Synthetic genetic networks based on neural models

Early works of system biology attempted to adopt design principles of artificial neural networks (ANNs), in which the state of a particular node depends on the combination of the states of other signals, to explain the emergent properties of gene networks and signaling pathways in biological systems. ANNs operate in the linear domain through a scalar multiplication, summation and a sigmoid activation function. Such behavior of ANNs is in contrast to biological systems which often exhibit logarithmic and power law input-output relations, where outcomes are dictated by relative fold-changes rather than absolute levels. Therefore, it was challenging to tightly connect model parameters of ANNs (weight, bias and detection threshold) to empirical biological parameters (Hill-coefficient, dissociation constant, protein level, and promoter basal level). In turn, synthetic biologists preferred to use other concepts of established engineering disciplines paradigms to program cellular behaviors. However, these approaches are often not suitable for achieving optimal behaviors in biological contexts. To overcome these challenges, we first redefine the basic computational structure of genetic processing units using a logarithmic transformative version of ANNs. Then, we experimentally implemented principles of ANNs in Escherichia coli to build a network that can take multiple decisions based on the fold change of a single input (e.g. analog-to-digital converter), and to study certain learning properties of genetically encoded functions (e.g. gradient descent). Realizing neural-like computing in individual cells marks an important step towards building synthetic biological systems that are adaptive via cellular machine learning, with implications for biotechnology and therapeutic applications.