

Background:

• Unsupervised clustering of cancer subtypes aids diagnosis, treatment, and prognosis.
• Multi-omics data offers a comprehensive view but presents challenges due to heterogeneity and noise.

Purpose of the Study:

• To develop a deep learning method for extracting cluster-friendly representations from multi-omics data.
• To improve the accuracy and robustness of cancer subtype discovery.

Main Methods:

• Proposed Multi-omics Clustering Variational Autoencoders (MCluster-VAEs), an end-to-end deep learning model.
• Utilized a unified network architecture with an attention mechanism for multi-omics data modeling.
• Employed a novel objective function based on Variational Bayes for clustering assignment estimation.

Main Results:

• MCluster-VAEs demonstrated superior performance compared to 12 state-of-the-art methods on TCGA benchmark datasets.
• Achieved an adjusted Rand index of ~0.78 for cancer category recognition on the Pan Cancer dataset, an 18% improvement.
• Survival analysis and clinical enrichment tests showed comparable or better results than common integrative approaches.

Conclusions:

• MCluster-VAEs is a powerful tool for dissecting complex multi-omics relationships.
• Provides new insights for cancer subtype discovery, potentially improving patient outcomes.
• Highlights the effectiveness of deep learning in integrating and analyzing heterogeneous multi-omics data.