A novel non-negative Bayesian stacking modeling method for Cancer survival prediction using high-dimensional omics...

Area of Science:

• Genomics
• Precision Medicine
• Computational Biology

Background:

• High-dimensional molecular data is crucial for cancer survival prediction.
• Carcinogenesis involves pathway-based pathogenesis, necessitating group-structured models for accurate prognosis.
• Existing single-model methods often lack robustness in prediction.

Purpose of the Study:

• To develop a novel survival stacking method integrating biological group information for robust cancer survival prediction.
• To enhance prediction robustness using high-dimensional omics data by leveraging pathway structures.
• To improve upon existing single-model prediction methods in genomics.

Main Methods:

• Introduced a survival stacking method using super learner to combine predictions from pathway-grouped biological features.
• Extended the super learner to non-negative Bayesian hierarchical generalized linear models and artificial neural networks.
• Compared the proposed method with Lasso Cox and group Lasso Cox using simulations and real-world data.

Main Results:

• The survival stacking method demonstrated superior and robust discrimination performance compared to single models on noisy simulated and real-world data.
• The non-negative Bayesian stacking approach successfully identified significant biological pathways and genes linked to cancer prognosis.
• The method proved effective in handling high-dimensional omics data for survival prediction.

Conclusions:

• A novel survival stacking strategy incorporating biological group information was proposed for cancer prognosis.
• The super learner was extended to Bayesian and artificial neural network models, enhancing sub-model combinations.
• The Bayesian stacking strategy offers favorable prediction and interpretation for complex survival data, aiding cancer target discovery.