Xinghai Li , Zhisen Wu , Lijing Zhang, Shengyang Tao
{"title":"Machine learning enables the prediction of amide bond synthesis based on small datasets","authors":"Xinghai Li , Zhisen Wu , Lijing Zhang, Shengyang Tao","doi":"10.3866/PKU.WHXB202309041","DOIUrl":null,"url":null,"abstract":"<div><div>Machine learning (ML) is progressively revealing notable advantages in chemical synthesis. However, the limited output of experimental data from traditional methods poses a bottleneck, impeding the widespread adoption of machine learning. Data from literature often leads to overly optimistic predictions, and obtaining thousands of experimental data points through experiments remains a substantial challenge. Using a small dataset of experimental data, we illustrated that machine learning algorithms can reliably predict the conversion rate of amide bond synthesis. We gathered hundreds of experimental data points for 9 aromatic amines and 12 organic acids using various coupling reagents and solvents in a 96-well plate high-throughput experimental setup. Subsequently, we derived 76 feature molecular descriptors from quantum chemical calculations and utilized them as inputs for training the machine learning model. Despite the inherent limitation of low data volume, the random forest algorithm demonstrated outstanding predictive performance (<em>R</em><sup>2</sup> > 0.95). Through comprehensive analysis of the reaction process employing importance analysis, shapley additive explanations (SHAP), and accumulated local effects (ALE) methods, we delved into the important factors influencing the reaction conversion rate. In predicting the conversion rate of unknown aromatic amine molecules, we discovered that incorporating a small amount of unknown molecule-related reaction data into the training set effectively enhances the model's predictive performance, even with a small dataset. By comparing models trained on different molecular descriptors such as density functional theory (DFT) and one-hot encoding, we validated the efficacy of adjusting the training set to improve prediction results. This study utilized a multitude of chemically meaningful feature descriptors and achieved more effective prediction results through multidimensional data analysis, offering valuable insights for machine learning-assisted chemical synthesis research in small datasets. In the near future, machine learning is poised to drive the intelligent development of organic chemistry.</div></div>","PeriodicalId":6964,"journal":{"name":"物理化学学报","volume":"41 2","pages":"Article 100010"},"PeriodicalIF":10.8000,"publicationDate":"2025-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"物理化学学报","FirstCategoryId":"92","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S1000681824000109","RegionNum":2,"RegionCategory":"化学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"CHEMISTRY, PHYSICAL","Score":null,"Total":0}
引用次数: 0
Abstract
Machine learning (ML) is progressively revealing notable advantages in chemical synthesis. However, the limited output of experimental data from traditional methods poses a bottleneck, impeding the widespread adoption of machine learning. Data from literature often leads to overly optimistic predictions, and obtaining thousands of experimental data points through experiments remains a substantial challenge. Using a small dataset of experimental data, we illustrated that machine learning algorithms can reliably predict the conversion rate of amide bond synthesis. We gathered hundreds of experimental data points for 9 aromatic amines and 12 organic acids using various coupling reagents and solvents in a 96-well plate high-throughput experimental setup. Subsequently, we derived 76 feature molecular descriptors from quantum chemical calculations and utilized them as inputs for training the machine learning model. Despite the inherent limitation of low data volume, the random forest algorithm demonstrated outstanding predictive performance (R2 > 0.95). Through comprehensive analysis of the reaction process employing importance analysis, shapley additive explanations (SHAP), and accumulated local effects (ALE) methods, we delved into the important factors influencing the reaction conversion rate. In predicting the conversion rate of unknown aromatic amine molecules, we discovered that incorporating a small amount of unknown molecule-related reaction data into the training set effectively enhances the model's predictive performance, even with a small dataset. By comparing models trained on different molecular descriptors such as density functional theory (DFT) and one-hot encoding, we validated the efficacy of adjusting the training set to improve prediction results. This study utilized a multitude of chemically meaningful feature descriptors and achieved more effective prediction results through multidimensional data analysis, offering valuable insights for machine learning-assisted chemical synthesis research in small datasets. In the near future, machine learning is poised to drive the intelligent development of organic chemistry.