A multi-omics supervised autoencoder for pan-cancer clinical outcome endpoints prediction

Area of Science:

• Computational biology
• Genomics
• Cancer research

Background:

• Advancements in sequencing technologies enable cost-effective collection of diverse cancer omics data.
• Existing multi-omics fusion methods often overlook the heterogeneity of different omics types, limiting the excavation of biological knowledge.
• Current approaches for predicting clinical outcomes using omics data may not fully leverage sample similarity or domain knowledge.

Purpose of the Study:

• To develop a method that learns effective omics-specific representations and fuses them to improve cancer outcome prediction.
• To address the limitations of existing methods in handling heterogeneous omics data and extracting biological insights.

Main Methods:

• Proposed MOSAE (Multi-omics Supervised Autoencoder), an autoencoder-based approach for multi-omics data fusion.
• Designed omics-specific autoencoders to generate representations tailored to the dimension of each omics type.
• Integrated a supervised autoencoder framework using clinical labels to learn both omics-specific and task-specific representations.
• Fused the learned representations from different omics for subsequent predictive tasks.

Main Results:

• Applied MOSAE to the TCGA Pan-Cancer dataset for predicting overall survival (OS), progression-free interval (PFI), disease-free interval (DFI), and cancer-specific death (DSS).
• MOSAE demonstrated superior predictive performance compared to traditional and state-of-the-art methods.
• Evaluated the contribution of each component of MOSAE, confirming positive impacts on predictive performance.

Conclusions:

• Multi-omics fusion is crucial for advancing precision and personalized medicine through accurate clinical outcome prediction.
• MOSAE provides a powerful framework for multi-omics fusion, generating omics-specific and task-specific representations.
• The method effectively improves predictive performance for clinical endpoint prediction tasks.