AVBAE-MODFR: A novel deep learning framework of embedding and feature selection on multi-omics data for pan-cancer...

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Oncology

Background:

• Multi-omics data integration is crucial for pan-cancer classification and clinical applications like precision medicine.
• High-dimensional and heterogeneous cancer multi-omics data present challenges for deep learning-based embedding and feature selection.
• Deep learning methods show promise for capturing complex, nonlinear relationships in multi-omics data.

Purpose of the Study:

• To propose a novel two-phase deep learning model, AVBAE-MODFR, for effective embedding and feature selection in pan-cancer classification.
• To enhance the quality of multi-omics data representation and improve feature ranking stability.
• To advance the clinical utility of multi-omics data integration for cancer diagnosis and treatment.

Main Methods:

• Developed AVBAE-MODFR, a two-phase deep learning model combining Adversarial Variational Bayes Autoencoder (AVBAE) for embedding and a Multi-omics Dual-net based Feature Ranking (MODFR) for selection.
• AVBAE utilizes a multi2multi autoencoder with an efficient discriminator for unsupervised representation learning.
• MODFR employs a multi2one selector network and an average gradient approach for robust multi-omics feature importance evaluation.

Main Results:

• AVBAE-MODFR demonstrated superior performance in pan-cancer classification compared to state-of-the-art methods on the TCGA dataset.
• The AVBAE phase effectively learned high-quality representations from multi-omics features.
• The MODFR phase provided stable and reliable feature ranking, mitigating bias from input feature drift.

Conclusions:

• AVBAE-MODFR offers a powerful and effective deep learning framework for integrating and analyzing cancer multi-omics data.
• The proposed model significantly improves embedding and feature selection for pan-cancer classification.
• This approach holds substantial potential for enhancing clinical applications in oncology, including diagnosis and precision medicine.