Autoencoders with shared and specific embeddings for multi-omics data integration.

Area of Science:

• Computational biology
• Bioinformatics
• Cancer genomics

Background:

• High-dimensional data integration is crucial for understanding cancer.
• Existing methods may not fully capture shared and specific information across datasets.

Purpose of the Study:

• To develop a novel autoencoder (AE) structure for integrating diverse cancer data types.
• To explicitly model shared and specific information using orthogonal embeddings.
• To evaluate the AE model's performance in cancer data integration and classification.

Main Methods:

• Proposed a novel autoencoder (AE) architecture with orthogonal loss for shared and specific embeddings.
• Utilized simulated data and The Cancer Genome Atlas (TCGA) cancer data for validation.
• Compared the AE model against existing AE structures and the Joint and Individual Variance Explained (JIVE) method.
• Assessed performance based on reconstruction loss and classification accuracy.

Main Results:

• The proposed AE models demonstrated slightly improved reconstruction loss compared to other AE structures.
• All AE models significantly enhanced classification accuracy over original features.
• The novel AE structure achieved high and consistent classification accuracy on training and testing sets.

Conclusions:

• The proposed autoencoder model effectively integrates multi-omics cancer data.
• Orthogonal constraints in AE embeddings improve data integration and downstream task performance.
• The novel AE approach shows superior or competitive performance compared to existing methods like JIVE and MOCSS.