Investigating Deep Learning Based Breast Cancer Subtyping Using Pan-Cancer and Multi-Omic Data

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Breast cancer comprises multiple subtypes impacting prognosis.
• Current stratification relies on limited gene expression data.
• Next Generation Sequencing generates vast omic data.

Purpose of the Study:

• Explore machine learning, specifically deep learning, for breast cancer subtyping.
• Investigate semi-supervised learning settings due to limited public data.
• Utilize multi-omic data (microRNA, copy number alterations).

Main Methods:

• Implemented and investigated supervised and semi-supervised deep learning architectures.
• Leveraged pan-cancer and non-cancer data for semi-supervised learning.
• Combined microRNA expression and copy number alteration data.

Main Results:

• Simpler models achieved comparable accuracy to deep semi-supervised approaches on gene expression data.
• Combining multi-omic data showed minimal accuracy improvement with deep models.
• Linear models confirmed known gene-subtype associations.

Conclusions:

• Deep learning models identify novel, non-linear relationships in multi-omic data for breast cancer subtyping.
• Further analysis on larger multi-omic datasets is warranted.
• Simpler models are effective for gene expression-based subtyping.