Knowledge-driven interpretable neural networks provide mechanistic insight

Area of Science:

• Bioinformatics
• Systems Biology
• Machine Learning

Background:

• Pathway analysis is crucial for understanding disease mechanisms but current methods lack mechanistic detail.
• Existing approaches often oversimplify complex biological interactions, limiting deep insights.

Purpose of the Study:

• To develop a knowledge-driven machine learning framework for in-depth pathway analysis.
• To enable interpretable modeling of biological interactions using omics data.
• To bridge predictive modeling with mechanistic interpretation for disease research.

Main Methods:

• A knowledge-driven machine learning framework embedding features into pathway graphs.
• Analytical modeling of biological reactions within pathways.
• Estimation of functional associations to model biological interactions.
• Feature selection agnostic approach utilizing full omics datasets.

Main Results:

• Generated interpretable feature hierarchies and subnetworks.
• Successfully applied to breast cancer microRNA-gene regulation and COVID-19 metabolomics data.
• Highlighted key immune and metabolic pathways associated with disease progression.

Conclusions:

• The framework provides a foundation for integrative pathway analysis, combining predictive power with mechanistic understanding.
• Enables deeper insights into molecular mechanisms underlying pathological changes.
• Offers a robust method for analyzing complex omics data in the context of biological pathways.