SUPREME: multiomics data integration using graph convolutional networks.

Area of Science:

• Computational biology
• Bioinformatics
• Cancer research

Background:

• Precision medicine requires grouping cancer patients by biological similarity into distinct subtypes.
• High-dimensional multiomics data necessitates integrative approaches for accurate cancer subtyping.
• Graph Neural Networks (GNNs) offer advanced methods for learning from graph-structured data, but existing tools have limitations.

Purpose of the Study:

• To develop an advanced node classification framework, SUPREME, for integrating multiple data modalities.
• To address limitations in existing integrative prediction tools for cancer subtyping.
• To improve the accuracy and biological relevance of cancer subtype identification.

Main Methods:

• Developed SUPREME, a node classification framework integrating multiple data modalities on graph-structured data.
• SUPREME generates patient embeddings from multiple similarity networks using multiomics features.
• Integrated embeddings with raw features to capture complementary signals for enhanced subtyping.

Main Results:

• SUPREME outperformed existing tools in breast cancer subtype prediction across three datasets.
• SUPREME-inferred subtypes demonstrated significant survival differences, often exceeding those of the ground truth.
• The framework showed superior performance compared to nine other approaches and demonstrated model-agnostic applicability on additional datasets.

Conclusions:

• SUPREME effectively utilizes multiomics data to uncover novel cancer subtype characteristics linked to survival differences.
• The framework has the potential to refine existing cancer subtype labels, which are often based on single data types.
• SUPREME advances the development of precision medicine by enabling more accurate and biologically meaningful cancer subtyping.