Toxicity Prediction Study of Small Molecules Based on Graph Attention Networks
Yaxin Tong, Qiwei Guo, Jianye Cui, Xin Peng · 2023
Small molecule toxicity prediction has a very important value in many fields such as drug design and chemical biology, the establishment of small molecule toxicity prediction model can help researchers focus on the research focus, improve the efficiency of research and development, in the future will certainly be widely used for people's scientific research and production and life convenience. However, traditional machine learning methods cannot directly use molecules as inputs, and how to accurately extract molecular features is also a difficult thing. To solve the problems above, this study proposes a model based on a graph attention network, by introducing the attention mechanism also mines the connection relationship between atoms and atoms, and aggregates the feature information of other nodes weighted by the attention coefficient. The algorithm is validated on the publicly available dataset Tox21 with AUC values of 0.958, 0.829, and 0.823 on the training, validation, and test sets. Compared with the traditional convolutional neural network-based model, the AUC value is improved by 0.5% on average, which better predicts molecular toxicity.