Learning vector quantized representation for cancer subtypes identification.

Area of Science:

• Computational biology
• Genomics
• Cancer research

Background:

• Accurate cancer subtyping is crucial for personalized therapy and prognosis.
• Omics data, like transcriptomics, are vital but present challenges such as sample scarcity and high dimensionality.
• Existing methods often struggle with these data issues, leading to overfitting and unreliable feature extraction.

Purpose of the Study:

• To develop a novel approach for cancer subtyping that addresses the limitations of omics data.
• To improve the accuracy and robustness of cancer subtyping for better patient outcomes.
• To extract meaningful discrete representations from complex cancer data for enhanced clustering.

Main Methods:

• Leveraging a Vector-Quantized Variational AutoEncoder (VQ-VAE), a powerful generative model.
• Extracting discrete latent representations from omics data that retain essential information for reconstruction.
• Applying these representations to improve subsequent clustering of cancer subtypes.

Main Results:

• Demonstrated significant and robust improvement in patient prognosis across multiple cancer datasets.
• Validated the effectiveness on 10 distinct cancer types, showcasing broad applicability.
• The proposed method outperformed prevalent subtyping systems in clinical relevance.

Conclusions:

• The VQ-VAE approach offers a flexible method that does not impose strict data distribution assumptions.
• The extracted latent features provide superior representations of transcriptomic data for different cancer subtypes.
• This leads to enhanced clustering performance applicable with various mainstream clustering techniques.