GRASMOS: Graph Signage Model Selection for Gene Regulatory Networks

Angelina Brilliantova, Hannah Miller, Ivona BezákováAAAI 2023 · DOI

Introduction

In the study of biological systems, Gene Regulatory Networks (GRNs) are typically represented as directed graphs where edges denote interactions between genes. However, these interactions are not merely connections; they possess "signs" indicating whether a gene activates (+) or represses (-) another. While significant progress has been made in modeling network topology, understanding the generative patterns of network signage remains a challenge. GRASMOS (Graph Signage Model Selection) addresses this by treating node roles as latent variables and employing Maximum Likelihood Estimation (MLE) to identify the biological logic governing these interactions.

Methodology

The GRASMOS framework evaluates four hierarchical models of signage:

  1. Node-Oblivious (NO): Signs are assigned randomly, independent of node identity.
  2. Source-Consistent (SC): Signs depend solely on the regulator (source gene).
  3. Target-Consistent (TC): Signs depend solely on the regulated gene (target).
  4. Bi-Node-Consistent (BNC): Signs depend on the latent groups of both the source and target genes.

To overcome the $2^n$ computational complexity of checking all node role combinations, the authors utilize Markov Chain Monte Carlo (MCMC) sampling for the BNC model and exact linear-time algorithms for simpler models. A key innovation is a 16x runtime reduction achieved by identifying low-degree nodes in sparse GRNs and calculating their contributions exactly.

Datasets

The framework was validated using both synthetic topologies and real-world biological databases. The synthetic data utilized Directed Scale-Free (DSF) models to mimic real GRN structures, while biological data was sourced from RegulonDB (E. coli) and SubtiWiki (B. subtilis).

Summary of the biological and synthetic datasets used in the study, detailing node and edge counts.

Key Results and Biological Insights

Model Performance and Accuracy

On synthetic data, GRASMOS demonstrated high precision in recovering original model parameters. The framework effectively minimized the L1-norm error while maintaining high accuracy in recovering the latent signage logic across various parameter configurations.

Real-World Application and Model Fit

The application of GRASMOS to E. coli and B. subtilis revealed distinct regulatory patterns. For both species, the Bi-Node-Consistent (BNC) model provided the best fit, reflecting the inherent complexity of bacterial regulation.

Comparison of model parameters and likelihood scores across the NO, SC, TC, and BNC models, identifying BNC as the best fit.

In B. subtilis, the Source-Consistent (SC) model performed remarkably well, supporting the "operon" hypothesis where clusters of genes are governed by a single regulator. Conversely, E. coli exhibited patterns suggesting more complex feed-forward control mechanisms where genes tend to activate those in different functional groups.

Conclusion

GRASMOS provides a robust, computationally efficient method for uncovering the logic of interaction signage in complex networks. By decoupling signage from topology, it allows for cross-species comparisons of regulatory logic even when physical network structures differ. Future work will focus on integrating topology generation with signage and accounting for context-dependent interactions and missing data.

Summary updated

BibTeX
@inproceedings{brilliantova2023grasmos,
  author = {Angelina Brilliantova and Hannah Miller and Ivona Bezáková},
  title = {{GRASMOS: Graph Signage Model Selection for Gene Regulatory Networks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i10.26457},
}
Figure