AIME

AIME extracts low-dimensional embeddings from multi-omics data using autoencoders while adjusting for confounding variables to enable integrative analysis of relationships between datasets such as microRNA and gene expression.

Key Features:

• Nonlinear Data Embedding: Uses autoencoder-based deep learning to capture nonlinear relationships between omics types, including microRNA and gene expression, compared with linear methods such as canonical correlation analysis (CCA) and partial least squares (PLS).
• Confounder Adjustment: Incorporates clinical confounding variables into the model to adjust their effects on the integrated embeddings.
• Informative Data Representation: Produces optimized low-dimensional representations that reflect intrinsic relationships between different omics datasets.
• Feature Ranking and Pairing: Identifies and ranks features by their contributions and detects paired features across data types via the learned embeddings.
• Simulation Studies Validation: Demonstrated through simulation studies to effectively identify major contributing features across datasets.

Scientific Applications:

• Integrative multi-omics analysis: Enables joint analysis of multiple omics datasets to reveal cross-omics relationships and interactions.
• Confounder-aware clinical studies: Supports analyses that require adjustment for clinical confounding variables to reduce spurious associations.
• Biomarker and therapeutic target discovery: Facilitates identification of candidate biomarkers and therapeutic targets via embedded feature ranking and pairing.

Methodology:

Trains autoencoders to learn compressed representations of input omics data while incorporating confounding variables directly into the model architecture; implemented using Keras with a TensorFlow backend.