DNA-based neural networks are emerging as a biocompatible and energy-saving computing paradigm, but they have long been constrained by the lack of hardware reusability. In this work, we establish an unprecedented noncomplementary DNA-based neural network that relies exclusively on reversible strand-displacement reactions, providing an inherently reusable computing framework capable of multibit pattern recognition. We present how mismatched DNA strands can be programmed into scalable noncomplementary perceptrons (NCPs) via strategically positioning mismatched bases and incorporating short “tagging” domains. We invented a noncomplementary “winner-take-all” module that enables multiway decision-making from NCP outputs. Moreover, by introducing lipid-oligonucleotide conjugates as removable inputs and bench-top column separation, we accomplished multicycle molecular pattern recognition using the noncomplementary neural network. This represents the first demonstration of a reusable DNA-based neural network. Our findings pave an avenue for molecular computing systems able to implement learning mechanisms and process dynamic biomedical information.
This work, titled "Reusable Noncomplementary DNA-Based Neural Network" was published in JACS.Chengjie Sun, a research assistant professor from the school of engineering at Westlake University, is the first author of this article, and Jianjun Cheng, a principle investigator from the school of engineering, is the corresponding author. This work was supported by the National Natural Science Foundation of China and“Pioneer”and“Leading Goose”R&D Program of Zhejiang.J.C.acknowledges the support of the New Cornerstone Investigator Program, New Cornerstone Science Foundation.
Original link: https://doi.org/10.1021/jacs.5c04886
