A graph convolutional neural network for gene expression data analysis with multiple gene networks.

Area of Science:

• Genomics
• Computational Biology
• Machine Learning

Background:

• Standard spectral graph convolutional neural networks (GCNs) typically utilize a single gene network.
• Genomic applications often involve multiple condition- or tissue-specific gene networks, posing challenges for existing GCNs in disease classification.
• Identifying the most informative gene networks for specific learning tasks is often difficult.

Purpose of the Study:

• To develop a deep multiple graph convolutional neural network (GCN) capable of integrating information from multiple gene networks for disease classification.
• To introduce measures for assessing gene importance and gene network relevance to the learning task.
• To evaluate the performance of the new method on real-world cancer datasets.

Main Methods:

• A deep multiple graph convolutional neural network (GCN) was developed, combining spectral GCNs for gene-gene relationships and feed-forward neural networks (FNNs) for gene-specific expression profiles.
• The method computes gene features as weighted averages of neighbors' features and extracts information from expression profiles.
• Measures for gene importance and gene network relative importance were devised.

Main Results:

• The deep multiple GCN achieved high classification accuracy on breast cancer and diffuse large B-cell lymphoma datasets.
• The method successfully prioritized important genes strongly associated with cancer.
• Performance was compared against standard FNN, GCN, and random forest, demonstrating superior results.

Conclusions:

• The developed deep multiple GCN effectively incorporates multiple gene networks for improved disease classification.
• The method provides valuable insights into gene and gene network contributions to the learning task.
• This approach shows significant promise for genomic applications requiring the integration of diverse biological network data.