Deeply integrating latent consistent representations in high-noise multi-omics data for cancer subtyping

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Cancer is a complex disease with high mortality, necessitating accurate subtyping for personalized treatment.
• Multi-omics data analysis is key to understanding cancer progression but faces challenges due to noise and integration difficulties.
• Existing methods struggle to capture consistent representations and integrate information effectively from noisy omics data.

Purpose of the Study:

• To develop a novel deep learning model for accurate cancer subtyping using multi-omics data.
• To address challenges in noise reduction and information integration within omics datasets.
• To improve the biological significance and interpretability of identified cancer subtypes.

Main Methods:

• Proposed a variational autoencoder-based deep learning model named Deeply Integrating Latent Consistent Representations (DILCR).
• Employed independent variational autoencoders and contrastive loss functions to extract latent consistent representations from noisy omics data.
• Utilized an Attention Deep Integration Network for effective cross-omics data integration and an Improved Deep Embedded Clustering algorithm for variable clustering.

Main Results:

• DILCR demonstrated superior performance in cancer subtyping across 10 diverse cancer datasets from The Cancer Genome Atlas.
• The model effectively captured consistent representations within omics data, outperforming 14 state-of-the-art integration methods.
• A case study on Kidney Renal Clear Cell Carcinoma identified biologically significant and interpretable cancer subtypes.

Conclusions:

• DILCR offers a robust framework for integrating multi-omics data for precise cancer subtyping.
• The model's ability to handle noisy data and extract consistent representations enhances its clinical applicability.
• Accurate subtyping using DILCR holds promise for advancing personalized cancer medicine and improving patient outcomes.