Modularity of developmental gene regulatory networks

Ten Tusscher and Hogeweg (2011) summarised the theoretical case against expecting modular networks to evolve. Theory had argued that evolution should neither produce nor preserve architecturally modular networks¹. Their account of the reason was that modular networks form only a small subset of the architectures able to perform a given function. Studies that did evolve modular networks had to use quite specific fitness targets¹. They also noted that other genome and network properties can arise as a neutral side effect of mutational dynamics, and that similar suggestions had been made for modularity¹. Against this, Kriegman, Cheney and Bongard (2018) stated that modular systems are more evolvable because evolution can improve one subsystem without disrupting others⁵.

Ten Tusscher and Hogeweg checked whether their parameter settings biased evolution toward densely connected, non-modular networks. They ran simulations with two times and five times higher transcription factor binding site deletion rates. Architectural modularity scores did not rise significantly, whether the original or the core networks were evaluated². They then continued the evolution of a previously evolved core network under the five-fold rate, and again saw no significant increase in Q values². They concluded that purely architectural methods suggest the two network types they compared are equally non-modular².

Other authors report modular structure in different settings. Sengupta, Stemmler and Friston (2013) wrote that neuronal networks obey Rent's rule and show hierarchical modularity that trades physical cost against topological complexity³. Lobo and Levin (2015) applied an automated inference method to wnt1 and wnt11-5 knock-down data. The model it found was a redundant modular network⁴. In that model, knocking down either gene alone gave the control pattern, while knocking down both gave a double-head worm, as in vivo⁴. Kriegman and colleagues observed that their evolved robots were modular by definition, since morphology genes were independent of control genes. The more successful lineages added differential canalization, in which some plastic traits became canalized and others stayed plastic⁵.

Krakauer and colleagues (2020) noted that network science has quantitative modularity measures built on partitions compared against a null model. Developmental definitions, such as those for limb formation or body segments, had not been given as quantitative measures⁶. Bongard and Levin (2023) went further. They called modular design in engineering a concession to human cognitive limits, and argued that if a system polycomputes, different observers may see different subsets of functions, in which case the question of how modular such a system is becomes ill-defined⁷. Blackiston and colleagues (2025) reported significant changes in network modularity before and after injury in aneural tissue. They made no mechanistic claim and left open whether the changes are instructive or an epiphenomenon of healing⁸.