【作者】 张宇；

【导师】 李燕；

【作者基本信息】 大连理工大学 ， 化学工艺， 2024， 硕士

【摘要】 表皮生长因子受体（EGFR）突变被确定为非小细胞肺癌（NSCLC）的驱动突变,但耐药性是关键问题。设计新的表皮生长因子受体抑制剂是一个迫切的需求。在本文中,我们构建了分类模型和回归模型来预测下一代表皮生长因子受体抑制剂。分类模型采用了八种机器学习方法和两种筛选方法。结果表明,在外部验证集中,支持向量机（SVM）模型的准确率为95.5%,受试者特征曲线（ROC）值为92.4%,马修斯系数（MCC）值为84.7%。在回归模型中使用递归特征消除（RFE）来选择描述符。建立了单一模型和组合模型来预测活性化合物的PIC₅₀值。对于表皮生长因子受体,随机森林与支持向量机结合模型（RF-RFE-SVM）的性能最佳,R²=0.93。通过分析特征的SHAPLEY值,通过模型解释,我们得到了新一代抑制剂可能所具备的结构特征,然后筛选出的28个化合物被用于构建药效谱和分子对接。我们得到了28种不同骨架所共有的三个药效团结构,通过分子对接,我们进一步研究了化合物与EGFR之间的相互作用和一些重要的氨基酸残基。在本文中,机器学习方法有效地预测了p IC₅₀值,并准确地区分了EGFR抑制剂和非抑制剂,加速了下一代EGFR抑制剂的开发。由于hERG心脏毒性是候选药物能否上市的重要标准,在本文中,我们搭建了hERG毒性预测模型,我们使用了3种不同指纹（ECFP、MACCS以及药效团）和片段相似性来作为特征输入,通过已确定的分子拆分为片段,探究片段对于分子毒性的影响,结合机器学习模型以及深度学习模型共搭建了24个分类模型。结果表明,轻量梯度提升机（LGBM）模型的准确率到达了85.8%,ROC值为85.5%,MCC值为71%。在本文中,我们完成了后续对于药物分子hERG毒性的准确预测,有助于筛选药物分子,对药物研发具有一定的意义。更多还原

【Abstract】 Epidermal growth factor receptor（EGFR）mutations are identified as driver mutations in non-small cell lung cancer（NSCLC）,but drug resistance is the key issue.Designing the new EGFR inhibitors is a crying need.In this work,classification models and regression models were constructed to predict the next-generation EGFR inhibitors.Eight machine-learning approaches and two screening methods were used in classification models.The results showed that the accuracy of SVM model was 95.5%,the ROC value was 92.4%,and the MCC value was 84.7%in the external validation set.Recursive feature elimination（RFE）was used to select the descriptors in regression models.Single and combined models were built to predict the PIC50 value of the activity compounds.As for EGFR,RF-RFE-SVM showed the best performance with R²=0.93.By analyzing the SHAPLEY value of features,through model interpretation,we obtained the possible structural features of the new-generation inhibitors,and then the 28 compounds screened were used to construct pharmacophore models and molecular docking.We obtained three pharmacophore structures shared by 28 different skeletons,and by molecular docking,we further explored the compounds’interactions with EGFR and some important amino acid residues.In this study,the p IC₅₀ values were effectively predicted and EGFR inhibitors and non-inhibitors were exactly distinguished by the machine learning approach,accelerating the development of the next-generation EGFR inhibitors.Since hERG cardiotoxicity is an important criterion for the marketability of drug candidates,in the paper,we built hERG toxicity prediction models.We used three different fingerprints（ECFP,MACCS,and pharmacophore）and fragment similarity as feature inputs to explore the effect of fragments on molecular toxicity by splitting the identified molecules into fragments,combining machine learning models and deep learning models,a total of 24classification models were constructed.The results show that the accuracy of the LGBM model reaches 85.8%,the ROC value is 85.5%,and the MCC value is 71%.In the paper,we accomplished the subsequent accurate prediction of hERG toxicity of drug molecules,which is helpful for screening drug molecules and has certain significance for drug development.更多还原