REFINED-CNN framework for survival prediction with high-dimensional features.

Area of Science:

• Genomics
• Machine Learning
• Pharmacogenomics
• Bioinformatics

Background:

• Accurate survival prediction in clinical trials using high-throughput genomics data is crucial but challenging.
• Current machine learning models often lack predictive performance and interpretability for this task.

Purpose of the Study:

• To extend the REFINED-CNN model for survival prediction using RNA sequencing data.
• To improve predictive performance and model interpretability in clinical trial survival analysis.

Main Methods:

• Mapping high-dimensional RNA sequencing data into REFINED images for Convolutional Neural Network (CNN) modeling.
• Utilizing transfer learning for adapting the REFINED-CNN survival model to new cancer types with limited patient data.
• Employing risk score backpropagation for local and global feature (gene) importance quantification.

Main Results:

• The REFINED-CNN survival model demonstrates robust and accurate predictions.
• The model shows effective adaptation to new tasks and cancer types via transfer learning.
• Risk score backpropagation provides interpretable insights into feature importance for survival prediction.

Conclusions:

• REFINED-CNN offers a powerful and interpretable approach for survival prediction in pharmacogenomics.
• The model's ability to leverage transfer learning and provide feature importance enhances its clinical utility.
• This method advances the application of deep learning in analyzing genomics data for clinical trial outcomes.