A denoised multi-omics integration framework for cancer subtype classification and survival prediction

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning in Oncology

Background:

• High-throughput sequencing generates vast, high-dimensional data for disease research.
• Training machine learning models with such data presents significant computational challenges.
• Integrating multi-omics data is crucial for a comprehensive understanding of complex diseases like cancer.

Purpose of the Study:

• To develop a denoised multi-omics integration framework for improved cancer prognosis prediction and subtype identification.
• To address the challenges of high-dimensional omics data in machine learning models.
• To identify potential cancer biomarkers through feature selection.

Main Methods:

• Feature Selection with Distribution (FSD): A distribution-based algorithm for feature denoising and dimension reduction.
• Attention Multi-Omics Integration (AttentionMOI): A novel framework for integrating diverse omics data.
• Application to 15 The Cancer Genome Atlas Program (TCGA) cancer datasets for survival prediction and subtype identification.

Main Results:

• FSD significantly improved model performance for both single-omic and multi-omic data analyses.
• AttentionMOI outperformed existing machine learning models and multi-omics integration methods in high-dimensional settings.
• The study successfully predicted cancer prognosis and identified distinct cancer subtypes using the proposed framework.

Conclusions:

• The proposed denoised multi-omics integration framework (FSD + AttentionMOI) offers a robust approach for analyzing high-dimensional omics data.
• This framework enhances the accuracy of cancer prognosis prediction and cancer subtype identification.
• Identified features by FSD hold potential as novel cancer biomarkers for clinical applications.