Learning to Train and to Explain a Deep Survival Model with Large-Scale Ovarian Cancer Transcriptomic Data

Area of Science:

• Oncology
• Genomics
• Bioinformatics

Background:

• Ovarian cancer presents poor outcomes and limited therapeutic options globally.
• Novel biomarkers are crucial for stratifying patients and predicting treatment response.
• Gene expression data offers potential for developing predictive outcome models.

Purpose of the Study:

• To develop deep learning-based outcome predictors using ovarian cancer gene expression data.
• To identify molecular pathways associated with patient survival.
• To stratify ovarian cancer patients into distinct risk groups for personalized therapy.

Main Methods:

• Utilized The Cancer Genome Atlas (TCGA) ovarian cancer transcriptomic data (372 patients, ~16,600 genes).
• Trained and evaluated deep learning survival models.
• Interpreted model outputs to derive gene contributions and molecular pathways.
• Validated pathway-based stratification on in-house (12 patients) and external (274 patients) datasets.

Main Results:

• Identified molecular pathways enabling stratification of TCGA patients into high-risk and low-risk groups (p=0.025).
• Validated stratification efficacy on an in-house dataset (p=0.229) and an external dataset (p=0.006).
• Demonstrated interpretability of deep learning models for uncovering survival-associated biological processes.

Conclusions:

• Deep learning models analyzing RNA-seq data can predict survival in ovarian cancer patients.
• These models effectively detect and interpret gene sets linked to survival outcomes.
• This approach offers a new avenue for biomarker discovery and personalized treatment in ovarian cancer.