Deep Learning data integration for better risk stratification models of bladder cancer

Area of Science:

• Oncology
• Bioinformatics
• Genomics

Background:

• Bladder cancer (BC) prognosis remains challenging due to patient heterogeneity.
• Accurate prediction of survival subtypes is crucial for personalized treatment strategies.

Purpose of the Study:

• To develop and validate an unsupervised multi-omics integration pipeline for predicting bladder cancer survival subtypes.
• To identify molecular features associated with different survival outcomes in BC.

Main Methods:

• Utilized a deep learning autoencoder algorithm for unsupervised multi-omics integration (mRNA, miRNA, methylation) from TCGA dataset.
• Developed a supervised classification model to predict survival subgroups.
• Validated the model's robustness using cross-validation and an independent validation dataset.

Main Results:

• Identified two distinct survival subtypes in bladder cancer with significant survival differences (p=8e-4).
• The high-risk subgroup showed enrichment of KRT6/14 overexpression and PI3K-Akt pathway activation.
• The pipeline demonstrated robust performance on training, testing, and validation datasets (p=0.03 and p=0.02, respectively).

Conclusions:

• The proposed multi-omics integration pipeline is an effective tool for inferring bladder cancer survival subtypes.
• This approach can aid in stratifying patients and guiding clinical decision-making for bladder cancer.