Deep learning-based cancer survival prognosis from RNA-seq data: approaches and evaluations

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Oncology

Background:

• Deep Learning (DL) models, initially for image processing, are now advancing medical research, particularly in cancer patient survival prognosis.
• DL-based Cox proportional hazards models utilize transcriptomic data for predicting cancer survival outcomes.

Purpose of the Study:

• To conduct a comprehensive analysis of DL-based survival prognosis models across various TCGA (The Cancer Genome Atlas) cancers.
• To evaluate and compare the performance of established DL models (Cox-nnet, DeepSurv) and a novel method (AECOX).

Main Methods:

• Utilized transcriptomic data from TCGA cancers.
• Applied and evaluated Deep Learning models: Cox-nnet, DeepSurv, and AECOX (AutoEncoder with Cox regression network).
• Assessed model performance using concordance index and log-rank test p-values.

Main Results:

• All evaluated DL models demonstrated competitive performance across 12 cancer types.
• Lower-dimensional representations from DL models' hidden layers facilitate feature reduction and visualization.
• Prognosis accuracy showed a negative correlation with overall survival and tumor mutation burden (TMB), indicating a link between survival, TMB, and prediction accuracy.

Conclusions:

• DL algorithms outperform traditional machine learning models in cancer prognosis.
• While model performance (concordance index) was comparable across DL methods, it varied significantly among different cancer types.
• Findings illuminate the interplay between patient characteristics, survival data, and predictive model learnability on a pan-cancer scale.