CustOmics: A versatile deep-learning based strategy for multi-omics integration.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Integrating multi-omics data is crucial for understanding disease mechanisms and developing predictive models.
• High-dimensional and heterogeneous omics data present significant computational challenges.
• Deep learning offers promising solutions for effective multi-omics data integration.

Purpose of the Study:

• To review existing autoencoder-based multi-omics integration strategies.
• To propose a novel, customizable two-phase deep learning approach for multi-omics integration.
• To enhance the interpretability of multi-omics data analysis using Shapley additive explanations.

Main Methods:

• A two-phase deep learning strategy: independent data source training followed by cross-modality interaction learning.
• Customizable autoencoder architecture adapted for multi-omics data integration.
• Integration with Shapley additive explanations for model interpretability.

Main Results:

• The proposed method efficiently integrates diverse omics sources, outperforming existing strategies.
• Demonstrated strong performance in cancer tasks, including tumor type and subtype classification, and survival prediction.
• Achieved excellent results across seven diverse TCGA datasets, providing interpretable insights.

Conclusions:

• The novel two-phase integration strategy effectively leverages individual omics data singularities.
• The customizable deep learning model offers a powerful and interpretable tool for multi-omics analysis in cancer research.
• The approach shows significant potential for advancing precision medicine through integrated omics data.