Benchmarking of Optimization Algorithms for Boolean Model Inference in Biomedicine

Biological processes in health and disease are regulated in great complexity, imposing significant challenges in understanding and modifying their behavior for healthcare applications. Boolean networks have become essential tools for modeling gene regulatory systems and understanding cellular decision-making processes, but their optimization for biological relevance and precision medicine remains challenging.

This study presents a comprehensive benchmark comparison of four prominent Boolean network optimization methods involving genetic algorithms, integer linear programming, and answer set programming, evaluating their performance across structural robustness, method reliability, and biological relevance using mean squared error (MSE) as the primary optimization criterion. Through systematic analysis of network reconstruction under varying perturbation levels (10-90%), we demonstrate that each method exhibits distinct performance profiles: answer set programming (ASP) achieves optimal topological similarity with computational efficiency, integer linear programming (ILP) produces reasonable MSE minimization but with high variance, genetic algorithms (GA) shows superior functional reconstruction stability despite longer computational times. Our results reveal critical limitations in current evaluation approaches, particularly the insufficient discriminatory power of F1 scores and Hamming distance metrics, and highlight fundamental trade-offs between data fitting accuracy and topological preservation.

The analysis demonstrates that no single optimization method dominates across all criteria, with all methods showing significant performance degradation at perturbation thresholds above 10-30%, suggesting that method selection should be application-specific and guided by requirements for computational efficiency, reconstruction accuracy, and robustness to uncertainty in prior knowledge.