Improved survival analysis by learning shared genomic information from pan-cancer data

Area of Science:

• Biomedical informatics
• Computational biology
• Cancer research

Background:

• Deep learning shows promise in biomedical tasks but faces challenges in cancer transcriptome survival analysis.
• High dimensionality of gene expression data relative to sample size leads to overfitting in deep learning models.

Purpose of the Study:

• To introduce VAECox, a novel deep learning architecture designed to overcome overfitting in cancer survival analysis.
• To leverage transfer learning for improved survival prediction using human cancer transcriptome data.

Main Methods:

• Developed VAECox, a deep learning architecture employing transfer learning and fine-tuning.
• Pre-trained a variational autoencoder on RNA-seq data from 20 TCGA datasets.
• Transferred and fine-tuned pre-trained weights for survival prediction models on individual datasets.

Main Results:

• VAECox outperformed traditional models like Cox Proportional Hazard (with LASSO/ridge) and Cox-nnet on 7 out of 10 TCGA datasets.
• Achieved superior performance based on the C-index metric.
• Demonstrated reduced overfitting and robust performance on unseen cancer patient samples due to transferred information.

Conclusions:

• Transfer learning effectively mitigates overfitting in deep learning-based cancer survival analysis.
• VAECox offers a robust and accurate approach for predicting patient survival from transcriptome data.
• The methodology provides a valuable tool for advancing precision oncology.