A Dataset of Benchmark Boolean Models for Gene Regulatory Networks Authors: Caya L. O. Hotstegs, Jose P. Llano, Hans A. Kestler Published: 2026 ABSTRACT Gene regulatory networks (GRNs) capture the processes involved in gene regulation. Boolean network (BN) modeling provides a simple but effective framework for understanding the dynamical behavior of GRNs. Although BNs have been widely studied and applied, algorithms and theoretical analyses are usually tested on ad hoc selected or artificially constructed models, which may introduce bias and fail to capture the essential structural and dynamical properties of real GRNs for which they are ultimately intended. Benchmarking offers standardized models for validation and comparison of computational methods and analyses. We construct benchmark BN models for GRNs of four major biological kingdoms: animals, bacteria, fungi, and plants. All models are built from empirically observed recurrent properties and motifs in GRNs. The proposed benchmark BNs provide a systematical and unbiased basis for evaluating algorithms and theoretical analyses. Background & Summary Modern systems biology requires us to move beyond the study of isolated gene interactions. Gene regulatory networks (GRNs) capture the complex mechanisms through which gene regulation occurs and can provide essential knowledge of system-level properties. Their proper study is facilitated by the use of mathematical network models. Boolean networks (BNs) are qualitative mathematical models that are used to represent GRNs. Originally introduced by Kauffman1, they have since become one of the most popular and powerful tools for analyzing GRNs2–8, and biological systems in general9–11. In a BN, each node in the network graph represents a biological entity, and the directed edges between them account for their influence, either activating (represented by a positive signed edge) or inhibiting (represented by a negative signed edge). Each node has a binary state, either on (1) or off (0) at a given time. A local Boolean function updates the state of each node based on the states of its neighbor nodes. The iterative application of all local functions defines the dynamics of the system whose analysis reveals properties of its long-term behavior. Specifically, the attractors of the system correspond to its stable configurations, which are often associated with phenotypes. The analysis of the dynamics of BNs has grown into a research field of its own, with countless works focusing on the study of their theoretical12–16 and computational properties5,17–20. A proper evaluation of many theoretical and computational studies requires testing on real GRNs. However, testing most often relies on arbitrarily chosen (or even ad hoc) BN models of GRNs. This can introduce bias, since such models may miss common properties and motifs appearing in real-world BN models of GRNs. Moreover, the large number of available BN models of GRNs makes it challenging to select a representative subset that covers the diverse structural properties without over- or under-representation. Benchmark models are constructed from common properties and provide a standard framework for algorithmic testing and theoretical analyses, ensuring reproducibility and enabling practical comparison of results across studies. Specifically, research has revealed properties and motifs common in real-world BN models of GRNs. These include the scale-freeness for out-degree distribution21,22, the appearance of feed-forward loops (FFLs) and feedback loops (FBLs)23–25, or the canalization of Boolean functions26,27, among others3,28. Despite their importance and feasibility, benchmark BN models have not yet been proposed, which makes their construction desirable. To address this limitation, in this work, we propose benchmark BN models for GRNs29. These networks are built considering the structural and dynamical properties and motifs empirically observed in the BN models of GRNs3. Specifically, we construct four signed network models to capture specific features of four major biological kingdoms: animals, bacteria, fungi, and plants. Our work provides a solid base for testing algorithms and enabling systematic comparison of BNs. Methods We construct the benchmark BNs29 by incorporating recurrent properties and motifs present in the GRNs as reported in prior meta-analytical studies3,21– 28,30–35. To do that, first, we present some necessary definitions and preliminaries. Secondly, we examine the methodology for constructing the benchmark BN models. Finally, we discuss the possible limitations of our approach. Definitions and preliminaries A simple, connected, directed graph G = (V, E) is given by its node set V = {1, 2, …, n} and its edge set E, where (u, v) ∈ E if there is a directed edge or arc from u ∈ V to v ∈ V. For every node u ∈ V, xu ∈ {0, 1} is its state variable or state value. A configuration is a vector of state values for all entities x = (x1, …, xn) ∈ {0, 1}n at a given time. Then, a (synchronous) Boolean network (BN) is a discrete dynamical system defined by the map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{F}}=({F}_{1},\ldots ,{F}_{n}):{\{0,1\}}^{n}\to {\{0,1\}}^{n},\qquad {\boldsymbol{F}}({\boldsymbol{x}})=({F}_{1}({\boldsymbol{x}}),\ldots ,{F}_{n}({\boldsymbol{x}})),$$\end{document}F=(F1,…,Fn):{0,1}n→{0,1}n,F(x)=(F1(x),…,Fn(x)),where Fu is the (Boolean) local update function associated with the node u ∈ V, whose expression only depends on the state values of the neighbors of u in the graph G = (V, E). Let a = (a1, …, an) ∈ {0, 1}n be a configuration of the system. When Fm(a) = a and Fk(a) ≠ a for every positive integer k < m, we say that a is an attractor of length m. If m = 1, then a is called a steady state or a fixed point. Attractors of length m > 1 are usually referred to as m-limit cycles or m-periodic orbits. Since the system is finite, the iteration of F eventually leads to an attractor, which makes them central for studying the dynamics of BNs. An edge is called a self-loop if it connects a node to itself. A path is a finite sequence of edges. The in-degree \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{v}^{{\rm{in}}}$$\end{document}kvin of a node v ∈ V is the number of its incoming edges (i.e., its regulatory inputs), while the out-degree \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{v}^{{\rm{out}}}$$\end{document}kvout is the number of its outgoing edges (i.e., its targets). The average connectivity of a network ⟨k⟩ is the mean in-degree of a node, i.e., the average number of regulators per node3. GRNs are typically scale-free networks, meaning that they exhibit a heavy-tailed power-law distribution for the out-degree3,11,21,28,30, that is, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{{\rm{out}}}(k)\propto {k}^{-\gamma },\qquad \,{\rm{being}}\,k\ge {k}_{\min }^{{\rm{out}}}\quad \,{\rm{and}}\,\quad \gamma > 1,$$\end{document}Pout(k)∝k-γ,beingk≥kminoutandγ>1,where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{\min }^{{\rm{out}}}$$\end{document}kminout is the minimum out-degree among all nodes. On the other hand, the in-degree distribution of GRNs approximates a Poisson distribution3. Signs can be assigned to the edges in G = (V, E) to obtain a signed interaction graph (G, σ), which represents the activating and inhibiting dependencies in BNs as defined by the local update functions. That is, each edge (u, v) ∈ E has a sign σ(u, v) = + 1 if u activates v, and σ(u, v) = − 1 if u inhibits v. Equivalently, the sign indicates whether the local update function Fu is increasing (positive) or decreasing (negative) with respect to xv. Additionally, for a path of nodes π ≡ u0∣u1∣…∣um, ui ∈ V, the path sign, σ(π), is the product of the signs of its edges. The signed interaction graph of a BN can be well-defined when all local update functions are regulatory, i.e., every variable appears once (positive or negative) in their disjunctive normal form. Aside from regulatory ones, canalizing Boolean functions are recurrent in modeling GRNs, since canalization promotes robustness to perturbations3,26,27. A Boolean function F is canalizing if there exists a canalizing variable xu, a canalizing input a ∈ {0, 1}, and a canalized output b ∈ {0, 1} satisfying \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F({x}_{1},\ldots ,{x}_{n})=\{\begin{array}{ll}b, & \,{\rm{if}}\,{x}_{u}=a,\\ {H}_{1}({x}_{1},\ldots ,{x}_{u-1},{x}_{u+1},\ldots ,{x}_{n})\ne b, & \,{\rm{otherwise}}.\end{array}$$\end{document}F(x1,…,xn)= {b,ifxu=a,H1(x1,…,xu-1,xu+1,…,xn)≠b,otherwise. If the function H1 is also canalizing, then F is 2-canalizing, and, in general, assuming the previous notation, if H1, …, Hk are canalizing, then F is k + 1-canalizing, and k + 1 is said to be the canalizing depth of F. For a fully canalizing function, the canalizing depth matches the in-degree. A feed-forward loop (FFL) in a directed graph G = (V, E) is a connected subgraph of three nodes u, v and w such that (u, v), (v, w) and (u, w) belong to E. That is, node u regulates w directly through the edge (u, v) and indirectly through the path u∣v∣w. When the signs of the interactions are considered, a FFL can be classified as coherent, if the signs of the direct edge (u, v) and the indirect path u∣v∣w coincide, indicating that both have the same regulatory influence; or incoherent, if the signs differ, meaning that they have opposing effects. There are 23 = 8 possible sign combinations for a FFL, four of them coherent and the other four incoherent. All eight FFL types are shown in Fig. 1.Fig. 1Types of coherent (1-4) and incoherent (5-8) FFLs. Types of coherent (1-4) and incoherent (5-8) FFLs. FFLs that share nodes or edges form a cluster3,30. Six possible clusters arise when considering two FFLs that share one node (Motifs 1–6 in Fig. 2). Six additional clusters are obtained when considering two FFLs that share two nodes (Motifs 7–12). For clusters with three nodes, there are two possible options composed of two FFLs (Motifs 13–14) and an additional one, consisting of three FFLs (Motif 15).Fig. 2Types of FFL clusters. Types of FFL clusters. A feed-back loop (FBL) of length m + 1 is a closed path u = u0|u1|…|um = u in G of m + 1 distinct nodes, i.e., a directed cycle. Since FBLs are paths, they can be positive or negative depending on whether the product of the signs of their edges is positive or negative. Positive FBLs are typically associated with steady-state attractors, while negative FBLs are generally linked to attractors of greater length31. Benchmark BNs construction methodology To build the proposed benchmark BNs, we constructed a unique synthetic unsigned network of size n = 20 that reproduces observed metrics of essential properties and motifs of GRNs3. This network, which we call the underlying network, is used as the base topology for all the signed benchmark networks across different kingdoms. To construct it, we first generated random directed graphs and iteratively adjusted them to meet empirical metrics. Specifically, in each iteration, we examined the out-degree, in-degree, and FFL cluster distributions of the generated graphs, as well as the average connectivity. Then, we applied automated edge modifications to approximate empirical distributions. Secondly, starting from the underlying network, we constructed four benchmark signed networks to represent the specific characteristics of the four major biological kingdoms: animals, bacteria, fungi, and plants. Signs, representing activating or inhibitory influences, were assigned to reproduce observed distributions of sign-based motifs3. In particular, these include the distribution of coherent and incoherent FFLs, the ratio of negative to positive interactions relative to in-degree, and the overall distribution of FBL signs. More specifically, for each of the four kingdoms, random sign assignments are applied to the edges forming FFLs to generate 25 partially signed network candidates that best match the observed distributions of coherent and incoherent FFLs3. This step is performed first since not all edges belong to FFLs in the underlying network. Each partially signed network is then used as a seed to build (totally) signed graphs by fitting the remaining sign-based motif distributions: the fraction of positive FBLs by cycle lengths 1 to 6, and the proportion of positive inputs per in-degree. To this end, a search algorithm is used that randomly assigns signs to free edges (those that do not belong to any FFL) until M edges remain. In parallel, L edges from FFLs are randomly selected according to their importance, measured by frequency of occurrence in FFLs, with less frequent edges being more likely to be chosen. Exhaustive search is then performed over the remaining M + L edges. The network with the smallest weighted error across the three target motif distributions is selected (with weights wFFL = wFBL = windeg = 1/3). The previous routine is repeated for 1000 iterations, and different values of M and L, and the best-fitting network is selected for each seed. Finally, the optimal networks for each seed are compared, and the overall best-fitting network is chosen. Lastly, to construct the BNs for each kingdom, we assigned a canalizing Boolean update function to every node in the benchmark signed network, excluding the external parameters. We used fully canalizing Boolean functions, as these are known to dominate in GRNs3. Specifically, we defined three alternative models based on the types of fully canalizing local update functions used in them. The first model uses only the logical OR and NOT operators, which corresponds to a so-called OR-NOT BN12,32. The second model uses, for each local function, either OR and NOT, or AND and NOT, which corresponds to a so-called AND-OR-NOT BN13,33,34. The third model is a network in which all local functions are canalizing and combine the operators AND, OR, and NOT27,35. Limitations Our proposed benchmark BN models capture common structural and dynamical properties observed across many GRNs but do not reproduce the mechanistic details of any specific GRN. Therefore, they do not replace testing on real GRNs, but rather provide valid and practical complementary models for computational evaluation and comparison. Data Records The dataset29 contains four benchmark signed network models, one for each of the biological kingdoms–animals, bacteria, fungi and plants. The underlying network is also provided. The dataset is organized in five subfolders. One subfolder contains the adjacency matrix of the underlying topology, provided as a CSV file, together with a visualization of the network. The remaining four subfolders are named according to the respective kingdom and contain the customized network models. Each of these folders contains (i) a CSV file with the adjacency matrix of the kingdom-specific network, (ii) a visualization of the network, and (iii) three TXT files specifying different Boolean network update rules. The complete dataset is available at 10.5281/zenodo.17406797. Data Overview As stated before, an underlying (unsigned) network is used as the base topology for the benchmark networks of all kingdoms. The underlying network comprises 19 interacting entities and one external parameter, and is represented by a graph with 20 nodes and 50 directed edges. With the underlying topology of the underlying network, four distinct benchmark signed networks are proposed, one for each kingdom, which are shown in Fig. 3.Fig. 3The benchmark signed networks for each of the four kingdoms. A green arc (pointed end) indicates an activating interaction, and a red arc (flat end) an inhibitory interaction. Bidirectional edges are green if both interactions are activating, red if both are inhibitory, and mixed if one is activating and the other inhibitory. The benchmark signed networks for each of the four kingdoms. A green arc (pointed end) indicates an activating interaction, and a red arc (flat end) an inhibitory interaction. Bidirectional edges are green if both interactions are activating, red if both are inhibitory, and mixed if one is activating and the other inhibitory. Lastly, for each benchmark signed network, three distinct benchmark BN models are proposed: an OR-NOT BN, an AND-OR-NOT BN, and a general canalizing BN. All models have different attractors and attractor lengths36 as displayed in Table 1.Table 1Type and number of attractors in the benchmark Boolean network models for every kingdom and Boolean update rule set.Animal KingdomOR-NOTAND-OR-NOTGeneral canalizingTypeNumberTypeNumberTypeNumberFixed point4**Fixed point2**2-periodic22-periodic2****4-periodic4Bacteria KingdomOR-NOTAND-OR-NOTGeneral canalizingTypeNumberTypeNumberTypeNumberFixed point2Fixed point6Fixed point10****2-periodic4Funghi KingdomOR-NOTAND-OR-NOTGeneral canalizingTypeNumberTypeNumberTypeNumberFixed point6Fixed point2Fixed point5**2-periodic4******3-periodic8Plant KingdomOR-NOTAND-OR-NOTGeneral canalizingTypeNumberTypeNumberTypeNumberFixed point2Fixed point6Fixed point6For OR-NOT update rules, only the Boolean functions OR and NOT are used. For AND-OR-NOT rule sets, each rule consists either exclusively of AND and NOT or exclusively of OR and NOT functions. Canalizing rules may use any combinations of AND, OR and NOT but still have to be canalizing. The type of an attractor defines if it is a fixed point or a periodic attractor with length l. * indicates that there are no attractors of a certain length. Table 1Type and number of attractors in the benchmark Boolean network models for every kingdom and Boolean update rule set.Animal KingdomOR-NOTAND-OR-NOTGeneral canalizingTypeNumberTypeNumberTypeNumberFixed point4**Fixed point2**2-periodic22-periodic2****4-periodic4Bacteria KingdomOR-NOTAND-OR-NOTGeneral canalizingTypeNumberTypeNumberTypeNumberFixed point2Fixed point6Fixed point10****2-periodic4Funghi KingdomOR-NOTAND-OR-NOTGeneral canalizingTypeNumberTypeNumberTypeNumberFixed point6Fixed point2Fixed point5**2-periodic4******3-periodic8Plant KingdomOR-NOTAND-OR-NOTGeneral canalizingTypeNumberTypeNumberTypeNumberFixed point2Fixed point6Fixed point6For OR-NOT update rules, only the Boolean functions OR and NOT are used. For AND-OR-NOT rule sets, each rule consists either exclusively of AND and NOT or exclusively of OR and NOT functions. Canalizing rules may use any combinations of AND, OR and NOT but still have to be canalizing. The type of an attractor defines if it is a fixed point or a periodic attractor with length l. * indicates that there are no attractors of a certain length. Type and number of attractors in the benchmark Boolean network models for every kingdom and Boolean update rule set. For OR-NOT update rules, only the Boolean functions OR and NOT are used. For AND-OR-NOT rule sets, each rule consists either exclusively of AND and NOT or exclusively of OR and NOT functions. Canalizing rules may use any combinations of AND, OR and NOT but still have to be canalizing. The type of an attractor defines if it is a fixed point or a periodic attractor with length l. * indicates that there are no attractors of a certain length. Technical Validation The underlying network has n = 20 nodes, which makes it tractable for analysis using current software tools36. The network has one external parameter, matching that empirically observed for networks of similar size3. Additionally, the underlying network exhibits an average connectivity of μ⟨k⟩ = 2.5, consistent with the observed mean average connectivity α⟨k⟩ = 2.56 for comparable models3. Network models of GRNs typically exhibit scale-free out-degree distributions, characterized by a few highly connected hub nodes and many nodes of low connectivity3,21,22. The underlying network follows this pattern and is consistent with power-law distribution (although small in number of nodes for significant statistical testing). The out-degree distribution is shown in Fig. 4 (left), along with the in-degree distribution (right).Fig. 4The out-degree distribution (left) and in-degree distribution (right) of the underlying network. The out-degree distribution (left) and in-degree distribution (right) of the underlying network. As mentioned previously, one of the main subgraph motifs in GRNs are FFLs23–25. FFLs usually form clusters, where the same edge belongs to multiple FFLs (see Figure 2). Such clusters are independent of the interaction signs and can therefore be analyzed for the underlying network. The distribution of FFL clusters in the underlying network approximates that observed for real GRN models3, as shown in Fig. 5.Fig. 5Frequency distribution of FFLs in the underlying network and distribution of relative frequencies in real GRN models3. Frequency distribution of FFLs in the underlying network and distribution of relative frequencies in real GRN models3. Benchmark signed network model per kingdom The assignment of a positive or negative sign to each of the three edges of a FFL (corresponding to an activating or an inhibitory interaction) yields 23 possible types. By considering this, FFLs can be classified as coherent or incoherent3,23,37 (see Figure 1). The distribution of FFL types in the benchmark signed networks matches that empirically observed in real GRN models3, which are shown in Fig. 6.Fig. 6Distribution of FFL types in the benchmark signed networks and corresponding distribution in real GRN models3 across kingdoms. Distribution of FFL types in the benchmark signed networks and corresponding distribution in real GRN models3 across kingdoms. The proportion of activating to inhibitory inputs by in-degree for each kingdom has also been analyzed in real GRN models3. Each of the proposed benchmark BN models reproduces this corresponding proportion, as shown in Fig. 7.Fig. 7Distribution of activating vs inhibitory interactions by in-degree in the benchmark signed networks and corresponding distribution in real GRN models3. Distribution of activating vs inhibitory interactions by in-degree in the benchmark signed networks and corresponding distribution in real GRN models3. Lastly, the proportion of activating vs inhibitory edges has also been analyzed for FBLs of different sizes3, showing that FBLs are enriched for inhibitory edges. Each benchmark BN model reproduces this pattern, as shown in Fig. 8.Fig. 8Distribution of activating vs inhibitory interactions by FBL length in the benchmark signed networks and corresponding distribution in real GRN models3 across kingdoms. Distribution of activating vs inhibitory interactions by FBL length in the benchmark signed networks and corresponding distribution in real GRN models3 across kingdoms. Benchmark Boolean network models per kingdom Different local update functions can produce very different dynamics for the same signed network. To account for this, for each benchmark signed network, we propose three benchmark BN models per kingdom, all sharing the same signed graph, but with different canalizing local update functions. All local update functions are chosen so that the canalization depth equals the number of inputs of each node, in accordance with empirical observations in real GRN models3,26,27. As previously stated, the first model corresponds to an OR-NOT BN12,32, the second model to an AND-OR-NOT BN13,33,34, and the third corresponds to a general canalizing BN27,35. The dynamics of the models range from fixed points to coexisting attractors of greater length, showing how tuning the local update functions can generate diverse behaviors and capture different real-world scenarios. The proposed models and associated dynamical analyzes are provided in 10.5281/zenodo.17406797.