Graph attention convolutional neural network model for chemical poisoning of honey bees' prediction.

Abstract The impact of pesticides on insect pollinators has caused worldwide concern. Both global bee decline and stopping the use of pesticides may have serious consequences for food security. Automated and accurate prediction of chemical poisoning of honey bees is a challenging task owing to a lack of understanding of chemical toxicity and introspection. Deep learning (DL) shows potential utility for general and highly variable tasks across fields. Here, we developed a new DL model of deep graph attention convolutional neural networks (GACNN) with the combination of undirected graph (UG) and attention convolutional neural networks (ACNN) to accurately classify chemical poisoning of honey bees. We used a training dataset of 720 pesticides and an external validation dataset of 90 pesticides, which is one order of magnitude larger than the previous datasets. We tested its performance in two ways: poisonous versus non-poisonous and GACNN versus other frequently-used machine learning models. The first case represents the accuracy in identifying bee poisonous chemicals. The second represents performance advantages. The GACNN achieved ~6% higher performance for predicting toxic samples and more stable with ~7% Matthews Correlation Coefficient (MCC) higher compared to all tested models, demonstrating GACNN is capable of accurately classifying chemicals and has considerable potential in practical applications. In addition, we also summarized and evaluated the mechanisms underlying the response of honey bees to chemical exposure based on the mapping of molecular similarity. Moreover, our cloud platform ( http://beetox.cn ) of this model provides low-cost universal access to information, which could vitally enhance environmental risk assessment.