Direct versus multi-scale genotype-phenotype encoding

Lobo, Beane and Levin (2012) posed the issue for planarian regeneration. They asked whether a direct encoding of target morphology exists, and noted that shape is commonly held to be an emergent property of cell interactions.¹ They pointed to worms in which a transient change in physiology produced a two-headed outcome. The DNA sequence was unchanged, yet the pattern persisted through later amputations. They suggested that the large-scale axial plan may be encoded in physiological networks, and so be open to non-genetic intervention.²

Watson and colleagues (2016) argued that the regulatory interactions of gene networks evolve like synaptic connections under correlation learning. On that view the networks hold a developmental memory that can store and recall several past phenotypes, as a Hopfield network does.³ They also showed that such networks can produce novel phenotypes that were never selected for but share structural regularities with past ones, such as new combinations of phenotypic modules. Szilágyi, Szabó, Santos and Szathmáry (2020) noted that evo-devo treats the genotype-phenotype mapping as highly non-linear.⁴ They argued that developmental memory of past selected phenotypes, together with attractors, suggests a non-linear genotype-phenotype map that can produce multiple distinct phenotypes, a view that evo-devo's treatment of the map also supports.⁴

Shreesha and Levin (2023) built a minimal artificial embryo: a one-dimensional array of integers scored by how monotonic it is. Their baseline used direct encoding, in which the array values specify the embryo's order of values directly. They compared it with a more realistic map in which cells have some competency.⁵ Hartl, Risi and Levin (2024) evolved neural cellular automata to build an 8×8 Czech flag. They found that direct-encoding and multi-scale solutions differ greatly under novel conditions.⁶ At near-zero noise, direct encoding found solutions fastest, which they partly attributed to how the cell types were defined.⁷ At noise levels of 0.375 and 0.5, the direct case found no solution in 2000 generations.⁸

Pio-Lopez, Hartl and Levin (2025) used the same kind of model for aging. They state that a moderate noise level can help evolution identify robust NCA parameters. If noise is too low, evolution most likely finds a direct encoding of the target pattern and neglects cellular competencies.⁹ If noise is too high, they report, it blocks a good solution and limits cell specification.⁹ They read noisy updates as accumulating misdifferentiated cells that corrupt neighbours' perceptions, a mechanism that accelerates aging but is not a cause in itself.⁹