A DNA-Level Convolutional Neural Network Based on Strand Displacement Reaction for Image Recognition

Over the past few decades, convolutional neural network (CNN) has found broad applications in image recognition. Nevertheless, the operational environment of CNN is facing significant challenges, including poor parallelism and high energy consumption due to the explosive growth in data scale and the rise of multi-modal data. DNA-based molecular computing systems, which are known for their high parallelism and low energy consumption, have the potential to provide an ideal operating platform for CNN. Here, we employed strand displacement reaction (SDR) to construct a DNA-level CNN. Firstly, we integrated DNA-based weighted-sum module, subtraction activation module, and reporter module using SDR to implement CNN, and designed the weighted shared boxes to perform the function of convolutional kernel. The feasibility and parallelism of the proposed DNA-level CNN were verified by simultaneous recognition of three different categories of images, including handwritten numbers, letters and Chinese characters. We further applied the DNA-level CNN to parallel recognition of multi-class medical images. Visual DSD software was employed to perform all simulation experiments. The efficient parallel computing characteristics of DNA-level CNN provides a new approach for image recognition, and show great potential in multi-class medical image recognition.