Multi-Omics Data Integration for Improved Cancer Subtyping via Denoising Autoencoder-Based Multi-Kernel Learning.

Area of Science:

• Computational biology
• Genomics
• Machine learning

Background:

• Cancer is a complex disease with diverse molecular profiles, necessitating precise patient stratification for effective treatment.
• Multi-omics data integration offers a powerful approach to understanding cancer heterogeneity and advancing precision medicine.

Purpose of the Study:

• To develop and validate a novel deep learning framework, DAE-MKL, for identifying molecular cancer subtypes using multi-omics data.
• To enhance the accuracy of cancer subtyping and improve prognostic stratification.

Main Methods:

• The DAE-MKL framework integrates denoising autoencoders (DAE) for feature extraction and multi-kernel learning (MKL) for subtype identification.
• DAE reduces data dimensionality and noise, while MKL enhances classification accuracy.
• The framework was applied to simulated data and real-world datasets from low-grade glioma (LGG) and kidney renal clear cell carcinoma (KIRC).

Main Results:

• DAE-MKL demonstrated superior performance in simulations, achieving significant improvements in Normalized Mutual Information (NMI).
• In LGG and KIRC datasets, DAE-MKL identified distinct subtypes with significant survival differences (LGG log-rank p = 3.99 × 10^-8, KIRC log-rank p = 3.33 × 10^-8).
• Potential cancer-related biomarkers were identified for the discovered subtypes.

Conclusions:

• DAE-MKL is an effective tool for identifying molecular subtypes in multi-omics cancer data.
• The framework successfully reduces data dimensionality and improves prognostic stratification.
• This method advances precision oncology by enabling more accurate patient subgrouping and personalized treatment strategies.