Reusable Noncomplementary DNA-Based Neural Network

Neural network computation is a cornerstone of modern
artificial
intelligence, with electronic software-based approaches achieving
widespread success due to their ability to enable continuous, iterative
learning on the same platform. DNA-based neural networks, with their
potential advantages in versatility, scalability, and energy efficiency,
offer a promising alternative to traditional systems. However, despite
significant advancements in pattern recognition and algorithm accuracy,
current DNA-based neural networks, relying on the complementary pairing
of DNA nucleobases, suffer from the nonreusability of their computing
materials. This limitation not only raises operational costs but also
restricts their capacity for implementing learning mechanisms. Here,
we introduce an unprecedented noncomplementary DNA-based perceptron
(NCP) computation strategy, marking the first successful demonstration
of a reusable DNA-based neural network. We present a “tagging”
strategy to facilitate the scaling-up of noncomplementary DNA-based
neural network. We show that 4-bit molecular pattern recognition can
be simply achieved through strand-displacement reactions between four
input strands and four pairs of noncomplementary DNA duplexes in the
NCP, with weighting values modulated by duplex concentrations. Furthermore,
a noncomplementary “winner-take-all” module enables
decision-making, as demonstrated in an “I Spy” game
task. Most importantly, by utilizing removable input strands (lipid–oligonucleotide
conjugates), our NCP-based neural network enables reliable multicycle
computations, overcoming the critical reusability challenge in DNA-based
neural network computation. This work pioneers reusability in DNA-based
neural networks, offering a practical path to molecular computing
systems with learning capabilities.