Integrate multi-omics data with biological interaction networks using Multi-view Factorization AutoEncoder (MAE).

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in genomics

Background:

• Vast multi-omics datasets (gene expression, methylation, etc.) are generated from disease profiling.
• Integrating these datasets is key to understanding complex molecular pathways.
• The 'big p, small n' problem challenges deep learning model generalizability with high-dimensional omics data.

Purpose of the Study:

• To develop a method for seamless integration of multi-omics data and domain knowledge.
• To address the overfitting and generalizability issues in deep learning for multi-omics data analysis.
• To improve the uncovering of relationships between molecular and clinical features.

Main Methods:

• Developed Multi-view Factorization AutoEncoder (MAE) with network constraints.
• Simultaneously learned feature and patient embeddings using deep representation learning.
• Incorporated biological domain knowledge (molecular interaction networks) as regularization terms.

Main Results:

• The MAE method effectively integrates multi-omics data with biological networks.
• Learned representations (features and patients) were constrained to improve generalizability.
• Experiments on TCGA datasets demonstrated the method's power in predicting clinical variables.

Conclusions:

• Incorporating biological domain knowledge as inductive biases mitigates deep learning overfitting.
• The MAE approach shows promise for integrating large-scale multi-omics data and biomedical knowledge.
• This facilitates a deeper understanding of molecular and clinical feature relationships in diseases.