AutoEncoder

AutoEncoder integrates multi-omics data, including gene expression, miRNA expression, and DNA methylation, with biological interaction networks to learn joint feature and patient embeddings for prediction of clinical variables in TCGA cancer datasets.

Key Features:

• Model architecture - Multi-view Factorization AutoEncoder (MAE): Employs a multi-view factorization autoencoder to jointly represent multiple molecular feature spaces.
• Integration with network constraints: Incorporates biological interaction networks as constraints during representation learning to link molecular features across views.
• Deep representation learning: Learns simultaneous feature and patient embeddings using deep learning to capture high-dimensional relationships despite "big p, small n" scenarios.
• Regularization with domain knowledge: Adds domain-knowledge-derived regularization terms to the training objective to introduce inductive bias and mitigate overfitting.
• Predictive evaluation on TCGA: Demonstrates capability to predict target clinical variables through experiments on The Cancer Genome Atlas (TCGA) datasets.

Scientific Applications:

• Cancer multi-omics integration: Integrates gene expression, miRNA expression, and DNA methylation to reveal cross-layer molecular interactions in cancer.
• Clinical variable prediction: Predicts clinical outcomes and target clinical variables from integrated multi-omics and network-constrained embeddings in TCGA datasets.
• Personalized medicine and target discovery: Supports identification of molecular relationships and potential therapeutic targets by linking molecular features to clinical outcomes.

Methodology:

Training of a multi-view factorization autoencoder with biological interaction network-derived constraints, deep learning to learn feature and patient representations, and regularization terms encoding domain knowledge.