Autoencoder techniques for survival analysis on renal cell carcinoma

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Survival outcomes are critical for evaluating cancer therapies.
• Identifying molecular predictors of survival is a key research area.
• High-dimensional patient data complicates survival analysis.

Purpose of the Study:

• To compress high-dimensional transcriptomic data using autoencoders for survival prediction.
• To apply statistical methods for predicting Progression-Free Survival (PFS).
• To enhance the interpretability of autoencoder models in oncology.

Main Methods:

• Utilized autoencoders to create latent features from transcriptomic data.
• Applied COX Proportional Hazards model with Breslow's estimator for survival analysis.
• Incorporated tabular and graph-based (protein-protein interactions) data.
• Analyzed mutual information for gene-feature association.

Main Results:

• Denoising autoencoders improved data reconstruction.
• Sparse autoencoders generated more meaningful latent representations.
• Combined penalties enhanced reconstruction and interpretability.
• Identified LRP2 and ACE2 as relevant genes for renal cell carcinoma.

Conclusions:

• Autoencoders are effective for managing high-dimensional data in oncology.
• Different autoencoder types offer distinct advantages for specific tasks.
• Combining autoencoder penalties improves model performance and interpretability.
• This approach aids in identifying molecular predictors of survival and therapeutic targets.