Performance Comparison of Deep Learning Autoencoders for Cancer Subtype Detection Using Multi-Omics Data.

Area of Science:

• Genomics
• Computational Biology
• Cancer Research

Background:

• Cancer is a heterogeneous disease driven by multiple pathways, leading to varied patient survival and drug responses.
• Accurate cancer subtyping is crucial for personalized medicine but often requires integrating multi-omics data.

Purpose of the Study:

• To compare the performance of different deep learning autoencoder models for cancer subtype detection.
• To evaluate the utility of autoencoders in identifying biologically relevant cancer subtypes and predicting survival outcomes.

Main Methods:

• Utilized four different autoencoder implementations for subtype detection on The Cancer Genome Atlas (TCGA) datasets across four cancer types.
• Employed the silhouette score to determine the optimal number of subtypes and compared feature selection and similarity measures.
• Analyzed differentially expressed genes within identified subtypes using the Glioblastoma multiforme (GBM) dataset.

Main Results:

• Detected cancer subtypes exhibiting significant differences in survival profiles across multiple cancer types.
• Identified differentially expressed genes in Glioblastoma multiforme subtypes, consistent with existing genomic studies.
• Demonstrated that autoencoder-based multi-omics data fusion effectively predicts patient subgroups and survival.

Conclusions:

• Deep learning autoencoders are effective tools for cancer subtype detection and characterization using multi-omics data.
• The identified subtypes and their associated genomic features provide insights into cancer biology and patient prognosis.
• This approach facilitates the prediction of patient subgroups and survival outcomes, supporting precision oncology.