Drug ADMET Prediction Method Based on Improved Graph Convolution Neural Network
Haoliang Xiao, Xiangyang Chen · 2022
To improve the performance of the determination model for evaluating drug absorption, distribution, metabolism, exclusion, and toxicity (ADMET), a drug ADMET prediction model based on graph convolution network is proposed from the two chemical properties of drug solubility and toxicity. At the same time, a multilayer perceptron and an attention mechanism are introduced to improve the model by using special information such as features in molecular bonds and the difference in interaction strength between adjacent atoms. The error of the improved model is reduced by about 24% and 25% on average on two public data sets of drug solubility, and the accuracy of three public data sets of drug toxicity is increased by about 7.8% on average and 9.6 % at most. Experiments show that the good performance of this method can be used for early drug ADMET prediction, thus providing a reference for computer-aided drug design.