Bipartite networks represent causality better than simple networks: evidence, algorithms, and applications

Area of Science:

• Systems Biology
• Bioinformatics
• Machine Learning

Background:

• Gene regulatory networks (GRNs) are models of gene interactions.
• GRNs are often inferred using machine learning and validated against experimental data.
• The utility of GRNs is often assessed by their ability to represent causality.

Purpose of the Study:

• To compare the predictive performance of GRNs inferred from time-series data with gold standard regulatory edges.
• To evaluate if current GRN inference methods accurately capture gene regulatory causality.
• To propose new goals for causality research in gene regulation.

Main Methods:

• Inferred non-linear machine learning models from time-series gene expression data across four species.
• Compared the predictive accuracy of these models against gold standard regulatory edges.
• Calculated the reduction in root mean square error (RMSE) to quantify performance improvements.

Main Results:

• Non-linear machine learning models inferred from time-series data demonstrated superior predictive performance.
• Improvements in predictive performance ranged from 5.3% to 25.3% in RMSE reduction compared to models based on gold standard edges.
• Established that current GRN inference methods may not fully capture causality.

Conclusions:

• Gene regulatory networks may not accurately characterize causality.
• Causality research should prioritize predictive accuracy.
• Proposed new research directions including parsimonious enumeration of predictive genes, identification of disjoint predictive gene sets, and bipartite network representations.