Local augmented graph neural network for multi-omics cancer prognosis prediction and analysis

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in oncology

Background:

• Cancer prognosis prediction is crucial for patient care and treatment planning.
• Multi-omics data and biological networks are increasingly used for cancer prognosis.
• Existing graph neural networks face limitations due to restricted neighboring gene information.

Purpose of the Study:

• To propose a novel local augmented graph convolutional network (LAGProg) for improved cancer prognosis prediction.
• To address the accuracy limitations of current graph neural networks in cancer prognosis.
• To enhance the representation of multi-omics features and model robustness.

Main Methods:

• Utilizing an augmented conditional variational autoencoder to generate enhanced multi-omics features.
• Integrating original and generated features into a graph convolutional neural network and Cox proportional risk model.
• Developing a two-layer graph convolutional network coupled with fully connected layers for risk prediction.

Main Results:

• LAGProg demonstrated effectiveness and efficiency across 15 TCGA datasets.
• Achieved an average 8.5% improvement in C-index values compared to state-of-the-art methods.
• Local augmentation enhanced multi-omics feature representation, robustness to missing data, and reduced over-smoothing.

Conclusions:

• LAGProg offers a significant advancement in cancer prognosis prediction accuracy.
• The local augmentation technique is key to improving model performance and reliability.
• Identified 13 prognostic markers for breast cancer, with 10 validated by existing literature.