Prediction of Nano–Bio Interactions through Convolutional Neural Network Analysis of Nanostructure Images

Artificial intelligence approaches, such as machine learning and deep learning, may predict nano–bio interactions. However, such a prediction is now hindered by the paucity of suitable nanodescriptors with applicable nanostructure annotation methods. Inspired by face recognition technology, we have developed a novel nanostructure annotation method to automatically convert nanostructures to images for convolutional neural network modeling. In this operation, nanostructure features were directly learned from nanoparticle images without complicated nanodescriptor calculations. The constructed convolutional neural network models were successfully used to predict physicochemical properties (i.e., logP and zeta potential) and biological activities (i.e., cellular uptake and protein adsorption) of 147 unique nanoparticles, including 123 gold nanoparticles, 12 platinum nanoparticles, and 12 palladium nanoparticles. Our nanostructure diversity and wide distribution of experimental values are beneficial for building predictive deep learning models. The deep learning models provide highly accurate predictions with all determination coefficients (R2) higher than 0.68 for both cross validation and external prediction. In addition, the constructed model is explainable because we can visualize how it learns from the class activation map. This approach enables a much more efficient end-to-end deep learning modality suitable for design of next generation nanomaterials.