Multi-label Deep Learning for Gene Function Annotation in Cancer Pathways.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Biocuration is crucial for biomedical research but struggles to keep pace with data generation.
• Pathway analysis offers insights into cancer biology.
• Existing biocuration methods are resource-intensive and lag behind data generation.

Purpose of the Study:

• To develop a deep-learning model for enhanced gene multi-function discovery and pathway completion.
• To address the limitations of current biocuration and pathway analysis methods in cancer research.

Main Methods:

• Implementation of a Stacked Denoising Autoencoder Multi-Label Learning (SdaMLL) model.
• Utilizing deep learning to capture robust intermediate representations and generate low-dimensional codes.
• Comparison of SdaMLL against classical multi-label algorithms.

Main Results:

• SdaMLL demonstrated superior performance compared to existing multi-label algorithms.
• Identified novel gene functions, including Fused in Sarcoma (FUS) and p27, potentially involved in cancer.
• These findings contribute to completing existing cancer pathways.

Conclusions:

• SdaMLL is an effective deep-learning tool for gene function discovery and pathway completion in cancer research.
• The identified gene functions offer new insights into cancer mechanisms, such as transcriptional misregulation and viral carcinogenesis.
• A visual tool is available for exploring these new gene functions in cancer pathways.