Canalization as a stabilizing principle of gene regulatory networks: a discrete dynamical systems perspective

TL;DR

This perspective addresses how canalization can stabilize gene regulatory networks (GRNs) despite pervasive noise and perturbations. It surveys discrete dynamical models (notably Boolean networks), formal definitions of canalization, and a suite of stability metrics that connect function-level canalization to network-level robustness, including Derrida-based analyses and coherence measures. The paper highlights multiple, interrelated canalization notions (depth, strength, input redundancy, collective canalization) and discusses their differential relevance for evolution, control, and network inference, while outlining challenges in relating theory to empirical data. It also extends the discussion to multistate discrete models, identifying conceptual and practical gaps and proposing directions for data-driven, cross-disciplinary research to bridge theory and experiment and to build comprehensive repositories for future work.

Abstract

Gene regulatory networks exhibit remarkable stability, maintaining functional phenotypes despite genetic and environmental perturbations. Discrete dynamical models, such as Boolean networks, provide systems biologists with a tractable framework to explore the mathematical underpinnings of this robustness. A key mechanism conferring stability is canalization. This perspective synthesizes historical insights, formal definitions of canalization in discrete dynamical models, quantitative measures of stability, illustrative applications, and emerging challenges at the interface of theory and experiment.

Figures (6)

• Figure 1: Example of a 2-node Boolean network $F(x_1,x_2) = (x_2,x_1)$. (A) Wiring diagram indicating that $x_1$ and $x_2$ regulate each other. (B) Boolean update rules in truth table format, containing one row for each state of the network. (C) Deterministic synchronous state transition graph, containing three attractors: two steady states and a 2-cycle. (D) Stochastic asynchronous state transition graph, containing the same two steady states but no cyclic attractor.
• Figure 2: Example of a Boolean nested canalizing function. (A) Truth table of the 3-input Boolean NCF $f(x_1,x_2,x_3) = x_1 \vee (x_2 \wedge x_3)$. Setting $x_1=1$ canalizes $f$ to $1$. Setting $x_2=0$ or $x_3=0$ canalizes the subfunction $f(x_1=0,x_2,x_3) = x_2 \wedge x_3$. This means $f$ is an NCF with layer structure $(1,2)$. (B) Canalizing properties can be derived from a Boolean (hyper)cube labeled according to $f$. The proportion of ($n-k$)-dimensional faces that are constant is the $k$-set canalizing proportion. For example, the constant (red) face $(1,\#,\#)$ indicates that $x_1=1$ canalizes $f$. Similarly, the constant (red) edges $(0,0,\#)$ and $(0,\#,0)$ indicate that $x_2=0$ or $x_3=0$ independently canalize the subfunction. (C) Two possible nested evaluation trees, highlighting that $x_1$ is in the most important canalizing layer and that $x_2$ and $x_3$ are equally important. (D) Reduced truth table where # indicates that a certain input does not matter. The edge effectiveness and input redundancy are computed from this reduced table.
• Figure 3: Relationships between canalization metrics for all non-degenerate Boolean functions with 2-5 inputs. (A) Observed proportion of functions with specific canalizing depth in 122 expert-curated Boolean GRN models kadelka2024meta. (B,C) Proportion of non-degenerate functions with each canalizing depth, shown on (B) linear and (C) log scales, computed using exact formulas kadelka2025number. As the number of inputs increases, functions with high canalizing depth become exponentially rarer. (D,E) Distribution of (D) canalizing strength and (E) normalized input redundancy for non-degenerate functions with specific canalizing depth, generated using BoolForgekadelka2025boolforge and 1000 random functions per distribution. The high correlations between canalizing depth, canalizing strength, and input redundancy indicate that these metrics capture related but not identical aspects of canalization.
• Figure 4: Difference between canalizing strength and normalized input redundancy. All 4-input non-degenerate Boolean functions were generated and analyzed using BoolForgekadelka2025boolforge.
• Figure 5: Canalization creates a stability gradient within Waddington's landscape. High canalization deepens the valleys that channel developing cells toward robust phenotypic outcomes (left–right gradient, top bar), reinforcing the stability of developmental trajectories. However, when multiple attractors in highly canalized networks coexist, they are concentrated in nearby regions of the state space. This crowding flattens the local landscape near attractors relative to mid-trajectory regions, leaving terminal cell states positioned closer to basin boundaries. As a consequence, mature phenotypes can be more susceptible to directed or coordinated perturbations even when the overall developmental funnel is highly robust. This "intra-valley stability gradient" offers a mechanistic interpretation of how canalization can simultaneously ensure reproducible development and permit regulated phenotype switching in contexts such as reprogramming, regeneration, and pathological transitions.