DMOIT: denoised multi-omics integration approach based on transformer multi-head self-attention mechanism.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Multi-omics data integration is vital for understanding biological complexity but faces challenges like data heterogeneity, high dimensionality, and noise.
• Existing methods struggle to capture intra- and inter-omics interactions effectively, leading to suboptimal performance in biological analyses.

Purpose of the Study:

• To introduce a novel Denoised Multi-Omics Integration approach (DMOIT) utilizing Transformer multi-head self-attention.
• To enhance the accuracy and robustness of multi-omics data analysis for biological insights.

Main Methods:

• DMOIT employs three modules: a generative adversarial imputation network for missing data, robust feature selection for noise reduction, and a novel multi-head self-attention (MHSA) architecture for enhanced intra-omics interaction capture.
• Model performance was validated on Cancer Genome Atlas (TCGA) datasets for cancer survival time and estrogen receptor status classification.

Main Results:

• DMOIT demonstrated superior performance compared to traditional machine learning methods and the MoGCN state-of-the-art integration method.
• The approach achieved higher accuracy and weighted F1 scores in both survival time and estrogen receptor status classification tasks.
• Comparative analysis against alternative MHSA architectures confirmed DMOIT's consistent outperformance across various cancer types and omics combinations.

Conclusions:

• DMOIT offers a robust and effective solution for multi-omics data integration, outperforming existing methods.
• The proposed approach shows significant potential as a valuable tool for diverse applications in biological research and precision medicine.