Department of Industrial and Systems Engineering, Dongguk University-Seoul, Seoul, 04620, South Korea.
Data Science Laboratory (DSLAB), Dongguk University-Seoul, Seoul, 04620, South Korea.
Sci Rep. 2023 Mar 2;13(1):957. doi: 10.1038/s41598-022-25701-5.
The water solubility of molecules is one of the most important properties in various chemical and medical research fields. Recently, machine learning-based methods for predicting molecular properties, including water solubility, have been extensively studied due to the advantage of effectively reducing computational costs. Although machine learning-based methods have made significant advances in predictive performance, the existing methods were still lacking in interpreting the predicted results. Therefore, we propose a novel multi-order graph attention network (MoGAT) for water solubility prediction to improve the predictive performance and interpret the predicted results. We extracted graph embeddings in every node embedding layer to consider the information of diverse neighboring orders and merged them by attention mechanism to generate a final graph embedding. MoGAT can provide the atomic-specific importance scores of a molecule that indicate which atoms significantly influence the prediction so that it can interpret the predicted results chemically. It also improves prediction performance because the graph representations of all neighboring orders, which contain diverse range of information, are employed for the final prediction. Through extensive experiments, we demonstrated that MoGAT showed better performance than the state-of-the-art methods, and the predicted results were consistent with well-known chemical knowledge.
分子的水溶性是各种化学和医学研究领域中最重要的性质之一。最近，由于有效降低计算成本的优势，基于机器学习的方法已被广泛用于预测分子性质，包括水溶性。尽管基于机器学习的方法在预测性能方面取得了重大进展，但现有的方法在解释预测结果方面仍然存在不足。因此，我们提出了一种新颖的多阶图注意网络（MoGAT）来进行水溶性预测，以提高预测性能并解释预测结果。我们在每个节点嵌入层中提取图嵌入，以考虑不同阶数的信息，并通过注意力机制对其进行合并，以生成最终的图嵌入。MoGAT 可以提供分子的原子特定重要性得分，指出哪些原子对预测有显著影响，从而可以从化学角度解释预测结果。它还通过利用所有邻近阶的图表示（包含不同范围的信息）来提高预测性能，从而进行最终预测。通过广泛的实验，我们证明了 MoGAT 比最先进的方法具有更好的性能，并且预测结果与众所周知的化学知识一致。
