A prognostic prediction model for ovarian cancer using a cross-modal view correlation discovery network

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Ovarian cancer presents diverse clinicopathological and molecular features, often with advanced spread at diagnosis.
• Early diagnosis and prognostic prediction are crucial for understanding pathogenesis and improving therapeutic outcomes.
• Multi-omics data in ovarian cancer exhibit heterogeneity, posing challenges for existing integration methods.

Purpose of the Study:

• To propose a novel deep learning model, MDCADON, for integrating multi-omics data to predict ovarian cancer prognosis.
• To address the limitations of existing methods in handling variability and inter-correlation within multi-omics data.
• To enhance survival analysis and guide treatment strategies for ovarian cancer patients.

Main Methods:

• Feature selection using random forest and LASSO regression on mRNA expression, DNA methylation, miRNA expression, and copy number variation (CNV).
• A multi-modal deep neural network for learning feature representations from omics and clinical data.
• A cross-modal view correlation discovery network to construct a multi-omics discovery tensor for exploring inter-relationships.

Main Results:

• MDCADON demonstrated superior performance compared to existing methods in predicting ovarian cancer prognosis.
• The model enables accurate survival analysis, aiding in the determination of patient follow-up treatment plans.
• Gene Ontology (GO) term and pathway analyses identified key genes and revealed underlying ovarian cancer mechanisms.

Conclusions:

• MDCADON offers a powerful new approach for ovarian cancer prognosis prediction by effectively integrating multi-omics data.
• The model's ability to explore inter-omics correlations provides insights into disease mechanisms.
• Findings support improved clinical decision-making and therapeutic guidance for ovarian cancer.