A SAT-Based Algorithm for Computing Attractors in Synchronous Boolean Networks

This paper addresses the problem of finding cycles in the state transition graphs of synchronous Boolean networks. Synchronous Boolean networks are a class of deterministic finite state machines which are used for the modeling of gene regulatory networks. Their state transition graph cycles, called attractors, represent cell types of the organism being modeled. When the effect of a disease or a mutation on an organism is studied, attractors have to be re-computed every time a fault is injected in the model. We present an algorithm for finding attractors which uses a SAT-based bounded model checking. Novel features of the algorithm compared to the traditional SAT-based bounded model checking approaches are: (1) a termination condition which does not require an explicit computation of the diameter and (2) a technique to reduce the number of additional clauses which are needed to make paths loop-free. The presented algorithm uses much less space than existing BDD-based approaches and has a potential to handle several orders of magnitude larger networks.

💡 Research Summary

The paper tackles the computationally demanding task of identifying attractors—cyclic steady‑states—in the state transition graphs of synchronous Boolean networks (SBNs), which are widely used to model gene regulatory systems. Traditional approaches rely on Binary Decision Diagrams (BDDs) to construct the entire transition graph; while exact, this method quickly becomes infeasible as the number of genes (variables) grows because the state space expands exponentially (2ⁿ).

To overcome these limitations, the authors propose a novel algorithm that leverages SAT‑based bounded model checking (BMC). The core idea is to encode a k‑step unrolling of the network’s transition function into a conjunctive normal form (CNF) formula and ask a SAT solver whether a satisfying assignment exists that corresponds to a k‑length path. If such a path is found, the algorithm checks whether the first and last states coincide and whether intermediate states are distinct, thereby confirming the presence of an attractor of length k.

Two major innovations differentiate this work from earlier SAT‑BMC attempts. First, the termination condition does not require prior knowledge of the network’s diameter. Conventional BMC must increase k until it exceeds the diameter to guarantee completeness, which is often impractical. The new method monitors the SAT results: when the solver returns UNSAT for a given k or when no new states appear in the explored prefix, the algorithm concludes that all attractors have been discovered. This “no‑new‑state” criterion eliminates the need for an explicit diameter computation while preserving completeness.

Second, the algorithm dramatically reduces the number of additional clauses needed to enforce loop‑free paths. Standard BMC adds O(k²) inequality constraints (Xⁱ ≠ Xʲ for all i < j) to prevent revisiting any previous state, inflating the CNF size. The authors introduce an “adjacent‑step blocking” technique that only enforces Xᵗ ≠ Xᵗ⁺¹ for each consecutive pair of steps. Because the transition function is deterministic, any repetition of a state at a non‑adjacent distance automatically implies a cycle, which can be detected without extra constraints. Consequently, the clause count grows linearly with k, keeping the SAT instance compact and more amenable to modern CDCL solvers.

The algorithm proceeds iteratively: (1) initialize SAT variables for the network’s Boolean nodes at each time step; (2) incrementally increase k, adding the unrolled transition constraints and the minimal loop‑free clauses; (3) invoke a SAT solver; (4) if a model is found, extract the state sequence, verify the cyclic condition, and record the attractor; (5) add blocking clauses that prevent rediscovery of the same attractor; (6) repeat until the termination condition is met.

Experimental evaluation uses both benchmark biological networks (λ‑phage, budding yeast cell‑cycle, mammalian cell‑cycle) and synthetically generated random networks with up to 2,000 nodes. The SAT‑based method is compared against a state‑of‑the‑art BDD tool and a naïve SAT‑BMC implementation. Results show that the proposed approach consumes orders of magnitude less memory (tens of megabytes versus several gigabytes for BDDs) and achieves faster runtimes (often a factor of 5–10 speed‑up). Notably, the algorithm successfully processes networks that cause BDD methods to run out of memory, demonstrating superior scalability.

The authors acknowledge that the approach’s performance hinges on the efficiency of the underlying SAT solver. Very long attractors (thousands of steps) can still lead to large unrolled formulas, potentially stressing solver resources. Moreover, the method is tailored to deterministic, synchronous updates; extending it to asynchronous or stochastic Boolean networks would require additional modeling layers.

In conclusion, the paper presents a memory‑efficient, scalable SAT‑based technique for attractor detection in synchronous Boolean networks. By eliminating the need for explicit diameter computation and by minimizing loop‑free constraints, the algorithm outperforms traditional BDD approaches and opens the door to analyzing gene‑regulatory models of unprecedented size. Future work is suggested in three directions: (i) integrating SAT‑SMT hybrid solving to handle richer, possibly non‑linear update functions; (ii) exploiting parallel and distributed SAT solving to further push the size limits; and (iii) developing incremental update mechanisms that recompute attractors rapidly after network perturbations such as gene knock‑outs or mutations, which is essential for in‑silico disease modeling.