Synthetic neuromorphic computing in living cells

Biological regulatory networks in cells and neuronal networks employ complex circuit topologies with highly interconnected nodes to perform sophisticated information processing. Despite the complexity of neuronal networks, their information processing and computational capabilities can be recapitulated using simplified models comprising repeated connected nodes, e.g., perceptrons, with decision-making capabilities and flexible weighted links. Here, we argue that analogous to their revolutionary impact on computing, neuro-inspired models can similarly transform synthetic gene circuit design in a manner that is reliable, efficient in resource utilization, and can be readily reconfigurable for new tasks. We introduce neuromorphic design for synthetic gene circuits by first defining the perceptgene, a perceptron that computes in the logarithmic domain, which enables efficient implementation of artificial neural networks in the cellular milieu. Working in Escherichia coli cells, we experimentally demonstrated logarithmic scale analog multiplication using a single perceptgene. We modified perceptgene parameters (weights and biases) to create devices that compute a log-transformed negative rectifier encoding the minimum operation, log-transformed positive rectifier encoding the maximum operation, and log-transformed average of analog inputs. We then created multi-layer perceptgene circuits that compute a majority function, perform analog-to-digital conversion, and implement a ternary switch. Experimental and theoretical analysis showed that our approach enables circuit optimization via artificial intelligence algorithms such as gradient descent and backpropagation. Realizing neural-like computing in the noisy resource-limited environments of individual cells marks an important step towards synthetic biological systems that can be engineered effectively via supervised ANN optimization algorithms.