A Hierarchical Graph Convolution Network for Representation Learning of Gene Expression Data.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Genomics

Background:

• High-dimensional gene expression data presents challenges like the curse of dimensionality and limited labeled samples.
• Existing deep learning methods struggle with interpretability and effective utilization of limited labeled gene expression data.
• Current semi-supervised graph convolution networks often analyze only feature or sample spaces, limiting their performance.

Purpose of the Study:

• To propose a novel transductive semi-supervised representation learning method for gene expression data.
• To address the challenges of high dimensionality, low sample size, and limited labeled data in gene expression analysis.
• To enhance the interpretability of deep learning models in biomedical applications.

Main Methods:

• Developed a hierarchical graph convolution network (HiGCN) integrating both feature and sample spaces.
• Constructed a feature graph using external knowledge and a sample graph via a similarity kernel.
• Employed two spatial-based graph convolutional networks (GCNs) for information aggregation.

Main Results:

• HiGCN effectively aggregates information from both feature and sample spaces.
• The model learns superior representations for gene expression data compared to existing methods.
• Performance on downstream tasks is significantly improved, particularly with scarce labeled samples.

Conclusions:

• HiGCN offers a powerful approach for analyzing high-dimensional gene expression data with limited labels.
• The method provides reliable interpretability by extracting important features.
• HiGCN advances semi-supervised learning for complex biological data analysis.