Chemcnet-based intelligent analysis method and system for organic synthesis

Abstract

The present application relates to the technical field of artificial intelligence in organic synthesis, in particular to an organic synthesis intelligent analysis method and system based on ChemCNet; the method comprises: 1) calculation of feature descriptors, converting chemical reaction conditions into feature descriptors according to software; 2) model establishment, constructing an attention-driven ChemCNet model combining deep learning and integrated boosting tree CatBoost, performing re-representation learning on the feature descriptors, intelligently predicting the yield, and analyzing the internal relationship between the yield and the reaction conditions; 3) visualizing and analyzing the ChemCNet model and its prediction results using feature importance and SHAP values. The method can reduce data redundancy, realize accurate prediction of chemical reaction yield, and conform to the concept of green chemistry and wisdom chemistry, realizing accurate prediction of yield on the basis of resource saving and intelligent analysis of the internal relationship between reaction conditions and yield.

Description

Technical Field

[0001] This invention relates to the field of organic synthesis technology in the field of artificial intelligence, and specifically to an intelligent analysis method and system for organic synthesis based on ChemCNet. Background Technology

[0002] Aromatic amines are fundamental products and building blocks in chemistry, and many active pharmaceutical ingredients contain aromatic amines, making them valuable targets in medicinal chemistry. Therefore, reactions involving the formation of aryl C-N bonds are increasingly used in the pharmaceutical field. C(sp2)-N bonds are ubiquitous in drug discovery, although classic strategies for forming these bonds, such as aromatic nucleophilic substitution reactions (S... N Ar reactions, Curtius rearrangements, and copper-catalyzed coupling reactions remain important shifts, but their limited scope restricts their widespread applicability. However, palladium-catalyzed Buchwald-Hartwig coupling reactions have dramatically changed this direction. This reaction is currently a highly efficient and versatile method for generating C-N bonds, and further optimization and development of this reaction hold great potential for application in drug development and large-scale production. In 2004, Robinson, GE, and others synthesized ZM549865, an effective orally active 5-HT1B receptor antagonist for treating depression and anxiety, based on a palladium-catalyzed Buchwald-Hartwig coupling synthesis strategy. In 2010, American scientist Richard Heck, Japanese scientists Eiichi Negishi, and Akira Suzuki were awarded the Nobel Prize in Chemistry for their outstanding contributions to the field of palladium-catalyzed cross-coupling reactions. Their work has wide applications in the pharmaceutical field, such as the preparation of antibiotics that work against drug-resistant bacteria and fluorescent labeling essential for DNA sequencing. In 2016, it was reported that at least 10% of medicinal chemistry papers published in 2014 used the Buchwald-Hartwig coupling reaction at least once. A research team at the Novartis Institute for Biomedical Research found in 2016 that the Buchwald-Hartwig coupling reaction is a commonly used reaction type by medicinal chemists over the past 40 years. Due to the large number of nitrogen-containing bioactive compounds, the integration of N-arylation reactions in the pharmaceutical industry is rapid. The simplicity of the synthetic procedure and the versatility of the resulting products also simplify the creation of compound libraries for medicinal chemistry, all of which demonstrate the importance of this reaction in pharmaceutical research and development.

[0003] However, this reaction route is complex, time-consuming, and costly. Traditional chemical experiments require extensive manual trial and error, consuming significant human, material, and financial resources. Furthermore, in the preparation of aromatic amines via the Buchwald-Hartwig coupling reaction, Pd metal is expensive and toxic, and the reaction may generate byproducts such as aromatic compounds, which are not only hazardous but also lead to low yields in the Buchwald-Hartwig coupling reaction. Therefore, utilizing artificial intelligence algorithms to accurately predict reaction yields, explore the internal relationship between reaction conditions and yields, improve production efficiency, save production costs, and simplify the large-scale manufacturing of special and complex materials is a highly meaningful endeavor.

[0004] With the rapid development of big data technology, chemical research is shifting towards a data-driven scientific discovery paradigm. Researchers use computer technology to perform calculations and encoding on chemical data, transforming it into computer-readable data forms, i.e., descriptors. Artificial intelligence algorithms excel at mining the inherent potential information within the vast amounts of experimental data generated in chemical reaction experiments, helping chemists make reasonable predictions and analyses, greatly improving the efficiency of chemical research and development, and driving the transformation and upgrading of traditional chemical research methods. In 2018, Ahneman et al. reported the prediction of reaction yields for the Buchwald-Hartwig coupling reaction using random forests, representing advanced research in the field of multidimensional chemical space prediction using machine learning methods. In 2021, Yanan Zhao et al. used a two-dimensional DCNN to predict reaction yields. However, Ahneman et al. used raw data as model input, resulting in high data dimensionality, redundancy, and computational cost; and their use of random forests as the prediction model lacked the feature learning concept of deep learning, failing to effectively mine deep features. Yanan Zhao et al. also used raw data, with a large number of network layers and nodes, making computation relatively complex, and fully connected layers were prone to overfitting. Therefore, the above methods all have different drawbacks to varying degrees. They cannot accurately uncover the correlation between reaction conditions and reaction yield in chemical experiments, resulting in a waste of human and experimental resources.

[0005] In summary, there is an urgent need for a method that can quickly and accurately uncover the correlation between reaction conditions and reaction yields in chemical experiments, avoid the waste of human and experimental resources, provide constructive advice to chemists, and promote the green and intelligent transformation and upgrading of chemical research. Summary of the Invention

[0006] To address the aforementioned technical problems, the present invention aims to provide an intelligent analysis method for organic synthesis based on ChemCNet, the specific technical solution of which is as follows:

[0007] 1) The calculation of the feature descriptors is based on the software converting the components of each coupling reaction into corresponding feature descriptors;

[0008] 2) The model establishment involves constructing the ChemCNet model to perform feature learning and intelligent regression prediction of yield for feature descriptors;

[0009] 3) Visualize and analyze the prediction results of the ChemCNet model using feature importance and SHAP value.

[0010] Preferably, step 1) includes:

[0011] (1.1) Import the chemical reactants and reagents into the software. The software automatically calculates the characteristic descriptor of each coupling reaction component and converts the chemical reactants into one-dimensional data.

[0012] (1.2) Combine RFE with CatBoost, use SHAP to score features and filter features, then combine the data to obtain a two-dimensional matrix and standardize the data.

[0013] Preferably, step (1.2) includes:

[0014] The RFE backward search method is selected, and the SHAP value is used as the feature evaluation standard to score the features. According to the principle of backward search, the feature with the lowest score is deleted, and then the model is built on the remaining features to obtain a new round of feature ranking. The feature with the lowest score is deleted again, and this process is repeated until the specified number of features is reached.

[0015] In each iteration, the set of remaining features is re-evaluated, and the score of each feature is adjusted during repeated iterations, ultimately presented as the model's prediction metric RMSE. After obtaining the selected features, these features are divided into training and test sets and matched with their corresponding yields. The data is then standardized. Finally, the selected dataset (1×24) is expanded to (1×25).

[0016] Preferably, step 2) includes:

[0017] (2.1) Construct the ChemCNet model; Construct an attention convolutional neural network, import the training set into the convolutional neural network for feature learning, mine deep features of the data, integrate attention into the convolutional neural network model, and calculate the loss function value of each iteration through continuous iterative learning until the loss function value is minimized, and save the trained model.

[0018] (2.2) Intelligent regression prediction of yield; The training set and the corresponding yield are imported into the convolutional neural network for feature learning. The output of the last fully connected layer is used as the input data of the CatBoost model for training and prediction, which is the final prediction result. The prediction effect of the model is evaluated by the coefficient of determination and the root mean square error.

[0019] (2.3) Perform out-of-sample prediction on the trained model. If the out-of-sample prediction is effective, the effectiveness and generalization ability of the trained model are verified.

[0020] (2.4) Users can adjust the parameters themselves according to the prediction effect and their own needs. If they are not satisfied, they can adjust the kernel size, number of layers, number of nodes and CatBoost related parameters of the convolutional neural network and return to step (2.2) until they are satisfied.

[0021] Preferably, step 3) includes:

[0022] (3.1) The importance ranking of the feature descriptors output by CatBoost is used to find the descriptors that have a significant impact on the reaction yield; the correlation between descriptors and reaction yield, as well as the internal relationship between the reaction yields of individual features, are analyzed using SHAP values ​​to provide users with decision-making information about organic chemical coupling reactions.

[0023] Preferably, step (2.1) includes:

[0024] (2.1.1) A convolutional neural network model is constructed by using an input layer, hidden layers and an output layer; the hidden layer includes 5 convolutional layers, each with the ReLU activation function and 3 fully connected layers;

[0025] (2.1.2) Add an attention layer after the last convolutional layer;

[0026] (2.1.3) Train the constructed model, set the number of iterations required for training the entire training set to p, and the amount of data read in at one time to q, where p≥1 and q≥1;

[0027] (2.1.4) Iterate the learning process multiple times and calculate the loss function value of the network model in each iteration. When the loss function value converges to the minimum, save the model parameters.

[0028] (2.1.5) The features extracted from the last fully connected layer are used as new output data and imported into the CatBoost model for training and prediction; this is the final prediction result; the coefficient of determination and root mean square error are used to evaluate the prediction performance of the model;

[0029] The attention layer uses ECA-Net; let the output of a convolutional block be X∈R. W×H×CWhere W, H, and C are the width, height, and channel dimensions, respectively, and R is the real number field; the non-dimensionality-reduced aggregated feature y∈R C Where R is the real number field; C is the channel dimension; the ECA module uses a frequency band matrix W. k Note when accessing the learning channel:

[0030]

[0031] This involves k×C parameters, y i The weight calculation only considers y i The interaction between y and its k neighbors, i.e. i The weights are: Where, ω i For y i The weights; σ is the Sigmoid function; For y i The interaction between it and its k neighbors (e.g., w) 1,1 ,...,w 1,k ); For y i The j-th adjacent channel; For y i The set of k adjacent channels;

[0032] Make all channels share the same learning parameters, that is:

[0033]

[0034] Where, ω i For y i The weights; σ is the Sigmoid function; For y i The interaction between it and its k neighbors (e.g., w) 1,1 ,...,w 1,k ); For y i The j-th adjacent channel; For y i The set of k adjacent channels;

[0035] Sharing the same learning parameters across all channels can be achieved using a one-dimensional convolution with a kernel size of k, i.e.:

[0036] ω=σ(C1D k (y))

[0037] Where C1D is a one-dimensional convolution; σ is the Sigmoid function;

[0038] Determine the kernel size k of the one-dimensional convolution; there exists a mapping φ between k and C: C = φ(k); the mapping is a linear function, i.e., φ(k) = γ * kb; γ is the coefficient of the first-order term in the linear function, and b is the constant term; then extend the linear function φ(k) = γ * kb to a nonlinear function, i.e.:

[0039] C=φ(k=2 (γ*k-b)

[0040] Then, given the channel dimension C, the kernel size k can be adaptively determined:

[0041]

[0042] in, express The closest odd number;

[0043] Finally, a fully connected layer is used to synthesize the extracted features; then, the features extracted by the fully connected layer are used as input data to the CatBoost model for regression prediction.

[0044] The CatBoost model uses the following approach: for each sample X k Train a single model M k The model consists of samples X that are not included. k The model was trained on the training set and used to estimate X. k The gradient on the sample is calculated, and this estimate is used to score the resulting tree; that is, the gradient of the model on the new sample is updated using the current model trained on the previous samples, providing an unbiased gradient; at each step t of the learning process, each model is interpreted as a model F. t Approximate value;

[0045] CatBoost first generates s+1 random sequences σ0, σ1, ..., σ1 for the training samples. s , where σ1,...,σ s It is used to construct a decision tree, σ0 is used to select the value of the leaf node, then an unbiased estimate of the gradient step size is used, and then standard GBDT is performed; in addition, CatBoost uses a symmetric tree structure as the basic predictor and uses the same splitting criterion throughout the tree.

[0046] This invention also provides an intelligent analysis system for organic synthesis based on ChemCNet, the system comprising:

[0047] The feature descriptor input and preprocessing module is used to convert the components of each coupling reaction into corresponding feature descriptors according to the software.

[0048] The intelligent analysis module based on the ChemCNet system is used to build ChemCNet models to perform feature learning and intelligent yield prediction for feature descriptors.

[0049] The visualization-based results interpretability analysis module is used to visualize and analyze the prediction results of the ChemCNet model using features such as importance and SHAP values.

[0050] The embodiments of the present invention have at least the following beneficial effects:

[0051] This invention addresses the lack of in-depth feature mining in traditional machine learning for predicting the yield of coupling reactions, proposing an intelligent analysis method and system for organic synthesis based on ChemCNet. This method incorporates deep learning's feature learning approach, leveraging attention-driven neural networks to deeply mine the essential features of the data. It replaces the ordinary fully connected output layer with a high-performance ensemble tree regressor, improving the model's prediction accuracy while avoiding overfitting. Users can adjust parameters to achieve optimal prediction results; the tree model obtains the importance ranking of feature descriptors, identifying those with significant yield impacts; and SHAP value analysis reveals the correlation between descriptors and reaction yields, as well as the internal relationships between individual feature reaction yields, providing users with comprehensive and multi-faceted decision-making information regarding organic chemical coupling reactions. This invention effectively assists chemists in analysis and prediction, achieving accurate yield predictions while enhancing the readability of the results. This invention is simple to operate and easy to implement, allowing users to conveniently and quickly obtain relatively accurate analysis results.

[0052] Furthermore, the ChemCNet model constructed in this invention improves prediction accuracy and enriches the semantic information of the network by adding attention to focus on key features without significantly increasing model complexity. Attached Figure Description

[0053] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a flowchart of the analytical method of the present invention;

[0055] Figure 2 The chemical reaction formulas and reaction components in the embodiments of the present invention are shown below;

[0056] Figure 2The symbols in the text are: Equation:Buchwald-Hartwig: coupling reaction and reaction components, Aryl: halide, Ligand: ligand, Base: substrate, Additive: additive. Detailed Implementation

[0057] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following detailed description, in conjunction with the accompanying drawings and preferred embodiments, provides a detailed explanation of the specific implementation methods, structures, features, and effects of the solutions proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0059] Please see Figure 1 The diagram illustrates a method flowchart provided by an embodiment of the present invention, which includes the following steps:

[0060] 1) The calculation of the feature descriptor is based on the calculation of the feature descriptor of each coupling reaction component by chemical software, and then converts it into one-dimensional experimental data.

[0061] Specifically, step 1) includes the following steps:

[0062] (1.1) Import the chemical reactants and reagents into the chemical software. The software automatically calculates the characteristic descriptor of each coupling reaction component and converts the chemical reactants into one-dimensional data.

[0063] In this embodiment, Figure 2 All reaction components of the Buchwald-Hartwig coupling reaction (including 23 additives, 15 halides, 3 substrates, and 4 ligands) were imported into chemical software. The software automatically calculated and extracted one-dimensional feature descriptors for each reaction component, resulting in 120 feature descriptors. This converted the chemical reactants into one-dimensional data. After removing invalid reactions and filling in missing values ​​with averages, a total of 3960 reaction data sets were obtained as experimental data.

[0064] (1.2) Combine RFE with CatBoost, use SHAP to score features and filter features, then combine the data to obtain a two-dimensional matrix and standardize the data.

[0065] Preferably, the RFE backward search method is selected, and the SHAP value is used as the feature evaluation standard to score the features. According to the principle of backward search, the feature with the lowest feature score is deleted, and then the model is built on the remaining features to obtain a new round of feature ranking. The feature with the lowest score is deleted again, and this process is repeated until the specified number of features is reached. The specified number of features is 24 in this embodiment, and the implementer can adjust it according to the specific situation.

[0066] In each iteration, the set of remaining features is re-evaluated, and the score of each feature is adjusted during repeated iterations, ultimately presented as the model's prediction metric, RMSE. After obtaining the selected features, these features are divided into training and test sets. In this embodiment, the feature descriptors are divided into training and test sets in a 7:3 ratio. In practice, the implementer can adjust the ratio and match it with the corresponding yield. The data is then standardized. The resulting dataset (1×24) is then expanded to (1×25); that is, any one dimension is randomly selected from the (1×24) dataset, copied, and placed in the last column to complete the dataset expansion. The purpose of expanding the dataset is to convert the descriptor set into the input format of the convolutional neural network so that its size can be adjusted later.

[0067] It should be noted that this invention believes there is a very close relationship between the model and the features. In order to maximize the model's benefits and reduce data redundancy, a feature selection algorithm related to the model is adopted to obtain a set of concise and comprehensive feature descriptors. In order to conform to the input format of the network, the data is combined to obtain a two-dimensional matrix and the data is standardized.

[0068] 2) The model is built by constructing the ChemCNet model to perform feature learning and intelligent regression prediction of yield for feature descriptors.

[0069] Specifically, step 2) includes:

[0070] (2.1) Construct the ChemCNet model; that is, construct an attention convolutional neural network according to the characteristics of the data, import the training set into the convolutional neural network for feature learning, deeply explore the deep features of the data, integrate attention into the convolutional neural network model, and calculate the loss function value of each iteration through continuous iterative learning until the loss function value is minimized, and save the trained model.

[0071] Preferably, step (2.1) specifically includes:

[0072] (2.1.1) A convolutional neural network model is constructed by input layer, hidden layer and output layer; the hidden layer includes 5 convolutional layers, each with ReLU activation function and 3 fully connected layers.

[0073] It should be noted that since each descriptor contains specific, non-negligible chemical information, it is necessary to retain as much data as possible. Therefore, based on the actual background meaning of the data, this invention chooses to remove the pooling layer. That is, a convolutional neural network model is built by simply stacking the input layer, hidden layer, and output layer; the hidden layer includes 5 convolutional layers, each with the ReLU activation function and 3 fully connected layers.

[0074] (2.1.2) Add an attention layer (ECA-Net) after the last convolutional layer. The purpose of adding the attention layer is to enable the model to focus on information at different scales and key features when extracting information.

[0075] (2.1.3) Import the data obtained in step (1.2) into the ChemCNet model; train the constructed model, set the number of iterations required for training the entire training set to p, and the amount of data read in at one time to q, where p≥1 and q≥1.

[0076] (2.1.4) Iterate the learning process multiple times and calculate the loss function value of the network model in each iteration. In this embodiment, the loss function is the mean squared error loss function. When the loss function value converges to the minimum, save the model parameters.

[0077] (2.1.5) The features extracted from the last fully connected layer are used as new output data and imported into the CatBoost model for training and prediction; this is the final prediction result. The coefficient of determination (R-squared, R0) is used. 2 The predictive performance of the model is evaluated using the root mean square error (RMSE).

[0078] The attention layer employs ECA-Net; after channel-level global average pooling without reducing dimensionality, ECA captures local cross-channel interaction information by considering each channel and its k neighbors; let the output of a convolutional block be X∈R. W ×H×C Where W, H, and C are the width, height, and channel dimensions, respectively; the channel dimension, also known as the number of channels, represents the number of filters; R is the real number field; and the non-dimensionality-reduced aggregated feature y∈R C Where R is the real number field; C is the channel dimension; the ECA module uses a band matrix W. k Note when accessing the learning channel:

[0079]

[0080] This involves k×C parameters and avoids complete independence between different groups in the equation, y i The weight calculation only considers y i The interaction between y and its k neighbors, i.e. i The weights are: Where, ω i For y i The weights; σ is the Sigmoid function; For y i The interaction between it and its k neighbors (e.g., w) 1 ,1 ,...,w 1,k ); For y i The j-th adjacent channel; For y i The set of k adjacent channels.

[0081] A more efficient approach is to have all channels share the same learning parameters, i.e.:

[0082]

[0083] Where, ω i For y i The weights; σ is the Sigmoid function; For y i The interaction between it and its k neighbors (e.g., w) 1,1 ,...,w 1,k ); For y i The j-th adjacent channel; For y i The set of k adjacent channels.

[0084] Sharing the same learning parameters across all channels can be achieved using a one-dimensional convolution with a kernel size of k, i.e.:

[0085] ω=σ(C1D k (y))

[0086] Where C1D is a one-dimensional convolution; σ is the Sigmoid function.

[0087] Since the ECA module aims to appropriately capture local cross-channel interactions, it is necessary to determine the coverage of the interactions (i.e., the kernel size k of the one-dimensional convolution). Clearly, k is related to the size of the channel C. Generally, larger sizes are better at capturing long-range dependencies, while smaller sizes tend to capture short-range interactions. Therefore, there exists a mapping φ: C = φ(k) between k and C. The simplest mapping is a linear function, i.e., φ(k) = γ*kb. However, relationships characterized by linear functions are too limited; on the other hand, it is well known that the channel dimension C is usually set to a power of 2. Therefore, this embodiment extends the linear function φ(k) = γ*kb to a non-linear function, introducing a possible solution, namely:

[0088] C=φ(k=2 (γ*k-b)

[0089] Then, given the channel dimension C, the kernel size k can be adaptively determined:

[0090]

[0091] in, express The closest odd number;

[0092] Finally, a fully connected layer is used to synthesize the extracted features; then, the features extracted by the fully connected layer are used as input data to the CatBoost model for regression prediction.

[0093] The CatBoost model uses the following approach: for each sample X k Train a single model M k The model consists of samples X that are not included. k The model was trained on the training set and used to estimate X. k The gradient on the sample is calculated, and this estimate is used to score the resulting tree; that is, the gradient of the model on the new sample is updated using the current model trained on the previous samples, providing an unbiased gradient; at each step t of the learning process, each model is interpreted as a model F. t Approximate value;

[0094] The CatBoost model, drawing inspiration from online learning algorithms, introduces an "artificial timeline"—a timeline based on the arrival time of training examples. This ensures that only previously seen examples are used when calculating statistical data, effectively preventing data leakage. Furthermore, CatBoost employs several permutations to enhance the algorithm's robustness.

[0095] CatBoost first generates s+1 random sequences σ0, σ1, ..., σ1 for the training samples. s , where σ1,...,σs It is used to construct a decision tree, σ0 is used to select the value of the leaf node, then an unbiased estimate of the gradient step size is used, and then standard GBDT is performed; in addition, CatBoost uses a symmetric tree structure as the basic predictor and uses the same splitting criterion throughout the tree.

[0096] It's important to note that in the GBDT framework, the decision tree construction process typically involves two stages: selecting the tree structure (i.e., splitting attributes) and calculating the leaf nodes. To select the optimal tree structure, this process enumerates different splits, constructs the tree using these splits, obtains the leaf node values, scores the tree, and selects the best split. CatBoost, on the other hand, first generates s+1 random sequences σ0, σ1, ..., σ... for the training samples. s , where σ1,...,σ s The algorithm is used to construct the decision tree, with σ0 used to select the values ​​of the leaf nodes. An unbiased estimate of the gradient step size is then applied, followed by standard GBDT. Furthermore, CatBoost uses a symmetric tree structure as the basic predictor; the same splitting criterion is used throughout the tree. This type of tree is balanced, less prone to overfitting, and significantly speeds up predictions during test time.

[0097] (2.2) Intelligent regression prediction of yield; The training set and corresponding yield are imported into a convolutional neural network for feature learning. The output of the last fully connected layer is used as the input data for training and prediction of the CatBoost model, which is the final prediction result; The coefficient of determination (R-squared, R) is used. 2 The predictive performance of the model is evaluated using the root mean square error (RMSE).

[0098] (2.3) Perform out-of-sample prediction on the trained model. If the out-of-sample prediction is effective, the effectiveness and generalization of the trained model are verified.

[0099] In this embodiment, a portion of the additives ( Figure 2 Out-of-sample predictions were performed on the 15th, 18th, 19th, 21st, and 22nd samples.

[0100] Specifically, the out-of-sample prediction involves dividing the dataset into two disjoint parts: one as a training set for model estimation and the other as a test set for prediction, thereby testing the model's generalization ability. This invention randomly selects five additives as unknown reaction conditions and uses the remaining known reaction conditions as training data to predict the yield of the unknown reaction conditions. Compared with the out-of-sample prediction results using random forests by Ahneman et al., this invention has a larger R-value. 2The smaller RMSE indicates that the method achieves better out-of-sample prediction performance, and the generalization ability of the model designed in this invention is improved.

[0101] (2.4) Users can adjust the parameters themselves according to the prediction effect and their own needs. If they are not satisfied, they can adjust the kernel size, number of layers, number of nodes and CatBoost related parameters of the convolutional neural network and return to step (2.2) until they are satisfied.

[0102] 3) Visualize and analyze the prediction results of the ChemCNet model using feature importance and SHAP value.

[0103] Specifically, step 3) includes:

[0104] (3.1) The importance ranking of the feature descriptors output by CatBoost is used to find the descriptors that have a significant impact on the reaction yield; the correlation between descriptors and reaction yield, as well as the internal relationship between the reaction yields of single features, are analyzed using SHAP values ​​to provide users with decision-making information about organic chemical coupling reactions; it can provide users with decision-making information about organic chemical coupling reactions from all aspects and multiple perspectives.

[0105] Simulation experiment:

[0106] The method of this invention is further demonstrated through simulation experiments, taking the Buchwald-Hartwig coupling reaction as an example (chemical reaction equation and reaction components are as follows). Figure 2 As shown in the figure, the descriptors for each reaction component were calculated and extracted using Spartan software. After calculation, 120 feature descriptors were obtained for each group of reactions, including 64 atomic descriptors, 28 molecular descriptors and 28 vibrational descriptors.

[0107] One-dimensional feature descriptors and two-dimensional feature descriptor matrices after feature preprocessing, along with their corresponding yields, were imported into some common machine learning models and the ChemCNet model for regression prediction; the simulation results are shown in Table 1.

[0108] Table 1

[0109]

[0110] Table 1 shows the comparison of ChemCNet's prediction accuracy with several machine learning algorithms: CatBoost, GBDT (Gradient Boosting Decision Tree), Random Forest, Decision Tree, and Linear Regression. The experimental results show that the ChemCNet model achieves higher R-values. 2 =0.97, RMSE=4.88, indicating that ChemCNet's regression prediction results are optimal and have the highest degree of fit to the true values. In summary, the regression prediction results of the ChemCNet model are superior to general machine learning algorithms, and can achieve relatively accurate predictions of reaction yields.

[0111] In summary, this invention organically combines neural networks with ensemble tree models and adds attention to further optimize the network structure. Users can adjust the parameters themselves and test them on a test set. If the test results are satisfactory, parameter tuning can be stopped, and the prediction results can be output.

[0112] The ChemCNet model offers more accurate predictions compared to traditional machine learning algorithms. It combines deep learning's feature learning approach with attention functionality, focusing on key features without significantly increasing model complexity. Furthermore, it utilizes an ensemble learning model as a regressor, improving prediction accuracy and generalization ability. ChemCNet replaces traditional fully connected layers with an ensemble tree model, effectively avoiding overfitting. Compared to typical deep neural networks, ChemCNet has a simpler structure, reducing operational difficulty and cost, and minimizing reliance on hardware. Unlike typical deep learning algorithms, ChemCNet also performs well on small sample sizes.

[0113] This invention also provides an intelligent analysis system for organic synthesis based on ChemCNet, the system comprising:

[0114] The feature descriptor input and preprocessing module is used to calculate the feature descriptor corresponding to each coupling reaction component based on the software.

[0115] The intelligent analysis module based on the ChemCNet system is used to build ChemCNet models to perform feature learning and intelligent regression prediction of yield for feature descriptors.

[0116] The visualization-based results interpretability analysis module is used to visualize and analyze the prediction results of the ChemCNet model using feature importance and SHAP values.

[0117] The implementation of an intelligent organic synthesis analysis system based on ChemCNet is the same as that of an intelligent organic synthesis analysis method based on ChemCNet. Since the specific implementation of the intelligent organic synthesis analysis method based on ChemCNet has been detailed in steps 1) to 3) above, it will not be repeated further.

[0118] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 of the technical features. Such 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 this application, and should all be included within the protection scope of this application.

Claims

1. A smart analysis method for organic synthesis based on ChemCNet, characterized in that, The method includes the following steps: 1) The calculation of the feature descriptors is based on the software converting the components of each coupling reaction into corresponding feature descriptors; 2) The model establishment involves constructing the ChemCNet model to perform feature learning and intelligent regression prediction of yield for feature descriptors; 3) Visualize and analyze the prediction results of the ChemCNet model using feature importance and SHAP values; Step 2) includes: (2.1) Construct the ChemCNet model; Construct an attention convolutional neural network, import the training set into the convolutional neural network for feature learning, mine deep features of the data, integrate attention into the convolutional neural network model, and calculate the loss function value of each iteration through continuous iterative learning until the loss function value is minimized, and save the trained model. (2.2) Intelligent regression prediction of yield; The training set and the corresponding yield are imported into the convolutional neural network for feature learning. The output of the last fully connected layer is used as the input data of the CatBoost model for training and prediction, which is the final prediction result. The prediction effect of the model is evaluated by the coefficient of determination and the root mean square error. (2.3) Perform out-of-sample prediction on the trained model. If the out-of-sample prediction is effective, the effectiveness and generalization ability of the trained model are verified. (2.4) Users can adjust the parameters themselves according to the prediction effect and their own needs. If they are not satisfied, they can adjust the kernel size, number of layers, number of nodes and CatBoost related parameters of the convolutional neural network and return to step (2.2) until they are satisfied.

2. The intelligent analysis method for organic synthesis based on ChemCNet according to claim 1, characterized in that, Step 1) includes: (1.1) Import the chemical reactants and reagents into the software. The software automatically calculates the characteristic descriptor of each coupling reaction component and converts the chemical reactants into one-dimensional data. (1.2) Combine RFE with CatBoost, use SHAP to score features and filter features, then combine the data to obtain a two-dimensional matrix and standardize the data.

3. The intelligent analysis method for organic synthesis based on ChemCNet according to claim 2, characterized in that, Step (1.2) includes: The RFE backward search method is selected, and the SHAP value is used as the feature evaluation standard to score the features. According to the principle of backward search, the feature with the lowest score is deleted, and then the model is built on the remaining features to obtain a new round of feature ranking. The feature with the lowest score is deleted again, and this process is repeated until the specified number of features is reached. In each iteration, the set of remaining features is re-evaluated, and the score of each feature is adjusted during repeated iterations, ultimately presented in the form of the model's prediction metric RMSE. After obtaining the filtered features, these features are divided into training and test sets and matched with the corresponding yields. The data is then standardized. Finally, the filtered dataset (1×24) is expanded to (1×25).

4. The intelligent analysis method for organic synthesis based on ChemCNet according to claim 1, characterized in that, Step 3) includes: (3.1) The importance ranking of the feature descriptors output by CatBoost is used to find the descriptors that have a significant impact on the reaction yield; the SHAP value is used to analyze the correlation between the descriptors and the reaction yield, as well as the internal relationship between the reaction yields of individual features, to provide users with decision-making information about organic chemical coupling reactions.

5. The intelligent analysis method for organic synthesis based on ChemCNet according to claim 1, characterized in that, Step (2.1) includes: (2.1.1) A convolutional neural network model is constructed by using an input layer, hidden layers and an output layer; the hidden layer includes 5 convolutional layers, each with the ReLU activation function and 3 fully connected layers; (2.1.2) Add an attention layer after the last convolutional layer; (2.1.3) Train the constructed model, set the number of iterations required for training the entire training set to p, and the amount of data read in at one time to q, where p≥1 and q≥1; (2.1.4) Iterate the learning process multiple times and calculate the loss function value of the network model in each iteration. When the loss function value converges to the minimum, save the model parameters. (2.1.5) The features extracted from the last fully connected layer are used as new output data and imported into the CatBoost model for training and prediction; this is the final prediction result; the coefficient of determination and root mean square error are used to evaluate the prediction performance of the model; The attention layer uses ECA-Net; the output of a convolutional block is set to... Where W, H, and C are the width, height, and channel dimensions, respectively, and R is the real number field; non-dimensionality reduction aggregated features. Where R is the real number field; C is the channel dimension; the ECA module uses a frequency band matrix. Note when accessing the learning channel: This involves k×C parameters. The weight calculation only considers The interaction between it and its k neighbors, i.e. The weights are: ;in, for The weights; For the Sigmoid function; for The interaction between it and its k neighbors; for The j-th adjacent channel; for The set of k adjacent channels; Make all channels share the same learning parameters, that is: in, for The weights; For the Sigmoid function; for The interaction between it and its k neighbors; for The j-th adjacent channel; for The set of k adjacent channels; Sharing the same learning parameters across all channels can be achieved using a one-dimensional convolution with a kernel size of k, i.e.: Wherein, C1D is a one-dimensional convolution; For the Sigmoid function; Determine the kernel size k of the one-dimensional convolution; there exists a mapping between k and C. The mapping is a linear function, that is... ; γ is the coefficient of the linear term in the linear function, and b is the constant term; then the linear function Extending this to nonlinear functions, that is: Then, given the channel dimension C, the kernel size k can be adaptively determined: in, express The closest odd number; Finally, a fully connected layer is used to synthesize the extracted features; then, the features extracted by the fully connected layer are used as input data to the CatBoost model for regression prediction. The CatBoost model uses the following approach: for each sample Train a single model The model consists of samples that are not included. The training set was used to train the model, and the model was used to estimate... The gradient on the sample is calculated, and this estimate is used to score the resulting tree; that is, the gradient of the model on the new sample is updated using the current model trained on the previous samples, providing an unbiased gradient; at each step t of the learning process, each model is interpreted as a model Approximate value; CatBoost first generates s+1 random sequences for the training samples. ,in Used to build decision trees The values ​​of the leaf nodes are selected, and then an unbiased estimate of the gradient step size is used, followed by standard GBDT. In addition, CatBoost uses a symmetric tree structure as the basic predictor and uses the same splitting criterion throughout the tree.

6. An intelligent analysis system for organic synthesis based on ChemCNet, characterized in that, The system includes, The feature descriptor input and preprocessing module is used to convert the components of each coupling reaction into corresponding feature descriptors according to the software. The intelligent analysis module based on the ChemCNet system is used to build ChemCNet models to perform feature learning and intelligent yield prediction for feature descriptors. The visualization-based results interpretability analysis module is used to visualize and analyze the prediction results of the ChemCNet model using features such as importance and SHAP values. Step 2) includes: (2.1) Construct the ChemCNet model; Construct an attention convolutional neural network, import the training set into the convolutional neural network for feature learning, mine deep features of the data, integrate attention into the convolutional neural network model, and calculate the loss function value of each iteration through continuous iterative learning until the loss function value is minimized, and save the trained model. (2.2) Intelligent regression prediction of yield; The training set and the corresponding yield are imported into the convolutional neural network for feature learning. The output of the last fully connected layer is used as the input data of the CatBoost model for training and prediction, which is the final prediction result. The prediction effect of the model is evaluated by the coefficient of determination and the root mean square error. (2.3) Perform out-of-sample prediction on the trained model. If the out-of-sample prediction is effective, the effectiveness and generalization ability of the trained model are verified. (2.4) Users can adjust the parameters themselves according to the prediction effect and their own needs. If they are not satisfied, they can adjust the kernel size, number of layers, number of nodes and CatBoost related parameters of the convolutional neural network and return to step (2.2) until they are satisfied.

Images