Improved prediction of gene expression through integrating cell signalling models with machine learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Predicting gene expression levels is a key challenge in bioinformatics.
• Current approaches include mechanistic models and machine learning (ML), each with limitations.
• Mechanistic models lack empirical data utilization, while ML methods underutilize biological knowledge.

Purpose of the Study:

• To overcome limitations of existing gene expression prediction methods.
• To integrate mechanistic cell signaling models with ML approaches.
• To enhance predictive accuracy and biological interpretability.

Main Methods:

• Augmented ML with similarity features derived from mechanistic cell signaling models using graph theory.
• Generated seven sets of similarity features.
• Learned multi-target regression models using these features.
• Stacked seven regression models into an ensemble prediction model.

Main Results:

• All generated similarity features significantly improved prediction accuracy over a baseline model.
• The stacked model demonstrated significant improvement over the baseline on 95% of genes in an independent test set.
• The integrated approach provided interpretable biological knowledge, such as the role of ERBB3 in MCF7 breast cancer cells.

Conclusions:

• Integrating mechanistic models as graphs enhances ML predictive performance for gene expression.
• This hybrid approach provides valuable biological insights for developing advanced mechanistic models.
• The method improves both prediction accuracy and biological understanding in gene expression analysis.