A Reusable Non-Complementary-DNA-Based Neural Network

The emulation of intelligence across diverse domains of the human brain has spurred the development of neural network based artificial intelligence. The computation of the DNA-based neural network has recently emerged as a focal point of research due to its versatility, scalability, energy efficiency and potentially other huge benefits and implications, as compared to electronic computation. Despite notable advancements, the development of the current DNA neural networks, based on the complementary pairing of DNA nucleobases, is restricted by the lack of reusability of the DNA computing materials, one of key bottlenecks impeding their progression towards neural network learning and evolution. As a result, even the state-of-the-art DNA neural network computations are limited to one-time use currently. Here we report the design of an unprecedented, reusable DNA based non-complementary perceptron (NCP) strategy that implements thresholding and weighted summation functions like neurons and the corresponding neural networks capable of 4-bit molecular pattern recognition. To facilitate the scaling-up of the non-complementary circuits, a modulated concept employing “tagging” domains is also coined. We demonstrate that non-complementary “winner-take-all” circuit can be rationally constructed with a non-complementary annihilator strand. Such NCP based neural network architecture is capable of 4-bit pattern recognition, evidenced by its success in playing the “I Spy” game. Most importantly, when removable input strands (lipid-oligonucleotide conjugates) are utilized, this NCP-based pattern recognition neural network shows high fidelity in multiple-cycle computing. This suggests a reusable DNA based NCP computation strategy as a potential conceptual breakthrough for the design of next-generation DNA computers.