Supervised, semi-supervised and unsupervised inference of gene regulatory networks

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Gene regulatory network inference from expression data is a complex challenge.
• Existing methods lack comprehensive evaluation across different approaches.
• Guidelines for practical application are needed.

Purpose of the Study:

• To extensively evaluate unsupervised, semi-supervised, and supervised gene regulatory network inference methods.
• To compare method performance on both simulated and experimental expression data.
• To provide practical guidelines for method selection.

Main Methods:

• Performed an extensive evaluation of various gene regulatory network inference algorithms.
• Utilized both simulated and experimental gene expression datasets.
• Compared prediction accuracies across unsupervised, semi-supervised, and supervised approaches.

Main Results:

• Unsupervised methods generally showed low prediction accuracies.
• The Z-SCORE method demonstrated notable accuracy on knockout data.
• Supervised methods consistently achieved the highest accuracies.
• Semi-supervised approaches with limited positive samples outperformed unsupervised methods.

Conclusions:

• Supervised methods are recommended for accurate gene regulatory network inference.
• The Z-SCORE method is a viable unsupervised option for specific data types (e.g., knockout).
• Further research into optimizing semi-supervised methods may improve their performance.