Deep learning-based approaches for multi-omics data integration and analysis.

Area of Science:

• Biomedical data science
• Artificial intelligence in healthcare
• Computational biology

Background:

• Deep learning and big data enable complex, heterogeneous data fusion and analysis.
• Multi-omics data (genomics, imaging) offer complementary insights for a holistic understanding.
• Integrating diverse data modalities can significantly improve predictive and classification tasks.

Purpose of the Study:

• To review and categorize deep learning-based multi-omics integration methods.
• To discuss the capabilities and emerging themes in the field.
• To provide a framework for understanding different deep learning architectures for omics data.

Main Methods:

• Categorization of deep learning approaches by architecture (non-generative and generative).
• Analysis of the unique strengths and weaknesses of each category.
• Discussion of emerging trends in multi-omics data integration.

Main Results:

• Deep learning methods are broadly classified into non-generative (feedforward, GCN, autoencoders) and generative (variational, GANs, pretrained models).
• Generative methods can enforce constraints, incorporate prior knowledge, and impute missing data modalities.
• Recent advances enable handling incomplete data and integrating non-traditional omics data, including imaging.

Conclusions:

• Expect growth in methods addressing data missingness, a common challenge.
• Integrating more diverse data types is anticipated to boost downstream task performance.
• A comprehensive view of samples through multi-modal integration will drive future biomedical discoveries.