Predicting the Fluorescence Properties of Hairpin-DNA-Templated Silver Nanoclusters via Deep Learning

DNA-templated silver nanoclusters (AgNCs) are widely used as fluorescent probes in various fields owing to their structural stability and fluorescence tunability. The selection of suitable DNA sequences and synthesis conditions with a predictable method is therefore needed, which benefits the preparation of DNA-templated AgNCs with the desired properties and further extension of their application in other fields. Here, we develop and propose a deep learning protocol based on recurrent neural network (RNN) algorithms to predict the fluorescence properties of hairpin-DNA-templated AgNCs (H-AgNCs). Using the stem sequence of hairpin DNA and experimental parameters as input features, the RNN model can identify the number of fluorescence emission peak of H-AgNCs and the fluorescence color of single-peaked H-AgNCs (spH-AgNCs) with the training accuracies of 81.4% and 83.0%, respectively. For each color category, the DNA base distributions of the hairpin stem on the testing set are consistent with those of the training set, indicating the good generalization ability of the constructed model. In addition, a prediction accuracy of 56% is experimentally verified using 20 testing samples, which is 2.3 times higher than that of the random probability, indicating the potential uses of the model for experimental guidance.