A method for predicting conversion rates of amide bond molecular synthesis based on machine learning

Abstract

The application discloses a kind of based on machine learning prediction amide bond molecular synthesis conversion rate method, belong to chemical technical field.The application is based on high-throughput experiment collection reaction conversion rate data, establishes data set, and utilizes quantum chemistry calculation means to extract the feature descriptor of reaction related molecule, and the model after training finds that random forest model only needs a small amount of data to reach good prediction accuracy.Through the prediction of unknown aromatic amine molecular conversion rate in amidation reaction, it is found that a small amount of relevant reaction data can be added in the training set of random forest model to significantly improve the prediction performance of unfamiliar data.This has important guiding significance for exploring the synthesis of amide bond and the future application of machine learning to the field of chemical synthesis.

Description

Technical Field

[0001] This invention belongs to the field of chemical technology, specifically relating to the synthesis of amide bond molecules in organic chemistry and a method for predicting the conversion rate of amide bond molecule synthesis based on machine learning. It is used to explore the important influencing factors of amide bond synthesis and how to use machine learning to quickly apply to the field of chemical synthesis. Background Technology

[0002] Compounds containing amide bonds have been widely studied and applied in chemistry, materials science, and biopharmaceuticals. Examples include proteins and DNA in biomolecules; polyacrylamide and nylon (polyamide) in polymer materials; and penicillin and acetaminophen in drug synthesis. The unique resonance structure of amides allows them to exhibit high stability under acidic, alkaline, and high-temperature conditions. In recent years, many anti-COVID-19 drug molecules also contain multiple amide bonds, such as Nirmatrelvir, Molnupiravir, and Azvudine. Amide bond synthesis has become one of the most frequently occurring reactions in drug synthesis. However, amidation reactions still face challenges such as high production costs, low atom utilization, and variable reaction conditions.

[0003] With the rapid development of computing power and artificial intelligence algorithms, machine learning is becoming an increasingly important digital technology. In recent years, machine learning has shown significant application prospects in the field of chemical molecular synthesis. For example, the Doyle group used data from 4608 Buchwald-Hartwig coupling reactions to train and test a machine learning model, and the trained random forest model could accurately predict the reaction yield; the Norquist group used 3955 unique complete reactions to train and test a machine learning model, and the trained support vector machine algorithm achieved a success rate of 89% in predicting the preparation strategy of selenite crystals.

[0004] Therefore, using a machine learning model built on a small dataset to predict the conversion rate of amide bond molecule synthesis, describing the structural features of compound molecules in the amidation reaction using molecular descriptors, and studying the important factors affecting the synthesis of amide bonds are of great significance for guiding subsequent amidation experiments. It also provides a reference method for the application of machine learning in the field of chemistry. Summary of the Invention

[0005] Therefore, the purpose of this invention is to provide a method for predicting the conversion rate of amide bond synthesis reactions using machine learning based on a small sample dataset. The optimal machine learning model is selected to study the important factors affecting amide bond synthesis. By slightly adjusting the training set data of the model, the accuracy of predicting the reaction conversion rate of unknown aromatic amine molecules in amidation reactions can be significantly improved.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for predicting the conversion rate of amide bond molecules based on machine learning includes the following steps:

[0008] (1) Using high-throughput synthesis technology, amidation reaction was carried out to obtain reaction conversion data of organic amines and organic carboxylic acids under different coupling reagents and solvents, and a dataset was established.

[0009] (2) The physicochemical properties of the reaction substrate molecules, coupling reagents and solvents are plotted and calculated using quantum chemical calculation software, and feature descriptors are extracted and organized. Spartan is preferred for plotting software, and Python scripts are preferred for parameter extraction.

[0010] (3) The dataset is integrated by computer language. 70% of the dataset is used as the training set and the remaining data is used as the validation set. A machine learning model is established to predict the conversion rate of amide bond synthesis reaction using machine learning algorithm. The optimal prediction model is obtained. R language is preferred as the computer language.

[0011] (4) Using the optimal prediction model obtained in step (3), the relative importance of the feature descriptors obtained in step (2) is evaluated using the toolkit in the computer language software. The size of the MSE value is preferably used as the importance evaluation standard, and the key factors affecting the synthesis of amide bonds are finally obtained.

[0012] (5) The conversion rate of unknown aromatic amine molecules is predicted by using the optimal prediction model. By adding a certain amount of simple reaction data related to unknown aromatic amine molecules to the training set, the prediction performance of the conversion rate of unknown aromatic amine molecules is significantly improved.

[0013] Based on the above technical solution, further, the coupling reagent in step (1) includes onium salts and organophosphorus compounds, and the organic carboxylic acids include fatty acids and aromatic acids; the solvent is a commonly used organic solvent; and the organic amine includes commonly used aromatic amines.

[0014] Based on the above technical solution, the dataset is further collected not using data reported in the literature, but using experimental data obtained through high-throughput synthesis technology to avoid systematic errors caused by inconsistent experimental conditions in the literature data. The data obtained through experiments is more reliable and systematic.

[0015] Based on the above technical solution, further, in step (2), molecular descriptors, vibrational descriptors and atomic descriptors are extracted using quantum chemical calculation software, totaling 76 feature descriptors.

[0016] Based on the above technical solution, the machine learning algorithms mentioned in step (3) further include Line Model, k-Nearest Neighbor (kNN), Bayes Generalized Linear Model (Bayes GLM), Support Vector Machine (SVM), Neural Network (Neural Network) and Random Forest (Random Forest); the preferred prediction model is Random Forest (RandomForest).

[0017] Based on the above technical solutions, the software used for machine learning includes sumlime_text code editing, Spartan'14V1.1.4 molecular drawing and DFT calculation, Anaconda3 providing Python 3, R and RStudio for statistical calculation and plotting.

[0018] Based on the above technical solution, further, in step (3), 10% to 70% of the dataset is used as the training set and 30% to 90% of the dataset is used as the validation set.

[0019] Based on the above technical solution, further, the toolkit in the computer language software mentioned in step (4) includes the randomForest toolkit in the R language.

[0020] Based on the above technical solution, further, the key factors affecting amide bond synthesis in step (4) include the electrostatic charge of *N1 on the aromatic amine and the vibrational intensity of the aromatic amine molecule, the dipole moment and *C2 nuclear magnetic resonance shift of the organic solvent, and the E of the organic carboxylic acid. HOMO And ellipticity, etc.

[0021] Based on the above technical solution, further, the unknown aromatic amine molecules mentioned in step (5) include p-iodoaniline, naphthylamine, p-toluidine, and p-chloroaniline.

[0022] Based on the above technical solution, further, in step (5), the reaction data information of the unknown aromatic amine molecule with 3-6 fatty acids is added to the training set for model training to optimize the model's judgment of new molecular characteristics, and then the model is used to predict the reaction of the unknown aromatic amine molecule with aromatic carboxylic acids.

[0023] Based on the above technical solution, further, according to R 2 The prediction results are represented by the coefficient of determination (R²) and RMSE (root mean square error). 2 It indicates the extent to which the model explains the response outcome through various descriptor variables; RMSE indicates the overall magnitude of the deviation between the predicted and observed values.

[0024] The advantages of this invention over the prior art are as follows:

[0025] 1. In the prediction of amide bond synthesis, the quantum chemical calculation software Spartan, combined with a Python script, can quickly calculate, extract, and organize the characteristic descriptors of reacting molecules in the amidation reaction. Based on 1632 small sample data, six machine learning algorithms, namely linear model, Bayesian generalized linear model, support vector machine, K-nearest neighbor, neural network, and random forest, were selected for the first time to predict the conversion rate of aromatic amines in the amidation reaction to a certain extent, verifying the generality of machine learning algorithms. The random forest model showed the best prediction performance in the amidation reaction.

[0026] 2. By changing the ratio of the training set to the validation set, it was found that the random forest model can provide good prediction performance with less reaction data. Therefore, the random forest model can be used to provide suggestions or ideas for reaction optimization when there is a small sample of data.

[0027] 3. Analysis of the importance results provided by the random forest model explored the factors influencing amide bond synthesis. The random forest model was used to predict reactions of unknown aromatic amine molecules in amidation reactions. It was found that adding a small amount of relevant reaction data to the training set of the random forest model significantly improved the prediction performance for unfamiliar data. This invention established a machine learning model based on small sample experimental data, successfully predicting the factors influencing amide bond synthesis, providing a new research approach for future amidation experimental research and the application of machine learning in the field of chemistry. Attached Figure Description

[0028] To more clearly illustrate the embodiments of the present invention, the accompanying drawings involved in the embodiments will be briefly described below.

[0029] Figure 1 This is a flowchart of the machine learning process in the method of the present invention.

[0030] Figure 2 The molecular structure diagram of the amidation reaction provided in the embodiments of the present invention is shown.

[0031] Figure 3 The distribution map of 1632 reaction data used in the embodiments of the present invention is shown.

[0032] Figure 4 The image shows the prediction results and experimental results fitting plots obtained using different machine learning algorithms according to embodiments of the present invention.

[0033] Figure 5 A graph showing the importance of feature descriptors for a random forest model.

[0034] Figure 6The figure shows the fitting results of the random forest model for the conversion rate of unknown aromatic amines and the experimental results. In the figure, A represents the types of aromatic amines in the training set and the test set, B represents the prediction results of the random forest model for the out-of-sample data set, and C represents the performance evaluation of the optimized random forest model. Detailed Implementation

[0035] The present invention will be described in detail below with reference to the embodiments. However, the implementation of the present invention is not limited thereto. Obviously, the embodiments described below are only some embodiments of the present invention. For those skilled in the art, other similar embodiments can be obtained without creative effort and all fall within the protection scope of the present invention.

[0036] Example 1

[0037] 1) Prepare 13 organic aromatic amines, 12 organic carboxylic acids, 4 coupling reagents, and triethylamine solution in 20 ml vials, transfer to a 96-well plate for amidation reaction, and perform sample analysis using high-performance liquid chromatography (HPLC) to collect data. This example yielded 1632 reaction data points. The reacting molecules involved in the amidation reactions are listed in the appendix. Figure 2 As shown in the attached figure, we used a heatmap to display the conversion information of aromatic amines in all amidation reactions. Figure 3 As shown.

[0038] 2) The detailed operation process in this embodiment is referenced from: https: / / github.com / doylelab / rxnpredict. This calculation program first appeared in the published literature (Predicting reaction performance in C–Ncross-coupling using machine learning, Science, 2018, 360, 186-190). We are just users of this program.

[0039] 3) We plotted the molecular structure using Spartan and saved it as a spinput format file. We then wrote a Python script to submit the saved spinput file to Spartan DFT, resulting in a series of TXT files containing physicochemical descriptors. These molecular descriptors include orbital energies (E0, E0, E0, E0, E0). HOMO and E LUMO (e.g., electronegativity, dipole moment, ellipticity, and surface area); atomic descriptors include the electrostatic charge and nuclear magnetic displacement of atoms; vibrational descriptors include vibrational frequency and vibrational intensity, of which there are 30 descriptors describing aromatic amines, 21 descriptors describing organic carboxylic acids, 10 descriptors describing coupling reagents, and 15 descriptors describing organic solvents.

[0040] Aromatic amine descriptors (n=30): *C1 NMR shift, *C1 electrostatic charge, *C2 NMR shift, *C2 electrostatic charge, *C3 NMR shift, *C3 electrostatic charge, *C4 NMR shift, *C4 electrostatic charge, *N1 electrostatic charge, E HOMO E LUMO Vibration frequency of V1, vibration intensity of V1, vibration frequency of V2, vibration intensity of V2, vibration frequency of V3, vibration intensity of V3, vibration frequency of V4, vibration intensity of V4, vibration frequency of V5, vibration intensity of V5, vibration frequency of V6, vibration intensity of V6, dipole moment, electronegativity, hardness, molecular volume, molecular mass, ellipticity, surface area.

[0041] Carboxylic acid descriptors (n=21): *C1 NMR shift, *C1 electrostatic charge, *C2 NMR shift, *C2 electrostatic charge, *O1 electrostatic charge, *O2 electrostatic charge, E HOMO E LUMO Vibrational frequency of V1, vibrational intensity of V1, vibrational frequency of V2, vibrational intensity of V2, vibrational frequency of V3, vibrational intensity of V3, dipole moment, electronegativity, hardness, molecular volume, molecular mass, ellipticity, and surface area.

[0042] Coupling reagent descriptors (n=10): *O1 electrostatic charge, E HOMO E LUMO Dipole moment, electronegativity, hardness, molecular volume, molecular mass, ellipticity, and surface area.

[0043] Solvent descriptors (n=15): *C1 NMR shift, *C1 electrostatic charge, *C2 NMR shift, *C2 electrostatic charge, E HOMO E LUMO V1 vibrational frequency, V1 vibrational intensity, dipole moment, electronegativity, hardness, molecular volume, molecular mass, ellipticity, and surface area.

[0044] 4) 70% of the collected experimental data was used as the training set and 30% as the validation set. Six algorithms were selected for model training: linear model, k-nearest neighbor, support vector machine, Bayesian generalized linear model, neural network, and random forest. The generalization ability of machine learning was verified, and the results are shown in the appendix. Figure 4 As shown, the random forest model exhibited the best predictive performance among the six models trained.

[0045] 5) Given that the Random Forest model exhibits the best predictive performance, we evaluated the relative importance of the descriptors used to build the model using the randomForest toolkit in R. The importance analysis results provided by the model are as follows: Figure 5 As shown, the larger the increase in MSE value, the more important this descriptor is in the model building process.

[0046] 6) Based on the importance analysis results from the random forest model, it was found that the characteristic descriptors of aromatic amines, organic carboxylic acids, and organic solvents all have a significant impact on the conversion rate of the amidation reaction. The top six important descriptors in the predicted reaction results are the electrostatic charge of *N1 on aromatic amines and the vibrational intensity of aromatic amine molecules, the dipole moment and *C2 nuclear magnetic resonance shift of organic solvents, and the Ei of organic carboxylic acids. HOMO And ellipticity.

[0047] 7) Based on the random forest model and its importance analysis results, this embodiment attempts to evaluate the predictive performance of four unfamiliar aromatic amines using training set data of nine aromatic amines. We selected p-iodoaniline, naphthylamine, p-toluidine, and p-chloroaniline (e.g., based on structural similarity) according to the principle of structural similarity. Figure 6 As shown in Figure A), the prediction results are as follows: Figure 6 As shown in B, through Figure 6 B. We found that the random forest model could not effectively predict the conversion rates of these four unfamiliar aromatic amines and did not achieve the expected results.

[0048] 8) We attempted to incorporate the reaction data of these four unfamiliar aromatic amines with six fatty acids into the training dataset for model training to optimize the random forest model's judgment of new molecular features. Then, we used this model to predict the reactions of these four aromatic amine molecules with aromatic carboxylic acids. The results are as follows: Figure 6 As shown in Figure C, the comparison of the results before and after demonstrates that this method can effectively predict the conversion rate of the new molecule to the aromatic carboxylic acid, and the model's prediction performance for each molecule is better than that of the previous model.

[0049] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the conversion rate of amide bond molecules based on machine learning, characterized in that, Includes the following steps: (1) Amidation reaction was carried out using high-throughput synthesis technology to obtain reaction conversion data of organic amines and organic carboxylic acids under different coupling reagents and solvents, and a dataset was established; (2) Physicochemical property parameters of reaction substrate molecules, coupling reagents and solvents were plotted and calculated using quantum chemical calculation software, and feature descriptors were extracted and organized by the program; (3) The dataset is integrated by computer language. 70% of the dataset is used as the training set and 30% as the validation set. A machine learning model is established to predict the conversion rate of amide bond synthesis reaction by using machine learning algorithm, and the optimal prediction model is obtained. (4) Using the optimal prediction model obtained in step (3), the relative importance of the feature descriptors obtained in step (2) is evaluated using the toolkit in the computer language software to obtain the key factors affecting the synthesis of amide bonds; (5) The conversion rate of unknown aromatic amine molecules is predicted by using the optimal prediction model. By adding a certain amount of simple reaction data related to unknown aromatic amine molecules to the training set, the prediction performance of the conversion rate of unknown aromatic amine molecules is significantly improved.

2. The method according to claim 1, characterized in that, The coupling reagents mentioned in step (1) include onium salts and organophosphorus compounds, and the organic carboxylic acids include fatty acids and aromatic acids; the solvents are commonly used organic solvents; and the organic amines include commonly used aromatic amines.

3. The method according to claim 1, characterized in that, In step (2), quantum chemical calculation software is used to extract molecular descriptors, vibrational descriptors, and atomic descriptors.

4. The method according to claim 1, characterized in that, The machine learning algorithms mentioned in step (3) include linear models, k-nearest neighbors, Bayesian generalized linear models, support vector machines, neural networks, and random forests.

5. The method according to claim 1, characterized in that, The software used for machine learning includes the sumlime text code editor, Spartan'14 V1.1.4 for molecular plotting and DFT calculations, and Anaconda3 for statistical calculations and plotting using Python 3, R, and RStudio.

6. The method according to claim 1, characterized in that, The toolkits in the computer language software mentioned in step (4) include the randomForest toolkit in computer languages.

7. The method according to claim 1, characterized in that, The key factors affecting amide bond synthesis mentioned in step (4) include the electrostatic charge of *N1 on the aromatic amine and the vibrational intensity of the aromatic amine molecule, the dipole moment and *C2 nuclear magnetic resonance shift of the organic solvent, and the E1 of the organic carboxylic acid. HOMO And ellipticity.

8. The method according to claim 1, characterized in that, The unknown aromatic amine molecules mentioned in step (5) include iodoaniline, naphthylamine, p-toluidine, and p-chloroaniline.

9. The method according to claim 1, characterized in that, In step (5), the reaction data of the unknown aromatic amine molecule with 3-6 fatty acids are added to the training set for model training to optimize the model’s judgment of new molecular characteristics. Then, the model is used to predict the reaction of the unknown aromatic amine molecule with aromatic carboxylic acids.

Images