A method for predicting ratio shift value based on machine learning

Abstract

The present invention discloses a method for predicting a ratio transfer value based on machine learning, which collects data of compounds, developing agents and ratio transfer values, establishes a ratio transfer value prediction model through a machine learning method, can quickly and accurately predict the ratio transfer value curve of a target compound under a target developing agent system, and gives an optimal developing agent ratio scheme according to the curve, so that the ratio transfer value obtained in the experiment is not too large or too small, greatly reducing the dependence of thin-layer chromatography analysis technology on the experience of the experimenter, solving the problem of developing agent selection, avoiding boring repeated experiments, greatly improving the efficiency of polarity determination, and saving time and labor costs. The present invention introduces machine learning technology into the field of experimental chemistry, making it possible to obtain the ratio transfer value of a compound without experimental conditions, and has wide applications and important significance in the fields of experimental chemistry, drug synthesis and analysis.

Description

Technical Field

[0001] The present invention relates to the field of organic chemical polarity determination, and in particular to a method for predicting a ratio shift value based on machine learning. Background Art

[0002] In the field of organic chemistry, the polarity of compounds is often experimentally determined by thin layer chromatography, and the relative polarity is expressed by the relative transfer value (also known as the Rf value). However, thin layer chromatography requires a high level of experience and operation of the experimenter, especially the selection of a suitable developing agent so that the relative transfer value obtained is not too large or too small, which makes the determination of polarity a time-consuming and labor-intensive task that requires repeated trial and error. At the same time, a huge amount of polarity data is generated every day in chemical laboratories around the world, and these data have not been well utilized, resulting in repeated measurements and waste of resources.

[0003] With the development of computer technology, machine learning algorithms have been widely used in big data analysis tasks. Among them, methods such as XGBoost, LightGBM, and neural networks have outstanding performance in data regression and prediction, and have the advantages of fast, stable, and strong generalization capabilities. In the field of chemistry, due to the relatively high cost of experimental data acquisition, machine learning algorithms often perform poorly. Therefore, how to overcome the above problems and develop a ratio shift value prediction algorithm based on machine learning with limited data to quickly and accurately predict the ratio shift value of a compound under a given developing agent from its structure and properties is an urgent problem to be solved. Summary of the invention

[0004] The purpose of the present invention is to provide a method for predicting the relative transfer value based on machine learning to solve the problems existing in the above-mentioned prior art, and to train a relative transfer value prediction model using limited data, so as to quickly and accurately predict the relative transfer value of an organic compound under a given developing agent. To achieve the above-mentioned purpose, the present invention provides the following technical solutions:

[0005] A method for predicting a ratio shift value based on machine learning comprises the following steps:

[0006] A. Data collection and cleaning: Collect compound, developer and ratio shift data and clean them to obtain a complete, non-duplicate, and outlier-free polarity data set.

[0007] B. Data preprocessing: Preprocess the cleaned data, including: digitizing the molecular formula of the compound, characterizing the molecular structure and properties of the compound through molecular fingerprints, dipole moments and molecular descriptors; using the developer ratio vector to represent the developer system and ratio; the information vector of each preprocessed data contains multiple dimensions, thereby generating an input information matrix.

[0008] C. Model training: The preprocessed data is divided into a training set and a validation set. The XGBoost, LightGBM, and neural network algorithms constrained by the Sigmoid function are used for training on the training set. The three trained models are integrated into an integrated model to predict the ratio shift value. During the training process, the validation set is used to adopt the cross-validation method to prevent overfitting.

[0009] D. Model prediction: Use the specific transfer value prediction model trained in step C to predict the specific transfer value curve of the target compound in the target developing agent system to provide guidance for experimental determination.

[0010] E. Feedback update: automatically update and correct the model based on experimental feedback data.

[0011] The data collected in the above step A include: the chemical formula of the compound, the developing solvent system and ratio used, and the ratio shift value of the compound under the developing solvent system and ratio; the method of cleaning the data includes: a) if any one of the chemical formula of the compound, the developing solvent system and ratio used, and the ratio shift value under the developing solvent system and ratio is missing, then remove the data; b) remove data with illegal ratio shift values ​​determined experimentally, that is, data with ratio shift values ​​greater than 1 or less than 0; c) remove duplicate data and retain only one of them or retain the average of the duplicate data.

[0012] In the above step B, the molecular fingerprint includes MACCSkeys fingerprint, Morgan fingerprint, etc., which can be selected according to the situation; the molecular descriptor includes molecular weight, topological polar surface area, number of rotatable bonds, number of hydrogen bond donors, number of hydrogen bond acceptors, lipid-water partition coefficient value, etc.; the dipole moment refers to the product of the distance between the positive and negative charge centers and the charge carried by the charge center. If N kinds of developing agents are involved, an N-dimensional vector is used to represent the developing agent system and ratio.

[0013] In the above step C, some data can be randomly selected as the test set, and the prediction ability of the model is determined by the coefficient of determination R of the model on the test set. 2 The calculation formula is as follows:

[0014]

[0015] Where N is the number of samples in the test set, y i is the true shift value, is the mean of the true ratio shift values, is the predicted ratio shift value; R 2 The closer it is to 1, the stronger the predictive ability of the model.

[0016] In step D, the target compound, target developing agent system and ratio to be predicted are first converted into an input information matrix and directly input into the trained integrated model to obtain the predicted ratio shift value.

[0017] As a further solution of the present invention, after predicting the ratio transfer value curve using the trained ratio transfer value prediction model, the model will give the optimal developing agent ratio scheme based on the curve, so that the ratio transfer value obtained in the experiment will not be too large or too small.

[0018] As a further solution of the present invention, for a variety of different developing agent systems, the developing agent information will be input into the prediction model in the form of a developing agent ratio vector.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] 1. The present invention establishes a transfer value prediction model through a machine learning method, which can quickly and accurately predict the transfer value curve of the target compound in the target developing agent system, and give the optimal developing agent ratio scheme according to the curve, so that the transfer value obtained in the experiment is not too large or too small, which greatly reduces the dependence of thin layer chromatography analysis technology on the experimenter's experience, solves the problem of developing agent selection, avoids boring repetitive experiments, greatly improves the efficiency of polarity determination, and saves time and labor costs.

[0021] 2. The present invention integrates XGBoost, LightGBM and neural network algorithms constrained by Sigmoid function, and utilizes the physical constraint that the range of the ratio shift value is between 0 and 1 and the technology of multi-algorithm integration to make the prediction model have stronger stability and prediction ability.

[0022] 3. The present invention can utilize new experimental data to periodically update and correct the original prediction model, effectively utilize massive polarity data, continuously enhance the generalization ability of the model, and avoid waste of resources.

[0023] 4. The present invention introduces machine learning technology into the field of experimental chemistry, making it possible to obtain the ratio value of a compound without experimental conditions. It has wide applications and important significance in the fields of experimental chemistry, drug synthesis and analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is an overall flow chart of the ratio shift value prediction method of the present invention.

[0025] Figure 2 The flowchart is a flow chart of data preprocessing and model training in the ratio shift value prediction method of the present invention.

[0026] Figure 3 This is a diagram of the overall prediction effect of an embodiment of the present invention on the prediction set.

[0027] Figure 4 It is a specific implementation effect diagram of the present invention, that is, the specific transfer value curve of the new compound in the target developing agent system is predicted by the trained specific transfer value prediction model. DETAILED DESCRIPTION

[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiment is only one embodiment of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0029] In general, the present invention proposes a method for predicting a ratio shift value based on machine learning, comprising the following steps:

[0030] 1) Data acquisition and cleaning

[0031] In this embodiment, the acquisition of the ratio shift value data is obtained by an automated thin layer chromatography analysis platform. Further, the artificial experimental data generated under a standardized experimental environment can also be obtained as valid data. Specifically, the data obtained include the chemical formula of the compound, the developing agent system and ratio used, and the ratio shift value under the system and ratio. In this embodiment, a total of 4928 data were obtained, including the ratio shift values ​​of 352 organic compounds under 2 developing agent systems, and the types of the organic compounds include ketones, heterocycles, ethers, halides, alcohols, etc. The developing agent system includes n-hexane (Hexane): ethyl acetate (EA) system and chloroform (DCM): methanol (MeOH) system. For the Hexane:EA system, there are 9 ratios, including 1:0, 50:1, 20:1, 5:1, 3:1, 1:1, 1:2 and 0:1. For the DCM:MeOH system, there are 5 ratios, including 1:0, 100:1, 50:1, 30:1, 20:1 and 10:1. Considering that there are some extreme cases in the experiment that lead to partial data loss, it is necessary to clean the data. The specific cleaning implementation method is:

[0032] a) If any of the chemical formula of the compound, the developing solvent system and ratio used, and the migration value under the system and ratio is missing, the data will be removed.

[0033] b) Remove data with illegal ratio shift values ​​determined experimentally, that is, data with ratio shift values ​​greater than 1 or less than 0.

[0034] c) Remove duplicate data and keep only one or the mean of the duplicate data.

[0035] After data cleaning, a total of 4114 data were qualified.

[0036] 2) Data preprocessing

[0037] In this embodiment, the cleaned data needs to be preprocessed to generate an input information matrix, such as Figure 2 As shown. Since the chemical formula cannot be directly input into the computer, this embodiment digitizes it into molecular fingerprints, molecular descriptors and dipole moments to characterize the structure and properties of the compound. Specifically, molecular fingerprints include MACCSkeys fingerprints, Morgan fingerprints, etc., which can be selected as appropriate; molecular descriptors include molecular weight, topological polar surface area, the number of rotatable bonds, the number of hydrogen bond donors, the number of hydrogen bond acceptors, and the value of lipid-water partition coefficient; dipole moment is the product of the distance between the positive and negative charge centers and the charge center, which is closely related to the polarity of the substance and is obtained by computational chemistry. In this embodiment, MACCSkeys molecular fingerprints and dipole moments are used, and molecular weight, topological polar surface area, the number of rotatable bonds, the number of hydrogen bond donors, the number of hydrogen bond acceptors, and the value of lipid-water partition coefficient are used as molecular descriptors.

[0038] Furthermore, since there are many commonly used developing agent systems, this embodiment uses a developing agent ratio vector to represent the system and ratio of the developing agent. Specifically, the types of developing agents include Hexane, EA, DCM, and MeOH, etc. Assuming that N developing agents are considered here, an N-dimensional vector is used to represent the proportional relationship between the developing agents. For example, [Hexane, EA, DCM, MeOH] = [0.75, 0.25, 0, 0] means that the system and ratio of the developing agent are Hexane: EA = 3:1. In this embodiment, after data preprocessing, the information vector of each data contains 178 dimensions, including 167 dimensions of molecular fingerprints, 4 dimensions of developing agent ratio vectors, and 7 dimensions of dipole moments and molecular descriptors. Overall, the entire data set is a 4114×178 information matrix.

[0039] Furthermore, the data will be stored in a proprietary compound polarity database after preprocessing, and the database will be automatically updated after periodically receiving new data.

[0040] 3) Model training

[0041] In this embodiment, an integrated algorithm of XGBoost, LightGBM and artificial neural network based on Sigmoid function constraints is used as the training model. Specifically, the Sigmoid function is:

[0042]

[0043] Considering the physical constraint that the range of the ratio shift value is between 0 and 1, and the range of the Sigmoid function is also (0,1), using the Sigmoid function as a physical constraint can improve the stability of model prediction. XGBoost, LightGBM and artificial neural networks can all be quickly implemented in Python and trained with training data. Suppose the trained models are g XGB ,g LGB and g ANN . Then the integrated model can be expressed as:

[0044]

[0045] Furthermore, during the training process, the training set and the validation set are divided, and the cross-validation method is used to prevent overfitting. During the training process of the neural network algorithm, early termination, dropout and other means are used to prevent overfitting and improve the performance of the model. In this embodiment, the ratios of 330 compounds in different developing solvent systems are randomly selected as the training set, with a total of 3372 data, the ratios of 20 compounds in different developing solvent systems are randomly selected as the validation set, with a total of 107 data, and the ratios of 10 compounds in different developing solvent systems are randomly selected as the test set, with a total of 112 data. The predictive ability of the model is determined by the coefficient of determination R of the model on the test set. 2 The calculation formula is as follows:

[0046]

[0047] Where N is the number of samples in the test set, y i is the true shift value, is the mean of the true ratio shift values, is the predicted ratio shift value. 2 The closer it is to 1, the stronger the prediction ability of the model. In this embodiment, the R 2 The value is 0.9469, which indicates that the model has strong predictive ability. On the test set, the scatter plot of the ratio shift value predicted by the ensemble model and the true ratio shift value is as follows: Figure 3 As shown, it can be seen that the prediction error of the model is very small, and the scatter points are basically on the y=x line.

[0048] Preferably, the selection of hyperparameters in the machine learning algorithm of the present invention can find the optimal parameter combination through grid-search.

[0049] 4) Model prediction

[0050] After the model training is completed, the trained model will be used for prediction. Specifically, for the target compound and the target developing agent system and ratio to be predicted, they are first converted into an input information matrix and directly input into the trained integrated model to obtain the predicted transfer ratio. In this embodiment, the transfer ratio curves of two new compounds in two different developing systems are shown in FIG. Figure 4 As shown. The triangle mark represents the experimental value, and the × mark represents the predicted value. This shows that for new compounds that the model has never seen, the prediction model can predict its specific transfer value curve in different development systems with high accuracy.

[0051] Furthermore, the model prediction supports the simultaneous input of multiple data, and the corresponding ratio shift value predictions will be obtained at the same time.

[0052] Furthermore, for the target compound, the prediction model can predict its specific transfer value curve under the target developing agent system, and provide guidance for configuring the developing agent according to the developing agent ratio corresponding to the curve specific transfer value between 0.3 and 0.7.

[0053] 5) Model update

[0054] In this embodiment, the prediction model will periodically add new data to the original model for update training to continuously improve the generalization ability of the model. Specifically, the source of the new data is experimental feedback and experimental data of the automated thin layer chromatography analysis platform. Furthermore, after each round of update, the test set will be used to evaluate the prediction effect of the model. If the evaluation is qualified, the update will be retained, and if it is unqualified, it will roll back to the version before the update. Specifically, when the coefficient of determination of the model prediction is greater than or equal to 85%, the model meets the standard, otherwise it does not meet the standard. In this embodiment, an update training is performed every time a database update is detected.

Claims

1. A method for predicting a ratio shift value based on machine learning, comprising the following steps: 1) Data collection and cleaning: Collect compound, developing agent and ratio transplantation data and perform data cleaning to obtain a complete, non-repetitive, and outlier-free polarity data set; 2) Data preprocessing: preprocessing the cleaned data, including: characterizing the molecular structure and properties of the compound by molecular fingerprints, dipole moments and molecular descriptors, and using a developing agent ratio vector to represent the developing agent system and ratio; the information vector of each piece of preprocessed data contains multiple dimensions, thereby generating an input information matrix; wherein, if the developing agent system involves N developing agents, an N-dimensional vector is used to represent the developing agent system and ratio; 3) Model training: The preprocessed data is divided into a training set and a validation set. The training set is trained using the XGBoost, LightGBM, and neural network algorithms constrained by the Sigmoid function. The three trained models are integrated into an integrated model to predict the ratio shift value. During the training process, the validation set is used to adopt the cross-validation method to prevent overfitting. 4) Model prediction: The specific transfer value prediction model trained in step 3) is used to predict the specific transfer value curve of the target compound in the target developing agent system.

2. The method for predicting ratio shift value according to claim 1, characterized in that: The data collected in step 1) include the chemical formula of the compound, the developing solvent system and ratio used, and the specific migration value of the compound under the developing solvent system and ratio; Data cleaning includes: a) if any of the chemical formula of the compound, the developing solvent system and ratio used, and the ratio shift value under the developing solvent system and ratio is missing, then remove the data; b) remove data with illegal ratio shift values ​​determined experimentally, that is, data with ratio shift values ​​greater than 1 or less than 0; c) remove duplicate data and only retain one or retain the average of the duplicate data.

3. The method for predicting ratio shift value according to claim 1, characterized in that: The molecular fingerprint in step 2) uses MACCSkeys fingerprint and / or Morgan fingerprint; the molecular descriptors include molecular weight, topological polar surface area, number of rotatable bonds, number of hydrogen bond donors, number of hydrogen bond acceptors, and lipid-water partition coefficient value; the dipole moment is the product of the distance between the positive and negative charge centers and the charge carried by the charge center.

4. The method for predicting ratio shift value according to claim 1, characterized in that: The Sigmoid function in step 3) is: The value range of Sigmoid function is (0,1); XGBoost, LightGBM and neural network algorithms are implemented in Python. The trained models are gXGB, g LGB and g ANN , then the integrated model is expressed as:

5. The method for predicting ratio shift value according to claim 1, characterized in that: In step 3), some data are randomly selected as the test set, and the predictive ability of the model is determined by the coefficient of determination R of the model on the test set. 2 The calculation formula is as follows: Where N is the number of samples in the test set, y i is the true shift value, is the mean of the true ratio shift values, is the predicted ratio shift value; R 2 The closer it is to 1, the stronger the predictive ability of the model.

6. The method for predicting ratio shift value according to claim 1, characterized in that: Step 4) For the target compound and target developing agent system and ratio to be predicted, they are first converted into an input information matrix and directly input into the trained integrated model to obtain the predicted ratio shift value.

7. The method for predicting ratio shift value according to claim 1, characterized in that: Step 4) After predicting the ratio transfer value curve using the trained ratio transfer value prediction model, an optimal developing agent ratio scheme is given according to the ratio transfer value curve.

8. The method for predicting ratio shift value according to claim 1, characterized in that: The ratio shift value prediction method also includes step 5) feedback update: automatically updating and correcting the ratio shift value prediction model according to the experimental feedback data.

Images