VAECox

VAECox applies deep learning with a pre-trained variational autoencoder and transfer learning to improve survival prediction from cancer RNA-seq transcriptome data, mitigating overfitting when the number of genes exceeds available patient samples.

Key Features:

• Deep Learning Integration: Incorporates a variational autoencoder (VAE) and deep neural networks for survival prediction from RNA-seq transcriptome data.
• Transfer Learning Approach: Pre-trains the VAE on RNA-seq data from 20 TCGA datasets to capture shared pan-cancer transcriptomic features.
• Fine-Tuning Mechanism: Transfers learned VAE weights to a survival prediction model and fine-tunes those weights on individual cancer datasets.
• Performance and Robustness: Achieves a higher C-index than Cox Proportional Hazard with LASSO and ridge penalties and Cox-nnet on 7 of 10 TCGA datasets.
• Reduction of Overfitting: Mitigates overfitting in high-dimensional genomic settings (genes > samples) by leveraging shared information across cancers.

Scientific Applications:

• Survival Prediction and Risk Stratification: Improves survival prediction and risk stratification in oncology using cancer RNA-seq transcriptome data.
• Pan-cancer Feature Extraction: Extracts shared transcriptomic features from TCGA RNA-seq to inform models for specific cancer types.
• Model Benchmarking: Supports comparative evaluation against Cox Proportional Hazard models with LASSO/ridge penalties and Cox-nnet using the C-index.

Methodology:

Pre-trains a variational autoencoder (VAE) on RNA-seq from 20 TCGA datasets, transfers the learned weights to a survival prediction model, fine-tunes weights on individual cancer datasets, and evaluates performance using the C-index compared to Cox Proportional Hazard with LASSO and ridge penalties and Cox-nnet on 10 TCGA datasets.