GraphPLA：一种预测蛋白质–配体结合亲和力的图神经网络方法

GraphPLA: A Graph Neural Networks Method to Predict Protein-Ligand Binding Affinity

摘要: 蛋白质–配体结合亲和力是蛋白质与配体相互作用强度的一个重要指标。准确预测蛋白质–配体结合亲和力对于发现药物的新应用至关重要。迄今为止，已经开发了许多计算技术来预测结合亲和力；然而，这些技术中的一部分需要并不常见的蛋白质三维结构，一些方法还将配体表示为不是分子的适当表示的SMILES串。为了避免这些问题，本文开发了一种名为GraphPLA的新模型，使用具有直接结合配体的独特特征的蛋白质结合口袋作为局部输入特征。还使用扩展卷积来捕捉蛋白质的多尺度远程相互作用，图神经网络来学习配体的图表示。实验结果表明，GraphPLA的RMSE为1.388，MAE为1.118，R为0.795，SD为1.345，CI为0.796，优于目前最先进的预测方法，可以有效预测蛋白质–配体结合亲和力。

Abstract: Protein-ligand binding affinity is an important indicator of the strength of protein-ligand interactions. Accurately predicting protein-ligand binding affinity is crucial for discovering new applications for drugs. To date, many computational techniques have been developed to predict binding affinity. However, some of these technologies require uncommon protein three-dimensional structures, and some methods also represent ligands as SMILES strings that are not appropriate representations of molecules. To avoid these issues, this paper develops a new model called GraphPLA, which uses protein binding pockets with unique features of directly binding ligands as local input features. Extended convolution is also used to capture multi-scale remote interactions of proteins, and a graph neural network is used to learn the graph representation of ligands. The experimental results show that the RMSE of GraphPLA is 1.388, MAE is 1.118, R is 0.795, SD is 1.345, and CI is 0.796, which is superior to the most advanced prediction methods and can effectively predict protein-ligand binding affinity.