Molecular image-convolutional neural network (CNN) assisted QSAR models for predicting contaminant reactivity toward OH radicals: Transfer learning, data augmentation and model interpretation
Molecular image-convolutional neural network (CNN) assisted QSAR models for predicting contaminant reactivity toward OH radicals: Transfer learning, data augmentation and model interpretation
复制标题
DOI：
10.1016/j.cej.2020.127998
复制
发表时间：
2021-03
影响因子：
15.1
通讯作者：
Shifa Zhong;Jiajie Hu;X. Yu;Huichun Zhang
中科院分区：
文献类型：
--
作者：
Shifa Zhong;Jiajie Hu;X. Yu;Huichun Zhang
在这项研究中，我们使用分子图像作为有机化合物的代表，并将其与卷积神经网络（CNN）相结合，以开发定量结构活性关系（QSAR），用于预测化合物对OH自由基的速率常数。我们应用迁移学习和数据增强来训练分子图像-CNN模型，并使用梯度加权类激活映射（Grad-CAM）方法来解释它们。结果表明，数据增强和迁移学习可以有效地提高模型的鲁棒性和预测性能，测试数据集上的均方根误差（RMSE）值（RMSE测试）从（0.395-0.45）至（0.284-0.339）在应用数据增强之后，应用迁移学习后，训练数据集上的RMSE（RMSE训练）从（0.452-0.592）降低到（0.123-0.151）。所获得的分子图像-CNN模型显示出与基于分子指纹的模型（RMSEtest0.30-0.35）相当的预测性能（RMSEtest0.284-0.339）。Grad-CAM解释表明，分子图像-CNN模型正确地选择了图像中的分子特征，并确定了影响反应性的关键官能团。适用性域分析表明，分子图像CNN模型比基于分子指纹的模型具有更广泛的适用性域，并且可以可靠地预测与训练数据集中的化合物的最大相似性超过0.85的任何新化合物的反应性。这项研究表明，分子图像CNN是开发环境应用QSAR的新工具，可用于构建可做出有意义预测的可信模型。
In this study, we used molecular images as a representation for organic compounds and combined them with a convolutional neural network (CNN) to develop quantitative structure-activity relationships (QSARs) for predicting compound rate constants toward OH radicals. We applied transfer learning and data augmentation to train molecular image-CNN models and the Gradient-weighted Class Activation Mapping (Grad-CAM) method to interpret them. Results showed that data augmentation and transfer learning can effectively enhance the robustness and predictive performance of the models, with the root-mean-square-error (RMSE) values on the test dataset (RMSEtest) decreasing from (0.395–0.45) to (0.284–0.339) after applying data augmentation, and the RMSE on the training dataset (RMSEtrain) decreasing from (0.452–0.592) to (0.123–0.151) after applying transfer learning. The obtained molecular image-CNN models showed comparative predictive performance (RMSEtest0.284–0.339) with the molecular fingerprint-based models (RMSEtest0.30–0.35). Grad-CAM interpretation showed that the molecular image-CNN models correctly chose the molecular features in the images and identified key functional groups that influenced the reactivity. The applicability domain analysis showed that the molecular image-CNN models have a broader applicability domain than molecular fingerprints-based models and the reactivity of any new compounds with a maximum similarity of over 0.85 to the compounds in the training dataset can be reliably predicted. This study demonstrated that molecular image-CNN is a new tool to develop QSARs for environmental applications and can be used to build trustful models that make meaningful predictions.