Hu Jinpeng, Yang Xin, Yao Chenggui, Zhang Mingxuan, Shen Shihang, Na Lijie, Zhao Qi
School of Computer Science and Software Engineering, University of Science and Technology Liaoning, Anshan, 114051, China.
College of Data Science, Jiaxing University, Jiaxing, 314000, China.
Interdiscip Sci. 2025 Jul 16. doi: 10.1007/s12539-025-00742-2.
Accurate mutagenicity prediction is crucial in drug discovery, but evaluating the mutagenicity of drug molecules is technically challenging, and experimental mutagenicity tests are both time-consuming and costly. In this study, we introduce a new deep learning predictive model called AMPred-MFG, designed to predict the mutagenicity of drug molecules. Our approach first combines multiple molecular fingerprint features with molecular graph features to obtain a comprehensive molecular representation. Next, considering the significance of molecular substructures on mutagenicity, we decompose the molecules into motifs to form motif-based graph. We use a graph transformer to extract features from these motif-based graphs and fuse motif graph features with molecular fingerprint features and molecular graph features to create the final molecular representation. Finally, we use a multilayer perceptron to classify the compounds into mutagens and non-mutagens. We evaluate the performance of AMPred-MFG through ten-fold cross-validation experiments and validate its robustness on external validation datasets. By comparing with several state-of-the-art algorithms, AMPred-MFG achieves the best results in terms of performance, with AUC value of 0.912, ACC of 0.835, SEN of 0.849, NPV of 0.811, PPV of 0.854, MCC of 0.665. In addition, we use attention scores to identify molecular fragments related to mutagenicity, highlighting the interpretability of AMPred-MFG. We believe that AMPred-MFG can act as a dependable tool for predicting mutagenicity, allowing for the assessment of both mutagens and non-mutagens during the early phases of drug development. AMPred-MFG is freely available at https://github.com/zhaoqi106/AMPred-MFG .
准确的致突变性预测在药物发现中至关重要，但评估药物分子的致突变性在技术上具有挑战性，并且实验性致突变性测试既耗时又昂贵。在本研究中，我们引入了一种名为AMPred-MFG的新型深度学习预测模型，旨在预测药物分子的致突变性。我们的方法首先将多个分子指纹特征与分子图特征相结合，以获得全面的分子表示。接下来，考虑到分子子结构对致突变性的重要性，我们将分子分解为基序以形成基于基序的图。我们使用图变换器从这些基于基序的图中提取特征，并将基序图特征与分子指纹特征和分子图特征融合，以创建最终的分子表示。最后，我们使用多层感知器将化合物分类为诱变剂和非诱变剂。我们通过十折交叉验证实验评估了AMPred-MFG的性能，并在外部验证数据集上验证了其稳健性。通过与几种最先进的算法进行比较，AMPred-MFG在性能方面取得了最佳结果，AUC值为0.912，ACC为0.835，SEN为0.849，NPV为0.811，PPV为0.854，MCC为0.665。此外，我们使用注意力分数来识别与致突变性相关的分子片段，突出了AMPred-MFG的可解释性。我们相信AMPred-MFG可以作为预测致突变性的可靠工具，允许在药物开发的早期阶段对诱变剂和非诱变剂进行评估。AMPred-MFG可在https://github.com/zhaoqi106/AMPred-MFG上免费获取。
