Cancer Subtyping via Embedded Unsupervised Learning on Transcriptomics Data.

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Cancer remains a leading cause of global mortality, necessitating accurate subtype classification for effective treatment.
• Deep learning methods show promise for automatic cancer subtyping but often suffer from data overfitting due to high dimensionality and scarcity.
• Existing unsupervised methods frequently rely on limiting Gaussianity assumptions, which are inadequate for small sample sizes.

Purpose of the Study:

• To investigate automatic cancer subtyping using an unsupervised learning approach.
• To address the overfitting issue in deep learning-based cancer subtyping systems.
• To develop a method that captures latent space features and models molecular cancer subtypes without strong distributional assumptions.

Main Methods:

• Proposed an unsupervised learning framework for cancer subtyping.
• Employed vector quantization to bypass the Gaussianity assumption common in small-sample unsupervised learning.
• Focused on directly constructing the underlying data distribution to generate sufficient data and alleviate overfitting.

Main Results:

• The proposed method effectively captures latent space features relevant to cancer subtypes.
• Demonstrated improved modeling of cancer subtype manifestation on a molecular basis.
• Experimental results confirmed the method's efficacy in addressing overfitting and enhancing subtyping accuracy.

Conclusions:

• Unsupervised learning, by constructing data distributions, offers a viable solution to overfitting in deep learning-based cancer subtyping.
• Vector quantization provides a robust alternative to Gaussianity assumptions, enabling better molecular subtyping from limited data.
• The developed method enhances the potential for accurate and reliable automatic cancer subtyping, crucial for clinical applications.