Redundancy and computational capacity of cellular agents

Lobo and Levin (2015) described a mismatch in developmental biology. They wrote that much has been learned about the molecular pathways regulating stem cell decision-making. They added that axial polarity, morphogenesis and persistent changes to the bodyplan still lack constructivist models.¹ In the same work they reported that no quantitative model reproduces more than a few main features of the planarian regeneration dataset. Watson and colleagues (2016) treated a related problem at the level of evolution. They described traits that evolve parameters of population structure controlling the amount of mixing, which they linked to fraternal transitions such as multicellularity.²

Kriegman, Blackiston, Levin and Bongard (2020) used the phrase computational capacity of cells in an engineering setting. They wrote that exploiting it to let cells function in novel configurations suggests a possibility, not an established result. They suggested this could produce synthetic morphology with complex novel anatomies, combining emergence with guided self-assembly.³ Their starting point was that living systems resist adopting new behaviours imposed on them. The same robustness shown in development and regeneration is what they proposed to use.

Zakirov and colleagues (2021) studied cell decisions in angiogenesis, where cells compete to become the migratory tip cell. They reported that filopodia keep the Notch selection process timely. In their account, filopodia movements alter receptor locations and activations, and this amplifies differences between cells in a positive feedback cycle.⁴ They argued this sensorimotor coordination could constitute a basal form of active perception. They also noted that lamella-type extensions and triangular veils have not been considered for this role. Each protrusion type might give a different mechanism, and cells might combine several of them when enriching sensory input for problem-solving.⁵

Hartl, Risi and Levin (2024) varied computational capacity directly in simulated cells. Each uni-cellular agent has an artificial neural network. The authors added independent copies of its policy module, which interprets the local environment and proposes a state update. This raises the number of evolvable parameters doing the same job. They defined a redundancy number R for the copy count.⁶ A cell's output is the average over its copies, so decisions are biased by several redundant paths, an idea they took from error-correcting codes.⁷ The cells receive no fitness feedback. They must evolve their own communication protocol so that single-cell decisions fit the global target pattern.⁸