Should we really use graph neural networks for transcriptomic prediction?

Area of Science:

• Bioinformatics
• Machine Learning
• Computational Biology

Background:

• Deep learning methods, particularly graph neural networks (GNNs), show promise for bioinformatics tasks like phenotype prediction.
• GNNs leverage gene networks to embed information on gene regulation and co-expression.
• A comprehensive benchmark comparing GNNs to standard machine learning methods for phenotype prediction is lacking.

Purpose of the Study:

• To conduct a reproducible benchmark comparing GNNs with standard machine learning methods for phenotype prediction.
• To evaluate the cost-benefit trade-off of using GNNs in bioinformatics.
• To identify factors influencing the performance of GNNs in phenotype prediction.

Main Methods:

• Developed a benchmark framework with clear, comparable policies for evaluating different machine learning methods.
• Tested various methods, including GNNs and simpler alternatives, on multiple datasets.
• Utilized controlled simulated datasets to analyze method performance.

Main Results:

• GNNs rarely offered significant improvements in prediction performance over simpler machine learning methods.
• The computational effort required for GNNs often exceeded the performance gains.
• Analysis on simulated data suggested that the quality of the input gene network may limit GNN predictive power.

Conclusions:

• The practical benefit of using GNNs for phenotype prediction is often limited, especially considering their computational demands.
• The predictive accuracy of GNNs may be constrained by the inherent quality and biological relevance of the gene networks used as input.
• Further research into improving gene network construction and quality is crucial for enhancing GNN performance in bioinformatics.