deepNF: deep network fusion for protein function prediction.

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• High-throughput experiments generate large-scale molecular and functional interaction networks.
• Network connectivity offers rich information for inferring gene and protein functions.
• Integrating heterogeneous networks to extract protein features for function prediction remains challenging.

Purpose of the Study:

• To propose deepNF, a novel network fusion method based on Multimodal Deep Autoencoders.
• To extract high-level protein features from multiple heterogeneous interaction networks for function prediction.
• To overcome limitations of shallow models in capturing complex network structures.

Main Methods:

• Implemented deepNF using Multimodal Deep Autoencoders to fuse STRING interaction networks.
• Employed separate network-type layers in early stages, converging to a single bottleneck layer.
• Extracted high-level protein features from the bottleneck layer for function prediction.

Main Results:

• deepNF successfully constructed a common low-dimensional representation of protein features.
• The method outperformed state-of-the-art approaches, including Mashup, on human and yeast STRING networks.
• Achieved substantial improvements in predicting Gene Ontology terms across various types and specificities.

Conclusions:

• deepNF provides an effective deep learning approach for integrating heterogeneous biological networks.
• The method enhances protein function prediction accuracy compared to existing network integration techniques.
• The developed deepNF tool is publicly available for research use.